Usage of documents embedding for Person Name Disambiguation on the Web

Abstract This article introduces a new approach to improve person names’ disambiguation on Web based search using document embedding. This new approach drastically simplifies the pre-processing stage, while also outperforming state-of-the-art methods in a monolingual scenario. The addition of a summarization step, even if it comes with a drawback of increasing the computational cost, was able to significantly improve the performance of the whole pipeline. The methods were evaluated on the subset of English pages of the MC4WePS corpus with different metrics: BCubed precision, BCubed recall and F0.5-score. Therefore, the F0.5-score was improved by 5 percent compared to state-of-the-art on the English pages of the MC4WePS corpus. This allowed to take into account a realistic proportion of social media pages in the mix of search engine results.