We evaluate the performance disparity of the Whisper and MMS families of ASR models across the VoxPopuli and Common Voice multilingual datasets, with an eye toward intersectionality. Our two most important findings are that model size, surprisingly, correlates logarithmically with worst-case performance disparities, meaning that larger (and better) models are less fair. We also observe the importance of intersectionality. In particular, models often exhibit significant performance disparity across binary gender for adolescents.
The performance cost of differential privacy has, for some applications, been shown to be higher for minority groups; fairness, conversely, has been shown to disproportionally compromise the privacy of members of such groups. Most work in this area has been restricted to computer vision and risk assessment. In response, we evaluate the impact of differential privacy on fairness across four diverse tasks, focusing on how attempts to mitigate privacy violations and between-group performance differences interact: Does privacy inhibit attempts to ensure fairness? To this end, we train (𝜀,𝛿)-differentially private models with empirical risk minimization and group distributionally robust training objectives. Consistent with previous findings, we find that differential privacy increases between-group performance differences in the baseline setting; more interestingly, differential privacy reduces between-group performance differences in the robust setting. We explain this by interpreting differential privacy as regularization.
Facts are subject to contingencies and can be true or false in different circumstances. One such contingency is time, wherein some facts mutate over a given period, e.g., the president of a country or the winner of a championship. Trustworthy language models ideally identify mutable facts as such and process them accordingly. We create MuLan, a benchmark for evaluating the ability of English language models to anticipate time-contingency, covering both 1:1 and 1:N relations. We hypothesize that mutable facts are encoded differently than immutable ones, hence being easier to update. In a detailed evaluation of six popular large language models, we consistently find differences in the LLMs’ confidence, representations, and update behavior, depending on the mutability of a fact. Our findings should inform future work on the injection of and induction of time-contingent knowledge to/from LLMs.
Large Vision Language Models can be used to assist visually impaired individuals by describing images they capture in their daily lives. Current evaluation datasets may not reflect the diverse cultural user backgrounds nor the situational context of this use case. To address this problem, we create a survey to determine caption preferences and propose a culture-centric evaluation benchmark by filtering VizWiz, an existing dataset with images taken by people who are blind. We then evaluate different models and prompts, investigating their reliability as visual assistants. While the evaluation results for state-of-the-art models seem promising, we identified some weak spots such as hallucinations and problems with conventional evaluation metrics. Our survey, data, code, and model outputs will be publicly available.
Rationales in the form of manually annotated input spans usually serve as ground truth when evaluating explainability methods in NLP. They are, however, time-consuming and often biased by the annotation process. In this paper, we debate whether human gaze, in the form of webcam-based eye-tracking recordings, poses a valid alternative when evaluating importance scores. We evaluate the additional information provided by gaze data, such as total reading times, gaze entropy, and decoding accuracy with respect to human rationale annotations. We compare WebQAmGaze, a multilingual dataset for information-seeking QA, with attention and explainability-based importance scores for 4 different multilingual Transformer-based language models (mBERT, distil-mBERT, XLMR, and XLMR-L) and 3 languages (English, Spanish, and German). Our pipeline can easily be applied to other tasks and languages. Our findings suggest that gaze data offers valuable linguistic insights that could be leveraged to infer task difficulty and further show a comparable ranking of explainability methods to that of human rationales.
Explainability methods are used to benchmark the extent to which model predictions align with human rationales i.e., are ‘right for the right reasons’. Previous work has failed to acknowledge, however, that what counts as a rationale is sometimes subjective. This paper presents what we think is a first of its kind, a collection of human rationale annotations augmented with the annotators demographic information. We cover three datasets spanning sentiment analysis and common-sense reasoning, and six demographic groups (balanced across age and ethnicity). Such data enables us to ask both what demographics our predictions align with and whose reasoning patterns our models’ rationales align with. We find systematic inter-group annotator disagreement and show how 16 Transformer-based models align better with rationales provided by certain demographic groups: We find that models are biased towards aligning best with older and/or white annotators. We zoom in on the effects of model size and model distillation, finding –contrary to our expectations– negative correlations between model size and rationale agreement as well as no evidence that either model size or model distillation improves fairness.
Humans can effortlessly understand the coordinate structure of sentences such as “Niels Bohr and Kurt Cobain were born in Copenhagen and Seattle, *respectively*”. In the context of natural language inference (NLI), we examine how language models (LMs) reason with respective readings (Gawron and Kehler, 2004) from two perspectives: syntactic-semantic and commonsense-world knowledge. We propose a controlled synthetic dataset WikiResNLI and a naturally occurring dataset NatResNLI to encompass various explicit and implicit realizations of “respectively”. We show that fine-tuned NLI models struggle with understanding such readings without explicit supervision. While few-shot learning is easy in the presence of explicit cues, longer training is required when the reading is evoked implicitly, leaving models to rely on common sense inferences. Furthermore, our fine-grained analysis indicates models fail to generalize across different constructions. To conclude, we demonstrate that LMs still lag behind humans in generalizing to the long tail of linguistic constructions.
In this work, we conduct a detailed analysis on the performance of legal-oriented pre-trained language models (PLMs). We examine the interplay between their original objective, acquired knowledge, and legal language understanding capacities which we define as the upstream, probing, and downstream performance, respectively. We consider not only the models’ size but also the pre-training corpora used as important dimensions in our study. To this end, we release a multinational English legal corpus (LeXFiles) and a legal knowledge probing benchmark (LegalLAMA) to facilitate training and detailed analysis of legal-oriented PLMs. We release two new legal PLMs trained on LeXFiles and evaluate them alongside others on LegalLAMA and LexGLUE. We find that probing performance strongly correlates with upstream performance in related legal topics. On the other hand, downstream performance is mainly driven by the model’s size and prior legal knowledge which can be estimated by upstream and probing performance. Based on these findings, we can conclude that both dimensions are important for those seeking the development of domain-specific PLMs.
Machine translation (MT) operates on the premise of an interlingua which abstracts away from surface form while preserving meaning. A decade ago the idea of using round-trip MT to guide grammatical error correction was proposed as a way to abstract away from potential errors in surface forms (Madnani et al., 2012). At the time, it did not pan out due to the low quality of MT systems of the day. Today much stronger MT systems are available so we re-evaluate this idea across five languages and models of various sizes. We find that for extra large models input augmentation through round-trip MT has little to no effect. For more ‘workable’ model sizes, however, it yields consistent improvements, sometimes bringing the performance of a base or large model up to that of a large or xl model, respectively. The round-trip translation comes at a computational cost though, so one would have to determine whether to opt for a larger model or for input augmentation on a case-by-case basis.
Over the years, many researchers have seemingly made the same observation: Brain and language model activations exhibit some structural similarities, enabling linear partial mappings between features extracted from neural recordings and computational language models. In an attempt to evaluate how much evidence has been accumulated for this observation, we survey over 30 studies spanning 10 datasets and 8 metrics. How much evidence has been accumulated, and what, if anything, is missing before we can draw conclusions? Our analysis of the evaluation methods used in the literature reveals that some of the metrics are less conservative. We also find that the accumulated evidence, for now, remains ambiguous, but correlations with model size and quality provide grounds for cautious optimism.
Language models may memorize more than just facts, including entire chunks of texts seen during training. Fair use exemptions to copyright laws typically allow for limited use of copyrighted material without permission from the copyright holder, but typically for extraction of information from copyrighted materials, rather than verbatim reproduction. This work explores the issue of copyright violations and large language models through the lens of verbatim memorization, focusing on possible redistribution of copyrighted text. We present experiments with a range of language models over a collection of popular books and coding problems, providing a conservative characterization of the extent to which language models can redistribute these materials. Overall, this research highlights the need for further examination and the potential impact on future developments in natural language processing to ensure adherence to copyright regulations. Code is at https://github.com/coastalcph/CopyrightLLMs.
Van Miltenburg et al. (2021) suggest NLP research should adopt preregistration to prevent fishing expeditions and to promote publication of negative results. At face value, this is a very reasonable suggestion, seemingly solving many methodological problems with NLP research. We discuss pros and cons - some old, some new: a) Preregistration is challenged by the practice of retrieving hypotheses after the results are known; b) preregistration may bias NLP toward confirmatory research; c) preregistration must allow for reclassification of research as exploratory; d) preregistration may increase publication bias; e) preregistration may increase flag-planting; f) preregistration may increase p-hacking; and finally, g) preregistration may make us less risk tolerant. We cast our discussion as a dialogue, presenting both sides of the debate.
Gender-neutral pronouns have recently been introduced in many languages to a) include non-binary people and b) as a generic singular. Recent results from psycholinguistics suggest that gender-neutral pronouns (in Swedish) are not associated with human processing difficulties. This, we show, is in sharp contrast with automated processing. We show that gender-neutral pronouns in Danish, English, and Swedish are associated with higher perplexity, more dispersed attention patterns, and worse downstream performance. We argue that such conservativity in language models may limit widespread adoption of gender-neutral pronouns and must therefore be resolved.
Logical approaches to representing language have developed and evaluated computational models of quantifier words since the 19th century, but today’s NLU models still struggle to capture their semantics. We rely on Generalized Quantifier Theory for language-independent representations of the semantics of quantifier words, to quantify their contribution to the errors of NLU models. We find that quantifiers are pervasive in NLU benchmarks, and their occurrence at test time is associated with performance drops. Multilingual models also exhibit unsatisfying quantifier reasoning abilities, but not necessarily worse for non-English languages. To facilitate directly-targeted probing, we present an adversarial generalized quantifier NLI task (GQNLI) and show that pre-trained language models have a clear lack of robustness in generalized quantifier reasoning.
Multilingual NLP models provide potential solutions to the digital language divide, i.e., cross-language performance disparities. Early analyses of such models have indicated good performance across training languages and good generalization to unseen, related languages. This work examines whether, between related languages, multilingual models are equally right for the right reasons, i.e., if interpretability methods reveal that the models put emphasis on the same words as humans. To this end, we provide a new trilingual, parallel corpus of rationale annotations for English, Danish, and Italian sentiment analysis models and use it to benchmark models and interpretability methods. We propose rank-biased overlap as a better metric for comparing input token attributions to human rationale annotations. Our results show: (i) models generally perform well on the languages they are trained on, and align best with human rationales in these languages; (ii) performance is higher on English, even when not a source language, but this performance is not accompanied by higher alignment with human rationales, which suggests that language models favor English, but do not facilitate successful transfer of rationales.
Learned self-attention functions in state-of-the-art NLP models often correlate with human attention. We investigate whether self-attention in large-scale pre-trained language models is as predictive of human eye fixation patterns during task-reading as classical cognitive models of human attention. We compare attention functions across two task-specific reading datasets for sentiment analysis and relation extraction. We find the predictiveness of large-scale pre-trained self-attention for human attention depends on ‘what is in the tail’, e.g., the syntactic nature of rare contexts. Further, we observe that task-specific fine-tuning does not increase the correlation with human task-specific reading. Through an input reduction experiment we give complementary insights on the sparsity and fidelity trade-off, showing that lower-entropy attention vectors are more faithful.
We present a benchmark suite of four datasets for evaluating the fairness of pre-trained language models and the techniques used to fine-tune them for downstream tasks. Our benchmarks cover four jurisdictions (European Council, USA, Switzerland, and China), five languages (English, German, French, Italian and Chinese) and fairness across five attributes (gender, age, region, language, and legal area). In our experiments, we evaluate pre-trained language models using several group-robust fine-tuning techniques and show that performance group disparities are vibrant in many cases, while none of these techniques guarantee fairness, nor consistently mitigate group disparities. Furthermore, we provide a quantitative and qualitative analysis of our results, highlighting open challenges in the development of robustness methods in legal NLP.
Recent studies have shown that language models pretrained and/or fine-tuned on randomly permuted sentences exhibit competitive performance on GLUE, putting into question the importance of word order information. Somewhat counter-intuitively, some of these studies also report that position embeddings appear to be crucial for models’ good performance with shuffled text. We probe these language models for word order information and investigate what position embeddings learned from shuffled text encode, showing that these models retain a notion of word order information. We show this is in part due to a subtlety in how shuffling is implemented in previous work – before rather than after subword segmentation. Surprisingly, we find even Language models trained on text shuffled after subword segmentation retain some semblance of information about word order because of the statistical dependencies between sentence length and unigram probabilities. Finally, we show that beyond GLUE, a variety of language understanding tasks do require word order information, often to an extent that cannot be learned through fine-tuning.
Various efforts in the Natural Language Processing (NLP) community have been made to accommodate linguistic diversity and serve speakers of many different languages. However, it is important to acknowledge that speakers and the content they produce and require, vary not just by language, but also by culture. Although language and culture are tightly linked, there are important differences. Analogous to cross-lingual and multilingual NLP, cross-cultural and multicultural NLP considers these differences in order to better serve users of NLP systems. We propose a principled framework to frame these efforts, and survey existing and potential strategies.
Large multilingual pretrained language models such as mBERT and XLM-RoBERTa have been found to be surprisingly effective for cross-lingual transfer of syntactic parsing models Wu and Dredze (2019), but only between related languages. However, source and training languages are rarely related, when parsing truly low-resource languages. To close this gap, we adopt a method from multi-task learning, which relies on automated curriculum learning, to dynamically optimize for parsing performance on outlier languages. We show that this approach is significantly better than uniform and size-proportional sampling in the zero-shot setting.
Around two thirds of NLP research at top venues is devoted exclusively to developing technology for speakers of English, most speech data comes from young urban speakers, and most texts used to train language models come from male writers. These biases feed into consumer technologies to widen existing inequality gaps, not only within, but also across, societies. Many have argued that it is almost impossible to mitigate inequality amplification. I argue that, on the contrary, it is quite simple to do so, and that counter-measures would have little-to-no negative impact, except for, perhaps, in the very short term.
The prototypical NLP experiment trains a standard architecture on labeled English data and optimizes for accuracy, without accounting for other dimensions such as fairness, interpretability, or computational efficiency. We show through a manual classification of recent NLP research papers that this is indeed the case and refer to it as the square one experimental setup. We observe that NLP research often goes beyond the square one setup, e.g, focusing not only on accuracy, but also on fairness or interpretability, but typically only along a single dimension. Most work targeting multilinguality, for example, considers only accuracy; most work on fairness or interpretability considers only English; and so on. Such one-dimensionality of most research means we are only exploring a fraction of the NLP research search space. We provide historical and recent examples of how the square one bias has led researchers to draw false conclusions or make unwise choices, point to promising yet unexplored directions on the research manifold, and make practical recommendations to enable more multi-dimensional research. We open-source the results of our annotations to enable further analysis.
In document classification for, e.g., legal and biomedical text, we often deal with hundreds of classes, including very infrequent ones, as well as temporal concept drift caused by the influence of real world events, e.g., policy changes, conflicts, or pandemics. Class imbalance and drift can sometimes be mitigated by resampling the training data to simulate (or compensate for) a known target distribution, but what if the target distribution is determined by unknown future events? Instead of simply resampling uniformly to hedge our bets, we focus on the underlying optimization algorithms used to train such document classifiers and evaluate several group-robust optimization algorithms, initially proposed to mitigate group-level disparities. Reframing group-robust algorithms as adaptation algorithms under concept drift, we find that Invariant Risk Minimization and Spectral Decoupling outperform sampling-based approaches to class imbalance and concept drift, and lead to much better performance on minority classes. The effect is more pronounced the larger the label set.
Pretrained language models can be queried for factual knowledge, with potential applications in knowledge base acquisition and tasks that require inference. However, for that, we need to know how reliable this knowledge is, and recent work has shown that monolingual English language models lack consistency when predicting factual knowledge, that is, they fill-in-the-blank differently for paraphrases describing the same fact. In this paper, we extend the analysis of consistency to a multilingual setting. We introduce a resource, mParaRel, and investigate (i) whether multilingual language models such as mBERT and XLM-R are more consistent than their monolingual counterparts;and (ii) if such models are equally consistent across languages. We find that mBERT is as inconsistent as English BERT in English paraphrases, but that both mBERT and XLM-R exhibit a high degree of inconsistency in English and even more so for all the other 45 languages.
Synthetic datasets have successfully been used to probe visual question-answering datasets for their reasoning abilities. CLEVR (John- son et al., 2017), for example, tests a range of visual reasoning abilities. The questions in CLEVR focus on comparisons of shapes, colors, and sizes, numerical reasoning, and existence claims. This paper introduces a minimally biased, diagnostic visual question-answering dataset, QLEVR, that goes beyond existential and numerical quantification and focus on more complex quantifiers and their combinations, e.g., asking whether there are more than two red balls that are smaller than at least three blue balls in an image. We describe how the dataset was created and present a first evaluation of state-of-the-art visual question-answering models, showing that QLEVR presents a formidable challenge to our current models. Code and Dataset are available at https://github.com/zechenli03/QLEVR
In recent years, the natural language processing (NLP) community has given increased attention to the disparity of efforts directed towards high-resource languages over low-resource ones. Efforts to remedy this delta often begin with translations of existing English datasets into other languages. However, this approach ignores that different language communities have different needs. We consider a group of low-resource languages, creole languages. Creoles are both largely absent from the NLP literature, and also often ignored by society at large due to stigma, despite these languages having sizable and vibrant communities. We demonstrate, through conversations with creole experts and surveys of creole-speaking communities, how the things needed from language technology can change dramatically from one language to another, even when the languages are considered to be very similar to each other, as with creoles. We discuss the prominent themes arising from these conversations, and ultimately demonstrate that useful language technology cannot be built without involving the relevant community.
NLP models are dependent on the data they are trained on, including how this data is annotated. NLP research increasingly examines the social biases of models, but often in the light of their training data and specific social biases that can be identified in the text itself. In this paper, we present an annotation experiment that is the first to examine the extent to which social bias is sensitive to how data is annotated. We do so by collecting annotations of arguments in the same documents following four different guidelines and from four different demographic annotator backgrounds. We show that annotations exhibit widely different levels of group disparity depending on which guidelines annotators follow. The differences are not explained by task complexity, but rather by characteristics of these demographic groups, as previously identified by sociological studies. We release a dataset that is small in the number of instances but large in the number of annotations with demographic information, and our results encourage an increased awareness of annotator bias.
We aim to learn language models for Creole languages for which large volumes of data are not readily available, and therefore explore the potential transfer from ancestor languages (the ‘Ancestry Transfer Hypothesis’). We find that standard transfer methods do not facilitate ancestry transfer. Surprisingly, different from other non-Creole languages, a very distinct two-phase pattern emerges for Creoles: As our training losses plateau, and language models begin to overfit on their source languages, perplexity on the Creoles drop. We explore if this compression phase can lead to practically useful language models (the ‘Ancestry Bottleneck Hypothesis’), but also falsify this. Moreover, we show that Creoles even exhibit this two-phase pattern even when training on random, unrelated languages. Thus Creoles seem to be typological outliers and we speculate whether there is a link between the two observations.
Logical approaches to representing language have developed and evaluated computational models of quantifier words since the 19th century, but today’s NLU models still struggle to capture their semantics. We rely on Generalized Quantifier Theory for language-independent representations of the semantics of quantifier words, to quantify their contribution to the errors of NLU models. We find that quantifiers are pervasive in NLU benchmarks, and their occurrence at test time is associated with performance drops. Multilingual models also exhibit unsatisfying quantifier reasoning abilities, but not necessarily worse for non-English languages. To facilitate directly-targeted probing, we present an adversarial generalized quantifier NLI task (GQNLI) and show that pre-trained language models have a clear lack of robustness in generalized quantifier reasoning.
The performance cost of differential privacy has, for some applications, been shown to be higher for minority groups fairness, conversely, has been shown to disproportionally compromise the privacy of members of such groups. Most work in this area has been restricted to computer vision and risk assessment. In this paper, we evaluate the impact of differential privacy on fairness across four tasks, focusing on how attempts to mitigate privacy violations and between-group performance differences interact Does privacy inhibit attempts to ensure fairness? To this end, we train epsilon, delta-differentially private models with empirical risk minimization and group distributionally robust training objectives. Consistent with previous findings, we find that differential privacy increases between-group performance differences in the baseline setting but more interestingly, differential privacy reduces between-group performance differences in the robust setting. We explain this by reinterpreting differential privacy as regularization.
Pretrained multilingual language models can help bridge the digital language divide, enabling high-quality NLP models for lower-resourced languages. Studies of multilingual models have so far focused on performance, consistency, and cross-lingual generalisation. However, with their wide-spread application in the wild and downstream societal impact, it is important to put multilingual models under the same scrutiny as monolingual models. This work investigates the group fairness of multilingual models, asking whether these models are equally fair across languages. To this end, we create a new four-way multilingual dataset of parallel cloze test examples (MozArt), equipped with demographic information (balanced with regard to gender and native tongue) about the test participants. We evaluate three multilingual models on MozArt –mBERT, XLM-R, and mT5– and show that across the four target languages, the three models exhibit different levels of group disparity, e.g., exhibiting near-equal risk for Spanish, but high levels of disparity for German.
Multilingual language models exhibit better performance for some languages than for others (Singh et al., 2019), and many languages do not seem to benefit from multilingual sharing at all, presumably as a result of poor multilingual segmentation (Pyysal o et al., 2020). This work explores the idea of learning multilingual language models based on clustering of monolingual segments. We show significant improvements over standard multilingual segmentation and training across nine languages on a question answering task, both in a small model regime and for a model of the size of BERT-base.
Søgaard (2020) obtained results suggesting the fraction of trees occurring in the test data isomorphic to trees in the training set accounts for a non-trivial variation in parser performance. Similar to other statistical analyses in NLP, the results were based on evaluating linear regressions. However, the study had methodological issues and was undertaken using a small sample size leading to unreliable results. We present a replication study in which we also bin sentences by length and find that only a small subset of sentences vary in performance with respect to graph isomorphism. Further, the correlation observed between parser performance and graph isomorphism in the wild disappears when controlling for covariants. However, in a controlled experiment, where covariants are kept fixed, we do observe a correlation. We suggest that conclusions drawn from statistical analyses like this need to be tempered and that controlled experiments can complement them by more readily teasing factors apart.
Large-scale language models such as ELMo and BERT have pushed the horizon of what is possible in semantic role labeling (SRL), solving the out-of-vocabulary problem and enabling end-to-end systems, but they have also introduced significant biases. We evaluate three SRL parsers on very simple transitive sentences with verbs usually associated with animate subjects and objects, such as, “Mary babysat Tom”: a state-of-the-art parser based on BERT, an older parser based on GloVe, and an even older parser from before the days of word embeddings. When arguments are word forms predominantly used as person names, aligning with common sense expectations of animacy, the BERT-based parser is unsurprisingly superior; yet, with abstract or random nouns, the opposite picture emerges. We refer to this as “common sense bias” and present a challenge dataset for evaluating the extent to which parsers are sensitive to such a bias. Our code and challenge dataset are available here: github.com/coastalcph/comte
Recent work in cross-topic argument mining attempts to learn models that generalise across topics rather than merely relying on within-topic spurious correlations. We examine the effectiveness of this approach by analysing the output of single-task and multi-task models for cross-topic argument mining, through a combination of linear approximations of their decision boundaries, manual feature grouping, challenge examples, and ablations across the input vocabulary. Surprisingly, we show that cross-topic models still rely mostly on spurious correlations and only generalise within closely related topics, e.g., a model trained only on closed-class words and a few common open-class words outperforms a state-of-the-art cross-topic model on distant target topics.
Creole languages such as Nigerian Pidgin English and Haitian Creole are under-resourced and largely ignored in the NLP literature. Creoles typically result from the fusion of a foreign language with multiple local languages, and what grammatical and lexical features are transferred to the creole is a complex process. While creoles are generally stable, the prominence of some features may be much stronger with certain demographics or in some linguistic situations. This paper makes several contributions: We collect existing corpora and release models for Haitian Creole, Nigerian Pidgin English, and Singaporean Colloquial English. We evaluate these models on intrinsic and extrinsic tasks. Motivated by the above literature, we compare standard language models with distributionally robust ones and find that, somewhat surprisingly, the standard language models are superior to the distributionally robust ones. We investigate whether this is an effect of over-parameterization or relative distributional stability, and find that the difference persists in the absence of over-parameterization, and that drift is limited, confirming the relative stability of creole languages.
Pretrained language models have been shown to encode relational information, such as the relations between entities or concepts in knowledge-bases — (Paris, Capital, France). However, simple relations of this type can often be recovered heuristically and the extent to which models implicitly reflect topological structure that is grounded in world, such as perceptual structure, is unknown. To explore this question, we conduct a thorough case study on color. Namely, we employ a dataset of monolexemic color terms and color chips represented in CIELAB, a color space with a perceptually meaningful distance metric. Using two methods of evaluating the structural alignment of colors in this space with text-derived color term representations, we find significant correspondence. Analyzing the differences in alignment across the color spectrum, we find that warmer colors are, on average, better aligned to the perceptual color space than cooler ones, suggesting an intriguing connection to findings from recent work on efficient communication in color naming. Further analysis suggests that differences in alignment are, in part, mediated by collocationality and differences in syntactic usage, posing questions as to the relationship between color perception and usage and context.
Negation is one of the most fundamental concepts in human cognition and language, and several natural language inference (NLI) probes have been designed to investigate pretrained language models’ ability to detect and reason with negation. However, the existing probing datasets are limited to English only, and do not enable controlled probing of performance in the absence or presence of negation. In response, we present a multilingual (English, Bulgarian, German, French and Chinese) benchmark collection of NLI examples that are grammatical and correctly labeled, as a result of manual inspection and reformulation. We use the benchmark to probe the negation-awareness of multilingual language models and find that models that correctly predict examples with negation cues, often fail to correctly predict their counter-examples without negation cues, even when the cues are irrelevant for semantic inference.
We evaluated a range of neural machine translation techniques developed specifically for low-resource scenarios. Unsuccessfully. In the end, we submitted two runs: (i) a standard phrase-based model, and (ii) a random babbling baseline using character trigrams. We found that it was surprisingly hard to beat (i), in spite of this model being, in theory, a bad fit for polysynthetic languages; and more interestingly, that (ii) was better than several of the submitted systems, highlighting how difficult low-resource machine translation for polysynthetic languages is.
This work introduces Itihasa, a large-scale translation dataset containing 93,000 pairs of Sanskrit shlokas and their English translations. The shlokas are extracted from two Indian epics viz., The Ramayana and The Mahabharata. We first describe the motivation behind the curation of such a dataset and follow up with empirical analysis to bring out its nuances. We then benchmark the performance of standard translation models on this corpus and show that even state-of-the-art transformer architectures perform poorly, emphasizing the complexity of the dataset.
This work shows that competitive translation results can be obtained in a constrained setting by incorporating the latest advances in memory and compute optimization. We train and evaluate large multilingual translation models using a single GPU for a maximum of 100 hours and get within 4-5 BLEU points of the top submission on the leaderboard. We also benchmark standard baselines on the PMI corpus and re-discover well-known shortcomings of translation systems and metrics.
The global geometry of language models is important for a range of applications, but language model probes tend to evaluate rather local relations, for which ground truths are easily obtained. In this paper we exploit the fact that in geography, ground truths are available beyond local relations. In a series of experiments, we evaluate the extent to which language model representations of city and country names are isomorphic to real-world geography, e.g., if you tell a language model where Paris and Berlin are, does it know the way to Rome? We find that language models generally encode limited geographic information, but with larger models performing the best, suggesting that geographic knowledge can be induced from higher-order co-occurrence statistics.
Negation scope resolution is key to high-quality information extraction from clinical texts, but so far, efforts to make encoders used for information extraction negation-aware have been limited to English. We present a universal approach to multilingual negation scope resolution, that overcomes the lack of training data by relying on disparate resources in different languages and domains. We evaluate two approaches to learn from these resources, training on combined data and training in a multi-task learning setup. Our experiments show that zero-shot scope resolution in clinical text is possible, and that combining available resources improves performance in most cases.
Most, if not all forms of ellipsis (e.g., so does Mary) are similar to reading comprehension questions (what does Mary do), in that in order to resolve them, we need to identify an appropriate text span in the preceding discourse. Following this observation, we present an alternative approach for English ellipsis resolution relying on architectures developed for question answering (QA). We present both single-task models, and joint models trained on auxiliary QA and coreference resolution datasets, clearly outperforming the current state of the art for Sluice Ellipsis (from 70.00 to 86.01 F1) and Verb Phrase Ellipsis (from 72.89 to 78.66 F1).
(CITATION) argued for using random splits rather than standard splits in NLP experiments. We argue that random splits, like standard splits, lead to overly optimistic performance estimates. We can also split data in biased or adversarial ways, e.g., training on short sentences and evaluating on long ones. Biased sampling has been used in domain adaptation to simulate real-world drift; this is known as the covariate shift assumption. In NLP, however, even worst-case splits, maximizing bias, often under-estimate the error observed on new samples of in-domain data, i.e., the data that models should minimally generalize to at test time. This invalidates the covariate shift assumption. Instead of using multiple random splits, future benchmarks should ideally include multiple, independent test sets instead; if infeasible, we argue that multiple biased splits leads to more realistic performance estimates than multiple random splits.
We evaluate two common conjectures in error analysis of NLP models: (i) Morphology is predictive of errors; and (ii) the importance of morphology increases with the morphological complexity of a language. We show across four different tasks and up to 57 languages that of these conjectures, somewhat surprisingly, only (i) is true. Using morphological features does improve error prediction across tasks; however, this effect is less pronounced with morphologically complex languages. We speculate this is because morphology is more discriminative in morphologically simple languages. Across all four tasks, case and gender are the morphological features most predictive of error.
Since the popularization of the Transformer as a general-purpose feature encoder for NLP, many studies have attempted to decode linguistic structure from its novel multi-head attention mechanism. However, much of such work focused almost exclusively on English — a language with rigid word order and a lack of inflectional morphology. In this study, we present decoding experiments for multilingual BERT across 18 languages in order to test the generalizability of the claim that dependency syntax is reflected in attention patterns. We show that full trees can be decoded above baseline accuracy from single attention heads, and that individual relations are often tracked by the same heads across languages. Furthermore, in an attempt to address recent debates about the status of attention as an explanatory mechanism, we experiment with fine-tuning mBERT on a supervised parsing objective while freezing different series of parameters. Interestingly, in steering the objective to learn explicit linguistic structure, we find much of the same structure represented in the resulting attention patterns, with interesting differences with respect to which parameters are frozen.
Automatic coreference resolution is understudied in Danish even though most of the Danish Dependency Treebank (Buch-Kromann, 2003) is annotated with coreference relations. This paper describes a conversion of its partial, yet well-documented, coreference relations into coreference clusters and the training and evaluation of coreference models on this data. To the best of our knowledge, these are the first publicly available, neural coreference models for Danish. We also present a new entity linking annotation on the dataset using WikiData identifiers, a named entity disambiguation (NED) dataset, and a larger automatically created NED dataset enabling wikily supervised NED models. The entity linking annotation is benchmarked using a state-of-the-art neural entity disambiguation model.
In order to preserve word-order information in a non-autoregressive setting, transformer architectures tend to include positional knowledge, by (for instance) adding positional encodings to token embeddings. Several modifications have been proposed over the sinusoidal positional encodings used in the original transformer architecture; these include, for instance, separating position encodings and token embeddings, or directly modifying attention weights based on the distance between word pairs. We first show that surprisingly, while these modifications tend to improve monolingual language models, none of them result in better multilingual language models. We then answer why that is: sinusoidal encodings were explicitly designed to facilitate compositionality by allowing linear projections over arbitrary time steps. Higher variances in multilingual training distributions requires higher compression, in which case, compositionality becomes indispensable. Learned absolute positional encodings (e.g., in mBERT) tend to approximate sinusoidal embeddings in multilingual settings, but more complex positional encoding architectures lack the inductive bias to effectively learn cross-lingual alignment. In other words, while sinusoidal positional encodings were designed for monolingual applications, they are particularly useful in multilingual language models.
Sentiment analysis systems have been shown to exhibit sensitivity to protected attributes. Round-trip translation, on the other hand, has been shown to normalize text. We explore the impact of round-trip translation on the demographic parity of sentiment classifiers and show how round-trip translation consistently improves classification fairness at test time (reducing up to 47% of between-group gaps). We also explore the idea of retraining sentiment classifiers on round-trip-translated data.
Using data from English cloze tests, in which subjects also self-reported their gender, age, education, and race, we examine performance differences of pretrained language models across demographic groups, defined by these (protected) attributes. We demonstrate wide performance gaps across demographic groups and show that pretrained language models systematically disfavor young non-white male speakers; i.e., not only do pretrained language models learn social biases (stereotypical associations) – pretrained language models also learn sociolectal biases, learning to speak more like some than like others. We show, however, that, with the exception of BERT models, larger pretrained language models reduce some the performance gaps between majority and minority groups.
Online conversations can sometimes take a turn for the worse, either due to systematic cultural differences, accidental misunderstandings, or mere malice. Automatically forecasting derailment in public online conversations provides an opportunity to take early action to moderate it. Previous work in this space is limited, and we extend it in several ways. We apply a pretrained language encoder to the task, which outperforms earlier approaches. We further experiment with shifting the training paradigm for the task from a static to a dynamic one to increase the forecast horizon. This approach shows mixed results: in a high-quality data setting, a longer average forecast horizon can be achieved at the cost of a small drop in F1; in a low-quality data setting, however, dynamic training propagates the noise and is highly detrimental to performance.
I highlight a simple failure mode of state-of-the-art machine reading systems: when contexts do not align with commonly shared beliefs. For example, machine reading systems fail to answer What did Elizabeth want? correctly in the context of ‘My kingdom for a cough drop, cried Queen Elizabeth.’ Biased by co-occurrence statistics in the training data of pretrained language models, systems predict my kingdom, rather than a cough drop. I argue such biases are analogous to human belief biases and present a carefully designed challenge dataset for English machine reading, called Auto-Locke, to quantify such effects. Evaluations of machine reading systems on Auto-Locke show the pervasiveness of belief bias in machine reading.
Summarization systems are ultimately evaluated by human annotators and raters. Usually, annotators and raters do not reflect the demographics of end users, but are recruited through student populations or crowdsourcing platforms with skewed demographics. For two different evaluation scenarios – evaluation against gold summaries and system output ratings – we show that summary evaluation is sensitive to protected attributes. This can severely bias system development and evaluation, leading us to build models that cater for some groups rather than others.
NLP models struggle with generalization due to sampling and annotator bias. This paper focuses on a different kind of bias that has received very little attention: guideline bias, i.e., the bias introduced by how our annotator guidelines are formulated. We examine two recently introduced dialogue datasets, CCPE-M and Taskmaster-1, both collected by trained assistants in a Wizard-of-Oz set-up. For CCPE-M, we show how a simple lexical bias for the word like in the guidelines biases the data collection. This bias, in effect, leads to poor performance on data without this bias: a preference elicitation architecture based on BERT suffers a 5.3% absolute drop in performance, when like is replaced with a synonymous phrase, and a 13.2% drop in performance when evaluated on out-of-sample data. For Taskmaster-1, we show how the order in which instructions are resented, biases the data collection.
Humans do not make inferences over texts, but over models of what texts are about. When annotators are asked to annotate coreferent spans of text, it is therefore a somewhat unnatural task. This paper presents an alternative in which we preprocess documents, linking entities to a knowledge base, and turn the coreference annotation task – in our case limited to pronouns – into an annotation task where annotators are asked to assign pronouns to entities. Model-based annotation is shown to lead to faster annotation and higher inter-annotator agreement, and we argue that it also opens up an alternative approach to coreference resolution. We present two new coreference benchmark datasets, for English Wikipedia and English teacher-student dialogues, and evaluate state-of-the-art coreference resolvers on them.
Frame-semantic annotations exist for a tiny fraction of the world’s languages, Wikidata, however, links knowledge base triples to texts in many languages, providing a common, distant supervision signal for semantic parsers. We present WikiBank, a multilingual resource of partial semantic structures that can be used to extend pre-existing resources rather than creating new man-made resources from scratch. We also integrate this form of supervision into an off-the-shelf frame-semantic parser and allow cross-lingual transfer. Using Google’s Sling architecture, we show significant improvements on the English and Spanish CoNLL 2009 datasets, whether training on the full available datasets or small subsamples thereof.
We present a named entity annotation for the Danish Universal Dependencies treebank using the CoNLL-2003 annotation scheme: DaNE. It is the largest publicly available, Danish named entity gold annotation. We evaluate the quality of our annotations intrinsically by double annotating the entire treebank and extrinsically by comparing our annotations to a recently released named entity annotation of the validation and test sections of the Danish Universal Dependencies treebank. We benchmark the new resource by training and evaluating competitive architectures for supervised named entity recognition (NER), including FLAIR, monolingual (Danish) BERT and multilingual BERT. We explore cross-lingual transfer in multilingual BERT from five related languages in zero-shot and direct transfer setups, and we show that even with our modestly-sized training set, we improve Danish NER over a recent cross-lingual approach, as well as over zero-shot transfer from five related languages. Using multilingual BERT, we achieve higher performance by fine-tuning on both DaNE and a larger Bokmål (Norwegian) training set compared to only using DaNE. However, the highest performance isachieved by using a Danish BERT fine-tuned on DaNE. Our dataset enables improvements and applicability for Danish NER beyond cross-lingual methods. We employ a thorough error analysis of the predictions of the best models for seen and unseen entities, as well as their robustness on un-capitalized text. The annotated dataset and all the trained models are made publicly available.
Large-scale pretrained language models are the major driving force behind recent improvements in perfromance on the Winograd Schema Challenge, a widely employed test of commonsense reasoning ability. We show, however, with a new diagnostic dataset, that these models are sensitive to linguistic perturbations of the Winograd examples that minimally affect human understanding. Our results highlight interesting differences between humans and language models: language models are more sensitive to number or gender alternations and synonym replacements than humans, and humans are more stable and consistent in their predictions, maintain a much higher absolute performance, and perform better on non-associative instances than associative ones.
Several approaches to neural speed reading have been presented at major NLP and machine learning conferences in 2017–20; i.e., “human-inspired” recurrent network architectures that learn to “read” text faster by skipping irrelevant words, typically optimizing the joint objective of minimizing classification error rate and FLOPs used at inference time. This paper reflects on the meaningfulness of the speed reading task, showing that (a) better and faster approaches to, say, document classification, already exist, which also learn to ignore part of the input (I give an example with 7% error reduction and a 136x speed-up over the state of the art in neural speed reading); and that (b) any claims that neural speed reading is “human-inspired”, are ill-founded.
The one-sided focus on English in previous studies of gender bias in NLP misses out on opportunities in other languages: English challenge datasets such as GAP and WinoGender highlight model preferences that are “hallucinatory”, e.g., disambiguating gender-ambiguous occurrences of ‘doctor’ as male doctors. We show that for languages with type B reflexivization, e.g., Swedish and Russian, we can construct multi-task challenge datasets for detecting gender bias that lead to unambiguously wrong model predictions: In these languages, the direct translation of ‘the doctor removed his mask’ is not ambiguous between a coreferential reading and a disjoint reading. Instead, the coreferential reading requires a non-gendered pronoun, and the gendered, possessive pronouns are anti-reflexive. We present a multilingual, multi-task challenge dataset, which spans four languages and four NLP tasks and focuses only on this phenomenon. We find evidence for gender bias across all task-language combinations and correlate model bias with national labor market statistics.
Cross-language differences in (universal) dependency parsing performance are mostly attributed to treebank size, average sentence length, average dependency length, morphological complexity, and domain differences. We point at a factor not previously discussed: If we abstract away from words and dependency labels, how many graphs in the test data were seen in the training data? We compute graph isomorphisms, and show that, treebank size aside, overlap between training and test graphs explain more of the observed variation than standard explanations such as the above.
Existing algorithms for aligning cross-lingual word vector spaces assume that vector spaces are approximately isomorphic. As a result, they perform poorly or fail completely on non-isomorphic spaces. Such non-isomorphism has been hypothesised to result from typological differences between languages. In this work, we ask whether non-isomorphism is also crucially a sign of degenerate word vector spaces. We present a series of experiments across diverse languages which show that variance in performance across language pairs is not only due to typological differences, but can mostly be attributed to the size of the monolingual resources available, and to the properties and duration of monolingual training (e.g. “under-training”).
Evaluation of grammatical error correction (GEC) systems has primarily focused on essays written by non-native learners of English, which however is only part of the full spectrum of GEC applications. We aim to broaden the target domain of GEC and release CWEB, a new benchmark for GEC consisting of website text generated by English speakers of varying levels of proficiency. Website data is a common and important domain that contains far fewer grammatical errors than learner essays, which we show presents a challenge to state-of-the-art GEC systems. We demonstrate that a factor behind this is the inability of systems to rely on a strong internal language model in low error density domains. We hope this work shall facilitate the development of open-domain GEC models that generalize to different topics and genres.
This work describes the system presented by the CoAStaL Natural Language Processing group at University of Copenhagen. The main system we present uses the same attention mechanism presented in (Yang et al., 2016). Our overall model architecture is also inspired by their hierarchical classification model and adapted to deal with classification in dialogue by encoding information at the turn level. We use different encodings for each turn to create a more expressive representation of dialogue context which is then fed into our classifier. We also define a custom preprocessing step in order to deal with language commonly used in interactions across many social media outlets. Our proposed system achieves a micro F1 score of 0.7340 on the test set and shows significant gains in performance compared to a system using dialogue level encoding.
Training neural sequence-to-sequence models with simple token-level log-likelihood is now a standard approach to historical text normalization, albeit often outperformed by phrase-based models. Policy gradient training enables direct optimization for exact matches, and while the small datasets in historical text normalization are prohibitive of from-scratch reinforcement learning, we show that policy gradient fine-tuning leads to significant improvements across the board. Policy gradient training, in particular, leads to more accurate normalizations for long or unseen words.
In Semantic Dependency Parsing (SDP), semantic relations form directed acyclic graphs, rather than trees. We propose a new iterative predicate selection (IPS) algorithm for SDP. Our IPS algorithm combines the graph-based and transition-based parsing approaches in order to handle multiple semantic head words. We train the IPS model using a combination of multi-task learning and task-specific policy gradient training. Trained this way, IPS achieves a new state of the art on the SemEval 2015 Task 18 datasets. Furthermore, we observe that policy gradient training learns an easy-first strategy.
In this tutorial, we provide a comprehensive survey of the exciting recent work on cutting-edge weakly-supervised and unsupervised cross-lingual word representations. After providing a brief history of supervised cross-lingual word representations, we focus on: 1) how to induce weakly-supervised and unsupervised cross-lingual word representations in truly resource-poor settings where bilingual supervision cannot be guaranteed; 2) critical examinations of different training conditions and requirements under which unsupervised algorithms can and cannot work effectively; 3) more robust methods for distant language pairs that can mitigate instability issues and low performance for distant language pairs; 4) how to comprehensively evaluate such representations; and 5) diverse applications that benefit from cross-lingual word representations (e.g., MT, dialogue, cross-lingual sequence labeling and structured prediction applications, cross-lingual IR).
In online discussion fora, speakers often make arguments for or against something, say birth control, by highlighting certain aspects of the topic. In social science, this is referred to as issue framing. In this paper, we introduce a new issue frame annotated corpus of online discussions. We explore to what extent models trained to detect issue frames in newswire and social media can be transferred to the domain of discussion fora, using a combination of multi-task and adversarial training, assuming only unlabeled training data in the target domain.
While rule-based detection of subject-verb agreement (SVA) errors is sensitive to syntactic parsing errors and irregularities and exceptions to the main rules, neural sequential labelers have a tendency to overfit their training data. We observe that rule-based error generation is less sensitive to syntactic parsing errors and irregularities than error detection and explore a simple, yet efficient approach to getting the best of both worlds: We train neural sequential labelers on the combination of large volumes of silver standard data, obtained through rule-based error generation, and gold standard data. We show that our simple protocol leads to more robust detection of SVA errors on both in-domain and out-of-domain data, as well as in the context of other errors and long-distance dependencies; and across four standard benchmarks, the induced model on average achieves a new state of the art.
Sequence tagging models for constituent parsing are faster, but less accurate than other types of parsers. In this work, we address the following weaknesses of such constituent parsers: (a) high error rates around closing brackets of long constituents, (b) large label sets, leading to sparsity, and (c) error propagation arising from greedy decoding. To effectively close brackets, we train a model that learns to switch between tagging schemes. To reduce sparsity, we decompose the label set and use multi-task learning to jointly learn to predict sublabels. Finally, we mitigate issues from greedy decoding through auxiliary losses and sentence-level fine-tuning with policy gradient. Combining these techniques, we clearly surpass the performance of sequence tagging constituent parsers on the English and Chinese Penn Treebanks, and reduce their parsing time even further. On the SPMRL datasets, we observe even greater improvements across the board, including a new state of the art on Basque, Hebrew, Polish and Swedish.
Parsers are available for only a handful of the world’s languages, since they require lots of training data. How far can we get with just a small amount of training data? We systematically compare a set of simple strategies for improving low-resource parsers: data augmentation, which has not been tested before; cross-lingual training; and transliteration. Experimenting on three typologically diverse low-resource languages—North Sámi, Galician, and Kazah—We find that (1) when only the low-resource treebank is available, data augmentation is very helpful; (2) when a related high-resource treebank is available, cross-lingual training is helpful and complements data augmentation; and (3) when the high-resource treebank uses a different writing system, transliteration into a shared orthographic spaces is also very helpful.
Unresolved coreference is a bottleneck for relation extraction, and high-quality coreference resolvers may produce an output that makes it a lot easier to extract knowledge triples. We show how to improve coreference resolvers by forwarding their input to a relation extraction system and reward the resolvers for producing triples that are found in knowledge bases. Since relation extraction systems can rely on different forms of supervision and be biased in different ways, we obtain the best performance, improving over the state of the art, using multi-task reinforcement learning.
The task of bilingual dictionary induction (BDI) is commonly used for intrinsic evaluation of cross-lingual word embeddings. The largest dataset for BDI was generated automatically, so its quality is dubious. We study the composition and quality of the test sets for five diverse languages from this dataset, with concerning findings: (1) a quarter of the data consists of proper nouns, which can be hardly indicative of BDI performance, and (2) there are pervasive gaps in the gold-standard targets. These issues appear to affect the ranking between cross-lingual embedding systems on individual languages, and the overall degree to which the systems differ in performance. With proper nouns removed from the data, the margin between the top two systems included in the study grows from 3.4% to 17.2%. Manual verification of the predictions, on the other hand, reveals that gaps in the gold standard targets artificially inflate the margin between the two systems on English to Bulgarian BDI from 0.1% to 6.7%. We thus suggest that future research either avoids drawing conclusions from quantitative results on this BDI dataset, or accompanies such evaluation with rigorous error analysis.
Representational Similarity Analysis (RSA) is a technique developed by neuroscientists for comparing activity patterns of different measurement modalities (e.g., fMRI, electrophysiology, behavior). As a framework, RSA has several advantages over existing approaches to interpretation of language encoders based on probing or diagnostic classification: namely, it does not require large training samples, is not prone to overfitting, and it enables a more transparent comparison between the representational geometries of different models and modalities. We demonstrate the utility of RSA by establishing a previously unknown correspondence between widely-employed pretrained language encoders and human processing difficulty via eye-tracking data, showcasing its potential in the interpretability toolbox for neural models.
Elazar and Goldberg (2018) showed that protected attributes can be extracted from the representations of a debiased neural network for mention detection at above-chance levels, by evaluating a diagnostic classifier on a held-out subsample of the data it was trained on. We revisit their experiments and conduct a series of follow-up experiments showing that, in fact, the diagnostic classifier generalizes poorly to both new in-domain samples and new domains, indicating that it relies on correlations specific to their particular data sample. We further show that a diagnostic classifier trained on the biased baseline neural network also does not generalize to new samples. In other words, the biases detected in Elazar and Goldberg (2018) seem restricted to their particular data sample, and would therefore not bias the decisions of the model on new samples, whether in-domain or out-of-domain. In light of this, we discuss better methodologies for detecting bias in our models.
Historical text normalization often relies on small training datasets. Recent work has shown that multi-task learning can lead to significant improvements by exploiting synergies with related datasets, but there has been no systematic study of different multi-task learning architectures. This paper evaluates 63 multi-task learning configurations for sequence-to-sequence-based historical text normalization across ten datasets from eight languages, using autoencoding, grapheme-to-phoneme mapping, and lemmatization as auxiliary tasks. We observe consistent, significant improvements across languages when training data for the target task is limited, but minimal or no improvements when training data is abundant. We also show that zero-shot learning outperforms the simple, but relatively strong, identity baseline.
Although the vast majority of knowledge bases (KBs) are heavily biased towards English, Wikipedias do cover very different topics in different languages. Exploiting this, we introduce a new multilingual dataset (X-WikiRE), framing relation extraction as a multilingual machine reading problem. We show that by leveraging this resource it is possible to robustly transfer models cross-lingually and that multilingual support significantly improves (zero-shot) relation extraction, enabling the population of low-resourced KBs from their well-populated counterparts.
This paper describes our contribution to the low-resource track of the BEA 2019 shared task on Grammatical Error Correction (GEC). Our approach to GEC builds on the theory of the noisy channel by combining a channel model and language model. We generate confusion sets from the Wikipedia edit history and use the frequencies of edits to estimate the channel model. Additionally, we use two pre-trained language models: 1) Google’s BERT model, which we fine-tune for specific error types and 2) OpenAI’s GPT-2 model, utilizing that it can operate with previous sentences as context. Furthermore, we search for the optimal combinations of corrections using beam search.
Neural machine translation models have little inductive bias, which can be a disadvantage in low-resource scenarios. Neural models have to be trained on large amounts of data and have been shown to perform poorly when only limited data is available. We show that using naive regularization methods, based on sentence length, punctuation and word frequencies, to penalize translations that are very different from the input sentences, consistently improves the translation quality across multiple low-resource languages. We experiment with 12 language pairs, varying the training data size between 17k to 230k sentence pairs. Our best regularizer achieves an average increase of 1.5 BLEU score and 1.0 TER score across all the language pairs. For example, we achieve a BLEU score of 26.70 on the IWSLT15 English–Vietnamese translation task simply by using relative differences in punctuation as a regularizer.
Can attention- or gradient-based visualization techniques be used to infer token-level labels for binary sequence tagging problems, using networks trained only on sentence-level labels? We construct a neural network architecture based on soft attention, train it as a binary sentence classifier and evaluate against token-level annotation on four different datasets. Inferring token labels from a network provides a method for quantitatively evaluating what the model is learning, along with generating useful feedback in assistance systems. Our results indicate that attention-based methods are able to predict token-level labels more accurately, compared to gradient-based methods, sometimes even rivaling the supervised oracle network.
We combine multi-task learning and semi-supervised learning by inducing a joint embedding space between disparate label spaces and learning transfer functions between label embeddings, enabling us to jointly leverage unlabelled data and auxiliary, annotated datasets. We evaluate our approach on a variety of tasks with disparate label spaces. We outperform strong single and multi-task baselines and achieve a new state of the art for aspect-based and topic-based sentiment analysis.
When learning POS taggers and syntactic chunkers for low-resource languages, different resources may be available, and often all we have is a small tag dictionary, motivating type-constrained unsupervised induction. Even small dictionaries can improve the performance of unsupervised induction algorithms. This paper shows that performance can be further improved by including data that is readily available or can be easily obtained for most languages, i.e., eye-tracking, speech, or keystroke logs (or any combination thereof). We project information from all these data sources into shared spaces, in which the union of words is represented. For English unsupervised POS induction, the additional information, which is not required at test time, leads to an average error reduction on Ontonotes domains of 1.5% over systems augmented with state-of-the-art word embeddings. On Penn Treebank the best model achieves 5.4% error reduction over a word embeddings baseline. We also achieve significant improvements for syntactic chunk induction. Our analysis shows that improvements are even bigger when the available tag dictionaries are smaller.
Sluice resolution in English is the problem of finding antecedents of wh-fronted ellipses. Previous work has relied on hand-crafted features over syntax trees that scale poorly to other languages and domains; in particular, to dialogue, which is one of the most interesting applications of sluice resolution. Syntactic information is arguably important for sluice resolution, but we show that multi-task learning with partial parsing as auxiliary tasks effectively closes the gap and buys us an additional 9% error reduction over previous work. Since we are not directly relying on features from partial parsers, our system is more robust to domain shifts, giving a 26% error reduction on embedded sluices in dialogue.
Most recent approaches to bilingual dictionary induction find a linear alignment between the word vector spaces of two languages. We show that projecting the two languages onto a third, latent space, rather than directly onto each other, while equivalent in terms of expressivity, makes it easier to learn approximate alignments. Our modified approach also allows for supporting languages to be included in the alignment process, to obtain an even better performance in low resource settings.
Learning attention functions requires large volumes of data, but many NLP tasks simulate human behavior, and in this paper, we show that human attention really does provide a good inductive bias on many attention functions in NLP. Specifically, we use estimated human attention derived from eye-tracking corpora to regularize attention functions in recurrent neural networks. We show substantial improvements across a range of tasks, including sentiment analysis, grammatical error detection, and detection of abusive language.
Unsupervised machine translation - i.e., not assuming any cross-lingual supervision signal, whether a dictionary, translations, or comparable corpora - seems impossible, but nevertheless, Lample et al. (2017) recently proposed a fully unsupervised machine translation (MT) model. The model relies heavily on an adversarial, unsupervised cross-lingual word embedding technique for bilingual dictionary induction (Conneau et al., 2017), which we examine here. Our results identify the limitations of current unsupervised MT: unsupervised bilingual dictionary induction performs much worse on morphologically rich languages that are not dependent marking, when monolingual corpora from different domains or different embedding algorithms are used. We show that a simple trick, exploiting a weak supervision signal from identical words, enables more robust induction and establish a near-perfect correlation between unsupervised bilingual dictionary induction performance and a previously unexplored graph similarity metric.
Most previous research in text simplification has aimed to develop generic solutions, assuming very homogeneous target audiences with consistent intra-group simplification needs. We argue that this assumption does not hold, and that instead we need to develop simplification systems that adapt to the individual needs of specific users. As a first step towards personalized simplification, we propose a framework for adaptive lexical simplification and introduce Lexi, a free open-source and easily extensible tool for adaptive, personalized text simplification. Lexi is easily installed as a browser extension, enabling easy access to the service for its users.
Cross-lingual representation learning is an important step in making NLP scale to all the world’s languages. Previous work on bilingual lexicon induction suggests that it is possible to learn cross-lingual representations of words based on similarities between images associated with these words. However, that work focused (almost exclusively) on the translation of nouns only. Here, we investigate whether the meaning of other parts-of-speech (POS), in particular adjectives and verbs, can be learned in the same way. Our experiments across five language pairs indicate that previous work does not scale to the problem of learning cross-lingual representations beyond simple nouns.
Neural part-of-speech (POS) taggers are known to not perform well with little training data. As a step towards overcoming this problem, we present an architecture for learning more robust neural POS taggers by jointly training a hierarchical, recurrent model and a recurrent character-based sequence-to-sequence network supervised using an auxiliary objective. This way, we introduce stronger character-level supervision into the model, which enables better generalization to unseen words and provides regularization, making our encoding less prone to overfitting. We experiment with three auxiliary tasks: lemmatization, character-based word autoencoding, and character-based random string autoencoding. Experiments with minimal amounts of labeled data on 34 languages show that our new architecture outperforms a single-task baseline and, surprisingly, that, on average, raw text autoencoding can be as beneficial for low-resource POS tagging as using lemma information. Our neural POS tagger closes the gap to a state-of-the-art POS tagger (MarMoT) for low-resource scenarios by 43%, even outperforming it on languages with templatic morphology, e.g., Arabic, Hebrew, and Turkish, by some margin.
Historical text normalization suffers from small datasets that exhibit high variance, and previous work has shown that multi-task learning can be used to leverage data from related problems in order to obtain more robust models. Previous work has been limited to datasets from a specific language and a specific historical period, and it is not clear whether results generalize. It therefore remains an open problem, when historical text normalization benefits from multi-task learning. We explore the benefits of multi-task learning across 10 different datasets, representing different languages and periods. Our main finding—contrary to what has been observed for other NLP tasks—is that multi-task learning mainly works when target task data is very scarce.
This work aims to contribute to our understanding of when multi-task learning through parameter sharing in deep neural networks leads to improvements over single-task learning. We focus on the setting of learning from loosely related tasks, for which no theoretical guarantees exist. We therefore approach the question empirically, studying which properties of datasets and single-task learning characteristics correlate with improvements from multi-task learning. We are the first to study this in a text classification setting and across more than 500 different task pairs.
Punctuation is a strong indicator of syntactic structure, and parsers trained on text with punctuation often rely heavily on this signal. Punctuation is a diversion, however, since human language processing does not rely on punctuation to the same extent, and in informal texts, we therefore often leave out punctuation. We also use punctuation ungrammatically for emphatic or creative purposes, or simply by mistake. We show that (a) dependency parsers are sensitive to both absence of punctuation and to alternative uses; (b) neural parsers tend to be more sensitive than vintage parsers; (c) training neural parsers without punctuation outperforms all out-of-the-box parsers across all scenarios where punctuation departs from standard punctuation. Our main experiments are on synthetically corrupted data to study the effect of punctuation in isolation and avoid potential confounds, but we also show effects on out-of-domain data.
Sluicing resolution is the task of identifying the antecedent to a question ellipsis. Antecedents are often sentential constituents, and previous work has therefore relied on syntactic parsing, together with complex linguistic features. A recent model instead used partial parsing as an auxiliary task in sequential neural network architectures to inject syntactic information. We explore the linguistic information being brought to bear by such networks, both by defining subsets of the data exhibiting relevant linguistic characteristics, and by examining the internal representations of the network. Both perspectives provide evidence for substantial linguistic knowledge being deployed by the neural networks.
Sentiment analysis models often rely on training data that is several years old. In this paper, we show that lexical features change polarity over time, leading to degrading performance. This effect is particularly strong in sparse models relying only on highly predictive features. Using predictive feature selection, we are able to significantly improve the accuracy of such models over time.
We introduce a novel discriminative latent-variable model for the task of bilingual lexicon induction. Our model combines the bipartite matching dictionary prior of Haghighi et al. (2008) with a state-of-the-art embedding-based approach. To train the model, we derive an efficient Viterbi EM algorithm. We provide empirical improvements on six language pairs under two metrics and show that the prior theoretically and empirically helps to mitigate the hubness problem. We also demonstrate how previous work may be viewed as a similarly fashioned latent-variable model, albeit with a different prior.
This paper presents a challenge to the community: Generative adversarial networks (GANs) can perfectly align independent English word embeddings induced using the same algorithm, based on distributional information alone; but fails to do so, for two different embeddings algorithms. Why is that? We believe understanding why, is key to understand both modern word embedding algorithms and the limitations and instability dynamics of GANs. This paper shows that (a) in all these cases, where alignment fails, there exists a linear transform between the two embeddings (so algorithm biases do not lead to non-linear differences), and (b) similar effects can not easily be obtained by varying hyper-parameters. One plausible suggestion based on our initial experiments is that the differences in the inductive biases of the embedding algorithms lead to an optimization landscape that is riddled with local optima, leading to a very small basin of convergence, but we present this more as a challenge paper than a technical contribution.
The best systems at the SemEval-16 and SemEval-17 community question answering shared tasks – a task that amounts to question relevancy ranking – involve complex pipelines and manual feature engineering. Despite this, many of these still fail at beating the IR baseline, i.e., the rankings provided by Google’s search engine. We present a strong baseline for question relevancy ranking by training a simple multi-task feed forward network on a bag of 14 distance measures for the input question pair. This baseline model, which is fast to train and uses only language-independent features, outperforms the best shared task systems on the task of retrieving relevant previously asked questions.
Previous work has suggested that parameter sharing between transition-based neural dependency parsers for related languages can lead to better performance, but there is no consensus on what parameters to share. We present an evaluation of 27 different parameter sharing strategies across 10 languages, representing five pairs of related languages, each pair from a different language family. We find that sharing transition classifier parameters always helps, whereas the usefulness of sharing word and/or character LSTM parameters varies. Based on this result, we propose an architecture where the transition classifier is shared, and the sharing of word and character parameters is controlled by a parameter that can be tuned on validation data. This model is linguistically motivated and obtains significant improvements over a monolingually trained baseline. We also find that sharing transition classifier parameters helps when training a parser on unrelated language pairs, but we find that, in the case of unrelated languages, sharing too many parameters does not help.
NLP tasks are often limited by scarcity of manually annotated data. In social media sentiment analysis and related tasks, researchers have therefore used binarized emoticons and specific hashtags as forms of distant supervision. Our paper shows that by extending the distant supervision to a more diverse set of noisy labels, the models can learn richer representations. Through emoji prediction on a dataset of 1246 million tweets containing one of 64 common emojis we obtain state-of-the-art performance on 8 benchmark datasets within emotion, sentiment and sarcasm detection using a single pretrained model. Our analyses confirm that the diversity of our emotional labels yield a performance improvement over previous distant supervision approaches.
Discourse segmentation is the first step in building discourse parsers. Most work on discourse segmentation does not scale to real-world discourse parsing across languages, for two reasons: (i) models rely on constituent trees, and (ii) experiments have relied on gold standard identification of sentence and token boundaries. We therefore investigate to what extent constituents can be replaced with universal dependencies, or left out completely, as well as how state-of-the-art segmenters fare in the absence of sentence boundaries. Our results show that dependency information is less useful than expected, but we provide a fully scalable, robust model that only relies on part-of-speech information, and show that it performs well across languages in the absence of any gold-standard annotation.
In recent past, NLP as a field has seen tremendous utility of distributional word vector representations as features in downstream tasks. The fact that these word vectors can be trained on unlabeled monolingual corpora of a language makes them an inexpensive resource in NLP. With the increasing use of monolingual word vectors, there is a need for word vectors that can be used as efficiently across multiple languages as monolingually. Therefore, learning bilingual and multilingual word embeddings/vectors is currently an important research topic. These vectors offer an elegant and language-pair independent way to represent content across different languages.This tutorial aims to bring NLP researchers up to speed with the current techniques in cross-lingual word representation learning. We will first discuss how to induce cross-lingual word representations (covering both bilingual and multilingual ones) from various data types and resources (e.g., parallel data, comparable data, non-aligned monolingual data in different languages, dictionaries and theasuri, or, even, images, eye-tracking data). We will then discuss how to evaluate such representations, intrinsically and extrinsically. We will introduce researchers to state-of-the-art methods for constructing cross-lingual word representations and discuss their applicability in a broad range of downstream NLP applications.We will deliver a detailed survey of the current methods, discuss best training and evaluation practices and use-cases, and provide links to publicly available implementations, datasets, and pre-trained models.
In cross-lingual dependency annotation projection, information is often lost during transfer because of early decoding. We present an end-to-end graph-based neural network dependency parser that can be trained to reproduce matrices of edge scores, which can be directly projected across word alignments. We show that our approach to cross-lingual dependency parsing is not only simpler, but also achieves an absolute improvement of 2.25% averaged across 10 languages compared to the previous state of the art.
We present UDP, the first training-free parser for Universal Dependencies (UD). Our algorithm is based on PageRank and a small set of specific dependency head rules. UDP features two-step decoding to guarantee that function words are attached as leaf nodes. The parser requires no training, and it is competitive with a delexicalized transfer system. UDP offers a linguistically sound unsupervised alternative to cross-lingual parsing for UD. The parser has very few parameters and distinctly robust to domain change across languages.
Discourse parsing is an integral part of understanding information flow and argumentative structure in documents. Most previous research has focused on inducing and evaluating models from the English RST Discourse Treebank. However, discourse treebanks for other languages exist, including Spanish, German, Basque, Dutch and Brazilian Portuguese. The treebanks share the same underlying linguistic theory, but differ slightly in the way documents are annotated. In this paper, we present (a) a new discourse parser which is simpler, yet competitive (significantly better on 2/3 metrics) to state of the art for English, (b) a harmonization of discourse treebanks across languages, enabling us to present (c) what to the best of our knowledge are the first experiments on cross-lingual discourse parsing.
While cross-lingual word embeddings have been studied extensively in recent years, the qualitative differences between the different algorithms remain vague. We observe that whether or not an algorithm uses a particular feature set (sentence IDs) accounts for a significant performance gap among these algorithms. This feature set is also used by traditional alignment algorithms, such as IBM Model-1, which demonstrate similar performance to state-of-the-art embedding algorithms on a variety of benchmarks. Overall, we observe that different algorithmic approaches for utilizing the sentence ID feature space result in similar performance. This paper draws both empirical and theoretical parallels between the embedding and alignment literature, and suggests that adding additional sources of information, which go beyond the traditional signal of bilingual sentence-aligned corpora, may substantially improve cross-lingual word embeddings, and that future baselines should at least take such features into account.
Multi-task learning (MTL) in deep neural networks for NLP has recently received increasing interest due to some compelling benefits, including its potential to efficiently regularize models and to reduce the need for labeled data. While it has brought significant improvements in a number of NLP tasks, mixed results have been reported, and little is known about the conditions under which MTL leads to gains in NLP. This paper sheds light on the specific task relations that can lead to gains from MTL models over single-task setups.
We address the challenge of cross-lingual POS tagger evaluation in absence of manually annotated test data. We put forth and evaluate two dictionary-based metrics. On the tasks of accuracy prediction and system ranking, we reveal that these metrics are reliable enough to approximate test set-based evaluation, and at the same time lean enough to support assessment for truly low-resource languages.
Recent work in geolocation has made several hypotheses about what linguistic markers are relevant to detect where people write from. In this paper, we examine six hypotheses against a corpus consisting of all geo-tagged tweets from the US, or whose geo-tags could be inferred, in a 19% sample of Twitter history. Our experiments lend support to all six hypotheses, including that spelling variants and hashtags are strong predictors of location. We also study what kinds of common nouns are predictive of location after controlling for named entities such as dolphins or sharks
Geolocation is the task of identifying a social media user’s primary location, and in natural language processing, there is a growing literature on to what extent automated analysis of social media posts can help. However, not all content features are equally revealing of a user’s location. In this paper, we evaluate nine name entity (NE) types. Using various metrics, we find that GEO-LOC, FACILITY and SPORT-TEAM are more informative for geolocation than other NE types. Using these types, we improve geolocation accuracy and reduce distance error over various famous text-based methods.
The problem of detecting scientific fraud using machine learning was recently introduced, with initial, positive results from a model taking into account various general indicators. The results seem to suggest that writing style is predictive of scientific fraud. We revisit these initial experiments, and show that the leave-one-out testing procedure they used likely leads to a slight over-estimate of the predictability, but also that simple models can outperform their proposed model by some margin. We go on to explore more abstract linguistic features, such as linguistic complexity and discourse structure, only to obtain negative results. Upon analyzing our models, we do see some interesting patterns, though: Scientific fraud, for examples, contains less comparison, as well as different types of hedging and ways of presenting logical reasoning.
We show that text readability prediction improves significantly from hard parameter sharing with models predicting first pass duration, total fixation duration and regression duration. Specifically, we induce multi-task Multilayer Perceptrons and Logistic Regression models over sentence representations that capture various aggregate statistics, from two different text readability corpora for English, as well as the Dundee eye-tracking corpus. Our approach leads to significant improvements over Single task learning and over previous systems. In addition, our improvements are consistent across train sample sizes, making our approach especially applicable to small datasets.
Syntactic annotation is costly and not available for the vast majority of the world’s languages. We show that sometimes we can do away with less labeled data by exploiting more readily available forms of mark-up. Specifically, we revisit an idea from Valentin Spitkovsky’s work (2010), namely that hyperlinks typically bracket syntactic constituents or chunks. We strengthen his results by showing that not only can hyperlinks help in low resource scenarios, exemplified here by Quechua, but learning from hyperlinks can also improve state-of-the-art NLP models for English newswire. We also present out-of-domain evaluation on English Ontonotes 4.0.
Automated processing of historical texts often relies on pre-normalization to modern word forms. Training encoder-decoder architectures to solve such problems typically requires a lot of training data, which is not available for the named task. We address this problem by using several novel encoder-decoder architectures, including a multi-task learning (MTL) architecture using a grapheme-to-phoneme dictionary as auxiliary data, pushing the state-of-the-art by an absolute 2% increase in performance. We analyze the induced models across 44 different texts from Early New High German. Interestingly, we observe that, as previously conjectured, multi-task learning can learn to focus attention during decoding, in ways remarkably similar to recently proposed attention mechanisms. This, we believe, is an important step toward understanding how MTL works.
Discourse segmentation is a crucial step in building end-to-end discourse parsers. However, discourse segmenters only exist for a few languages and domains. Typically they only detect intra-sentential segment boundaries, assuming gold standard sentence and token segmentation, and relying on high-quality syntactic parses and rich heuristics that are not generally available across languages and domains. In this paper, we propose statistical discourse segmenters for five languages and three domains that do not rely on gold pre-annotations. We also consider the problem of learning discourse segmenters when no labeled data is available for a language. Our fully supervised system obtains 89.5% F1 for English newswire, with slight drops in performance on other domains, and we report supervised and unsupervised (cross-lingual) results for five languages in total.
Keyphrase boundary classification (KBC) is the task of detecting keyphrases in scientific articles and labelling them with respect to predefined types. Although important in practice, this task is so far underexplored, partly due to the lack of labelled data. To overcome this, we explore several auxiliary tasks, including semantic super-sense tagging and identification of multi-word expressions, and cast the task as a multi-task learning problem with deep recurrent neural networks. Our multi-task models perform significantly better than previous state of the art approaches on two scientific KBC datasets, particularly for long keyphrases.
We propose a novel approach to cross-lingual part-of-speech tagging and dependency parsing for truly low-resource languages. Our annotation projection-based approach yields tagging and parsing models for over 100 languages. All that is needed are freely available parallel texts, and taggers and parsers for resource-rich languages. The empirical evaluation across 30 test languages shows that our method consistently provides top-level accuracies, close to established upper bounds, and outperforms several competitive baselines.
We launch the SemDaX corpus which is a recently completed Danish human-annotated corpus available through a CLARIN academic license. The corpus includes approx. 90,000 words, comprises six textual domains, and is annotated with sense inventories of different granularity. The aim of the developed corpus is twofold: i) to assess the reliability of the different sense annotation schemes for Danish measured by qualitative analyses and annotation agreement scores, and ii) to serve as training and test data for machine learning algorithms with the practical purpose of developing sense taggers for Danish. To these aims, we take a new approach to human-annotated corpus resources by double annotating a much larger part of the corpus than what is normally seen: for the all-words task we double annotated 60% of the material and for the lexical sample task 100%. We include in the corpus not only the adjucated files, but also the diverging annotations. In other words, we consider not all disagreement to be noise, but rather to contain valuable linguistic information that can help us improve our annotation schemes and our learning algorithms.
Natural-language processing of historical documents is complicated by the abundance of variant spellings and lack of annotated data. A common approach is to normalize the spelling of historical words to modern forms. We explore the suitability of a deep neural network architecture for this task, particularly a deep bi-LSTM network applied on a character level. Our model compares well to previously established normalization algorithms when evaluated on a diverse set of texts from Early New High German. We show that multi-task learning with additional normalization data can improve our model’s performance further.
Several recent studies have shown that eye movements during reading provide information about grammatical and syntactic processing, which can assist the induction of NLP models. All these studies have been limited to English, however. This study shows that gaze and part of speech (PoS) correlations largely transfer across English and French. This means that we can replicate previous studies on gaze-based PoS tagging for French, but also that we can use English gaze data to assist the induction of French NLP models.
We experiment with different ways of training LSTM networks to predict RST discourse trees. The main challenge for RST discourse parsing is the limited amounts of training data. We combat this by regularizing our models using task supervision from related tasks as well as alternative views on discourse structures. We show that a simple LSTM sequential discourse parser takes advantage of this multi-view and multi-task framework with 12-15% error reductions over our baseline (depending on the metric) and results that rival more complex state-of-the-art parsers.
We present two new NER datasets for Twitter; a manually annotated set of 1,467 tweets (kappa=0.942) and a set of 2,975 expert-corrected, crowdsourced NER annotated tweets from the dataset described in Finin et al. (2010). In our experiments with these datasets, we observe two important points: (a) language drift on Twitter is significant, and while off-the-shelf systems have been reported to perform well on in-sample data, they often perform poorly on new samples of tweets, (b) state-of-the-art performance across various datasets can be obtained from crowdsourced annotations, making it more feasible to “catch up” with language drift.
Several works in Natural Language Processing have recently looked into part-of-speech annotation of Twitter data and typically used their own data sets. Since conventions on Twitter change rapidly, models often show sample bias. Training on a combination of the existing data sets should help overcome this bias and produce more robust models than any trained on the individual corpora. Unfortunately, combining the existing corpora proves difficult: many of the corpora use proprietary tag sets that have little or no overlap. Even when mapped to a common tag set, the different corpora systematically differ in their treatment of various tags and tokens. This includes both pre-processing decisions, as well as default labels for frequent tokens, thus exhibiting data bias and label bias, respectively. Only if we address these biases can we combine the existing data sets to also overcome sample bias. We present a systematic study of several Twitter POS data sets, the problems of label and data bias, discuss their effects on model performance, and show how to overcome them to learn models that perform well on various test sets, achieving relative error reduction of up to 21%.
We present DSim, a new sentence aligned Danish monolingual parallel corpus extracted from 3701 pairs of news telegrams and corresponding professionally simplified short news articles. The corpus is intended for building automatic text simplification for adult readers. We compare DSim to different examples of monolingual parallel corpora, and we argue that this corpus is a promising basis for future development of automatic data-driven text simplification systems in Danish. The corpus contains both the collection of paired articles and a sentence aligned bitext, and we show that sentence alignment using simple tf*idf weighted cosine similarity scoring is on line with state―of―the―art when evaluated against a hand-aligned sample. The alignment results are compared to state of the art for English sentence alignment. We finally compare the source and simplified sides of the corpus in terms of lexical and syntactic characteristics and readability, and find that the one―to―many sentence aligned corpus is representative of the sentence simplifications observed in the unaligned collection of article pairs.