COLING 2024 论文列表

Previous studies have highlighted the importance of vaccination as an effective strategy to control the transmission of the COVID-19 virus. It is crucial for policymakers to have a comprehensive understanding of the public’s stance towards vaccination on a large scale. However, attitudes towards COVID-19 vaccination, such as pro-vaccine or vaccine hesitancy, have evolved over time on social media. Thus, it is necessary to account for possible temporal shifts when analysing these stances. This study aims to examine the impact of temporal concept drift on stance detection towards COVID-19 vaccination on Twitter. To this end, we evaluate a range of transformer-based models using chronological (splitting the training, validation, and test sets in order of time) and random splits (randomly splitting these three sets) of social media data. Our findings reveal significant discrepancies in model performance between random and chronological splits in several existing COVID-19-related datasets; specifically, chronological splits significantly reduce the accuracy of stance classification. Therefore, real-world stance detection approaches need to be further refined to incorporate temporal factors as a key consideration.

Large language models (LLMs) have demonstrated remarkable capabilities across various NLP tasks. However, their computational costs are prohibitively high. To address this issue, previous research has attempted to distill the knowledge of LLMs into smaller models by generating annotated data. Nonetheless, these works have mainly focused on the direct use of LLMs for text generation and labeling, without fully exploring their potential to comprehend the target task and acquire valuable knowledge. In this paper, we propose EvoKD: Evolving Knowledge Distillation, which leverages the concept of active learning to interactively enhance the process of data generation using large language models, simultaneously improving the task capabilities of small domain model (student model). Different from previous work, we actively analyze the student model’s weaknesses, and then synthesize labeled samples based on the analysis. In addition, we provide iterative feedback to the LLMs regarding the student model’s performance to continuously construct diversified and challenging samples. Experiments and analysis on different NLP tasks, namely, text classification and named entity recognition show the effectiveness of EvoKD.

While Large Language Models (LLMs) excel at the Winograd Schema Challenge (WSC), a coreference resolution task testing common-sense reasoning through pronoun disambiguation, they struggle with instances that feature minor alterations or rewording. To address this, we introduce EvoGrad, an open-source platform that harnesses a human-in-the-loop approach to create a dynamic dataset tailored to such altered WSC instances. Leveraging ChatGPT’s capabilities, we expand our task instances from 182 to 3691, setting a new benchmark for diverse common-sense reasoning datasets. Additionally, we introduce the error depth metric, assessing model stability in dynamic tasks. Our results emphasize the challenge posed by EvoGrad: Even the best performing LLM, GPT-3.5, achieves an accuracy of 65.0% with an average error depth of 7.2, a stark contrast to human performance of 92.8% accuracy without perturbation errors. This highlights ongoing model limitations and the value of dynamic datasets in uncovering them.

Human annotation is costly and impractical when it comes to scarcely labeled data. Besides, the presence of biased language in well-known benchmarks notably misleads predictive models to perform incredibly well, not because of the model capability but due to the hidden false correlations in the linguistic corpus. Motivated by this, we propose a neutralized Knowledge Transfer framework (NKT) to equip pre-trained language models with neutralized transferability. Specifically, we construct debiased multi-source corpora (CV and EL) for two exemplary knowledge transfer tasks: claim verification and evidence learning, respectively. To counteract biased language, we design a neutralization mechanism in the presence of label skewness. We also design a label adaptation mechanism in light of the mixed label systems in the multi-source corpora. In extensive experiments, the proposed NKT framework shows effective transferability contrarily to the disability of dominant baselines, particularly in the zero-shot cross-domain transfer setting.

Developmental stages are a linguistic concept claiming that language learning, despite its large inter-individual variance, generally progresses in an ordered, step-like manner. At the core of research has been the acquisition of verb placement by learners, as conceptualized within Processability Theory (Pienemann, 1989). The computational implementation of a system detecting developmental stages is a prerequisite for an automated analysis of L2 language development. However, such an implementation faces two main challenges. The first is the lack of a fully fleshed out, coherent linguistic specification of the stages. The second concerns the translation of the linguistic specification into computational procedures that can extract clauses from learner-produced text and assign them to a developmental stage based on verb placement. Our contribution provides the necessary linguistic specification of the stages as well as detaiiled discussion and recommendations regarding computational implementation.

Narrative reasoning relies on the understanding of eventualities in story contexts, which requires a wealth of background world knowledge. To help machines leverage such knowledge, existing solutions can be categorized into two groups. Some focus on implicitly modeling eventuality knowledge by pretraining language models (LMs) with eventuality-aware objectives. However, this approach breaks down knowledge structures and lacks interpretability. Others explicitly collect world knowledge of eventualities into structured eventuality-centric knowledge graphs (KGs). However, existing research on leveraging these knowledge sources for free-texts is limited. In this work, we propose an initial comprehensive framework called EventGround, which aims to tackle the problem of grounding free-texts to eventuality-centric KGs for contextualized narrative reasoning. We identify two critical problems in this direction: the event representation and sparsity problems. We provide simple yet effective parsing and partial information extraction methods to tackle these problems. Experimental results demonstrate that our approach consistently outperforms baseline models when combined with graph neural network (GNN) or large language model (LLM) based graph reasoning models. Our framework, incorporating grounded knowledge, achieves state-of-the-art performance while providing interpretable evidence.

The embedding-based retrieval (EBR) approach is widely used in mainstream search engine retrieval systems and is crucial in recent retrieval-augmented methods for eliminating LLM illusions. However, existing EBR models often face the “semantic drift” problem and insufficient focus on key information, leading to a low adoption rate of retrieval results in subsequent steps. This issue is especially noticeable in real-time search scenarios, where the various expressions of popular events on the Internet make real-time retrieval heavily reliant on crucial event information. To tackle this problem, this paper proposes a novel approach called EER, which enhances real-time retrieval performance by improving the dual-encoder model of traditional EBR. We incorporate contrastive learning to accompany pairwise learning for encoder optimization. Furthermore, to strengthen the focus on critical event information in events, we include a decoder module after the document encoder, introduce a generative event triplet extraction scheme based on prompt-tuning, and correlate the events with query encoder optimization through comparative learning. This decoder module can be removed during inference. Extensive experiments demonstrate that EER can significantly improve the real-time search retrieval performance. We believe that this approach will provide new perspectives in the field of information retrieval. The codes and dataset are available at https://github.com/open-event-hub/Event-enhanced_Retrieval.

Event representation learning plays a crucial role in numerous natural language processing (NLP) tasks, as it facilitates the extraction of semantic features associated with events. Current methods of learning event representation based on contrastive learning processes positive examples with single-grain random masked language model (MLM), but fall short in learn information inside events from multiple aspects. In this paper, we introduce multi-grained contrastive learning and triple-mixture of experts (MCTM) for event representation learning. Our proposed method extends the random MLM by incorporating a specialized MLM designed to capture different grammatical structures within events, which allows the model to learn token-level knowledge from multiple perspectives. Furthermore, we have observed that mask tokens with different granularities affect the model differently, therefore, we incorporate mixture of experts (MoE) to learn importance weights associated with different granularities. Our experiments demonstrate that MCTM outperforms other baselines in tasks such as hard similarity and transitive sentence similarity, highlighting the superiority of our method.

Cross-lingual transfer-learning is widely used in Event Extraction for low-resource languages and involves a Multilingual Language Model that is trained in a source language and applied to the target language. This paper studies whether the typological similarity between source and target languages impacts the performance of cross-lingual transfer, an under-explored topic. We first focus on Basque as the target language, which is an ideal target language because it is typologically different from surrounding languages. Our experiments on three Event Extraction tasks show that the shared linguistic characteristic between source and target languages does have an impact on transfer quality. Further analysis of 72 language pairs reveals that for tasks that involve token classification such as entity and event trigger identification, common writing script and morphological features produce higher quality cross-lingual transfer. In contrast, for tasks involving structural prediction like argument extraction, common word order is the most relevant feature. In addition, we show that when increasing the training size, not all the languages scale in the same way in the cross-lingual setting. To perform the experiments we introduce EusIE, an event extraction dataset for Basque, which follows the Multilingual Event Extraction dataset (MEE). The dataset and code are publicly available.

Traditional evaluation methods for Grammatical Error Correction (GEC) fail to fully capture the full range of system capabilities and objectives. The emergence of large language models (LLMs) has further highlighted the shortcomings of these evaluation strategies, emphasizing the need for a paradigm shift in evaluation methodology. In the current study, we perform a comprehensive evaluation of various GEC systems using a recently published dataset of Swedish learner texts. The evaluation is performed using established evaluation metrics as well as human judges. We find that GPT-3 in a few-shot setting by far outperforms previous grammatical error correction systems for Swedish, a language comprising only about 0.1% of its training data. We also found that current evaluation methods contain undesirable biases that a human evaluation is able to reveal. We suggest using human post-editing of GEC system outputs to analyze the amount of change required to reach native-level human performance on the task, and provide a dataset annotated with human post-edits and assessments of grammaticality, fluency and meaning preservation of GEC system outputs.

Lexical translation consistency is one of the most common discourse phenomena in Chinese-to-English document-level translation. To better evaluate the performance of lexical translation consistency, previous researches assumes that all repeated source words should be translated consistently. However, constraining translations of repeated source words to be consistent will hurt word diversity and human translators tend to use different words in translation. Therefore, in this paper we construct a test set of 310 bilingual news articles to properly evaluate lexical translation consistency. We manually differentiate those repeated source words whose translations are consistent into two types: true consistency and false consistency. Then based on the constructed test set, we evaluate the performance of lexical translation consistency for several typical NMT systems.

Pretrained language models and large language models are increasingly used to assist in a great variety of natural language tasks. In this work, we explore their use in evaluating the quality of alternative corpus annotation schemes. For this purpose, we analyze two alternative annotations of the Turkish BOUN treebank, versions 2.8 and 2.11, in the Universal Dependencies framework using large language models. Using a suitable prompt generated using treebank annotations, large language models are used to recover the surface forms of sentences. Based on the idea that the large language models capture the characteristics of the languages, we expect that the better annotation scheme would yield the sentences with higher success. The experiments conducted on a subset of the treebank show that the new annotation scheme (2.11) results in a successful recovery percentage of about 2 points higher. All the code developed for this work is available at https://github.com/boun-tabi/eval-ud .

We investigate GPT-SW3, a generative language model for the Nordic languages, to assess its understanding of the low-resourced Faroese language. Our aim is to demonstrate the advantages of using language-family-specific generative models to augment data for related languages with fewer resources. We evaluate GPT-SW3 by prompting it for Faroese to English translation in a zero, one, and few-shot setting. We assess such translations with an ensemble score consisting of an arithmetic average between the BLEU and a semantic similarity score (SBERT). Moreover, we challenge the model’s Faroese language understanding capabilities on a small dataset of curated Faroese trick sentences. There, we make a qualitative comparison of the model’s performance with respect to Open AI’s GPT-3.5 and GPT-4, demonstrating the advantages of using a language-family-specific generative model for navigating non-trivial scenarios. We evaluate the pipeline thus created and use it, as a proof of concept, to create an automatically annotated Faroese semantic textual similarity (STS) dataset.

Human evaluation is a critical component in machine translation system development and has received much attention in text translation research. However, little prior work exists on the topic of human evaluation for speech translation, which adds additional challenges such as noisy data and segmentation mismatches. We take the first steps to fill this gap by conducting a comprehensive human evaluation of the results of several shared tasks from the last International Workshop on Spoken Language Translation (IWSLT 2023). We propose an effective evaluation strategy based on automatic resegmentation and direct assessment with segment context. Our analysis revealed that: 1) the proposed evaluation strategy is robust and scores well-correlated with other types of human judgements; 2) automatic metrics are usually, but not always, well-correlated with direct assessment scores; and 3) COMET as a slightly stronger automatic metric than chrF, despite the segmentation noise introduced by the resegmentation step systems. We release the collected human-annotated data in order to encourage further investigation.

Pre-trained Large Acoustic Models, when fine-tuned, have largely shown to improve the performances in various tasks related to spoken language technologies. However, their evaluation has been mostly on datasets that contain English or other widely spoken languages, and their potential for novel under-resourced languages is not fully known. In this work, four novel under-resourced tribal languages that do not have a standard writing system were introduced and the application of such large pre-trained models was assessed to document such languages using Automatic Speech Recognition and Direct Speech-to-Text Translation systems. The transcriptions for these tribal languages were generated by adapting scripts from those languages that held a prominent presence in the geographical regions where these tribal languages are spoken. The results from this study suggest a viable direction to document these languages in the electronic domain by using Spoken Language Technologies that incorporate LAMs. Additionally, this study helped in understanding the varying performances exhibited by the Large Acoustic Model between these four languages. This study not only informs the adoption of appropriate scripts for transliterating spoken-only languages based on the language family but also aids in making informed decisions in analyzing the behavior of particular Large Acoustic Model in linguistic contexts.

Research projects incorporating spoken data require either a selection of existing speech corpora, or they plan to record new data. In both cases, recordings need to be transcribed to make them accessible to analysis. Underestimating the effort of transcribing can be risky. Automatic Speech Recognition (ASR) holds the promise to considerably reduce transcription effort. However, few studies have so far attempted to evaluate this potential. The present paper compares efforts for manual transcription vs. correction of ASR-output. We took recordings from corpora of varying settings (interview, colloquial talk, dialectal, historic) and (i) compared two methods for creating orthographic transcripts: transcribing from scratch vs. correcting automatically created transcripts. And (ii) we evaluated the influence of the corpus characteristics on the correcting efficiency. Results suggest that for the selected data and transcription conventions, transcribing and correcting still take equally long with 7 times real-time on average. The more complex the primary data, the more time has to be spent on corrections. Despite the impressive latest developments in speech technology, to be a real help for conversation analysts or dialectologists, ASR systems seem to require even more improvement, or we need sufficient and appropriate data for training such systems.

This paper replicates and evaluates the word expansion (WE) method for sentiment lexicon generation from Rheault et al. (2016), applying it to two novel corpora of parliamentary speech from Denmark and Bulgaria. GloVe embeddings and vector similarity are leveraged to expand synonym seed lists with domain-specific terms from the speech corpora. The resulting Danish and Bulgarian lexica are compared to other multilingual lexica by analyzing a gold standard of speech excerpts annotated for sentiment. WE correlates best with hand-coded annotations for Danish, while a machine-translated Lexicoder dictionary does best for Bulgarian. WE performance is also found to be very sensitive to processing and scoring techniques, though this is also an issue with the other lexica. Overall, automatic lexicon translation best balances computational complexity and accuracy across both languages, but robust language-agnosticism remains elusive. Theoretical and practical problems of WE are discussed.

Rationales in the form of manually annotated input spans usually serve as ground truth when evaluating explainability methods in NLP. They are, however, time-consuming and often biased by the annotation process. In this paper, we debate whether human gaze, in the form of webcam-based eye-tracking recordings, poses a valid alternative when evaluating importance scores. We evaluate the additional information provided by gaze data, such as total reading times, gaze entropy, and decoding accuracy with respect to human rationale annotations. We compare WebQAmGaze, a multilingual dataset for information-seeking QA, with attention and explainability-based importance scores for 4 different multilingual Transformer-based language models (mBERT, distil-mBERT, XLMR, and XLMR-L) and 3 languages (English, Spanish, and German). Our pipeline can easily be applied to other tasks and languages. Our findings suggest that gaze data offers valuable linguistic insights that could be leveraged to infer task difficulty and further show a comparable ranking of explainability methods to that of human rationales.

Sentence embeddings produced by Pretrained Language Models (PLMs) have received wide attention from the NLP community due to their superior performance when representing texts in numerous downstream applications. However, the high dimensionality of the sentence embeddings produced by PLMs is problematic when representing large numbers of sentences in memory- or compute-constrained devices. As a solution, we evaluate unsupervised dimensionality reduction methods to reduce the dimensionality of sentence embeddings produced by PLMs. Our experimental results show that simple methods such as Principal Component Analysis (PCA) can reduce the dimensionality of sentence embeddings by almost 50%, without incurring a significant loss in performance in multiple downstream tasks. Surprisingly, reducing the dimensionality further improves performance over the original high dimensional versions for the sentence embeddings produced by some PLMs in some tasks.