COLING 2024 论文列表

The paper reports on a series of experiments aiming at probing LeBenchmark, a pretrained acoustic model trained on 7k hours of spoken French, for syntactic information. Pretrained acoustic models are increasingly used for downstream speech tasks such as automatic speech recognition, speech translation, spoken language understanding or speech parsing. They are trained on very low level information (the raw speech signal), and do not have explicit lexical knowledge. Despite that, they obtained reasonable results on tasks that requires higher level linguistic knowledge. As a result, an emerging question is whether these models encode syntactic information. We probe each representation layer of LeBenchmark for syntax, using the Orféo treebank, and observe that it has learnt some syntactic information. Our results show that syntactic information is more easily extractable from the middle layers of the network, after which a very sharp decrease is observed.

We investigate the behavior of methods using linear projections to remove information about a concept from a language representation, and we consider the question of what happens to a dataset transformed by such a method. A theoretical analysis and experiments on real-world and synthetic data show that these methods inject strong statistical dependencies into the transformed datasets. After applying such a method, the representation space is highly structured: in the transformed space, an instance tends to be located near instances of the opposite label. As a consequence, the original labeling can in some cases be reconstructed by applying an anti-clustering method.

Large Language Models (LLMs) are now being considered as judges of high efficiency to evaluate the quality of answers generated by candidate models. However, their judgments may be influenced by complex scenarios and inherent biases, raising concerns about their reliability. This study aims to bridge this gap by introducing four unexplored factors and examining the performance of LLMs as judges, namely answer quantity, inducing statements, judging strategy, and judging style. Additionally, we introduce a new dimension of question difficulty to provide a more comprehensive understanding of LLMs’ judgments across varying question intricacies. We employ ChatGPT, GPT-4, Gemini, and Claude-2 as judges and conduct experiments on Vicuna Benchmark and MT-bench. Our study reveals that LLMs’ judging abilities are susceptible to the influence of these four factors, and analyzing from the newly proposed dimension of question difficulty is highly necessary. We also provide valuable insights into optimizing LLMs’ performance as judges, enhancing their reliability and adaptability across diverse evaluation scenarios.

This paper investigates what insights about linguistic features and what knowledge about the structure of natural language can be obtained from the encodings in transformer language models. In particular, we explore how BERT encodes the government relation between constituents in a sentence. We use several probing classifiers, and data from two morphologically rich languages. Our experiments show that information about government is encoded across all transformer layers, but predominantly in the early layers of the model. We find that, for both languages, a small number of attention heads encode enough information about the government relations to enable us to train a classifier capable of discovering new, previously unknown types of government, never seen in the training data. Currently, data is lacking for the research community working on grammatical constructions, and government in particular. We release the Government Bank—a dataset defining the government relations for thousands of lemmas in the languages in our experiments.

Predicting the compositionality of noun compounds such as climate change and tennis elbow is a vital component in natural language understanding. While most previous computational methods that automatically determine the semantic relatedness between compounds and their constituents have applied a synchronic perspective, the current study investigates what diachronic changes in contexts and semantic topics of compounds and constituents reveal about the compounds’ present-day degrees of compositionality. We define a binary classification task that utilizes two diachronic vector spaces based on contextual co-occurrences and semantic topics, and demonstrate that diachronic changes in cosine similarities – measured over context or topic distributions – uncover patterns that distinguish between compounds with low and high present-day compositionality. Despite fewer dimensions in the topic models, the topic space performs on par with the co-occurrence space and captures rather similar information. Temporal similarities between compounds and modifiers as well as between compounds and their prepositional paraphrases predict the compounds’ present-day compositionality with accuracy >0.7.

Non-information seeking questions (NISQ) capture the subtle dynamics of human discourse. In this work, we utilize a dataset of over 1,500 information-seeking question(ISQ) and NISQ to evaluate human and machine performance on classifying fine-grained NISQ types. We introduce the first publicly available corpus focused on annotating both ISQs and NISQs as an initial benchmark. Additionally, we establish competitive baselines by assessing diverse systems, including Generative Pre-Trained Transformer Language models, on a new question classification task. Our results demonstrate the inherent complexity of making nuanced NISQ distinctions. The dataset is publicly available at https://github.com/YaoSun0422/NISQ_dataset.git

Hierarchical Text Classification (HTC) is a challenging task which aims to extract the labels in a tree structure corresponding to a given text. Discriminative methods usually incorporate the hierarchical structure information into the encoding process, while generative methods decode the features according to it. However, the data distribution varies widely among different categories of samples, but current methods ignore the data imbalance, making the predictions biased and susceptible to error propagation. In this paper, we propose an **IM**plicitly **A**ugmented **G**enerativ **E** framework with distribution modification for hierarchical text classification (**IMAGE**). Specifically, we translate the distributions of original samples along various directions through implicit augmentation to get more diverse data. Furthermore, given the scarcity of the samples of tail classes, we adjust their distributions by transferring knowledge from other classes in label space. In this way, the generative framework learns a better beginning of the feature sequence without a prediction bias and avoids being misled by its wrong predictions for head classes. Experimental results show that **IMAGE** obtains competitive results compared with state-of-the-art methods and prove its superiority on unbalanced data.

Corpora are important assets within the natural language processing (NLP) and linguistics communities, as they allow the training of models and corpus-based studies of languages. However, corpora do not exist for many languages and language varieties, such as Acadian French. In this paper, we first show that off-the-shelf NLP systems perform more poorly on Acadian French than on standard French. An Acadian French corpus could, therefore, potentially be used to improve NLP models for this dialect. Then, leveraging web-as-corpus methodologies, specifically BootCaT, domain crawling, and social media scraping, we create three corpora of Acadian French. To evaluate these corpora, drawing on the linguistic literature on Acadian French, we propose 22 statistical corpus-based measures of the extent to which a corpus is Acadian French. We use these measures to compare these newly built corpora to known Acadian French text and find that all three corpora include some traces of Acadian French.

Sign language is an effective non-verbal communication mode for the hearing-impaired people. Since the video-based sign language detection models have high requirements for enough lighting and clear background, current wearing glove-based sign language models are robust for poor light and occlusion situations. In this paper, we annotate a new dataset of Word-based Wearable Chinese Sign Languag (WW-CSL) gestures. Specifically, we propose a three-form (e.g., sequential sensor data, gesture video, and gesture text) scheme to represent dynamic CSL gestures. Guided by the scheme, a total of 3,000 samples were collected, corresponding to 100 word-based CSL gestures. Furthermore, we present a transformer-based baseline model to fuse 2 inertial measurement unites (IMUs) and 10 flex sensors for the wearable CSL detection. In order to integrate the advantage of video-based and wearable glove-based CSL gestures, we also propose a transformer-based Multi-Modal CSL Detection (MM-CSLD) framework which adeptly integrates the local sequential sensor data derived from wearable-based CSL gestures with the global, fine-grained skeleton representations captured from video-based CSL gestures simultaneously.

Evaluation plays a significant role in modern natural language processing. Most modern NLP benchmarks consist of arbitrary sets of tasks that neither guarantee any generalization potential for the model once applied outside the test set nor try to minimize the resource consumption needed for model evaluation. This paper presents a theoretical instrument and a practical algorithm to calculate similarity between benchmark tasks, we call this similarity measure “Vygotsky distance”. The core idea of this similarity measure is that it is based on relative performance of the “students” on a given task, rather that on the properties of the task itself. If two tasks are close to each other in terms of Vygotsky distance the models tend to have similar relative performance on them. Thus knowing Vygotsky distance between tasks one can significantly reduce the number of evaluation tasks while maintaining a high validation quality. Experiments on various benchmarks, including GLUE, SuperGLUE, CLUE, and RussianSuperGLUE, demonstrate that a vast majority of NLP benchmarks could be at least 40% smaller in terms of the tasks included. Most importantly, Vygotsky distance could also be used for the validation of new tasks thus increasing the generalization potential of the future NLP models.

In recent years, there has been great interest in multimodal inference. We concentrate on visual-textual entailment (VTE), a critical task in multimodal inference. VTE is the task of determining entailment relations between an image and a sentence. Several deep learning-based approaches have been proposed for VTE, but current approaches struggle with accurately handling quantities. On the other hand, one promising approach, one based on logical inference that can successfully deal with large quantities, has also been proposed. However, that approach uses automated theorem provers, increasing the computational cost for problems involving many entities. In addition, that approach cannot deal well with lexical differences between the semantic representations of images and sentences. In this paper, we present a logic-based VTE system that overcomes these drawbacks, using model checking for inference to increase efficiency and knowledge injection to perform more robust inference. We create a VTE dataset containing quantities and negation to assess how well VTE systems understand such phenomena. Using this dataset, we demonstrate that our system solves VTE tasks with quantities and negation more robustly than previous approaches.

Image-text retrieval is a fundamental task to bridge the semantic gap between natural language and vision. Recent works primarily focus on aligning textual meanings with visual appearance. However, they often overlook the semantic discrepancy caused by syntactic structure in natural language expressions and relationships among visual entities. This oversight would lead to sub-optimal alignment and degraded retrieval performance, since the underlying semantic dependencies and object interactions remain inadequately encoded in both textual and visual embeddings. In this paper, we propose a novel Visual-Linguistic Dependency Encoding (VL-DE) framework, which explicitly models the dependency information among textual words and interaction patterns between image regions, improving the discriminative power of cross-modal representations for more accurate image-text retrieval. Specifically, VL-DE enhances textual representations by considering syntactic relationships and dependency types, and visual representations by attending to its spatially neighboring regions. Cross-attention mechanism is then introduced to aggregate aligned region-word pairs into image-text similarities. Analysis on Winoground, a dataset specially designed to measure vision-linguistic compositional structure reasoning, shows that VL-DE outperforms existing methods, demonstrating its effectiveness at this task. Comprehensive experiments on two benchmarks, Flickr30K and MS-COCO, further validates the competitiveness of our approach.

Due to privacy restrictions, there’s a shortage of publicly available speech recognition datasets in the medical domain. In this work, we present VietMed - a Vietnamese speech recognition dataset in the medical domain comprising 16h of labeled medical speech, 1000h of unlabeled medical speech and 1200h of unlabeled general-domain speech. To our best knowledge, VietMed is by far the world’s largest public medical speech recognition dataset in 7 aspects: total duration, number of speakers, diseases, recording conditions, speaker roles, unique medical terms and accents. VietMed is also by far the largest public Vietnamese speech dataset in terms of total duration. Additionally, we are the first to present a medical ASR dataset covering all ICD-10 disease groups and all accents within a country. Moreover, we release the first public large-scale pre-trained models for Vietnamese ASR, w2v2-Viet and XLSR-53-Viet, along with the first public large-scale fine-tuned models for medical ASR. Even without any medical data in unsupervised pre-training, our best pre-trained model XLSR-53-Viet generalizes very well to the medical domain by outperforming state-of-the-art XLSR-53, from 51.8% to 29.6% WER on test set (a relative reduction of more than 40%). All code, data and models are made publicly available here.

Lexical-syntactic flexibility, in the form of conversion (or zero-derivation) is a hallmark of English morphology. In conversion, a word with one part of speech is placed in a non-prototypical context, where it is coerced to behave as if it had a different part of speech. However, while this process affects a large part of the English lexicon, little work has been done to establish the degree to which language models capture this type of generalization. This paper reports the first study on the behavior of large language models with reference to conversion. We design a task for testing lexical-syntactic flexibility—the degree to which models can generalize over words in a construction with a non-prototypical part of speech. This task is situated within a natural language inference paradigm. We test the abilities of five language models—two proprietary models (GPT-3.5 and GPT-4), three open source model (Mistral 7B, Falcon 40B, and Llama 2 70B). We find that GPT-4 performs best on the task, followed by GPT-3.5, but that the open source language models are also able to perform it and that the 7-billion parameter Mistral displays as little difference between its baseline performance on the natural language inference task and the non-prototypical syntactic category task, as the massive GPT-4.

This paper investigates the impact of corpus creation decisions on large multi-lingual geographic web corpora. Beginning with a 427 billion word corpus derived from the Common Crawl, three methods are used to improve the quality of sub-corpora representing specific language-country pairs like New Zealand English: (i) the agreement of independent language identification systems, (ii) hash-based deduplication, and (iii) location-specific outlier detection. The impact of each of these steps is then evaluated at the language level and the country level by using corpus similarity measures to compare each resulting corpus with baseline data sets. The goal is to understand the impact of upstream data cleaning decisions on downstream corpora with a specific focus on under-represented languages and populations. The evaluation shows that the validity of sub-corpora is improved with each stage of cleaning but that this improvement is unevenly distributed across languages and populations. This result shows how standard corpus creation techniques can accidentally exclude under-represented populations.

Out-of-distribution (OOD) detection plays a crucial role in ensuring the safety and reliability of deep neural networks in various applications. While there has been a growing focus on OOD detection in visual data, the field of textual OOD detection has received less attention. Only a few attempts have been made to directly apply general OOD detection methods to natural language processing (NLP) tasks, without adequately considering the characteristics of textual data. In this paper, we delve into textual OOD detection with Transformers. We first identify a key problem prevalent in existing OOD detection methods: the biased representation learned through the maximization of the conditional likelihood p(y|x) can potentially result in subpar performance. We then propose a novel variational inference framework for OOD detection (VI-OOD), which maximizes the likelihood of the joint distribution p(x, y) instead of p(y|x). VI-OOD is tailored for textual OOD detection by efficiently exploiting the representations of pre-trained Transformers. Through comprehensive experiments on various text classification tasks, VI-OOD demonstrates its effectiveness and wide applicability. Our code has been released at https://github.com/liam0949/LLM-OOD.

Verb detection is a fundamental task in natural language processing that involves identifying the action or state expressed by a verb in a sentence. However, in Uzbek language, verb detection is challenging due to the complexity of its morphology and the agglutinative nature of the language. In this paper, we propose a rule-based approach for verb detection in Uzbek texts based on affixes/suffixes. Our method is based on a set of rules that capture the morphological patterns of verb forms in Uzbek language. We evaluate the proposed approach on a dataset of Uzbek texts and report an F1-score of 0.97, which outperforms existing methods for verb detection in Uzbek language. Our results suggest that rule-based approaches can be effective for verb detection in Uzbek texts and have potential applications in various natural language processing tasks.

This paper focuses on improving the performance of machine translation for manga (Japanese-style comics). In manga machine translation, text consists of a sequence of speech bubbles and each speech bubble is translated individually. However, each speech bubble itself does not contain sufficient information for translation. Therefore, previous work has proposed methods to use contextual information, such as the previous speech bubble, speech bubbles within the same scene, and corresponding scene images. In this research, we propose two new approaches to capture broader contextual information. Our first approach involves scene-based translation that considers the previous scene. The second approach considers broader context information, including details about the work, author, and manga genre. Through our experiments, we confirm that each of our methods improves translation quality, with the combination of both methods achieving the highest quality. Additionally, detailed analysis reveals the effect of zero-anaphora resolution in translation, such as supplying missing subjects not mentioned within a scene, highlighting the usefulness of longer contextual information in manga machine translation.

Hierarchical text classification (HTC) is a challenging subtask of multi-label classification due to its complex taxonomic structure. Nearly all recent HTC works focus on how the labels are structured but ignore the sub-structure of ground-truth labels according to each input text which contains fruitful label co-occurrence information. In this work, we introduce this local hierarchy with an adversarial framework. We propose a HiAdv framework that can fit in nearly all HTC models and optimize them with the local hierarchy as auxiliary information. We test on two typical HTC models and find that HiAdv is effective in all scenarios and is adept at dealing with complex taxonomic hierarchies. Further experiments demonstrate that the promotion of our framework indeed comes from the local hierarchy and the local hierarchy is beneficial for rare classes which have insufficient training data.