Comparing Speech and Text Classification on ICNALE

Comparing Speech and Text Classification on ICNALE Sergiu Nisioi Center for Computational Linguistics, Faculty of Mathematics and Computer Science, University of Bucharest Academiei 14, Bucaharest, Romania sergiu.nisioi@gmail.com Abstract In this paper we explore and compare a speech and text classification approach on a corpus of native and non-native English speakers. We experiment on a subset of the International Corpus Network of Asian Learners of English containing the recorded speeches and the equivalent text transcriptions. Our results suggest a high correlation between the spoken and written classification results, showing that native accent is highly correlated with grammatical structures found in text. Keywords: Native Language Identification, Accent Recognition, Learner Corpora, Text Classification 1. Introduction When students learn a new language, they commonly make use of the grammatical rules specific to their native lan- guage (NL) to produce utterances in the target language (TL). The learning process is shaped by a so-called cross- linguistic influence (Kellerman and Sharwood-Smith, 1986; Arabski, 2006) that involves not only the NL of a learner, but also additional other languages he may have acquired before TL. On one hand, this linguistic information can contribute to a better assimilation of new grammatical rules and on the other hand, it can also impede the developmental process by erroneously transferring into the TL utterances. Interlanguage is a system of grammatical rules that emerges when learners - both children and adults - express meaning in the TL (Selinker and Rutherford, 2014). It represents a complete linguistic system, covering aspects such as vocab- ulary use, morphology, phonology or syntax. Herein we plan to investigate two aspects of interlanguage in foreign students of English: the first regards their English accent and its distinctiveness and the second one is related to the particular syntactical patterns occurring in speakers that share the same native language. In particular, we make use of the International Corpus Network of Asian Learners of English (Ishikawa, 2013)1 to train a machine learning classifier on the speeches and transcripts corresponding to students from different Asian countries. In order to clas- sify the speeches we repurpose a set of features previously used for gender and affect detection. Therefore, we estab- lish a first baseline for ICNALE on both speech and text and provide a comparative analysis between these two distinct aspects of interlanguage. Identifying the native language (NLI) can be of crucial im- portance for a wide range of NLP applications, from train- ing better language models to more robust speech recogni- tion systems that can “comprehend” non-native speech/text and accent. Error detection and correction are also key tasks that can be improved on account of native language information. Furthermore, from a second language acqui- 1ICNALE - http://language.sakura.ne.jp/ icnale/ sition perspective, NLI tools can help consolidate previ- ous linguistic hypotheses and improve the quality of lan- guage teaching and learning: systems can help students self-evaluate and prevent them for making mistakes while teachers can track the learning process and the recurring problems more easily. 2. Previous Work Native language identification is a prolific …

Content truncated to 500 words