AACL 2023 Papers

Investigating the behaviour of Machine Reading Comprehension (MRC) models under various types of test-time perturbations can shed light on the enhancement of their robustness and generalisation capability, despite the su-perhuman performance they have achieved on existing benchmark datasets. In this paper, we study the robustness of contemporary MRC systems to context paraphrasing, i.e., whether these models are still able to correctly answer the questions once the reading passages have been paraphrased. To this end, we systematically design a pipeline to semi-automatically generate perturbed MRC instances which ultimately lead to the creation of a paraphrased test set. We conduct experiments on this dataset with six state-of-the-art neural MRC models and we find that even the minimum performance drop of all these models exceeds 41%, whereas human performance remains high. Re-training models with augmented perturbed examples results in improved robustness, though the performance remains lower than on the original dataset. These results demonstrate that the existing high-performing MRC systems are still far away from real language understanding 1 .

Recent advancements in neural language models have revolutionized natural language understanding. However, many languages still face the risk of being left behind without the benefits of such advancements, potentially leading to their extinction. One such language is Azerbaijani in Iran, which suffers from limited digital resources and a lack of alignment be-tween spoken and written forms. In contrast, Azerbaijani in the Republic of Azerbaijan has seen more resources and is not considered as low-resource as its Iranian counterpart. In this context, our research focuses on the computational progress made in Iranian Azerbaijani language. We propose a transliteration model that leverages an Azerbaijani parallel dataset, effectively bridging the gap between the Latin and Persian scripts. By enabling seamless communication between these two scripts, our model facilitates cultural exchange and serves as a valuable tool for transfer learning. The effectiveness of our approach surpasses traditional rule-based methods, as evidenced by the significant improvements in performance metrics. We observe a minimum 15% increase in BLEU scores and a reduction of at least 1/3 in edit distance. Furthermore, our model’s online demo is accessible at https://azeri.parsi.ai/ .

Iranian Azerbaijani is a dialect of the Azer-baijani language spoken by more than 16% of the population in Iran (>14 million). Unfortunately, a lack of computational resources is one of the factors that puts this language and its rich culture at risk of extinction. This work aims to create fundamental natural language processing (NLP) resources and pipelines for the processing and analysis of Iranian Azerbai-jani introducing standard datasets and starter models for various NLP tasks such as language modeling, text classification, part-of-speech (POS) tagging, and machine translation. The proposed resources have been curated and pre-processed to facilitate the development of NLP models for Iranian Azerbaijani and provide a strong baseline for further research and development. This study is an example of bridging the gap in NLP for low-resource languages and promoting the advancement of language technologies in underrepresented languages. To the best of our knowledge, for the first time, this paper presents major infrastructures for the processing and analysis of Iranian Azer-baijani, with the ultimate goal of improving communication and information access for millions of individuals. Furthermore, our translation model’s online demo is accessible at https://azeri.parsi.ai/ .

Open-domain table question answering aims to provide answers to a question by retrieving and extracting information from a large collection of tables. Existing studies of open-domain table QA either directly adopt text retrieval methods or consider the table structure only in the encoding layer for table retrieval, which may cause syntactical and structural information loss during table scoring. To address this issue, we propose a syntax- and structure-aware retrieval method for the open-domain table QA task. It provides syntactical representations for the question and uses the structural header and value representations for the tables to avoid the loss of fine-grained syntactical and structural information. Then, a syntactical-to-structural aggregator is used to obtain the matching score between the question and a candidate table by mimicking the human retrieval process. Experimental results show that our method achieves the state-of-the-art on the NQ-tables dataset and overwhelms strong baselines on a newly curated open-domain Text-to-SQL dataset.

Subword segmentation methods are the predominant solution to vocab sparsity in NMT. However, they cannot currently handle capitalization well. We re-encode case to allow the perplexity-driven SPM unigram language model algorithm to learn how to segment capitalization. Since naturally occurring data accurately describes the prevalence of capi-talization but underestimates the importance humans ascribe to capitalization robustness, we propose data augmentation to fill this gap. We demonstrate that our proposed method improves translation quality on ALL CAPS, lower cased, and Title Case, while maintaining quality on standard test sets. In contrast to prior work, our proposed method has minimal impact on decoding speed. We release our code: github.com/marian-nmt/sentencepiece .

Event extraction from multimodal documents is an important yet under-explored problem. One challenge faced by this task is the scarcity of paired image-text datasets, making it diffi-cult to fully exploit the strong representation power of multimodal language models. In this paper, we present Theia, an end-to-end multimodal event extraction framework that can be trained on incomplete data. Specifically, we couple a generation-based event extraction model with a customised image synthesizer that can generate images from text. Our model leverages capabilities of pre-trained vision-language models and can be trained on incomplete (i.e. text-only) data. Experimental results on existing multimodal datasets demonstrate the effectiveness of our approach for both syn-thesising missing data and extracting events over state-of-the-art approaches.

In intent detection tasks, leveraging meaningful semantic information from intent labels can be particularly beneficial for few-shot scenarios. However, existing few-shot intent detection methods either ignore the intent labels, (e.g. treating intents as indices) or do not fully utilize this information (e.g. only using part of the intent labels). In this work, we present an end-to-end One-to-All system that enables the comparison of an input utterance with all label candidates. The system can then fully utilize label semantics in this way. Experiments on three few-shot intent detection tasks demonstrate that One-to-All is especially effective when the training resource is extremely scarce, achieving state-of-the-art performance in 1-, 3- and 5-shot settings. Moreover, we present a novel pretraining strategy for our model that utilizes indirect supervision from paraphrasing, enabling zero-shot cross-domain generalization on intent detection tasks. Our code is at https://github.com/jiangshdd/AllLablesTogether.

Previous attempts to incorporate a mention detection step into end-to-end neural coreference resolution for English have been hampered by the lack of singleton mention span data as well as other entity information. This paper presents a coreference model that learns singletons as well as features such as entity type and information status via a multi-task learning-based approach. This approach achieves new state-of-the-art scores on the OntoGUM benchmark (+2.7 points) and increases robustness on multiple out-of-domain datasets (+2.3 points on average), likely due to greater generalizability for mention detection and utilization of more data from singletons when compared to only coreferent mention pair matching.

Both standalone language models (LMs) as well as LMs within downstream-task systems have been shown to generate statements which are factually untrue. This problem is especially severe for low-resource languages, where training data is scarce and of worse quality than for high-resource languages. In this opinion piece, we argue that LMs in their current state will never be fully trustworthy in critical settings and suggest a possible novel strategy to handle this issue: by building LMs such that can cite their sources - i.e., point a user to the parts of their training data that back up their outputs. We first discuss which current NLP tasks would or would not benefit from such models. We then highlight the expected benefits such models would bring, e.g., quick verifiability of statements. We end by outlining the individual tasks that would need to be solved on the way to developing LMs with the ability to cite. We hope to start a discussion about the field's current approach to building LMs, especially for low-resource languages, and the role of the training data in explaining model generations.

For sequence-to-sequence tasks it is challenging to combine individual system outputs. Further, there is also often a mismatch between the decoding criterion and the one used for assessment. Minimum Bayes' Risk (MBR) decoding can be used to combine system outputs in a manner that encourages better alignment with the final assessment criterion. This paper examines MBR decoding for Grammatical Error Correction (GEC) systems, where performance is usually evaluated in terms of edits and an associated F-score. Hence, we propose a novel MBR loss function directly linked to this form of criterion. Furthermore, an approach to expand the possible set of candidate sentences is described. This builds on a current max-voting combination scheme, as well as individual edit-level selection. Experiments on three popular GEC datasets and with state-of-the-art GEC systems demonstrate the efficacy of the proposed MBR approach. Additionally, the paper highlights how varying reward metrics within the MBR decoding framework can provide control over precision, recall, and the F-score in combined GEC systems.

Resources for building NLP applications, such as data and models, are usually only created and curated for a limited set of high resource languages. Thus, the ability to transfer knowledge to a new language is a key way in which to enable access to NLP technology for a wider population. This paper presents a framework to perform zero-shot inference in a target language by using cross-lingual retrieval from another language where limited annotated data for a comparable domain is available. Results on two large-scale multilingual datasets show that, in this setup, this framework improves over fine-tuning multilingual models or translating annotated data, and achieves results relatively close to fine-tuning the model on the target language directly. These results show that models can be transferred efficiently across languages for a given task and domain, even for languages not covered by multilingual model training approaches.

Large language models (LLMs) have shown remarkable performance on a variety of multilingual NLP tasks. Code-switching is one of the most convenient styles of communication in multilingual communities. It is known to present several challenges to the existing language models and task-specific models. In this paper, we evaluate the capability of multi-lingual LLMs for the code-switched machine translation (CSMT) task in traditional and novel settings and present our insights. We observe that ChatGPT outperforms other LLMs and shows competitive performance to the supervised fine-tuned models. Though promising, ChatGPT shows major limitations, such as high gender bias, stereotypes, and factual inconsistencies. It further demands a multi-dimensional large-scale evaluation of the multilingual LLMs for code-switched languages.

We report a problem that one faces when using ChatGPT in similar languages, taking Malay and Indonesian as examples: ChatGPT often responds to prompts in Malay (the language with fewer speakers) in Indonesian (the language with more speakers). We examined ChatGPT’s identification (LangID) ability to find out whether this language choice problem arises from LangID errors. The results show that LangID errors alone cannot explain the problem’s severity. By comparing the patterns of responses to Malay prompts and those to Javanese prompts, we conclude that the problem happens mainly because ChatGPT does not treat Malay and Indonesian equally as distinct languages. Rather, it behaves as if Malay were a non-standard variety of Indonesian. We also discuss social issues the language choice problem causes and possible solutions to them.

Prompt tuning (PT) is an effective approach to adapting pre-trained language models to downstream tasks. Without a good initialization, prompt tuning doesn't perform well under few-shot settings. So pre-trained prompt tuning (PPT) is proposed to initialize prompts by leveraging pre-training data. We propose MetaPT (Meta-learned Prompt Tuning) to further improve PPT's initialization by considering latent structure within the pre-training data. Specifically, we introduce the structure by first clustering pre-training data into different auxiliary tasks with unsupervised methods. Then we use these tasks to pre-train prompts with a meta-learning algorithm. Such a process can make prompts learn a better initialization by discovering commonalities among these auxiliary tasks. We evaluate our method on seven downstream tasks. Our MetaPT achieves better and more stable performance than the state-of-the-art method.

Since the popularization of BiLSTMs and Transformer-based bidirectional encoders, state-of-the-art syntactic parsers have lacked incrementality, requiring access to the whole sentence and deviating from human language processing. This paper explores whether fully incremental dependency parsing with modern architectures can be competitive. We build parsers combining strictly left-to-right neural encoders with fully incremental sequence-labeling and transition-based decoders. The results show that fully incremental parsing with modern architectures considerably lags behind bidirectional parsing, noting the challenges of psycholinguistically plausible parsing.

Two distinct approaches have been proposed for relational triple extraction - pipeline and joint. Joint models, which capture interactions across triples, are the more recent development, and have been shown to outperform pipeline models for sentence-level extraction tasks. Document-level extraction is a more challenging setting where interactions across triples can be long-range, and individual triples can also span across sentences. Joint models have not been applied for document-level tasks so far. In this paper, we benchmark state-of-the-art pipeline and joint extraction models on sentence-level as well as document-level datasets. Our experiments show that while joint models outperform pipeline models significantly for sentence-level extraction, their performance drops sharply below that of pipeline models for the document-level dataset.

Volatility, a crucial statistical measure in the financial market, serves as an indicator of financial instrument risk. Accurate volatility capture aids in predicting stock movements and is valuable in derivative trading, such as options trading. While recent research focuses on volatility forecasting using earnings call transcriptions, most approaches rely on end-to-end models that directly process textual or vocal data. However, limited efforts have been made to simulate the reading and comprehension processes of financial professionals, thereby enhancing the capabilities of language models. To address this gap, we propose a general numeral attachment dataset designed to train language models to understand earnings calls with the expertise of professionals. Additionally, we introduce a pre-training process that improves the semantic understanding of earnings calls. Experimental results demonstrate that our pre-trained language model enhances the accuracy of 3-day volatility forecasting.

Pre-trained language models trained on large-scale data have learned serious levels of social biases. Consequently, various methods have been proposed to debias pre-trained models. Debiasing methods need to mitigate only discriminatory bias information from the pre-trained models, while retaining information that is useful for the downstream tasks. In previous research, whether useful information is retained has been confirmed by the performance of downstream tasks in debiased pre-trained models. On the other hand, it is not clear whether these benchmarks consist of data pertaining to social biases and are appropriate for investigating the impact of debiasing. For example in gender-related social biases, data containing female words (e.g. ``she, female, woman''), male words (e.g. ``he, male, man''), and stereotypical words (e.g. ``nurse, doctor, professor'') are considered to be the most affected by debiasing. If there is not much data containing these words in a benchmark dataset for a target task, there is the possibility of erroneously evaluating the effects of debiasing. In this study, we compare the impact of debiasing on performance across multiple downstream tasks using a wide-range of benchmark datasets that containing female, male, and stereotypical words. Experiments show that the effects of debiasing are consistently \emph{underestimated} across all tasks. Moreover, the effects of debiasing could be reliably evaluated by separately considering instances containing female, male, and stereotypical words than all of the instances in a benchmark dataset.

Evaluation of QA systems is very challenging and expensive, with the most reliable approach being human annotations of correctness of answers for questions. Recent works (AVA, BEM) have shown that transformer LM encoder based similarity metrics transfer well for QA evaluation, but they are limited by the usage of a single correct reference answer. We propose a new evaluation metric: SQuArE (Sentence-level QUestion AnsweRing Evaluation), using multiple reference answers (combining multiple correct and incorrect references) for sentence-form QA. We evaluate SQuArE on both sentence-level extractive (Answer Selection) and generative (GenQA) QA systems, across multiple academic and industrial datasets, and show that it outperforms previous baselines and obtains the highest correlation with human annotations.