QUESTION ANSWERING

A Question Answering system provides concrete answers to questions. For example, if a user asks the question "Which country won the last World Cup?", the system must give "Spain" as an answer, and a set of websites in which it is possible to find information on the Football World Cup and/or the Spanish national team. This is a different behaviour from the one that the usual search engines such as Google or Yahoo!, in which the answer is a set of websites that the user must look through to find by himself the information he is searching. Besides the question answering systems are provided with any type of bases, and not only websites or specific databases.

In the current Question Answering systems (most of them are in an experimental state) it is possible to identify a sequence of common operations that would make it possible to define a generic architecture based on the following components:

• Question analysis. To be able to answer a question, it is necessary to have some characteristics among which, in almost all the cases, is the type of question and the type of answer expected. The type of question determines if you want to find a concrete fact (for example, "Which country won the 2006 last World Cup?"), and in this case you speak about factual questions, or if you want to get the definition of a concept (for example, "what is the osteoporosis?"), in which you speak about definition questions. Depending on the system considered, other types are included such as when it is a closed list question, that is, if a list of values is expected as an answer (for example, "which countries form the European Union?"). The type of answer expected that is, if you search for a person, an organization, a concrete date, etc. is also determined in this stage.

• Information retrieval. The Question Answering systems working on web contents need a mechanism that provides a list of websites that could contain the answer. These systems are named information retrieval systems and among them there are the most famous search engines.

• Selection of passages. When you have documents likely to have an answer, you need to study them to select the sentences or passages that could contain an answer to the question asked. The selected passages will be the ones used as the following component input.

• Answers extraction. Finally, you must look through the sentences containing the answers to extract the precise answer that the user wants. In this stage what is used is the type expected in the answer, as well as the type of question, to select the concrete words that form the answer or the sentence or passage that constitutes it.

As it can be guessed, in order to carry out this search process successfully, it is necessary to have a technology to make profound linguistic analysis of the texts involved. At Daedalus we work in the integration of our linguistic technology in a complete Question Answering system.

Before this initiative, at Daedalus we had already developed solutions with partial versions of this type of systems. The basic idea is to index little sized texts (passages – for example, paragraphs – instead of complete documents), and that is why it is possible to find better the answer to a determined consultation. Concretely, we have used these solutions to answer automatically consultations from Frequently Asked Questions (FAQ), of which we had versions in different languages, and also from user's guides. This is the type of problems you face in the call centres. Our reference client for this technology is Linguaserve.

Do you want more information about the Daedalus products applicable to Question Answering?