COLING 2024 论文列表

Weakly-supervised Phrase Grounding (WPG) is an emerging task of inferring the fine-grained phrase-region matching, while merely leveraging the coarse-grained sentence-image pairs for training. However, existing studies on WPG largely ignore the implicit phrase-region matching relations, which are crucial for evaluating the capability of models in understanding the deep multimodal semantics. To this end, this paper proposes an Implicit-Enhanced Causal Inference (IECI) approach to address the challenges of modeling the implicit relations and highlighting them beyond the explicit. Specifically, this approach leverages both the intervention and counterfactual techniques to tackle the above two challenges respectively. Furthermore, a high-quality implicit-enhanced dataset is annotated to evaluate IECI and detailed evaluations show the great advantages of IECI over the state-of-the-art baselines. Particularly, we observe an interesting finding that IECI outperforms the advanced multimodal LLMs by a large margin on this implicit-enhanced dataset, which may facilitate more research to evaluate the multimodal LLMs in this direction.

Due to the broad range of social media platforms, the requirements of abusive language detection systems are varied and ever-changing. Already a large set of annotated corpora with different properties and label sets were created, such as hate or misogyny detection, but the form and targets of abusive speech are constantly evolving. Since, the annotation of new corpora is expensive, in this work we leverage datasets we already have, covering a wide range of tasks related to abusive language detection. Our goal is to build models cheaply for a new target label set and/or language, using only a few training examples of the target domain. We propose a two-step approach: first we train our model in a multitask fashion. We then carry out few-shot adaptation to the target requirements. Our experiments show that using already existing datasets and only a few-shots of the target task the performance of models improve both monolingually and across languages. Our analysis also shows that our models acquire a general understanding of abusive language, since they improve the prediction of labels which are present only in the target dataset and can benefit from knowledge about labels which are not directly used for the target task.

Current language models require a lot of training data to obtain high performance. For Relation Classification (RC), many datasets are domain-specific, so combining datasets to obtain better performance is non-trivial. We explore a multi-domain training setup for RC, and attempt to improve performance by encoding domain information. Our proposed models improve > 2 Macro-F1 against the baseline setup, and our analysis reveals that not all the labels benefit the same: The classes which occupy a similar space across domains (i.e., their interpretation is close across them, for example “physical”) benefit the least, while domain-dependent relations (e.g., “part-of”) improve the most when encoding domain information.

This paper presents a new perspective on framing through the lens of speech acts and investigates how politicians make use of different pragmatic speech act functions in political debates. To that end, we created a new resource of German parliamentary debates, annotated with fine-grained speech act types. Our hierarchical annotation scheme distinguishes between cooperation and conflict communication, further structured into six subtypes, such as informative, declarative or argumentative-critical speech acts, with 14 fine-grained classes at the lowest level. We present classification baselines on our new data and show that the fine-grained classes in our schema can be predicted with an avg. F1 of around 82.0%. We then use our classifier to analyse the use of speech acts in a large corpus of parliamentary debates over a time span from 2003–2023.

This work investigates the acquisition of formal linguistic competence by neural language models, hypothesizing that languages with complex grammar, such as Basque, present substantial challenges during the pre-training phase. Basque is distinguished by its complex morphology and flexible word order, potentially complicating grammar extraction. In our analysis, we evaluated the grammatical knowledge of BERT models trained under various pre-training configurations, considering factors such as corpus size, model size, number of epochs, and the use of lemmatization. To assess this grammatical knowledge, we constructed the BL2MP (Basque L2 student-based Minimal Pairs) test set. This test set consists of minimal pairs, each containing both a grammatically correct and an incorrect sentence, sourced from essays authored by students at different proficiency levels in the Basque language. Additionally, our analysis explores the difficulties in learning various grammatical phenomena, the challenges posed by flexible word order, and the influence of the student’s proficiency level on the difficulty of correcting grammar errors.

This paper investigates the rational thinking capability of Large Language Models (LLMs) in multi-round argumentative debates by exploring the impact of fallacious arguments on their logical reasoning performance. More specifically, we present Logic Competence Measurement Benchmark (LOGICOM), a diagnostic benchmark to assess the robustness of LLMs against logical fallacies. LOGICOM involves two agents: a persuader and a debater engaging in a multi-round debate on a controversial topic, where the persuader tries to convince the debater of the correctness of its claim. First, LOGICOM assesses the potential of LLMs to change their opinions through reasoning. Then, it evaluates the debater’s performance in logical reasoning by contrasting the scenario where the persuader employs logical fallacies against one where logical reasoning is used. We use this benchmark to evaluate the performance of GPT-3.5 and GPT-4 using a dataset containing controversial topics, claims, and reasons supporting them. Our findings indicate that both GPT-3.5 and GPT-4 can adjust their opinion through reasoning. However, when presented with logical fallacies, GPT-3.5 and GPT-4 are erroneously convinced 41% and 69% more often, respectively, compared to when logical reasoning is used. Finally, we introduce a new dataset containing over 5k pairs of logical vs. fallacious arguments.

Large language models (LLMs) have drawn great attention from the field of natural language processing and beyond, due to their impressive capability of autoregressive modeling, yet bringing an obvious problem, i.e., the largely increased latency. An emerging idea to alleviate this problem is speculative decoding, which first uses a draft model to draft tokens autoregressively and then makes the target model verify these tokens in parallel. The draft model is typically smaller than the target model, and it essentially trades generation quality for speed. Thereby, speculative decoding can be viewed as a speculative game for the target model in term of verification failures. That is, the lengthy draft tokens proposed by the small draft models could fail in the verification stage. Naturally, a critical question arises: how speculative can speculative decoding be, or in other words, how small can an adequate draft model be and how large can an appropriate number of draft tokens be? This work aims to investigate these questions and demonstrate how the scale of the draft model and the number of draft tokens would have an impact on the overall latency of the speculative decoding. We theoretically show that neither of above two factors will be infinitely speculative. Namely, there is a certain turning point for each of them. We then empirically show that the scale of the draft model could be 10-20× smaller than the target model and the optimal number of draft tokens should lie in 3-5.

Question-answering (QA) on hybrid scientific tabular and textual data deals with scientific information, and relies on complex numerical reasoning. In recent years, while tabular QA has seen rapid progress, understanding their robustness on scientific information is lacking due to absence of any benchmark dataset. To investigate the robustness of the existing state-of-the-art QA models on scientific hybrid tabular data, we propose a new dataset, “SciTabQA”, consisting of 822 question-answer pairs from scientific tables and their descriptions. With the help of this dataset, we assess the state-of-the-art Tabular QA models based on their ability (i) to use heterogeneous information requiring both structured data (table) and unstructured data (text) and (ii) to perform complex scientific reasoning tasks. In essence, we check the capability of the models to interpret scientific tables and text. Our experiments show that “SciTabQA” is an innovative dataset to study question-answering over scientific heterogeneous data. We benchmark three state-of-the-art Tabular QA models, and find that the best F1 score is only 0.462.

This paper concerns the pressing need to understand and manage inappropriate language within the evolving human-robot interaction (HRI) landscape. As intelligent systems and robots transition from controlled laboratory settings to everyday households, the demand for polite and culturally sensitive conversational abilities becomes paramount, especially for younger individuals. This study explores data cleaning methods, focussing on rudeness and contextual similarity, to identify and mitigate inappropriate language in real-time interactions. State-of-the-art natural language models are also evaluated for their proficiency in discerning rudeness. This multifaceted investigation highlights the challenges of handling inappropriate language, including its tendency to hide within idiomatic expressions and its context-dependent nature. This study will further contribute to the future development of AI systems capable of engaging in intelligent conversations and upholding the values of courtesy and respect across diverse cultural and generational boundaries.

Previous work has showcased the intriguing capability of large language models (LLMs) in retrieving facts and processing context knowledge. However, only limited research exists on the layer-wise capability of LLMs to encode knowledge, which challenges our understanding of their internal mechanisms. In this paper, we devote the first attempt to investigate the layer-wise capability of LLMs through probing tasks. We leverage the powerful generative capability of ChatGPT to construct probing datasets, providing diverse and coherent evidence corresponding to various facts. We employ V-usable information as the validation metric to better reflect the capability in encoding context knowledge across different layers. Our experiments on conflicting and newly acquired knowledge show that LLMs: (1) prefer to encode more context knowledge in the upper layers; (2) primarily encode context knowledge within knowledge-related entity tokens at lower layers while progressively expanding more knowledge within other tokens at upper layers; and (3) gradually forget the earlier context knowledge retained within the intermediate layers when provided with irrelevant evidence. Code is publicly available at https://github.com/Jometeorie/probing_llama.

Subword tokenization has become the prevailing standard in the field of natural language processing (NLP) over recent years, primarily due to the widespread utilization of pre-trained language models. This shift began with Byte-Pair Encoding (BPE) and was later followed by the adoption of SentencePiece and WordPiece. While subword tokenization consistently outperforms character and word-level tokenization, the precise factors contributing to its success remain unclear. Key aspects such as the optimal segmentation granularity for diverse tasks and languages, the influence of data sources on tokenizers, and the role of morphological information in Indo-European languages remain insufficiently explored. This is particularly pertinent for biomedical terminology, characterized by specific rules governing morpheme combinations. Despite the agglutinative nature of biomedical terminology, existing language models do not explicitly incorporate this knowledge, leading to inconsistent tokenization strategies for common terms. In this paper, we seek to delve into the complexities of subword tokenization in French biomedical domain across a variety of NLP tasks and pinpoint areas where further enhancements can be made. We analyze classical tokenization algorithms, including BPE and SentencePiece, and introduce an original tokenization strategy that integrates morpheme-enriched word segmentation into existing tokenization methods.

Out-of-distribution (OOD) detection plays a vital role in enhancing the reliability of machine learning models. As large language models (LLMs) become more prevalent, the applicability of prior research on OOD detection that utilized smaller-scale Transformers such as BERT, RoBERTa, and GPT-2 may be challenged, due to the significant differences in the scale of these models, their pre-training objectives, and the paradigms used for inference. This paper initiates a pioneering empirical investigation into the OOD detection capabilities of LLMs, focusing on the LLaMA series ranging from 7B to 65B in size. We thoroughly evaluate commonly used OOD detectors, examining their performance in both zero-grad and fine-tuning scenarios. Notably, we alter previous discriminative in-distribution fine-tuning into generative fine-tuning, aligning the pre-training objective of LLMs with downstream tasks. Our findings unveil that a simple cosine distance OOD detector demonstrates superior efficacy, outperforming other OOD detectors. We provide an intriguing explanation for this phenomenon by highlighting the isotropic nature of the embedding spaces of LLMs, which distinctly contrasts with the anisotropic property observed in smaller BERT family models. The new insight enhances our understanding of how LLMs detect OOD data, thereby enhancing their adaptability and reliability in dynamic environments. We have released the source code at https://github.com/Awenbocc/LLM-OOD for other researchers to reproduce our results.

Neural language models, which reach state-of-the-art results on most natural language processing tasks, are trained on large text corpora that inevitably contain value-burdened content and often capture undesirable biases, which the models reflect. This case study focuses on the political biases of pre-trained encoders in Czech and compares them with a representative value survey. Because Czech is a gendered language, we also measure how the grammatical gender coincides with responses to men and women in the survey. We introduce a novel method for measuring the model’s perceived political values. We find that the models do not assign statement probability following value-driven reasoning, and there is no systematic difference between feminine and masculine sentences. We conclude that BERT-sized models do not manifest systematic alignment with political values and that the biases observed in the models are rather due to superficial imitation of training data patterns than systematic value beliefs encoded in the models.

Pre-trained masked language models, such as BERT, perform strongly on a wide variety of NLP tasks and have become ubiquitous in recent years. The typical way to use such models is to fine-tune them on downstream data. In this work, we aim to study how the difference in domains between the pre-trained model and the task effects its final performance. We first devise a simple mechanism to quantify the domain difference (using a cloze task) and use it to partition our dataset. Using these partitions of varying domain discrepancy, we focus on answering key questions around the impact of discrepancy on final performance, robustness to out-of-domain test-time examples and effect of domain-adaptive pre-training. We base our experiments on a large-scale openly available e-commerce dataset, and our findings suggest that in spite of pre-training the performance of BERT degrades on datasets with high domain discrepancy, especially in low resource cases. This effect is somewhat mitigated by continued pre-training for domain adaptation. Furthermore, the domain-gap also makes BERT sensitive to out-of-domain examples during inference, even in high resource tasks, and it is prudent to use as diverse a dataset as possible during fine-tuning to make it robust to domain shift.

The attention mechanism, a cornerstone of state-of-the-art neural models, faces computational hurdles in processing long sequences due to its quadratic complexity. Consequently, research efforts in the last few years focused on finding more efficient alternatives. Among them, Hyena (Poli et al., 2023) stands out for achieving competitive results in both language modeling and image classification, while offering sub-quadratic memory and computational complexity. Building on these promising results, we propose ConfHyena, a Conformer whose encoder self-attentions are replaced with an adaptation of Hyena for speech processing, where the long input sequences cause high computational costs. Through experiments in automatic speech recognition (for English) and translation (from English into 8 target languages), we show that our best ConfHyena model significantly reduces the training time by 27%, at the cost of minimal quality degradation (∼1%), which, in most cases, is not statistically significant.

We investigate disputes in the United Nations Security Council (UNSC) by studying the linguistic means of expressing conflicts. As a result, we present the UNSC Conflict Corpus (UNSCon), a collection of 87 UNSC speeches that are annotated for conflicts. We explain and motivate our annotation scheme and report on a series of experiments for automatic conflict classification. Further, we demonstrate the difficulty when dealing with diplomatic language - which is highly complex and often implicit along various dimensions - by providing corpus examples, readability scores, and classification results.

The authorship of the Homeric poems has been a matter of debate for centuries. Computational approaches such as language modeling exist that can aid experts in making crucial headway. We observe, however, that such work has, thus far, only been carried out at the level of lengthier excerpts, but not individual verses, the level at which most suspected interpolations occur. We address this weakness by presenting a corpus of Homeric verses, each complemented with a score quantifying linguistic unexpectedness based on Perplexity. We assess the nature of these scores by exploring their correlation with named entities, the frequency of character n-grams, and (inverse) word frequency, revealing robust correlations with the latter two. This apparent bias can be partly overcome by simply dividing scores for unexpectedness by the maximum term frequency per verse.

This work revisits the topic of jointly parsing constituency and dependency trees, i.e., to produce compatible constituency and dependency trees simultaneously for input sentences, which is attractive considering that the two types of trees are complementary in representing syntax. The original work of Zhou and Zhao (2019) performs joint parsing only at the inference phase. They train two separate parsers under the multi-task learning framework (i.e., one shared encoder and two independent decoders). They design an ad-hoc dynamic programming-based decoding algorithm of O(n5) time complexity for finding optimal compatible tree pairs. Compared to their work, we make progress in three aspects: (1) adopting a much more efficient decoding algorithm of O(n4) time complexity, (2) exploring joint modeling at the training phase, instead of only at the inference phase, (3) proposing high-order scoring components to promote constituent-dependency interaction. We conduct experiments and analysis on seven languages, covering both rich-resource and low-resource scenarios. Results and analysis show that joint modeling leads to a modest overall performance boost over separate modeling, but substantially improves the complete matching ratio of whole trees, thanks to the explicit modeling of tree compatibility.

Cross-modal retrieval is an important yet challenging task due to the semantic discrepancy between visual content and language. To measure the correlation between images and text, most existing research mainly focuses on learning global or local correspondence, failing to explore fine-grained local-global alignment. To infer more accurate similarity scores, we introduce a novel High Order Semantic Alignment (HOSA) model that can provide complementary and comprehensive semantic clues. Specifically, to jointly learn global and local alignment and emphasize local-global interaction, we employ tensor-product (t-product) operation to reconstruct one modal’s representation based on another modal’s information in a common semantic space. Such a cross-modal reconstruction strategy would significantly enhance inter-modal correlation learning in a fine-grained manner. Extensive experiments on two benchmark datasets validate that our model significantly outperforms several state-of-the-art baselines, especially in retrieving the most relevant results.