An Analysis of Radicals-based Features in Subjectivity Classification on Simplified Chinese Sentences

PACLIC 28 !495 An Analysis of Radicals-based Features in Subjectivity Classification on Simplified Chinese Sentences First Author Affiliation / Address line 1 Affiliation / Address line 2 Affiliation / Address line 3 email@domain Second Author Affiliation / Address line 1 Affiliation / Address line 2 Affiliation / Address line 3 email@domain Abstract Chinese radicals are linguistic elements smaller than Chinese characters1. Normally, a radical is a semantic category and almost all characters contain radicals or are radicals themselves. In subjectivity classification on sentences, we can use radicals to represent characters, which reduce the scale of word space while keep the subjectivity information. In this paper, we manually labeled a charac- ter set to build a high-quality radical-character mapping, and then the mapping is used to gen- eralize character-based features with radicals. In experiments, we at first evaluated the per- formance when directly generalizing charac- ters with radicals, and then offer a hypothesis that can reduce noises. Experiments show that this approach based on our hypothesis can reduce feature space while keep or improve the performance, which is es- pecially useful when the training samples are scarce. keyword: sentiment analysis, subjectivity classi- fication, radical, Chinese character 1 Introduction In sentiment analysis, an important task is subjec- tivity classification on sentences, which means clas- sify sentences as subjective or objective. This step’s performance greatly affects the following process- ing that is related with polarity or emotion etc. Here 1We use the terminology “character” for “Chinese charac- ter” in this paper. 1. !C![WI!å‹©!"øè""èà¢" Most classrooms of Yongle elementary school nearby are also classified as dangerous buildings. 2. M÷)„Øf\[!Ü" Boarding school life makes students tired and their parents irritable. we offer two sentences from NTCIR6 training cor- pus for subjectivity classification. For the first sentence, although à(dangerous) is a sentiment character, it is used to modify building, so the semantic emphasis of à¢(dangerous build- ing) is building, and also à¢(dangerous building) is somewhat known as a term, so normally is re- garded as objective, thus the whole sentence is la- beled as objective. For the second sentence, the subjectivity mainly comes from \(tired) and Ü(irritable). These two characters are also with different level of subjec- tivity. “tired” is a physical experience, compared with Ü(irritable), it is somewhat “objective”. But Ü(irritable) can surely make the sentence subjec- tive. If we take a further step, we can see that Ü(irritable) has a Chinese radical ª(fire), which can derive concepts of sentiment from linguistic per- spective. In real system, to label subjectivity sentences is of high cost especially when high quality is required. As we know, the size of common Chinese characters is around several thousands, while the size of radi- cals in Chinese is only around several hundreds, if we use the radicals to generalize character, it may overcome to some extent the sparseness problem Copyright 2014 by Ge Xu and Chu-Ren Huang 28th Pacific Asia Conference on Language, Information and Computation pages 495–502 Chu-Ren Huang PACLIC 28 !496 when training model and reduce the time …

Content truncated to 500 words