COLING 2024 论文列表

Research on conceptual abstraction has investigated the differences in contextual distributions, or “contextual variability,” of abstract and concrete concept words (e.g., *love* vs. *cat*). Empirical studies on this topic show that abstract words tend to occur in diverse linguistic contexts, while concrete words are typically constrained within more homogeneous contexts. Nonetheless, these investigations have somewhat overlooked a factor that influences both abstract and concrete concepts: *Categorial Specificity*, which denotes the inclusiveness of a category (e.g., *ragdoll* vs. *mammal*). We argue that more specific words are tied to narrower domains, independently or whether they are concrete or abstract, thus resulting in a diminished degree of contextual variability when compared to generic terms. In this study, we used distributional models to investigate the interplay between contextual variability, concreteness, specificity, and their interaction. Analyzing 676 English nouns, we found that contextual variability is explained by both concreteness and specificity: more specific words have closer contexts, while generic words, whether abstract or concrete, exhibit less related contexts.

Hate speech is infamously one of the most demanding topics in Natural Language Processing, as its multifacetedness is accompanied by a handful of challenges, such as multilinguality and cross-linguality. Hate speech has a subjective aspect that intensifies when referring to different cultures and different languages. In this respect, we design a pipeline that will help us explore the possibility of the creation of a parallel multilingual hate speech dataset, using machine translation. In this paper, we evaluate how/whether this is feasible by assessing the quality of the translations, calculating the toxicity levels of original and target texts, and calculating correlations between the newly obtained scores. Finally, we perform a qualitative analysis to gain further semantic and grammatical insights. With this pipeline we aim at exploring ways of filtering hate speech texts in order to parallelize sentences in multiple languages, examining the challenges of the task.

This short paper focuses on the coverage of eventive entries (verbs, predicates, etc.) of some well-known lexical semantic resources when applied to random running texts taken from the internet. While coverage gaps are often reported for manually created lexicons (which is the case of most semantically-oriented lexical ones), it was our aim to quantify these gaps, cross-lingually, on a new purely textual resource set produced by the HPLT Project from crawled internet data. Several English, German, Spanish and Czech lexical semantic resources (which, for the most part, focus on verbs and predicates) have been selected for this experiment. We also describe the challenges related to the fact that these resources are (to a varying extent) semantically oriented, meaning that the texts have to be preprocessed to obtain lemmas (base forms) and some types of MWEs before the coverage can be reasonably evaluated, and thus the results are necessarily only approximate. The coverage of these resources, with some exclusions as described in the paper, range from 41.00% to 97.33%, confirming the need to expand at least some - even well-known - resources to cover the prevailing source of today’s textual resources with regard to lexical units describing events or states (or possibly other eventive mentions).

Text feedback from urban scenes is a crucial tool for pedestrians to understand surroundings, obstacles, and safe pathways. However, existing image captioning datasets often concentrate on the overall image description and lack detailed scene descriptions, overlooking features for pedestrians walking on urban streets. We developed a new dataset to assist pedestrians in urban scenes using 360-degree camera images. Through our dataset of Text360Nav, we aim to provide textual feedback from machinery visual perception such as 360-degree cameras to visually impaired individuals and distracted pedestrians navigating urban streets, including those engrossed in their smartphones while walking. In experiments, we combined our dataset with multimodal generative models and observed that models trained with our dataset can generate textual descriptions focusing on street objects and obstacles that are meaningful in urban scenes in both quantitative and qualitative analyses, thus supporting the effectiveness of our dataset for urban pedestrian navigation.

Narratives have been the subject of extensive research across various scientific fields such as linguistics and computer science. However, the scarcity of freely available datasets, essential for studying this genre, remains a significant obstacle. Furthermore, datasets annotated with narratives components and their morphosyntactic and semantic information are even scarcer. To address this gap, we developed the Text2Story Lusa datasets, which consist of a collection of news articles in European Portuguese. The first datasets consists of 357 news articles and the second dataset comprises a subset of 117 manually densely annotated articles, totaling over 50 thousand individual annotations. By focusing on texts with substantial narrative elements, we aim to provide a valuable resource for studying narrative structures in European Portuguese news articles. On the one hand, the first dataset provides researchers with data to study narratives from various perspectives. On the other hand, the annotated dataset facilitates research in information extraction and related tasks, particularly in the context of narrative extraction pipelines. Both datasets are made available adhering to FAIR principles, thereby enhancing their utility within the research community.

The rapid proliferation of multimedia content has necessitated the development of effective multimodal video retrieval systems. Multimodal video retrieval is a non-trivial task involving retrieval of relevant information across different modalities, such as text, audio, and visual. This work aims to improve multimodal retrieval by guiding the creation of a shared embedding space with task-specific contrastive loss functions. An important aspect of our work is to propose a model that learns retrieval cues for the textual query from multiple modalities both separately and jointly within a hierarchical architecture that can be flexibly extended and fine-tuned for any number of modalities. To this end, the loss functions and the architectural design of the model are developed with a strong focus on increasing the mutual information between the textual and cross-modal representations. The proposed approach is quantitatively evaluated on the MSR-VTT and YouCook2 text-to-video retrieval benchmark datasets. The results showcase that the approach not only holds its own against state-of-the-art methods, but also outperforms them in a number of scenarios, with a notable relative improvements from baseline in R@1, R@5 and R@10 metrics.

Evaluating Text Style Transfer (TST) is a complex task due to its multi-faceted nature. The quality of the generated text is measured based on challenging factors, such as style transfer accuracy, content preservation, and overall fluency. While human evaluation is considered to be the gold standard in TST assessment, it is costly and often hard to reproduce. Therefore, automated metrics are prevalent in these domains. Nonetheless, it is uncertain whether and to what extent these automated metrics correlate with human evaluations. Recent strides in Large Language Models (LLMs) have showcased their capacity to match and even exceed average human performance across diverse, unseen tasks. This suggests that LLMs could be a viable alternative to human evaluation and other automated metrics in TST evaluation. We compare the results of different LLMs in TST evaluation using multiple input prompts. Our findings highlight a strong correlation between (even zero-shot) prompting and human evaluation, showing that LLMs often outperform traditional automated metrics. Furthermore, we introduce the concept of prompt ensembling, demonstrating its ability to enhance the robustness of TST evaluation. This research contributes to the ongoing efforts for more robust and diverse evaluation methods by standardizing and validating TST evaluation with LLMs.

Web-crawled corpora are essential resources for linguistic and NLP research, offering far more data than is available from curated corpora. However, they often contain a great deal of low-quality texts which can complicate research and degrade the quality of pre-trained language models. Therefore, they are typically filtered, e.g. by applying rules or classifiers. In this paper, we compare the effectiveness of various text filtering classifiers and measure their impact on language model performance for three medium-resource languages. We present TQ-IS, an Icelandic text quality dataset consisting of 2,000 web-crawled documents, in which spans of low-quality text have been manually identified and labeled. We then evaluate a perplexity-based classifier, a supervised classifier trained on TQ-IS, and a self-supervised classifier trained to discern between documents from curated and web-crawled corpora on Icelandic, Estonian and Basque. We find that these classifiers obtain F1 scores of 94.48%, 99.01% and 93.40%, respectively, when evaluated on the TQ-IS dataset. Furthermore, our results show that while adding filtered web-crawled text to a pre-training corpus can improve downstream performance for pre-trained language models, any improvement is likely to remain modest unless the web-crawled corpus is significantly larger in size.

Argument synthesis aims to generate rational claims, representing a fundamental objective in this field. While existing models excel in summarizing arguments and engaging in debates, we observe a critical gap in their ability to generate accurate arguments that incorporate forward-looking perspectives. In light of this observation, this paper introduces a novel task called “forward-looking claim planning.” We delve into this task by exploring the efficacy of well-performing classification and generation models. Furthermore, we propose several customized preprocessing methods that yield substantial performance improvements. Through comprehensive discussion and analysis, we also outline a future research agenda for the forward-looking claim planning task.

Reasoning over the Temporal Knowledge Graph (TKG) that predicts facts in the future has received much attention. Most previous works attempt to model temporal dynamics with knowledge graphs and graph convolution networks. However, these methods lack the consideration of high-order interactions between objects in TKG, which is an important factor to predict future facts. To address this problem, we introduce dynamic hypergraph embedding for temporal knowledge graph reasoning. Specifically, we obtain high-order interactions by constructing hypergraphs based on temporal knowledge graphs at different timestamps. Besides, we integrate the differences caused by time into the hypergraph representation in order to fit TKG. Then, we adapt dynamic meta-embedding for temporal hypergraph representation that allows our model to choose the appropriate high-order interactions for downstream reasoning. Experimental results on public TKG datasets show that our method outperforms the baselines. Furthermore, the analysis part demonstrates that the proposed method brings good interpretation for the predicted results.

A wide body of research is concerned with the semantics of narratives, both in terms of understanding narratives and generating fictional narratives and stories. We provide a dataset of summaries to be used as a proxy for entire stories or for the analysis of the summaries themselves. Our dataset consists of a total of 96,831 individual summaries across 29,505 stories. We intend for the dataset to be used for training and evaluation of embedding representations for stories, specifically the stories’ narratives. The summary data is harvested from five different language versions of Wikipedia. Our dataset comes with rich metadata, which we extract from Wikidata, enabling a wide range of applications that operate on story summaries in conjunction with metadata. To set baseline results, we run retrieval experiments on the dataset, exploring the capability of similarity models in retrieving summaries of the same story. For this retrieval, a crucial element is to not place too much emphasis on the named entities, as this can enable retrieval of other summaries for the same work without taking the narrative into account.

Large Language Models (LLMs) have achieved impressive results in Machine Translation by simply following instructions, even without training on parallel data. However, LLMs still face challenges on low-resource languages due to the lack of pre-training data. In real-world situations, humans can become proficient in their native languages through abundant and meaningful social interactions and can also learn foreign languages effectively using well-organized textbooks. Drawing inspiration from human learning patterns, we introduce the Translate After LEarNing Textbook (TALENT) approach, which aims to enhance LLMs’ ability to translate low-resource languages by learning from a textbook. TALENT follows a step-by-step process: (1) Creating a Textbook for low-resource languages. (2) Guiding LLMs to absorb the Textbook’s content for Syntax Patterns. (3) Enhancing translation by utilizing the Textbook and Syntax Patterns. We thoroughly assess TALENT’s performance using 112 low-resource languages from FLORES-200 with two LLMs: ChatGPT and BLOOMZ. Evaluation across three different metrics reveals that TALENT consistently enhances translation performance by 14.8% compared to zero-shot baselines. Further analysis demonstrates that TALENT not only improves LLMs’ comprehension of low-resource languages but also equips them with the knowledge needed to generate accurate and fluent sentences in these languages.

News headline generation is a crucial task in increasing productivity for both the readers and producers of news. This task can easily be aided by automated News headline-generation models. However, the presence of irrelevant headlines in scraped news articles results in sub-optimal performance of generation models. We propose that relevance-based headline classification can greatly aid the task of generating relevant headlines. Relevance-based headline classification involves categorizing news headlines based on their relevance to the corresponding news articles. While this task is well-established in English, it remains under-explored in low-resource languages like Telugu due to a lack of annotated data. To address this gap, we present TeClass, the first-ever human-annotated Telugu news headline classification dataset, containing 78,534 annotations across 26,178 article-headline pairs. We experiment with various baseline models and provide a comprehensive analysis of their results. We further demonstrate the impact of this work by fine-tuning various headline generation models using TeClass dataset. The headlines generated by the models fine-tuned on highly relevant article-headline pairs, showed about a 5 point increment in the ROUGE-L scores. To encourage future research, the annotated dataset as well as the annotation guidelines will be made publicly available.

Finetuning pretrained language models (LMs) have enabled appealing performance on a diverse array of tasks. The intriguing task-agnostic property has driven a shifted focus from task-specific to task-agnostic distillation of LMs. While task-agnostic, compute-efficient, performance-preserved LMs can be yielded by task-agnostic distillation, previous studies mainly sit in distillation of either encoder-only LMs (e.g., BERT) or decoder-only ones (e.g., GPT) yet largely neglect that distillation of encoder-decoder LMs (e.g., T5) can posit very distinguished behaviors. Frustratingly, we discover that existing task-agnostic distillation methods can fail to handle the distillation of encoder-decoder LMs. To the demand, we explore a few paths and uncover a path named as MiniEnD that successfully tackles the distillation of encoder-decoder LMs in a task-agnostic fashion. We examine MiniEnD on language understanding and abstractive summarization. The results showcase that MiniEnD is generally effective and is competitive compared to other alternatives. We further scale MiniEnD up to distillation of 3B encoder-decoder language models with interpolated distillation. The results imply the opportunities and challenges in distilling large language models (e.g., LLaMA).

Since paraphrasing is an ill-defined task, the term “paraphrasing” covers text transformation tasks with different characteristics. Consequently, existing paraphrasing studies have applied quite different (explicit and implicit) criteria as to when a pair of texts is to be considered a paraphrase, all of which amount to postulating a certain level of semantic or lexical similarity. In this paper, we conduct a literature review and propose a taxonomy to organize the 25 identified paraphrasing (sub-)tasks. Using classifiers trained to identify the tasks that a given paraphrasing instance fits, we find that the distributions of task-specific instances in the known paraphrase corpora vary substantially. This means that the use of these corpora, without the respective paraphrase conditions being clearly defined (which is the normal case), must lead to incomparable and misleading results.

In this paper, we propose a novel evaluation paradigm for Targeted Syntactic Evaluations, where we assess how well language models can recognize linguistic phenomena situated at different levels of the Chomsky hierarchy. Specifically, we create formal languages that abstract four syntactic phenomena in natural languages, each identified at a different level of the Chomsky hierarchy, and use these to evaluate the capabilities of language models: (1) (Adj)ˆn NP type, (2) NPˆn VPˆn type, (3) Nested Dependency type, and (4) Cross Serial Dependency type. We first train three different language models (LSTM, Transformer LM, and Stack-RNN) on language modeling tasks and then evaluate them using pairs of a positive and a negative sentence by investigating whether they can assign a higher probability to the positive sentence than the negative one. Our result demonstrated that all language models have the ability to capture the structural patterns of the (Adj)ˆn NP type formal language. However, LSTM and Transformer LM failed to capture NPˆn VPˆn type language and no architectures can recognize nested dependency and Cross Serial dependency correctly. Neural language models, especially Transformer LMs, have exhibited high performance across a multitude of downstream tasks, leading to the perception that they possess an understanding of natural languages. However, our findings suggest that these models may not necessarily comprehend the syntactic structures that underlie natural language phenomena such as dependency. Rather, it appears that they may extend grammatical rules equivalent to regular grammars to approximate the rules governing dependencies.

Stance detection is a fundamental task in Natural Language Processing (NLP). It is challenging due to diverse expressions and topics related to the targets from multiple domains. Recently, prompt-tuning has been introduced to convert the original task into a cloze-style prediction task, achieving impressive results. Many prompt-tuning-based methods focus on one or two classic scenarios with concrete external knowledge enhancement. However, when facing intricate information in multi-domain stance detection, these methods cannot be adaptive to multi-domain semantics. In this paper, we propose a novel target-adaptive consistency enhanced prompt-tuning method (TCP) for stance detection with multiple domains. TCP incorporates target knowledge and prior knowledge to construct target-adaptive verbalizers for diverse domains and employs pilot experiments distillation to enhance the consistency between verbalizers and model training. Specifically, to capture the knowledge from multiple domains, TCP uses a target-adaptive candidate mining strategy to obtain the domain-related candidates. Then, TCP refines them with prior attributes to ensure prediction consistency. The Pre-trained Language Models (PLMs) in prompt-tuning are with large-scale parameters, while only changing the verbalizer without corresponding tuning has a limited impact on the training process. Target-aware pilot experiments are conducted to enhance the consistency between the verbalizer and training by distilling the target-adaptive knowledge into prompt-tuning. Extensive experiments and ablation studies demonstrate that TCP outperforms the state-of-the-art methods on nine stance detection datasets from multiple domains.

Researchers have attempted to mitigate lexical bias in toxic language detection (TLD). However, existing methods fail to disentangle the “useful” and “misleading” impact of lexical bias on model decisions. Therefore, they do not effectively exploit the positive effects of the bias and lead to a degradation in the detection performance of the debiased model. In this paper, we propose a Counterfactual Causal Debiasing Framework (CCDF) to mitigate lexical bias in TLD. It preserves the “useful impact” of lexical bias and eliminates the “misleading impact”. Specifically, we first represent the total effect of the original sentence and biased tokens on decisions from a causal view. We then conduct counterfactual inference to exclude the direct causal effect of lexical bias from the total effect. Empirical evaluations demonstrate that the debiased TLD model incorporating CCDF achieves state-of-the-art performance in both accuracy and fairness compared to competitive baselines applied on several vanilla models. The generalization capability of our model outperforms current debiased models for out-of-distribution data.

Recent research shows that pre-trained language models (PLMs) suffer from “prompt bias” in factual knowledge extraction, i.e., prompts tend to introduce biases toward specific labels. Prompt bias presents a significant challenge in assessing the factual knowledge within PLMs. Therefore, this paper aims to improve the reliability of existing benchmarks by thoroughly investigating and mitigating prompt bias. We show that: 1) all prompts in the experiments exhibit non-negligible bias, with gradient-based prompts like AutoPrompt and OptiPrompt displaying significantly higher levels of bias; 2) prompt bias can amplify benchmark accuracy unreasonably by overfitting the test datasets, especially on imbalanced datasets like LAMA. Based on these findings, we propose a representation-based approach to mitigate the prompt bias during inference time. Specifically, we first estimate the biased representation using prompt-only querying, and then remove it from the model’s internal representations to generate the debiased representations, which are used to produce the final debiased outputs. Experiments across various prompts, PLMs, and benchmarks show that our approach can not only correct the overfitted performance caused by prompt bias, but also significantly improve the prompt retrieval capability (up to 10% absolute performance gain). These results indicate that our approach effectively alleviates prompt bias in knowledge evaluation, thereby enhancing the reliability of benchmark assessments. Hopefully, our plug-and-play approach can be a golden standard to strengthen PLMs toward reliable knowledge bases. Code and data are released in https://github.com/FelliYang/PromptBias.