INDONESIAN TASK-ORIENTED DIALOGUE SYSTEM USING END-TO-END APPROACH

A Task-oriented dialogue system (ToDS) is a conversational agent designed to communicate with users in natural language, assisting them in completing various user tasks i.e. making restaurant reservations or purchasing tickets. Recently, many researchers are interested in using the end-to-end app...

全面介紹

Saved in:
書目詳細資料
主要作者: Dehan Al Kautsar, Muhammad
格式: Final Project
語言:Indonesia
在線閱讀:https://digilib.itb.ac.id/gdl/view/73923
標簽: 添加標簽
沒有標簽, 成為第一個標記此記錄!
機構: Institut Teknologi Bandung
語言: Indonesia
實物特徵
總結:A Task-oriented dialogue system (ToDS) is a conversational agent designed to communicate with users in natural language, assisting them in completing various user tasks i.e. making restaurant reservations or purchasing tickets. Recently, many researchers are interested in using the end-to-end approach to build the ToDS framework, as it offers simplicity compared to the other approach. However, since there is no ToDS corpus available in Indonesia, which is an under-represented language, an Indonesian ToDS benchmark has never been conducted. The main focus of this thesis involves the construction of two ToDS training corpora specifically designed for the Indonesian language. Additionally, the thesis includes conducting experiments using two distinct ToDS frameworks, while utilizing the manually created corpora. The total number of constructed dialogues for the Indonesian ToDS corpus in this thesis amounts to 999 dialogues, originating from two English ToDS corpora named CamRest and SMD. The experiments conducted in this study evaluate the frameworks in three types of experiments: monolingual, bilingual, and cross-lingual transfer learning. The end-to-end frameworks tested in this thesis are Sequicity and MinTL, which differ in the type of word embedding employed by each framework. The experiment results reveal that the language adaptation process incorporating two sets of Indonesian and English corpora during the training phase of the ToDS framework has a positive impact on the model's capability to successfully complete user-given tasks. Furthermore, bilingual experiments consistently yield the best metric values across all experiments. However, both frameworks struggle to achieve satisfactory results in cross-lingual experiments due to the language disparity between the training and testing corpora.