We evaluate the ability of large language models (LLMs) to provide PropBank semantic role label annotations across different realizations of the same verbs in transitive, intransitive, and middle voice constructions. In order to assess the meta-linguistic capabilities of LLMs as well as their ability to glean such capabilities through in-context learning, we evaluate the models in a zero-shot setting, in a setting where it is given three examples of another verb used in transitive, intransitive, and middle voice constructions, and finally in a setting where it is given the examples as well as the correct sense and roleset information. We find that zero-shot knowledge of PropBank annotation is almost nonexistent. The largest model evaluated, GPT-4, achieves the best performance in the setting where it is given both examples and the correct roleset in the prompt, demonstrating that larger models can ascertain some meta-linguistic capabilities through in-context learning. However, even in this setting, which is simpler than the task of a human in PropBank annotation, the model achieves only 48% accuracy in marking numbered arguments correctly. To ensure transparency and reproducibility, we publicly release our dataset and model responses.
All existing transformer-based approaches to NLP using subword tokenisation algorithms encode whitespace (word boundary information) through the use of special space symbols (such as ## or _) forming part of tokens. These symbols have been shown to a) lead to reduced morphological validity of tokenisations, and b) give substantial vocabulary redundancy. As such, removing these symbols has been shown to have a beneficial effect on the processing of morphologically complex words for transformer encoders in the pretrain-finetune paradigm. In this work, we explore whether word boundary information is at all useful to such models. In particular, we train transformer encoders across four different training scales, and investigate several alternative approaches to including word boundary information, evaluating on two languages (English and Finnish) with a range of tasks across different domains and problem set-ups: sentence classification datasets, NER (for token-level classification), and two classification datasets involving complex words (Superbizarre and FLOTA). Overall, through an extensive experimental setup that includes the pre-training of 35 models, we find no substantial improvements from our alternative approaches, suggesting that modifying tokenisers to remove word boundary information isn’t leading to a loss of useful information.
Large language models, comprising billions of parameters and pre-trained on extensive web-scale corpora, have been claimed to acquire certain capabilities without having been specifically trained on them. These capabilities, referred to as “emergent abilities,” have been a driving force in discussions regarding the potentials and risks of language models. A key challenge in evaluating emergent abilities is that they are confounded by model competencies that arise through alternative prompting techniques, including in-context learning, which is the ability of models to complete a task based on a few examples. We present a novel theory that explains emergent abilities, taking into account their potential confounding factors, and rigorously substantiate this theory through over 1000 experiments. Our findings suggest that purported emergent abilities are not truly emergent, but result from a combination of in-context learning, model memory, and linguistic knowledge. Our work is a foundational step in explaining language model performance, providing a template for their efficient use and clarifying the paradox of their ability to excel in some instances while faltering in others. Thus, we demonstrate that their capabilities should not be overestimated.
Large Language Models (LLMs) have been developed without a theoretical framework, yet we posit that evaluating and improving LLMs will benefit from the development of theoretical frameworks that enable comparison of the structures of human language and the model of language built up by LLMs through the processing of text. In service of this goal, we develop the Construction Grammar Schematicity (“CoGS”) corpus of 10 distinct English constructions, where the constructions vary with respect to schematicity, or in other words the level to which constructional slots require specific, fixed lexical items, or can be filled with a variety of elements that fulfill a particular semantic role of the slot. Our corpus constructions are carefully curated to range from substantive, frozen constructions (e.g., Let-alone) to entirely schematic constructions (e.g., Resultative). The corpus was collected to allow us to probe LLMs for constructional information at varying levels of abstraction. We present our own probing experiments using this corpus, which clearly demonstrate that even the largest LLMs are limited to more substantive constructions and do not exhibit recognition of the similarity of purely schematic constructions. We publicly release our dataset, prompts, and associated model responses.
Code-switching is a prevalent linguistic phenomenon in which multilingual individuals seamlessly alternate between languages. Despite its widespread use online and recent research trends in this area, research in code-switching presents unique challenges, primarily stemming from the scarcity of labelled data and available resources. In this study we investigate how pre-trained Language Models handle code-switched text in three dimensions: a) the ability of PLMs to detect code-switched text, b) variations in the structural information that PLMs utilise to capture code-switched text, and c) the consistency of semantic information representation in code-switched text. To conduct a systematic and controlled evaluation of the language models in question, we create a novel dataset of well-formed naturalistic code-switched text along with parallel translations into the source languages. Our findings reveal that pre-trained language models are effective in generalising to code-switched text, shedding light on abilities of these models to generalise representations to CS corpora. We release all our code and data, including the novel corpus, at https://github.com/francesita/code-mixed-probes.
This paper measures the skew in how well two families of LLMs represent diverse geographic populations. A spatial probing task is used with geo-referenced corpora to measure the degree to which pre-trained language models from the OPT and BLOOM series represent diverse populations around the world. Results show that these models perform much better for some populations than others. In particular, populations across the US and the UK are represented quite well while those in South and Southeast Asia are poorly represented. Analysis shows that both families of models largely share the same skew across populations. At the same time, this skew cannot be fully explained by sociolinguistic factors, economic factors, or geographic factors. The basic conclusion from this analysis is that pre-trained models do not equally represent the world’s population: there is a strong skew towards specific geographic populations. This finding challenges the idea that a single model can be used for all populations.
Large language models (LLMs) have shown promising results in a wide array of generative NLP tasks, such as summarization and machine translation. In the context of narrative generation, however, existing models still do not capture factors that contribute to producing consistent text. For instance, it is logical that a piece of text or a story should be uniformly readable throughout and that this form of complexity should be controllable. As such, if the complexity of an input text prompt is rated first-grade reading level in the Flesch Reading Ease test, then the generated text continuing the plot should also be within this range of complexity. With this in mind, we introduce Uniform Complexity for Text Generation (UCTG), a new benchmark test which raises the challenge of making generative models observe uniform linguistic properties with respect to prompts. We experiment with over 150+ linguistically and cognitively motivated features for evaluating text complexity in humans and generative models. From our results, we find that models such as GPT-2 struggle to preserve the complexity of input prompts used in its generations, even if finetuned with professionally written texts.
Readability metrics and standards such as Flesch Kincaid Grade Level (FKGL) and the Common European Framework of Reference for Languages (CEFR) exist to guide teachers and educators to properly assess the complexity of educational materials before administering them for classroom use. In this study, we select a diverse set of open and closed-source instruction-tuned language models and investigate their performances in writing story completions and simplifying narratives—tasks that teachers perform—using standard-guided prompts controlling text readability. Our extensive findings provide empirical proof of how globally recognized models like ChatGPT may be considered less effective and may require more refined prompts for these generative tasks compared to other open-sourced models such as BLOOMZ and FlanT5—which have shown promising results.
Article prediction is a task that has long defied accurate linguistic description. As such, this task is ideally suited to evaluate models on their ability to emulate native-speaker intuition. To this end, we compare the performance of native English speakers and pre-trained models on the task of article prediction set up as a three way choice (a/an, the, zero). Our experiments with BERT show that BERT outperforms humans on this task across all articles. In particular, BERT is far superior to humans at detecting the zero article, possibly because we insert them using rules that the deep neural model can easily pick up. More interestingly, we find that BERT tends to agree more with annotators than with the corpus when inter-annotator agreement is high but switches to agreeing more with the corpus as inter-annotator agreement drops. We contend that this alignment with annotators, despite being trained on the corpus, suggests that BERT is not memorising article use, but captures a high level generalisation of article use akin to human intuition.
This paper presents the shared task on Multilingual Idiomaticity Detection and Sentence Embedding, which consists of two subtasks: (a) a binary classification task aimed at identifying whether a sentence contains an idiomatic expression, and (b) a task based on semantic text similarity which requires the model to adequately represent potentially idiomatic expressions in context. Each subtask includes different settings regarding the amount of training data. Besides the task description, this paper introduces the datasets in English, Portuguese, and Galician and their annotation procedure, the evaluation metrics, and a summary of the participant systems and their results. The task had close to 100 registered participants organised into twenty five teams making over 650 and 150 submissions in the practice and evaluation phases respectively.
We compare sequential fine-tuning with a model for multi-task learning in the context where we are interested in boosting performance on two of the tasks, one of which depends on the other. We test these models on the FigLang2022 shared task which requires participants to predict language inference labels on figurative language along with corresponding textual explanations of the inference predictions. Our results show that while sequential multi-task learning can be tuned to be good at the first of two target tasks, it performs less well on the second and additionally struggles with overfitting. Our findings show that simple sequential fine-tuning of text-to-text models is an extraordinarily powerful method of achieving cross-task knowledge transfer while simultaneously predicting multiple interdependent targets. So much so, that our best model achieved the (tied) highest score on the task.
Deep neural models, in particular Transformer-based pre-trained language models, require a significant amount of data to train. This need for data tends to lead to problems when dealing with idiomatic multiword expressions (MWEs), which are inherently less frequent in natural text. As such, this work explores sample efficient methods of idiomaticity detection. In particular we study the impact of Pattern Exploit Training (PET), a few-shot method of classification, and BERTRAM, an efficient method of creating contextual embeddings, on the task of idiomaticity detection. In addition, to further explore generalisability, we focus on the identification of MWEs not present in the training data. Our experiments show that while these methods improve performance on English, they are much less effective on Portuguese and Galician, leading to an overall performance about on par with vanilla mBERT. Regardless, we believe sample efficient methods for both identifying and representing potentially idiomatic MWEs are very encouraging and hold significant potential for future exploration.
Toxicity is pervasive in social media and poses a major threat to the health of online communities. The recent introduction of pre-trained language models, which have achieved state-of-the-art results in many NLP tasks, has transformed the way in which we approach natural language processing. However, the inherent nature of pre-training means that they are unlikely to capture task-specific statistical information or learn domain-specific knowledge. Additionally, most implementations of these models typically do not employ conditional random fields, a method for simultaneous token classification. We show that these modifications can improve model performance on the Toxic Spans Detection task at SemEval-2021 to achieve a score within 4 percentage points of the top performing team.
This paper describes our submission to SemEval 2021 Task 2. We compare XLM-RoBERTa Base and Large in the few-shot and zero-shot settings and additionally test the effectiveness of using a k-nearest neighbors classifier in the few-shot setting instead of the more traditional multi-layered perceptron. Our experiments on both the multi-lingual and cross-lingual data show that XLM-RoBERTa Large, unlike the Base version, seems to be able to more effectively transfer learning in a few-shot setting and that the k-nearest neighbors classifier is indeed a more powerful classifier than a multi-layered perceptron when used in few-shot learning.
This paper describes the participation of the UoB-NLP team in the ProfNER-ST shared subtask 7a. The task was aimed at detecting the mention of professions in social media text. Our team experimented with two methods of improving the performance of pre-trained models: Specifically, we experimented with data augmentation through translation and the merging of multiple language inputs to meet the objective of the task. While the best performing model on the test data consisted of mBERT fine-tuned on augmented data using back-translation, the improvement is minor possibly because multi-lingual pre-trained models such as mBERT already have access to the kind of information provided through back-translation and bilingual data.
The CogNLP-Sheffield submissions to the CMCL 2021 Shared Task examine the value of a variety of cognitively and linguistically inspired features for predicting eye tracking patterns, as both standalone model inputs and as supplements to contextual word embeddings (XLNet). Surprisingly, the smaller pre-trained model (XLNet-base) outperforms the larger (XLNet-large), and despite evidence that multi-word expressions (MWEs) provide cognitive processing advantages, MWE features provide little benefit to either model.
This paper measures the impact of increased exposure on whether learned construction grammars converge onto shared representations when trained on data from different registers. Register influences the frequency of constructions, with some structures common in formal but not informal usage. We expect that a grammar induction algorithm exposed to different registers will acquire different constructions. To what degree does increased exposure lead to the convergence of register-specific grammars? The experiments in this paper simulate language learning in 12 languages (half Germanic and half Romance) with corpora representing three registers (Twitter, Wikipedia, Web). These simulations are repeated with increasing amounts of exposure, from 100k to 2 million words, to measure the impact of exposure on the convergence of grammars. The results show that increased exposure does lead to converging grammars across all languages. In addition, a shared core of register-universal constructions remains constant across increasing amounts of exposure.
Despite their success in a variety of NLP tasks, pre-trained language models, due to their heavy reliance on compositionality, fail in effectively capturing the meanings of multiword expressions (MWEs), especially idioms. Therefore, datasets and methods to improve the representation of MWEs are urgently needed. Existing datasets are limited to providing the degree of idiomaticity of expressions along with the literal and, where applicable, (a single) non-literal interpretation of MWEs. This work presents a novel dataset of naturally occurring sentences containing MWEs manually classified into a fine-grained set of meanings, spanning both English and Portuguese. We use this dataset in two tasks designed to test i) a language model’s ability to detect idiom usage, and ii) the effectiveness of a language model in generating representations of sentences containing idioms. Our experiments demonstrate that, on the task of detecting idiomatic usage, these models perform reasonably well in the one-shot and few-shot scenarios, but that there is significant scope for improvement in the zero-shot scenario. On the task of representing idiomaticity, we find that pre-training is not always effective, while fine-tuning could provide a sample efficient method of learning representations of sentences containing MWEs.
While lexico-semantic elements no doubt capture a large amount of linguistic information, it has been argued that they do not capture all information contained in text. This assumption is central to constructionist approaches to language which argue that language consists of constructions, learned pairings of a form and a function or meaning that are either frequent or have a meaning that cannot be predicted from its component parts. BERT’s training objectives give it access to a tremendous amount of lexico-semantic information, and while BERTology has shown that BERT captures certain important linguistic dimensions, there have been no studies exploring the extent to which BERT might have access to constructional information. In this work we design several probes and conduct extensive experiments to answer this question. Our results allow us to conclude that BERT does indeed have access to a significant amount of information, much of which linguists typically call constructional information. The impact of this observation is potentially far-reaching as it provides insights into what deep learning methods learn from text, while also showing that information contained in constructions is redundantly encoded in lexico-semantics.
This paper presents our submission to Task 2 of the Workshop on Noisy User-generated Text. We explore improving the performance of a pre-trained transformer-based language model fine-tuned for text classification through an ensemble implementation that makes use of corpus level information and a handcrafted feature. We test the effectiveness of including the aforementioned features in accommodating the challenges of a noisy data set centred on a specific subject outside the remit of the pre-training data. We show that inclusion of additional features can improve classification results and achieve a score within 2 points of the top performing team.
Much as the social landscape in which languages are spoken shifts, language too evolves to suit the needs of its users. Lexical semantic change analysis is a burgeoning field of semantic analysis which aims to trace changes in the meanings of words over time. This paper presents an approach to lexical semantic change detection based on Bayesian word sense induction suitable for novel word sense identification. This approach is used for a submission to SemEval-2020 Task 1, which shows the approach to be capable of the SemEval task. The same approach is also applied to a corpus gleaned from 15 years of Twitter data, the results of which are then used to identify words which may be instances of slang.
The growing popularity and applications of sentiment analysis of social media posts has naturally led to sentiment analysis of posts written in multiple languages, a practice known as code-switching. While recent research into code-switched posts has focused on the use of multilingual word embeddings, these embeddings were not trained on code-switched data. In this work, we present word-embeddings trained on code-switched tweets, specifically those that make use of Spanish and English, known as Spanglish. We explore the embedding space to discover how they capture the meanings of words in both languages. We test the effectiveness of these embeddings by participating in SemEval 2020 Task 9: Sentiment Analysis on Code-Mixed Social Media Text. We utilised them to train a sentiment classifier that achieves an F-1 score of 0.722. This is higher than the baseline for the competition of 0.656, with our team (codalab username francesita) ranking 14 out of 29 participating teams, beating the baseline.
Pre-trained language model word representation, such as BERT, have been extremely successful in several Natural Language Processing tasks significantly improving on the state-of-the-art. This can largely be attributed to their ability to better capture semantic information contained within a sentence. Several tasks, however, can benefit from information available at a corpus level, such as Term Frequency-Inverse Document Frequency (TF-IDF). In this work we test the effectiveness of integrating this information with BERT on the task of identifying abuse on social media and show that integrating this information with BERT does indeed significantly improve performance. We participate in Sub-Task A (abuse detection) wherein we achieve a score within two points of the top performing team and in Sub-Task B (target detection) wherein we are ranked 4 of the 44 participating teams.
The explosive growth and popularity of Social Media has revolutionised the way we communicate and collaborate. Unfortunately, this same ease of accessing and sharing information has led to an explosion of misinformation and propaganda. Given that stance detection can significantly aid in veracity prediction, this work focuses on boosting automated stance detection, a task on which pre-trained models have been extremely successful on, as on several other tasks. This work shows that the task of stance detection can benefit from feature based information, especially on certain under performing classes, however, integrating such features into pre-trained models using ensembling is challenging. We propose a novel architecture for integrating features with pre-trained models that address these challenges and test our method on the RumourEval 2019 dataset. This method achieves state-of-the-art results with an F1-score of 63.94 on the test set.
Prior work has demonstrated that question classification (QC), recognizing the problem domain of a question, can help answer it more accurately. However, developing strong QC algorithms has been hindered by the limited size and complexity of annotated data available. To address this, we present the largest challenge dataset for QC, containing 7,787 science exam questions paired with detailed classification labels from a fine-grained hierarchical taxonomy of 406 problem domains. We then show that a BERT-based model trained on this dataset achieves a large (+0.12 MAP) gain compared with previous methods, while also achieving state-of-the-art performance on benchmark open-domain and biomedical QC datasets. Finally, we show that using this model’s predictions of question topic significantly improves the accuracy of a question answering system by +1.7% P@1, with substantial future gains possible as QC performance improves.
The idea that a shift in concreteness within a sentence indicates the presence of a metaphor has been around for a while. However, recent methods of detecting metaphor that have relied on deep neural models have ignored concreteness and related psycholinguistic information. We hypothesis that this information is not available to these models and that their addition will boost the performance of these models in detecting metaphor. We test this hypothesis on the Metaphor Detection Shared Task 2020 and find that the addition of concreteness information does in fact boost deep neural models. We also run tests on data from a previous shared task and show similar results.
The automatic identification of propaganda has gained significance in recent years due to technological and social changes in the way news is generated and consumed. That this task can be addressed effectively using BERT, a powerful new architecture which can be fine-tuned for text classification tasks, is not surprising. However, propaganda detection, like other tasks that deal with news documents and other forms of decontextualized social communication (e.g. sentiment analysis), inherently deals with data whose categories are simultaneously imbalanced and dissimilar. We show that BERT, while capable of handling imbalanced classes with no additional data augmentation, does not generalise well when the training and test data are sufficiently dissimilar (as is often the case with news sources, whose topics evolve over time). We show how to address this problem by providing a statistical measure of similarity between datasets and a method of incorporating cost-weighting into BERT when the training and test sets are dissimilar. We test these methods on the Propaganda Techniques Corpus (PTC) and achieve the second highest score on sentence-level propaganda classification.
We present a system for Answer Selection that integrates fine-grained Question Classification with a Deep Learning model designed for Answer Selection. We detail the necessary changes to the Question Classification taxonomy and system, the creation of a new Entity Identification system and methods of highlighting entities to achieve this objective. Our experiments show that Question Classes are a strong signal to Deep Learning models for Answer Selection, and enable us to outperform the current state of the art in all variations of our experiments except one. In the best configuration, our MRR and MAP scores outperform the current state of the art by between 3 and 5 points on both versions of the TREC Answer Selection test set, a standard dataset for this task.
We present in this paper a purely rule-based system for Question Classification which we divide into two parts: The first is the extraction of relevant words from a question by use of its structure, and the second is the classification of questions based on rules that associate these words to Concepts. We achieve an accuracy of 97.2%, close to a 6 point improvement over the previous State of the Art of 91.6%. Additionally, we believe that machine learning algorithms can be applied on top of this method to further improve accuracy.