Knowledge-Enriched Domain Specific Chatbot on Low-resource Language
Rizal Setya Perdana, Putra Pandu Adikara, Indriati Indriati, Diva Kurnianingtyas · 2022
This study presents architecture in learning human-machine conversation using a human language known as chatbots. Given an utterance, a chatbot attempt to reply using human language by mimicking human intelligence in communication. Previous works in this domain were mostly implemented in English, thus it remains a problem when bringing to a low-resource language, e.g., Indonesian. The limited availability of human-to-human dialog history led to difficulty in the learning process of machine learning algorithms. Therefore, this study proposed a Knowledgeable Chatbot (KC), an enhanced pipeline that enables to receiving of transferred knowledge from another task. A data augmentation pipeline is proposed to handle the limited number of available. To deal with the low-resource language, this study proposed to incorporate a pre-trained language model to gain contextualized language understanding. As this research can be categorized as preliminary, extensive experiments are required to prove the effectiveness of each part. Standard automatic metrics for information retrieval and classification prove that KC excels in the ablation study.