International Conference on Computer and Knowledge Engineering

Home / 14th International Conference on Computer and Knowledge Engineering

Enhancing Persian Word Sense Disambiguation with Large Language Models: Techniques and Applications

Authors :

Fatemeh Zahra Arshia¹ Saeedeh Sadat Sadidpour²

1- Faculty of Electronic & Computer Engineering, Malek Ashtar University of Technology 2- Faculty of Electronic & Computer Engineering, Malek Ashtar University of Technology

Keywords :

Word Sense Disambiguation (WSD)،Large Language Models (LLMs)،Persian Disambiguation

Abstract :

WSD means the task of word sense disambiguation, which is a very important task in NLP. It assigns not only the meaningful word to the source text but also the proper meaning of the word according to the context. Hence, it is key to the proper accomplishment of NLP in Persian—a language rich in morphology and great polysemy. The recent improvements in LMs have greatly advanced the capabilities of NLP, opening further improvement avenues in WSD performance. This paper presents the integration of LLMs for improving WSD in Persian, considering the linguistic challenges related to this language. In this study, we consider four models of the Persian language: FaBERT, AriaBERT, GPT-2 Persian, and PersianMind-v1.0. We use the supervised fine-tuning method on the SBU-WSD-Corpus. Our methodology will consist of preprocessing the Persian WSD corpus, then fine-tuning the models with the mentioned corpus, and measuring their performance. Results indicate that methods using LLMs significantly improve WSD accuracy against traditional methods, with FaBERT achieving the best accuracy. We have further expounded on their real-life applications, such as sentiment analysis, to show the consequential effect of this advancement on general NLP tasks. The study is concluded with some insights into future research directions, underlining the potential role that LLMs can play in further transforming WSD and related fields.

Papers List

List of archived papers

Impact of Oversampling Methods on Imbalanced Dataset for Software Fault Prediction

Alireza Abiri - Alireza Tajary - Mansoor Fateh

A Comprehensive Approach to SMS Spam Filtering Integrating Embedded and Statistical Features

Shaghayegh Hosseinpour - Mohammad Reza Keyvanpour

Assessing Users' Influence on Respondents in Conversation Quality: A Quantitative Study on Reddit Based on the Cooperative Principle

Afsaneh Habibi - Fattaneh Taghiyareh

LightFedSelect: A Lightweight Framework for Byzantine-Robust Federated Learning

Seyed Saeed Razavi - Seyed Arsalan Vasegh Rahim Parvar - Soroosh Dadashi Pakdeh - Mohammad Matin Rezaeifard - Morteza Mollaie Chafi - Reza Ebrahimi Atani

Leveraging Self-Supervised Models for Automatic Whispered Speech Recognition

Aref Farhadipour - Homa Asadi - Volker Dellwo

An Automated Visual Defect Segmentation for Flat Steel Surface Using Deep Neural Networks

Dorna Nourbakhsh Sabet - Mohammad Reza Zarifi - Javad Khoramdel - Yasamin Borhani - Esmaeil Najafi

Camouflage Object Segmentation with Attention-Guided Pix2Pix and Boundary Awareness

Erfan Akbarnezhad Sany - Fatemeh Naserizadeh - Parsa Sinichi - Seyyed Abed Hosseini

I-ACS: An Improved Ant Colony System to Solve the Time-Dependent Orienteering Problem

Zahra Bakhshandeh - Morteza Keshtkaran

Real-Time Forecasting Using Mixed Frequency Time-Series Data

Armin Khayati - Mohammad Taheri - Koorush Ziarati

Stock market prediction using multi-objective optimization

Mahshid Zolfaghari - Hamid Fadishei - Mohsen Tajgardan - Reza Khoshkangini