Fluid Annotation: A Granularity-aware Annotation Tool for Chinese Word Fluidity

Fluid Annotation: A Granularity-aware Annotation Tool for Chinese Word Fluidity Shu-Kai Hsieh1, Yu-Hsiang Tseng2, Chih-Yao Lee3, Chiung-Yu Chiang4 1,3,4Graduate Institute of Linguistics, 2Department of Psychology 1,2,3,4Knowledge-Yielding Language Engineering (KYLE) National Taiwan University, Taipei, Taiwan {shukai, seantyh, chiyaolee, sonnejoy}@gmail.com Abstract This paper presents a novel word granularity-aware annotation framework for Chinese. Anchored in current functionalist linguistics, this model rearranges the boundary of word segmentation and linguistic annotation, and gears toward a deeper understanding of lexical units and their behavior. The web-based annotation UI also supports flexible annotation tasks for various linguistic and affective phenomena. Keywords: Wordhood, annotation, Chinese Word Segmentation 1. Introduction Word segmentation has been one of the most important preprocessing NLP tasks in the pipeline-alike architecture for languages without explicit word delimiter in their writ- ten forms. The engineering treatment of word segmenta- tion naturally leads to the requirement of the existing gold standard, including a presumably agreeable standard and a word-segmented corpus based on the standard. Unfortu- nately, this long-standing rationale does not provide a con- vincing argument that concurs with current findings of cog- nitive science. Words as conventionalized symbols which presents the function by which meaning is attached to form. However, the basic units of cognition are clearly not words. Theoreti- cal and empirical advances in the past decade have revealed that word meanings are only pointers to coherent chunks of encyclopedic knowledge (Malt and Wolff, 2010). In the light of reading task,(Liu et al., 2013) show that Chinese reader did not follow the segmentation rules, and tended to chunk single words into large information units, imply- ing that word meanings sometimes work against the way knowledge is organized in memory. In this paper, we argue that word-meaning pair is fluid in nature, whose granularity (in terms of the length of the word) is influenced by its underlying ontology (paradig- matic dimension), surrounding context (syntagmatic di- mension) and real-world application (pragmatic force). Un- der this view, word segmentation can be considered as wordhood annotation, disentangling itself from the error- prone role in the NLP pipeline architecture. 2. Review Word segmentation has been a thorny issue in NLP for many decades. In addition to structural ambiguity reso- lution and unknown word detection, the current focus is concerned with propagation error and domain adaption. As the pre-processing task in the pipeline architecture, word segmentation errors can propagate to later process- ing stages. To handle with this, joint approaches exploiting various machine learning models including the latest neural network have been proposed (Lyu et al., 2016; Shao et al., 2017). Second, it has been recognized that different appli- cations and domains have different calls for different granu- larities of word segmentation. Recent neural domain adap- tation approaches also work through cross-domain embed- dings to improve the cross-domain performance (Cai and Zhao, 2016; Zhang et al., 2014). However, a critical exam- ination of the underlying assumption, and their assessment in the light of naturally occurring linguistic data, reveal its inherent contradictions (Taylor, 2012). In the follow- ing sections, we introduce …

Content truncated to 500 words