Computing Thresholds of Linguistic Saliency

Computing Thresholds of Linguistic Saliency* Siaw-Fong Chunga , Kathleen Ahrensa Chung-Ping Chengb, Chu-Ren Huangc , Petr Šimonc,d aGraduate Institute of Linguistics, National Taiwan University, No.1, Roosevelt Road, Section 4, Taipei 106, Taiwan. bDepartment of Psychology, National Chengchi University, No. 64, ZhiNan Road, Section 2, Taipei 11605, Taiwan. cInstitute of Linguistics, Academia Sinica, No. 128, Academia Road, Section 2, Nangang, Taipei 115, Taiwan. dTIGP-CLCLP, Institute of Linguistics, Academia Sinica, No. 128, Academia Road, Section 2, Nangang, Taipei 115, Taiwan. f91142002@ntu.edu.tw, kathleenahrens@yahoo.com, cpcheng@nccu.edu.tw, churen@gate.sinica.edu.tw, petr.simon@gmail.com Abstract. We propose and test several computational methods to automatically determine possible saliency cut-off points in Sketch Engine (Kilgarriff and Tugwell, 2001). Sketch Engine currently displays collocations in descending importance, as well as according to grammatical relations. However, Sketch Engine does not provide suggestions for a cut-off point such that any items above this cut-off point may be considered significantly salient. This proposal suggests improvement to the present Sketch Engine interface by calculating three different cut-off point methods, so that the presentation of results can be made more meaningful to users. In addition, our findings also contribute to linguistic analyses based on empirical data. Keywords: saliency, cut-off point, threshold, collocations 1. Introduction All lexical resources, at the point of their design, will take into consideration whether the resources are useful to a target group. For example, WordNet (Fellbaum, 1998) was originally designed for psychologists, but later was used extensively by computational linguists. Similarly, corpora such as British National Corpus (BNC), the Academia Sinica Corpus of Mandarin Chinese (Chen et al., 1996) and the Gigaword corpus were also designed for the use of target groups such as lexicographers, linguists, language teachers, language learners, etc. These corpora usually provide some forms of statistical analyses so that users will be able to summarize their research results quickly. For example, many corpora provide collocational measures such as Mutual Information values (Church and Hanks, 1989) so that collocated words can be sorted according to their frequency of co-occurrence. Sketch Engine (Kilgarriff and Tugwell, 2001) is a powerful resource which displays search summary in collocated patterns, as well as according to grammatical relations. However, like many other resources, Sketch Engine is unable to determine which of the results in the list are meaningful linguistically. Therefore, when provided with collocation lists, most linguists report the top “few,” based on their preferences. Some linguists report the top one or two and keep the rest in appendixes. In fact, the current search summary from corpora or lexical resources does not give enough * We would like to thank Professor Shu-Chuan Tseng for her comments on this paper, and for suggesting the idea underlying third method for calculating cut-off points. Copyright 2007 by Siaw-Fong Chung, Kathleen Ahrens, Chung-Ping Cheng, Chu-Ren Huang and Petr Šimon. 126 information regarding which of the collocational patterns are significantly different from the bottom words. In this paper, a research question is asked, i.e., whether or not one can select top rankings from linguistic results using principled measures. This selection of top rankings …

Content truncated to 500 words