Synthetic Textual Features for the Large-Scale Detection of Basic-level Categories in English and Mandarin

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8294–8305 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 8294 Synthetic Textual Features for the Large-Scale Detection of Basic-level Categories in English and Mandarin Yiwen Chen and Simone Teufel Department of Computer Science and Technology University of Cambridge CB3 0FD JJ Thomson Road UK {yc429,sht25}@cl.cam.ac.uk Abstract Basic-level categories (BLC) are an important psycholinguistic concept introduced by Rosch et al. (1976); they are defined as the most in- clusive categories for which a concrete men- tal image of the category as a whole can be formed, and also as those categories which are acquired early in life. Rosch’s original algo- rithm for detecting BLC (called cue-validity) is based on the availability of semantic fea- tures such as ‘has tail’ for ‘cat’, and has re- mained untested at large. An at-scale algo- rithm for the automatic determination of BLC exists, but it operates without Rosch-style se- mantic features, and is thus unable to ver- ify Rosch’s hypothesis. We present the first method for the detection of BLC at scale that makes use of Rosch-style semantic features. For both English and Mandarin, we test three methods of generating such features for any synset within Wordnet (WN): extraction of tex- tual features from Wikipedia pages, Distribu- tional Memory (DM) and BART. The best of our methods outperforms the current SoA in BLC detection, with an accuracy of English BLC detection of 75.0%, and of Mandarin BLC detection 80.7% on a test set. When ap- plied to all of WordNet, our model predicts that 1,118 synsets in English Wordnet (1.4%) are BLC, far fewer than existing methods, and with a precision improvement of over 200% over these. As well as confirming the useful- ness of Rosch’s cue validity algorithm, we also developed and evaluated our own new indica- tor for BLC, which models the fact that BLC features tend to be BLC themselves. 1 Introduction Rosch et al. (1976) introduced the concept of Ba- sic Level Categories (BLC) to the psycholinguistic literature. She defined basic-level categories as the most general (inclusive) categories for which a concrete image can be formed, and hypothesised that the definition of concrete categories during Figure 1: Simplified Taxonomy of Concepts and Fea- tures. child development is based on semantic features of each individual in the child’s perception. She per- formed separate human prediction experiments on shape, affordance (what she called motor program) and semantic features. Rosch also presented an algorithm for finding the level in a hierarchy where BLC concepts are located, based on an information- theoretically defined property called cue validity, which is calculable if one has access to semantic features for each concept. BLC categories are those that are maximally dis- tinguished from their sibling categories by many features with strong distinctiveness. For example, in Fig. 1, the sibling BLC categories dog and cat have features such as good sense of smell and me- ows. Higher-level categories (“super”) are more general and contain features such …

Content truncated to 500 words