Fine-tuning for Named Entity Recognition Using Part-of-Speech Tagging

Fine-tuning for Named Entity Recognition Using Part-of-Speech Tagging Masaya Suzuki Kanako Komiya Minoru Sasaki Hiroyuki Shinnou Ibaraki University 4-12-1, Nakanarusawa, Hitachi, Ibaraki, 316-8511, Japan {17nm708n, kanako.komiya.nlp}@vc.ibaraki.ac.jp {minoru.sasaki.01, hiroyuki.shinnou.0828}@vc.ibaraki.ac.jp Abstract In recent years, machine learning methods beyond the confines of conventional super- vised learning have been used along with deep learning methods and intensively investigated. Fine-tuning, which improves the performance of one task by re-learning using the weights of a model learned for another task as ini- tial values, is one such example. This paper proposes fine-tuning named entity recognition (NER) using part-of-speech tagging. The ex- periments revealed that fine-tuning improves the performance of NER. They also revealed that there was no performance improvement even if POS tag set included tags that corre- sponded to an NE tag, when there was a dif- ference in definitions between these tags. 1 Introduction In recent years, many researchers use machine learn- ing methods beyond the confines of conventional su- pervised learning along with deep learning meth- ods and intensively investigate them. In general, it is difficult to learn a model of a task with high performance using small training data when we use conventional supervised learning. However, if we had larger corpora of the other tasks that are similar to and related to the task, it could help to achieve high performance. Fine-tuning, which improves the performance of one task by re-learning using the weights of a model learned for another task as ini- tial values, is one way to realize it. Named entity recognition (NER) involves seek- ing to locate and classify elements in texts into predefined categories such as the names of people, organizations, and locations and is one of the se- quential labeling tasks. NER is related to part-of- speech (POS) tagging because POS tagging is also a sequential labeling task and POS are popular fea- tures for NER. Therefore, fine-tuning for NER us- ing POS tagging would be a good way to learn an NER model with high performance when we have a small NE corpus and a large POS corpus at the same time. This paper proposes fine-tuning NER using POS tagging. We targeted at Japanese NER and POS tagging in the current study. There are many researches on Japanese NER like (Komiya et al., 2018) and (Iwakura et al., 2016). In addi- tion, many researchers investigated NER using mul- titasking learning or joint learning like (Qu et al., 2016) and (Peng and Dredze, 2015). We investi- gated Japanese NER using fine-tuning (See Section 2). The proposed model is developed on the basis of a model described in (Ma and Hovy, 2016). We simplified the model by excluding a CNN for the character-level representations (See Section 3). We evaluated precision, recall, and F-measure based on the gold standard with experiments using the mod- els (described in Section 4). We discuss the results by comparing them to those of the existing method without fine-tuning (See Section 5) and conclude our work (See Section 6). 2 Related Work NER has …

Content truncated to 500 words