Design of a Learner Corpus for Listening and Speaking Performance

Design of a Learner Corpus for Listening and Speaking Performance Katsunori Kotani Takehiko Yoshimi Kansai Gaidai University Ryukoku University 16-1 Nakamiyahigashino-cho, Hirakata, 1-5 Yokotani, Seta, Otsu, Osaka, Japan 573-1001 Shiga, Japan 520-7729 kkotani@kansaigaidai.ac.jp Abstract A learner corpus is a useful resource for developing automatic assessment techniques for implementation in a computer-assisted language learning system. However, presently, learner corpora are only helpful in terms of evaluating the accuracy of learner output (speaking and writing). Therefore, the present study proposes a learner corpus annotated with evaluation results regarding the accuracy and fluency of performance in speaking (output) and listening (input). 1 Introduction The linguistic properties of learners of English as a foreign language (EFL), which are different from those of native speakers, have been identified through the analysis of output compiled in learner corpora (Sugiura et al. 2007, Friginal et al. 2013, Barron and Black 2014). These properties have been used to statistically classify learners’ output into a range of proficiency levels (Thewissen 2013). Thus, a learner corpus is a useful linguistic resource for developing assessment techniques that are implementable in a computer-assisted language learning (CALL) system. Although the contribution of learner corpora is well acknowledged (Granger 2009), previous learner corpora are limited in that learners’ outputs have only been examined in terms of linguistic accuracy. As noted by Housen et al. (2012), learners’ performance should be analyzed in terms of both accuracy and fluency. On the one hand, it is true that a proficient learner uses a target language accurately; however, on the other hand, a trade-off is often observed, as a learner speaks grammatically correct sentences (high accuracy), but does so at an unnaturally slow speech rate (low fluency) (Brand and Götz 2011, Chang 2012). Another limitation is typically seen in the target skill. Most previous learner corpora cover output skills in spoken or written language. From the viewpoint of communicative competence in spoken language, learners need to be proficient not only in speaking (output), but also in listening (input). Although speaking proficiency is well correlated with listening proficiency, a gap between these proficiencies is also known to exist (Liao et al. 2010, Liu and Costanzo 2013), as a learner may comprehend some sentences containing lexical and syntactic items that are difficult for them to actually articulate. Because previous learner corpora have been limited in terms of speaking, a spoken learner corpus that demonstrates accuracy and fluency in speaking and listening is needed. The present study proposes to build a spoken learner corpus by annotating relevant information on sentences that learners spoke and listened to, respectively. Another limitation is seen in the scope of corpus data analysis. Although the learner corpus of Kotani et al. (2015) addressed listening, it was only capable of providing listening comprehension data for analysis at the text level, because that was the level at which comprehension was examined. However, identifying which linguistic properties affect listening comprehension through text-level analysis is difficult. To identify learners’ linguistic PACLIC 29 351 29th Pacific Asia Conference on Language, Information …

Content truncated to 500 words