Approaches to implement and evaluate aggregated search

Aggregated search or aggregated retrieval can be seen as a third paradigm for information retrieval following the boolean retrieval paradigm and the ranked retrieval paradigm. In the first two, we are returned respectively sets and ranked lists of search results. It is up to the time-poor user to scroll this set/list, scan within different documents and assemble his/her information need. Alternatively, aggregated search not only aims the identification of relevant information nuggets, but also the assembly of these nuggets into a coherent answer. In this work, we present at first an analysis of related work to aggregated search which is analyzed with a general framework composed of three steps: query dispatching, nugget retrieval and result aggregation. Existing work is listed aside different related domains such as relational search, federated search, question answering, natural language generation, etc. Within the possible research directions, we have then focused on two directions we believe promise the most namely : relational aggregated search and crossvertical aggregated search. ∙ Relational aggregated search targets relevant information, but also relations between relevant information nuggets which are to be used to assemble reasonably the final answer. In particular, there are three types of queries which would easily benefit from this paradigm: attribute queries (e.g. president of France, GDP of Italy, major of Glasgow, . . . ), instance queries (e.g. France, Italy, Glasgow, Nokia e72, . . . ) and class queries (countries, French cities, Nokia mobile phones, . . . ). We call these queries as relational queries and we tackle with three important problems concerning the information retrieval and aggregation for these types of queries. First, we propose an attribute retrieval approach after arguing that attribute retrieval is one of the crucial problems to be solved. Our approach relies on the HTML tables in the Web. It is capable to identify useful and relevant tables which are used to extract relevant attributes for whatever queries. The different experimental results show that our approach is effective, it can answer many queries with high coverage and it outperforms state of the art techniques. Second, we deal with result aggregation where we are given relevant instances and attributes for a given query. The problem is particularly interesting for class queries where the final answer will be a table