Constructing novel datasets for intent detection and ner in a korean healthcare advice system: guidelines and empirical results

초록

The demand for intelligent dialogue systems has increased rapidly in recent years. However, building such systems involves numerous complicated processes, including training data construction for language understanding. Although training data is essential in dialogue systems, it isn't easy to find consistent guidelines that can facilitate such construction. In this paper, we propose a systematic construction process for NLU datasets with guidelines, especially considering the specificity of dialogue data. Two new datasets are constructed following the proposed procedure. We suggest using question-answering (QA) data instead of dialogue data to overcome the data shortage issue. To the best of our knowledge, this is the first attempt at using QA data for constructing a dialogue system. The process is demonstrated with a concrete example from the healthcare domain, which has rarely been considered when studying dialogue systems. The target system aims to diagnose illnesses based on user symptoms and provide healthcare, the diagnostic techniques for which were written in Korean. We present in detail the method to define intent and entity types and slots from the QA data. The effectiveness of our approach is verified by the experimental results from two crucial language understanding tasks: query intent detection and medical entity recognition. The tasks were performed using four variations of a state-of-the-art language representation model called bidirectional encoder representations from transformers. We obtained a satisfactory result on both tasks with a best f1 score of 0.84 and 0.92 for the intent detection and medical entity recognition.

키워드