JP2000148770A - Device and method for classifying question documents and record medium where program wherein same method is described is recorded - Google Patents

Abstract

PROBLEM TO BE SOLVED: To retrieve a category matching the question contents of a question document with high precision by using a feature quantity extracted from constituent elements of the question document corresponding to an answer document included in the category when the feature quantity of the extracted category is calculated. SOLUTION: Individual documents are extracted from an answer document set and documents are put together form a question document set (S1, S2). Those individual documents are classified (S3) after they are decomposed into words and morpheme analysis is carried out to calculate feature vectors. After the classified answer documents are held (S4), feature vectors are calculated as to the question documents and classification category feature quantities are calculated (S6) while made to correspond to the answer documents. Then a new question document is extracted (S7), a feature vector is calculated as to the new document to calculate the adaptivity to the held classification category (S8), and a specific number of matching categories from the top are gathered (S9).

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention provides a method of answering a large number of inquiries, extracting combinations of similar contents from past inquiries and answer histories, and supporting or automating the answering operation. The present invention relates to an apparatus and method for classifying an inquiry document and a recording medium on which a program describing the method is recorded.

[0002]

2. Description of the Related Art With the development of natural language processing technology and the improvement of computer processing capability, it is possible to extract documents having similar contents from a large set of stored documents and classify them into a plurality of categories. Has become. A technique for selecting a category having the highest matching degree from existing categories for a newly given document is also known.

The following method is known as a method for classifying a document set. First, a document to be classified is decomposed into elements in units of character strings, words, and phrases, and a feature amount is calculated based on a combination of the elements. Next, the similarities of the feature amounts are obtained for all the combinations of documents, and clusters are formed in order from the combination having the highest similarity. This process is repeated until a certain fitness or a cluster size is reached, and the cluster is used as a classification result.

Various methods have been devised as a method of calculating the feature amount. For example, as described above, after decomposing a document into elements in units of character strings, words, and phrases, the weights of the elements are obtained based on the frequency of occurrence of each element in a document set and the frequency of occurrence in the document. There is known a method of expressing a feature amount by a vector formed by each element and its weight.

[0005] Alternatively, the degree of relevance between elements included in the document is calculated by a predetermined method, and the elements are appropriately arranged in an n-dimensional vector space so that the element having a high degree of relevance is located nearby.
There is also known a method of calculating a feature vector of a document based on a sum of vectors formed by elements included in the document.

[0006] As a method of calculating the similarity between feature quantities, when a feature quantity is expressed as a vector, a method of calculating a fitness based on an inner product or cosine of two vectors is widely used. .

[0007]

In each of the conventional techniques, a document to be classified is used as an information source, and classification is performed based on elements constituting the document. However, in an inquiry document received from an unspecified number of people,
Since the vocabulary and expressions used vary from person to person, the components of the document may differ even if the contents of the inquiry are the same.

For this reason, it is difficult to classify a set of inquiry documents with high accuracy according to the contents of the inquiry.

SUMMARY OF THE INVENTION The present invention has been made in view of the above-described problems in the prior art, and has as its object to perform classification according to the contents of an inquiry.

[0010]

Means for Solving the Problems In the prior art, in each case, a document itself to be classified is used as an information source, and classification is performed based on elements constituting the document. The classification is performed based on the elements constituting the answer document corresponding to the inquiry document, and the elements constituting the inquiry document are used in calculating the feature amount of each category obtained as a result of the classification.

As described above, since the vocabulary and expressions used in the inquiry document are various, the feature amounts extracted therefrom may not match the contents of the inquiry.

On the other hand, in the task of answering an inquiry received from an unspecified number of people, the number of people who create answer documents is smaller than the number of people who create inquiry documents, and the use of terms The vocabulary and expressions that appear in the answer document are more uniform than the vocabulary and expressions that appear in the query document because they are often unified among the respondents. Answer documents using vocabulary and expressions tend to be created. Therefore, even if the vocabulary included in the inquiry document is different, if the contents of the inquiry are the same, it can be expected that the feature amount extracted from the corresponding answer document shows a high similarity.

Further, since the vocabulary and the expression tend to be different between the answer document and the query document, it is not possible to search for a category that matches the new query document by using the features extracted from the response document. However, in the present invention, when calculating the feature amount of the extracted category, since the feature amount extracted from the component of the query document corresponding to the answer document included in the category is used, a new query document is used. With the input as an input, it is possible to search for a cacodeli conforming to the contents of the inquiry document with high accuracy.

[0014]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A description will now be given of an implementation example of a query document classification apparatus according to the present invention.

FIG. 1 shows a system configuration of this embodiment.

In the figure, reference numeral 1 denotes a set of answer documents, a set of past answer documents, 2 denotes a set of inquiry documents, a set of past query documents, and 3 denotes a set of answer document classifying means and 1 , A feature vector is calculated by performing a morphological analysis on the contents of the answer document, and the contents of the answer document set 1 are classified, 4 is an answer document classification category, 5 is a category feature amount, 6 is a category feature amount calculating means, 7 Indicates a new inquiry document, 8 indicates a matching category selection unit, and 9 indicates a matching category selection result. Hereinafter, each component shown in FIG. 1 will be described in order.

The answer document classifying means 3 is realized by the flow shown in FIG.

In FIG. 2, reference numerals 1, 3, and 4 correspond to those in FIG. 1, 31 is a morphological analysis unit, 32 is a response document feature vector calculation unit, 33 is a classification target list, and 34 is a similar document extraction unit. I have.

As shown in FIG. 2, the individual contents of the answer document set 1 are subjected to morphological analysis in the morphological analysis unit 31 to decompose each sentence of the answer document into words. A feature vector FV is calculated based on each word. The result is stored as a classification target list 33. Next, the similar document extracting unit 34 extracts two elements from the classification target list 33, calculates the similarity, extracts a combination having a large similarity, and obtains a cluster. Get category 4. This will be described more specifically.

The morphological analysis unit 31 decomposes each sentence in the answer document into words using a known technique. The answer document feature vector calculation unit 32 calculates a feature vector FV based on each word included in the answer document.

FV (i) = (w (i, 1), ..., w (i, j), ..., w (i, N)) w (i, j) = tf (i, j ) ^* log (M / df (j)) where i is the document number, N is the total number of words in the answer document set, tf (i, j) is the number of occurrences of word j in document i, M
Is the total number of answer documents, and df (j) is the number of occurrences of word j in the document set.

The feature vector FV (i) of the document i is expressed as a vector on a space having dimensions equal to the number of words constituting the document set. The elements of each dimension indicate how strong the document is with the elements of that dimension. The term tf (i, j) is the number of occurrences of word j in document i, which is an important word that characterizes a document if the word appears repeatedly in the document. It means to be positioned.

The term log (M / df (j)) is a number obtained by taking the logarithm of the reciprocal of the frequency of occurrence of word j. This term becomes more important as the frequency of occurrence of the word throughout the entire document decreases. It means to interpret it as a word. However, if the reciprocal of the appearance frequency is regarded as the importance as it is, a word whose appearance frequency is 1/100 is 100 times more important, which is unnatural, and is therefore logarithmic. This calculation method is described in G. It is based on the tf ^* idf method developed by Salton, and is widely used in text search.

In this embodiment, the feature quantity is expressed by a vector having elements as words in general. However, the components of the vector can be selected by a method using only independent words as an element, an affix or a compound word. And a method of using a word string included in a noun phrase as an element. Various other calculation methods of the feature amount are considered. Any method can be used as long as it is a quantity calculated based on the components of the document and the similarity between the quantities can be calculated. The present invention can also be applied to

The similar document extracting unit 34 sorts the classification target list 3
Two elements are extracted from 3 and the similarity is calculated for all combinations.

The similarity R (i, k) between the element i and the element k is obtained by the following equation.

R (i, k) = FV (i) · FV (k) / (| FV (i) | · | FV (k) |) FV (i) and FV (k) are the features of the document, respectively. FV (i) · FV (k) is the vector inner product, | FV (i)
| · | FV (k) | means the product of the magnitudes of the respective vectors. That is, R (i, k) represents the cosine of the two vectors. This is based on the idea that the smaller the angle between two vectors, the higher the similarity.

In the initial state, all answer documents are set as a classification target list. A combination having the highest similarity is extracted from the classification target list, and a cluster C composed of elements of the combination is generated. When the element of the cluster C is i, k, the feature vector of one cluster C is calculated by the following equation.

FV (C) = (u (i) ^* FV (i) + u (k) ^* FV (k)) /
(u (i) + u (k)) Here, u (i) is 1 when i is a document, and is the number of documents included in i when i is a cluster.

This finds the center of gravity of the feature vector. Since a cluster is a group including a plurality of documents, the weight of the cluster is weighted by the number of documents included in the cluster to determine the center of gravity of the vector.

Next, i and k are removed from the classification target list 33, a cluster C is added, and a combination having the highest similarity is again extracted from the classification target list to generate a cluster.

The above process is repeated until the maximum value of the similarity falls below a predetermined threshold. When the repetition is completed, clusters that are not elements of other clusters are enumerated, and a set of documents included in each cluster is set as a similar document set.

The classification category feature quantity calculating means 6 is realized by the flow shown in FIG.

In FIG. 3, reference numerals 2, 4, and 6 correspond to FIG. 1, 61 is an inquiry document / answer document correspondence management unit, 62 is a classification category feature vector calculation unit, 63 is a classification category feature vector, and 64 is a morpheme. It represents an analysis unit.

The inquiry document / answer document correspondence management unit 61
For all answer documents included in each classification category, an inquiry document corresponding to the answer document is acquired from answer document classification category 4. Each sentence of the inquiry document is managed in a state where it is decomposed into words by the inquiry document set 2 and the morphological analysis unit 64. Classification category feature vector calculation unit 62
Finds the following feature vector FV and uses it as the classification category feature vector 63.

[0036]

(Equation 1)

Where i is the number of the category, N is the total number of words in the query document set, and C (i) is the category
The number of answer documents contained in i, tf (c, j) is the word in the query document corresponding to answer document c contained in the classification category.
j is the number of appearances, M is the total number of query documents, and df (j) is the number of occurrences of word j in the query document set.

The classification category is obtained as a group of a plurality of similar documents. In order to represent a feature vector of this classification category, information on a word included in a document included therein is used. Basically, the concept is the same as finding the feature vector of an individual document, but since it includes multiple documents, the feature vector of the entire category is calculated by adding up the number of occurrences of the words contained in each document. I'm asking.

The matching category selecting means 8 is realized by the flow shown in FIG.

In FIG. 4, reference numerals 4, 7, and 8 correspond to FIG. 1, 81 is a morphological analysis unit, 82 is a new query document feature amount calculation unit, 83 is a new query document feature vector, and 84 is a fitness degree calculation unit. , 85 indicates a matching category list, and 86 indicates a sort processing unit.

The morphological analysis unit 81 extracts each sentence of the newly given inquiry document from the new inquiry document 7 and breaks it down into words.

The new query document feature vector calculation unit 82 calculates the following query document feature vector QV as a new query document feature vector 83 based on the word information extracted by the morphological analysis unit 81.

QV = (tf (1),..., Tf (j),..., Tf (N)) where tf (j) is the number of appearances of word j in the new query document, and N is This is the total number of words in the query document set.

The relevance calculator 84 converts the feature vector obtained by the new query document feature calculator 82 and all the classification categories obtained by the classification category feature calculator 6 from the answer document classification category 4. The extracted category is calculated by the following formula to calculate the degree of conformity with the feature vector of the classification category, and a conforming category list 85 is obtained.

Score (i) = QV · FV (i) / (| QV | · | FV (i) |) where QV is the feature vector of the new query document, FV (i)
Is a feature vector of the classification category i.

It is intended to find a classification category in which many words overlap the new inquiry document. Therefore, QV
Is once decomposed into a vector of words and then score
Is calculated. The score is calculated as the cosine between the feature vector of each category and the feature vector of the new query document.

The sort processing unit 86 includes
4. Sorting is performed in ascending order of the degree of fitness calculated in step 4, and a predetermined number of classification categories are extracted from the top and output as search results.

In the above description, the inquiry document classifying apparatus has been described. However, it is possible to consider a query document classifying method corresponding to the inquiry document classifying apparatus, and to execute a program describing the query document classifying method. The program can be recorded on a recording medium in advance.

FIG. 5 shows a processing flow representing a method of classifying an inquiry document according to the present invention. Steps (S1) to (S4) and step (S
Steps 6) to (S9) correspond to the processes of the units indicated by reference numerals 1 to 4 and 6 to 9 shown in FIG. Step (S1): Individual documents are extracted from answer document set 1.

Step (S2): A document is extracted from the inquiry document set 2.

Step (S3): Each document is decomposed into words, morphological analysis is performed, feature vectors are calculated, and classification is performed.

Step (S4): The classified answer document is held.

Step (S6): The feature vector is calculated for the inquiry document, and the classification category feature amount is calculated in association with the answer document. Then, the process returns to step (S4) and holds.

Step (S7): A new inquiry document is extracted.

Step (S8): A feature vector is calculated for the new inquiry document, and the degree of matching with the classification category held in step (S6) is calculated.

Step (S9): A predetermined number of matching categories are collected from the top.

The inquiry classification method shown in FIG. 5 can be described in the form of a program, and the program can be recorded on a recording medium. Therefore, the present invention includes the recorded recording medium in the technical scope.

[0058]

As described above, according to the present invention, the accumulated inquiry / answer document information is classified into a plurality of categories based on the contents thereof, and the contents of the inquiry / response document information are added to the newly received inquiry. Can be selected with high accuracy.

For example, at the inquiry window of a company,
When we receive a query similar to the one we answered earlier,
It is possible to extract the case that was answered last time with high accuracy. However, a company's inquiry counter often answers a very large number of inquiries, and there may be a very large number of similar cases. At that time, if similar cases are presented from one end, it would be a burden on the worker. Therefore, similar cases are once grouped into categories, and this category is presented to the worker as a search result. As a result, the operator can select correct information from a smaller number of candidates, thereby obtaining an effect of reducing the work load.

[Brief description of the drawings]

FIG. 1 is a system block diagram of an inquiry document classification device of the present invention.

FIG. 2 is a processing flow of an answer document classifying means of the inquiry document classifying apparatus of the present invention.

FIG. 3 is a processing flow of a classification category feature amount calculation unit of the inquiry document classification device of the present invention.

FIG. 4 is a processing flow of a matching category selection unit of the inquiry document classification device of the present invention.

FIG. 5 is a processing flow of an inquiry document classification method according to the present invention.

[Explanation of symbols]

 1 Answer Document Set 2 Inquiry Document Set 3 Answer Document Classification Means 4 Answer Document Classification Category 5 Classification Category Feature 6 Classification Category Feature Calculation Means 7 New Inquiry Document 8 Matching Category Selection Means 9 Matching Category Selection Result

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Masayuki Sugizaki 3-19-2 Nishi-Shinjuku, Shinjuku-ku, Tokyo Within Japan Telegraph and Telephone Corporation (72) Inventor Kazuo Tanaka 3-192-1, Nishi-Shinjuku, Shinjuku-ku, Tokyo No. Nippon Telegraph and Telephone Corporation F-term (reference) 5B075 ND03 ND36 NK06 NK32 UU06

Claims (3)

[Claims] An apparatus for managing a plurality of inquiry documents and answer documents for each inquiry in association with each other, wherein a set of answer documents is classified into a plurality of categories according to a feature amount extracted from each answer document. Answer document classifying means; and classification category feature quantity calculating means for calculating the feature quantity of the category using the feature quantity of the query document corresponding to each answer document included in each category classified by the answer document classifying means. And for a newly given query document, a feature amount obtained by the classification category feature amount calculation unit from the categories classified by the answer document classification unit matches the feature amount of the new inquiry document. And a matching category selecting means for extracting an object to be searched. 2. A method for managing a plurality of inquiry documents and answer documents corresponding to each inquiry in association with each other, wherein a set of answer documents is classified into a plurality of categories according to a feature amount extracted from each answer document. An answer document classifying step; and using a feature amount of an inquiry document corresponding to each answer document included in each category classified by the answer document classifying step, calculating a feature amount of the category, a classification category feature amount calculating step. And for a newly given query document, a feature obtained in the classification category feature calculation step from among the categories classified in the answer document classification step matches the feature quantity of the new query document. A matching category selecting step of extracting a matching document. 3. A recording medium in which a method of managing a plurality of inquiry documents and response documents to each inquiry in association with each other is described in the form of a program, and the management method includes the step of managing a set of response documents. An answer document classifying step of classifying the answer document into a plurality of categories according to the feature amount extracted from each answer document; and characteristics of the query document corresponding to each answer document included in each category classified by the answer document classifying step. Calculating a feature quantity of the category using the quantity, a category category feature quantity calculating step, and for a newly given query document, the classification category is selected from the categories classified by the answer document classifying step. A matching category selecting step of extracting a matching feature amount obtained in the feature amount calculating step with the feature amount of the new query document. Law and described in the form of a program, a recording medium characterized by recording the program.