Browse Prior Art Database

Use of Ontology in Question-Answering

IP.com Disclosure Number: IPCOM000013791D
Original Publication Date: 2000-Mar-01
Included in the Prior Art Database: 2003-Jun-18
Document File: 1 page(s) / 40K

Publishing Venue

IBM

Abstract

Disclosed is a system and method for using an ontology in a Question-Answering system. A perpetual problem in Information Retrieval (IR) is that the vocabulary used by the user is often different than that used by the authors of the texts being searched. A particular vari- ation of this problem is particularly frequent in Question-Answering (QA): the user is seeking one or more members of a category, and uses the category name in a question, yet the texts contain just the member names. For example, the question might be "What animals are found in Africa" and the texts may talk about African lions and giraffes, but not mention the word animal. Since QA systems in particular, and IR systems in general, work by finding the greatest number of matches between texts and queries, this mis-match is a problem. The invention here is to use the technique of Predictive Annotation (PA) in conjunction with an ontology. An ontology can be represented in many ways, including trees, semantic networks or just plain lists of terms. The functionality is simply to establish a set of relation- ships of the form class-subclass or class-member. Predictive Annotation is the process of aug- menting the text prior to indexing with labels indicating the class of object the text represents, and modifying queries to include these labels instead of the question words. Thus PA would cause, for example, "14,692 feet", in the text"...ascent of the Matterhorn (14,692 feet)..." to be annotated as LENGTH$, and the question "How tall is the Matterhorn" to be transformed to "LENGTH$ Matterhorn", thus enabling the traditional bag-of-words matching performed by search engines to work. By extending this technique to a large number of class-member relationships, a much wider variety of questions can be accurately answered than previously. To use the African example again, if we suppose an ontology lists the fact that lions, giraffes and others are all animals, then an annotator prior to indexing may annotate each occurrence with the label ANIMAL$. The query processor will translate "What animal..." or "Which animal..." to ANIMAL$, thus ensuring the desired match will occur during search. Given an ontology, there are several ways of finding related terms. Taking the giraffe as a starting point, there are specific kinds of giraffe (e.g. Masai giraffe), other animal names (e.g. lion, zebra, antelope), parent class (e.g. mammal), grandparent class (e.g. animal) great- granparent class (e.g. living thing) and so on. We choose to use as the annotation what is known as the 'basic category' in this chain, which is 'animal'. 1

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 56% of the total text.

Page 1 of 1

Use of Ontology in Question-Answering

    Disclosed is a system and method for using an ontology in a Question-Answering system. A perpetual problem in Information Retrieval (IR) is that the vocabulary used by the user is often different than that used by the authors of the texts being searched. A particular vari- ation of this problem is particularly frequent in Question-Answering (QA): the user is seeking one or more members of a category, and uses the category name in a question, yet the texts contain just the member names. For example, the question might be "What animals are found in Africa" and the texts may talk about African lions and giraffes, but not mention the word animal. Since QA systems in particular, and IR systems in general, work by finding the greatest number of matches between texts and queries, this mis-match is a problem.

    The invention here is to use the technique of Predictive Annotation (PA) in conjunction with an ontology. An ontology can be represented in many ways, including trees, semantic networks or just plain lists of terms. The functionality is simply to establish a set of relation- ships of the form class-subclass or class-member. Predictive Annotation is the process of aug- menting the text prior to indexing with labels indicating the class of object the text represents, and modifying queries to include these labels instead of the question words. Thus PA would cause, for example, "14,692 feet", in the text"...ascent of the Matterhorn...