Transfer of Frames from English FrameNet to Construct Chinese FrameNet: A Bilingual Corpus-Based Approach

Transfer of Frames from English FrameNet to Construct Chinese FrameNet: A Bilingual Corpus-based Approach Tsung-Han Yang, Hen-Hsen Huang, An-Zi Yen, and Hsin-Hsi Chen Department of Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan {thyang, hhhuang, azyen}@nlg.csie.ntu.edu.tw, hhchen@ntu.edu.tw Abstract Current publicly available Chinese FrameNet has a relatively low coverage of frames and lexical units compared with FrameNet in other languages. Frames are incompletely specified for some lexical units, and some critical lexical elements are even missing. That results in suitable frames cannot be triggered and filled in practical applications. This paper presents an automatic approach to constructing Chinese FrameNet. We first capture the mapping between English lexical entries and their Chinese counterparts in a large scale sentence-aligned English-Chinese bilingual corpus. Then, a semantic transfer approach is proposed based on word alignments applied to a large balanced bilingual corpus. The resource currently covers 779 frames and 36k lexical units. We apply it to annotate diary and tweet, and achieve overall 86% success rate to provide frame recommendations that are acceptable by annotators. The success rates in terms of source types are 95% and 80% for diaries and tweets respectively. Keywords: FrameNet, Chinese, frame semantics, lexical units 1. Introduction Semantic role labeling (SRL) is a fundamental task for many NLP applications. Given a context, SRL is aimed at identifying the semantic roles, or the set of semantic properties and relationships defined over a lexical unit (LU) or a target. The resources such as FrameNet (Baker et al., 1998; Fillmore et al., 2003) and PropBank (Kingsbury and Palmer, 2002) storing abundant information about lexical and predicate-argument semantics have advanced the field of semantic analysis further and made it possible for learning algorithms to train upon. 1.1 FrameNet FrameNet is a lexical database with rich human-annotated semantic content based on the linguistic theory of Frame Semantics proposed by Fillmore (1982). FrameNet provides both human and machine readable structure for semantic frames, associated with frame elements (FEs), lexical units (LUs), and sample sentences. The words that evoke a specific frame are called LUs or targets. Accompanied with each frame, there are a set of FEs defining semantic roles that is meaningful to that frame. Most importantly, FrameNet provides over 200,000 human annotated sentences associated to more than 1,200 semantic frames, and forms a great resource for analyzing semantic structures in natural language. 1.2 FrameNet in Chinese Unlike English FrameNet, which has been in operation since 1997, Chinese FrameNet (You et al, 2005) started constructing the resource in 2005. To date, Chinese FrameNet (CFN) contains 3,947 LUs, 323 semantic frames, and 20,000 annotated sentences. Compared with English FrameNet, in which 13,638 LUs, 1,221 semantic frames, and 200,000 annotated sentences are provided, Chinese FrameNet is considerably smaller. In other words, Chinese FrameNet has a much low coverage of frames and lexical units, and results in limited applications. 1.3 FrameNet Construction in Other Languages Efforts have been made to construct FrameNet resources in other languages. Most of which construct their resources with human annotation one by one …

Content truncated to 500 words