Our multilingual capacity is central to human cognition. Adult second language learning, in particular, sheds light on learning across the lifespan with far-reaching consequences for language learning practice as well as intercultural communication. This project investigates importantly how the mother tongue (L1) influences second language learning (L2) by looking at the linguistic distance between L1 and L2, whether L1 and L2 belong to the same language family and share typological features or whether they are typologically unrelated. It does so innovatively through the use of big data from online language learning, namely by looking at learner progress in English as a second language (L2 English) through the writings of thousands of learners from all around the world. Large scale L2 English data from online language learning provides an empirical bridge between developmental lab-based second language acquisition research and language proficiency in education. The research will help educators target the language background of learners with benefits particularly for learners from linguistically distant languages (e.g. Chinese, Japanese), at a time when language skills are becoming increasingly important [1].

We propose to exploit the recently developed Parametric Comparison Method (PCM) [2,4] from the subfield of linguistic typology and geneaology to provide a typological framework for the investigation of linguistic distance and its impact on L2. We will investigate 10 typologically diverse mother-tongues (L1s) such as Brazilian-Portuguese, Russian, Chinese and Arabic for thousands of language learners in the EF-Cambridge Open Language Database (EFCAMDAT) [10], an open access corpus developed at Cambridge. EFCAMDAT consists of L2 writings submitted to the online school of EF Education First, an international school of English as a foreign language. Available at http://corpus.mml.cam.ac.uk/efcamdat, EFCAMDAT is the largest open access corpus of its kind, with 1.2 million scripts summing 71.8 million words, and continues to grow. Our objective is to incorporate the impact of linguistic distance into current L2 learnability theories.

People

• Dora Alexopoulou (Principal Investigator)
• Ianthi Tsimpli (Co-Investigator)
• Xiaobin Chen (Research Associate)

To measure L1-L2 distance, we adopt the Parametric Comparison Method (PCM) [2-4]. Following the Principles and Parameters framework [15], PMC uses binary parameters to model cross-linguistic variation. Linguistic distance is measured through (the co-efficient of) identities and differences in parameter values. PMC yields measures refined enough to differentiate between as many as 28 languages and successfully distinguish between language genealogies [4]. Importantly, PMC defines implicational relations between parameters allowing us to capture clusters of superficial properties under few parameters.

We will use the PCM to characterise 10 typologically diverse L1s (Spanish, French, Italian, Brazilian, German, Chinese, Russian, Japanese, Korean, Arabic, Turkish). For instance, based on 63 (nominal) parameters, the PMC distance between English and German is 0,1111 (40 identical parameter settings:5 differences, 18 not applicable), with Spanish 0,1818 (36:8), Russian 0,3143 (24:11) and Arabic 0,4048 (24, 16). We will document the impact of PMC distance on individual features in morphology and syntax (e.g. article, verbal endings, subcategorisation frames etc.), syntactic complexity and the interface with semantics and discourse (aspectual distinctions, anaphoric marking). A variety of parameters will be used, mega, meso and micro [16].

To obtain a dataset rich enough for an investigation of typological effects across proficiency with significant learner numbers we will use EFCAMDAT. EFCAMDAT contains 128 distinct tasks across the proficiency spectrum drawing from learners across 170 nationalities. The empirical research will consist of a comprehensive corpus analysis exploiting EFCAMDAT (and smaller corpora like ICLE [17]). Psycholinguistic experiments will distinguish avoidance strategies [18,19] or possible data sparseness in production from true unlearnability. Outputs: The main output will be a book synthesizing the typological perspective with current learnability hypotheses, presenting an empirical investigation of unique scope.