COLING 2024 Papers

Interviews are an effective method to elicit critical skills to perform particular processes in various domains. In order to understand the knowledge structure of these domain-specific processes, we consider semantic role and predicate annotation based on Frame Semantics. We introduce a dataset of interview dialogues with experts in the culinary and gardening domains, each annotated with semantic frames. This dataset consists of (1) 308 interview dialogues related to the culinary domain, originally assembled by Okahisa et al. (2022), and (2) 100 interview dialogues associated with the gardening domain, which we newly acquired. The labeling specifications take into account the domain-transferability by adopting domain-agnostic labels for frame elements. In addition, we conducted domain transfer experiments from the culinary domain to the gardening domain to examine the domain transferability with our dataset. The experimental results showed the effectiveness of our domain-agnostic labeling scheme.

Domain adaption has been widely adapted for cross-domain sentiment analysis to transfer knowledge from the source domain to the target domain. Whereas, most methods are proposed under the assumption that the target (test) domain is known, making them fail to generalize well on unknown test data that is not always available in practice. In this paper, we focus on the problem of domain generalization for cross-domain sentiment analysis. Specifically, we propose a backdoor adjustment-based causal model to disentangle the domain-specific and domain-invariant representations that play essential roles in tackling domain shift. First, we rethink the cross-domain sentiment analysis task in a causal view to model the causal-and-effect relationships among different variables. Then, to learn an invariant feature representation, we remove the effect of domain confounders (e.g., domain knowledge) using the backdoor adjustment. A series of experiments over many homologous and diverse datasets show the great performance and robustness of our model by comparing it with the state-of-the-art domain generalization baselines.

Recent studies have demonstrated that the ability of dense retrieval models to generalize to target domains with different distributions is limited, which contrasts with the results obtained with interaction-based models. Prior attempts to mitigate this challenge involved leveraging adversarial learning and query generation approaches, but both approaches nevertheless resulted in limited improvements. In this paper, we propose to combine the query-generation approach with a self-supervision approach in which pseudo-relevance labels are automatically generated on the target domain. To accomplish this, a T5-3B model is utilized for pseudo-positive labeling, and meticulous hard negatives are chosen. We also apply this strategy on conversational dense retrieval model for conversational search. A similar pseudo-labeling approach is used, but with the addition of a query-rewriting module to rewrite conversational queries for subsequent labeling. This proposed approach enables a model’s domain adaptation with real queries and documents from the target dataset. Experiments on standard dense retrieval and conversational dense retrieval models both demonstrate improvements on baseline models when they are fine-tuned on the pseudo-relevance labeled data.

Curriculum Learning (CL) has been emerged as an effective technique for improving the performances and reducing the cost of pre-training Large Language Models (LLMs). The efficacy of CL demonstrated in different scenarios is in the training LLMs by organizing examples from the simplest to the most complex. Although improvements have been shown extensively, this approach was used for pre-training, leaving novel fine-tuning approaches such as instruction-tuning unexplored. In this paper, we propose a novel complexity measure to empower the instruction-tuning method using the CL paradigm. To complement previous works, we propose cognitively motivated measures to determine the complexity of training demonstrations used in the instruction-tuning paradigm. Hence, we experiment with the proposed heuristics first in English and then in other languages. The downstream results show that delivering training examples by complexity ranking is also effective for instruction tuning, as it improves downstream performance while reducing costs. Furthermore, the technique can be easily transferred to languages other than English, e.g., Italian and French, without any adaptation, maintaining functionality and effectiveness.

he present study introduces the knowledge-augmented generator, which is specifically designed to produce information that remains grounded in contextual knowledge, regardless of alterations in the context. Previous research has predominantly focused on examining hallucinations stemming from static input, such as in the domains of summarization or machine translation. However, our investigation delves into the faithfulness of generative question answering in the presence of dynamic knowledge. Our objective is to explore the existence of hallucinations arising from parametric memory when contextual knowledge undergoes changes, while also analyzing the underlying causes for their occurrence. In order to efficiently address this issue, we propose a straightforward yet effective measure for detecting such hallucinations. Intriguingly, our investigation uncovers that all models exhibit a tendency to generate previous answers as hallucinations. To gain deeper insights into the underlying causes of this phenomenon, we conduct a series of experiments that verify the critical role played by context in hallucination, both during training and testing, from various perspectives.

Recently, ChatGPT has demonstrated remarkable performance in various downstream tasks such as open-domain question answering, machine translation, and code generation. As a general-purpose task solver, an intriguing inquiry arises: Does ChatGPT itself know that it does not know, without any access to internal states? In response to this query, we present an initial evaluation of ChatGPT for black-box calibration. We designed three types of proxy confidence, from three perspectives to assess its performance. Experiments are conducted on five datasets, spanning four tasks, and the results show that ChatGPT has a degree of capability for black-box calibration. Specifically, proxy confidence displayed a significantly positive Pearson correlation (95.16%) with accuracy in the TruthfulQA dataset, while revealing a negative correlation in the ModAr dataset. We delved deeper into ChatGPT’s black-box calibration ability by examining failure cases in the ModAr dataset. Our analysis revealed that ChatGPT’s tendency to exhibit overconfidence may stem from its reliance on semantic priors. Furthermore, we investigated why ChatGPT performs relatively well in TruthfulQA. The findings suggest that ChatGPT might implicitly acquire calibration skills during the reinforcement learning process, rather than relying solely on simplistic heuristics.

Document-level Event Extraction (DEE) is a vital task in NLP as it seeks to automatically recognize and extract event information from a document. However, current approaches often overlook intricate relationships among events and subtle correlations among arguments within a document, which can significantly impact the effectiveness of event type recognition and the extraction of cross-sentence arguments in DEE task. This paper proposes a novel Correlation Association Interactive Network (CAINet), comprising two key components: event relationship graph and argument correlation graph. In particular, the event relationship graph models the relationship among various events through structural associations among event nodes and sentence nodes, to improve the accuracy of event recognition. On the other hand, the arguments correlation graph models the correlations among arguments by quantifying the strength of association among arguments, to effectively aggregate cross-sentence arguments, contributing to the overall success of DEE. Furthermore, we use the large language model to execute DEE task experiments. Experimental results show the proposed CAINet outperforms existing state-of-the-art models and large language models in terms of F1-score across two benchmark datasets.

The Document Set Expansion (DSE) task involves identifying relevant documents from large collections based on a limited set of example documents. Previous research has highlighted Positive and Unlabeled (PU) learning as a promising approach for this task. However, most PU methods rely on the unrealistic assumption of knowing the class prior for positive samples in the collection. To address this limitation, this paper introduces a novel PU learning framework that utilizes intractable density estimation models. Experiments conducted on PubMed and Covid datasets in a transductive setting showcase the effectiveness of the proposed method for DSE. Code is available from https://github.com/Beautifuldog01/Document-set-expansion-puDE.

We present a novel task of document-level script event prediction, which aims to predict the next event given a candidate list of narrative events in long-form documents. To enable this, we introduce DocSEP, a challenging dataset in two new domains - contractual documents and Wikipedia articles, where timeline events may be paragraphs apart and may require multi-hop temporal and causal reasoning. We benchmark existing baselines and present a novel architecture called DocScript to learn sequential ordering between events at the document scale. Our experimental results on the DocSEP dataset demonstrate that learning longer-range dependencies between events is a key challenge and show that contemporary LLMs such as ChatGPT and FlanT5 struggle to solve this task, indicating their lack of reasoning abilities for understanding causal relationships and temporal sequences within long texts.

Table-text document (e.g., financial reports) understanding has attracted increasing attention in recent two years. TAT-DQA is a realistic setting for the understanding of visually-rich table-text documents, which involves answering associated questions requiring discrete reasoning. Most existing work relies on token-level semantics, falling short in the reasoning across document elements such as quantities and dates. To address this limitation, we propose a novel Doc2SoarGraph model that exploits element-level semantics and employs Semantic-oriented hierarchical Graph structures to capture the differences and correlations among different elements within the given document and question. Extensive experiments on the TAT-DQA dataset reveal that our model surpasses the state-of-the-art conventional method (i.e., MHST) and large language model (i.e., ChatGPT) by 17.73 and 6.49 points respectively in terms of Exact Match (EM) metric, demonstrating exceptional effectiveness.

While compositional interpretation is the core of language understanding, humans also derive meaning via inference. For example, while the phrase “the blue hat” introduces a blue hat into the discourse via the direct composition of “blue” and “hat,” the same discourse entity is introduced by the phrase “the blue color of this hat” despite the absence of any local composition between “blue” and “hat.” Instead, we infer that if the color is blue and it belongs to the hat, the hat must be blue. We tested the performance of neural language models and humans on such inferentially driven conceptual compositions, eliciting probability estimates for a noun in a minimally composed phrase, “This blue hat”, following contexts that had introduced the conceptual combinations of those nouns and adjectives either syntactically or inferentially. Surprisingly, our findings reveal significant disparities between the performance of neural language models and human judgments. Among the eight models evaluated, RoBERTa, BERT-large, and GPT-2 exhibited the closest resemblance to human responses, while other models faced challenges in accurately identifying compositions in the provided contexts. Our study reveals that language models and humans may rely on different approaches to represent and compose lexical items across sentence structure. All data and code are accessible at https://github.com/wangshaonan/BlueHat.

Gender bias has been widely studied by the NLP community. However, other more subtle variations of it, such as mansplaining, have yet received little attention. Mansplaining is a discriminatory behaviour that consists of a condescending treatment or discourse towards women. In this paper, we introduce and analyze Well, actually..., a corpus of 886 mansplaining stories experienced by women. We analyze the corpus in terms of features such as offensiveness, sentiment or misogyny, among others. We also explore to what extent Large Language Models (LLMs) can understand and identify mansplaining and other gender-related microaggressions. Specifically, we experiment with ChatGPT-3.5-Turbo and LLaMA-2 (13b and 70b), with both targeted and open questions. Our findings suggest that, although they can identify mansplaining to some extent, LLMs still struggle to point out this attitude and will even reproduce some of the social patterns behind mansplaining situations, for instance by praising men for giving unsolicited advice to women.

Large, curated, web-crawled corpora play a vital role in training language models (LMs). They form the lion’s share of the training data in virtually all recent LMs, such as the well-known GPT, LLaMA and XLM-RoBERTa models. However, despite this importance, relatively little attention has been given to the quality of these corpora. In this paper, we compare four of the currently most relevant large, web-crawled corpora (CC100, MaCoCu, mC4 and OSCAR) across eleven lower-resourced European languages. Our approach is two-fold: first, we perform an intrinsic evaluation by performing a human evaluation of the quality of samples taken from different corpora; then, we assess the practical impact of the qualitative differences by training specific LMs on each of the corpora and evaluating their performance on downstream tasks. We find that there are clear differences in quality of the corpora, with MaCoCu and OSCAR obtaining the best results. However, during the extrinsic evaluation, we actually find that the CC100 corpus achieves the highest scores. We conclude that, in our experiments, the quality of the web-crawled corpora does not seem to play a significant role when training LMs.

Despite the superior performance, Large Language Models (LLMs) require significant computational resources for deployment and use. To overcome this issue, quantization methods have been widely applied to reduce the memory footprint of LLMs as well as increase the inference rate. However, a major challenge is that low-bit quantization methods often lead to performance degradation. It is important to understand how quantization impacts the capacity of LLMs. Different from previous studies focused on overall performance, this work aims to investigate the impact of quantization on emergent abilities, which are important characteristics that distinguish LLMs from small language models. Specifically, we examine the abilities of in-context learning, chain-of-thought reasoning, and instruction-following in quantized LLMs. Our empirical experiments show that these emergent abilities still exist in 4-bit quantization models, while 2-bit models encounter severe performance degradation on the test of these abilities. To improve the performance of low-bit models, we conduct two special experiments: (1) fine-gained impact analysis that studies which components (or substructures) are more sensitive to quantization, and (2) performance compensation through model fine-tuning. Our work derives a series of important findings to understand the impact of quantization on emergent abilities and sheds light on the possibilities of extremely low-bit quantization for LLMs.

Previous methods on knowledge base question generation (KBQG) primarily focus on refining the quality of a single generated question. However, considering the remarkable paraphrasing ability of humans, we believe that diverse texts can express identical semantics through varied expressions. The above insights make diversifying question generation an intriguing task, where the first challenge is evaluation metrics for diversity. Current metrics inadequately assess the aforementioned diversity. They calculate the ratio of unique n-grams in the generated question, which tends to measure duplication rather than true diversity. Accordingly, we devise a new diversity evaluation metric, which measures the diversity among top-k generated questions for each instance while ensuring their relevance to the ground truth. Clearly, the second challenge is how to enhance diversifying question generation. To address this challenge, we introduce a dual model framework interwoven by two selection strategies to generate diverse questions leveraging external natural questions. The main idea of our dual framework is to extract more diverse expressions and integrate them into the generation model to enhance diversifying question generation. Extensive experiments on widely used benchmarks for KBQG show that our approach can outperform pre-trained language model baselines and text-davinci-003 in diversity while achieving comparable performance with ChatGPT.

Traditional automated metrics for evaluating conditional natural language generation rely on pairwise comparisons between a single generated text and the best-matching gold-standard reference. This method is effective when ground truth data diversity can be attributed to noise, however, it falls short when diversity in references holds valuable contextual information, as in visual description or summarization, as it does not evaluate the ability of a model to generate text matching the diversity of the ground truth samples. In this paper, we challenge the adequacy of existing metrics in such semantically diverse contexts and introduce a novel approach for evaluating conditional language generation models, leveraging a family of meta-metrics that build on existing pairwise distance functions. These meta-metrics assess not just single-samples, but distributions of reference and model-generated captions using small sample sets. We demonstrate our approach through a case study of visual description in the English language which reveals not only how current models prioritize single-description quality over diversity, but further sheds light on the impact of sampling methods and temperature settings on description quality and diversity.

Multiple Choice Questions (MCQs) are very common in both high-stakes and low-stakes examinations, and their effectiveness in assessing students relies on the quality and diversity of distractors, which are the incorrect answer options provided alongside the correct answer. Motivated by the progress in generative language models, we propose a two-step automatic distractor generation approach which is based on text to text transfer transformer models. Unlike most of previous methods for distractor generation, our approach does not rely on the correct answer options. Instead, it first generates both correct and incorrect answer options, and then discriminates potential correct options from distractors. Identified distractors are finally categorised based on semantic similarity scores into separate clusters, and the cluster heads are selected as our final distinct distractors. Experiments on two publicly available datasets show that our approach outperforms previous models both in the case of single-word answer options and longer-sequence reading comprehension questions.

Continual Few-Shot Relation Learning (CFRL) aims to learn an increasing number of new relational patterns from a data stream. However, due to the limited number of samples and the continual training mode, this method frequently encounters the catastrophic forgetting issues. The research on causal inference suggests that this issue is caused by the loss of causal effects from old data during the new training process. Inspired by the causal graph, we propose a unified causal framework for CFRL to restore the causal effects. Specifically, we establish two additional causal paths from old data to predictions by having the new data and memory data collide with old data separately in the old feature space. This augmentation allows us to preserve causal effects effectively and enhance the utilization of valuable information within memory data, thereby alleviating the phenomenon of catastrophic forgetting. Furthermore, we introduce a self-adaptive weight to achieve a delicate balance of causal effects between the new and old relation types. Extensive experiments demonstrate the superiority of our method over existing state-of-the-art approaches in CFRL task settings. Our codes are publicly available at: https://github.com/ywh140/CECF.

Free-text explanations are crucial for enhancing the interpretability of AI models. However, training models to generate high-quality free-text explanations is challenging, primarily due to the requirement of a substantial amount of human-written explanations, which can be expensive. Recently, Large language models (LLMs) like ChatGPT and GPT-4 have made remarkable progress in various NLP tasks while also providing explanations alongside their answers. Leveraging LLMs for data labeling offers a more cost-effective alternative. However, a key concern arises from the fact that the answers provided by LLMs are not entirely accurate, potentially introducing noise to both task outputs and explanation generation. To remedy this, we propose a new mechanism, Distillation with Explanations from LLMs. we observe that despite the incorrectness in LLMs-generated answers, their explanations are consistent with their answers. Leveraging this consistency, our method combines the ground truth labels and answers-explanations generated by LLMs, to simultaneously generate more accurate answers and the corresponding free-text explanations. Experimental results demonstrate that our approach achieves improved predictive performance and also generates explanations that exhibit greater alignment with the model’s task outputs.