COLING 2024 论文列表

Adversarial examples, capable of misleading machine learning models into making erroneous predictions, pose significant risks in safety-critical domains such as crisis informatics, medicine, and autonomous driving. To counter this, we introduce a novel textual adversarial example method that identifies falsely learned word indicators by leveraging explainable AI methods as importance functions on incorrectly predicted instances, thus revealing and understanding the weaknesses of a model. To evaluate the effectiveness of our approach, we conduct a human and a transfer evaluation and propose a novel adversarial training evaluation setting for better robustness assessment. While outperforming current adversarial example and training methods, the results also show our method’s potential in facilitating the development of more resilient transformer models by detecting and rectifying biases and patterns in training data, showing baseline improvements of up to 23 percentage points in accuracy on adversarial tasks. The code of our approach is freely available for further exploration and use.

Persuasive dialogue is a type of dialogue commonly used in human daily life in scenarios such as promotion and sales. Its purpose is to influence the decision, attitude or behavior of another person through the dialogue process. Persuasive automated dialogue systems can be applied in a variety of fields such as charity, business, education, and healthcare. Regardless of their amazing abilities, Large Language Models (LLMs) such as ChatGPT still have limitations in persuasion. There is few research dedicated to persuasive dialogue in the current research of automated dialogue systems. In this paper, we introduce a persuasive automated dialogue system. In the system, a context-aware persuasion strategy selection module makes dialogue system flexibly use different persuasion strategies to persuade users; Then a natural language generation module is used to output a response. We also propose a persuasiveness prediction model to automatically evaluate the persuasiveness of generated text. Experimental results show that our dialogue system can achieve better performance on several automated evaluation metrics than baseline models.

The awareness of multi-cultural human values is critical to the ability of language models (LMs) to generate safe and personalized responses. However, this awareness of LMs has been insufficiently studied, since the computer science community lacks access to the large-scale real-world data about multi-cultural values. In this paper, we present WorldValuesBench, a globally diverse, large-scale benchmark dataset for the multi-cultural value prediction task, which requires a model to generate a rating response to a value question based on demographic contexts. Our dataset is derived from an influential social science project, World Values Survey (WVS), that has collected answers to hundreds of value questions (e.g., social, economic, ethical) from 94,728 participants worldwide. We have constructed more than 20 million examples of the type "(demographic attributes, value question) → answer” from the WVS responses. We perform a case study using our dataset and show that the task is challenging for strong open and closed-source models. On merely 11.1%, 25.0%, 72.2%, and 75.0% of the questions, Alpaca-7B, Vicuna-7B-v1.5, Mixtral-8x7B-Instruct-v0.1, and GPT-3.5 Turbo can respectively achieve <0.2 Wasserstein 1-distance from the human normalized answer distributions. WorldValuesBench opens up new research avenues in studying limitations and opportunities in multi-cultural value awareness of LMs.

Dictionary definitions play a prominent role in a wide range of NLP tasks, for instance by providing additional context about the meaning of rare and emerging terms. Many dictionaries also provide examples to illustrate the prototypical usage of words, which brings further opportunities for training or enriching NLP models. The intrinsic qualities of dictionaries, and related lexical resources such as glossaries and encyclopedias, are however still not well-understood. While there has been significant work on developing best practices, such guidance has been aimed at traditional usages of dictionaries (e.g. supporting language learners), and it is currently unclear how different quality aspects affect the NLP systems that rely on them. To address this issue, we compare WordNet, the most commonly used lexical resource in NLP, with a variety of dictionaries, as well as with examples that were generated by ChatGPT. Our analysis involves human judgments as well as automatic metrics. We furthermore study the quality of word embeddings derived from dictionary examples, as a proxy for downstream performance. We find that WordNet’s examples lead to lower-quality embeddings than those from the Oxford dictionary. Surprisingly, however, the ChatGPT generated examples were found to be most effective overall.

Event Detection (ED) is a task of automatically extracting multi-class trigger words. The understanding of word sense is crucial for ED. In this paper, we utilize context-specific commonsense knowledge to strengthen word sense modeling. Specifically, we leverage a Context-specific Knowledge Selector (CKS) to select the exact commonsense knowledge of words from a large knowledge base, i.e., ConceptNet. Context-specific selection is made in terms of the relevance of knowledge to the living contexts. On this basis, we incorporate the commonsense knowledge into the word-level representations before decoding. ChatGPT is an ideal generative CKS when the prompts are deliberately designed, though it is cost-prohibitive. To avoid the heavy reliance on ChatGPT, we train an offline CKS using the predictions of ChatGPT over a small number of examples (about 9% of all). We experiment on the benchmark ACE-2005 dataset. The test results show that our approach yields substantial improvements compared to the BERT baseline, achieving the F1-score of about 78.3%. All models, source codes and data will be made publicly available.

Multimodal learning is generally expected to make more accurate predictions than text-only analysis. Here, although various methods for fusing multimodal inputs have been proposed for sentiment analysis tasks, we found that they may be inhibiting their fusion methods, which are based on attention-based language models, from learning non-verbal modalities, because non-verbal ones are isolated from the linguistic semantics and contexts and do not include them, meaning that they are unsuitable for applying attention to text modalities during the fusion phase. To address this issue, we propose Word-aware Modality Stimulation Fusion (WA-MSF) for facilitating integration of non-verbal modalities with the text modality. The Modality Stimulation Unit layer (MSU-layer) is the core concept of WA-MSF; it integrates language contexts and semantics into non-verbal modalities, thereby instilling linguistic essence into these modalities. Moreover, WA-MSF uses aMLP in the fusion phase in order to utilize spatial and temporal representations of non-verbal modalities more effectively than transformer fusion. In our experiments, WA-MSF set a new state-of-the-art level of performance on sentiment prediction tasks.

Fine-tuning Pre-trained Language Models (PLMs) is a popular Natural Language Processing (NLP) paradigm for addressing Named Entity Recognition (NER) tasks. However, neural network models often demonstrate poor generalization capabilities due to significant disparities between the knowledge learned by PLMs and the distribution of the target dataset, as well as data scarcity issues. In addition, token omission in predictions due to insufficient learning remains a challenge in NER. In this paper, we propose a kNN retrieval enhancement algorithm (WkNER) that incorporates word segmentation information to enhance the model’s generalization ability and alleviate the problem of missing entity tokens in prediction. The introduction of word segmentation information is used to preliminarily determine the boundaries of entities and alleviate the common prediction errors of missing tokens within entities made by the fine-tuned model. Secondly, we find that non-entities in the retrieval table contain a large amount of redundant information, and explore the effects of introducing non-entity information of different scales on the model. Experimental results show that our proposed method significantly improves the performance of baseline models, and achieves better or compared recognition accuracy than previous state-of-the-art models in multiple public Chinese and English datasets. Especially in low-resource scenarios, our method achieves higher accuracy on 20% of the dataset than the original method on the full dataset.

We present a comprehensive computational study of the under-investigated phenomenon of personal name compounds (PNCs) in German such as Willkommens-Merkel (‘Welcome-Merkel’). Prevalent in news, social media, and political discourse, PNCs are hypothesized to exhibit an evaluative function that is reflected in a more positive or negative perception as compared to the respective personal full name (such as Angela Merkel). We model 321 PNCs and their corresponding full names at discourse level, and show that PNCs bear an evaluative nature that can be captured through a variety of computational methods. Specifically, we assess through valence information whether a PNC is more positively or negatively evaluative than the person’s name, by applying and comparing two approaches using (i) valence norms and (ii) pre-trained language models (PLMs). We further enrich our data with personal, domain-specific, and extra-linguistic information and perform a range of regression analyses revealing that factors including compound and modifier valence, domain, and political party membership influence how a PNC is evaluated.

The task of Split and Rephrase, which splits a complex sentence into multiple simple sentences with the same meaning, improves readability and enhances the performance of downstream tasks in natural language processing (NLP). However, while Split and Rephrase can be improved using a text-to-text generation approach that applies encoder-decoder models fine-tuned with a large-scale dataset, it still suffers from hallucinations and under-splitting. To address these issues, this paper presents a simple and strong data refinement approach. Here, we create WikiSplit++ by removing instances in WikiSplit where complex sentences do not entail at least one of the simpler sentences and reversing the order of reference simple sentences. Experimental results show that training with WikiSplit++ leads to better performance than training with WikiSplit, even with fewer training instances. In particular, our approach yields significant gains in the number of splits and the entailment ratio, a proxy for measuring hallucinations.

The factuality of large language model (LLMs) tends to decay over time since events posterior to their training are “unknown” to them. One way to keep models up-to-date could be factual update: the task of inserting, replacing, or removing certain simple (atomic) facts within the model. To study this task, we present WikiFactDiff, a dataset that describes the evolution of factual knowledge between two dates as a collection of simple facts divided into three categories: new, obsolete, and static. We describe several update scenarios arising from various combinations of these three types of basic update. The facts are represented by subject-relation-object triples; indeed, WikiFactDiff was constructed by comparing the state of the Wikidata knowledge base at 4 January 2021 and 27 February 2023. Those fact are accompanied by verbalization templates and cloze tests that enable running update algorithms and their evaluation metrics. Contrary to other datasets, such as zsRE and CounterFact, WikiFactDiff constitutes a realistic update setting that involves various update scenarios, including replacements, archival, and new entity insertions. We also present an evaluation of existing update algorithms on WikiFactDiff.

Given the high prevalence of mental disorders and the significant diagnostic delays and difficulties in patient follow-up, voice biomarkers hold the promise of improving access to care and therapeutic follow-up for people with psychiatric disorders. Yet, despite many years of successful research in the field, none of these voice biomarkers are implemented in clinical practice. Beyond the reductive explanation of the lack of explainability of the involved machine learning systems, we look for arguments in the epistemology and sociology of psychiatry. We show that the estimation of diagnoses, the major task in the literature, is of little interest to both clinicians and patients. After tackling the common misbeliefs about diagnosis in psychiatry in a didactic way, we propose a paradigm shift towards the estimation of clinical symptoms and signs, which not only address the limitations raised against diagnosis estimation but also enable the formulation of new machine learning tasks. We hope that this paradigm shift will empower the use of vocal biomarkers in clinical practice. It is however conditional on a change in database labeling practices, but also on a profound change in the speech processing community’s practices towards psychiatry.

The task of quote attribution seeks to pair textual utterances with the name of their speakers. Despite continuing research efforts on the task, models are rarely evaluated systematically against previous models in comparable settings on the same datasets. This has resulted in a poor understanding of the relative strengths and weaknesses of various approaches. In this work we formalize the task of quote attribution, and in doing so, establish a basis of comparison across existing models. We present an exhaustive benchmark of known models, including natural extensions to larger LLM base models, on all available datasets in both English and Chinese. Our benchmarking results reveal that the CEQA model attains state-of-the-art performance among all supervised methods, and ChatGPT, operating in a four-shot setting, demonstrates performance on par with or surpassing that of supervised methods on some datasets. Detailed error analysis identify several key factors contributing to prediction errors.

Bragging is the act of uttering statements that are likely to be positively viewed by others and it is extensively employed in human communication with the aim to build a positive self-image of oneself. Social media is a natural platform for users to employ bragging in order to gain admiration, respect, attention and followers from their audiences. Yet, little is known about the scale of bragging online and its characteristics. This paper employs computational sociolinguistics methods to conduct the first large scale study of bragging behavior on Twitter (U.S.) by focusing on its overall prevalence, temporal dynamics and impact of demographic factors. Our study shows that the prevalence of bragging decreases over time within the same population of users. In addition, younger, more educated and popular users in the U.S. are more likely to brag. Finally, we conduct an extensive linguistics analysis to unveil specific bragging themes associated with different user traits.

This paper introduces the corpus for the novel task of presupposition generation - a natural language generation problem where a model produces a list of presuppositions carried by the given input sentence, in the context of the presented research - given the cross-examination question. Two datasets, PECaN (Presupposition, Entailment, Contradiction and Neutral) and PGen (Presuppostion Generation), are designed to fine-tune existing BERT (CITATION) and T5 (CITATION) models for classification and generation tasks. Various corpora construction methods are proposed ranging from manual annotations, prompting the GPT 3.0 model, to augmenting data from the existing corpora. The fine-tuned models achieved high accuracy on the novel Presupposition as Natural Language Inference (PNLI) task which extends the traditional Natural Language Inference (NLI) incorporating instances of presupposition into classification. T5 outperforms BERT by broad margin achieving an overall accuracy of 84.35% compared to 71.85% of BERT, and specifically when classifying presuppositions (93% vs 73% respectively). Regarding presupposition generation, we observed that despite the limited amount of data used for fine-tuning, the model displays an emerging proficiency in generation presuppositions reaching ROUGE scores of 43.47, adhering to systematic patterns that mirror valid strategies for presupposition generation, although failed to generate the complete lists.

Decoding semantic meanings from brain activity has attracted increasing attention. Neurolinguists have found that semantic perception is open to multisensory stimulation, as word meanings can be delivered by both auditory and visual inputs. Prior work which decodes semantic meanings from neuroimaging data largely exploits brain activation patterns triggered by stimulation in cross-modality (i.e. text-audio pairs, text-picture pairs). Their goal is to develop a more sophisticated computational model to probing what information from the act of language understanding is represented in human brain. While how the brain receiving such information influences decoding performance is underestimated. This study dissociates multisensory integration of word understanding into written text, spoken text and image perception respectively, exploring the decoding efficiency and reliability of unisensory information in the brain representation. The findings suggest that, in terms of unisensory, decoding is most successful when semantics is represented in pictures, but the effect disappears in the case of congeneric words which share a related meaning. These results reveal the modality dependence and multisensory enhancement in the brain decoding methodology.

Most existing approaches for unsupervised bilingual lexicon induction (BLI) depend on good quality static or contextual embeddings requiring large monolingual corpora for both languages. However, unsupervised BLI is most likely to be useful for low-resource languages (LRLs), where large datasets are not available. Often we are interested in building bilingual resources for LRLs against related high-resource languages (HRLs), resulting in severely imbalanced data settings for BLI. We first show that state-of-the-art BLI methods in the literature exhibit near-zero performance for severely data-imbalanced language pairs, indicating that these settings require more robust techniques. We then present a new method for unsupervised BLI between a related LRL and HRL that only requires inference on a masked language model of the HRL, and demonstrate its effectiveness on truly low-resource languages Bhojpuri and Magahi (with <5M monolingual tokens each), against Hindi. We further present experiments on (mid-resource) Marathi and Nepali to compare approach performances by resource range, and release our resulting lexicons for five low-resource Indic languages: Bhojpuri, Magahi, Awadhi, Braj, and Maithili, against Hindi.

Sarcasm recognition is challenging because it needs an understanding of the true intention, which is opposite to or different from the literal meaning of the words. Prior work has addressed this challenge by developing a series of methods that provide richer contexts, e.g., sentiment or cultural nuances, to models. While shown to be effective individually, no study has systematically evaluated their collective effectiveness. As a result, it remains unclear to what extent additional contexts can improve sarcasm recognition. In this work, we explore the improvements that existing methods bring by incorporating more contexts into a model. To this end, we develop a framework where we can integrate multiple contextual cues and test different approaches. In evaluation with four approaches on three sarcasm recognition benchmarks, we achieve existing state-of-the-art performances and also demonstrate the benefits of sequentially adding more contexts. We also identify inherent drawbacks of using more contexts, highlighting that in the pursuit of even better results, the model may need to adopt societal biases.

In this paper we propose a new framework and new methods for the reference-free evaluation of topic segmentation systems directly in the embedding space. Specifically, we define a common framework for reference-free, embedding-based topic segmentation metrics, and show how this applies to an existing metric. We then define new metrics, based on a previously defined cohesion score, Average Relative Proximity. Using this approach, we show that Large Language Models (LLMs) yield features that, if used correctly, can strongly correlate with traditional topic segmentation metrics based on costly and rare human annotations, while outperforming existing reference-free metrics borrowed from clustering evaluation in most domains. We then show that smaller language models specifically fine-tuned for different sentence-level tasks can outperform LLMs several orders of magnitude larger. Via a thorough comparison of our metric’s performance across different datasets, we see that conversational data present the biggest challenge in this framework. Finally, we analyse the behaviour of our metrics in specific error cases, such as those of under-generation and moving of ground truth topic boundaries, and show that our metrics behave more consistently than other reference-free methods.

In this paper, we investigate the role of arguments in the automatic scoring of cohesion in argumentative essays. The feature analysis reveals that in argumentative essays, the lexical cohesion between claims is more important to the overall cohesion, while the evidence is expected to be diverse and divergent. Our results show that combining features related to argument segments and cohesion features improves the performance of the automatic cohesion scoring model trained on a transformer. The cohesion score is also learned more accurately in a multi-task learning process by adding the automatic segmentation of argumentative elements as an auxiliary task. Our findings contribute to both the understanding of cohesion in argumentative writing and the development of automatic feedback.