INDONESIAN OPEN DOMAIN ANSWER FINDER ON TEXT DOCUMENT USING R-NET ARCHITECTURE

• Main Author: YUDI UTAMA - NIM : 13514011 , MICKY
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/28912
• Institution: Institut Teknologi Bandung

Answer finder system aims to answer user's questions by searching answers from related passages. In Indonesian, researches related to answer finder system that use Deep Neural Networks is still very rare. In English, the state of the art architecture for answer finder is R-NET. R-NET is divided into 4 stages. The first stage is encoding which applies BiRNN to the vector of passages and questions. The second stage is question-passage matching which aims to provide information about the context of the question to the passage vector. The third stage is self-passage matching which aims to give information to the question-aware passage representation about the passage context. The final step is the answer pointer layer which aims to return the starting point and end point from the passage as an answer. <br /> <br /> <br /> <br /> <br /> This final project uses the R-NET architecture to build an Indonesian open domain answer finder. Modifications are made by adding POS Tag and NER information to the input layer of the R-NET architecture. Experiments were carried out in three stages. In the first stage, the experiment was carried out on three parameters which are hyperparameters of the R-NET architecture, namely hidden layer dimension, dropout rate, and initial learning rate. The second stage of the experiment was conducted on parameters related to word embedding, namely word embedding technique, word embedding dimension and the use of pretrained character embedding. The model built using parameters obtained in the first and second stage experiments is the baseline model. In the third stage, the experiment was carried out by adding POS Tag and NER information to the baseline model. <br /> <br /> <br /> <br /> <br /> By using 2,495 training data and 312 test data, the model with the best performance is the baseline model with addition of NER information. The metric score obtained from the model is EM-score 56.7308%, F1-score 68.3129% and MRR score 60.4274%. The addition of NER information increases the performance of the model built because most of the data has a factoid answer type which is namedentity. Adding POS Tag information does not affect the performance of the answer finder. <br />