COLING 2024 论文列表

Detecting synthetically generated text in the wild has become increasingly difficult with advances in Natural Language Generation techniques and the proliferation of freely available Large Language Models (LLMs). Social media and news sites can be flooded with synthetically generated misinformation via tweets and posts while authentic users can inadvertently spread this text via shares and retweets. Most modern natural language processing techniques designed to detect synthetically generated text focus primarily on long-form content, such as news articles, or incorporate stylometric characteristics and metadata during their analysis. Unfortunately, for short form text like tweets, this information is often unavailable, usually detached from its original source, displayed out of context, and is often too short or informal to yield significant information from stylometry. This paper proposes a method of detecting synthetically generated tweets via a Transformer architecture and incorporating unique style-based features. Additionally, we have created a new dataset consisting of human-generated and Large Language Model generated tweets for 4 topics and another dataset consisting of tweets paraphrased by 3 different paraphrase models.

In this paper, we propose a new setting for generating product descriptions from images, augmented by marketing keywords. It leverages the combined power of visual and textual information to create descriptions that are more tailored to the unique features of products. For this setting, previous methods utilize visual and textual encoders to encode the image and keywords and employ a language model-based decoder to generate the product description. However, the generated description is often inaccurate and generic since same-category products have similar copy-writings, and optimizing the overall framework on large-scale samples makes models concentrate on common words yet ignore the product features. To alleviate the issue, we present a simple and effective Multimodal In-Context Tuning approach, named ModICT, which introduces a similar product sample as the reference and utilizes the in-context learning capability of language models to produce the description. During training, we keep the visual encoder and language model frozen, focusing on optimizing the modules responsible for creating multimodal in-context references and dynamic prompts. This approach preserves the language generation prowess of large language models (LLMs), facilitating a substantial increase in description diversity. To assess the effectiveness of ModICT across various language model scales and types, we collect data from three distinct product categories within the E-commerce domain. Extensive experiments demonstrate that ModICT significantly improves the accuracy (by up to 3.3% on Rouge-L) and diversity (by up to 9.4% on D-5) of generated results compared to conventional methods. Our findings underscore the potential of ModICT as a valuable tool for enhancing the automatic generation of product descriptions in a wide range of applications. Data and code are at https://github.com/HITsz-TMG/Multimodal-In-Context-Tuning

The automatic translation of spoken language into pictogram units can facilitate communication involving individuals with language impairments. However, there is no established translation formalism or publicly available datasets for training end-to-end speech translation systems. This paper introduces the first aligned speech, text, and pictogram translation dataset ever created in any language. We provide a French dataset that contains 230 hours of speech resources. We create a rule-based pictogram grammar with a restricted vocabulary and include a discussion of the strategic decisions involved. It takes advantage of an in-depth linguistic study of resources taken from the ARASAAC website. We validate these rules through multiple post-editing phases by expert annotators. The constructed dataset is then used to experiment with a Speech-to-Pictogram cascade model, which employs state-of-the-art Automatic Speech Recognition models. The dataset is freely available under a non-commercial licence. This marks a starting point to conduct research into the automatic translation of speech into pictogram units.

We report the creation of the first high-quality corpus of Aromanian - an endangered Romance language spoken in the Balkans - and the equivalent sentence-aligned translations into Romanian, English, and French. The corpus is released publicly using several orthographic standards and consists in short stories collected in the ‘70s in Romania. Additionally, we provide an corpus-based analysis of Aromanian linguistic particularities and the overall demographic and political context which impacts the contemporary development of the language.

In this paper, we explore physical commonsense reasoning of large language models (LLMs) and propose a specific methodology to evaluate low-level understanding of the physical world. Specifically, the goal is to create a test set to analyze physical commonsense reasoning in large language models for Italian and focus on a trustworthy analysis of the results. To that end, we present a tiered Italian dataset, called Graded Italian Annotated dataset (GITA), written and thoroughly annotated by a professional linguist, which allows us to concentrate on three different levels of commonsense understanding. Moreover, we create a semi-automated system to complete the accurate annotation of the dataset. We also validate our dataset by carrying out three tasks with a multilingual model (XLM-RoBERTa) and propose a qualitative analysis of the results. We found out that, although the model may perform at high-level classification tasks, its easoning is inconsistent and unverifiable, since it does not capture intermediate evidence.

Precise understanding of free-text radiology reports through localised extraction of clinical findings can enhance medical imaging applications like computer-aided diagnosis. We present a new task, that of segmenting radiology reports into topically meaningful passages (segments) and a transformer-based model that both segments reports into semantically coherent segments and classifies each segment using a set of 37 radiological abnormalities, thus enabling fine-grained analysis. This contrasts with prior work that performs classification on full reports without localisation. Trained on over 2.7 million unlabelled chest X-ray reports and over 28k segmented and labelled reports, our model achieves state-of-the-art performance on report segmentation (0.0442 WinDiff) and multi-label classification (0.84 report-level macro F1) over 37 radiological labels and 8 NLP-specific labels. This work establishes new benchmarks for fine-grained understanding of free-text radiology reports, with precise localisation of semantics unlocking new opportunities to improve computer vision model training and clinical decision support. We open-source our annotation tool, model code and pretrained weights to encourage future research.

We present a corpus of 100 documents, named OBSINFOX, selected from 17 sources of French press considered unreliable by expert agencies, annotated using 11 labels by 8 annotators. By collecting more labels than usual, by more annotators than is typically done, we can identify features that humans consider as characteristic of fake news, and compare them to the predictions of automated classifiers. We present a topic and genre analysis using Gate Cloud, indicative of the prevalence of satire-like text in the corpus. We then use the subjectivity analyzer VAGO, and a neural version of it, to clarify the link between ascriptions of the label Subjective and ascriptions of the label Fake News. The annotated dataset is available online at the following url: https://github.com/obs-info/obsinfox Keywords: Fake News, Multi-Labels, Subjectivity, Vagueness, Detail, Opinion, Exaggeration, French Press

Studies on literary quality have constantly stimulated the interest of critics, both in theoretical and empirical fields. To examine the perceived quality of literary works, some approaches have focused on data annotated through crowd-sourcing platforms, and others relied on available expert annotated data. In this work, we contribute to the debate by presenting a dataset collecting quality judgments on 9,000 19th and 20th century English-language literary novels by 3,150 predominantly Anglophone authors. We incorporate expert opinions and crowd-sourced annotations to allow comparative analyses between different literary quality evaluations. We also provide several textual metrics chosen for their potential connection with literary reception and engagement. While a large part of the texts is subjected to copyright, we release quality and reception measures together with stylometric and sentiment data for each of the 9,000 novels to promote future research and comparison.

According to the United Nations Development Programme, gender inequality is a metric that is composed of three dimensions: reproductive health, empowerment, and the labour market. Gender inequality is an obstacle to equal opportunities in society as a whole. In this paper we present our work-in-progress of designing and playing a physical game with digital elements. We currently conduct Conversation Analysis of transcribed speech of 58567 words and documenting bias. We also test OpenAI’s ChatGPT for bias in quiz-like gender-related questions.

Generating coherent and credible explanations remains a significant challenge in the field of AI. In recent years, researchers have delved into the utilization of entailment trees to depict explanations, which exhibit a reasoning process of how a hypothesis is deduced from the supporting facts. However, existing models often overlook the importance of generating intermediate conclusions with logical consistency from the given facts, leading to inaccurate conclusions and undermining the overall credibility of entailment trees. To address this limitation, we propose the logical pattern memory pre-trained model (LMPM). LMPM incorporates an external memory structure to learn and store the latent representations of logical patterns, which aids in generating logically consistent conclusions. Furthermore, to mitigate the influence of logically irrelevant domain knowledge in the Wikipedia-based data, we introduce an entity abstraction approach to construct the dataset for pre-training LMPM. The experimental results highlight the effectiveness of our approach in improving the quality of entailment tree generation. By leveraging logical entailment patterns, our model produces more coherent and reasonable conclusions that closely align with the underlying premises.

Semantic role labeling is an essential component of semantic and syntactic processing of natural languages, which reveals the predicate-argument structure of the language. Despite its importance, semantic role labeling for the Korean language has not been studied extensively. One notable issue is the lack of uniformity among data annotation strategies across different datasets, which often lack thorough rationales. In this study, we suggest an annotation strategy for Korean semantic role labeling that is in line with the previously proposed linguistic theories as well as the distinct properties of the Korean language. We further propose a simple yet viable conversion strategy from the Sejong verb dictionary to a CoNLL-style dataset for Korean semantic role labeling. Experiment results using a transformer-based sequence labeling model demonstrate the reliability and trainability of the converted dataset.

Preparing historical newspaper collections is a complicated endeavour, consisting of multiple steps that have to be carefully adapted to the specific content in question, including imaging, layout prediction, optical character recognition, and linguistic annotation. To address the high costs associated with the process, we present a lightweight approach to producing high-quality corpora and apply it to a massive collection of Slovenian historical newspapers from the 18th, 19th and 20th century resulting in a billion-word giga-corpus. We start with noisy OCR-ed data produced by different technologies in varying periods by the National and University Library of Slovenia. To address the inherent variability in the quality of textual data, a challenge commonly encountered in digital libraries globally, we perform a targeted post-digitisation correction procedure, coupled with a robust curation mechanism for noisy texts via language model inference. Subsequently, we subject the corrected and filtered output to comprehensive linguistic annotation, enriching the corpus with part-of-speech tags, lemmas, and named entity labels. Finally, we perform an analysis through topic modeling at the noun lemma level, along with a frequency analysis of the named entities, to confirm the viability of our corpus preparation method.

Cross-lingual pre-training methods mask and predict tokens in multilingual text to generalize diverse multilingual information. However, due to the lack of sufficient aligned multilingual resources in the pre-training process, these methods may not fully explore the multilingual correlation of masked tokens, resulting in the limitation of multilingual information interaction. In this paper, we propose a lifelong multilingual multi-granularity semantic alignment approach, which continuously extracts massive aligned linguistic units from noisy data via a maximum co-occurrence probability algorithm. Then, the approach releases a version of the multilingual multi-granularity semantic alignment resource, supporting seven languages, namely English, Czech, German, Russian, Romanian, Hindi and Turkish. Finally, we propose how to use this resource to improve the translation performance on WMT14 18 benchmarks in twelve directions. Experimental results show an average of 0.3 1.1 BLEU improvements in all translation benchmarks. The analysis and discussion also demonstrate the superiority and potential of the proposed approach. The resource used in this work will be publicly available.

This paper introduces a large annotated corpus of public domain German poetry, covering the time period from 1600 to the 1920s with 65k poems. We describe how the corpus was compiled, how it was cleaned (including duplicate detection), and how it looks now in terms of size, format, temporal distribution, and automatic annotation. Besides metadata, the corpus contains reliable annotation of tokens, syllables, part-of-speech, and meter and verse measure. Finally, we give some statistics on the annotation and an overview of other poetry corpora.

Knowledge-based, open-domain dialogue generation aims to build chit-chat systems that talk to humans using mined support knowledge. Many types and sources of knowledge have previously been shown to be useful as support knowledge. Even in the era of large language models, response generation grounded in knowledge retrieved from additional up-to-date sources remains a practically important approach. While prior work using single-source knowledge has shown a clear positive correlation between the performances of knowledge selection and response generation, there are no existing multi-source datasets for evaluating support knowledge retrieval. Further, prior work has assumed that the knowledge sources available at test time are the same as during training. This unrealistic assumption unnecessarily handicaps models, as new knowledge sources can become available after a model is trained. In this paper, we present a high-quality benchmark named multi-source Wizard of Wikipedia (Ms.WoW) for evaluating multi-source dialogue knowledge selection and response generation. Unlike existing datasets, it contains clean support knowledge, grounded at the utterance level and partitioned into multiple knowledge sources. We further propose a new challenge, dialogue knowledge plug-and-play, which aims to test an already trained dialogue model on using new support knowledge from previously unseen sources in a zero-shot fashion.

Text Simplification enhances the readability of texts for specific audiences. However, automated models may introduce unwanted content or omit essential details, necessitating a focus on maintaining faithfulness to the original input. Furthermore, existing simplified corpora contain instances of low faithfulness. Motivated by this issue, we present a new Japanese simplification corpus designed to prioritize faithfulness. Our collection comprises 7,075 paired sentences simplified from newspaper articles. This process involved collaboration with language education experts who followed guidelines balancing readability and faithfulness. Through corpus analysis, we confirmed that our dataset preserves the content of the original text, including personal names, dates, and city names. Manual evaluation showed that our corpus robustly maintains faithfulness to the original text, surpassing other existing corpora. Furthermore, evaluation by non-native readers confirmed its readability to the target audience. Through the experiment of fine-tuning and in-context learning, we demonstrated that our corpus enhances faithful sentence simplification.

Aspect-Based Sentiment Analysis ( ABSA) aims to identify terms or multiword expressions (MWEs) on which sentiments are expressed and the sentiment polarities associated with them. The development of supervised models has been at the forefront of research in this area. However, training these models requires the availability of manually annotated datasets which is both expensive and time-consuming. Furthermore, the available annotated datasets are tailored to a specific domain, language, and text type. In this work, we address this notable challenge in current state-of-the-art ABSA research. We propose a hybrid approach for Aspect Based Sentiment Analysis using transfer learning. The approach focuses on generating weakly-supervised annotations by exploiting the strengths of both large language models (LLM) and traditional syntactic dependencies. We utilise syntactic dependency structures of sentences to complement the annotations generated by LLMs, as they may overlook domain-specific aspect terms. Extensive experimentation on multiple datasets is performed to demonstrate the efficacy of our hybrid method for the tasks of aspect term extraction and aspect sentiment classification.

This paper introduces TVB-HKSL-News, a new Hong Kong sign language (HKSL) dataset collected from a TV news program over a period of 7 months. The dataset is collected to enrich resources for HKSL and support research in large-vocabulary continuous sign language recognition (SLR) and translation (SLT). It consists of 16.07 hours of sign videos of two signers with a vocabulary of 6,515 glosses (for SLR) and 2,850 Chinese characters or 18K Chinese words (for SLT). One signer has 11.66 hours of sign videos and the other has 4.41 hours. One objective in building the dataset is to support the investigation of how well large-vocabulary continuous sign language recognition/translation can be done for a single signer given a (relatively) large amount of his/her training data, which could potentially lead to the development of new modeling methods. Besides, most parts of the data collection pipeline are automated with little human intervention; we believe that our collection method can be scaled up to collect more sign language data easily for SLT in the future for any sign languages if such sign-interpreted videos are available. We also run a SOTA SLR/SLT model on the dataset and get a baseline SLR word error rate of 34.08% and a baseline SLT BLEU-4 score of 23.58 for benchmarking future research on the dataset.

Aspect category sentiment analysis (ACSA) aims to simultaneously detect aspect categories and their corresponding sentiment polarities (category-sentiment pairs). Some recent studies have used pre-trained generative models to complete ACSA and achieved good results. However, for ACSA, generative models still face three challenges. First, addressing the missing predictions in ACSA is crucial, which involves accurately predicting all category-sentiment pairs within a sentence. Second, category-sentiment pairs are inherently a disordered set. Consequently, the model incurs a penalty even when its predictions are correct, but the predicted order is inconsistent with the ground truths. Third, different aspect categories should focus on relevant sentiment words, and the polarity of the aspect category should be the aggregation of the polarities of these sentiment words. This paper proposes a hierarchical generative model with a coverage mechanism using sequence-to-set learning to tackle all three challenges simultaneously. Our model’s superior performance is demonstrated through extensive experiments conducted on several datasets.