Position-Aligned translation model for citation recommendation

The goal of a citation recommendation system is to suggest some references for a snippet in an article or a book, and this is very useful for both authors and the readers. The citation recommendation problem can be cast as an information retrieval problem, in which the query is the snippet from an article, and the relevant documents are the cited articles. In reality, the citation snippet and the cited articles may be described in different terms, and this makes the citation recommendation task difficult. Translation model is very useful in bridging the vocabulary gap between queries and documents in information retrieval. It can be trained on a collection of query and document pairs, which are assumed to be parallel. However, such training data contains much noise: a relevant document usually contains some relevant parts along with irrelevant ones. In particular, the citation snippet may only mention only some parts of the cited article's content. To cope with this problem, in this paper, we propose a method to train translation models on such noisy data, called position-aligned translation model. This model tries to align the query to the most relevant parts of the document, so that the estimated translation probabilities could rely more on them. We test this model in a citation recommendation task for scientific papers. Our experiments show that the proposed method can significantly improve the previous retrieval methods based on translation models.