COLING 2024 论文列表

Deep neural networks (DNNs) are notoriously vulnerable to adversarial attacks that place carefully crafted perturbations on normal examples to fool DNNs. To better understand such attacks, a characterization of the features carried by adversarial examples is needed. In this paper, we tackle this challenge by inspecting the subspaces of sample features through spectral analysis. We first empirically show that the features of either clean signals or adversarial perturbations are redundant and span in low-dimensional linear subspaces respectively with minimal overlap, and the classical low-dimensional subspace projection can suppress perturbation features out of the subspace of clean signals. This makes it possible for DNNs to learn a subspace where only features of clean signals exist while those of perturbations are discarded, which can facilitate the distinction of adversarial examples. To prevent the residual perturbations that is inevitable in subspace learning, we propose an independence criterion to disentangle clean signals from perturbations. Experimental results show that the proposed strategy enables the model to inherently suppress adversaries, which not only boosts model robustness but also motivates new directions of effective adversarial defense.

Large Language Models (LLMs) have demonstrated impressive performances across various NLP tasks with just a few prompts via in-context learning. Previous studies have emphasized the pivotal role of well-chosen examples in in-context learning, as opposed to randomly selected instances that exhibits unstable results.A successful example selection scheme depends on multiple factors, while in the context of LLMs-based machine translation, the common selection algorithms only consider the single factor, i.e., the similarity between the example source sentence and the input sentence.In this paper, we introduce a novel approach to use multiple translational factors for in-context example selection by using monotone submodular function maximization.The factors include surface/semantic similarity between examples and inputs on both source and target sides, as well as the diversity within examples.Importantly, our framework mathematically guarantees the coordination between these factors, which are different and challenging to reconcile.Additionally, our research uncovers a previously unexamined dimension: unlike other NLP tasks, the translation part of an example is also crucial, a facet disregarded in prior studies.Experiments conducted on BLOOMZ-7.1B and LLAMA2-13B, demonstrate that our approach significantly outperforms random selection and robust single-factor baselines across various machine translation tasks.

We propose a sub-table rescorer (STR) to improve the performance of an inner table retriever (ITR)-based inference for the table question answering. Tabular language model (TLM) truncates the sequence of a long table due to their input token limits. It leads to accuracy degradation. To solve this problem, ITR extracts sub-table candidates, which correspond to a part of an entire greater original table on the basis of relevance scores to the question for each of the columns and rows. Then, the topN longest sub-tables are selected. Our proposed STR estimates the relevance score between a question and each sub-table. In this work, we explored two different methods to integrate STR to the ITR-based inference. In the first method, STR rescores sub-table candidates, and the topN sub-tables are chosen. Then, TLM outputs the most confident answer. In the second method, the score calculated by STR is interpolated with the score calculated by TLM. Then, the most confident answer is chosen. In the experiment, we evaluate the performance on the WikiTableQuestions dataset. By applying STR to the ITR-based inference, we observed 4.4% and 6.3% relative reductions in error rate in the rescoring- and score-fusion-based methods, respectively.

Unsupervised text style transfer aims to modify the style of a sentence while preserving its content without parallel corpora. Existing approaches attempt to separate content from style, but some words contain both content and style information. It makes them difficult to disentangle, where unsatisfactory disentanglement results in the loss of the content information or the target style. To address this issue, researchers adopted a “cycle reconstruction” mechanism to maintain content information, but it is still hard to achieve satisfactory content preservation due to incomplete disentanglement. In this paper, we propose a new disentanglement-based method, StyleFlow, which effectively avoids the loss of contents through a better cycle reconstruction via a reversible encoder. The reversible encoder is a normalizing flow that can not only produce output given input but also infer the exact input given the output reversely. We design a stack of attention-aware coupling layers, where each layer is reversible and adopts the attention mechanism to improve the content-style disentanglement. Moreover, we propose a data augmentation method based on normalizing flow to enhance the training data. Our experiments on sentiment transfer and formality transfer tasks show that StyleFlow outperforms strong baselines on both content preservation and style transfer.

Employing pre-trained generation models for cross-domain aspect-based sentiment classification has recently led to large improvements. However, they ignore the importance of syntactic structures, which have shown appealing effectiveness in classification based models. Different from previous studies, efficiently encoding the syntactic structure in generation model is challenging because such models are pretrained on natural language, and modeling structured data may lead to catastrophic forgetting of distributional knowledge. In this study, we propose a novel structure-aware generation model to tackle this challenge. In particular, a prompt-driven strategy is designed to bridge the gap between different domains, by capturing implicit syntactic information from the input and output sides. Furthermore, the syntactic structure is explicitly encoded into the structure-aware generation model, which can effectively learn domain-irrelevant features based on syntactic pivot features. Empirical results demonstrate the effectiveness of the proposed structure-aware generation model over several strong baselines. The results also indicate the proposed model is capable of leveraging the input syntactic structure into the generation model.

Over the past few years, we have witnessed remarkable advancements in Code Pre-trained Models (CodePTMs). These models achieved excellent representation capabilities by designing structure-based pre-training tasks for code. However, how to enhance the absorption of structural knowledge when fine-tuning CodePTMs still remains a significant challenge. To fill this gap, in this paper, we present SAT, a novel structure-enhanced and plug-and-play fine-tuning method for CodePTMs. We first propose a structure loss to quantify the difference between the information learned by CodePTMs and the knowledge extracted from code structure. Specifically, we use the attention scores from Transformer layer as the learned information, and the shortest path length between leaves in abstract syntax trees as the structural knowledge. Subsequently, multi-task learning is introduced to improve the performance of fine-tuning. Experiments conducted on four pre-trained models and two generation tasks demonstrate the effectiveness of our proposed method as a plug-and-play solution. Furthermore, we observed that SAT can benefit CodePTMs more with limited training data.

The task of address matching involves linking unstructured addresses to standard ones in a database. The challenges presented by this task are manifold: misspellings, incomplete information, and variations in address content are some examples. While there have been previous studies on entity matching in natural language processing, for the address matching solution, existing approaches still rely on string-based similarity matching or manually-designed rules. In this paper, we propose StructAM, a novel method based on pre-trained language models (LMs) and graph neural networks to extract the textual and structured information of the addresses. The proposed method leverages the knowledge acquired by large language models during the pre-training phase, and refines it during the fine-tuning process on the address domain, to obtain address-specific semantic features. Meanwhile, it also applies an attribute attention mechanism based on Graph Sampling and Aggregation (GraphSAGE) module to capture internal hierarchy information of the address text. To further enhance the accuracy of our algorithm in dirty settings, we incorporate spatial coordinates and contextual information from the surrounding area as auxiliary guidance. We conduct extensive experiments on real-world datasets from four different countries and the results show that StructAM outperforms state-of-the-art baseline approaches for address matching.

This study addresses the critical issue of Natural Language Processing in low-resource languages such as Hindi, which, despite having substantial number of speakers, is limited in linguistic resources. The paper focuses on Word Sense Disambiguation, a fundamental NLP task that deals with polysemous words. It introduces a novel Hindi WSD dataset in the modern WiC format, enabling the training and testing of contextualized models. The primary contributions of this work lie in testing the efficacy of multilingual models to transfer across languages and hence to handle polysemy in low-resource languages, and in providing insights into the minimum training data required for a viable solution. Experiments compare different contextualized models on the WiC task via transfer learning from English to Hindi. Models purely transferred from English yield poor 55% accuracy, while fine-tuning on Hindi dramatically improves performance to 90% accuracy. This demonstrates the need for language-specific tuning and resources like the introduced Hindi WiC dataset to drive advances in Hindi NLP. The findings offer valuable insights into addressing the NLP needs of widely spoken yet low-resourced languages, shedding light on the problem of transfer learning in these contexts.

Storytelling, i.e., the use of of anecdotes and personal experiences, plays a crucial role in everyday argumentation. This is particularly true for the highly controversial debates that spark in times of crisis - where the focus of the discussion is on heterogeneous aspects of everyday life. For individuals, stories can have a strong persuasive power; for a larger collective, stories can help decision-makers to develop strategies for addressing the challenges people are facing, especially in times of crisis. In this paper, we analyse the use of storytelling in the COVID-19 discourse. We carry out our analysis on three publicly available Reddit datasets, for a total of 367K comments. We automatically annotate the Reddit datasets by detecting spans containing storytelling and classifying them into: a) personal vs. general – is the story experienced by the speaker? b) argumentative function (Does the story clarify a problem, potentially consisting in harm to a specific group? Does it exemplify a solution to a problem, or does it establish the credibility of the speaker?), and c) topic. We then carry out an analysis which establishes the relevance of storytelling in the COVID discourse and further uncovers interactions between topics and types of stories associated to them.

Modern Greek is normally written in the Greek alphabet. In informal online messages, however, Greek is often written using characters available on Latin-character keyboards, a form known as Greeklish. Originally used to bypass the lack of support for the Greek alphabet in older computers, Greeklish is now also used to avoid switching languages on multilingual keyboards, hide spelling mistakes, or as a form of slang. There is no consensus mapping, hence the same Greek word can be written in numerous different ways in Greeklish. Even native Greek speakers may struggle to understand (or be annoyed by) Greeklish, which requires paying careful attention to context to decipher. Greeklish may also be a problem for NLP models trained on Greek datasets written in the Greek alphabet. Experimenting with a range of statistical and deep learning models on both artificial and real-life Greeklish data, we find that: (i) prompting large language models (e.g., GPT-4) performs impressively well with few- or even zero-shot training, outperforming several fine-tuned encoder-decoder models; however (ii) a twenty years old statistical Greeklish transliteration model is still very competitive; and (iii) the problem is still far from having been solved; (iv) nevertheless, downstream Greek NLP systems that need to cope with Greeklish, such as moderation classifiers, can benefit significantly even with the current non-perfect transliteration systems. We make all our code, models, and data available and suggest future improvements, based on an analysis of our experimental results.

Recently, the autoregressive framework based on large language models (LLMs) has achieved excellent performance in controlling the generated text to adhere to the required style. These methods guide LLMs through prompt learning to generate target text in an autoregressive manner. However, this manner possesses lower controllability and suffers from the challenge of accumulating errors, where early prediction inaccuracies might influence subsequent word generation. Furthermore, existing prompt-based methods overlook specific region editing, resulting in a deficiency of localized control over input text. To overcome these challenges, we propose a novel three-stage prompt-based approach for specific region editing. To alleviate the issue of accumulating errors, we transform the text style transfer task into a text infilling task, guiding the LLMs to modify only a small portion of text within the editing region to achieve style transfer, thus reducing the number of autoregressive iterations. To achieve an effective specific editing region, we adopt both prompt-based and word frequency-based strategies for region selection, subsequently employing a discriminator to validate the efficacy of the selected region. Experiments conducted on several publicly competitive datasets for text style transfer task confirm that our proposed approach achieves state-of-the-art performance. Keywords: text style transfer, natural language generation, large language models

An entailment tree is a structured reasoning path that clearly demonstrates the process of deriving hypotheses through multiple steps of inference from known premises. It enhances the interpretability of QA systems. Existing methods for generating entailment trees typically employ iterative frameworks to ensure reasoning faithfulness. However, they often suffer from the issue of false feasible steps, where selected steps appear feasible but actually lead to incorrect intermediate conclusions. Moreover, the existing iterative frameworks do not consider error-prone search branches, resulting in error propagation. In this work, we propose SPEH: an iterative entailment tree generation framework with Step feasibility Perception and state Error Handling mechanisms. Step Feasibility Perception enables the model to learn how to choose steps that are not false feasible. State Error Handling includes error detection and backtracking, allowing the model to correct errors when entering incorrect search branches. Experimental results demonstrate the effectiveness of our approach in improving the generation of entailment trees.

Social media platforms are rich sources of opinionated content. Stance detection allows the automatic extraction of users’ opinions on various topics from such content. We focus on zero-shot stance detection, where the model’s success relies on (a) having knowledge about the target topic; and (b) learning general reasoning strategies that can be employed for new topics. We present Stance Reasoner, an approach to zero-shot stance detection on social media that leverages explicit reasoning over background knowledge to guide the model’s inference about the document’s stance on a target. Specifically, our method uses a pre-trained language model as a source of world knowledge, with the chain-of-thought in-context learning approach to generate intermediate reasoning steps. Stance Reasoner outperforms the current state-of-the-art models on 3 Twitter datasets, including fully supervised models. It can better generalize across targets, while at the same time providing explicit and interpretable explanations for its predictions.

Automatic generation of code from natural language descriptions has emerged as one of the main use cases of large language models (LLMs). This has also led to a proliferation of datasets to track progress in the reliability of code generation models, including domains such as programming challenges and common data science tasks. However, existing datasets primarily target the use of code generation models to aid expert programmers in writing code. In this work, we consider a domain of code generation which is more frequently used by users without sophisticated programming skills: translating English descriptions to spreadsheet formulas that can be used to do everyday data processing tasks. We extract naturalistic instructions from StackOverflow posts and manually verify and standardize the corresponding spreadsheet formulas. We use this dataset to evaluate an off-the-shelf code generation model (GPT 3.5 text-davinci-003) as well as recently proposed pragmatic code generation procedures and find that Code Reviewer reranking (Zhang et al., 2022) performs best among the evaluated methods but still frequently generates formulas that differ from human-generated ones.

Speech-to-Speech and Speech-to-Text translation are currently dynamic areas of research. In our commitment to advance these fields, we present SpeechAlign, a framework designed to evaluate the underexplored field of source-target alignment in speech models. The SpeechAlign framework has two core components. First, to tackle the absence of suitable evaluation datasets, we introduce the Speech Gold Alignment dataset, built upon a English-German text translation gold alignment dataset. Secondly, we introduce two novel metrics, Speech Alignment Error Rate (SAER) and Time-weighted Speech Alignment Error Rate (TW-SAER), which enable the evaluation of alignment quality within speech models. While the former gives equal importance to each word, the latter assigns weights based on the length of the words in the speech signal. By publishing SpeechAlign we provide an accessible evaluation framework for model assessment, and we employ it to benchmark open-source Speech Translation models. In doing so, we contribute to the ongoing research progress within the fields of Speech-to-Speech and Speech-to-Text translation.

Automatic Speech Recognition (ASR) can be a valuable tool to document endangered languages. However, building ASR tools for these languages poses several difficult research challenges, notably data scarcity. In this paper, we show the whole process of creating a useful ASR tool for language documentation scenarios. We publish the first speech corpus for Khinalug, an endangered language spoken in Northern Azerbaijan. The corpus consists of 2.67 hours of labeled data from recordings of spontaneous speech about various topics. As Khinalug is an extremely low-resource language, we investigate the benefits of multilingual models for self-supervised learning and supervised learning and achieve the performance of 6.65 Character Error Rate (CER) points and 25.53 Word Error Rate (WER) points. The benefits of multilingual models are further validated through experimentation with three additional under-resourced languages. Lastly, this work conducts quality assessments with linguists on new recordings to investigate the model’s usefulness in language documentation. We observe an evident degradation for new recordings, indicating the importance of enhancing model robustness. In addition, we find the inaudible content is the main cause of wrong ASR predictions, suggesting relating work on incorporating contextual information.

Despite the growing demand for digital therapeutics for children with Autism Spectrum Disorder (ASD), there is currently no speech corpus available for Korean children with ASD. This paper introduces a speech corpus specifically designed for Korean children with ASD, aiming to advance speech technologies such as pronunciation and severity evaluation. Speech recordings from speech and language evaluation sessions were transcribed, and annotated for articulatory and linguistic characteristics. Three speech and language pathologists rated these recordings for social communication severity (SCS) and pronunciation proficiency (PP) using a 3-point Likert scale. The total number of participants will be 300 for children with ASD and 50 for typically developing (TD) children. The paper also analyzes acoustic and linguistic features extracted from speech data collected and completed for annotation from 73 children with ASD and 9 TD children to investigate the characteristics of children with ASD and identify significant features that correlate with the clinical scores. The results reveal some speech and linguistic characteristics in children with ASD that differ from those in TD children or another subgroup of ASD categorized by clinical scores, demonstrating the potential for developing automatic assessment systems for SCS and PP.

Italy exhibits rich linguistic diversity across its territory due to the distinct regional languages spoken in different areas. Recent advances in self-supervised learning provide new opportunities to analyze Italy’s linguistic varieties using speech data alone. This includes the potential to leverage representations learned from large amounts of data to better examine nuances between closely related linguistic varieties. In this study, we focus on automatically identifying the geographic region of origin of speech samples drawn from Italy’s diverse language varieties. We leverage self-supervised learning models to tackle this task and analyze differences and similarities between Italy’s regional languages. In doing so, we also seek to uncover new insights into the relationships among these diverse yet closely related varieties, which may help linguists understand their interconnected evolution and regional development over time and space. To improve the discriminative ability of learned representations, we evaluate several supervised contrastive learning objectives, both as pre-training steps and additional fine-tuning objectives. Experimental evidence shows that pre-trained self-supervised models can effectively identify regions from speech recording. Additionally, incorporating contrastive objectives during fine-tuning improves classification accuracy and yields embeddings that distinctly separate regional varieties, demonstrating the value of combining self-supervised pre-training and contrastive learning for this task.

This paper introduces a novel annotation framework for the fine-grained modeling of Noun Phrases’ (NPs) genericity in natural language. The framework is designed to be simple and intuitive, making it accessible to non-expert annotators and suitable for crowd-sourced tasks. Drawing from theoretical and cognitive literature on genericity, this framework is grounded in established linguistic theory. Through a pilot study, we created a small but crucial annotated dataset of 324 sentences, serving as a foundation for future research. To validate our approach, we conducted an evaluation comparing our continuous annotations with existing binary annotations on the same dataset, demonstrating the framework’s effectiveness in capturing nuanced aspects of genericity. Our work offers a practical resource for linguists, providing a first annotated dataset and an annotation scheme designed to build real-language datasets that can be used in studies on the semantics of genericity, and NLP practitioners, contributing to the development of commonsense knowledge repositories valuable in enhancing various NLP applications.