Thai Knowledge-Augmented Language Model Adaptation (ThaiKALA)

Large language models have exhibited considerable prowess in diverse NLP tasks, demonstrating promising performance. However, they still have limitations in effectively capturing domain-specific knowledge and contextually-relevant information, resulting in hallucination issues. To address these challenges, This paper presents ThaiKALA, a framework designed for the Thai language to augment domain-specific knowledge into the language model. The framework utilizes three modules to handle Thai language specifically: event extraction, a self-defined ID database, and a multilingual language model. To confirm the performance, the framework is also evaluated with strong generative baselines like GPT-3 and GPT-3.5-turbo-16k. As a result, ThaiKALA, with only Entity Memory, outperforms all baselines including GPT-3 and GPT-3.5 in extractive Question Answering (EQA) tasks, achieving a higher exact match (42.48%) and competitive F1 scores (67.07%). These results demonstrate that ThaiKALA is effective in enhancing the language model's performance on Thai extractive QA by augmenting the extracted knowledge.