ANSWER FINDER YES / NO IN INDONESIAN LANGUAGE BASED ON INFERENCE USING DEEP NEURAL NETWORK

• Main Author: Rahmawati, Sari
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/44159
• Institution: Institut Teknologi Bandung

Answer finder built on this system have yes / no answers to confirm the location of an object and object comparison. The system is made to answer questions by inference on a set of sentences. In the case of the search for answers used to build 4 Deep Neural Network architectures, that are RNN, BiLSTM, End to End Memory Network and Dynamic Memory Network. The RNN and BiLSTM models are sequence to sequence models used with fixed-length vectors. Both of these models have weaknesses in documents with sequences that are too long. The next model is the End to End memory network which has a memory mechanism that can write and read relevant document vectors more effectively, this model has advantages in faster performance compared to other models. The final model is the Dynamic Memory Network that works by using attention memory to solve problems experienced in the two previous models. Based on the experimental results, the Deep Neural Network method has good accuracy results in the case of object size comparison by getting 90% results in the End to End Memory Network architecture, Dynamic Memory Network and BiLSTM. In the object comparison dataset the best results are shown in End to End Memory Network and Dynamic Memory Network with an average of 90% and get the highest results 92% in test case 22 with a learning rate of 0.001 and a dropout value of 0.2. The results of 92% were obtained by the DMN case with a learning rate of 0.02 and a dropout of 0.5. The case of finding the location of the object gets good results on End to End Memory Network and BiLSTM with accuracy values above 80% in some experiments. The results of experiments conducted also have higher accuracy compared to other studies in Indonesian for reasoning cases. For End to End Memory Network architecture, it has more consistent results for each test case given with an average f1 score of 77%.