AACL 2022 Papers

In this paper, we describe KreolMorisienMT, a dataset for benchmarking machine translation quality of Mauritian Creole. Mauritian Creole (Kreol Morisien) is a French-based creole and a lingua franca of the Republic of Mauritius. KreolMorisienMT consists of a parallel corpus between English and Kreol Morisien, French and Kreol Morisien and a monolingual corpus for Kreol Morisien. We first give an overview of Kreol Morisien and then describe the steps taken to create the corpora. Thereafter, we benchmark Kreol Morisien ↔ English and Kreol Morisien ↔ French models leveraging pre-trained models and multilingual transfer learning. Human evaluation reveals our systems’ high translation quality.

Virtual adversarial training (VAT) is a powerful approach to improving robustness and performance, leveraging both labeled and unlabeled data to compensate for the scarcity of labeled data. It is adopted on lots of vision and language classification tasks. However, for tasks with structured output (e.g., dependency parsing), the application of VAT is nontrivial due to the intrinsic proprieties of structures: (1) the non-sparse problem and (2) exponential complexity. Against this background, we propose the Sparse Parse Adjustment (spa) algorithm and successfully applied VAT to the dependency parsing task. spa refers to the learning algorithm which combines the graph-based dependency parsing model with VAT in an exact computational manner and enhances the dependency parser with controllable and adjustable sparsity. Empirical studies show that the TreeCRF parser optimized using outperforms other methods without sparsity regularization.

Recent work has shown advantages of incorporating knowledge graphs (KGs) into BERT for various NLP tasks. One common way is to feed entity embeddings as an additional input during pre-training. There are two limitations to such a method. First, to train the entity embeddings to include rich information of factual knowledge, it typically requires access to the entire KG. This is challenging for KGs with daily changes (e.g., Wikidata). Second, it requires a large scale pre-training corpus with entity annotations and high computational cost during pre-training. In this work, we efficiently construct entity embeddings only from the type knowledge, that does not require access to the entire KG. Although the entity embeddings contain only local information, they perform very well when combined with context. Furthermore, we show that our entity embeddings, constructed from BERT’s input embeddings, can be directly incorporated into the fine-tuning phase without requiring any specialized pre-training. In addition, these entity embeddings can also be constructed on the fly without requiring a large memory footprint to store them. Finally, we propose task-specific models that incorporate our entity embeddings for entity linking, entity typing, and relation classification. Experiments show that our models have comparable or superior performance to existing models while being more resource efficient.

Being able to train Named Entity Recognition (NER) models for emerging topics is crucial for many real-world applications especially in the medical domain where new topics are continuously evolving out of the scope of existing models and datasets. For a realistic evaluation setup, we introduce a novel COVID-19 news NER dataset (COVIDNEWS-NER) and release 3000 entries of hand annotated strongly labelled sentences and 13000 auto-generated weakly labelled sentences. Besides the dataset, we propose CONTROSTER, a recipe to strategically combine weak and strong labels in improving NER in an emerging topic through transfer learning. We show the effectiveness of CONTROSTER on COVIDNEWS-NER while providing analysis on combining weak and strong labels for training. Our key findings are: (1) Using weak data to formulate an initial backbone before tuning on strong data outperforms methods trained on only strong or weak data. (2) A combination of out-of-domain and in-domain weak label training is crucial and can overcome saturation when being training on weak labels from a single source.

Speech translation for subtitling (SubST) is the task of automatically translating speech data into well-formed subtitles by inserting subtitle breaks compliant to specific displaying guidelines. Similar to speech translation (ST), model training requires parallel data comprising audio inputs paired with their textual translations. In SubST, however, the text has to be also annotated with subtitle breaks. So far, this requirement has represented a bottleneck for system development, as confirmed by the dearth of publicly available SubST corpora. To fill this gap, we propose a method to convert existing ST corpora into SubST resources without human intervention. We build a segmenter model that automatically segments texts into proper subtitles by exploiting audio and text in a multimodal fashion, achieving high segmentation quality in zero-shot conditions. Comparative experiments with SubST systems respectively trained on manual and automatic segmentations result in similar performance, showing the effectiveness of our approach.

Several multi-hop reading comprehension datasets have been proposed to resolve the issue of reasoning shortcuts by which questions can be answered without performing multi-hop reasoning. However, the ability of multi-hop models to perform step-by-step reasoning when finding an answer to a comparison question remains unclear. It is also unclear how questions about the internal reasoning process are useful for training and evaluating question-answering (QA) systems. To evaluate the model precisely in a hierarchical manner, we first propose a dataset, HieraDate, with three probing tasks in addition to the main question: extraction, reasoning, and robustness. Our dataset is created by enhancing two previous multi-hop datasets, HotpotQA and 2WikiMultiHopQA, focusing on multi-hop questions on date information that involve both comparison and numerical reasoning. We then evaluate the ability of existing models to understand date information. Our experimental results reveal that the multi-hop models do not have the ability to subtract two dates even when they perform well in date comparison and number subtraction tasks. Other results reveal that our probing questions can help to improve the performance of the models (e.g., by +10.3 F1) on the main QA task and our dataset can be used for data augmentation to improve the robustness of the models.

Despite the promising evaluation results by knowledge distillation (KD) in natural language understanding (NLU) and sequence-to-sequence (seq2seq) tasks, KD for causal language modeling (LM) remains a challenge. In this paper, we present a novel perspective of knowledge distillation by proposing plug and play knowledge distillation (PP-KD) to improve a (student) kNN-LM that is the state-of-the-art in causal language modeling by leveraging external logits from either a powerful or a heterogeneous (teacher) LM. Unlike conventional logit-based KD where the teacher’s knowledge is built-in during training, PP-KD is plug and play: it stores the teacher’s knowledge (i.e., logits) externally and uses the teacher’s logits of the retrieved k-nearest neighbors during kNN-LM inference at test time. In contrast to marginal perplexity improvement by logit-based KD in conventional neural (causal) LM, PP-KD achieves a significant improvement, enhancing the kNN-LMs in multiple language modeling datasets, showing a novel and promising perspective for causal LM distillation.

A bottleneck to developing Semantic Parsing (SP) models is the need for a large volume of human-labeled training data. Given the complexity and cost of human annotation for SP, labeled data is often scarce, particularly in multilingual settings. Large Language Models (LLMs) excel at SP given only a few examples, however LLMs are unsuitable for runtime systems which require low latency. In this work, we propose CLASP, a simple method to improve low-resource SP for moderate-sized models: we generate synthetic data from AlexaTM 20B to augment the training set for a model 40x smaller (500M parameters). We evaluate on two datasets in low-resource settings: English PIZZA, containing either 348 or 16 real examples, and mTOP cross-lingual zero-shot, where training data is available only in English, and the model must generalize to four new languages. On both datasets, we show significant improvements over strong baseline methods.

The Entity and Relation Extraction (ERE) task includes two basic sub-tasks: Named Entity Recognition and Relation Extraction. In the last several years, much work focused on joint approaches for the common perception that the pipeline approach suffers from the error propagation problem. Recent work reconsiders the pipeline scheme and shows that it can produce comparable results. To systematically study the pros and cons of these two schemes. We design and test eight pipeline and joint approaches to the ERE task. We find that with the same span representation methods, the best joint approach still outperforms the best pipeline model, but improperly designed joint approaches may have poor performance. We hope our work could shed some light on the pipeline-vs-joint debate of the ERE task and inspire further research.

Automatic readability assessment (ARA) aims at classifying the readability level of a passage automatically. In the past, manually selected linguistic features are used to classify the passages. However, as the use of deep neural network surges, there is less work focusing on these linguistic features. Recently, many works integrate linguistic features with pre-trained language model (PLM) to make up for the information that PLMs are not good at capturing. Despite their initial success, insufficient analysis of the long passage characteristic of ARA has been done before. To further investigate the promotion of linguistic features on PLMs in ARA from the perspective of passage length, with commonly used linguistic features and abundant experiments, we find that: (1) Linguistic features promote PLMs in ARA mainly on long passages. (2) The promotion of the features on PLMs becomes less significant when the dataset size exceeds 750 passages. (3) By analyzing commonly used ARA datasets, we find Newsela is actually not suitable for ARA. Our code is available at https://github.com/recorderhou/linguistic-features-in-ARA.

We investigate in this paper how correlations between occupations and gendered-pronouns can be affected and changed by adding negation in bias probes, or changing the grammatical tense of the verbs in the probes. We use a set of simple bias probes in Norwegian and English, and perform 16 different probing analysis, using four Norwegian and four English pre-trained language models. We show that adding negation to probes does not have a considerable effect on the correlations between gendered-pronouns and occupations, supporting other works on negation in language models. We also show that altering the grammatical tense of verbs in bias probes do have some interesting effects on models’ behaviours and correlations. We argue that we should take grammatical tense into account when choosing bias probes, and aggregating results across tenses might be a better representation of the existing correlations.

Stance Recognition over proposals is the task of automatically detecting whether a comment on a specific proposal is in favor of this proposal, against this proposal or that neither inference is likely. The dataset that we propose to use is an online debating platform inaugurated in 2021, where users can submit proposals and comment over proposals or over other comments. It contains 4.2k proposals and 20k comments focused on various topics. Every comment and proposal can come written in another language, with more than 40% of the proposal/comment pairs containing at least two languages, creating a unique intra-multilingual setting. A portion of the data (more than 7k comment/proposal pairs, in 26 languages) was annotated by the writers with a self-tag assessing whether they are in favor or against the proposal. Another part of the data (without self-tag) has been manually annotated: 1206 comments in 6 morphologically different languages (fr, de, en, el, it, hu) were tagged, leading to a Krippendorff’s α of 0.69. This setting allows defining an intra-multilingual and multi-target stance classification task over online debates.

Spell-checkers are core applications in language learning and normalisation, which may enormously contribute to language revitalisation and language teaching in the context of indigenous communities. Spell-checking as a generation task, however, requires large amount of data, which is not feasible for endangered languages, such as the languages spoken in Peruvian Amazonia. We propose here augmentation methods for various misspelling types as a strategy to train neural spell-checking models and we create an evaluation resource for four indigenous languages of Peru: Shipibo-Konibo, Asháninka, Yánesha, Yine. We focus on special errors that are significant for learning these languages, such as phoneme-to-grapheme ambiguity, grammatical errors (gender, tense, number, among others), accentuation, punctuation and normalisation in contexts where two or more writing traditions co-exist. We found that an ensemble model, trained with augmented data from various types of error achieves overall better scores in most of the error types and languages. Finally, we released our spell-checkers as a web service to be used by indigenous communities and organisations to develop future language materials.

Topic-sensitive query set expansion is an important area of research that aims to improve search results for information retrieval. It is particularly crucial for queries related to sensitive and emerging topics. In this work, we describe a method for query set expansion about emerging topics using vector space interpolation. We use a transformer model called OPTIMUS, which is suitable for vector space manipulation due to its variational autoencoder nature. One of our proposed methods – Dirichlet interpolation shows promising results for query expansion. Our methods effectively generate new queries about the sensitive topic by incorporating set-level diversity, which is not captured by traditional sentence-level augmentation methods such as paraphrasing or back-translation.

We propose a method for controlling the difficulty of a sentence based on deep reinforcement learning. Although existing models are trained based on the word-level difficulty, the sentence-level difficulty has not been taken into account in the loss function. Our proposed method generates sentences of appropriate difficulty for the target audience through reinforcement learning using a reward calculated based on the difference between the difficulty of the output sentence and the target difficulty. Experimental results of English text simplification show that the proposed method achieves a higher performance than existing approaches. Compared to previous studies, the proposed method can generate sentences whose grade-levels are closer to those of human references estimated using a fine-tuned pre-trained model.

In light of the prominence of Pre-trained Language Models (PLMs) across numerous downstream tasks, shedding light on what they learn is an important endeavor. Whereas previous work focuses on assessing in-domain knowledge, we evaluate the generalization ability in biased scenarios through component combinations where it could be easy for the PLMs to learn shortcuts from the training corpus. This would lead to poor performance on the testing corpus, which is combinationally reconstructed from the training components. The results show that PLMs are able to overcome such distribution shifts for specific tasks and with sufficient data. We further find that overfitting can lead the models to depend more on biases for prediction, thus hurting the combinational generalization ability of PLMs.

A lack of large-scale human-annotated data has hampered the hierarchical discourse parsing of Chinese. In this paper, we present GCDT, the largest hierarchical discourse treebank for Mandarin Chinese in the framework of Rhetorical Structure Theory (RST). GCDT covers over 60K tokens across five genres of freely available text, using the same relation inventory as contemporary RST treebanks for English. We also report on this dataset’s parsing experiments, including state-of-the-art (SOTA) scores for Chinese RST parsing and RST parsing on the English GUM dataset, using cross-lingual training in Chinese and English with multilingual embeddings.

We propose a benchmark to assess the capability of large language models to reason with conventional metaphors. Our benchmark combines the previously isolated topics of metaphor detection and commonsense reasoning into a single task that requires a model to make inferences by accurately selecting between the literal and metaphorical register. We examine the performance of state-of-the-art pre-trained models on binary-choice tasks and find a large discrepancy between the performance of small and very large models, going from chance to near-human level. We also analyse the largest model in a generative setting and find that although human performance is approached, careful multiple-shot prompting is required.

The Transformer Language Model is a powerful tool that has been shown to excel at various NLP tasks and has become the de-facto standard solution thanks to its versatility. In this study, we employ pre-trained document embeddings in an Active Learning task to group samples with the same labels in the embedding space on a legal document corpus. We find that the calculated class embeddings are not close to the respective samples and consequently do not partition the embedding space in a meaningful way. In addition, we explore using the class embeddings as an Active Learning strategy with dramatically reduced results compared to all baselines.