KMITL

Permanent URI for this communityhttps://dspace.kmitl.ac.th/handle/123456789/1

Browse

Search Results

Now showing 1 - 6 of 6

Some of the metrics are blocked by your
consent settings
Thai Generalized Alignment Task (TGAT): A Corpus and Comparative Study for Hallucination Detection in Thai
(2025-01-01)
Wangwon, Sivakorn
;
Pitijaroonpong, Rapepong
;
Chuangkrud, Piyawat
;
Damrongrat, Chaianun
;
Kongyoung, Sarawoot
Large Language Models (LLMs) face critical reliability challenges due to hallucination-the generation of factually inaccurate content. While hallucination detection has advanced for major languages, Thai remains underserved, lacking specialized alignment models and datasets. We introduce the Thai Generalized Alignment Task (TGAT), a seven-subtask corpus tailored for Thai hallucination detection, spanning Fact Verification, Natural Language Inference, Information Retrieval, Question Answering, Summarization, Semantic Textual Similarity, and Paraphrase Identification. We conduct a comparative study across transformer families-encoder-only, encoder-decoder, and decoder-only-evaluated on the test set of each sub-task to analyze architectural compatibility and generalization for Thai hallucination detection. We also perform an ablation study to quantify how the presence of each subtask dataset affects the overall average performance, clarifying the role of dataset composition. Our experiments show that decoder-only models consistently outperform encoder-decoder and encoder-only alternatives under a standardized zero-shot prompting setup, establishing strong baselines and offering guidance for model selection and dataset design in low-resource settings.
Some of the metrics are blocked by your
consent settings
Pantip Multi-turn Datasets Generating from Thai Large Social Platform Forum Using Sentence Similarity Techniques
(2024-01-01)
Sae-Oueng, Anon
;
Kerdthaisong, Kun
;
Sukhantharat, Kittisak
;
Phasook, Pakawat
;
Chuangkrud, Piyawat
Fine-tuning Large Language Models (LLMs) for specific domains is crucial. However, lack of Thai open dialogues presents a major challenge. For the major challenge, this study proposes a novel methodology for extracting and constructing multi-turn conversational data from existing Thai large social platform, named Pantip. Our approach implements semantic matching algorithms to identify and compile both single-turn and multi-turn dialogues. By employing a cosine similarity threshold ≥ 0.3, we yields contextually coherent conversation pairs directly from the source data. The outcome dataset represents real Thai conversation styles, which could improve how accurately fine-tuned language models reflect Thai social and linguistic norms. Our approach introduces a new way to use existing publicly available data to create training datasets, which is especially valuable for languages with limited resources. Chaotic Pantip datasets can be contributed to the development of more culturally attuned and linguistically precise Thai language models, potentially advancing culturally-specific natural language processing.
Some of the metrics are blocked by your
consent settings
ThaiBKD: Effective of Continual Pre-Training LLM in Thai Language Based on Knowledge Dataset
(2024-01-01)
Phasook, Pakawat
;
Pranee, Jessada
;
Limcharoen, Chananyu
;
Sukhantharat, Kittisak
;
Saeoueng, Anon
Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide range of natural language processing (NLP) tasks. Existing works emphasize the importance of continual pretraining with high-quality knowledge data to enhance LLM performance. This paper investigates the comparative effectiveness of using meticulously curated, high-quality knowledge datasets - comprising scholarly articles, journalism, medical expertise, financial data, legal documents, and verified library resources - against general data sourced from social media and diverse domain sources, particularly from the popular Thai blog platform Pantip, as pre-training data for LLMs.The evaluation results show that ThaiBKD-based LLM, trained on clean and domain-specific data, consistently outperforms Pantip-based LLM in most aspects of Thai language evaluation tasks, demonstrating more effective reasoning capabilities and achieving higher scores across various benchmarks. Notably, the Thai Language Based on Knowledge Dataset (ThaiBKD) LLM outperforms or higher than the performance of GPT-3.5 Turbo in a versatile evaluation set, particularly in tasks requiring specialized knowledge.However, the Pantip-based LLM exhibits substantial strengths in culturally nuanced tasks, such as XCOPA, XNLI, and Belebele, where its understanding of informal and diverse language structures from social media provides a competitive edge. These findings highlight the nuanced trade-offs between data quality, domain specificity, and cultural relevance, underscoring the need for strategic data selection in the pre-training of advanced Large language Models and Continual-pretraining Large Language Models.
Some of the metrics are blocked by your
consent settings
When are Latent Topics Useful for Text Mining?: Enriching Bag-of-Words Representations with Information Extraction in Thai News Articles
(2023-01-01)
Kanungsukkasem, Nont
;
Chuangkrud, Piyawat
;
Pitichotchokphokhin, Pimpitcha
;
Damrongrat, Chaianun
;
Leelanupab, Teerapong
The Bag-of-Words (BOW) model is simple but one of the successful representations of text documents. This model, however, suffers from the sparse matrix, in which most of the elements are zero. Topic modeling is an unsupervised learning method that can represent text documents in a low-dimensional space. Latent Dirichlet Allocation (LDA) is a topic modeling technique used for topic extraction and data exploration, with interpretable output. This paper presents a thorough study of potential benefits of applying LDA, as a feature extraction, to topic discovery and document classification in Thai news articles, comparing with TF–IDF and Word2Vec. We also studied how much of the top Thai terms extracted from LDA with the different numbers of topics can be interpretable and meaningful, and can be a representative of the corpus. Besides, a set of Topic Coherence measures were included in our study to estimate the degree of semantic similarity of extracted topics. To compare the performance and optimization time of classification of features from the different feature extraction methods, various types of classifiers, e.g., Logistic Regression, Random Forest, XGBoosting, etc., were experimented.
Some of the metrics are blocked by your
consent settings
A Non-Entity Approach for Intent-Based Classification: A Case Study of Thai News
(2022-01-01)
Damrongrat, Chaianun
;
Sangkeettrakarn, Chatchawal
;
Piyatumrong, Apivadee
;
Trankultaweekoon, Kanokorn
;
Seekhem, Pachit
Conjunction and stop words are normally ignored in text classification task that is content-based, such as classifying news into entertainment or sports. However, they are useful in this study, since the content and the intention of the document are independent. This paper studies intent-based classification that specifically desires to classify the author's intention of Thai news article into three intents, 'inform', 'explain', and 'give solution'. These three intents subtly co-exist with the content of the article and thus is our classification challenge. Our experiments confirm that intent-based classification needs a different approach from those techniques used for content-based classification. Accordingly, we propose a new pipeline for Thai intent-based classification such that conjunction and others can play a significant role above entity. Our contributions include (1) proving the need for a new methodology to handle intent-based classification and (2) proposing the Non-Entity data processing approach to managing intent-based classification problems. The proposed methodology shows partially promising results. Nonetheless, flags for enhancement are also discussed in the conclusion for future works.
Some of the metrics are blocked by your
consent settings
Keyword-Text Graph Representation for Short Text Classification
(2021-01-01)
Chuanakrud, Piyawat
;
Leelanupab, Teerapong
;
Damrongrat, Chaianun
;
Kanunasukkasem, Nont
Short text classification is an essential task in Natural Language Processing. This task is widely applied to many applications, such as spam filtering, question-answering, artificial conversational agent, sentiment analysis, review mining, etc. Short texts usually encounter a great challenge for classification due to data sparseness as they do not provide sufficient contextual information. In this paper, we introduce Keyword-Text Graph Convolutional Networks (KwTGCN) for short text classification. We also propose a method to identify keywords by estimating word distribution over different categories. These category keywords are then used to build a special keyword-text graph of short text corpus. We employ Graph Convolutional Network (GCN) and our keyword-text graph to generate the representation of short text corpus based on the relations of document-keyword and document-word as well as the word co-occurrence. This document, word and keyword representation is further used as an input feature for the next layer of short text classification. The experimental results on multiple benchmark datasets show that our proposed model outperforms the state-of-the-art models for short text classification in multiple attempts.

KMITL

Browse

Filters

Settings

Sort By

Results per page

Search Results