PanPhon: A Resource for Mapping IPA Segments to Articulatory Feature Vectors

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3475–3484, Osaka, Japan, December 11-17 2016. PanPhon: A Resource for Mapping IPA Segments to Articulatory Feature Vectors David R. Mortensen, Patrick Littell, Akash Bharadwaj, Kartik Goyal, Chris Dyer, Lori Levin Carnegie Mellon University Language Technologies Institute 5000 Forbes Ave., Pittsburgh PA 15213 United State of America dmortens@cs.cmu.edu, plittell@cs.cmu.edu, akashb@cs.cmu.edu, kartikgo@cs.cmu.edu, cdyer@cs.cmu.edu, lsl@cs.cmu.edu Abstract This paper contributes to a growing body of evidence that—when coupled with appropriate machine-learning techniques—linguistically motivated, information-rich representations can out- perform one-hot encodings of linguistic data. In particular, we show that phonological features outperform character-based models using the PanPhon resource. PanPhon is a database relating over 5,000 IPA segments to 21 subsegmental articulatory features. We show that this database boosts performance in various NER-related tasks. Phonologically aware, neural CRF models built on PanPhon features are able to perform comparably to character-based models on monolin- gual Spanish and Turkish NER tasks. On transfer models (as between Uzbek and Turkish) they have been shown to perform better. Furthermore, PanPhon features also contribute measurably to Orthography-to-IPA conversion tasks. 1 Introduction This paper introduces PanPhon1, a resource consisting of a database that relates over 5,000 IPA segments (simple and complex) to their definitions in terms of 21 articulatory features (see Tab. 1) as well as a Python package for interacting with this database and manipulating the representations that it provides. While our previous publications (summarized in §4) have described experiments using it, this is the first full description of PanPhon. Combined with a sister package, Epitran2, it allows the conversion of syl son cons cont delrel lat nas strid voi sg cg ant cor distr lab hi lo back round tense long /p/ - - + - - - - 0 - - - + - 0 + - - - - 0 - /pʰ/ - - + - - - - 0 - + - + - 0 + - - - - 0 - /pʲ/ - - + - - - - 0 - - - + - 0 + + - - - 0 - /pʰʲ/ - - + - - - - 0 - + - + - 0 + + - - - 0 - Table 1: Illustration of IPA segments and feature vectors from PanPhon orthographic texts to sequences of articulatory feature vectors. The Epitran-Panphon pipeline is illustrated in Fig. 1. The input to Epitran consists of word tokens in orthographic representation. Take, for example, the Spanish word Madrid. Epitran converts this string to a phonemic (not phonetic) representation in IPA, in this case /madɾid/. Epitran then calls a PanPhon function to convert this IPA string into a sequence of feature vectors. It then returns this sequence, aligned with the orthographic representation, capitalization, and Unicode character category features. This work is licensed under a Creative Commons Attribution 4.0 International Licence. Licence details: http:// creativecommons.org/licenses/by/4.0/ 1https://github.com/dmort27/panphon 2https://github.com/dmort27/epitran 3475 tokens Epitran: orthography to IPA PanPhon: IPA to feature vector sequences feature vectors (aligned …

Content truncated to 500 words