Knowledge Graph Question Answering on Thai Retail Banking Products

Khongcharoen, W., Saetia, C., Chalothorn, T., Buabthong, P., "Question Answering Over Knowledge Graphs for Thai Retail Banking Products." 2022 17th International Joint Symposium on Artificial Intelligence and Natural Language Processing (iSAI-NLP), (2022). DOI: 10.1109/iSAI-NLP56921.2022.9960247

Our first NLP publication

tl;dr

• We looked at how to apply a method called "multi-hop KGQA" to the Thai language. This method helps the model reach answer nodes that are more than one hop away from the starting node.
• We manually built a knowledge graph based on banking products covering credit cards, debit cards, and deposit accounts.
• For typical single-answer-node questions, the model did pretty well in our tests, scoring 80.8 out of 100.
• Adding more data to the knowledge graph helped improve the performance.
• The work shows how this method can be applied to the Thai language and potentially help chatbots answer complex questions, especially in sectors like banking.

Comments

• The results indicate weaker performance for the translate dataset. Using translated dataset seems to be a common approach from other work when developing models in low-resource language setting. However, directly using the translator that I could potentially introduce propagation error from the translation model. On the other hand developing language specific model will pose a challenge when scaling the dataset.
• Using representation directly from other multilingual large language models could help address this problem, but human intervention maybe required in the early stages of the research.