English Island Constraints Revisited:Experimental vs. Deep Learning Approach

This paper examined English island constraints with a deep learning model and compared the analysis results with those of Lee and Park (2018). This paper adopted the BERTLARGE model which was pretrained with the CoLA dataset. The dataset which was used in the current study were composed of 480 sentences. Among them, 80 sentences came from the target sentences in Lee and Park (2018), and 400 sentences from the CoLA dataset. These 480 sentences were used as an input to the BERTLARGE model, and the acceptability scores were calculated. Unlike previous deep learning models for syntactic acceptability, however, this paper calculated the probability where the given sentence could be acceptable, and the probability was converted into the acceptability scores. After the acceptability scores were obtained for all the target sentences, they were normalized into the z-scores. Statistical analyses were applied to them to compare the results of the BERTLARGE model and the experimental results in Lee and Park (2018). As a result, it was found that the results of the BERTLARGE model showed patterns similar to those in Sprouse et al. (2012) and Lee and Park (2018). Through the analysis, the followings were observed: (i) deep learning models could represent human beings’ syntactic acceptability to a considerable extends, (ii) the BERTLARGE model clearly identified the English island constraints, and (iii) both main factors (Island and Location) and their interaction (Island:Location) influenced the acceptability scores of island sentences.