United States

The rise of video content on social media has led to toxic speech spread, necessitating effec tive moderation. This study addresses detect ing toxic speech in Bangla and Bangla-English videos using multimodal data and deep learn ing techniques. The BanTSS dataset, with 431 videos and 2021 annotated utterances, supports this research. We propose the BanVATLLM framework, a multimodal architecture integrat ing audio, video, and text data. Utilizing ad vanced models like Whisper, MMS, Video MAE, Timesformer, and ChatGPT-3.5, Ban VATLLMshows high accuracy in classifying toxicity, sentiment, and severity, with high F1 scores and Fleiss’ Kappa scores. Results in clude 95.78% F1 and 95.72% accuracy for toxi city, 88.27% F1 and 88.55% accuracy for sever ity, and 84.85% F1 and 83.86% accuracy for sentiment, enhancing detection in low-resource languages.

EMNLP 2024

BanVATLLM and BanTSS: A Multimodal Framework and a Dataset for Detecting Toxic Speech in Bangla and Bangla-English Videos

vidio

multi model

large language model

toxicity

audio

text

### Welcome!
The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. Our Virtual Poster Sessions will take place online Thursday, August 22, 2024.

You are required to register for this event. **Please register [here](https://2024.aclweb.org/registration). **

If you have already registered, please check your inbox for an email from Underline granting you access to ACL 2024 content.

Please register!

ACL 2024

The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. More information will be announced soon.

## Welcome to EMNLP 2024! 
We are excited to welcome you to one of the most prominent conferences in the field of Natural Language Processing. This year, EMNLP 2024 is being held in a hybrid format,
offering both virtual and in-person participation in beautiful Miami. Due to a record-breaking number of submissions, we've expanded the total number of accepted papers to accommodate more cutting-edge research from around the globe.
### [Conference Handbook](https://drive.google.com/file/d/1WPROgxjLAC96AJL7Ugy0tEnYm7dkrbHt/view?usp=sharing)

You are required to register for this event. **Please register [here](https://2024.emnlp.org/registration/).** The EMNLP 2024 event page on Underline will be open to public one week prior to the event.

EMNLP 2024 will take place in Miami, Florida from Nov 12th to Nov 16th, 2024, at the Hyatt Regency Miami Hote and on Underline for remote participants.

Large Language Models (LLMs) show promising learning and reasoning abilities. Compared to other NLP tasks, multilingual and multi-label emotion evaluation is under-explored in LLMs. In this paper, we present EthiopicEmotion, a multi-label emotion with intensity dataset for four Ethiopian languages, namely Amharic (amh), Afan Oromo (oro), Somali (som), and Tigrinya (tir). We perform extensive experiments with additional English multi-label emotion data from SemEval 2018: Affect in tweets work. Our evaluation includes zero-shot and in-context learning (ICL) from large language models. The result shows that accurate multi-label emotion classification is still insufficient, even for high-resource languages (English), and there is a large gap between the performance of English and low-resource languages.

EthiopicEmotion: Multi-label Emotion Dataset with Large Language Models Evaluation

NLP is severely hampered by a scarcity of digital resources. This is especially true for Amharic, a language with few resources but a rich morphology. In response, a total of 67,739 Amharic news documents from 8 different categories are gathered from web sources. A baseline document categorization experiment is carried out to validate the usability of the obtained corpora from various domains.

Document Classification for the Under-resourced Amharic Language

African languages are not well-represented in Natural Language Processing (NLP). The main reason is a lack of resources for training models. Low-resource languages, such as Amharic and Ge’ez, cannot benefit from modern NLP methods because of the lack of high-quality datasets. This paper presents AGE, an open-source tripartite alignment of Amharic, Ge’ez, and English parallel dataset. Additionally, we introduced a novel, 1,000 Ge’ez-centered sentences sourced from areas such as news and novels. Furthermore, we developed a model from a multilingual pre-trained language model, which brings 12.29 and 30.66 for English-Ge’ez and Ge’ez to English, respectively, and 9.39 and 12.29 for Amharic-Ge’ez and Ge’ez-Amharic respectively.

AGE: Amharic, Ge’ez and English Parallel Dataset

This study presents a comprehensive, long-term project to explore the effectiveness of various prompting techniques in detecting dialogical mental manipulation. We implement Chain-of-Thought prompting with Zero-Shot and Few-Shot settings on a binary mental manipulation detection task, building upon existing work conducted with Zero-Shot and Few-Shot prompting. Our primary objective is to decipher why certain prompting techniques display superior performance so as to craft a novel framework tailored for the detection of mental manipulation. Preliminary findings suggest that advanced prompting techniques may not be suitable for more complex models if they are not trained through example-based learning.

Enhanced Detection of Conversational Mental Manipulation Through Advanced Prompting Techniques

Jack and the BeansTALK: Towards Question Answering in Plant Biology

Jack and the BeansTALK : Towards Question Answering in Plant Biology

Aspect-based sentiment Analysis (ABSA) is an advanced NLP task that identifies sentiments related to specific aspects of a product or service, offering more detailed consumer insights than general sentiment analysis. The proposed research introduces a hybrid model, Instruct-DeBERTa, which combines InstructABSA for aspect term extraction (ATE) and DeBERTa-V3-baseabsa-V1 for aspect sentiment classification (ASC). Using datasets from SemEval Restaurant 2014,2015,2016 and SemEval Laptop 2014, the model demonstrated improved performance across domains. Further enhancements included category classification using cosine similarity, linear layers with ReLU activation, regularization methods, and optimized attention heads for the hospitality domain. These improvements address existing model limitations, providing a comprehensive solution for analyzing consumer feedback, valuable for enhancing customer satisfaction and product development.

Instruct-DeBERTa: A Hybrid Approach for Enhanced Aspect-Based Sentiment Analysis with Category Extraction

Ranking a set of items based on their relevance to a given query is a core problem in search and recommendation. Transformer-based ranking models are the state-of-the-art approaches for such tasks, but they score each query-item independently, ignoring the joint context of other relevant items. This leads to sub-optimal ranking accuracy and high computational costs. We propose Cross-encoders with Joint Efficient Modeling (CROSS-JEM), a novel ranking approach that enables transformer-based models to jointly score multiple items for a query, maximizing parameter utilization. CROSS-JEM leverages (a) redundancies and token overlaps to jointly score multiple items (short-text phrases in search and recommendations), and (b) a novel training objective that models ranking probabilities. CROSS-JEM achieves state-of-the-art accuracy on publicly available ranking benchmarks with over 4x-lower ranking latency compared to the baselines.

CROSS-JEM: Accurate and Efficient Cross-encoders for Short-text Ranking Tasks

Latin America (LatAm) is mainly represented by developing or under-developed countries, and given that their investment in research and development may be lacking, their presence in high-impact research communities such as computational linguistics (CL) could be marginal. This work aims to measure the presence of LatAm researchers in the CL community and improve the visibility of those underrepresented investigators. We extracted the metadata of all ACL Anthology publications with at least one LatAm researcher affiliated with an institution located in an LatAm country at the time of the publication. We found that only a small percentage (2.4 \%) of CL publications have affiliations with institutions based in LatAm and that Alexander Gelbukh (Mexico) and Luciana Benotti (Argentina) are the most productive researchers. Our analysis also reveals that some countries in the region have not contributed to any CL publications. Despite these challenges, our results highlight the potential for growth and improvement. By shedding light on the underrepresentation of LatAm researchers in CL, this study aims to promote greater visibility and inclusivity within the community, ultimately fostering a more diverse and vibrant research landscape.

The faces of Latin American research on computational linguistics

Using medical data for Deep Learning models can be highly beneficial, but protecting sensitive and personal patient information in the clinical field is critical. One of the most common ways to use this data while protecting patient privacy is by generating synthetic text with Large Language Models (LLMs) using differential privacy (DP). Although DP techniques, such as the Differentially Private Stochastic Gradient Descent (DP-SGD), are often assumed to guarantee privacy, they require specific conditions to be met. This study shows how memorization in LLMs can occur when these privacy guarantees are compromised, potentially leading to the leakage of personal and sensitive information in generated clinical re017 ports. If these gaps are addressed, DP could offer more reliable safeguards for clinical data, improving privacy without sacrificing utility

Analyzing the Leakage of Personal Information in Synthetic Clinical Spanish Texts

Modern language models predict the next token in the sequence by considering the past text through a powerful function. However, language models have no explicit mechanism that allows them to spend computation time for planning long-distance future text, leading to a suboptimal token prediction. In this paper, we propose a planner that predicts a latent plan for many sentences into the future. By sampling multiple plans at once, we condition the language model on an accurate approximation of the distribution of text continuations, which leads to better next token prediction accuracy. In effect, this allows trading computation time for prediction accuracy.

Next from EMNLP 2024

EthiopicEmotion: Multi-label Emotion Dataset with Large Language Models Evaluation

Similar lecture

ToxVidLM: A Multimodal Framework for Toxicity Detection in Code-Mixed Videos