Evaluasi Kinerja Penggabungan Knowledge Graph Embedded-Based Question Answering dan TransP pada Data Freebase

Abstract

In the past few years, data storage and analysis using graph keep increasing. One of the implementation of this is knowledge graph. There are many methods proposed on information extraction from knowledge graph, one of them is natural language question answering. However, all of the researches around question answering use direct query to find the answer. Knowledge Graph Embedding-based Question Answering (KEQA) is the latest method that implements deep learning and embedding to answer questions. Experiments demonstrate that KEQA outperforms other question answering methods. Despite having high accuracy, KEQA still uses simple and outdated embedding method.

Knowledge graph embedding is one of the method for knowledge graph representation where the entities and relations are represented in vector (embedding) using deep learning. Many proposed embedding methods do not really consider the depth of a knowledge graph. TransP is a proposed method that consider the indirect relationship to represent a knowledge graph. Experimental results show that TransP outperforms other embedding methods in the task given. Based on this, KEQA will be built using TransP with the expectation that the accuracy of KEQA will increase.

Based on the result of the experiment, TransP achieves Mean Rank of 5.390,25 and HIT10 of 28,5%. After that, KEQA with embedding can achieve up to 88,89% accuracy, and KEQA without embedding can achieve up to 88,89% accuracy. Experiment also shows that scoring parameters value with affect KEQA with embedding. In conclusion, TransP can increase the accuracy of KEQA.