COLING 2024 论文列表

Recent advances in generative language modeling applied to discrete speech tokens presented a new avenue for text-to-speech (TTS) synthesis. These speech language models (SLMs), similarly to their textual counterparts, are scalable, probabilistic, and context-aware. While they can produce diverse and natural outputs, they sometimes face issues such as unintelligibility and the inclusion of non-speech noises or hallucination. As the adoption of this innovative paradigm in speech synthesis increases, there is a clear need for an in-depth evaluation of its capabilities and limitations. In this paper, we evaluate TTS from a discrete token-based SLM, through both automatic metrics and listening tests. We examine five key dimensions: speaking style, intelligibility, speaker consistency, prosodic variation, spontaneous behaviour. Our results highlight the model’s strength in generating varied prosody and spontaneous outputs. It is also rated higher in naturalness and context appropriateness in listening tests compared to a conventional TTS. However, the model’s performance in intelligibility and speaker consistency lags behind traditional TTS. Additionally, we show that increasing the scale of SLMs offers a modest boost in robustness. Our findings aim to serve as a benchmark for future advancements in generative SLMs for speech synthesis.

Modern contextual lemmatizers often rely on automatically induced Shortest Edit Scripts (SES), namely, the number of edit operations to transform a word form into its lemma. In fact, different methods of computing SES have been proposed as an integral component in the architecture of several state-of-the-art contextual lemmatizers currently available. However, previous work has not investigated the direct impact of SES in the final lemmatization performance. In this paper we address this issue by focusing on lemmatization as a token classification task where the only input that the model receives is the word-label pairs in context, where the labels correspond to previously induced SES. Thus, by modifying in our lemmatization system only the SES labels that the model needs to learn, we may then objectively conclude which SES representation produces the best lemmatization results. We experiment with seven languages of different morphological complexity, namely, English, Spanish, Basque, Russian, Czech, Turkish and Polish, using multilingual and language-specific pre-trained masked language encoder-only models as a backbone to build our lemmatizers. Comprehensive experimental results, both in- and out-of-domain, indicate that computing the casing and edit operations separately is beneficial overall, but much more clearly for languages with high-inflected morphology. Notably, multilingual pre-trained language models consistently outperform their language-specific counterparts in every evaluation setting.

The application of self-supervision to speech representation learning has garnered significant interest in recent years, due to its scalability to large amounts of unlabeled data. However, much progress, both in terms of pre-training and downstream evaluation, has remained concentrated in monolingual models that only consider English. Few models consider other languages, and even fewer consider indigenous ones. In this work, benchmark the efficacy of large SSL models on 6 indigenous America languages: Quechua, Guarani , Bribri, Kotiria, Wa’ikhana, and Totonac on low-resource ASR. Our results show surprisingly strong performance by state-of-the-art SSL models, showing the potential generalizability of large-scale models to real-world data.

Deep learning models have performed well on many NLP tasks. However, their internal mechanisms are typically difficult for humans to understand. The development of methods to explain models has become a key issue in the reliability of deep learning models in many important applications. Various saliency explanation methods, which give each feature of input a score proportional to the contribution of output, have been proposed to determine the part of the input which a model values most. Despite a considerable body of work on the evaluation of saliency methods, whether the results of various evaluation metrics agree with human cognition remains an open question. In this study, we propose a new human-based method to evaluate saliency methods in NLP by crowdsourcing. We recruited 800 crowd workers and empirically evaluated seven saliency methods on two datasets with the proposed method. We analyzed the performance of saliency methods, compared our results with existing automated evaluation methods, and identified notable differences between NLP and computer vision (CV) fields when using saliency methods. The instance-level data of our crowdsourced experiments and the code to reproduce the explanations are available at https://github.com/xtlu/lreccoling_evaluation.

This paper proposes an analysis of prompting strategies for grammatical error correction (GEC) with selected large language models (LLM) based on language proficiency. GEC using generative LLMs has been known for overcorrection where results obtain higher recall measures than precision measures. The writing examples of English language learners may be different from those of native speakers. Given that there is a significant differences in second language (L2) learners’ error types by their proficiency levels, this paper attempts to reduce overcorrection by examining the interaction between LLM’s performance and L2 language proficiency. Our method focuses on zero-shot and few-shot prompting and fine-tuning models for GEC for learners of English as a foreign language based on the different proficiency. We investigate GEC results and find that overcorrection happens primarily in advanced language learners’ writing (proficiency C) rather than proficiency A (a beginner level) and proficiency B (an intermediate level). Fine-tuned LLMs, and even few-shot prompting with writing examples of English learners, actually tend to exhibit decreased recall measures. To make our claim concrete, we conduct a comprehensive examination of GEC outcomes and their evaluation results based on language proficiency.

Word embeddings and Language models are the building blocks of modern Deep Neural Network-based Natural Language Processing. They are extensively explored in high-resource languages and provide state-of-the-art (SOTA) performance for a wide range of downstream tasks. Nevertheless, these word embeddings are not explored in languages such as Assamese, where resources are limited. Furthermore, there has been limited study into the performance evaluation of these word embeddings for low-resource languages in downstream tasks. In this research, we explore the current state of Assamese pre-trained word embeddings. We evaluate these embeddings’ performance on sequence labeling tasks such as Parts-of-speech and Named Entity Recognition. In order to assess the efficiency of the embeddings, experiments are performed utilizing both ensemble and individual word embedding approaches. The ensembling approach that uses three word embeddings outperforms the others. In the paper, the outcomes of the investigations are described. The results of this comparative performance evaluation may assist researchers in choosing an Assamese pre-trained word embedding for subsequent tasks.

Modern Large Language Models (LLMs) have showcased remarkable prowess in various tasks necessitating sophisticated cognitive behaviors. Nevertheless, a paradoxical performance discrepancy is observed, where these models underperform in seemingly elementary tasks like relation extraction and event extraction due to two issues in conventional evaluation. (1) The imprecision of existing evaluation metrics that struggle to effectively gauge semantic consistency between model outputs and ground truth, and (2) The inherent incompleteness of evaluation benchmarks, primarily due to restrictive human annotation schemas, resulting in underestimated LLM performances. Inspired by the principles in subjective question correction, we propose a new evaluation method, SQC-Score. This method innovatively utilizes LLMs, fine-tuned through subjective question correction data, to refine matching between model outputs and golden labels. Additionally, by incorporating a Natural Language Inference (NLI) model, SQC-Score enriches golden labels, addressing benchmark incompleteness by acknowledging correct yet previously omitted answers. Results on three information extraction tasks show that SQC-Score is more preferred by human annotators than the baseline metrics. Utilizing SQC-Score, we conduct a comprehensive evaluation of the state-of-the-art LLMs and provide insights for future research for information extraction. Dataset and associated codes can be accessed at our <a href=https://github.com/THU-KEG/SQC-Score> GitHub repository </a>.

Discriminatory gender biases have been found in Pre-trained Language Models (PLMs) for multiple languages. In Natural Language Inference (NLI), existing bias evaluation methods have focused on the prediction results of one specific label out of three labels, such as neutral. However, such evaluation methods can be inaccurate since unique biased inferences are associated with unique prediction labels. Addressing this limitation, we propose a bias evaluation method for PLMs, called NLI-CoAL, which considers all the three labels of NLI task. First, we create three evaluation data groups that represent different types of biases. Then, we define a bias measure based on the corresponding label output of each data group. In the experiments, we introduce a meta-evaluation technique for NLI bias measures and use it to confirm that our bias measure can distinguish biased, incorrect inferences from non-biased incorrect inferences better than the baseline, resulting in a more accurate bias evaluation. We create the datasets in English, Japanese, and Chinese, and successfully validate the compatibility of our bias measure across multiple languages. Lastly, we observe the bias tendencies in PLMs of different languages. To our knowledge, we are the first to construct evaluation datasets and measure PLMs’ bias from NLI in Japanese and Chinese.

Recent advances in large language models (LLMs) have shown they can match or surpass finetuned models on many natural language processing tasks. Currently, more studies are being carried out to assess whether this performance carries over across different languages. In this paper, we present a thorough evaluation of LLMs for the less well-researched code-switching translation setting, where inputs include a mixture of different languages. We benchmark the performance of six state-of-the-art LLMs across seven datasets, with GPT-4 and GPT-3.5 displaying strong ability relative to supervised translation models and commercial engines. GPT-4 was also found to be particularly robust against different code-switching conditions. Several methods to further improve code-switching translation are proposed including leveraging in-context learning and pivot translation. Through our code-switching experiments, we argue that LLMs show promising ability for cross-lingual understanding.

Large language models like ChatGPT have recently shown a great promise in performing several tasks, including hate speech detection. However, it is crucial to comprehend the limitations of these models to build robust hate speech detection systems. To bridge this gap, our study aims to evaluate the strengths and weaknesses of the ChatGPT model in detecting hate speech at a granular level across 11 languages. Our evaluation employs a series of functionality tests that reveals various intricate failures of the model which the aggregate metrics like macro F1 or accuracy are not able to unfold. In addition, we investigate the influence of complex emotions, such as the use of emojis in hate speech, on the performance of the ChatGPT model. Our analysis highlights the shortcomings of the generative models in detecting certain types of hate speech and highlighting the need for further research and improvements in the workings of these models.

Systems that automatically generate subtitles from video are gradually entering subtitling workflows, both for supporting subtitlers and for accessibility purposes. Even though robust metrics are essential for evaluating the quality of automatically-generated subtitles and for estimating potential productivity gains, there is limited research on whether existing metrics, some of which directly borrowed from machine translation (MT) evaluation, can fulfil such purposes. This paper investigates how well such MT metrics correlate with measures of post-editing (PE) effort in automatic subtitling. To this aim, we collect and publicly release a new corpus containing product-, process- and participant-based data from post-editing automatic subtitles in two language pairs (en→de,it). We find that different types of metrics correlate with different aspects of PE effort. Specifically, edit distance metrics have high correlation with technical and temporal effort, while neural metrics correlate well with PE speed.

The European Language Grid (ELG) is a cloud platform for the whole European Language Technology community. While the EU project that developed the platform successfully concluded in June 2022, the ELG initiative has continued. This article provides a description of the current state of ELG in terms of user adoption and number of language resources and technologies available in early 2024. It also provides an overview of the various activities with regard to ELG since the end of the project and since the publication of the ELG book, especially the co-authors’ attempt to integrate the ELG platform into various data space initiatives. The article also provides an overview of the Digital Language Equality (DLE) dashboard and the current state of DLE in Europe.

Large language models (LLMs) have gained popularity recently due to their outstanding performance in various downstream Natural Language Processing (NLP) tasks. However, low-resource languages are still lagging behind current state-of-the-art (SOTA) developments in the field of NLP due to insufficient resources to train LLMs. Ethiopian languages exhibit remarkable linguistic diversity, encompassing a wide array of scripts, and are imbued with profound religious and cultural significance. This paper introduces EthioLLM – multilingual large language models for five Ethiopian languages (Amharic, Ge’ez, Afan Oromo, Somali, and Tigrinya) and English, and Ethiobenchmark – a new benchmark dataset for various downstream NLP tasks. We evaluate the performance of these models across five downstream NLP tasks. We open-source our multilingual language models, new benchmark datasets for various downstream tasks, and task-specific fine-tuned language models and discuss the performance of the models. Our dataset and models are available at the https://huggingface.co/EthioNLP repository.

Ethical reasoning is a crucial skill for Large Language Models (LLMs). However, moral values are not universal, but rather influenced by language and culture. This paper explores how three prominent LLMs – GPT-4, ChatGPT, and Llama2Chat-70B – perform ethical reasoning in different languages and if their moral judgement depend on the language in which they are prompted. We extend the study of ethical reasoning of LLMs by (CITATION) to a multilingual setup following their framework of probing LLMs with ethical dilemmas and policies from three branches of normative ethics: deontology, virtue, and consequentialism. We experiment with six languages: English, Spanish, Russian, Chinese, Hindi, and Swahili. We find that GPT-4 is the most consistent and unbiased ethical reasoner across languages, while ChatGPT and Llama2Chat-70B show significant moral value bias when we move to languages other than English. Interestingly, the nature of this bias significantly vary across languages for all LLMs, including GPT-4.

Rigorous evaluation of the causal effects of semantic features on language model predictions can be hard to achieve for natural language reasoning problems. However, this is such a desirable form of analysis from both an interpretability and model evaluation perspective, that it is valuable to investigate specific patterns of reasoning with enough structure and regularity to identify and quantify systematic reasoning failures in widely-used models. In this vein, we pick a portion of the NLI task for which an explicit causal diagram can be systematically constructed: the case where across two sentences (the premise and hypothesis), two related words/terms occur in a shared context. In this work, we apply causal effect estimation strategies to measure the effect of context interventions (whose effect on the entailment label is mediated by the semantic monotonicity characteristic) and interventions on the inserted word-pair (whose effect on the entailment label is mediated by the relation between these words). Extending related work on causal analysis of NLP models in different settings, we perform an extensive interventional study on the NLI task to investigate robustness to irrelevant changes and sensitivity to impactful changes of Transformers. The results strongly bolster the fact that similar benchmark accuracy scores may be observed for models that exhibit very different behaviour. Moreover, our methodology reinforces previously suspected biases from a causal perspective, including biases in favour of upward-monotone contexts and ignoring the effects of negation markers.

Existing methods for complexity estimation are typically developed for entire documents. This limitation in scope makes them inapplicable for shorter pieces of text, such as health assessment tools. These typically consist of lists of independent sentences, all of which are too short for existing methods to apply. The choice of wording in these assessment tools is crucial, as both the cognitive capacity and the linguistic competency of the intended patient groups could vary substantially. As a first step towards creating better tools for supporting health practitioners, we develop a two-step approach for estimating lexical complexity that does not rely on any pre-annotated data. We implement our approach for the Norwegian language and verify its effectiveness using statistical testing and a qualitative evaluation of samples from real assessment tools. We also investigate the relationship between our complexity measure and certain features typically associated with complexity in the literature, such as word length, frequency, and the number of syllables.

Neural machine translation requires large number of parallel sentences along with in-domain parallel data to attain best results. Nevertheless, no scientific parallel corpus for English-Persian language pair is available. In this paper, a parallel corpus called Esposito is introduced, which contains 3.5 million parallel sentences in the scientific domain for English-Persian language pair. In addition, we present a manually validated scientific test set that might serve as a baseline for future studies. We show that a system trained using Esposito along with other publicly available data improves the baseline on average by 7.6 and 8.4 BLEU scores for En->Fa and Fa->En directions, respectively. Additionally, domain analysis using the 5-gram KenLM model revealed notable distinctions between our parallel corpus and the existing generic parallel corpus. This dataset will be available to the public upon the acceptance of the paper.

Acceptability is one of the General Language Understanding Evaluation Benchmark (GLUE) probing tasks proposed to assess the linguistic capabilities acquired by a deep-learning transformer-based language model (LM). In this paper, we introduce the Spanish Corpus of Linguistic Acceptability EsCoLA. EsCoLA has been developed following the example of other linguistic acceptability data sets for English, Italian, Norwegian or Russian, with the aim of having a complete GLUE benchmark for Spanish. EsCoLA consists of 11,174 sentences and their acceptability judgements as found in well-known Spanish reference grammars. Additionally, all sentences have been annotated with the class of linguistic phenomenon the sentence is an example of, also following previous practices. We also provide as task baselines the results of fine-tuning four different language models with this data set and the results of a human annotation experiment. Results are also analyzed and commented to guide future research. EsCoLA is released under a CC-BY 4.0 license and freely available at https://doi.org/10.34810/data1138.

Chinese Spelling Check (CSC) aims to detect and correct error tokens in Chinese contexts, which has a wide range of applications. However, it is confronted with the challenges of insufficient annotated data and the issue that previous methods may actually not fully leverage the existing datasets. In this paper, we introduce our plug-and-play retrieval method with error-robust information for Chinese Spelling Check (RERIC), which can be directly applied to existing CSC models. The datastore for retrieval is built completely based on the training data, with elaborate designs according to the characteristics of CSC. Specifically, we employ multimodal representations that fuse phonetic, morphologic, and contextual information in the calculation of query and key during retrieval to enhance robustness against potential errors. Furthermore, in order to better judge the retrieved candidates, the n-gram surrounding the token to be checked is regarded as the value and utilized for specific reranking. The experiment results on the SIGHAN benchmarks demonstrate that our proposed method achieves substantial improvements over existing work.