COLING 2024 Papers

Although several methods were proposed to address the problem of automated essay scoring (AES) in the last 50 years, there is still much to desire in terms of effectiveness. Large Language Models (LLMs) are transformer-based models that demonstrate extraordinary capabilities on various tasks. In this paper, we test the ability of LLMs, given their powerful linguistic knowledge, to analyze and effectively score written essays. We experimented with two popular LLMs, namely ChatGPT and Llama. We aim to check if these models can do this task and, if so, how their performance is positioned among the state-of-the-art (SOTA) models across two levels, holistically and per individual writing trait. We utilized prompt-engineering tactics in designing four different prompts to bring their maximum potential on this task. Our experiments conducted on the ASAP dataset revealed several interesting observations. First, choosing the right prompt depends highly on the model and nature of the task. Second, the two LLMs exhibited comparable average performance in AES, with a slight advantage for ChatGPT. Finally, despite the performance gap between the two LLMs and SOTA models in terms of predictions, they provide feedback to enhance the quality of the essays, which can potentially help both teachers and students.

Natural language offers an appealing alternative to formal logics as a vehicle for representing knowledge. However, using natural language means that standard methods for automated reasoning can no longer be used. A popular solution is to use transformer-based language models (LMs) to directly reason about knowledge expressed in natural language, but this has two important limitations. First, the set of premises is often too large to be directly processed by the LM. This means that we need a retrieval strategy which can select the most relevant premises when trying to infer some conclusion. Second, LMs have been found to learn shortcuts and thus lack robustness, putting in doubt to what extent they actually understand the knowledge that is expressed. Given these limitations, we explore the following alternative: rather than using LMs to perform reasoning directly, we use them to learn embeddings of individual assertions. Reasoning is then carried out by manipulating the learned embeddings. We show that this strategy is feasible to some extent, while at the same time also highlighting the limitations of directly fine-tuning LMs to learn the required embeddings.

Textual domain is a crucial property within the Natural Language Processing (NLP) community due to its effects on downstream model performance. The concept itself is, however, loosely defined and, in practice, refers to any non-typological property, such as genre, topic, medium or style of a document. We investigate the core notion of domains via human proficiency in identifying related intrinsic textual properties, specifically the concepts of genre (communicative purpose) and topic (subject matter). We publish our annotations in TGeGUM: A collection of 9.1k sentences from the GUM dataset (Zeldes, 2017) with single sentence and larger context (i.e., prose) annotations for one of 11 genres (source type), and its topic/subtopic as per the Dewey Decimal library classification system (Dewey, 1979), consisting of 10/100 hierarchical topics of increased granularity. Each instance is annotated by three annotators, for a total of 32.7k annotations, allowing us to examine the level of human disagreement and the relative difficulty of each annotation task. With a Fleiss’ kappa of at most 0.53 on the sentence level and 0.66 at the prose level, it is evident that despite the ubiquity of domains in NLP, there is little human consensus on how to define them. By training classifiers to perform the same task, we find that this uncertainty also extends to NLP models.

The use of propaganda has spiked on mainstream and social media, aiming to manipulate or mislead users. While efforts to automatically detect propaganda techniques in textual, visual, or multimodal content have increased, most of them primarily focus on English content. The majority of the recent initiatives targeting medium to low-resource languages produced relatively small annotated datasets, with a skewed distribution, posing challenges for the development of sophisticated propaganda detection models. To address this challenge, we carefully develop the largest propaganda dataset to date, ArPro, comprised of 8K paragraphs from newspaper articles, labeled at the text span level following a taxonomy of 23 propagandistic techniques. Furthermore, our work offers the first attempt to understand the performance of large language models (LLMs), using GPT-4, for fine-grained propaganda detection from text. Results showed that GPT-4’s performance degrades as the task moves from simply classifying a paragraph as propagandistic or not, to the fine-grained task of detecting propaganda techniques and their manifestation in text. Compared to models fine-tuned on the dataset for propaganda detection at different classification granularities, GPT-4 is still far behind. Finally, we evaluate GPT-4 on a dataset consisting of six other languages for span detection, and results suggest that the model struggles with the task across languages. We made the dataset publicly available for the community.

If a person firmly believes in a non-factual statement, such as “The Earth is flat”, and argues in its favor, there is no inherent intention to deceive. As the argumentation stems from genuine belief, it may be unlikely to exhibit the linguistic properties associated with deception or lying. This interplay of factuality, personal belief, and intent to deceive remains an understudied area. Disentangling the influence of these variables in argumentation is crucial to gain a better understanding of the linguistic properties attributed to each of them. To study the relation between deception and factuality, based on belief, we present the DeFaBel corpus, a crowd-sourced resource of belief-based deception. To create this corpus, we devise a study in which participants are instructed to write arguments supporting statements like “eating watermelon seeds can cause indigestion”, regardless of its factual accuracy or their personal beliefs about the statement. In addition to the generation task, we ask them to disclose their belief about the statement. The collected instances are labelled as deceptive if the arguments are in contradiction to the participants’ personal beliefs. Each instance in the corpus is thus annotated (or implicitly labelled) with personal beliefs of the author, factuality of the statement, and the intended deceptiveness. The DeFaBel corpus contains 1031 texts in German, out of which 643 are deceptive and 388 are non-deceptive. It is the first publicly available corpus for studying deception in German. In our analysis, we find that people are more confident in the persuasiveness of their arguments when the statement is aligned with their belief, but surprisingly less confident when they are generating arguments in favor of facts. The DeFaBel corpus can be obtained from https://www.ims.uni-stuttgart.de/data/defabel .

Clinical data in hospitals are increasingly accessible for research through clinical data warehouses. However these documents are unstructured and it is therefore necessary to extract information from medical reports to conduct clinical studies. Transfer learning with BERT-like models such as CamemBERT has allowed major advances for French, especially for named entity recognition. However, these models are trained for plain language and are less efficient on biomedical data. Addressing this gap, we introduce CamemBERT-bio, a dedicated French biomedical model derived from a new public French biomedical dataset. Through continual pre-training of the original CamemBERT, CamemBERT-bio achieves an improvement of 2.54 points of F1-score on average across various biomedical named entity recognition tasks, reinforcing the potential of continual pre-training as an equally proficient yet less computationally intensive alternative to training from scratch. Additionally, we highlight the importance of using a standard evaluation protocol that provides a clear view of the current state-of-the-art for French biomedical models.

We present Camel Morph MSA, the largest open-source Modern Standard Arabic morphological analyzer and generator. Camel Morph MSA has over 100K lemmas, and includes rarely modeled morphological features of Modern Standard Arabic with Classical Arabic origins. Camel Morph MSA can produce ∼1.45B analyses and ∼535M unique diacritizations, almost an order of magnitude larger than SAMA (Maamouri et al., 2010c), in addition to having ∼36% less OOV rate than SAMA on a 10B word corpus. Furthermore, Camel Morph MSA fills the gaps of many lemma paradigms by modeling linguistic phenomena consistently. Camel Morph MSA seamlessly integrates with the Camel Tools Python toolkit (Obeid et al., 2020), ensuring ease of use and accessibility.

Recent advancements in large language models (LLMs) and their emergent capabilities make LLM a promising reference-free evaluator on the quality of natural language generation, and a competent alternative to human evaluation. However, hindered by the closed-source or high computational demand to host and tune, there is a lack of practice to further calibrate an off-the-shelf LLM-based evaluator towards better human alignment. In this work, we propose AutoCalibrate, a multi-stage, gradient-free approach to automatically calibrate and align an LLM-based evaluator toward human preference. Instead of explicitly modeling human preferences, we first implicitly encompass them within a set of human labels. Then, an initial set of scoring criteria is drafted by the language model itself, leveraging in-context learning on different few-shot examples. To further calibrate this set of criteria, we select the best performers and re-draft them with self-refinement. Our experiments on multiple text quality evaluation datasets illustrate a significant improvement in correlation with expert evaluation through calibration. Our comprehensive qualitative analysis conveys insightful intuitions and observations on the essence of effective scoring criteria.

Most existing topic models rely on bag-of-words (BOW) representation, which limits their ability to capture word order information and leads to challenges with out-of-vocabulary (OOV) words in new documents. Contextualized word embeddings, however, show superiority in word sense disambiguation and effectively address the OOV issue. In this work, we introduce a novel neural topic model called the Contextlized Word Topic Model (CWTM), which integrates contextualized word embeddings from BERT. The model is capable of learning the topic vector of a document without BOW information. In addition, it can also derive the topic vectors for individual words within a document based on their contextualized word embeddings. Experiments across various datasets show that CWTM generates more coherent and meaningful topics compared to existing topic models, while also accommodating unseen words in newly encountered documents.

Empathetic response generation endeavors to empower dialogue systems to perceive speakers’ emotions and generate empathetic responses accordingly. Psychological research demonstrates that emotion, as an essential factor in empathy, encompasses trait emotions, which are static and context-independent, and state emotions, which are dynamic and context-dependent. However, previous studies treat them in isolation, leading to insufficient emotional perception of the context, and subsequently, less effective empathetic expression. To address this problem, we propose Combining Trait and State emotions for Empathetic Response Model (CTSM). Specifically, to sufficiently perceive emotions in dialogue, we first construct and encode trait and state emotion embeddings, and then we further enhance emotional perception capability through an emotion guidance module that guides emotion representation. In addition, we propose a cross-contrastive learning decoder to enhance the model’s empathetic expression capability by aligning trait and state emotions between generated responses and contexts. Both automatic and manual evaluation results demonstrate that CTSM outperforms state-of-the-art baselines and can generate more empathetic responses. Our code is available at https://github.com/wangyufeng-empty/CTSM

Sentence Simplification aims to make sentences easier to read and understand. With most effort on corpus development focused on English, the amount of annotated data is limited in Chinese. To address this need, we introduce CSSWiki, an open-source dataset for Chinese sentence simplification based on Wikipedia. This dataset contains 1.6k source sentences paired with their simplified versions. Each sentence pair is annotated with operation tags that distinguish between linguistic and content modifications. We analyze differences in annotation scheme and data statistics between CSSWiki and existing datasets. We then report baseline sentence simplification performance on CSSWiki using zero-shot and few-shot approaches with Large Language Models.

Naively assuming English as a source language may hinder cross-lingual transfer for many languages by failing to consider the importance of language contact. Some languages are more well-connected than others, and target languages can benefit from transferring from closely related languages; for many languages, the set of closely related languages does not include English. In this work, we study the impact of source language for cross-lingual transfer, demonstrating the importance of selecting source languages that have high contact with the target language. We also construct a novel benchmark dataset for close contact Chinese-Japanese-Korean-Vietnamese (CJKV) languages to further encourage in-depth studies of language contact. To comprehensively capture contact between these languages, we propose to integrate Romanized transcription beyond textual scripts via Contrastive Learning objectives, leading to enhanced cross-lingual representations and effective zero-shot cross-lingual transfer.

This paper describes the automated generation of CounterNarratives (CNs) for Hate Speech (HS) in Spanish using GPT-based models. Our primary objective is to evaluate the performance of these models in comparison to human capabilities. For this purpose, the English CONAN Multitarget corpus is taken as a starting point and we use the DeepL API to automatically translate into Spanish. Two GPT-based models, GPT-3 and GPT-4, are applied to the HS segment through a few-shot prompting strategy to generate a new CN. As a consequence of our research, we have created a high quality corpus in Spanish that includes the original HS-CN pairs translated into Spanish, in addition to the CNs generated automatically with the GPT models and that have been evaluated manually. The resulting CONAN-MT-SP corpus and its evaluation will be made available to the research community, representing the most extensive linguistic resource of CNs in Spanish to date. The results demonstrate that, although the effectiveness of GPT-4 outperforms GPT-3, both models can be used as systems to automatically generate CNs to combat the HS. Moreover, these models consistently outperform human performance in most instances.

Dialogue act (DA) recognition is usually realized from a speech signal that is transcribed and segmented into text. However, only a little work in DA recognition from images exists. Therefore, this paper concentrates on this modality and presents a novel DA recognition approach for image documents, namely comic books. To the best of our knowledge, this is the first study investigating dialogue acts from comic books and represents the first steps to building a model for comic book understanding. The proposed method is composed of the following steps: speech balloon segmentation, optical character recognition (OCR), and DA recognition itself. We use YOLOv8 for balloon segmentation, Google Vision for OCR, and Transformer-based models for DA classification. The experiments are performed on a newly created dataset comprising 1,438 annotated comic panels. It contains bounding boxes, transcriptions, and dialogue act annotation. We have achieved nearly 98% average precision for speech balloon segmentation and exceeded the accuracy of 70% for the DA recognition task. We also present an analysis of dialogue structure in the comics domain and compare it with the standard DA datasets, representing another contribution of this paper.

Trainable metrics for machine translation evaluation have been scoring the highest correlations with human judgements in the latest meta-evaluations, outperforming traditional lexical overlap metrics such as BLEU, which is still widely used despite its well-known shortcomings. In this work we look at COMET, a prominent neural evaluation system proposed in 2020, to analyze the extent of its language support restrictions, and to investigate strategies to extend this support to new, under-resourced languages. Our case study focuses on English-Maltese and Spanish-Basque. We run a crowd-based evaluation campaign to collect direct assessments and use the annotated dataset to evaluate COMET-22, further fine-tune it, and to train COMET models from scratch for the two language pairs. Our analysis suggests that COMET’s performance can be improved with fine-tuning, and that COMET can be highly susceptible to the distribution of scores in the training data, which especially impacts low-resource scenarios.

Generative query rewrite generates reconstructed query rewrites using the conversation history while rely heavily on gold rewrite pairs that are expensive to obtain. Recently, few-shot learning is gaining increasing popularity for this task, whereas these methods are sensitive to the inherent noise due to limited data size. Besides, both attempts face performance degradation when there exists language style shift between training and testing cases. To this end, we study low-resource generative conversational query rewrite that is robust to both noise and language style shift. The core idea is to utilize massive unlabeled data to make further improvements via a contrastive co-training paradigm. Specifically, we co-train two dual models (namely Rewriter and Simplifier) such that each of them provides extra guidance through pseudo-labeling for enhancing the other in an iterative manner. We also leverage contrastive learning with data augmentation, which enables our model pay more attention on the truly valuable information than the noise. Extensive experiments demonstrate the superiority of our model under both few-shot and zero-shot scenarios. We also verify the better generalization ability of our model when encountering language style shift.

Extracting structured event knowledge, including event triggers and corresponding arguments, from military texts is fundamental to many applications, such as intelligence analysis and decision assistance. However, event extraction in the military field faces the data scarcity problem, which impedes the research of event extraction models in this domain. To alleviate this problem, we propose CMNEE, a large-scale, document-level open-source Chinese Military News Event Extraction dataset. It contains 17,000 documents and 29,223 events, which are all manually annotated based on a pre-defined schema for the military domain including 8 event types and 11 argument role types. We designed a two-stage, multi-turns annotation strategy to ensure the quality of CMNEE and reproduced several state-of-the-art event extraction models with a systematic evaluation. The experimental results on CMNEE fall shorter than those on other domain datasets obviously, which demonstrates that event extraction for military domain poses unique challenges and requires further research efforts. Our code and data can be obtained from https://github.com/Mzzzhu/CMNEE. Keywords: Corpus,Information Extraction, Information Retrieval, Knowledge Discovery/Representation

Metaphor is a prominent linguistic device in human language and literature, as they add color, imagery, and emphasis to enhance effective communication. This paper introduces a large-scale high quality annotated Chinese Metaphor Corpus, which comprises around 28K sentences drawn from a diverse range of Chinese literary sources, such as poems, prose, song lyrics, etc. To ensure the accuracy and consistency of our annotations, we introduce a comprehensive set of guidelines. These guidelines address the facets of metaphor annotation, including identifying tenors, vehicles, and grounds to handling the complexities of similes, personifications, juxtapositions, and hyperboles. Breaking tradition, our approach to metaphor generation emphasizes tenors and their distinct features rather than the conventional combination of tenors and vehicles. By integrating “ground” as a CoT (Chain of Thoughts) input, we are able to generate metaphors that resonate more with real-world intuition. We test generative models such as Belle, Baichuan, and Chinese-alpaca-33B using our annotated corpus. These models are able to generate creative and fluent metaphor sentences more frequently induced by selected samples from our dataset, demonstrating the value of our corpus for Chinese metaphor research.

Detecting offensive content in internet memes is challenging as it needs additional contextual knowledge. While previous works have only focused on detecting offensive memes, classifying them further into implicit and explicit categories depending on their severity is still a challenging and underexplored area. In this work, we present an end-to-end multitask model for addressing this challenge by empirically investigating two correlated tasks simultaneously: (i) offensive meme detection and (ii) explicit-implicit offensive meme detection by leveraging the two self-supervised pre-trained models. The first pre-trained model, referred to as the “knowledge encoder,” incorporates contextual knowledge of the meme. On the other hand, the second model, referred to as the “fine-grained information encoder”, is trained to understand the obscure psycho-linguistic information of the meme. Our proposed model utilizes contrastive learning to integrate these two pre-trained models, resulting in a more comprehensive understanding of the meme and its potential for offensiveness. To support our approach, we create a large-scale dataset, CM-Off-Meme, as there is no publicly available such dataset for the code-mixed Hindi-English (Hinglish) domain. Empirical evaluation, including both qualitative and quantitative analysis, on the CM-Off-Meme dataset demonstrates the effectiveness of the proposed model in terms of cross-domain generalization.