GENERATION OF ABSTRACT MEANING REPRESENTATION WITH GRAPH SEQUENCE ITERATIVE INFERENCE APPROACH

• Main Author: Tobing Alexandro, Christian
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/78302
• Institution: Institut Teknologi Bandung

This paper presents an Indonesian AMR parser that uses a one-stage parsing approach. One- stage parsing is a technique for parsing natural language sentences into Abstract Meaning Representation (AMR) in a single pass. This approach has not been previously developed for Indonesian AMR parsers. Therefore, this paper aims to implement the generation of AMR graphs using one-stage parsing. The one-stage parsing algorithm used is Graph Sequence Iterative Inference (GSII). GSII is an algorithm that parses a sentence into AMR by iteratively building a graph representation of the sentence by making decisions based on the input text and the state of the graph at that step. Each step completes several tasks such as predicting concepts, predicting relations, and predicting relation labels. An additional method used is to use a previous Indonesian AMR parser to generate a silver dataset that is used for the training process. Based on the experimental results, the best performance of the Indonesian AMR parser model was achieved using the IndoBERT-base-p2 model on the embedding process for the input sentence and also the addition of a silver dataset during the training process. The model with the best configuration obtained a SMATCH score of 0.87 for the simple sentence test data.