A Simple Yet Effective Approach for Named Entity Recognition from Transcribed Broadcast News

摘要

Automatic speech transcriptions pose serious challenges for NLP systems due to various peculiarities in the data. In this paper, we propose a simple approach for NER on speech transcriptions which achieves good results despite the peculiarities. The novelty of our approach is that it emphasizes on the maximum exploitation of the tokens, as they are, in the data. We developed a system for participating in the "NER on Transcribed Broadcast News" (closed) task of the EVALITA 2011 evaluation campaign where it was one of the best systems obtaining an F1-score of 57.02 on the automatic speech transcription test data. On the manual transcriptions of the same test data (although having no sentence boundary and punctuation symbol), the system achieves an F1-score of 73.54 which is quite high considering the fact that the system is language independent and uses no external dictionaries, gazetteers or ontologies.