Non-factoid question-answering system and method

• US 10,496,928 B2
• Filed: 05/15/2014
• Issued: 12/03/2019
• Est. Priority Date: 06/27/2013
• Status: Active Grant

• Current Assignee: National Institute of Information and Communications Technology
• Original Assignee: National Institute of Information and Communications Technology
• Inventors: Oh, Jonghoon, Torisawa, Kentaro, Hashimoto, Chikara, Sano, Motoki, De Saeger, Stijn, Ootake, Kiyonori
• Primary Examiner(s): Chang, Li Wu
• Application Number:
  US14/900,758
• Publication Number:
  US 20160155058A1
• Time in Patent Office: 2,028 Days
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 20/10 using kernel methods, e.g. support vector machines [SVM]
• CPC Class Codes: G06N 5/041 Abduction

Abstract

A question-answering system capable of appropriately answering a non-factoid question. The question-answering system responsive to a question, for taking out answer candidates from an object document storage unit; extracting plausible ones from the answer candidates; recognizing causal relation expression included in extracted answer candidates; determining whether or not a recognized causal relation is relevant as an answer; generating a feature vector related to a combination of the question and the searched answer candidate; and for calculating, when a feature vector is given, a score indicating a degree of how correct an answer candidate is as an answer to the question, for the question and the answer candidate as a source of the feature vector, and for outputting, as the answer to the question, an answer candidate having the highest score.

Claims

1. A non-factoid question-answering system connected to a document storage for storing a plurality of computer-readable documents, the system comprising:

a processor configured to;

receive a non-factoid question as an input;

responsive to the question, retrieve answer candidates to the question from said document storage;

create question and answer candidates comprising combinations of said question and each of the answer candidates;

for each question and answer candidate, generate a set of prescribed features comprising;

calculating a first set of features obtained from respective morpheme information and syntactic information;

in the answer of the question and answer candidate, selecting a causal relation expression based on a matching relationship; and

calculating a second set of features obtained from said selected causal relation expression based on a polarity of the selected causal relation expression, wherein said set of prescribed features includes said first set of features and said second set of features;

for the set of prescribed features and the question and the answer candidate used as a base for generating the set of features, calculate a score representing a degree of plausibility of the answer candidate as a correct answer to the question;

output an answer candidate having the highest score as an answer to the question;

specify, in each of said answer candidates, an expression serving as a clue for specifying a casual relation expression;

specify a causal relation expression consisting of a cause part and a result part of causal relation connected by a specified clue expression, in each of said answer candidates;

determine whether a combination of a noun and a polarity of a predicate on which the noun depends, included in the result part of said specified causal relation expression, matches a combination of a noun and a polarity of a predicate on which the noun depends, included in the question;

determine a relevance of the specified causal relation expression in each of said answer candidates based on word matching and dependency tree matching between each of said answer candidates and said question, and

output, as features of said second feature set, information representing a result of the determination.

2. The non-factoid question-answering system according to claim 1, wherein the processor is further configured to:

responsive to an input of said question, classify the question to any of a prescribed plurality of types,

responsive to an input of said question, divide each of the documents stored in said document storage to sets of sentences each consisting of one or more consecutive sentences and outputting the sets, and

for each set of sentences output, calculate a score in accordance with a standard determined by a result of classification, and outputting a set of sentences having the highest score as said answer candidates.

3. The non-factoid question-answering system according to claim 1, wherein the processor is further configured to:

output, as a set of said second features, an n-gram of a word of the specified causal relation expression, an n-gram of a sub-tree of the dependency tree in the causal relation, or a combination of a polarity of a predicate in said causal relation and meaning of a noun depending on the predicate, or any combination of these.

4. The non-factoid question-answering system according to claim 1, wherein the processor is further configured to:

output, as a set of said second features, the number of words, sub-trees of dependency tree, pairs of a noun and polarities (excitatory/inhibitory) of predicates on which the noun depends on in a dependency tree, shared by said question and a result part of the specified causal relation expression, or any combination of these.

5. A non-factoid question-answering method performed by a processor and executed by using document storage for storing a plurality of computer-readable documents, the method comprising:

receive a non-factoid question as an input

responsive to the question, retrieve using the processor answer candidates to the question from said document storage;

using the processor, create question and answer candidates comprising combinations of said question and each of the answer candidates;

for each question and answer candidate, generate using the processor a set of prescribed features comprising;

calculating a first set of features obtained from respective morpheme information and syntactic information;

in the answer of the question and answer candidate, selecting a causal relation expression based on a matching relationship; and

calculating a second set of features obtained from said selected causal relation expression based on a polarity of the selected causal relation expression, wherein said set of prescribed features includes said first set of features and said second set of features;

for the set of prescribed features and the question and the answer candidate used as a base for generating the set of features, calculate a score representing a degree of plausibility of the answer candidate as a correct answer to the question;

output an answer candidate having the highest score as an answer to the question;

specify, in each of said answer candidates, an expression serving as a clue for specifying a causal relation expression;

specify a causal relation expression consisting of a cause part and a result part of causal relation connected by a specified clue expression, in each of said answer candidates;

determine whether a combination of a noun and a polarity of a predicate, on which the noun depends, included in the result part of said specified causal relation expression, matches a combination of a noun and a polarity of a predicate on which the noun depends, included in the question;

determine a relevance of the specified causal relation expression in each of said answer candidates based on word matching and dependency tree matching between each of said answer candidates and said question, and

output, as features of said second feature set, information representing a result of the determination.