This paper presents the Chinchunmei team’s contributions to the WASSA2024 Shared-Task 1: Empathy Detection and Emotion Classification. We participated in Tracks 1, 2, and 3 to predict empathetic scores based on dialogue, article, and essay content. We choose Llama3-8b-instruct as our base model. We developed three supervised fine-tuning schemes: standard prediction, role-play, and contrastive prediction, along with an innovative scoring calibration method called Contrastive Reasoning Calibration during inference. Pearson Correlation was used as the evaluation metric across all tracks. For Track 1, we achieved 0.43 on the devset and 0.17 on the testset. For Track 2 emotion, empathy, and polarity labels, we obtained 0.64, 0.66, and 0.79 on the devset and 0.61, 0.68, and 0.58 on the testset. For Track 3 empathy and distress labels, we got 0.64 and 0.56 on the devset and 0.33 and 0.35 on the testset.

In the realm of conversational empathy and emotion prediction, emotions are frequently categorized into multiple levels. This study seeks to enhance the performance of emotion prediction models by incorporating the Pearson correlation coefficient as a regularization term within the loss function. This regularization approach ensures closer alignment between predicted and actual emotion levels, mitigating extreme predictions and resulting in smoother and more consistent outputs. Such outputs are essential for capturing the subtle transitions between continuous emotion levels. Through experimental comparisons between models with and without Pearson regularization, our findings demonstrate that integrating the Pearson correlation coefficient significantly boosts model performance, yielding higher correlation scores and more accurate predictions. Our system officially ranked 9th at the Track 2: CONV-turn. The code for our model can be found at Link .

This paper proposes a novel ensemble approach that combines Graph Neural Networks (GNNs) and LightGBM to enhance personality prediction based on the personality Big 5 model. By integrating BERT embeddings from user essays with knowledge graph-derived embeddings, our method accurately captures rich semantic and relational information. Additionally, a special loss function that combines Mean Squared Error (MSE), Pearson correlation loss, and contrastive loss to improve model performance is introduced. The proposed ensemble model, made of Graph Convolutional Networks (GCNs), Graph Attention Networks (GATs), and LightGBM, demonstrates superior performance over other models, with significant improvements in prediction accuracy for the Big Five personality traits achieved. Our system officially ranked 2^nd at the Track 4: PER track.

This paper presents our participation to the WASSA 2024 Shared Task on Empathy Detection and Emotion Classification and Personality Detection in Interactions. We focus on Track 2: Empathy and Emotion Prediction in Conversations Turns (CONV-turn), which consists of predicting the perceived empathy, emotion polarity and emotion intensity at turn level in a conversation. In the method, we conduct BERT and DeBERTa based finetuning, implement the CombinedLoss which consists of a structured contrastive loss and Pearson loss, adopt adversarial training using Fast Gradient Method (FGM). This method achieved Pearson correlation of 0.581 for Emotion,0.644 for Emotional Polarity and 0.544 for Empathy on the test set, with the average value of 0.590 which ranked 4th among all teams. After submission to WASSA 2024 competition, we further introduced the segmented mix-up for data augmentation, boosting for ensemble and regression experiments, which yield even better results: 0.6521 for Emotion, 0.7376 for EmotionalPolarity, 0.6326 for Empathy in Pearson correlation on the development set. The implementation and fine-tuned models are publicly-available at https://github.com/hyy-33/hyy33-WASSA-2024-Track-2.

Emotion expression is one of the essential traits of conversations. It may be self-related or caused by another speaker. The variety of reasons may serve as a source of the further emotion causes: conversation history, speaker’s emotional state, etc. Inspired by the most recent advances in Chain-of-Thought, in this work, we exploit the existing three-hop reasoning approach (THOR) to perform large language model instruction-tuning for answering: emotion states (THOR-state), and emotion caused by one speaker to the other (THOR-cause). We equip THORcause with the reasoning revision (RR) for devising a reasoning path in fine-tuning. In particular, we rely on the annotated speaker emotion states to revise reasoning path. Our final submission, based on Flan-T5-base (250M) and the rule-based span correction technique, preliminary tuned with THOR-state and fine-tuned with THOR-cause-rr on competition training data, results in 3rd and 4th places (F1-proportional) and 5th place (F1-strict) among 15 participating teams. Our THOR implementation fork is publicly available: https://github.com/nicolay-r/THOR-ECAC

SemEval-2024 Task 8 introduces the challenge of identifying machine-generated texts from diverse Large Language Models (LLMs) in various languages and domains. The task comprises three subtasks: binary classification in monolingual and multilingual (Subtask A), multi-class classification (Subtask B), and mixed text detection (Subtask C). This paper focuses on Subtask A & B. To tackle this task, this paper proposes two methods: 1) using traditional machine learning (ML) with natural language preprocessing (NLP) for feature extraction, and 2) fine-tuning LLMs for text classification. For fine-tuning, we use the train datasets provided by the task organizers. The results show that transformer models like LoRA-RoBERTa and XLM-RoBERTa outperform traditional ML models, particularly in multilingual subtasks. However, traditional ML models performed better than transformer models for the monolingual task, demonstrating the importance of considering the specific characteristics of each subtask when selecting an appropriate approach.

One of the key challenges in Natural Language Generation (NLG) is “hallucination,” in which the generated output appears fluent and grammatically sound but may contain incorrect information. To address this challenge, “SemEval-2024 Task 6 - SHROOM, a Shared-task on Hallucinations and Related Observable Overgeneration Mistakes” is introduced. This task focuses on detecting overgeneration hallucinations in texts generated from Large Language Models for various NLG tasks. To tackle this task, this paper proposes two methods: (1) hypothesis-target similarity, which measures text similarity between a generated text (hypothesis) and an intended reference text (target), and (2) a SelfCheckGPT-based method to assess hallucinations via predefined prompts designed for different NLG tasks. Experiments were conducted on the dataset provided in this task. The results show that both of the proposed methods can effectively detect hallucinations in LLM-generated texts with a possibility for improvement.

Headline Generation is an essential task in Natural Language Processing (NLP), where models often exhibit limited ability to accurately interpret numerals, leading to inaccuracies in generated headlines. This paper introduces CoT-NumHG, a training strategy leveraging the Chain of Thought (CoT) paradigm for Supervised Fine-Tuning (SFT) of large language models. This approach is aimed at enhancing numeral perception, interpretability, accuracy, and the generation of structured outputs. Presented in SemEval-2024 Task 7 (task 3): Numeral-Aware Headline Generation (English), this challenge is divided into two specific subtasks. The first subtask focuses on numerical reasoning, requiring models to precisely calculate and fill in the missing numbers in news headlines, while the second subtask targets the generation of complete headlines. Utilizing the same training strategy across both subtasks, this study primarily explores the first subtask as a demonstration of our training strategy. Through this competition, our CoT-NumHG-Mistral-7B model attained an accuracy rate of 94%, underscoring the effectiveness of our proposed strategy.

In this study, we introduce an MLP approach for extracting multimodal cause utterances in conversations, utilizing the multimodal conversational emotion causes from the ECF dataset. Our research focuses on evaluating a bi-modal framework that integrates video and audio embeddings to analyze emotional expressions within dialogues. The core of our methodology involves the extraction of embeddings from pre-trained models for each modality, followed by their concatenation and subsequent classification via an MLP network. We compared the accuracy performances across different modality combinations including text-audio-video, video-audio, and audio only.

In SemEval-2023 Task 1, a task of applying Word Sense Disambiguation in an image retrieval system was introduced. To resolve this task, this work proposes three approaches: (1) an unsupervised approach considering similarities between word senses and image captions, (2) a supervised approach using a Siamese neural network, and (3) a self-supervised approach using a Bayesian personalized ranking framework. According to the results, both supervised and self-supervised approaches outperformed the unsupervised approach. They can effectively identify correct images of ambiguous words in the dataset provided in this task.

Legal documents tend to be large in size. In this paper, we provide an experiment with attention-based approaches complemented by certain document processing techniques for judgment prediction. For the prediction of explanation, we consider this as an extractive text summarization problem based on an output of (1) CNN with attention mechanism and (2) self-attention of language models. Our extensive experiments show that treating document endings at first results in a 2.1% improvement in judgment prediction across all the models. Additional content peeling from non-informative sentences allows an improvement of explanation prediction performance by 4% in the case of attention-based CNN models. The best submissions achieved 8’th and 3’rd ranks on judgment prediction (C1) and prediction with explanation (C2) tasks respectively among 11 participating teams. The results of our experiments are published

Named Entity Recognition (NER) is a subtask of Natural Language Processing (NLP) that involves identifying and categorizing named entities. The result annotation makes unstructured natural texts applicable for other NLP tasks, including information retrieval, question answering, and machine translation. NER is also essential in legal as an initial stage in extracting relevant entities. However, legal texts contain domain-specific named entities, such as applicants, defendants, courts, statutes, and articles. The latter makes standard named entity recognizers incompatible with legal documents. This paper proposes an approach combining multiple models’ results via a voting mechanism for unique entity identification in legal texts. This endeavor focuses on extracting legal named entities, and the specific assignment (task B) is to create a legal NER system for unique entity annotation in legal documents. The results of our experiments and system implementation are published in https://github.com/SuperEDG/Legal_Project.

In human languages, there are many presuppositional constructions that impose a constrain on the taxonomic relations between two nouns depending on their order. These constructions create a challenge in validating taxonomic relations in real-world contexts. In SemEval2022-Task3 Presupposed Taxonomies: Evaluating Neural Network Semantics (PreTENS), the organizers introduced a task regarding validating the taxonomic relations within a variety of presuppositional constructions. This task is divided into two subtasks: classification and regression. Each subtask contains three datasets in multiple languages, i.e., English, Italian and French. To tackle this task, this work proposes to fine-tune different BERT-based models pre-trained on different languages. According to the experimental results, the fine-tuned BERT-based models are effective compared to the baselines in classification. For regression, the fine-tuned models show promising performance with the possibility of improvement.

Sarcasm has gained notoriety for being difficult to detect by machine learning systems due to its figurative nature. In this paper, Bidirectional Encoder Representations from Transformers (BERT) model has been used with ensemble loss made of cross-entropy loss and negative log-likelihood loss to classify whether a given sentence is in English and Arabic tweets are sarcastic or not. From the results obtained in the experiments, our proposed BERT with ensemble loss achieved superior performance when applied to English and Arabic test datasets. For the validation dataset, our model performed better on the Arabic dataset but failed to outperform the baseline method (made of BERT with only a single loss function) when applied on the English validation set.

Most question answering tasks focuses on predicting concrete answers, e.g., named entities. These tasks can be normally achieved by understanding the contexts without additional information required. In Reading Comprehension of Abstract Meaning (ReCAM) task, the abstract answers are introduced. To understand abstract meanings in the context, additional knowledge is essential. In this paper, we propose an approach that leverages the pre-trained BERT Token embeddings as a prior knowledge resource. According to the results, our approach using the pre-trained BERT outperformed the baselines. It shows that the pre-trained BERT token embeddings can be used as additional knowledge for understanding abstract meanings in question answering.

Humour detection is an interesting but difficult task in NLP. Because humorous might not be obvious in text, it can be embedded into context, hide behind the literal meaning and require prior knowledge to understand. We explored different shallow and deep methods to create a humour detection classifier for task 7-1a. Models like Logistic Regression, LSTM, MLP, CNN were used, and pre-trained models like DistilBert were introduced to generate accurate vector representation for textual data. We focused on applying multi-scale strategy on modelling, and compared different models. Our best model is the DistilBERT+MultiScale CNN, it used different sizes of CNN kernel to get multiple scales of features, which achieved 93.7% F1-score and 92.1% accuracy on the test set.

Crowdsourcing has been ubiquitously used for annotating enormous collections of data. However, the major obstacles to using crowd-sourced labels are noise and errors from non-expert annotations. In this work, two approaches dealing with the noise and errors in crowd-sourced labels are proposed. The first approach uses Sharpness-Aware Minimization (SAM), an optimization technique robust to noisy labels. The other approach leverages a neural network layer called softmax-Crowdlayer specifically designed to learn from crowd-sourced annotations. According to the results, the proposed approaches can improve the performance of the Wide Residual Network model and Multi-layer Perception model applied on crowd-sourced datasets in the image processing domain. It also has similar and comparable results with the majority voting technique when applied to the sequential data domain whereby the Bidirectional Encoder Representations from Transformers (BERT) is used as the base model in both instances.

SemEval Task 4 Commonsense Validation and Explanation Challenge is to validate whether a system can differentiate natural language statements that make sense from those that do not make sense. Two subtasks, A and B, are focused in this work, i.e., detecting against-common-sense statements and selecting explanations of why they are false from the given options. Intuitively, commonsense validation requires additional knowledge beyond the given statements. Therefore, we propose a system utilising pre-trained sentence transformer models based on BERT, RoBERTa and DistillBERT architectures to embed the statements before classification. According to the results, these embeddings can improve the performance of the typical MLP and LSTM classifiers as downstream models of both subtasks compared to regular tokenised statements. These embedded statements are shown to comprise additional information from external resources which help validate common sense in natural language.

Memes are widely used on social media. They usually contain multi-modal information such as images and texts, serving as valuable data sources to analyse opinions and sentiment orientations of online communities. The provided memes data often face an imbalanced data problem, that is, some classes or labelled sentiment categories significantly outnumber other classes. This often results in difficulty in applying machine learning techniques where balanced labelled input data are required. In this paper, a Gaussian Mixture Model sampling method is proposed to tackle the problem of class imbalance for the memes sentiment classification task. To utilise both text and image data, a multi-modal CNN-LSTM model is proposed to jointly learn latent features for positive, negative and neutral category predictions. The experiments show that the re-sampling model can slightly improve the accuracy on the trial data of sub-task A of Task 8. The multi-modal CNN-LSTM model can achieve macro F1 score 0.329 on the test set.