The relationship between reading literary novels and predictive inference generation: A corpus-based approach employing a corpus of Japanese novels

This study examined the relationship between reading literary novels and generating predictive inferences by analyzing a corpus of Japanese novels. Latent semantic analysis (LSA) was used to capture the statistical structure of the corpus. Then, the authors asked 74 Japanese college students to generate predictive inferences (e.g., “The newspaper burned”) in response to Japanese event sentences (e.g., “A newspaper fell into a bonfire”) and obtained more than 5,000 predicted events. The analysis showed a significant relationship between LSA similarity between the event sentences and the predicted events and frequency of the predicted events. This result suggests that exposure to literary works may help develop readers’ inference generation skills. In addition, two vector operation methods for sentence vector constructions from word vectors were compared: the “Average” method and the “Predication Algorithm” method (Kintsch, 2001). The results support the superiority of the Predication Algorithm method over the Average method.