Named Entity Disambiguation via Probabilistic Graphical Model with Embedding Features

Abstract

Listen

Translate article

Named entity disambiguation (NED) is the task of linking ambiguous mentions in text to their corresponding entities in a given knowledge base, such as Wikipedia. State-of-the-art NED solutions harness neural networks to generate abstract representations, i.e., embeddings, of mentions and entities, based on which the disambiguation process can be achieved by finding entity with the most similar representation to mention. Nevertheless, the coherence among mentions, and their corresponding entities, is yet neglected. To fill this gap, in this work, we put forward intra, an approach effectively integrating embedding features into a collective disambiguation framework, i.e., probabilistic graphical model. Markov Chain Monte Carlo sampling and SampleRank algorithm are implemented for model parameters learning and inference. We evaluate intra on existing dataset against several state-of-the-art NED systems, which validates the effectiveness of our proposed method.