SEMANTIC TEXTUAL SIMILARITY (STS) FOR INDONESIAN SENTENCE USING SIAMESE NEURAL NETWORK

• Main Author: Baptiso Sorlawan, Agung
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/54226
• Institution: Institut Teknologi Bandung

Semantic Textual Similarity (STS) is a task in natural language processing that deals with determining how similar two sentences are. STS is a very important component in solving other natural language processing tasks such as semantic search, summarization, question answering, plagiarism detection, and information extraction. One of the architecture which is the focus of this research that can be used to model STS is Siamese Neural Network (SNN). One of the most important components in STS is the encoder. The encoder maps sentences to numerical vectors. In this research, experiments are being made to various kinds of SNN encoder. Other than that, experiments also being done to the other components of SNN as well, which are the pooling layer and the objective function. The dataset used in these experiments is acquired from Prosa.ai which contains frequently asked questions (FAQ) sentences. From the experiments, the best STS model achieves an f1­score of 0,9723 which is better than the baseline. That model is SNN with IndoBERT encoder, MEAN + CLS pooling and regression objective function.