A Semantic Search Model Using Word Embedding, POS Tagging, and Named Entity Recognition
Hyunwoo Yoo, Minseok Kang, Kyung-Whan Oh · 2018
We are in trouble to remember the title of a movie. Even though we remember some scenes, it is difficult to recall words in the movie title. We also try to find the movie title with some web search systems, but in most cases, it is unsatisfactory. It is because most search systems available now do not work well for multiple word description. To overcome this problem, this study considers to use word embeddings, part-of-speech (POS) tagging, and named entity recognition (NER). Word embedding models have shown high potential in capturing a precise syntactic and semantic word relationships. POS tagging is a traditional technique to analyze a sentence to put each word a POS tag. NER model also provides a technique to analyze a sentence with named entity tag. This study evaluates the proposed system by conducting an experiment to search a movie title with several descriptive words for a specific movie. As an initial step, ontologies of 30 movies are developed. Then, it is asked to provide three, five, and seven words about any of the 30 movies. Finally, with these three, five, and seven words, it is tested if the proposed system can find the correct movie title. A total of 50 sample tests is repeated and the result shows that the hit ratio is about 50 to 80%. This hit ratio is, in fact, a very high one compared to those of Google and Naver in Korea.