Improving the Performance of a Named Entity Extractor by Applying a Stacking Scheme

Description

In this paper we investigate the way of improving the performance of a Named Entity Extraction (NEE) system by applying machine learning techniques and corpus transformation. The main resources used in our experiments are the publicly available tagger TnT and a corpus of Spanish texts in which named entities occurrences are tagged with BIO tags. We split the NEE task into two subtasks 1) Named Entity Recognition (NER) that involves the identification of the group of words that make up the name of an entity and 2) Named Entity Classification (NEC) that determines the category of a named entity. We have focused our work on the improvement of the NER task, generating four different taggers with the same training corpus and combining them using a stacking scheme. We improve the baseline of the NER task (Fβ=1 value of 81.84) up to a value of 88.37. When a NEC module is added to the NER system the performance of the whole NEE task is also improved. A value of 70.47 is achieved from a baseline of 66.07.

Additional details