Named Entity Recognition (NER) is an important subtask of informationextraction that seeks to locate and recognise named entities. Despite recentachievements, we still face limitations with correctly detecting andclassifying entities, prominently in short and noisy text, such as Twitter. Animportant negative aspect in most of NER approaches is the high dependency onhand-crafted features and domain-specific knowledge, necessary to achievestate-of-the-art results. Thus, devising models to deal with suchlinguistically complex contexts is still challenging. In this paper, we proposea novel multi-level architecture that does not rely on any specific linguisticresource or encoded rule. Unlike traditional approaches, we use featuresextracted from images and text to classify named entities. Experimental testsagainst state-of-the-art NER for Twitter on the Ritter dataset presentcompetitive results (0.59 F-measure), indicating that this approach may leadtowards better NER models.
展开▼
