COLING 2024 论文列表

Contemporary neural speech synthesis models have indeed demonstrated remarkable proficiency in synthetic speech generation as they have attained a level of quality comparable to that of human-produced speech. Nevertheless, it is important to note that these achievements have predominantly been verified within the context of high-resource languages such as English. Furthermore, the Tacotron and FastSpeech variants show substantial pausing errors when applied to the Korean language, which affects speech perception and naturalness. In order to address the aforementioned issues, we propose a novel framework that incorporates comprehensive modeling of both syntactic and acoustic cues that are associated with pausing patterns. Remarkably, our framework possesses the capability to consistently generate natural speech even for considerably more extended and intricate out-of-domain (OOD) sentences, despite its training on short audio clips. Architectural design choices are validated through comparisons with baseline models and ablation studies using subjective and objective metrics, thus confirming model performance.

African American English (AAE) has received recent attention in the field of natural language processing (NLP). Efforts to address bias against AAE in NLP systems tend to focus on lexical differences. When the unique structures of AAE are considered, the solution is often to remove or neutralize the differences. This work leverages knowledge about the unique linguistic structures to improve automatic disambiguation of habitual and non-habitual meanings of “be” in naturally produced AAE transcribed speech. Both meanings are employed in AAE but examples of Habitual be are rare in already limited AAE data. Generally, representing additional syntactic information improves semantic disambiguation of habituality. Using an ensemble of classical machine learning models with a representation of the unique POS and dependency patterns of Habitual be, we show that integrating syntactic information improves the identification of habitual uses of “be” by about 65 F1 points over a simple baseline model of n-grams, and as much as 74 points. The success of this approach demonstrates the potential impact when we embrace, rather than neutralize, the structural uniqueness of African American English.

With the development of the Internet, social media has produced a large amount of user-generated data, which brings new challenges for humor computing. Traditional humor computing research mainly focuses on the content, while neglecting the information of interaction relationships in social media. In addition, both content and users are important in social media, while existing humor computing research mainly focuses on content rather than people. To address these problems, we model the information transfer and entity interactions in social media as a heterogeneous graph, and create the first dataset which introduces the social context information - HumorWB, which is collected from Chinese social media - Weibo. Two humor-related tasks are designed in the dataset. One is a content-oriented humor recognition task, and the other is a novel humor evaluation task. For the above tasks, we purpose a graph-based model called SCOG, which uses heterogeneous graph neural networks to optimize node representation for downstream tasks. Experimental results demonstrate the effectiveness of feature extraction and graph representation learning methods in the model, as well as the necessity of introducing social context information.

Counter-narrative generation, i.e., the generation of fact-based responses to hate speech with the aim of correcting discriminatory beliefs, has been demonstrated to be an effective method to combat hate speech. However, its effectiveness is limited by the resource-intensive nature of dataset construction processes and only focuses on the primary language. To alleviate this problem, we propose a Korean Hate Speech Counter Punch (KHSCP), a cost-effective counter-narrative generation method in the Korean language. To this end, we release the first counter-narrative generation dataset in Korean and pose two research questions. Under the questions, we propose an effective augmentation method and investigate the reasonability of a large language model to overcome data scarcity in low-resource environments by leveraging existing resources. In this regard, we conduct several experiments to verify the effectiveness of the proposed method. Our results reveal that applying pre-existing resources can improve the generation performance by a significant margin. Through deep analysis on these experiments, this work proposes the possibility of overcoming the challenges of generating counter-narratives in low-resource environments.

Open Information Extraction (OIE) is a structure prediction (SP) task in Natural Language Processing (NLP) that aims to extract structured n-ary tuples - usually subject-relation-object triples - from free text. The word embeddings in the input text can be enhanced with linguistic features, usually Part-of-Speech (PoS) and Syntactic Dependency Parse (SynDP) labels. However, past enhancement techniques cannot leverage the power of pre-trained language models (PLMs), which themselves have been hardly used for OIE. To bridge this gap, we are the first to leverage linguistic features with a Seq2Seq PLM for OIE. We do so by introducing two methods - Weighted Addition and Linearized Concatenation. Our work gives any neural OIE architecture the key performance boost from both PLMs and linguistic features in one go. In our settings, this shows wide improvements of up to 24.9%, 27.3% and 14.9% on Precision, Recall and F1 scores respectively over the baseline. Beyond this, we address other important challenges in the field: to reduce compute overheads with the features, we are the first ones to exploit Semantic Dependency Parse (SemDP) tags; to address flaws in current datasets, we create a clean synthetic dataset; finally, we contribute the first known study of OIE behaviour in SP models.

We explore the task of event extraction and classification by harnessing the power of distant supervision. We present a novel text labeling method that leverages the redundancy of temporal information in a data lake. This method enables the creation of a large programmatically annotated corpus, allowing the training of transformer models using distant supervision. This aims to reduce expert annotation time, a scarce and expensive resource. Our approach utilizes temporal redundancy between structured sources and text, enabling the design of a replicable framework applicable to diverse real-world databases and use cases. We employ this method to create multiple silver datasets to reconstruct key events in cancer patients’ pathways, using clinical notes from a cohort of 380,000 oncological patients. By employing various noise label management techniques, we validate our end-to-end approach and compare it with a baseline classifier built on expert-annotated data. The implications of our work extend to accelerating downstream applications, such as patient recruitment for clinical trials, treatment effectiveness studies, survival analysis, and epidemiology research. While our study showcases the potential of the method, there remain avenues for further exploration, including advanced noise management techniques, semi-supervised approaches, and a deeper understanding of biases in the generated datasets and models.

This article addresses methodological issues related to developing domain corpora and a terminological database from scratch. We present an ongoing project focused on creating an Estonian-English Remote Sensing Termbase. First, we describe the compilation process of the Estonian Remote Sensing Corpus 2022 , which served as the primary data source for the termbase. The corpus was compiled by crawling the web and adding files using the Corpus Query System Sketch Engine (Kilgarriff et al., 2004). In the next step, we employed the Term Extraction module (Kilgarriff et al., 2014; Fišer et al., 2016; Blahuš et al., 2023) to identify terms, which were subsequently registered in the Estonian Remote Sensing Termbase using the Dictionary Writing System Ekilex (Tavast et al., 2018). For each term, we provided definitions, variants, and usage contexts. In the final stage, remote sensing experts reviewed and edited the terms, their variants, and usage contexts. Finally, we provide insights and outline directions for future work in this area.

Thanks to the development of pre-trained sequence-to-sequence (seq2seq) models (e.g., BART), recent studies on AMR parsing often regard this task as a seq2seq translation problem by linearizing AMR graphs into AMR token sequences in pre-processing and recovering AMR graphs from sequences in post-processing. Seq2seq AMR parsing is a relatively simple paradigm but it unavoidably loses structural information among AMR tokens. To compensate for the loss of structural information, in this paper we explicitly leverage AMR structure in the decoding phase. Given an AMR graph, we first project the structure in the graph into an AMR token graph, i.e., structure among AMR tokens in the linearized sequence. The structures for an AMR token could be divided into two parts: structure in prediction history and structure in future. Then we propose to model structure in prediction history via a graph attention network (GAT) and learn structure in future via a multi-task scheme, respectively. Experimental results show that our approach significantly outperforms a strong baseline and achieves performance with 85.5 ±0.1 and 84.2 ±0.1 Smatch scores on AMR 2.0 and AMR 3.0, respectively

Aspect-Based Sentiment Analysis (ABSA) stands as a crucial task in predicting the sentiment polarity associated with identified aspects within text. However, a notable challenge in ABSA lies in precisely determining the aspects’ boundaries (start and end indices), especially for long ones, due to users’ colloquial expressions. We propose DiffusionABSA, a novel diffusion model tailored for ABSA, which extracts the aspects progressively step by step. Particularly, DiffusionABSA gradually adds noise to the aspect terms in the training process, subsequently learning a denoising process that progressively restores these terms in a reverse manner. To estimate the boundaries, we design a denoising neural network enhanced by a syntax-aware temporal attention mechanism to chronologically capture the interplay between aspects and surrounding text. Empirical evaluations conducted on eight benchmark datasets underscore the compelling advantages offered by DiffusionABSA when compared against robust baseline models. Our code is publicly available at https://github.com/Qlb6x/DiffusionABSA.

Bilingual dictionaries present several challenges, especially for sign languages and oral languages, where multimodality plays a role. We deployed and tested the first bilingual Peruvian Sign Language (LSP) - Spanish Online Dictionary. The first feature allows the user to introduce a text and receive as a result a list of videos whose glosses are related to the input text or Spanish word. The second feature allows the user to sign in front of the camera and shows the five most probable Spanish translations based on the similarity between the input sign and gloss-labeled sign videos used to train a machine learning model. These features are constructed in a design and architecture that differentiates among the coincidence for the Spanish text searched, the sign gloss, and Spanish translation. We explain in depth how these concepts or database columns impact the search. Similarly, we share the challenges of deploying a real-world machine learning model for isolated sign language recognition through Amazon Web Services (AWS).

Recent work shows large language models can be prompted to generate useful rationales for commonsense question answering (CQA), which can improve the performance of both themselves and other models. However, the cost of deployment and further tuning is relatively expensive for the large models. Some work explores to distill the the rationale-generation ability to convenient small-sized models, yet it typically requires human-authored QA instances during the distillation. In this paper, we propose a novel framework that leverages both knowledge graphs and large language models to synthesize rationale-augmented CQA data. Based on it, we train Leros, a model that can generate helpful rationales to assist generic QA models to accomplish unseen CQA tasks. Empirical results demonstrate Leros can substantially enhance the performance of QA models on five unseen CQA benchmarks, providing better gains than both same-sized counterpart models trained with downstream data and 10x larger language models. Our work reveals a novel way to integrate knowledge from both knowledge graphs and large language models into smaller models. The codes and synthesized resources are publicly available at https://github.com/wchrepo/leros.

This paper presents preliminary experiments for the lemmatisation of unedited, Byzantine Greek epigrams. This type of Greek is quite different from its classical ancestor, mostly because of its orthographic inconsistencies. Existing lemmatisation algorithms display an accuracy drop of around 30pp when tested on these Byzantine book epigrams. We conducted seven different lemmatisation experiments, which were either transformer-based or based on neural edit-trees. The best performing lemmatiser was a hybrid method combining transformer-based embeddings with a dictionary look-up. We compare our results with existing lemmatisers, and provide a detailed error analysis revealing why unedited, Byzantine Greek is so challenging for lemmatisation.

Resource scarcity in Neural Machine Translation is a challenging problem in both industry applications and in the support of less-spoken languages represented, in the worst case, by endangered and low-resource languages. Many Data Augmentation methods rely on additional linguistic sources and software tools but these are often not available in less favoured language. For this reason, we present USKI (Unaligned Sentences Keytokens pre-traIning), a pre-training strategy that leverages the relationships and similarities that exist between unaligned sentences. By doing so, we increase the dataset size of endangered and low-resource languages by the square of the initial quantity, matching the typical size of high-resource language datasets such as WMT14 En-Fr. Results showcase the effectiveness of our approach with an increase on average of 0.9 BLEU across the benchmarks using a small fraction of the entire unaligned corpus, suggesting the importance of the research topic and the potential of a currently under-utilized resource and under-explored approach.

Argument mining has typically been researched for specific corpora belonging to concrete languages and domains independently in each research work. Human argumentation, however, has domain- and language-dependent linguistic features that determine the content and structure of arguments. Also, when deploying argument mining systems in the wild, we might not be able to control some of these features. Therefore, an important aspect that has not been thoroughly investigated in the argument mining literature is the robustness of such systems to variations in language and domain. In this paper, we present a complete analysis across three different languages and three different domains that allow us to have a better understanding on how to leverage the scarce available corpora to design argument mining systems that are more robust to natural language variations.

Gradient-based explanation methods are increasingly used to interpret neural models in natural language processing (NLP) due to their high fidelity. Such methods determine word-level importance using dimension-level gradient values through a norm function, often presuming equal significance for all gradient dimensions. However, in the context of Aspect-based Sentiment Analysis (ABSA), our preliminary research suggests that only specific dimensions are pertinent. To address this, we propose the Information Bottleneck-based Gradient (IBG) explanation framework for ABSA. This framework leverages an information bottleneck to refine word embeddings into a concise intrinsic dimension, maintaining essential features and omitting unrelated information. Comprehensive tests show that our IBG approach considerably improves both the models’ performance and the explanations’ clarity by identifying sentiment-aware features.

For nearly the past forty years, there has been discussion regarding whether symbolic representations are involved in morphological inflection, a debate commonly known as the Past Tense Debate. The previous literature has extensively explored whether neural models, which do not use symbolic representations can process morphological inflection like humans. However, current research interest has shifted towards whether neural models can acquire morphological inflection like humans. In this paper, we trained neural models, the recurrent neural network (RNN) with attention and the transformer, and a symbolic model, the Minimal Generalization Learner (MGL), under a human-like learning environment. Evaluating the models from the perspective of language acquisition, we found that while the transformer and the MGL exhibited some human-like characteristics, the RNN with attention did not demonstrate human-like behavior across all the evaluation metrics considered in this study. Furthermore, none of the models accurately inflected verbs in the same manner as humans in terms of morphological inflection direction. These results suggest that these models fall short as cognitive models of morphological inflection.

Empathetic leadership communication plays a pivotal role in modern workplaces as it is associated with a wide range of positive individual and organizational outcomes. This paper introduces LeadEmpathy, an innovative expert-annotated German dataset for modeling empathy in written leadership communication. It features a novel theory-based coding scheme to model cognitive and affective empathy in asynchronous communication. The final dataset comprises 770 annotated emails from 385 participants who were allowed to rewrite their emails after receiving recommendations for increasing empathy in an online experiment. Two independent annotators achieved substantial inter-annotator agreement of >= .79 for all categories, indicating that the annotation scheme can be applied to produce high-quality, multidimensional empathy ratings in current and future applications. Beyond outlining the dataset’s development procedures, we present a case study on automatic empathy detection, establishing baseline models for predicting empathy scores in a range of ten possible scores that achieve a Pearson correlation of 0.816 and a mean squared error of 0.883. Our dataset is available at https://github.com/caisa-lab/LEAD-empathy-dataset.

This paper proposes LayoutLLM, a more flexible document analysis method for understanding imaged documents. Visually Rich Document Understanding tasks, such as document image classification and information extraction, have gained significant attention due to their importance. Existing methods have been developed to enhance document comprehension by incorporating pre-training awareness of images, text, and layout structure. However, these methods require fine-tuning for each task and dataset, and the models are expensive to train and operate. To overcome this limitation, we propose a new LayoutLLM that integrates these with large-scale language models (LLMs). By leveraging the strengths of existing research in document image understanding and LLMs’ superior language understanding capabilities, the proposed model, fine-tuned with multimodal instruction datasets, performs an understanding of document images in a single model. Our experiments demonstrate improvement over the baseline model in various document analysis tasks.

Among the various pre-trained neural language models that are popular today, dropout is already an indispensable regularization technique. To solve the inconsistency between training and inference caused by the randomness of dropout, some studies use consistency training to regularize dropout at the output layer. In this paper, we propose a novel Layer-wise Regularized Dropout (LR-Drop), which is specially designed for Transformer-based Language models. Specifically, LR-Drop layer-wise regularizes each Transformer layer using the consistency training strategy. Each training sample passes through the two siamese sub-models sampled by dropout, and then LR-Drop forces the hidden states, multi-head attention matrices, and output distribution of the two siamese sub-models to be consistent. The proposed LR-Drop can be regarded as a “self-distillation” framework, in which each sub-model generated by dropout is the other’s “teacher” model and “student” model. Through extensive experiments on 8 natural language understanding datasets, 6 neural machine translation datasets, and 1 abstractive summarization dataset (a total of 15 datasets), we show that LR-Drop achieves superior performances, including state-of-the-art results.