Quality Exploration of PLMs for Summary Recommendation Considering Tweet Acceleration for Earthquake Events

The present study outlined a summary recommender through analyzing the quality metrics of extractive Pre-trained Language Models (PLMs), such as Distilbert, T5, and Pegasus, for situational tweets of four earthquake events that occurred at distinct periods and in various geographical locations. The research took into account reference-free automatic quality evaluation factors, such as JS-Divergence and Rouge-1, 2, & L, based on content selection, linguistic quality, & informativeness of generated summaries at times concerning the corpus data. The amount of information density for each earthquake event is retrieved in terms of tweet acceleration levels. Further, the quality metrics of PLM-generated summaries are examined on such acceleration levels, considering the performance superiority and variability for earthquake events. Based on the interpretation, a recommendation approach has been outlined to suggest more appropriate PLM/PLMs for summary production for seismic occurrences at different acceleration levels. From results, it is evident that the summary produced Distilbert is more superior in Content Selection with minute Word Ordering at Low & Medium acceleration and along with Pegasus at High acceleration level. Whereas summaries of T5 & Pegasus mostly enclose the context of the tweet corpus on the mentioned acceleration levels in the crisis timeline.