[2G6-GS-6] Language media processing:

[2G6-GS-6-05] Orthographic Input Map for L2 Learners and L1 Map

A Comparison of the English Word Vectors of the Trained Model fastText to those generated from MEXT-approved Textbooks for Elementary Schools

〇Jun Kaneko¹, Takashi Otsuki², Takayuki Sakaguchi², Jesse Sokolovsky¹, Hiroki Shoyama³, Dai Inoue⁴ (1. Mie University, 2. Yamagata University, 3. The Junior High School attached to the Faculty of Education, Mie University, 4. The Elementary School attached to the Faculty of Education, Mie University)

Keywords:Embedding, Text Mining, Word Vectors, Education Aid

In response to the question of what can be done to “see” a native speaker’s own sense of language, a previous study developed L1 Map, an L1 vocabulary 3D map for English. At that time, it was found that connections between words which were natural to native speakers of English might not be intuitive to Japanese learners of English. In order to carry out a quantitative comparison, an “L2 Map” of the language sense of Japanese learners of English is necessary. However, as L2 language data would be difficult to collect, the present study instead used input which is central to L2 formation in Japan: that of elementary school English textbooks. Language data were drawn from MEXT-certified fifth- and sixth-grade English textbooks from six publishers. Word vectors were generated from these combined texts using fastText. A visualization was subsequently created through dimensionality reduction by reducing the word vectors to three dimensions. This visualization, called the Orthographic Input Map for L2 Learners, was compared with L1 Map (version 2.0). As a result, notable differences were found in the relations between specific words. Awareness of these differences may provide insights beneficial to both teachers in the classroom and students studying independently.