Feature-Rich Part-Of-Speech Tagging Using Deep Syntactic and Semantic Analysis

Proceedings of Recent Advances in Natural Language Processing, pages 224–231, Hissar, Bulgaria, Sep 7–9 2015. Feature-Rich Part-Of-Speech Tagging Using Deep Syntactic and Semantic Analysis Luchezar Jackov Institute for Bulgarian Language Bulgarian Academy of Sciences lucho@skycode.com Abstract This paper describes the implementation, im- provement and evaluation of the machine translation (MT) system proposed by Jackov (2014) when used as a feature-rich part-of- speech (POS) tagger for Bulgarian. The sys- tem does not rely on POS tagging for mor- phological disambiguation. Instead, all ambi- guities are considered in parsing hypotheses that are scored and the best one is used for tag- ging. The system does not use automatic train- ing on annotated corpora. Manually and auto- matically compiled linguistic resources are used for hypothesis derivation and scoring. BulTreeBank manually annotated corpus (Simov and Osenova, 2004) was used for eval- uation, error detection and improvement. 1 Introduction Part-of-speech (POS) tagging is the activity of labeling the words of a text with contextual tags describing the various grammatical features of the specific word usage. This is not trivial since many word forms are homonymous to other word forms. For instance, “water” is a noun in “I drink water” and a verb in “They water the gar- den”. Linguists normally classify the words into at least eight basic POS classes: noun, pronoun, adjective, verb, adverb, preposition, conjunction, and interjection. Sometimes the list is extended with numerals, determiners, particles, etc. but the number of classes rarely exceeds 15. Computational linguistics works with a larger inventory of POS tags, e.g., the Penn Treebank (Marcus et al., 1993) uses 48 tags: 36 for part-of- speech, and 12 for punctuation and currency symbols. The increase in the number of tags is partially due to finer granularity, e.g., there are special tags for determiners, particles, modal verbs, cardinal numerals, foreign words, existen- tial there, etc., but also to the desire to encode morphological information as part of the tags. POS tagging poses major challenges for mor- phologically complex languages whose tagsets encode a lot of additional morpho-syntactic fea- tures (for most of the basic POS categories), e.g., gender, number, person, etc. For example, Bul- TreeBank (Simov and Osenova, 2004) for Bul- garian uses 680 tags, while the Prague Depen- dency Treebank (Hajič, 1998) for Czech has over 1,400 tags (Georgiev et al., 2012). POS tagging is a form of disambiguation and in many cases a deep syntactic and semantic analysis is needed for correct tagging. An interesting approach for deep syntactic and semantic disambiguation was presented by Jack- ov (2014). However, the paper indicated that no evaluation of the system has been made. The goal of this paper is to present an evaluation of this system by using it as a feature-rich morpho- logical tagger for Bulgarian and comparing the system output to the BulTreeBank manually an- notated corpus for Bulgarian (Simov and Oseno- va, 2004). The proposed approach considers the input text as a sequence of tokens. Then for each token all possible lemmas are derived. Lemma se- quences …

Content truncated to 500 words