Disambiguating named entities (NE) in running text to their correct interpretations in a specific knowledge base (KB) is an important problem in NLP. This paper presents two collective disambiguation approaches using a graph representation where possible KB candidates for NE textual mentions are represented as nodes and the coherence relations between different NE candidates are represented by edges. Each node has a local confidence score and each edge has a weight. The first approach uses Page-Rank (PR) to rank all nodes and selects a candidate based on PR score combined with local confidence score. The second approach uses an adapted Clique Partitioning technique to find the most weighted clique and expands this clique until all NE textual mentions are disambiguated. Experiments on 27,819 NE textual mentions show the effectiveness of both approaches, outperforming both baseline and state-of-the-art approaches.
展开▼