Hierarchical Attention Retrieval (HAR) is a method of efficiently retrieving information for factoid and non-factoid queries (where traditional keyword-based retrieval models do not work very well). HAR uses deep attention at the word, sentence, and document levels. At the word-level, cross-attention enables the identification of relevant query terms. At the sentence and document levels, long and short documents can be retrieved.
Not every part of a document is relevant for answering a query. Sometimes a short answer may be all that is required. In this case, returning a large document with a small amount of relevant information is counter-productive. So, it’s important to model the interaction of words as opposed to focusing only their presence.