Identifying Prepositional Phrases in Chinese Patent Texts with Rule-based and CRF Methods

Identifying Prepositional Phrases in Chinese Patent Texts with Rule-based and CRF Methods Hongzheng Li and Yaohong Jin Institute of Chinese Information Processing, Beijing Normal University 19, Xinjiekou Wai St., Haidian District, Beijing, 100875, China lihongzheng@mail.bnu.edu.cn, jinyaohong@bnu.edu.cn Abstract Identification of prepositional phrases (PP) has been an issue in the field of Natural Language Processing (NLP). In this paper, towards Chinese patent texts, we present a rule-based method and a CRF-based method to identify the PPs. In the rule-based method, according to the special features and expressions of PPs, we manually write targeted formal identification rules; in the CRF approach, after labelling the sentences with features, a typical CRF toolkit is exploited to train the model for identifying PPs. We then conduct some experiments to test the performance of the two methods, and final precision rates are over 90%, indicating the proposed methods are effective and feasible. 1 Introduction In recent years, patent text information processing (such as patent machine translation) has gradually become an important application field of natural language processing (NLP), and has aroused widespread attention. Prepositional phrases (PPs), as an important type of phrase, are widely distributed in Chinese patent text, in which the vast majority serve as adverbial components. According to (Li, et al., 2014), in a random sample of 500 Chinese patent sentences, the number of sentences containing PPs are 226, accounting for 45.2% of the total sample, indicating the high proportion of PPs. In the sentence S = W1,W2,W3......Wn, assuming the string Wi, Wi+1......Wj is the PP to be identified, the main task of identifying PP is to recognize the word Wi and Wj as left and right boundaries of PP, and identify the whole string as PP chunk. Since Wi is the preposition itself, thus the key issue is to determine the position of Wj. There exists some following difficulties in identifying PP of Chinese patent texts: (1) Different with other domain texts, PPs in the patent texts are much longer, with more characters. According to (Gan, et al., 2005; Hu, 2015), the average length of PPs in news texts has 4.9 characters, while 12.3 characters in patent texts. On the other hand, PPs tend to have much more complex structures, which can be composed of prepositions and various kinds of phrases, or even clauses. (2) Prepositions in Chinese are usually multi- category words, they can also serve as nouns, quantifiers, adjectives, conjunctions and verbs in different contexts. (3) Several parallel or nested PPs can appear in the same sentence. Here is an example sentence in the patent texts: 本发明[PP1 在条件允许的情况下][ PP2 通过 [PP3 为不同区域]提供预测信息]而提出了许 多更加准确的结果。 (The invention has proposed more accurate results [PP1 under the permitted condition] {PP2 by providing forecast information [PP3 for various regions.]}) As shown, two parallel PP1 and PP2 appear together in the same sentence, where PP2 also includes a nested PP3. PACLIC 29 143 29th Pacific Asia Conference on Language, Information and Computation pages 143 - 149 Shanghai, China, October 30 - November 1, 2015 Copyright 2015 by Hongzheng …

Content truncated to 500 words