Traditional document clustering algorithms consider text-based features such as unique word count, concept count, etc. to cluster documents. Meanwhile, event mining is the extraction of specific events, their related sub-events, and the associated semantic relations from documents. This work discusses an approach to event mining through clustering. The Universal Networking Language -based subgraph, a semantic representation of the document, is used as the input for clustering. Our research focuses on exploring the use of three different feature sets for event clustering (...) and comparing the approaches used for specific event mining. In our previous work, the clustering algorithm used UNL-based event semantics to represent event context for clustering. However, this approach resulted in different events with similar semantics being clustered together. Hence, instead of considering only UNL event semantics, we considered assigning additional weights to similarity between event contexts with event-related attributes such as time, place, and persons. Although we get specific events in a single cluster, sub-events related to the specific events are not necessarily in a single cluster. Therefore, to improve our cluster efficiency, connective terms between two sentences and their representation as UNL subgraphs were also considered for similarity determination. By combining UNL semantics, event-specific arguments similarity, and connective term concepts between sentences, we were able to obtain clusters for specific events and their sub-events. We have used 112 000 Tamil documents from the Forum for Information Retrieval Evaluation data corpus and achieved good results. We have also compared our approach with the previous state-of-the-art approach for Router-RCV1 corpus and achieved 30% improvements in precision. (shrink)

Coreference resolution is a challenging natural language processing task, and it is difficult to identify the correct mentions of an entity that can be any noun or noun phrase. In this article, a semisupervised, two-stage pattern-based bootstrapping approach is proposed for the coreference resolution task. During Stage 1, the possible mentions are identified using word-based features, and during Stage 2, the correct mentions are identified by filtering the non-coreferents of an entity using statistical measures and graph-based features. Whereas the existing (...) approaches use morphosyntactic and number/gender agreement features, the proposed approach uses semantic graph-based context-level semantics and nested noun phrases in the correct mentions identification. Moreover, mentions without the number/gender information are identified, using the context-based features of the semantic graph. The evaluation performed for the coreference resolution shows significant improvements, when compared with the word association-based bootstrapping systems. (shrink)