VERIFICATION OF UNANSWERABLE QUESTIONS IN THE QUESTION ANSWERING SYSTEM USING SENTENCE-BERT AND COSINE SIMILARITY

• Main Author: Alfarizy, Giffari
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/66650
• Institution: Institut Teknologi Bandung

With the growing quantity of human-machine interaction in modern times, the need for a question answering system is also felt to improve the quality of human- machine interaction. One part of the question answering study is reading comprehension where the system must be able to determine the correct answer from the given document and state if there is no answer in the document. ELECTRA as previous research is state-of-the-art on the SQuAD 2.0. However, this research can still be developed again because ELECTRA on the SQuAD 2.0 dev dataset still makes 1,287 answers errors in determining whether the question is answerable or unanswerable. This study proposes a modification to ELECTRA in verifying answers by adding similarity parameters. Sentence-BERT which is a state-of-the- art in sentence embedding is used to get the sentence similarity between context and prediction that have been converted into declarative sentences using the rule-based method. The experimental results show that the proposed method improve the performance of exact match and F1 scores of ELECTRA.