Seminar in Computational Linguistics

 In this talk I present recent and current work on the influence of document context on sentence acceptability judgments. Bernardy, Lappin, and Lau (2018) (BLL) report a puzzling compression effect in which context raises the mean ratings of sentences with low out-of-context mean assessments, but lowers the scores of sentences that are highly rated out of context. Their enriched LSTM language models achieve fairly strong Pearson coefficient correlations with mean human judgments for both out-of-context and in-context test sentences, with the former correlation being higher than the latter. The compression effect also shows up in an unrelated task that Burzzoni and Lappin (2019) report. I consider two possible explanations of the compression effect. I present some of the difficulties in experimental design that we have encountered in trying to reproduce this effect and test the two explanations for it. Finally, I discuss the set up of a new experiment that we are currently doing in an effort to clarify the role of document context on acceptability judgments. We will be applying the DNN models that (BLL) develop to these crowd sourced annotated test sets.