Evaluating LSA sensibility to disclosure in learners' interactions

Social technologies have been effectively applied in distant learning platforms to engage students in social interaction and active learning. Students' interaction is viewed as a process through which learners demonstrate their expertise and a transfer of assistance to their peers. Within this context, students may ask for peers' feedback when they encounter problems that they cannot solve themselves. In response to his request, a learner may receive a huge volume of peers' feedback which is not usually all positive, relevant and favourable to his learning. Discarding negative feedback (bullying, demeaning and those making an individual vulnerable by self-disclosing personal data) aligns to the first principle of e-learning environment which is to ensure for the learners to interact and collaborate free from fear in a safe e-learning environment. Scrutinizing the learners' natural language interactions requires the use of artificial intelligence techniques, namely Latent semantic Analysis (LSA). In this work, we analyse LSA's sensibility to the quality of peers' text-based interactions and its ability to discard negative and self-disclosing feedback. To do that, we built different LSA models by varying the approach parameters. We theorize that there would be positive correlations between LSA measures derived from the latent semantic space and human judges' rating. Regression analysis shows that once the LSA parameters that better represent the human judgments of feedback relevance and disclosure have been considered, LSA reliably predicts the human scores (r=.64, p<;.001).