COLING 2024 论文列表

While nationality is a pivotal demographic element that enhances the performance of language models, it has received far less scrutiny regarding inherent biases. This study investigates nationality bias in ChatGPT (GPT-3.5), a large language model (LLM) designed for text generation. The research covers 195 countries, 4 temperature settings, and 3 distinct prompt types, generating 4,680 discourses about nationality descriptions in Chinese and English. Automated metrics were used to analyze the nationality bias, and expert annotators alongside ChatGPT itself evaluated the perceived bias. The results show that ChatGPT’s generated discourses are predominantly positive, especially compared to its predecessor, GPT-2. However, when prompted with negative inclinations, it occasionally produces negative content. Despite ChatGPT considering its generated text as neutral, it shows consistent self-awareness about nationality bias when subjected to the same pair-wise comparison annotation framework used by human annotators. In conclusion, while ChatGPT’s generated texts seem friendly and positive, they reflect the inherent nationality biases in the real world. This bias may vary across different language versions of ChatGPT, indicating diverse cultural perspectives. The study highlights the subtle and pervasive nature of biases within LLMs, emphasizing the need for further scrutiny.

Large Language Models (LLMs) are showing emerging abilities, and one of the latest recognized ones deals with their ability to reason and answer questions from tabular data. Although there are some available datasets to assess question answering systems on tabular data, they are not large and diverse enough to properly assess the capabilities of LLMs. To this end, we propose DataBench, a benchmark composed of 65 real-world datasets over several domains, including 20 human-generated questions per dataset, totaling 1300 questions and answers overall. Using this benchmark, we perform a large-scale empirical comparison of several open and closed source models, including both code-generating and in-context learning models. The results highlight the current gap between open-source and closed-source models, with all types of model having room for improvement even in simple boolean questions or involving a single column.

We present QueryNER, a manually-annotated dataset and accompanying model for e-commerce query segmentation. Prior work in sequence labeling for e-commerce has largely addressed aspect-value extraction which focuses on extracting portions of a product title or query for narrowly defined aspects. Our work instead focuses on the goal of dividing a query into meaningful chunks with broadly applicable types. We report baseline tagging results and conduct experiments comparing token and entity dropping for null and low recall query recovery. Challenging test sets are created using automatic transformations and show how simple data augmentation techniques can make the models more robust to noise. We make the QueryNER dataset publicly available.

Legal professionals often grapple with navigating lengthy legal judgements to pinpoint information that directly address their queries. This paper focus on this task of extracting relevant paragraphs from legal judgements based on the query. We construct a specialized dataset for this task from the European Court of Human Rights (ECtHR) using the case law guides. We assess the performance of current retrieval models in a zero-shot way and also establish fine-tuning benchmarks using various models. The results highlight the significant gap between fine-tuned and zero-shot performance, emphasizing the challenge of handling distribution shift in the legal domain. We notice that the legal pre-training handles distribution shift on the corpus side but still struggles on query side distribution shift, with unseen legal queries. We also explore various Parameter Efficient Fine-Tuning (PEFT) methods to evaluate their practicality within the context of information retrieval, shedding light on the effectiveness of different PEFT methods across diverse configurations with pre-training and model architectures influencing the choice of PEFT method.

Spoken content is abundant – including podcasts, meeting transcripts, and TikTok-like short videos. And yet, many important tasks like summarization are often designed for written content rather than the looser, noiser, and more disfluent style of spoken content. Hence, we aim in this paper to quantify the impact of disfluency on spoken content summarization. Do disfluencies negatively impact the quality of summaries generated by existing approaches? And if so, to what degree? Coupled with these goals, we also investigate two methods towards improving summarization in the presence of such disfluencies. We find that summarization quality does degrade with an increase in these disfluencies and that a combination of multiple disfluency types leads to even greater degradation. Further, our experimental results show that naively removing disfluencies and augmenting with special tags can worsen the summarization when used for testing, but that removing disfluencies for fine-tuning yields the best results. We make the code available at https://github.com/mariateleki/Quantifying-Impact-Disfluency.

The questionnaire is a professional research methodology used for both qualitative and quantitative analysis of human opinions, preferences, attitudes, and behaviors. However, designing and evaluating questionnaires demands significant effort due to their intricate and complex structure. Questionnaires entail a series of questions that must conform to intricate constraints involving the questions, options, and overall structure. Specifically, the questions should be relevant and specific to the given research topic and intent. The options should be tailored to the questions, ensuring they are mutually exclusive, completed, and ordered sensibly. Moreover, the sequence of questions should follow a logical order, grouping similar topics together. As a result, automatically generating questionnaires presents a significant challenge and this area has received limited attention primarily due to the scarcity of high-quality datasets. To address these issues, we present Qsnail, the first dataset specifically constructed for the questionnaire generation task, which comprises 13,168 human-written questionnaires gathered from online platforms. We further conduct experiments on Qsnail, and the results reveal that retrieval models and traditional generative models do not fully align with the given research topic and intents. Large language models, while more closely related to the research topic and intents, exhibit significant limitations in terms of diversity and specificity. Despite enhancements through the chain-of-thought prompt and finetuning, questionnaires generated by language models still fall short of human-written questionnaires. Therefore, questionnaire generation is challenging and needs to be further explored. The dataset will be published in the future.

We present Qabas, a novel open-source Arabic lexicon designed for NLP applications. The novelty of Qabas lies in its synthesis of 110 lexicons. Specifically, Qabas lexical entries (lemmas) are assembled by linking lemmas from 110 lexicons. Furthermore, Qabas lemmas are also linked to 12 morphologically annotated corpora (about 2M tokens), making it the first Arabic lexicon to be linked to lexicons and corpora. Qabas was developed semi-automatically, utilizing a mapping framework and a web-based tool. Compared with other lexicons, Qabas stands as the most extensive Arabic lexicon, encompassing about 58K lemmas (45K nominal lemmas, 12.5K verbal lemmas, and 473 functional-word lemmas). Qabas is open-source and accessible online at https://sina.birzeit.edu/qabas

The paper describes a dataset composed of two sub-corpora from two different sources in Italian. The QUEEREOTYPES corpus includes social media texts regarding LGBTQIA+ individuals, behaviors, ideology and events. The texts were collected from Facebook and Twitter in 2018 and were annotated for the presence of stereotypes, and orthogonal dimensions (such as hate speech, aggressiveness, offensiveness, and irony in one sub-corpus, and stance in the other). The resource was developed by Natural Language Processing researchers together with activists from an Italian LGBTQIA+ not-for-profit organization. The creation of the dataset allows the NLP community to study stereotypes against marginalized groups, individuals and, ultimately, to develop proper tools and measures to reduce the online spread of such stereotypes. A test for the robustness of the language resource has been performed by means of 5-fold cross-validation experiments. Finally, text classification experiments have been carried out with a fine-tuned version of AlBERTo (a BERT-based model pre-trained on Italian tweets) and mBERT, obtaining good results on the task of stereotype detection, suggesting that stereotypes towards different targets might share common traits.

Chain-of-Thought prompting has improved reasoning capability of large language models (LLM). However, it still is challenging to guarantee the effectiveness and stability for questions requiring complicated reasoning. Recently, Plan-and-Solve prompting enhances the reasoning capability for complex questions by planning the solution steps firstly and then solving them step by step, but it suffers the difficulty to represent and execute the problem-solving logic of complex questions. To deal with these challenges, in this work, we propose a novel Plan-and-Solve prompting method based on Question Decomposition Meaning Representation (QDMR). Specifically, this method first allows the LLM to generate a QDMR graph to represent the problem-solving logic, which is a directed acyclic graph composed of sub-questions. Then, the LLM generates a specific solving process based on the QDMR graph. When solving each sub-question, it can locate the preceding sub-questions and their answers according to the QDMR graph, and then utilize this information for solution. Compared with existing Plan-and-Solve prompting techniques, our method can not only represent the problem-solving logic of complicated questions more accurately with the aid of QDMR graph, but also deliver the dependence information accurately for different solution steps according to the QDMR graph. In addition, with the supervised fine-tuning on the Allen Institute dataset, the decomposing capability of LLM for complicated questions can be considerably enhanced. Extensive experiments show that our method has achieve a great significance in arithmetic reasoning and commonsense reasoning task by comparing the classical Chain-of-Thought prompting and Plan-and-Solve prompting techniques, and the improvements achieved are even greater for problems with more reasoning steps.

This paper presents the creation of the Qatari Corpus of Argumentative Writing (QCAW) as an annotated L1 Arabic and L2 English bilingual writer corpus. It comprises 200,000 tokens of argumentative writing by Qatari university students in L1 Arabic and L2 English. The corpus includes 195 essays written by 195 students, 159 females and 36 males. The students were native Arabic speakers proficient in English as a second language. The corpus is divided into Arabic and English sections, accompanied by part-of-speech annotated files. The Metadata contains information about the students (gender, major, first and second languages) and the essays (text serial numbers, word limits, genre, writing date, time spent, and location). The paper outlines the steps for collecting and analysing the corpus, including details on essay writers, topic selection, pre-analysis text modifications, proficiency level, gender, and major ratings. Statistical analyses were applied to examine the corpus. The QCAW offers a valuable bilingual data source authored by the same students in Arabic and English, with implications for further research

Temporal relation annotation in the clinical domain is crucial yet challenging due to its workload and the medical expertise required. In this paper, we propose a novel annotation method that integrates event start-points ordering and question-answering (QA) as the annotation format. By focusing only on two points on a timeline, start-points ordering reduces ambiguity and simplifies the relation set to be considered during annotation. QA as annotation recasts temporal relation annotation into a reading comprehension task, allowing annotators to use natural language instead of the formalisms commonly adopted in temporal relation annotation. Based on our method, most of the relations in a document are inferable from a significantly smaller number of explicitly annotated relations, showing the efficiency of our proposed method. Using these inferred relations, we develop a temporal relation classification model that achieves a 0.72 F1 score. Also, by decomposing the annotation process into QA generation and QA validation, our method enables collaboration among medical experts and non-experts. We obtained high inter-annotator agreement (IAA) scores, which indicate the positive prospect of such collaboration in the annotation process. Our annotated corpus, annotation tool, and trained model are publicly available: https://github.com/seiji-shimizu/qa-start-ordering.

We introduce PyRater, an open-source Python toolkit designed for analysing corpora annotations. When creating new annotated language resources, probabilistic models of annotation are the state-of-the-art solution for identifying the best annotators, retrieving the gold standard, and more generally separating annotation signal from noise. PyRater offers a unified interface for several such models and includes an API for the addition of new ones. Additionally, the toolkit has built-in functions to read datasets with multiple annotations and plot the analysis outcomes. In this work, we also demonstrate a novel application of PyRater to zero-shot classifiers, where it effectively selects the best-performing prompt. We make PyRater available to the research community.

Humor is an intricate part of verbal communication and dealing with this kind of phenomenon is essential to building systems that can process language at large with all of its complexities. In this paper, we introduce Puntuguese, a new corpus of punning humor in Portuguese, motivated by previous works showing that currently available corpora for this language are still unfit for Machine Learning due to data leakage. Puntuguese comprises 4,903 manually-gathered punning one-liners in Brazilian and European Portuguese. To create negative examples that differ exclusively in terms of funniness, we carried out a micro-editing process, in which all jokes were edited by fluent Portuguese speakers to make the texts unfunny. Finally, we did some experiments on Humor Recognition, showing that Puntuguese is considerably more difficult than the previous corpus, achieving an F1-Score of 68.9%. With this new dataset, we hope to enable research not only in NLP but also in other fields that are interested in studying humor; thus, the data is publicly available.

Linguistic data, a component critical not only for research in a variety of fields but also for the development of various Natural Language Processing (NLP) applications, can contain personal information. As a result, its accessibility is limited, both from a legal and an ethical standpoint. One of the solutions is the pseudonymization of the data. Key stages of this process include the identification of sensitive elements and the generation of suitable surrogates in a way that the data is still useful for the intended task. Within this paper, we conduct an analysis of tagsets that have previously been utilized in anonymization and pseudonymization. We also investigate what kinds of Personally Identifiable Information (PII) appear in various domains. These reveal that none of the analyzed tagsets account for all of the PII types present cross-domain at the level of detailedness seemingly required for pseudonymization. We advocate for a universal system of tags for categorizing PIIs leading up to their replacement. Such categorization could facilitate the generation of grammatically, semantically, and sociolinguistically appropriate surrogates for the kinds of information that are considered sensitive in a given domain, resulting in a system that would enable dynamic pseudonymization while keeping the texts readable and useful for future research in various fields.

Structured pruning is an effective technique for compressing pre-trained language models (PLMs), reducing model size and improving inference speed for efficient deployment. However, most of existing pruning algorithms require retraining, leading to additional computational overhead. While some retraining-free approaches have been proposed for classification tasks, they still require a fully fine-tuned model for the task, and may cause catastrophic performance degradation on generative tasks. To address these challenges, we propose P-pruning (pre-pruning), an innovative task-specific compression framework. P-pruning prunes redundant modules of PLMs before fine-tuning, reducing the costs associated with fine-tuning. We also introduce a pruning algorithm for this framework, which includes two techniques: (1) module clustering, which clusters the outputs of all heads and neurons based on the task input; and (2) centroid selection, which identifies the most salient element in each cluster and prunes the others. We apply our method to BERT and GPT-2 and evaluate its effectiveness on GLUE, SQuAD, WikiText-2, WikiText-103, and PTB datasets. Experimental results demonstrate that our approach achieves higher performance in both classification and generative tasks, while also reducing the time required for fine-tuning.

Few-shot Event Detection (FSED) is a meaningful task due to the limited labeled data and expensive manual labeling. Some prompt-based methods are used in FSED. However, these methods require large GPU memory due to the increased length of input tokens caused by concatenating prompts, as well as additional human effort for designing verbalizers. Moreover, they ignore instance and prompt biases arising from the confounding effects between prompts and texts. In this paper, we propose a prototype-based prompt-instance Interaction with causal Intervention (2xInter) model to conveniently utilize both prompts and verbalizers and effectively eliminate all biases. Specifically, 2xInter first presents a Prototype-based Prompt-Instance Interaction (PPII) module that applies an interactive approach for texts and prompts to reduce memory and regards class prototypes as verbalizers to avoid design costs. Next, 2xInter constructs a Structural Causal Model (SCM) to explain instance and prompt biases and designs a Double-View Causal Intervention (DVCI) module to eliminate these biases. Due to limited supervised information, DVCI devises a generation-based prompt adjustment for instance intervention and a Siamese network-based instance contrasting for prompt intervention. Finally, the experimental results show that 2xInter achieves state-of-the-art performance on RAMS and ACE datasets.

Abstractive summarization models learned with maximum likelihood estimation (MLE) have long been guilty of generating unfaithful facts alongside ambiguous focus. Improved paradigm under the guidance of reference-identified words, i.e., guided summarization, has exhibited remarkable advantages in overcoming this problem. However, it suffers limited real applications since the prophetic guidance is practically agnostic at inference. In this paper, we introduce a novel teacher-student framework, which learns a regular summarization model to mimic the behavior of being guided by prophecy for boosting abstractive summaries. Specifically, by training in probability spaces to follow and distinguish a guided teacher model, a student model learns the key to generating teacher-like quality summaries without any guidance. We refer to this process as prophecy distillation, and it breaks the limitations of both standard and guided summarization. Through extensive experiments, we show that our method achieves new or matched state-of-the-art on four well-known datasets, including ROUGE scores, faithfulness, and saliency awareness. Human evaluations are also carried out to evidence these merits. Furthermore, we conduct empirical studies to analyze how the hyperparameters setting and the guidance choice affect TPG performance.

Although language models (LMs) have boosted the performance of Question Answering, they still need plenty of data. Data annotation, in contrast, is a time-consuming process. This especially applies to Question Answering, where possibly large documents have to be parsed and annotated with questions and their corresponding answers. Furthermore, Question Answering models often only work well for the domain they were trained on. Since annotation is costly, we argue that domain-agnostic knowledge from LMs, such as linguistic understanding, is sufficient to create a well-curated dataset. With this motivation, we show that using large language models can improve Question Answering performance on various datasets in the few-shot setting compared to state-of-the-art approaches. For this, we perform data generation leveraging the Prompting framework, suggesting that language models contain valuable task-agnostic knowledge that can be used beyond the common pre-training/fine-tuning scheme. As a result, we consistently outperform previous approaches on few-shot Question Answering.

Large language models (LLMs) have been applied to a wide range of data-to-text generation tasks, including tables, graphs, and time-series numerical data-to-text settings. While research on generating prompts for structured data such as tables and graphs is gaining momentum, in-depth investigations into prompting for time-series numerical data are lacking. Therefore, this study explores various input representations, including sequences of tokens and structured formats such as HTML, LaTeX, and Python-style codes. In our experiments, we focus on the task of Market Comment Generation, which involves taking a numerical sequence of stock prices as input and generating a corresponding market comment. Contrary to our expectations, the results show that prompts resembling programming languages yield better outcomes, whereas those similar to natural languages and longer formats, such as HTML and LaTeX, are less effective. Our findings offer insights into creating effective prompts for tasks that generate text from numerical sequences.