COLING 2024 论文列表

In this paper, we present AnCast, an intuitive and efficient tool for evaluating graph-based meaning representations (MR). AnCast implements evaluation metrics that are well understood in the NLP community, and they include concept F1, unlabeled relation F1, labeled relation F1, and weighted relation F1. The efficiency of the tool comes from a novel anchor broadcast alignment algorithm that is not subject to the trappings of local maxima. We show through experimental results that the AnCast score is highly correlated with the widely used Smatch score, but its computation takes only about 40% the time.

We empirically study the ability of a Large Language Model (gpt-3.5-turbo-instruct) to understand morphologically complex words. In our experiments, we looked at a variety of tasks to analyse German compounds with regard to compositional word formation and derivation, such as identifying the head noun of existing and novel compounds, identifying the shared verb stem between two words, or recognizing words constructed with inappropriately used derivation morphemes as invalid. Our results show that the language model is generally capable of solving most tasks, except for the task of identifying ill-formed word forms. While the model demonstrated a good overall understanding of complex words and their word-internal structure, the results also suggest that there is no formal knowledge of derivational rules, but rather an interpretation of the observed word parts to derive the meaning of a word.

Large language models (LLMs) such as Llama 2 perform very well on tasks that involve both natural language and source code, particularly code summarization and code generation. We show that for the task of code summarization, the performance of these models on individual examples often depends on the amount of (subword) token overlap between the code and the corresponding reference natural language descriptions in the dataset. This token overlap arises because the reference descriptions in standard datasets (corresponding to docstrings in large code bases) are often highly similar to the names of the functions they describe. We also show that this token overlap occurs largely in the function names of the code and compare the relative performance of these models after removing function names versus removing code structure. We also show that using multiple evaluation metrics like BLEU and BERTScore gives us very little additional insight since these metrics are highly correlated with each other.

The discourse surrounding climate change on social media platforms has emerged as a significant avenue for understanding public sentiments, perspectives, and engagement with this critical global issue. The unavailability of publicly available datasets, coupled with ignoring the multi-aspect analysis of climate discourse on social media platforms, has underscored the necessity for further advancement in this area. To address this gap, in this paper, we present an extensive exploration of the intricate realm of climate change discourse on Twitter, leveraging a meticulously annotated ClimaConvo dataset comprising 15,309 tweets. Our annotations encompass a rich spectrum, including aspects like relevance, stance, hate speech, the direction of hate, and humor, offering a nuanced understanding of the discourse dynamics. We address the challenges inherent in dissecting online climate discussions and detail our comprehensive annotation methodology. In addition to annotations, we conduct benchmarking assessments across various algorithms for six tasks: relevance detection, stance detection, hate speech identification, direction and target, and humor analysis. This assessment enhances our grasp of sentiment fluctuations and linguistic subtleties within the discourse. Our analysis extends to exploratory data examination, unveiling tweet distribution patterns, stance prevalence, and hate speech trends. Employing sophisticated topic modeling techniques uncovers underlying thematic clusters, providing insights into the diverse narrative threads woven within the discourse. The findings present a valuable resource for researchers, policymakers, and communicators seeking to navigate the intricacies of climate change discussions. The dataset and resources for this paper are available at https://github.com/shucoll/ClimaConvo.

The ever-growing number of people suffering from mental distress has motivated significant research initiatives towards automated depression estimation. Despite the multidisciplinary nature of the task, very few of these approaches include medical professionals in their research process, thus ignoring a vital source of domain knowledge. In this paper, we propose to bring the domain experts back into the loop and incorporate their knowledge within the gold-standard DAIC-WOZ dataset. In particular, we define a novel transformer-based architecture and analyse its performance in light of our expert annotations. Overall findings demonstrate a strong correlation between the psychological tendencies of medical professionals and the behavior of the proposed model, which additionally provides new state-of-the-art results.

Recent advances in large language models (LLMs) have enabled users to generate fluent and seemingly convincing text. However, these models have uneven performance in different languages, which is also associated with undesirable societal biases toward marginalized populations. Specifically, there is relatively little work on Japanese models, despite it being the thirteenth most widely spoken language. In this work, we first develop three Japanese language prompts to probe LLMs’ understanding of Japanese names and their association between gender and occupations. We then evaluate a variety of English, multilingual, and Japanese models, correlating the models’ outputs with occupation statistics from the Japanese Census Bureau from the last 100 years. Our findings indicate that models can associate Japanese names with the correct gendered occupations when using constrained decoding. However, with sampling or greedy decoding, Japanese language models have a preference for a small set of stereotypically gendered occupations, and multilingual models, though trained on Japanese, are not always able to understand Japanese prompts.

In this paper, we investigate and evaluate large language models’ capability in location prediction. We present experimental results with four models—FLAN-T5, FLAN-UL2, FLAN-Alpaca, and ChatGPT—in various instruction finetuning and exemplar settings. We analyze whether taking into account the context—tweets published before and after the tweet mentioning a location—is beneficial. Additionally, we conduct an ablation study to explore whether instruction modification is beneficial. Lastly, our qualitative analysis sheds light on the errors made by the best-performing model.

Interpretation methods provide saliency scores indicating the importance of input words for neural summarization models. Prior work has analyzed models by comparing them to human behavior, often using eye-gaze as a proxy for human attention in reading tasks such as classification. This paper presents a framework to analyze the model behavior in summarization by comparing it to human summarization behavior using eye-gaze data. We examine two research questions: RQ1) whether model saliency conforms to human gaze during summarization and RQ2) how model saliency and human gaze affect summarization performance. For RQ1, we measure conformity by calculating the correlation between model saliency and human fixation counts. For RQ2, we conduct ablation experiments removing words/sentences considered important by models or humans. Experiments on two datasets with human eye-gaze during summarization partially confirm that model saliency aligns with human gaze (RQ1). However, ablation experiments show that removing highly-attended words/sentences from the human gaze does not significantly degrade performance compared with the removal by the model saliency (RQ2).

Word Sense Disambiguation (WSD) is a key task in Natural Language Processing (NLP), aiming to assign the correct meaning (sense) to a word in context. However, traditional WSD systems rely on WordNet as the underlying sense inventory, often differentiating meticulously between subtle nuances of word meanings, which may lead to excessive complexity and reduced practicality of WSD systems in today’s NLP. Indeed, current Pretrained Language Models (PLMs) do seem to be able to perform disambiguation, but it is not clear to what extent, or to what level of granularity, they actually operate. In this paper, we address these points and, firstly, introduce a new large-scale resource that leverages homonymy relations to systematically cluster WordNet senses, effectively reducing the granularity of word senses to a very coarse-grained level; secondly, we use this resource to train Homonymy Disambiguation systems and investigate whether PLMs are inherently able to differentiate coarse-grained word senses. Our findings demonstrate that, while state-of-the-art models still struggle to choose the correct fine-grained meaning of a word in context, Homonymy Disambiguation systems are able to differentiate homonyms with up to 95% accuracy scores even without fine-tuning the underlying PLM. We release our data and code at https://github.com/SapienzaNLP/homonymy-wsd.

Pre-training and fine-tuning large language models (LMs) is currently the state-of-the-art methodology for enabling data-scarce downstream tasks. However, the derived models still tend to replicate and perpetuate social biases. To understand this process in more detail, this paper investigates the actual effects of learning downstream tasks on moral bias in LMs. We develop methods to assess the agreement of LMs to explicitly codified norms in both pre-training and fine-tuning stages. Even if a pre-trained foundation model exhibits consistent norms, we find that introducing downstream tasks may indeed lead to unexpected inconsistencies in norm representation. Specifically, we observe two phenomena during fine-tuning across both masked and causal LMs: (1) pre-existing moral bias may be mitigated or amplified even when presented with opposing views and (2) prompt sensitivity may be negatively impacted. We provide empirical evidence of models deteriorating into conflicting states, where contradictory answers can easily be triggered by slight modifications in the input sequence. Our findings thus raise concerns about the general ability of LMs to mitigate moral biases effectively.

Chain-of-Thought (CoT) prompting combined with large language models (LLM) has shown great potential in improving performance on challenging reasoning tasks. While understanding why CoT prompting is effective is crucial for the application and improvement of CoT prompting, few studies have addressed this issue. Besides, almost no prior work has conducted theoretical analysis on CoT prompting in the context of black-box models. In this paper, we approach the analysis of CoT prompting in black-box LLMs from an information-theoretic perspective. Specifically, we propose a new metric, EPVI (Estimated Pointwise V-Information), which extends the concept of pointwise V-information to black-box models, quantifying the label-relevant new information introduced by CoT prompting beyond the pre-existing information in the input. Based on this, we conduct a series of experiments at both the task and instance levels to analyze CoT prompting, demonstrating that the effectiveness of CoT prompting can be attributed to its capacity to influence the difficulty of model inference by augmenting or reducing the model-usable information. Furthermore, we show that selecting high-quality demonstrations of CoT reasoning based on EPVI can improve the downstream performance of reasoning tasks.

In this paper, we investigate how new bilingual vocabulary is acquired through Iterative Back-Translation (IBT), which is known as a data augmentation method for machine translation from monolingual data of both source and target languages. To reveal the acquisition process, we first identify the word translation pairs in test data that do not exist in a bilingual data but do only in two monolingual data, then observe how many pairs are successfully translated by the translation model trained through IBT. We experimented on it with domain adaptation settings on two language pairs. Our experimental evaluation showed that more than 60% of the new bilingual vocabulary is successfully acquired through IBT along with the improvement in the translation quality in terms of BLEU. It also revealed that new bilingual vocabulary was gradually acquired by repeating IBT iterations. From the results, we present our hypothesis on the process of new bilingual vocabulary acquisition where the context of the words plays a critical role in the success of the acquisition.

Clarifying the effects of subjective ideas on group performance is essential for future dialogue systems to improve mutual understanding among humans and group creativity. However, there has been little focus on dialogue research on quantitatively analyzing the effects of the quality and quantity of subjective information contained in dialogues on group performance. We hypothesize that the more subjective information interlocutors exchange, the better the group performance in collaborative work. We collected dialogues between drivers and engineers in motorsports when deciding how the car should be tuned as a suitable case to verify this hypothesis. Our analysis suggests that the greater the amount of subjective information (which we defined as “sensation”) in the driver’s utterances, the greater the race performance and driver satisfaction with the car’s tuning. The results indicate that it is essential for the development of dialogue research to create a corpus of situations that require high performance through collaboration among experts with different backgrounds but who have mastered their respective fields.

A preprocessing task such as tokenization and sentence boundary detection (SBD) has commonly been considered as NLP challenges that have already been solved. This perception is due to their generally good performance and the presence of pre-tokenized data. However, it’s important to note that the low error rates of current methods are mainly specific to certain tasks, and rule-based tokenization can be difficult to use across different systems. Despite being subtle, these limitations are significant in the context of the NLP pipeline. In this paper, we introduce a novel evaluation algorithm for the preprocessing task, including both tokenization and SBD results. This algorithm aims to enhance the reliability of evaluations by reevaluating the counts of true positive cases for F1 measures in both preprocessing tasks jointly. It achieves this through an alignment-based approach inspired by sentence and word alignments used in machine translation. Our evaluation algorithm not only allows for precise counting of true positive tokens and sentence boundaries but also combines these two evaluation tasks into a single organized pipeline. To illustrate and clarify the intricacies of this calculation and integration, we provide detailed pseudo-code configurations for implementation. Additionally, we offer empirical evidence demonstrating how sentence and word alignment can improve evaluation reliability and present case studies to further support our approach.

Traditional Chinese character is an important carrier of Chinese culture, and is still actively used in many areas. Automatic conversion between traditional and simplified Chinese characters can help modern people understand traditional culture and facilitate communication among different regions. Previous conversion methods rely on rule-based mapping or shallow feature-based machine learning models, which struggle to convert simplified characters with different origins and constructing training data is costly. In this study, we propose an unsupervised adaptive context-aware conversion model that learns to convert between simplified and traditional Chinese characters under a denoising auto-encoder framework requiring no labeled data. Our model includes a Latent Generative Adversarial Encoder that transforms vectors to a latent space with generative adversarial network, which adds noise as an inevitable side effect, Based on which a Context-aware Semantic Reconstruction Decoder restores the original input while considering a broader range of context with a pretrained language model. Additionally, we propose to apply early exit mechanism during inference to reduce the computation complexity and improve the generalization ability. To test the effectiveness of our model, we construct a high quality test dataset with simplified-traditional Chinese character text pairs. Experiment results and extensive analysis demonstrate that our model outperforms strong unsupervised baselines and yields better conversion result for one-to-many cases.

Lexical Simplification (LS) aims to simplify text at the lexical level. Existing methods rely heavily on annotated data, making it challenging to apply in low-resource scenarios. In this paper, we propose a novel LS method without parallel corpora. This method employs an Adversarial Editing System with guidance from a confusion loss and an invariance loss to predict lexical edits in the original sentences. Meanwhile, we introduce an innovative LLM-enhanced loss to enable the distillation of knowledge from Large Language Models (LLMs) into a small-size LS system. From that, complex words within sentences are masked and a Difficulty-aware Filling module is crafted to replace masked positions with simpler words. At last, extensive experimental results and analyses on three benchmark LS datasets demonstrate the effectiveness of our proposed method.

A document representation model that has been used for years in NLP and Text Mining tasks is TF-IDF (Term Frequency-Inverse Document Frequency). This model is indeed effective for various tasks like Information Retrieval and Document Classification. However, it may fall short when it comes to capturing the deeper semantic and contextual meaning of a text, which is where Transformer-based Pre-trained Language Models (PLMs) such as BERT have been gaining significant traction in recent years. Despite this, these models also face specific challenges related to Transformers and their attention mechanism limits, especially when dealing with long documents. Therefore, this paper proposes a novel approach to exploit the advantages of the TF-IDF representation while incorporating semantic context, by introducing a Latent Concept Frequency-Inverse Document Frequency (LCF-IDF) document representation model. Its effectiveness is tested with respect to the Long Document Classification task. The results obtained show promising performance of the proposed solution compared to TF-IDF and BERT-like representation models, including those specifically for long documents such as Longformer as well as those designed for particular domains, especially when it comes to Single Label Multi-Class (SLMC) classification.

Time-Sensitive Question Answering (TSQA) is to answer questions qualified for a certain timestamp based on the given document. It is split into easy and hard modes depending on whether the document contain time qualifiers mentioned in the question. While existing models have performed well on easy mode, their performance is significant reduced for answering hard time-sensitive questions, whose time qualifiers are implicit in the document. An intuitive idea is to match temporal events in the given document by treating time-sensitive question as a temporal event of missing objects. However, not all temporal events extracted from the document have explicit time qualifiers. In this paper, we propose an Event-AL framework, in which a graph pruning model is designed to locate the timespan of implicit temporal events by capturing temporal relation between events. Moreover, we present an abductive reasoning module to determine proper objects while providing explanations. Besides, as the same relation may be scattered throughout the document in diverse expressions, a relation-based prompt is introduced to instructs LLMs in extracting candidate temporal events. We conduct extensive experiment and results show that Event-AL outperforms strong baselines for hard time-sensitive questions, with a 12.7% improvement in EM scores. In addition, it also exhibits great superiority for multi-answer and beyond hard time-sensitive questions.

To assist in the documentation of Čakavian, an endangered language variety closely related to Croatian, we test four currently available ASR models that are trained with Croatian data and assess their performance in the transcription of Čakavian audio data. We compare the models’ word error rates, analyze the word-level error types, and showcase the most frequent Deletion and Substitution errors. The evaluation results indicate that the best-performing system for transcribing Čakavian was a CTC-based variant of the Conformer model.