JP2007115118A - Reputation information extraction method, apparatus and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To acquire an attribute expression candidate independently from an attribute expression list. <P>SOLUTION: This method comprises dividing an input sentence to words; retrieving and acquiring, in reference to an evaluation expression list describing evaluation expressions that are evaluations related to properties of an object thing, an evaluation expression from a divided word string; acquiring, for the evaluation expression, attribute expression candidates that are properties of the object thing by paying attention to a sentence structure; calculating, for each attribute expression candidate, an attribute likelihood showing the adequecy as the object thing in reference to word information related to the object thing and attribute expression candidates in a plurality of documents; and taking an attribute expression candidate having an attribute likelihood higher than a predetermined threshold as attribute expression and extracting a set of this attribute expression and an evaluation expression corresponding to the attribute expression as reputation information. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a reputation information extraction method, apparatus, and program, and more particularly, to a reputation information extraction method, apparatus, and program for extracting reputation information related to an object such as a product from a sentence.

  There are many Web pages including sentences in which a reputation about a product, a person, and the like (hereinafter referred to as “target thing”) is written. For example, for a mobile phone, a sentence including a reputation such as “battery is good” or “cute design” is written. If you can extract as a reputation information a set of attribute expressions that express the characteristics of the subject matter, such as “battery possession” and evaluation expressions, such as “good”, from such a sentence containing reputation, Reputation information extracted as reference information is useful. For this reason, research has been conducted to extract reputation information about the subject matter from sentences.

  Previously, we created a list of attribute expressions that represent the properties of the target object such as “color” and “weight”, and a list of evaluation expressions that represent evaluations for attribute expressions such as “good” and “beautiful” ( For example, refer to Non-Patent Document 1), and a method of extracting a pair of attribute expression and evaluation expression as reputation information by pattern matching from a sentence including reputation by referring to the attribute expression list and the evaluation expression list. (For example, refer nonpatent literature 2).

Further, as a method that does not use an attribute expression list, a method of determining whether a sentence includes a reputation from a sentence having an expression related to the target thing has been proposed (for example, see Non-Patent Document 3).
“Collecting Evaluation Expressions for Opinion Extraction” Proceedings of the Society of Language Processing, Vol.12, No.3, pp.203-222, 2005 “Extraction of Opinion Information from Web Document Set and Summary Generation Based on Focus” The 10th Annual Conference of the Association for Natural Language Processing, pp.644-647, 2004 “Extracting Opinion Sentences from Web Bulletin Boards by Automatic Acquisition of Domain Feature Words” The 11th Annual Conference of the Language Processing Society, pp.672-675, 2005

  However, since there are a wide variety of attribute expressions and the number of attribute expressions continues to increase with the appearance of new products, all the attribute expressions are registered in the attribute expression list even using the method described in Non-Patent Document 1 above. It is not possible. Therefore, in the method such as Non-Patent Document 2 described above, reputation information is extracted if an expression representing the nature of the object is included in the sentence including the reputation if the expression is not registered in the attribute expression list. I can't. For example, if there is an attribute expression list in which attribute expressions related to movies are described, and an actor first appearing in a movie named X appears in a new movie, X is not registered in the attribute expression list, so “X is cool” Although the sentence contains the reputation information, the reputation information cannot be extracted.

  Further, in the method of Non-Patent Document 3 for determining whether a sentence includes a reputation without using an attribute expression list, if an expression related to the target thing and an expression that easily represents the reputation are included, the target Even a sentence having an attribute expression that does not represent the nature of a thing is determined to be a sentence that includes a reputation for the subject matter. For example, in a document written about a movie, if there is a sentence “The program in which X appears is also interesting,” this sentence contains the expression “X” and reputation related to the subject matter. Since the expression “interesting” that is easy to represent is included, even a sentence having an attribute expression “program” that does not represent the nature of the movie is determined to be a sentence that includes a reputation about the movie. As a result, this method extracts reputation information that does not describe the object.

  The present invention has been made in view of the above points, and by obtaining attribute expression candidates without depending on the attribute expression list, reputation information that cannot be extracted because the attribute expression is not included in the attribute expression list. In addition to extracting text, candidate attribute expressions corresponding to evaluation expressions are acquired by focusing on the structure of the sentence, and the target objects are obtained by using the target object in a plurality of documents and the word information related to the attribute expressions for the acquired attribute expression candidates. It is an object of the present invention to provide a reputation information extraction method, apparatus, and program capable of extracting reputation information that is correctly stated about a subject matter by examining the appropriateness of the property.

  FIG. 1 is a diagram for explaining the principle of the present invention.

The present invention (Claim 1) is a reputation information extraction method for extracting reputation information about an object including a product from an input sentence,
A word dividing step (step 1) for dividing the inputted sentence into words;
By referring to the evaluation expression list stored in the evaluation expression list storing the evaluation expressions that are evaluations related to the properties of the target object, the evaluation expression is searched from the word string divided in the word dividing step and searched. An evaluation expression acquisition step (step 2) of acquiring the evaluated evaluation expression;
An attribute expression candidate acquisition step (step 3) for acquiring an attribute expression candidate that is a property of the target thing with reference to the sentence structure storage means that stores the sentence structure for the evaluation expression;
Attribute likelihood calculating step for calculating attribute likelihood representing appropriateness as a target thing with reference to word information storage means storing word information related to the target thing and attribute candidate in a plurality of documents for the attribute expression candidate (Step 4),
A reputation information extraction step (step 5) is performed in which an attribute expression candidate having an attribute likelihood higher than a predetermined threshold is defined as an attribute expression, and a set of the attribute expression and an evaluation expression corresponding to the attribute expression is extracted as reputation information. .

The present invention (Claim 2) is the reputation information extraction method of Claim 1,
In the attribute likelihood calculation step, the word appearance frequency is used as the word information.

  FIG. 2 is a principle configuration diagram of the present invention.

The present invention (Claim 3) is a reputation information extraction device that extracts reputation information about an object including a product from an input sentence,
An evaluation expression list storage means 6 for storing an evaluation expression list in which evaluation expressions that are evaluations related to the properties of the object are described;
Sentence structure storage means 7 storing the sentence structure;
Word information storage means 8 for storing word information related to a target thing and attribute expression candidates in a plurality of documents;
Word dividing means 1 for dividing an inputted sentence into words;
Evaluation expression acquisition means 2 for searching for an evaluation expression from the word string divided by the word dividing means 1 with reference to the evaluation expression list stored in the evaluation expression list storage means 6 and acquiring the searched evaluation expression When,
For the evaluation expression, referring to the sentence structure storage means 7, the attribute expression candidate acquisition means 3 for acquiring the attribute expression candidate that is the property of the target thing,
Attribute likelihood calculating means 4 for calculating the attribute likelihood representing the appropriateness as the target object with reference to the word information storage means 8 for the attribute expression candidate,
Reputation information extraction means 5 for extracting attribute expression candidates whose attribute likelihood is higher than a predetermined threshold as attribute expressions and extracting a combination of the attribute expression and an evaluation expression corresponding to the attribute expression as reputation information.

Moreover, this invention (Claim 4) is the reputation information extraction apparatus of Claim 3,
The word information storage means 8 stores the appearance frequency of words as word information,
The attribute likelihood calculating means 4
The attribute likelihood is calculated using the word appearance frequency in the word information storage means.

The present invention (Claim 5) includes an evaluation expression list storage unit that stores an evaluation expression list in which an evaluation expression that is an evaluation related to the property of a target object is described;
A sentence structure storage means for storing the sentence structure;
Word information storage means for storing word information related to a target thing and attribute expression candidates in a plurality of documents;
A reputation information extracting program that causes a computer having the function to function as the reputation information extracting device according to claim 3 or 4.

  As described above, according to the present invention, since the attribute expression candidates are acquired without depending on the attribute expression list, it is possible to extract more reputation information than the conventional technology that refers to the attribute expression list.

  In addition, the attribute expression candidate corresponding to the evaluation expression is acquired by paying attention to the structure of the sentence, and the acquired attribute expression candidate is used as a property of the target object using word information regarding the target object and the attribute expression in a plurality of documents. In order to check the appropriateness, it is possible to correctly extract the reputation information described about the subject matter rather than the conventional technique that does not refer to the attribute expression list.

  In the present invention, since the attribute expression is automatically acquired, it is possible to reduce the cost of creating the attribute expression list necessary in the conventional technique for referring to the attribute expression list.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  FIG. 3 shows a configuration of a reputation information extracting device according to an embodiment of the present invention.

  The reputation information extraction apparatus shown in the figure includes a word division unit 1, an evaluation expression acquisition unit 2, an attribute expression candidate acquisition unit 3, an attribute likelihood calculation unit 4, a reputation information extraction unit 5, an evaluation expression list storage unit 6, and a sentence structure. It consists of a database 7 and a word information database 8.

  The evaluation expression list storage unit 6 stores an evaluation expression list in which evaluation expressions that are evaluations related to the properties of the target object are described.

  The sentence structure database 7 stores a sentence structure.

  The word information database 8 stores word information related to target things and attribute expressions in a plurality of documents.

  The word dividing unit 1 divides the input word, acquires a word string, and passes it to the evaluation expression acquiring unit 2 together with the input target object.

  The evaluation expression acquisition unit 2 refers to the evaluation expression list in which the evaluation expression that is the evaluation related to the property of the target object stored in the evaluation expression list storage unit 6 is referred to, and searches the evaluation expression from the word string in the sentence. The retrieved evaluation expression is acquired and passed to the attribute expression candidate acquisition unit 3.

  The attribute expression acquisition unit 3 acquires attribute expression candidates for the evaluation expression by paying attention to the sentence structure stored in the sentence structure database 7 and passes the attribute expression candidates to the attribute likelihood calculation unit 4.

  The attribute likelihood calculation unit 4 refers to the word information database 8 in which word information related to the target thing and the attribute expression in a plurality of documents is stored with respect to the attribute expression candidate, and represents the attribute likelihood indicating appropriateness as the attribute expression. Is calculated and passed to the reputation information extraction unit 5.

  The reputation information extraction unit 5 extracts attribute expression candidates whose attribute likelihood is higher than the threshold as attribute expressions, and extracts and outputs a combination of attribute expressions and corresponding evaluation expressions as reputation information.

  The operation in the above configuration will be described below.

  FIG. 4 is a flowchart of the overall operation of the reputation information extracting apparatus according to the embodiment of the present invention.

  Step 100) First, the word dividing unit 1 divides an inputted sentence into words to obtain a word string.

  Step 200) Next, the evaluation expression acquisition unit 2 refers to the evaluation expression list in the evaluation expression list storage unit 6 in which the evaluation expression that is the evaluation regarding the property of the target object is described, and the evaluation expression is obtained from the word string in the sentence. Search and obtain the searched evaluation expression.

  Step 300) Next, the attribute expression candidate acquisition unit 3 acquires attribute expression candidates for the evaluation expression obtained in Step 200 by paying attention to the sentence structure stored in the sentence structure database 7.

  Step 400) Next, the attribute likelihood calculation unit 4 refers to the word information database 8 in which word information related to the target thing and the attribute expression in a plurality of documents is stored for the attribute expression candidate, and the appropriate attribute expression is obtained. Attribute likelihood representing the height is calculated.

  Step 500) Finally, the reputation information extraction unit 5 extracts attribute expression candidates whose attribute likelihoods calculated in step 400 are higher than the threshold value as attribute expressions, and extracts pairs of attribute expressions and corresponding evaluation expressions as reputation information. To do.

  It should be noted that the extracted reputation information may be displayed on the display device of the user or stored in the storage means.

  Hereinafter, embodiments of the present invention will be described with reference to the above-described flowcharts using specific examples together with the drawings.

  In the following example, first, a case where reputation information is extracted from the sentence shown in FIG. 5 will be described.

  Further, it is assumed that “Space Wars”, which is the name of a movie, is given as an object. The target thing may be given from the beginning as described above, or may be acquired from the input sentence by referring to the target expression list.

  In step 100, the word division unit 1 divides the sentence of FIG. 5 into words to obtain a word string. Here, it is assumed that word division is performed using an existing morphological analysis technique to obtain notation and part-of-speech information for each word. An example of dividing into words is shown in FIG.

  In step 200, the evaluation expression acquisition unit 2 searches the evaluation expression list from the word string in the sentence by referring to the evaluation expression list in which the evaluation expression that is the evaluation regarding the property of the target thing in the evaluation expression list storage unit 6 is described. , Get the retrieved evaluation expression. An example of the evaluation expression list is shown in FIG. From the word string of FIG. 6 obtained by the word dividing unit 1, it is checked whether the evaluation expression list of FIG. 7 is included. If the evaluation expression is included, the evaluation expression is acquired. Here, it is checked whether or not the evaluation expression is included in the entire word string in FIG. 6, but it may be checked from a part of the word string. Since the fifth evaluation expression “beautiful” in FIG. 7 is included in the word string in FIG. 6, “beautiful” is acquired as the evaluation expression.

  In step 300, the attribute expression candidate acquisition unit 3 acquires attribute expression candidates by paying attention to the sentence structure stored in the sentence structure database 7. An example of the structure of a sentence stored in the sentence structure database 7 is shown in FIG. In FIG. 8, as a sentence structure, any particle of “ha”, “ga”, and “mo” appears after a noun W1 in the sentence, and an evaluation expression W3 immediately after or after one or more adverbs W2 Focusing on whether or not the structure appears, if it has the structure, the first appearing noun W1 is acquired as an attribute expression candidate. The structure of the sentence of interest may be one or a plurality as in this example. The method for acquiring candidate attribute expressions is not limited to this, as long as the focus is on the sentence structure. Dependency analysis is used to analyze the sentence structure, and the nouns included in the phrase related to the evaluation expression are attributed. It may be acquired as an expression candidate. The word string in FIG. 6 has the sentence structure shown in FIG. 8 in which the particle “ga” appears after the noun “CG”, and the evaluation expression “beautiful” appears across one adverb “very”. Therefore, “CG”, which is a noun that appears first, is acquired as an attribute expression candidate.

  In step 400, the attribute likelihood calculation unit 4 refers to the word expression database 8 that stores word information related to the target thing and the attribute expression in a plurality of documents for the attribute expression candidate, and expresses the appropriateness as the attribute expression. Attribute likelihood is calculated. The word information database 8 divides a word into each of a plurality of documents different from the input sentence, and obtains and stores the word appearance frequency in each document. An example of the word information database 8 is shown in FIG. Note that the word information to be used is not limited to the appearance frequency of the word. The position information of the word in the document, the part of speech information of the word, the information of the category to which the word belongs are used. Or information combining them may be used. From the word information database 8, the number A of documents in which the frequency of the target thing and the attribute expression candidate is both 1 or more, the number B of documents in which the frequency of the target thing is 1 or more and the frequency of the attribute expression candidate is 0, and the target expression The number C of documents having a frequency of 0 and the frequency of candidate attribute expressions being 1 or more and the number D of documents having a target expression and candidate attribute expressions having a frequency of 0 are obtained, and the attribute is obtained from these values by the following formula: A likelihood L is calculated.

L = 2A / (2A + B + C) Formula (1)
The attribute likelihood calculation method is not limited to the above formula, and based on the frequency stored in the word information database 8, “H. Shutze, Dimensions of Meaning, Proceedings of Supercomputing 92, pp. 787-796, 1992 "may be created, and the distance between the target expression and the concept vector of the attribute expression candidate may be used as the attribute likelihood.

  Specifically, taking the document with the document number “2” in FIG. 9 as an example, the frequency of the word “Space Wars” in the document is “2”, the frequency of the word “CG” is “1”, and the word “ The frequency of “popcorn”, “today”, and “scene” is “0”. Based on this word information database 8, each value in FIG. 10 is calculated. With respect to the attribute expression candidate, it is checked which of the documents A to D corresponds to, and the number of documents corresponding to each of the documents corresponds to each value in FIG. For example, taking “CG” as an example, the document with document number 1 is D, the document with document number 2 is A, the document with document number 3 is B, the document with document number 4 is C,.

The attribute likelihood is calculated for each attribute expression candidate using FIG. 10 and the above equation (1). When the attribute likelihood L is calculated for the attribute expression candidate “CG”,
L = 2 × 28 / (2 × 28 + 152 + 93) ≈0.19
It becomes.

  In step 500, the reputation information extraction unit 5 extracts attribute expression candidates having attribute likelihoods equal to or greater than a threshold value as attribute expressions, and extracts a combination of the attribute expression and the corresponding evaluation expression as reputation information. Here, the threshold value is 0.15. As described above, the threshold value may be a fixed value for all the input sentences, or may be set to a different value for the input sentence so that 20% of the attribute expression candidates are equal to or greater than the threshold value. The attribute likelihood of the attribute expression candidate “CG” is 0.19, which is a value equal to or greater than the threshold value, so the attribute expression candidate “CG” is an attribute expression. Therefore, a set of reputation information of the attribute expression “CG” and the corresponding evaluation expression “beautiful” is extracted from the sentence of FIG.

  Next, the case where reputation information is extracted from the sentence shown in FIG. 11 will be described.

  It is assumed that “Space Wars” is given as the subject matter as before.

  In step 100, the word division unit 1 divides the sentence in FIG. 11 into words to obtain a word string. The result of dividing into words is shown in FIG.

  In step 200, the evaluation expression acquisition unit 2 searches the evaluation expression list from the word string in the sentence by referring to the evaluation expression list in which the evaluation expression that is the evaluation regarding the property of the target thing in the evaluation expression list storage unit 6 is described. , Get the retrieved evaluation expression. Since the third evaluation expression “delicious” in FIG. 7 is included in the word string in FIG. 12, “delicious” is acquired as the evaluation expression.

  In step 300, the attribute expression candidate acquisition unit 3 acquires attribute expression candidates by paying attention to the sentence structure stored in the sentence structure database 7. The word string in FIG. 11 has the sentence structure shown in FIG. 8 in which the particle “ga” appears after the noun “popcorn”, and the evaluation expression “oi” appears immediately after that. The first pop-up noun “popcorn” is acquired as an attribute expression candidate.

In step 400, the attribute likelihood calculation unit 4 refers to the word expression database 8 that stores word information related to the target thing and the attribute expression in a plurality of documents for the attribute expression candidate, and expresses the appropriateness as the attribute expression. Attribute likelihood is calculated. By calculating the attribute likelihood L for the attribute expression candidate “popcorn” using FIG. 9 and the above equation (1),
L = 2 × 10 / (2 × 10 + 170 + 50) ≈0.08
It becomes.

  In step 500, the reputation information extraction unit 5 extracts attribute expression candidates having attribute likelihoods equal to or greater than a threshold value as attribute expressions, and extracts a combination of the attribute expression and the corresponding evaluation expression as reputation information. Similarly to the previous case, if the threshold value is 0.15, the attribute likelihood of the attribute expression candidate “popcorn” is 0.08, which is a value less than the threshold value, so the attribute expression candidate “popcorn” does not become an attribute expression. Therefore, reputation information is not extracted from the sentence of FIG.

  As described above, since the attribute expression candidates are obtained by focusing on the sentence structure without depending on the attribute expression list, reputation information is not extracted because the attribute expression is not included in the attribute expression list. Disappear. Moreover, the attribute likelihood of the attribute expression candidate appropriate as the property of the target thing such as “CG” is high with respect to the acquired attribute expression candidate using word information related to the target thing and the attribute expression in a plurality of documents. The attribute likelihood is calculated so that the attribute likelihood of the attribute expression candidate that is not appropriate as the property of the target object such as “popcorn” has a low value, and only the attribute expression candidate having a high attribute likelihood is attributed. Since it is expressed, it is possible to correctly extract the reputation information described about the object.

  The series of operations shown in the above embodiments and examples is constructed as a program, installed in a computer used as a reputation information extraction device, and executed by a control means such as a CPU, or via a network. It is also possible to distribute.

  Also, the constructed program is stored in a hard disk device connected to a computer used as a reputation information extracting device, a portable storage medium such as a flexible disk, a CD-ROM, etc., and installed and executed on the computer. Is also possible.

  The present invention is not limited to the above-described embodiments and examples, and various modifications and applications can be made within the scope of the claims.

  The present invention can be used for surveys of customer satisfaction about products.

It is a figure for demonstrating the principle of this invention. It is a principle block diagram of this invention. It is a block diagram of the reputation information extraction apparatus in one embodiment of this invention. It is a flowchart of the whole operation | movement of the reputation information extraction apparatus in one embodiment of this invention. It is an example (the 1) of the input sentence of one Example of this invention. It is an example (the 1) of the word division | segmentation result by the word division part of one Example of this invention. It is an example of the evaluation expression list of one Example of this invention. It is an example of the sentence structure stored in the sentence structure database of one Example of this invention. It is an example of the word information database of one Example of this invention. It is an example of the value used for attribute likelihood calculation acquired from the word information database of one Example of this invention. It is an example (the 2) of the sentence of one Example of this invention. It is an example (the 2) of the word division | segmentation result by the word division part of one Example of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Word division means, Word division part 2 Evaluation expression acquisition means, Evaluation expression acquisition part 3 Attribute expression candidate acquisition means, Attribute expression candidate acquisition part 4 Attribute likelihood calculation extraction means, Attribute likelihood calculation extraction part 5 Reputation information extraction means, Reputation information extraction unit 6 Evaluation expression list storage unit, evaluation expression list storage unit 7 Sentence structure storage unit, sentence structure database 8 Word information storage unit, word information database

Claims (5)

A reputation information extraction method for extracting reputation information about an object including a product from an input sentence,
A word dividing step of dividing the input sentence into words;
With reference to the evaluation expression list described in the evaluation expression, which is an evaluation related to the properties of the target object, stored in the evaluation expression list storage means, the evaluation expression is searched from the word string divided in the word dividing step, An evaluation expression acquisition step of acquiring the searched evaluation expression;
An attribute expression candidate acquisition step of acquiring an attribute expression candidate that is a property of the target thing with reference to a sentence structure storage unit that stores a sentence structure for the evaluation expression;
Attribute likelihood calculation for calculating attribute likelihood representing appropriateness as a target thing with reference to word information storage means storing word information related to the target thing and attribute expression candidates in a plurality of documents for the attribute expression candidate Steps,
Reputation information extraction step for extracting a set of attribute expression and an evaluation expression corresponding to the attribute expression as reputation information as an attribute expression candidate whose attribute likelihood is higher than a predetermined threshold;
Reputation information extraction method characterized by performing. In the attribute likelihood calculating step,
The reputation information extraction method according to claim 1, wherein a word appearance frequency is used as the word information. A reputation information extraction device that extracts reputation information about an object including a product from an input sentence,
An evaluation expression list storage means for storing an evaluation expression list in which evaluation expressions that are evaluations related to the properties of the object are described;
A sentence structure storage means for storing the sentence structure;
Word information storage means for storing word information related to a target thing and attribute expression candidates in a plurality of documents;
Word dividing means for dividing the inputted sentence into words;
Evaluation expression acquisition means for retrieving an evaluation expression from the word string divided by the word dividing means by referring to the evaluation expression list stored in the evaluation expression list storage means and acquiring the searched evaluation expression When,
With respect to the evaluation expression, referring to the sentence structure storage means, attribute expression candidate acquisition means for acquiring an attribute expression candidate that is a property of the target thing,
For the attribute expression candidate, referring to the word information storage means, attribute likelihood calculating means for calculating an attribute likelihood representing appropriateness as a target object,
Reputation information extraction means for extracting a set of attribute expressions and evaluation expressions corresponding to the attribute expressions as reputation information, with the attribute expression candidates having an attribute likelihood higher than a predetermined threshold as attribute expressions;
A reputation information extraction device characterized by comprising: The word information storage means stores the appearance frequency of words as the word information,
The attribute likelihood calculating means includes:
Calculating the attribute likelihood using the appearance frequency of the word in the word information storage means;
The reputation information extraction device according to claim 3. An evaluation expression list storage means for storing an evaluation expression list in which evaluation expressions that are evaluations related to the properties of the object are described;
A sentence structure storage means for storing the sentence structure;
Word information storage means for storing word information related to a target thing and attribute expression candidates in a plurality of documents;
A computer having
5. A reputation information extraction program that functions as the reputation information extraction device according to claim 3.

Images