Domain Adaptation with Filtering for Named Entity Extraction of Japanese Anime-Related Words

Proceedings of Recent Advances in Natural Language Processing, pages 291–297, Hissar, Bulgaria, Sep 7–9 2015. Domain Adaptation with Filtering for Named Entity Extraction of Japanese Anime-Related Words Kanako KOMIYA 1 Daichi EDAMURA 2 Ryuta TAMURA 2 Minoru SASAKI 1 Hiroyuki SHINNOU 1 Yoshiyuki KOTANI 2 Ibaraki University 1 Tokyo University of Agriculture and Technology 2 {kkomiya, msasaki, shinnou}@mx.ibaraki.ac.jp edaedaichi@gmail.com, {50015646125@st, kotani@cc}.tuat.ac.jp Abstract We developed a system to extract Japanese anime-related words, i.e., Japanese NEs (named entities) in the anime-related do- main. Since the NEs in the area, such as the titles of anime or the names of charac- ters, were domain-specific, we started by building a tagged corpus and then used it for the experiments. We examined to see if the existing corpora were useful to im- prove the results. The experiments con- ducted using Conditional Random Fields showed that the effect of domain adap- tation varied according to the genres of the corpora, but the filtering of the source data not only reduced the time for training but also assisted in the domain adaptation work. 1 Introduction Japanese pop culture such as that exemplified by manga, anime, and gaming has recently gained popularity with the younger generations. They are commercially important; if the NEs (named enti- ties) in the area, such as the titles of the anime and the names of the character could be automatically obtained, they would be useful for the product search, identification, or recommendation. There- fore, we developed a system to extract Japanese NEs in these areas using Conditional Random Fields (CRF). Since the NEs in the area (see Sec- tion 3), such as the titles of the anime or the names of the characters, are domain-specific, we build a tagged corpus in the anime domain (see Sec- tion 5). We examined to see if the existing cor- pora were useful with the anime corpus using do- main adaptation, since tagging of corpora is time- consuming. The experiments (see Sections 4 and 6) showed that the effect of domain adaptation var- ied according to the genres of the corpora, but fil- tering the source data not only reduced the time for training but also asisted in the domain adapta- tion work. The F-measure was the best when the newspaper and anime corpora were used simulta- neously with the domain adaptation after the filter- ing of the newspaper data (see Section 7). 2 Related Work Named entity recognition (NER), which involves seeking to locate and classify elements in text into predefined categories, such as the names of people, organizations, and locations, has recently been intensively studied. There are two types of NER methods, i.e., NER using pattern match- ing and NER using supervised machine learning. NER using pattern matching finds elements in text that match the manually predefined patterns, i.e., the character strings that tend to co-occur with the NEs, e.g., “Mr.” or “University” (Takemoto et al., 2001). There are some works on analyzing or creating these patterns; (Lertcheva and Aroon- manakun, 2011) have analyzed the …

Content truncated to 500 words