Neural Ranking Models for Information Retrieval

The objective of this thesis is to apply shallow or deep neural networks to information retrieval. Text retrieval task consists on ranking text documents in response to a query. This task comes with some challenges: (1) query-document vocabulary mismatch problem and semantic understanding (a good IR system may consider the terms “hot” and “warm” related, as well as “dog” and “puppy”, but must also distinguish that a user who submits the query “hot dog” is not looking for a “warm puppy”); (2) to interpret words based on context (given a query about Prime Minister of UK it may be obvious from context whether it refers to John Major or Theresa May); (3) to find relevant documents that may also contain large sections of irrelevant text; etc. In this thesis the student will explore the use of shallow and deep neural networks to overcome these challenges of information retrieval tasks. Most existing shallow neural methods for IR focus on inexact matching using term embeddings, comparing the query with the document directly in the embedding space or using embeddings for query expansion. An alternative is to directly incorporate word embeddings within deep neural network models, developing an end-to-end neural network architecture for information retrieval.