US20140142920A1

US20140142920A1 - Method and apparatus for Utilizing Structural Information in Semi-Structured Documents to Generate Candidates for Question Answering Systems

Info

Publication number: US20140142920A1
Application number: US12/191,251
Authority: US
Inventors: Jennifer Chu-Carroll; James J. Fan; David Angelo Ferrucci
Original assignee: International Business Machines Corp
Current assignee: International Business Machines Corp
Priority date: 2008-08-13
Filing date: 2008-08-13
Publication date: 2014-05-22
Also published as: US20150254244A1; US9529845B2

Abstract

An approach to candidate answer generation by leveraging structural information in semi-structured resources, such as the title of a document and anchor texts in a document.

Description

FIELD OF THE INVENTION

The invention relates generally to question answering systems used to generate candidate answers by consulting a possibly heterogeneous collection of structured, semi-structured, and unstructured information resources.

RELATED ART

Most question answering (QA) systems suffer from two significant deficiencies. First, the systems rely on the question analysis component correctly identifying the semantic type of the answer and the named entity recognizer correctly identifying the correct answer as that semantic type. Failure at either stage produces an error from which the system cannot recover.
Second, most QA systems are not amenable to questions without answer types, such as “What was the Parthenon converted into in 1460?” For such questions, oftentimes all noun phrases from the search output are extracted, leading to a large number of extraneous and at times non-sensible candidate answers in the context of the question.

SUMMARY

The present invention provides an approach to candidate answer generation by leveraging structural information in semi-structured resources, such as the title of a document and anchor texts in a document.
In one aspect, the invention provides a method for candidate generation for question answering including receiving a natural language question and formulating queries used to retrieve search results including documents and passages that are relevant to answering the natural language question; extracting from the search results potential answers to the natural language question; and scoring and ranking the answers to produce a final ranked list of answers with associated confidence scores.
The method for candidate generation for question answering includes receiving at least one document or passage together with its provenance information; accessing a semi-structured source of information based on the provenance; retrieving substructures/entities including a title of a document and anchor text from the passage within the document; applying a normalization operation, such as replacing the html symbol “&nsp;” with a space character or removing the disambiguation field in a Wikipedia article titles (e.g. removing the text in parenthesis for title Titanic (1997 film)), to the substructure/entity (e.g. titles and anchor texts); and returning the resulting list of candidate answers.
The approach improves upon previous generation methods by producing candidate answers in a context-dependent fashion without requiring high accuracy in answer type detection and named entity recognition. The approach is applicable to questions with both definitive semantic answer types as well as untyped questions, and in the latter case, improves overall system efficiency by generating a significantly smaller set of candidate answers through leveraging context-dependent structural information.
The objects and features of the present invention, which are believed to be novel, are set forth with particularity in the appended claims. The present invention, both as to its organization and manner of operation, together with further objects and advantages, may best be understood by reference to the following description, taken in connection with the accompanying drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

The foregoing features and other features of the present invention will now be described with reference to the drawings. In the drawings, the same components have the same reference numerals. The illustrated embodiment is intended to illustrate, but not to limit the invention. The drawings include the following Figures:

FIG. 1 illustrates the components of a canonical question answering system and its workflow;

FIG. 2 is a flow diagram illustrating an approach to candidate answer generation by leveraging structural information in semi-structured resources, such as the title of a document; and

FIG. 3 is a flow diagram illustrating an approach to candidate answer generation by leveraging structural information in semi-structured resources, such as anchor texts in a document.

DETAILED DESCRIPTION

The present invention may be described herein in terms of various components and processing steps. It should be appreciated that such components and steps may be realized by any number of hardware and software components configured to perform the specified functions. For example, the present invention may employ various electronic control devices, visual display devices, input terminals and the like, which may carry out a variety of functions under the control of one or more control systems, microprocessors or other control devices.
In addition, the present invention may be practiced in any number of contexts and the exemplary embodiments relating to a searching system and method as described herein are merely a few of the exemplary applications for the invention. The processing steps may be conducted with one or more computer-based systems through the use of one or more algorithms.
FIG. 1 illustrates the components of a QA system 100 and its workflow, including question analysis component 104, search component 106, candidate generation component 108 and answer selection component 110.
In operation, the question analysis component 104 receives a natural language question 102, for example, “Who is the 42nd president of the United States?” Question analysis component 104 analyzes the question to produce, minimally, the semantic type of the expected answer (in this example, “president”), and optionally other analysis results for downstream processing.
The search component 106 formulates queries from the output of question analysis and consults various resources, for example, the world wide web and databases 107, to retrieve documents, passages, database tuples, and the like, that are relevant to answering the question.
The candidate generation component 108 then extracts from the search results potential answers to the question, which are then scored and ranked by the answer selection component 110 to produce a final ranked list of answers 112 with associated confidence scores.
Candidate generation component 108 is an important component in question answering systems in which potential answers to a given question are extracted from the search results. In a typical question answering system, candidate answers are identified based on the semantic type match between the answer type as determined by the question analysis component 104 and entities extracted from the search results via a named entity recognizer. For example, for the sample question “Who is the 42nd president of the United States?” all candidate answers will be of the semantic type US president.
FIGS. 2 and 3 are flow diagrams illustrating an approach to candidate answer generation by leveraging structural information in semi-structured resources, such as the title of a document and anchor texts in a document. This approach improves upon previous candidate generation methods by producing candidate answers in a context-dependent fashion without the reliance requiring high accuracy in on the high accuracy required of answer type detection and named entity recognition.
This approach is applicable to questions with both definitive semantic answer types as well as untyped questions, and improves overall system efficiency by generating a significantly smaller set of candidate answers through leveraging context-dependent structural information.
In certain types of documents, such as Encyclopedia articles and the like, the document title is an excellent candidate answer for properties described in the article about the title entity. For example, in a document about the band “The First Edition”, the following facts are provided: “The First Edition was a rock band, stalwart members being Kenny Rogers, Mickey Jones, and Terry Williams. The band formed in 1967, with noted folk musicians Mike Settle and the operatically trained Thelma Camacho completing the lineup” and “The First Edition were (outside of Mickey Jones) made up of former New Christy Minstrels who felt creatively stifled.” Given the question “What is the rock band formed by Kenny Rogers and other members of the New Christy Minstrels in 1967?” the search component 106 of QA system 100 is likely to include a document 202, for example “The First Edition” or passage texts 204 extracted from document 202 among its search results.
In one embodiment of the present invention, candidate generation component 108 performs document title approach 200 by extracting candidate answers from search results. If the search results include documents 202, then the “title field” of these documents, such as “The First Edition”, is extracted using title retrieval component 208 as illustrated in FIG. 2. One implementation of title approach 200 may be done through a database table lookup. If the search results include passage texts 204, then the documents that contain passage texts 202 a are retrieved through document retrieval component 206. This may be accomplished by retrieving the provenance information of the passage texts. Once the documents containing the passage texts 202 a are obtained, the titles may be retrieved using title retrieval component 208 and are used as candidate answers 210.
In the event that search component 106 returns document 202 as its search result, title retrieval component 208 returns the title of document 202 as a candidate answer 210.
In the event that search component 106 returns a passage 204 (for example, a short 1-3 sentence text snippet), then a document 202 a from which passage 204 has been extracted is searched for and identified using document retrieval component 206.
Document retrieval component 206 is configured to match passage 204 against a set of free-text records. These records could be any type of mainly unstructured text, such as newspaper articles, real estate records or paragraphs in a manual. Passages 204 may range from multi-sentence full descriptions of an information need to a few words.
Once document 202 a has been identified, title retrieval component 208 returns the title of document 202 a as a candidate answer 210.
In another embodiment, candidate generation component 108 includes anchor text retrieval approach 300, which leverages anchor texts found in a passage/document 302 to extract candidate answers 210 from text retrieved from passage/document 302. Anchor texts are text strings highlighted in a document to indicate hyperlinks to other documents.
As illustrated in FIG. 3, candidate generation component 108 uses anchor texts 304 as a candidate generation mechanism in QA system 100. For each document-oriented search result (i.e. passage or document 302) from search component 106, retrieve component 306 identifies the document from which the retrieved text has been extracted. Retrieve component 306 then retrieves all anchor texts 304 that are present in document 302. An implementation of using anchor texts 304 may be through a database lookup in which the database stores pairs of a document ID and a list of all anchor texts in that particular document. Given a search result, such as passage/document 302, the document ID of passage/document 302 may either be obtained through provenance information in the search result, if available, or retrieved again through retrieve component 306. Once the document ID is identified, the list of anchor texts 304 in that document may be obtained through a simple database query. The subset of the list of anchor texts that are present in passage/document 302 are selected as candidate answers 210.
Next, in match component 308, anchor texts 304 are matched against the retrieved text and all anchor texts 304 that are present in the retrieved text are selected as candidate answers 210.
It should be understood that the approaches described above regarding FIGS. 2 and 3 to candidate answer generation by leveraging structural information in semi-structured resources, such as the title of a document and anchor texts in a document, may be used separately or may be combined. For example, in one embodiment, anchor text retrieval approach 300 in which candidate generation component 108 uses anchor texts 304 as a candidate generation mechanism in QA system 100, may be applied to the candidate answers 210 extracted using document title approach 200. This is performed by treating each extracted candidate answer 210 from approach 200 as a search result (i.e. passage/document 302) and further extracting anchor text sub-candidates from within candidate answers 210. For example, for candidate answer 210 “List of Deserts in Australia”, the candidate answer 210 “Australia” may be generated.
The invention has been disclosed in an illustrative manner. Accordingly, the terminology employed throughout should be read in an exemplary rather than a limiting manner. Although minor modifications of the invention will occur to those of ordinary skill in the art, it shall be understood that what is intended to be circumscribed within the scope of the patent warranted hereon are all such embodiments that reasonably fall within the scope of the advancement to the art hereby contributed, and that scope shall not be restricted, except in light of the appended claims and their equivalents.

Claims

1. A method for candidate generation for question answering comprising:

receiving a natural language question and formulating queries used to retrieve search results including documents and passages that are relevant to answering the natural language question;

receiving at least one document or passage together with its provenance information;

accessing a semi-structured source of information based on the provenance;

retrieving substructures/entities including a title of a document and anchor texts from the passage within the document;

extracting from the search results candidate answers to the natural language question;

selecting the title and the anchor texts as candidate answers;

applying a normalization operation to the substructure/entity;

scoring and ranking the answers to produce a final ranked list of candidate answers with associated confidence scores; and

returning the resulting list of the candidate answers.