Solidarity is a crucial concept to understand social relations in societies. In this study, we investigate the frequency of (anti-)solidarity towards women and migrants in German parliamentary debates between 1867 and 2022. Using 2,864 manually annotated text snippets, we evaluate large language models (LLMs) like Llama 3, GPT-3.5, and GPT-4. We find that GPT-4 outperforms other models, approaching human annotation accuracy. Using GPT-4, we automatically annotate 18,300 further instances and find that solidarity with migrants outweighs anti-solidarity but that frequencies and solidarity types shift over time. Most importantly, group-based notions of (anti-)solidarity fade in favor of compassionate solidarity, focusing on the vulnerability of migrant groups, and exchange-based anti-solidarity, focusing on the lack of (economic) contribution. This study highlights the interplay of historical events, socio-economic needs, and political ideologies in shaping migration discourse and social cohesion.

Large language models (LLMs) have revolutionized NLP research. Notably, in-context learning enables their use as evaluation metrics for natural language generation, making them particularly advantageous in low-resource scenarios and time-restricted applications. In this work, we introduce PrExMe, a large-scale Prompt Exploration for Metrics, where we evaluate more than 720 prompt templates for open-source LLM-based metrics on machine translation (MT) and summarization datasets, totalling over 6.6M evaluations. This extensive comparison (1) benchmarks recent open-source LLMs as metrics and (2) explores the stability and variability of different prompting strategies. We discover that, on the one hand, there are scenarios for which prompts are stable. For instance, some LLMs show idiosyncratic preferences and favor to grade generated texts with textual labels while others prefer to return numeric scores. On the other hand, the stability of prompts and model rankings can be susceptible to seemingly innocuous changes. For example, changing the requested output format from “0 to 100” to "-1 to +1” can strongly affect the rankings in our evaluation. Our study contributes to understanding the impact of different prompting approaches on LLM-based metrics for MT and summarization evaluation, highlighting the most stable prompting patterns and potential limitations.

Natural Language Generation (NLG), and more generally generative AI, are among the currently most impactful research fields. Creative NLG, such as automatic poetry generation, is a fascinating niche in this area. While most previous research has focused on forms of the Turing test when evaluating automatic poetry generation — can humans distinguish between automatic and human generated poetry — we evaluate the diversity of automatically generated poetry (with a focus on quatrains), by comparing distributions of generated poetry to distributions of human poetry along structural, lexical, semantic and stylistic dimensions, assessing different model types (word vs. character-level, general purpose LLMs vs. poetry-specific models), including the very recent LLaMA3-8B, and types of fine-tuning (conditioned vs. unconditioned). We find that current automatic poetry systems are considerably underdiverse along multiple dimensions — they often do not rhyme sufficiently, are semantically too uniform and even do not match the length distribution of human poetry. Our experiments reveal, however, that style-conditioning and character-level modeling clearly increases diversity across virtually all dimensions we explore. Our identified limitations may serve as the basis for more genuinely diverse future poetry generation models.

State-of-the-art trainable machine translation evaluation metrics like xCOMET achieve high correlation with human judgment but rely on large encoders (up to 10.7B parameters), making them computationally expensive and inaccessible to researchers with limited resources. To address this issue, we investigate whether the knowledge stored in these large encoders can be compressed while maintaining quality. We employ distillation, quantization, and pruning techniques to create efficient xCOMET alternatives and introduce a novel data collection pipeline for efficient black-box distillation. Our experiments show that, using quantization, xCOMET can be compressed up to three times with no quality degradation. Additionally, through distillation, we create an 278M-sized xCOMET-lite metric, which has only 2.6% of xCOMET-XXL parameters, but retains 92.1% of its quality. Besides, it surpasses strong small-scale metrics like COMET-22 and BLEURT-20 on the WMT22 metrics challenge dataset by 6.4%, despite using 50% fewer parameters. All code, dataset, and models are available online.

State-of-the-art natural language generation evaluation metrics are based on black-box language models. Hence, recent works consider their explainability with the goals of better understandability for humans and better metric analysis, including failure cases. In contrast, we explicitly leverage explanations to boost the metrics’ performance. In particular, we perceive feature importance explanations as word-level scores, which we convert, via power means, into a segment-level score. We then combine this segment-level score with the original metric to obtain a better metric. Our tests show improvements for multiple metrics across MT and summarization datasets. While improvements on machine translation are small, they are strong for summarization. Notably, BMX with the LIME explainer and preselected parameters achieves an average improvement of 0.087 points in Spearman correlation on the system-level evaluation of SummEval.

While summarization has been extensively researched in natural language processing (NLP), cross-lingual cross-temporal summarization (CLCTS) is a largely unexplored area that has the potential to improve cross-cultural accessibility and understanding. This article comprehensively addresses the CLCTS task, including dataset creation, modeling, and evaluation. We (1) build the first CLCTS corpus with 328 instances for hDe-En (extended version with 455 instances) and 289 for hEn-De (extended version with 501 instances), leveraging historical fiction texts and Wikipedia summaries in English and German; (2) examine the effectiveness of popular transformer end-to-end models with different intermediate fine-tuning tasks; (3) explore the potential of GPT-3.5 as a summarizer; and (4) report evaluations from humans, GPT-4, and several recent automatic evaluation metrics. Our results indicate that intermediate task fine-tuned end-to-end models generate bad to moderate quality summaries while GPT-3.5, as a zero-shot summarizer, provides moderate to good quality outputs. GPT-3.5 also seems very adept at normalizing historical text. To assess data contamination in GPT-3.5, we design an adversarial attack scheme in which we find that GPT-3.5 performs slightly worse for unseen source documents compared to seen documents. Moreover, it sometimes hallucinates when the source sentences are inverted against its prior knowledge with a summarization accuracy of 0.67 for plot omission, 0.71 for entity swap, and 0.53 for plot negation. Overall, our regression results of model performances suggest that longer, older, and more complex source texts (all of which are more characteristic for historical language variants) are harder to summarize for all models, indicating the difficulty of the CLCTS task. Regarding evaluation, we observe that both the GPT-4 and BERTScore correlate moderately with human evaluations, implicating great potential for future improvement.

This study, conducted as part of the ReproHum project, aimed to replicate and evaluate the experiment presented in “Language Model as an Annotator: Exploring DialoGPT for Dialogue Summarization” by Feng et al. (2021). By employing DialoGPT, BART, and PGN models, the study assessed dialogue summarization’s informativeness. Based on the ReproHum project’s baselines, we conducted a human evaluation for the AIMI dataset, aiming to compare the results of the original study with our own experiments. Our objective is to contribute to the research on human evaluation and the reproducibility of the original study’s findings in the field of Natural Language Processing (NLP). Through this endeavor, we seek to enhance understanding and establish reliable benchmarks in human evaluation methodologies within the NLP domain.

Purpose: Based on the examples of English and German, we investigate to what extent parsers trained on modern variants of these languages can be transferred to older language levels without loss. Methods: We developed a treebank called DoTT (https://github.com/texttechnologylab/DoTT) which covers, roughly, the time period from 1800 until today, in conjunction with the further development of the annotation tool DependencyAnnotator. DoTT consists of a collection of diachronic corpora enriched with dependency annotations using 3 parsers, 6 pre-trained language models, 5 newly trained models for German, and two tag sets (TIGER and Universal Dependencies). To assess how the different parsers perform on texts from different time periods, we created a gold standard sample as a benchmark. Results: We found that the parsers/models perform quite well on modern texts (document-level LAS ranging from 82.89 to 88.54) and slightly worse on older texts, as expected (average document-level LAS 84.60 vs. 86.14), but not significantly. For German texts, the (German) TIGER scheme achieved slightly better results than UD. Conclusion: Overall, this result speaks for the transferability of parsers to past language levels, at least dating back until around 1800. This very transferability, it is however argued, means that studies of language change in the field of dependency syntax can draw on dependency distance but miss out on some grammatical phenomena.

State-of-the-art poetry generation systems are often complex. They either consist of task-specific model pipelines, incorporate prior knowledge in the form of manually created constraints, or both. In contrast, end-to-end models would not suffer from the overhead of having to model prior knowledge and could learn the nuances of poetry from data alone, reducing the degree of human supervision required. In this work, we investigate end-to-end poetry generation conditioned on styles such as rhyme, meter, and alliteration. We identify and address lack of training data and mismatching tokenization algorithms as possible limitations of past attempts. In particular, we successfully pre-train ByGPT5, a new token-free decoder-only language model, and fine-tune it on a large custom corpus of English and German quatrains annotated with our styles. We show that ByGPT5 outperforms other models such as mT5, ByT5, GPT-2 and ChatGPT, while also being more parameter efficient and performing favorably compared to humans. In addition, we analyze its runtime performance and demonstrate that it is not prone to memorization. We make our code, models, and datasets publicly available.

Protecting privacy in contemporary NLP models is gaining in importance. So does the need to mitigate social biases of such models. But can we have both at the same time? Existing research suggests that privacy preservation comes at the price of worsening biases in classification tasks. In this paper, we explore the extent to which this tradeoff really holds when we incorporate both privacy preservation and de-biasing techniques into training text generation models. How does improving the model along one dimension affect the other dimension as well as the utility of the model? We conduct an extensive set of experiments that include bias detection, privacy attacks, language modeling, and performance on downstream tasks.

Efficiency is a key property to foster inclusiveness and reduce environmental costs, especially in an era of LLMs. In this work, we provide a comprehensive evaluation of efficiency for MT evaluation metrics. Our approach involves replacing computation-intensive transformers with lighter alternatives and employing linear and quadratic approximations for alignment algorithms on top of LLM representations. We evaluate six (reference-free and reference-based) metrics across three MT datasets and examine 16 lightweight transformers. In addition, we look into the training efficiency of metrics like COMET by utilizing adapters. Our results indicate that (a) TinyBERT provides the optimal balance between quality and efficiency, (b) CPU speed-ups are more substantial than those on GPU; (c) WMD approximations yield no efficiency gains while reducing quality and (d) adapters enhance training efficiency (regarding backward pass speed and memory requirements) as well as, in some cases, metric quality. These findings can help to strike a balance between evaluation speed and quality, which is essential for effective NLG systems. Furthermore, our research contributes to the ongoing efforts to optimize NLG evaluation metrics with minimal impact on performance. To our knowledge, ours is the most comprehensive analysis of different aspects of efficiency for MT metrics conducted so far.

The vast majority of evaluation metrics for machine translation are supervised, i.e., (i) are trained on human scores, (ii) assume the existence of reference translations, or (iii) leverage parallel data. This hinders their applicability to cases where such supervision signals are not available. In this work, we develop fully unsupervised evaluation metrics. To do so, we leverage similarities and synergies between evaluation metric induction, parallel corpus mining, and MT systems. In particular, we use an unsupervised evaluation metric to mine pseudo-parallel data, which we use to remap deficient underlying vector spaces (in an iterative manner) and to induce an unsupervised MT system, which then provides pseudo-references as an additional component in the metric. Finally, we also induce unsupervised multilingual sentence embeddings from pseudo-parallel data. We show that our fully unsupervised metrics are effective, i.e., they beat supervised competitors on 4 out of our 5 evaluation datasets. We make our code publicly available.

Recently, there has been a growing interest in designing text generation systems from a discourse coherence perspective, e.g., modeling the interdependence between sentences. Still, recent BERT-based evaluation metrics are weak in recognizing coherence, and thus are not reliable in a way to spot the discourse-level improvements of those text generation systems. In this work, we introduce DiscoScore, a parametrized discourse metric, which uses BERT to model discourse coherence from different perspectives, driven by Centering theory. Our experiments encompass 16 non-discourse and discourse metrics, including DiscoScore and popular coherence models, evaluated on summarization and document-level machine translation (MT). We find that (i) the majority of BERT-based metrics correlate much worse with human rated coherence than early discourse metrics, invented a decade ago; (ii) the recent state-of-the-art BARTScore is weak when operated at system level—which is particularly problematic as systems are typically compared in this manner. DiscoScore, in contrast, achieves strong system-level correlation with human ratings, not only in coherence but also in factual consistency and other aspects, and surpasses BARTScore by over 10 correlation points on average. Further, aiming to understand DiscoScore, we provide justifications to the importance of discourse coherence for evaluation metrics, and explain the superiority of one variant over another. Our code is available at https://github.com/AIPHES/DiscoScore.

We report our efforts in identifying a set of previous human evaluations in NLP that would be suitable for a coordinated study examining what makes human evaluations in NLP more/less reproducible. We present our results and findings, which include that just 13% of papers had (i) sufficiently low barriers to reproduction, and (ii) enough obtainable information, to be considered for reproduction, and that all but one of the experiments we selected for reproduction was discovered to have flaws that made the meaningfulness of conducting a reproduction questionable. As a result, we had to change our coordinated study design from a reproduce approach to a standardise-then-reproduce-twice approach. Our overall (negative) finding that the great majority of human evaluations in NLP is not repeatable and/or not reproducible and/or too flawed to justify reproduction, paints a dire picture, but presents an opportunity for a rethink about how to design and report human evaluations in NLP.

We consider the end-to-end abstract-to-title generation problem, exploring seven recent transformer based models (including ChatGPT) fine-tuned on more than 30k abstract-title pairs from NLP and machine learning (ML) venues. As an extension, we also consider the harder problem of generating humorous paper titles. For the latter, we compile the first large-scale humor annotated dataset for scientific papers in the NLP/ML domains, comprising 2.6k titles. We evaluate all models using human and automatic metrics. Our human evaluation suggests that our best end-to-end system per-forms similarly to human authors (but arguably slightly worse). Generating funny titles is more difficult, however, and our automatic systems clearly underperform relative to humans and often learn dataset artefacts of humor. Finally, ChatGPT, without any fine-tuning, performs on the level of our best fine-tuned system.

Generative large language models (LLMs) have seen many breakthroughs over the last year. With an increasing number of parameters and pre-training data, they have shown remarkable capabilities to solve tasks with minimal or no task-related examples. Notably, LLMs have been successfully employed as evaluation metrics in text generation tasks. Strategies employed in this context differ in the choice of input prompts, the selection of samples for demonstration, and the methodology used to construct scores grading the generations. Approaches often differ in the input prompts, the samples that are selected for demonstration and the construction process of scores from the output. Within this context, we introduce the Eval4NLP 2023 shared task that asks participants to explore such approaches for machine translation evaluation and summarization eval- uation. Specifically, we select a list of allowed LLMs and disallow fine-tuning to ensure a focus on prompting. We test the approaches of the participants on a new reference-free test-set spanning 3 language pairs for machine transla- tion as well as a summarization dataset. Further, we present an overview of the approaches taken by the participants, present their results on the test set and analyze paths for future work. Fi- nally, as a separate track, we perform a human evaluation of the plausibility of explanations given by the LLMs and its effect on model performance. We make parts of our code and datasets available.

In this paper, we propose a retrieval-augmented in-context learning for natural language generation (NLG) evaluation. This method allows practitioners to utilize large language models (LLMs) for various NLG evaluation tasks without any fine-tuning. We apply our approach to Eval4NLP 2023 Shared Task in translation evaluation and summarization evaluation subtasks. The findings suggest that retrieval-augmented in-context learning is a promising approach for creating LLM-based evaluation metrics for NLG. Further research directions include exploring the performance of various publicly available LLM models and identifying which LLM properties help boost the quality of the metric.

Recently proposed BERT-based evaluation metrics for text generation perform well on standard benchmarks but are vulnerable to adversarial attacks, e.g., relating to information correctness. We argue that this stems (in part) from the fact that they are models of semantic similarity. In contrast, we develop evaluation metrics based on Natural Language Inference (NLI), which we deem a more appropriate modeling. We design a preference-based adversarial attack framework and show that our NLI based metrics are much more robust to the attacks than the recent BERT-based metrics. On standard benchmarks, our NLI based metrics outperform existing summarization metrics, but perform below SOTA MT metrics. However, when combining existing metrics with our NLI metrics, we obtain both higher adversarial robustness (15%–30%) and higher quality metrics as measured on standard benchmarks (+5% to 30%).

Machine translation is natural language generation (NLG) problem of translating source text from one language to another. As every task in machine learning domain it requires to have evaluation metric. The most obvious one is human evaluation but it is expensive in case of money and time consumption. In last years with appearing of pretrained transformer architectures and large language models (LLMs) state-of-the-art results in automatic machine translation evaluation got a huge quality step in terms of correlation with expert assessment. We introduce MRE-Score, seMantically-informed Regression Encoder Score, the approach with constructing automatic machine translation evaluation system based on regression encoder and contrastive pretraining for downstream problem.

Reproducibility is of utmost concern in machine learning and natural language processing (NLP). In the field of natural language generation (especially machine translation), the seminal paper of Post (2018) has pointed out problems of reproducibility of the dominant metric, BLEU, at the time of publication. Nowadays, BERT-based evaluation metrics considerably outperform BLEU. In this paper, we ask whether results and claims from four recent BERT-based metrics can be reproduced. We find that reproduction of claims and results often fails because of (i) heavy undocumented preprocessing involved in the metrics, (ii) missing code and (iii) reporting weaker results for the baseline metrics. (iv) In one case, the problem stems from correlating not to human scores but to a wrong column in the csv file, inflating scores by 5 points. Motivated by the impact of preprocessing, we then conduct a second study where we examine its effects more closely (for one of the metrics). We find that preprocessing can have large effects, especially for highly inflectional languages. In this case, the effect of preprocessing may be larger than the effect of the aggregation mechanism (e.g., greedy alignment vs. Word Mover Distance).

We report the results of the WMT 2022 shared task on Quality Estimation, in which the challenge is to predict the quality of the output of neural machine translation systems at the word and sentence levels, without access to reference translations. This edition introduces a few novel aspects and extensions that aim to enable more fine-grained, and explainable quality estimation approaches. We introduce an updated quality annotation scheme using Multidimensional Quality Metrics to obtain sentence- and word-level quality scores for three language pairs. We also extend the Direct Assessments and post-edit data (MLQE-PE) to new language pairs: we present a novel and large dataset on English-Marathi, as well as a zero-shot test set on English-Yoruba. Further, we include an explainability sub-task for all language pairs and present a new format of a critical error detection task for two new language pairs. Participants from 11 different teams submitted altogether 991 systems to different task variants and language pairs.

The evaluation of recent embedding-based evaluation metrics for text generation is primarily based on measuring their correlation with human evaluations on standard benchmarks. However, these benchmarks are mostly from similar domains to those used for pretraining word embeddings. This raises concerns about the (lack of) generalization of embedding-based metrics to new and noisy domains that contain a different vocabulary than the pretraining data. In this paper, we examine the robustness of BERTScore, one of the most popular embedding-based metrics for text generation. We show that (a) an embedding-based metric that has the highest correlation with human evaluations on a standard benchmark can have the lowest correlation if the amount of input noise or unknown tokens increases, (b) taking embeddings from the first layer of pretrained models improves the robustness of all metrics, and (c) the highest robustness is achieved when using character-level embeddings, instead of token-based embeddings, from the first layer of the pretrained model.

We introduce the well-established social scientific concept of social solidarity and its contestation, anti-solidarity, as a new problem setting to supervised machine learning in NLP to assess how European solidarity discourses changed before and after the COVID-19 outbreak was declared a global pandemic. To this end, we annotate 2.3k English and German tweets for (anti-)solidarity expressions, utilizing multiple human annotators and two annotation approaches (experts vs. crowds). We use these annotations to train a BERT model with multiple data augmentation strategies. Our augmented BERT model that combines both expert and crowd annotations outperforms the baseline BERT classifier trained with expert annotations only by over 25 points, from 58% macro-F1 to almost 85%. We use this high-quality model to automatically label over 270k tweets between September 2019 and December 2020. We then assess the automatically labeled data for how statements related to European (anti-)solidarity discourses developed over time and in relation to one another, before and during the COVID-19 crisis. Our results show that solidarity became increasingly salient and contested during the crisis. While the number of solidarity tweets remained on a higher level and dominated the discourse in the scrutinized time frame, anti-solidarity tweets initially spiked, then decreased to (almost) pre-COVID-19 values before rising to a stable higher level until the end of 2020.

Evaluation in NLP is usually done by comparing the scores of competing systems independently averaged over a common set of test instances. In this work, we question the use of averages for aggregating evaluation scores into a final number used to decide which system is best, since the average, as well as alternatives such as the median, ignores the pairing arising from the fact that systems are evaluated on the same test instances. We illustrate the importance of taking the instancelevel pairing of evaluation scores into account and demonstrate, both theoretically and empirically, the advantages of aggregation methods based on pairwise comparisons, such as the Bradley–Terry (BT) model, a mechanism based on the estimated probability that a given system scores better than another on the test set. By re-evaluating 296 real NLP evaluation setups across four tasks and 18 evaluation metrics, we show that the choice of aggregation mechanism matters and yields different conclusions as to which systems are state of the art in about 30% of the setups. To facilitate the adoption of pairwise evaluation, we release a practical tool for performing the full analysis of evaluation scores with the mean, median, BT, and two variants of BT (Elo and TrueSkill), alongside functionality for appropriate statistical testing.

In this paper, we introduce the Eval4NLP-2021 shared task on explainable quality estimation. Given a source-translation pair, this shared task requires not only to provide a sentence-level score indicating the overall quality of the translation, but also to explain this score by identifying the words that negatively impact translation quality. We present the data, annotation guidelines and evaluation setup of the shared task, describe the six participating systems, and analyze the results. To the best of our knowledge, this is the first shared task on explainable NLP evaluation metrics. Datasets and results are available at https://github.com/eval4nlp/SharedTask2021.

Cross-lingual representations have the potential to make NLP techniques available to the vast majority of languages in the world. However, they currently require large pretraining corpora or access to typologically similar languages. In this work, we address these obstacles by removing language identity signals from multilingual embeddings. We examine three approaches for this: (i) re-aligning the vector spaces of target languages (all together) to a pivot source language; (ii) removing language-specific means and variances, which yields better discriminativeness of embeddings as a by-product; and (iii) increasing input similarity across languages by removing morphological contractions and sentence reordering. We evaluate on XNLI and reference-free MT evaluation across 19 typologically diverse languages. Our findings expose the limitations of these approaches—unlike vector normalization, vector space re-alignment and text normalization do not achieve consistent gains across encoders and languages. Due to the approaches’ additive effects, their combination decreases the cross-lingual transfer gap by 8.9 points (m-BERT) and 18.2 points (XLM-R) on average across all tasks and languages, however.

In this work, we design an end-to-end model for poetry generation based on conditioned recurrent neural network (RNN) language models whose goal is to learn stylistic features (poem length, sentiment, alliteration, and rhyming) from examples alone. We show this model successfully learns the ‘meaning’ of length and sentiment, as we can control it to generate longer or shorter as well as more positive or more negative poems. However, the model does not grasp sound phenomena like alliteration and rhyming, but instead exploits low-level statistical cues. Possible reasons include the size of the training data, the relatively low frequency and difficulty of these sublexical phenomena as well as model biases. We show that more recent GPT-2 models also have problems learning sublexical phenomena such as rhyming from examples alone.

This paper describes our contribution to the Shared Task ReproGen by Belz et al. (2021), which investigates the reproducibility of human evaluations in the context of Natural Language Generation. We selected the paper “Generation of Company descriptions using concept-to-text and text-to-text deep models: data set collection and systems evaluation” (Qader et al., 2018) and aimed to replicate, as closely to the original as possible, the human evaluation and the subsequent comparison between the human judgements and the automatic evaluation metrics. Here, we first outline the text generation task of the paper of Qader et al. (2018). Then, we document how we approached our replication of the paper’s human evaluation. We also discuss the difficulties we encountered and which information was missing. Our replication has medium to strong correlation (0.66 Spearman overall) with the original results of Qader et al. (2018), but due to the missing information about how Qader et al. (2018) compared the human judgements with the metric scores, we have refrained from reproducing this comparison.

This paper presents our submissions to the WMT2021 Shared Task on Quality Estimation, Task 1 Sentence-Level Direct Assessment. While top-performing approaches utilize massively multilingual Transformer-based language models which have been pre-trained on all target languages of the task, the resulting insights are limited, as it is unclear how well the approach performs on languages unseen during pre-training; more problematically, these approaches do not provide any solutions for extending the model to new languages or unseen scripts—arguably one of the objectives of this shared task. In this work, we thus focus on utilizing massively multilingual language models which only partly cover the target languages during their pre-training phase. We extend the model to new languages and unseen scripts using recent adapter-based methods and achieve on par performance or even surpass models pre-trained on the respective languages.

Evaluation metrics are a key ingredient for progress of text generation systems. In recent years, several BERT-based evaluation metrics have been proposed (including BERTScore, MoverScore, BLEURT, etc.) which correlate much better with human assessment of text generation quality than BLEU or ROUGE, invented two decades ago. However, little is known what these metrics, which are based on black-box language model representations, actually capture (it is typically assumed they model semantic similarity). In this work, we use a simple regression based global explainability technique to disentangle metric scores along linguistic factors, including semantics, syntax, morphology, and lexical overlap. We show that the different metrics capture all aspects to some degree, but that they are all substantially sensitive to lexical overlap, just like BLEU and ROUGE. This exposes limitations of these novelly proposed metrics, which we also highlight in an adversarial test scenario.

We probe the layers in multilingual BERT (mBERT) for phylogenetic and geographic language signals across 100 languages and compute language distances based on the mBERT representations. We 1) employ the language distances to infer and evaluate language trees, finding that they are close to the reference family tree in terms of quartet tree distance, 2) perform distance matrix regression analysis, finding that the language distances can be best explained by phylogenetic and worst by structural factors and 3) present a novel measure for measuring diachronic meaning stability (based on cross-lingual representation variability) which correlates significantly with published ranked lists based on linguistic approaches. Our results contribute to the nascent field of typological interpretability of cross-lingual text representations.

We introspect black-box sentence embeddings by conditionally generating from them with the objective to retrieve the underlying discrete sentence. We perceive of this as a new unsupervised probing task and show that it correlates well with downstream task performance. We also illustrate how the language generated from different encoders differs. We apply our approach to generate sentence analogies from sentence embeddings.

Adversarial attacks are label-preserving modifications to inputs of machine learning classifiers designed to fool machines but not humans. Natural Language Processing (NLP) has mostly focused on high-level attack scenarios such as paraphrasing input texts. We argue that these are less realistic in typical application scenarios such as in social media, and instead focus on low-level attacks on the character-level. Guided by human cognitive abilities and human robustness, we propose the first large-scale catalogue and benchmark of low-level adversarial attacks, which we dub Zéroe, encompassing nine different attack modes including visual and phonetic adversaries. We show that RoBERTa, NLP’s current workhorse, fails on our attacks. Our dataset provides a benchmark for testing robustness of future more human-like NLP models.

This paper describes the system Clustering on Manifolds of Contextualized Embeddings (CMCE) submitted to the SemEval-2020 Task 1 on Unsupervised Lexical Semantic Change Detection. Subtask 1 asks to identify whether or not a word gained/lost a sense across two time periods. Subtask 2 is about computing a ranking of words according to the amount of change their senses underwent. Our system uses contextualized word embeddings from MBERT, whose dimensionality we reduce with an autoencoder and the UMAP algorithm, to be able to use a wider array of clustering algorithms that can automatically determine the number of clusters. We use Hierarchical Density Based Clustering (HDBSCAN) and compare it to Gaussian MixtureModels (GMMs) and other clustering algorithms. Remarkably, with only 10 dimensional MBERT embeddings (reduced from the original size of 768), our submitted model performs best on subtask 1 for English and ranks third in subtask 2 for English. In addition to describing our system, we discuss our hyperparameter configurations and examine why our system lags behind for the other languages involved in the shared task (German, Swedish, Latin). Our code is available at https://github.com/DavidRother/semeval2020-task1

Most approaches to emotion analysis of social media, literature, news, and other domains focus exclusively on basic emotion categories as defined by Ekman or Plutchik. However, art (such as literature) enables engagement in a broader range of more complex and subtle emotions. These have been shown to also include mixed emotional responses. We consider emotions in poetry as they are elicited in the reader, rather than what is expressed in the text or intended by the author. Thus, we conceptualize a set of aesthetic emotions that are predictive of aesthetic appreciation in the reader, and allow the annotation of multiple labels per line to capture mixed emotions within their context. We evaluate this novel setting in an annotation experiment both with carefully trained experts and via crowdsourcing. Our annotation with experts leads to an acceptable agreement of k = .70, resulting in a consistent dataset for future large scale analysis. Finally, we conduct first emotion classification experiments based on BERT, showing that identifying aesthetic emotions is challenging in our data, with up to .52 F1-micro on the German subset. Data and resources are available at https://github.com/tnhaider/poetry-emotion.

We study unsupervised multi-document summarization evaluation metrics, which require neither human-written reference summaries nor human annotations (e.g. preferences, ratings, etc.). We propose SUPERT, which rates the quality of a summary by measuring its semantic similarity with a pseudo reference summary, i.e. selected salient sentences from the source documents, using contextualized embeddings and soft token alignment techniques. Compared to the state-of-the-art unsupervised evaluation metrics, SUPERT correlates better with human ratings by 18- 39%. Furthermore, we use SUPERT as rewards to guide a neural-based reinforcement learning summarizer, yielding favorable performance compared to the state-of-the-art unsupervised summarizers. All source code is available at https://github.com/yg211/acl20-ref-free-eval.

Evaluation of cross-lingual encoders is usually performed either via zero-shot cross-lingual transfer in supervised downstream tasks or via unsupervised cross-lingual textual similarity. In this paper, we concern ourselves with reference-free machine translation (MT) evaluation where we directly compare source texts to (sometimes low-quality) system translations, which represents a natural adversarial setup for multilingual encoders. Reference-free evaluation holds the promise of web-scale comparison of MT systems. We systematically investigate a range of metrics based on state-of-the-art cross-lingual semantic representations obtained with pretrained M-BERT and LASER. We find that they perform poorly as semantic encoders for reference-free MT evaluation and identify their two key limitations, namely, (a) a semantic mismatch between representations of mutual translations and, more prominently, (b) the inability to punish “translationese”, i.e., low-quality literal translations. We propose two partial remedies: (1) post-hoc re-alignment of the vector spaces and (2) coupling of semantic-similarity based metrics with target-side language modeling. In segment-level MT evaluation, our best metric surpasses reference-based BLEU by 5.7 correlation points.

Sentence encoders map sentences to real valued vectors for use in downstream applications. To peek into these representations—e.g., to increase interpretability of their results—probing tasks have been designed which query them for linguistic knowledge. However, designing probing tasks for lesser-resourced languages is tricky, because these often lack largescale annotated data or (high-quality) dependency parsers as a prerequisite of probing task design in English. To investigate how to probe sentence embeddings in such cases, we investigate sensitivity of probing task results to structural design choices, conducting the first such large scale study. We show that design choices like size of the annotated probing dataset and type of classifier used for evaluation do (sometimes substantially) influence probing outcomes. We then probe embeddings in a multilingual setup with design choices that lie in a ‘stable region’, as we identify for English, and find that results on English do not transfer to other languages. Fairer and more comprehensive sentence-level probing evaluation should thus be carried out on multiple languages in the future.

A substantial overlap of coreferent mentions in the CoNLL dataset magnifies the recent progress on coreference resolution. This is because the CoNLL benchmark fails to evaluate the ability of coreference resolvers that requires linking novel mentions unseen at train time. In this work, we create a new dataset based on CoNLL, which largely decreases mention overlaps in the entire dataset and exposes the limitations of published resolvers on two aspects—lexical inference ability and understanding of low-level orthographic noise. Our findings show (1) the requirements for embeddings, used in resolvers, and for coreference resolutions are, by design, in conflict and (2) adversarial approaches are sometimes not legitimate to mitigate the obstacles, as they may falsely introduce mention overlaps in adversarial training and test sets, thus giving an inflated impression for the improvements.

Obstacles hindering the development of capsule networks for challenging NLP applications include poor scalability to large output spaces and less reliable routing processes. In this paper, we introduce: (i) an agreement score to evaluate the performance of routing processes at instance-level; (ii) an adaptive optimizer to enhance the reliability of routing; (iii) capsule compression and partial routing to improve the scalability of capsule networks. We validate our approach on two NLP tasks, namely: multi-label text classification and question answering. Experimental results show that our approach considerably improves over strong competitors on both tasks. In addition, we gain the best results in low-resource settings with few training instances.

Peer review is a core element of the scientific process, particularly in conference-centered fields such as ML and NLP. However, only few studies have evaluated its properties empirically. Aiming to fill this gap, we present a corpus that contains over 4k reviews and 1.2k author responses from ACL-2018. We quantitatively and qualitatively assess the corpus. This includes a pilot study on paper weaknesses given by reviewers and on quality of author responses. We then focus on the role of the rebuttal phase, and propose a novel task to predict after-rebuttal (i.e., final) scores from initial reviews and author responses. Although author responses do have a marginal (and statistically significant) influence on the final scores, especially for borderline papers, our results suggest that a reviewer’s final score is largely determined by her initial score and the distance to the other reviewers’ initial scores. In this context, we discuss the conformity bias inherent to peer reviewing, a bias that has largely been overlooked in previous research. We hope our analyses will help better assess the usefulness of the rebuttal phase in NLP conferences.

Visual modifications to text are often used to obfuscate offensive comments in social media (e.g., “!d10t”) or as a writing style (“1337” in “leet speak”), among other scenarios. We consider this as a new type of adversarial attack in NLP, a setting to which humans are very robust, as our experiments with both simple and more difficult visual perturbations demonstrate. We investigate the impact of visual adversarial attacks on current NLP systems on character-, word-, and sentence-level tasks, showing that both neural and non-neural models are, in contrast to humans, extremely sensitive to such attacks, suffering performance decreases of up to 82%. We then explore three shielding methods—visual character embeddings, adversarial training, and rule-based recovery—which substantially improve the robustness of the models. However, the shielding methods still fall behind performances achieved in non-attack scenarios, which demonstrates the difficulty of dealing with visual attacks.

A robust evaluation metric has a profound impact on the development of text generation systems. A desirable metric compares system output against references based on their semantics rather than surface forms. In this paper we investigate strategies to encode system and reference texts to devise a metric that shows a high correlation with human judgment of text quality. We validate our new metric, namely MoverScore, on a number of text generation tasks including summarization, machine translation, image captioning, and data-to-text generation, where the outputs are produced by a variety of neural and non-neural systems. Our findings suggest that metrics combining contextualized representations with a distance measure perform the best. Such metrics also demonstrate strong generalization capability across tasks. For ease-of-use we make our metrics available as web service.

Deep learning models continuously break new records across different NLP tasks. At the same time, their success exposes weaknesses of model evaluation. Here, we compile several key pitfalls of evaluation of sentence embeddings, a currently very popular NLP paradigm. These pitfalls include the comparison of embeddings of different sizes, normalization of embeddings, and the low (and diverging) correlations between transfer and probing tasks. Our motivation is to challenge the current evaluation of sentence embeddings and to provide an easy-to-access reference for future research. Based on our insights, we also recommend better practices for better future evaluations of sentence embeddings.

Due to its semantic succinctness and novelty of expression, poetry is a great test-bed for semantic change analysis. However, so far there is a scarcity of large diachronic corpora. Here, we provide a large corpus of German poetry which consists of about 75k poems with more than 11 million tokens, with poems ranging from the 16th to early 20th century. We then track semantic change in this corpus by investigating the rise of tropes (‘love is magic’) over time and detecting change points of meaning, which we find to occur particularly within the German Romantic period. Additionally, through self-similarity, we reconstruct literary periods and find evidence that the law of linear semantic change also applies to poetry.

We investigate whether and where multi-task learning (MTL) can improve performance on NLP problems related to argumentation mining (AM), in particular argument component identification. Our results show that MTL performs particularly well (and better than single-task learning) when little training data is available for the main task, a common scenario in AM. Our findings challenge previous assumptions that conceptualizations across AM datasets are divergent and that MTL is difficult for semantic or higher-level tasks.

Argument mining is a core technology for enabling argument search in large corpora. However, most current approaches fall short when applied to heterogeneous texts. In this paper, we present an argument retrieval system capable of retrieving sentential arguments for any given controversial topic. By analyzing the highest-ranked results extracted from Web sources, we found that our system covers 89% of arguments found in expert-curated lists of arguments from an online debate portal, and also identifies additional valid arguments.

Argumentation mining (AM) requires the identification of complex discourse structures and has lately been applied with success monolingually. In this work, we show that the existing resources are, however, not adequate for assessing cross-lingual AM, due to their heterogeneity or lack of complexity. We therefore create suitable parallel corpora by (human and machine) translating a popular AM dataset consisting of persuasive student essays into German, French, Spanish, and Chinese. We then compare (i) annotation projection and (ii) bilingual word embeddings based direct transfer strategies for cross-lingual AM, finding that the former performs considerably better and almost eliminates the loss from cross-lingual transfer. Moreover, we find that annotation projection works equally well when using either costly human or cheap machine translations. Our code and data are available at http://github.com/UKPLab/coling2018-xling_argument_mining.

Non-literal language phenomena such as idioms or metaphors are commonly studied in isolation from each other in NLP. However, often similar definitions and features are being used for different phenomena, challenging the distinction. Instead, we propose to view the detection problem as a generalized non-literal language classification problem. In this paper we investigate multi-task learning for related non-literal language phenomena. We show that in contrast to simply joining the data of multiple tasks, multi-task learning consistently improves upon four metaphor and idiom detection tasks in two languages, English and German. Comparing two state-of-the-art multi-task learning architectures, we also investigate when soft parameter sharing and learned information flow can be beneficial for our related tasks. We make our adapted code publicly available.

In this paper, we tackle four different tasks of non-literal language classification: token and construction level metaphor detection, classification of idiomatic use of infinitive-verb compounds, and classification of non-literal particle verbs. One of the tasks operates on the token level, while the three other tasks classify constructions such as “hot topic” or “stehen lassen” (“to allow sth. to stand” vs. “to abandon so.”). The two metaphor detection tasks are in English, while the two non-literal language detection tasks are in German. We propose a simple context-encoding LSTM model and show that it outperforms the state-of-the-art on two tasks. Additionally, we experiment with different embeddings for the token level metaphor detection task and find that 1) their performance varies according to the genre, and 2) word2vec embeddings perform best on 3 out of 4 genres, despite being one of the simplest tested model. In summary, we present a large-scale analysis of a neural model for non-literal language classification (i) at different granularities, (ii) in different languages, (iii) over different non-literal language phenomena.

We consider unsupervised cross-lingual transfer on two tasks, viz., sentence-level argumentation mining and standard POS tagging. We combine direct transfer using bilingual embeddings with annotation projection, which projects labels across unlabeled parallel data. We do so by either merging respective source and target language datasets or alternatively by using multi-task learning. Our combination strategy considerably improves upon both direct transfer and projection with few available parallel sentences, the most realistic scenario for many low-resource target languages.

Activation functions play a crucial role in neural networks because they are the nonlinearities which have been attributed to the success story of deep learning. One of the currently most popular activation functions is ReLU, but several competitors have recently been proposed or ‘discovered’, including LReLU functions and swish. While most works compare newly proposed activation functions on few tasks (usually from image classification) and against few competitors (usually ReLU), we perform the first largescale comparison of 21 activation functions across eight different NLP tasks. We find that a largely unknown activation function performs most stably across all tasks, the so-called penalized tanh function. We also show that it can successfully replace the sigmoid and tanh gates in LSTM cells, leading to a 2 percentage point (pp) improvement over the standard choices on a challenging NLP task.

This paper describes our approach to the SemEval 2017 Task 10: Extracting Keyphrases and Relations from Scientific Publications, specifically to Subtask (B): Classification of identified keyphrases. We explored three different deep learning approaches: a character-level convolutional neural network (CNN), a stacked learner with an MLP meta-classifier, and an attention based Bi-LSTM. From these approaches, we created an ensemble of differently hyper-parameterized systems, achieving a micro-F₁-score of 0.63 on the test data. Our approach ranks 2nd (score of 1st placed system: 0.64) out of four according to this official score. However, we erroneously trained 2 out of 3 neural nets (the stacker and the CNN) on only roughly 15% of the full data, namely, the original development set. When trained on the full data (training+development), our ensemble has a micro-F₁-score of 0.69. Our code is available from https://github.com/UKPLab/semeval2017-scienceie.

Argument mining has become a popular research area in NLP. It typically includes the identification of argumentative components, e.g. claims, as the central component of an argument. We perform a qualitative analysis across six different datasets and show that these appear to conceptualize claims quite differently. To learn about the consequences of such different conceptualizations of claim for practical applications, we carried out extensive experiments using state-of-the-art feature-rich and deep learning systems, to identify claims in a cross-domain fashion. While the divergent conceptualization of claims in different datasets is indeed harmful to cross-domain classification, we show that there are shared properties on the lexical level as well as system configurations that can help to overcome these gaps.

We investigate neural techniques for end-to-end computational argumentation mining (AM). We frame AM both as a token-based dependency parsing and as a token-based sequence tagging problem, including a multi-task learning setup. Contrary to models that operate on the argument component level, we find that framing AM as dependency parsing leads to subpar performance results. In contrast, less complex (local) tagging models based on BiLSTMs perform robustly across classification scenarios, being able to catch long-range dependencies inherent to the AM problem. Moreover, we find that jointly learning ‘natural’ subtasks, in a multi-task learning setup, improves performance.

This paper relates to the challenge of morphological tagging and lemmatization in morphologically rich languages by example of German and Latin. We focus on the question what a practitioner can expect when using state-of-the-art solutions out of the box. Moreover, we contrast these with old(er) methods and implementations for POS tagging. We examine to what degree recent efforts in tagger development are reflected by improved accuracies ― and at what cost, in terms of training and processing time. We also conduct in-domain vs. out-domain evaluation. Out-domain evaluations are particularly insightful because the distribution of the data which is being tagged by a user will typically differ from the distribution on which the tagger has been trained. Furthermore, two lemmatization techniques are evaluated. Finally, we compare pipeline tagging vs. a tagging approach that acknowledges dependencies between inflectional categories.

We analyze the performance of encoder-decoder neural models and compare them with well-known established methods. The latter represent different classes of traditional approaches that are applied to the monotone sequence-to-sequence tasks OCR post-correction, spelling correction, grapheme-to-phoneme conversion, and lemmatization. Such tasks are of practical relevance for various higher-level research fields including digital humanities, automatic text correction, and speech recognition. We investigate how well generic deep-learning approaches adapt to these tasks, and how they perform in comparison with established and more specialized methods, including our own adaptation of pruned CRFs.

We study the role of the second language in bilingual word embeddings in monolingual semantic evaluation tasks. We find strongly and weakly positive correlations between down-stream task performance and second language similarity to the target language. Additionally, we show how bilingual word embeddings can be employed for the task of semantic language classification and that joint semantic spaces vary in meaningful ways across second languages. Our results support the hypothesis that semantic language similarity is influenced by both structural similarity as well as geography/contact.