FLAIRS Poster Abstracts

Efforts have been made in the last decade for collecting and annotating datasets for textual similarity. But the problem is that some of the linguistic features which are very important in practical applications (such as negation) covered in existing textual similarity corpora generated automatically from collections of texts or collected using Mechanical Turk appear to be very limited. There is still a need for a balanced, linguistically richer, and larger corpus to better facilitate the development of robust approaches to the task of semantic similarity. Here, we present a web based tool called “SEM-COLLECTOR” developed with the aim of collecting pairs of texts that may or may not be semantically similar from people who have very good understanding of the problem domain and have the linguistic background necessary to create a high-quality textual similarity corpus. Using SEM-COLLECTOR, users can contribute in a various ways. A user can add new sentence pairs which are similar in meaning or which may have some differences, and provide a rating with respect to their similarity using a value in the range of [0, 10]. On the other hand, the tool can ask users to paraphrase for a given sentence and then ask to rate the similarity. In addition, one can simply browse through the list of existing sentence pairs and annotate for similarity. In each case, the user is asked to select the important aspect of each pair (such as, lexical overlap, temporal information) that is important for creating diverse and balanced datasets. Multilingual Twitter Summarization