JP2010271800A - Device, method, and program for classifying reply document - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a reply document classification device which can classify reply documents in a short time, by also using a keyword which is not included in retrieval keywords. <P>SOLUTION: A word and a part of speech are combined, a conditional pattern as a pattern showing prescribed conditions is stored, and when a sentence included in an input text group includes the stored conditional pattern, a conditional keyword as the array of words arranged in front of the conditional pattern is extracted, and when the conditional pattern is not included, a nominative case keyword as a word included in a clause which becomes a nominative case is extracted, and appearance frequency, in which a keyword under consideration appears as the conditional keyword, and appearance frequency, in which the keyword appears as the nominative case keyword, are totaled for each keyword included in the conditional keyword and the nominative case keyword, and when both of appearance frequency satisfy prescribed conditions, the keyword under consideration is a classification keyword. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to an apparatus for classifying answers to a question, and more particularly to an answer document classification apparatus that classifies and provides a plurality of answers to a question for each keyword.

  Conventionally, keyword search is often used as a means for searching for information published on a network such as the Internet. The keyword search is a service in which one or a plurality of search keywords are input from a user, and a document including all or any of the input search keywords is searched and output. The number of documents output as search results increases as the search keyword is a general word, and it is difficult for the user to find information suitable for the purpose. For this purpose, the user performs a task of narrowing down search results by adding a search keyword or specifying a search keyword that is characteristic of the field to be searched.

  Also, there are cases where users use Q & A community sites where users exchange questions and answers when they cannot find information that suits their purpose well by keyword search, or when they want to consult people. In the Q & A community site, as shown in FIG. 5, responses 33, 34, and 35 are posted by a plurality of respondents to one question 32 posted by a user who wants to ask questions, and communication is performed on the site. . The Q & A community site also provides a keyword search function, and the search target can be specified by specifying only a question or both a question and an answer. Also in this case, if the search keyword is a general word, the number of search results increases, and it is not easy to search for information suitable for the purpose.

  In order to make it easier to search for information that suits the purpose, there is a method of presenting search results in groups. For example, for a search keyword input by a user, a keyword that is often input adjacent to the search keyword input by the user is extracted as a related word from a history of search keywords input in the past. There has been known an invention that displays search results by grouping for each related word (for example, see Patent Document 1). According to the present invention, there is an advantage that the user does not need to think about a keyword for narrowing down the search result by himself and only has to select it.

Further, a clustering technique can be used as a method for classifying answer documents. The clustering technique is a method of calculating a similarity between sentences when a sentence set is given and creating a group for each similar sentence. As the similarity between sentences, a technique is widely used in which sentences are represented by sentence vectors based on the word appearance frequency, and the cosine similarity between sentence vectors is applied. In other words, the sentence vector sentence _{d n}

If represented by, v is word set _{W = {w 1, w 2} , ..., w v} indicates the total number of words _{in, x ni} indicates the weight of a word _{w i} in sentence _{d n.} At this time, the similarity between the sentence d _j and the sentence d _k is the angle formed by each sentence vector.

It is represented by The weight of w _i is the tf obtained by multiplying the appearance frequency tf (term frequency) in the sentence of the word as it is or by multiplying the appearance frequency tf by the logarithm idf obtained by dividing the word appearance number by the total sentence number. -Idf (term frequency / inverse document) is used. That is, sentences with high similarity mean sentences having similar word weight trends.

Japanese Patent No. 4009937

  Originally, the purpose of users of Q & A community sites is to obtain answers to questions, so it is desirable to classify and provide answer documents. When the prior art described in Patent Document 1 is applied to a Q & A community site, the user who searches is often not familiar with the field of information he / she wants to know, and an appropriate keyword is not always input. For this reason, there is a problem that the answer document cannot be classified with high accuracy only by using the search keyword as a keyword for classification.

  Also, when answer documents are classified by clustering technology, the processing time is proportional to the square of the number of sentences. Therefore, if there are many sentence sets of keyword search results, it takes time to output the classification results. is there.

  An object of the present invention is to provide an answer document classification device, an answer document classification method, and a program that can classify answer documents in a short time using a keyword that is not included in a search keyword.

  The present invention relates to a pattern composed of combinations of words and parts of speech in an answer document classification apparatus for classifying an answer document to a question document for each group including a classification keyword that is a keyword representing the group, and a predetermined pattern. A condition pattern accumulating means for accumulating a condition pattern which is a pattern indicating the condition of the condition, and whether or not a sentence included in the input text set includes the condition pattern accumulated in the condition pattern accumulating means. Condition pattern determining means for determining, condition keyword extracting means for extracting a condition keyword that is an array of words arranged in front of the condition pattern if the condition pattern is included, and the condition pattern Otherwise, a prominent keyword that extracts a prominent keyword that is a word contained in the phrase that becomes the prominent For each keyword included in the condition keyword and the main keyword, the number of appearances that is the number of times that the keyword of interest appears as the condition keyword, and the number of appearances that is the number of times that the keyword appears as the main keyword An answer document having classification keyword determination means for determining that the keyword of interest is a classification keyword when the number of appearances of the conditional keyword and the number of appearances of the main keyword satisfy a predetermined condition It is a classification device.

  According to the present invention, a user who wants to search for an answer document about a question can classify the answer document using a keyword that is not input as a search keyword, and the questioner simply selects the classified information. Thus, there is an effect that desired information can be acquired.

  In addition, according to the present invention, since the classification keyword is determined based on the number of appearances of the keyword included in the answer document set, there is an effect that the processing is fast.

It is a figure explaining the principle of this invention. It is a figure which shows the structure of the reply document classification | category apparatus 100 which is Example 1 of this invention. It is a figure which shows the example of the condition pattern which the condition pattern storage means 11 has accumulate | stored. 3 is a flowchart showing processing of a condition pattern determination unit 12. It is a figure which shows the structure of a Q & A community site. It is a figure which shows the morphological analysis result which the condition pattern determination means 12 determined. In Example 1, it is a figure which shows the example of the result of the dependency analysis which the conditional keyword extraction means 13 extracted. It is a figure which shows operation | movement of the main character keyword extraction means. FIG. 10 is an explanatory diagram of the operation of the classification keyword determination unit 15. 4 is a flowchart showing the operation of a classification keyword determination unit 19. It is a block diagram which shows the answer document classification | category apparatus 200 which is Example 2 of this invention. It is a figure which shows the example of accumulation | storage of the classification keyword storage means 18 in Example 2. FIG.

  The modes for carrying out the invention are the following examples.

  In the present invention, a classification keyword that is a keyword representing a group and is a keyword for classifying an answer document is extracted from the answer document. Since respondents who are familiar with the questioned content may provide detailed information using words or phrases that are not included in the question document, the classification keywords are extracted from the answer document. Answer documents for question documents are grouped for each classification keyword and provided to the user. Thus, the user can easily obtain desired information using the classification keyword as a clue.

  The present invention relates to a classification that is a keyword representing a group from keywords (words) included in a given set of answer documents in an answer document classification apparatus that groups answer documents for a question document for each classification keyword. Determine keywords.

  First, the characteristics of the answer document will be described. For example, consider the case where there is the following question document.

“I am planning to go to theme park X in June. I'm planning to prepare an umbrella or a pair in case of rain, which is better?”
Assume that three respondents answered the following questions.

Respondent A: “Umbrellas are annoying if there are many people.
Respondent B: “If you are an adult, you can use an umbrella.
Respondent C: “I think adults can use umbrellas.”
Respondent A states his opinion within the scope of the contents described in the question document. Respondent B adds information about whether he is “adult only” or “children are together” and is not included in the question document, and states his / her opinion in each case.

  The questioner does not know at the time of the question that the recommended information varies depending on whether "adults only" or "children are together", but the classification keywords in the answer document are "adults only" and "children together." If the keyword "is presented, the classification keyword can be selected in accordance with the user's situation, and an accurate answer document can be referred to.

  Respondent C stated “Adults are” and does not use the condition pattern “If ○○”. Note that the “condition pattern” is a clause that clearly indicates a condition.

  In the answer document by the respondent C, the phrase “adult is” is implicitly described including the condition “in the case of an adult” implicitly. In the answer document, “XX is”, which is a case, is often used to explain XX. Therefore, it is considered that keywords suitable for classifying response documents are also included in XX of the phrase “XX is”.

  FIG. 1 is a flowchart illustrating the principle of the present invention.

  First, it is determined whether or not the condition pattern is described for each sentence included in the input text set (S1). The “condition pattern” is, for example, “if” in “if it is XX” in the sentence “if it is XX, it is XX”.

  In addition, the following pattern is also considered as a condition pattern.

“If” in “If ○○”,
“In the case of ○○”, “In the case of”
“If it ’s ○○”, “It ’s”.

  If it is determined that the condition pattern is described in the sentence, the keyword included in the word “XX” in the phrase indicating the condition “if XX” is extracted as the “condition keyword” (S2). ). That is, the “condition keyword” is a keyword included in a word excluding the condition pattern from the clause including the condition pattern. Also, if it is determined that the condition pattern is not described, it is checked whether there is a clause that becomes a major, such as “XX”, and if there is a clause that becomes a major, the phrase that becomes the major “ The keywords included in “OO” are extracted as “main keywords” (S3).

  That is, the “prominent keyword” is a noun phrase included in the phrase that becomes the prominent.

  The condition keyword and the main keyword are composed of at least one word.

  Next, among the keywords extracted from the entire set of input texts, the number of appearances, which is the number of times that the focused keyword appears as the conditional keyword, and the number of appearances, which is the number of appearances as the main keyword If the number of appearances of the conditional keyword and the number of appearances of the main keyword satisfy a predetermined condition, it is determined that the keyword of interest is a classification keyword (S4).

  That is, the “classification keyword” is a keyword when the number of appearances of the condition keyword and the number of appearances of the main keyword satisfy a predetermined condition.

  FIG. 2 is a diagram showing the configuration of the answer document classification apparatus 100 that is Embodiment 1 of the present invention.

  The answer document classification device 100 is connected to the text input means 20 and the text output means 40, and the condition pattern storage means 11, the condition pattern determination means 12, the condition keyword extraction means 13, the main character keyword extraction means 14, Classification keyword determination means 15.

  When a set of answer documents is input from the text input means 20, the condition pattern determination means 12 divides the input answer document into sentence units, determines whether each sentence includes a condition pattern, and the condition pattern Is sent to the condition keyword extraction means 13, and a sentence not containing the condition pattern is sent to the main keyword extraction means 14.

  The condition keyword extraction unit 13 receives a sentence of the morphological analysis result including the condition pattern and information indicating the detected condition pattern from the condition pattern determination unit 12, and selects a keyword that exists in front of a location that matches the condition pattern. These are means for extracting as condition keywords and storing them in the storage device.

  The main keyword extracting unit 14 receives a sentence determined not to include the condition pattern from the condition pattern determining unit 12, performs dependency analysis, and stores the sentence in the storage device. In this analysis result, it is checked whether there is a main phrase, and if there is a main phrase, the keywords (noun phrases) included in the main phrase are extracted as “main keywords”. And store it in the storage device.

  The classification keyword determination means 15 inputs the condition keyword extracted by the condition keyword extraction means 13 and the main keyword extracted by the main keyword extraction means 14, and for each keyword, a predetermined condition (formula (1), formula described later) It is determined whether or not the condition (2) is satisfied, a keyword satisfying the preset condition is stored in the storage device as a classification keyword, and the classification keyword is output. Specifically, the classification keyword determination unit 15 sends the classification keyword, the answer sentence ID including the keyword, and the sentence number information to the text output unit 40 in combination.

  A question / answer document storage unit 30 is connected between the text input unit 20 and the text output unit 40. The question / answer document storage means 30 stores a large number of question documents and answer documents posted on the Q & A community site.

  The text input means 20 acquires a set of answer documents that meet a specific condition, which will be described later, from the question / answer document storage means 30 and sends it to the condition pattern determination means 12.

  A condition for acquiring a set of answer documents that meet a specific condition described later is a set of answer documents that answered a question including a specific keyword.

  The text output means 40 inputs the classification keyword output by the classification keyword determination means 15, classifies the answer document for each classification keyword, stores it in the storage device, and outputs it.

  FIG. 3 is a diagram showing an example of condition patterns stored in the condition pattern storage unit 11.

  The condition pattern accumulating unit 11 accumulates in advance the word to be extracted as the condition pattern and the appearance order pattern of this word as shown in FIG. In the accumulation example shown in FIG. 3, one condition pattern is described in one line. One condition pattern is a pattern in which “part of speech: reading” is connected by a comma, and words appear successively from the left. The word pattern on the fourth line, “determinant: de, verb stem: a, verb inflection ending: les, verb suffix: ba” is a description for extracting the conditional pattern “if”.

  When a set of answer documents is input from the text input means 20, the condition pattern determination means 12 divides the input answer document into sentence units, and determines whether each sentence includes a condition pattern. Then, the sentence including the condition pattern is sent to the condition keyword extracting means 13, and the sentence not including the condition pattern is sent to the main keyword extracting means 14. Each answer document is given an answer sentence ID that uniquely identifies the answer document.

  FIG. 4 is a flowchart showing the operation of the condition pattern determination unit 12.

  First, all the condition patterns stored in the condition pattern storage unit 11 are read (S51), and for the response document received from the text input unit 20, for each response document, the response document corresponding to the response text ID is divided into sentence units. Then, an appearance number in the answer document is given to the divided sentence (S52, S53).

  Subsequently, if there is an unprocessed sentence number (S54), one unprocessed sentence number is selected and a morpheme analysis process is performed (S55). It is checked whether or not there is a place where the part of speech and the reading of the word of the morphological analysis result matches the condition pattern (S56), and if there is a place matching the condition pattern (S56, YES), the answer sentence ID to be processed and The sentence number and the part where the sentence and the condition pattern match are sent to the condition keyword extracting means 13 (S57). If there is no portion that matches the condition pattern (S56, NO), the answer text ID to be processed, the sentence number, and the sentence are sent to the main keyword extracting means 14 (S58).

  FIG. 5 is a diagram illustrating a configuration example of the Q & A community site.

  A specific operation when the condition pattern determination unit 12 inputs the answer documents 33, 34, and 35 shown in FIG. 5 will be described. The sentence IDs of the answer documents 33, 34, and 35 are set to 33, 34, and 35, respectively. First, an unprocessed answer sentence ID 33 is selected (S52, S53). Since the selected answer sentence ID 33 is composed of one sentence (S53), one sentence is selected and a morphological analysis process is performed. (S54, S55).

  FIG. 6 is a diagram illustrating a morphological analysis result determined by the condition pattern determination unit 12.

  In the morpheme analysis result 61, the answer sentence ID is described in the first line, the sentence number is described in the second line, and the notation, part of speech, and reading for each morpheme are described in the third and subsequent lines. It is checked whether or not the morphological analysis result 61 includes a word appearance pattern (a pattern in which a word appears) that matches the condition pattern (S56). Since the morpheme analysis result 61 shown in FIG. 6 does not have a location that matches the condition pattern (S56, NO), the morpheme analysis result 61 is sent to the main keyword extraction means 14.

  Since there is no other sentence in the reply sentence ID 33 (S54, NO), the next unprocessed reply document 34 is selected (S52). The sentence with the answer sentence ID 34 is divided into two sentences (S53), the first sentence is selected, and morphological analysis processing is performed (S55). When it is checked whether or not the morphological analysis result 62 includes a word appearance pattern that matches the condition pattern (S56), a location 63 that matches the condition pattern “determinant: Nara” is found.

  Since there is a location 63 that matches the condition pattern (S56, YES), “*” is added to indicate the location that matches the condition pattern in the morphological analysis result 62, and the result is sent to the condition keyword extraction unit 13 ( S57). Subsequently, the processing of S54 to S56 is executed for the second sentence (morpheme analysis result 64) of the sentence with the answer sentence ID 34, and a portion 65 that matches the condition pattern is found.

  Next, the process of S52-S56 is performed about the sentence of reply sentence ID35, and since a condition pattern does not exist in the morphological analysis result 66 (S56, NO), the morphological analysis result 66 is sent to the main keyword extraction means 14. . Since there is no unprocessed answer sentence ID (S52, NO), the process is terminated.

  The condition keyword extraction unit 13 receives from the condition pattern determination unit 12 a morphological analysis result sentence including the condition pattern and information indicating a location (detected condition pattern) that matches the condition pattern, and matches the condition pattern. A keyword existing in front of the location is extracted as a condition keyword.

The following [Method 1] to [Method 3] can be considered as methods for extracting the condition keyword.
[Method 1] A method of extracting a forward noun closest to the detected condition pattern as a condition keyword,
[Method 2] A method of extracting a word / phrase excluding a condition pattern from a clause including the detected condition pattern as a condition keyword,
[Method 3] A method of extracting, as a condition keyword, an addition of N clauses related to the clause including the detected condition pattern in [Method 2] above.

  The integer value N in [Method 3] may be set in advance or may be set by the user.

  FIG. 7 is a diagram illustrating an example of a result of dependency analysis extracted by the conditional keyword extraction unit 13 in the first embodiment.

  That is, FIG. 7 is a diagram showing a result 71 of dependency analysis of the morphological analysis result 62 of the sentence number 1 of the answer sentence ID 34 and a result 72 of dependency analysis of the morphological analysis result 64 of the sentence number 2. It is. In the result 71 of the dependency analysis, “adult” is extracted as a conditional keyword in [Method 1], and “adult only” is extracted as a conditional keyword in [Method 2]. Then, since there is no clause related to “if only adults”, “only adults” is extracted as a conditional keyword.

  In the result 72 of the dependency analysis, “together” is extracted as a condition keyword in the above [Method 1], and “together” is extracted as a condition keyword in the above [Method 2], and in the above [Method 3]. If the number of clauses related to the clause including the condition pattern is N = 1, “children together” is extracted as the condition keyword. The extracted condition keyword is associated with the extracted answer sentence ID and sentence number, and passed to the classification keyword determining means 15.

  The main keyword extraction unit 14 receives from the condition pattern determination unit 12 a sentence determined not to include the condition pattern, and performs dependency analysis. In this analysis result, it is checked whether there is a main phrase, and if there is a main phrase, the keywords (noun phrases) included in the main phrase are extracted as “main keywords”. . In the present invention, the case “OO is” is extracted as the main phrase. The case “XX is” may be an action subject, but is also used to describe the noun phrase “XX”. For this reason, the part that explains the noun phrase “XX” in the answer document is considered to be useful information for the questioner, and the case “XX” is extracted. I will do it.

  As the main keyword, N clauses related to the main clause were added as in [Method 3] above when extracting only the noun phrase part “XX” of the C case and when extracting the conditional keyword. You may make it extract a thing as a main keyword.

  FIG. 8 is a diagram illustrating the operation of the main keyword extracting unit 14.

  FIG. 8A is a diagram showing a result of dependency analysis of the answer sentence ID 33 when, for example, sentences with the answer sentence IDs 33 and 35 are input to the main keyword extracting unit 14. FIG. 8B is a diagram illustrating the result of dependency analysis of the sentence with the answer sentence ID 35.

  In the example shown in FIG. 8A, the phrase that becomes the main phrase is “umbrella is”, and the noun “umbrella” is extracted as the main keyword from this phrase. Similarly, in the example shown in FIG. 8 (2), the phrase that is the main phrase is “adult is”, and the noun “adult” is extracted from this phrase as the main keyword. The extracted main character keywords “umbrella” and “adult” are respectively associated with the answer sentence ID and the sentence number and sent to the classification keyword determination means 15 as “classification keyword”. Note that the classification keyword, the answer sentence ID, and the sentence number are associated with each other and sent to the classification keyword determination unit 15, and the answer sentence ID and the sentence number are not included in the classification keyword.

  The classification keyword determination means 15 inputs the condition keyword, answer sentence ID and sentence number extracted by the condition keyword extraction means 13, and the main keyword, answer sentence ID and sentence number extracted by the main keyword extraction means 14, and for each keyword, It is determined whether or not the conditions shown in the following expressions (1) and (2) are satisfied, and the keywords that satisfy the preset conditions are output as classification keywords.

  That is, the “classification keyword” is a keyword in which the condition keyword and the main keyword satisfy a preset condition.

  The preset condition is a condition according to the number of appearances X (w) that is the number of times that the keyword w appears as a conditional keyword and the number of appearances Y (w) that is the number of times that the keyword w appears as a main keyword. is there. For example, the following formula (1) is used as a condition to determine the determination formula F (w).

F (w) = αX (w) + (1 1α) Y (w) (1)
F (w) ≧ th (2)
Here, when 0 ≦ α ≦ 1 and α is set to a value close to 1, the above formula (1) becomes a higher value for a keyword having a larger number of appearances as a conditional keyword. Furthermore, according to the above formula (2), only keywords having a value equal to or greater than a preset threshold th can be determined as classification keywords.

  When counting up the number of appearances, only the keywords that exactly match are counted as the number of appearances, and when they partially match (partial match), they may be counted as the number of appearances. Below, the case where it counts as the frequency | count of appearance when it corresponds partially is demonstrated.

  FIG. 9 is an explanatory diagram of the operation of the classification keyword determination unit 15.

  For example, let us consider a case where two words “only adult” and “small child” are extracted by the conditional keyword extracting means, and two words “umbrella” and “adult” are extracted as the main keywords by the main keyword extracting means 14.

  The condition keyword information 91 is described in one line in the order of “condition keyword answer sentence ID sentence number”, separated by spaces. The main character keyword information 92 is described in one line in the order of “main character keyword answer sentence ID sentence number” in a space delimiter.

  FIG. 10 is a flowchart showing the operation of the classification keyword determination means 15.

  The classification keyword determination means 15 selects a classification keyword candidate (S101). The classification keyword candidates are a set of keywords from which duplicates are removed from the input condition keywords and main keywords.

  Four types described in the classification keyword candidate column in the total result 93 are selected. Next, if there is an unprocessed classification keyword candidate (S102, YES), one classification keyword candidate is selected, and the condition keyword appearance count and the main keyword appearance count of this classification keyword candidate are initialized to 0 ( S103).

  Next, the inputted condition keywords are selected one by one as “comparison keywords”, and the selected comparison keywords are compared with classification keyword candidates. For convenience, the condition keyword to be compared with the classification keyword candidate is expressed as “comparison keyword”.

  If there is an unprocessed conditional keyword (S104), one unprocessed conditional keyword is selected as a comparison keyword and compared with a classification keyword candidate (S105).

  As a result of this comparison, if the comparison keyword is a complete match or the classification keyword candidate includes a comparison keyword (partial match) (S106, YES), the condition keyword appearance count is incremented by 1 (S107). In other cases (S106, NO), the process returns to S104, and the processes of S104 to S107 are executed until there is no unprocessed conditional keyword. If there is no unprocessed conditional keyword (S104, NO), the process proceeds to a process of comparing the main keyword similarities.

  If there is an unprocessed main keyword (S108, YES), one unprocessed main keyword is selected as a comparison keyword and compared with the classified keyword Kyosuke (S109). As a result of this comparison, if the comparison keyword is a complete match, or if the comparison keyword is included in the classification keyword candidate (partial match) (S110, YES), the number of appearances of the main keyword is increased by 1 (S111). In other cases (S110, NO), the process returns to S108, and the processes of S108 to S111 are executed until there is no unprocessed main keyword.

  When there is no unprocessed main keyword (S108, NO), the process proceeds to the comparison processing of the next classification keyword candidate. A process (S102 to S111) of selecting an unprocessed keyword and counting the number of appearances of the condition keyword and the number of appearances of the main keyword is executed for all the classified keyword candidates, and the process ends.

  By the above comparison, the total result 93 shown in FIG. 9 can be obtained.

  In the example of the condition keyword information 91 and the main character keyword information 92 shown in FIG. 9, first, the classification keyword candidate “adult only” is selected (S102) and compared with the input condition keyword. When the condition keyword “adult only” is selected and compared (S105), the condition keyword appearance count is incremented by 1 (S107) because it matches completely.

  Next, although it is compared with the condition keyword “small child”, since the condition of S106 is not satisfied, it is checked whether or not an unprocessed condition keyword exists (S104). If there is no unprocessed condition keyword (S104, NO), a comparison process with the main keyword is executed. The unprocessed main keyword “umbrella” is selected and compared with the classification keyword candidate “adult only” (S109). As a result of the comparison, the condition of S110 that is a complete match or a partial match is not satisfied, and the process returns to S108. Since there is an unprocessed main keyword “adult”, “adult” is selected and compared with the classification keyword candidate “adult only” (S109).

  Since the comparison keyword “adult” is included in the classification keyword candidate “adult only”, the number of appearances of the main keyword is increased by one. Since there is no unprocessed main keyword (S108, NO), an unprocessed classification keyword candidate “small child” is selected (S103), and the processes of (S104 to S111) are executed. As a result, the condition keyword appearance count increases by 1 for the classification keyword candidate “small child” to 1, and the main keyword appearance count increases by 1 for the classification keyword supplement “umbrella”.

  In addition, comparing the classification keyword supplement “Adult” with the condition keyword “Adult only”, there is a partial match, but “Adult only” is not included in “Adult”, so the number of occurrences of the condition keyword is counted. I can't. When the classification keyword supplement “adult” and the main keyword “adult” are compared with each other, they match completely. As a result of processing in this way, a total result 93 shown in FIG. 9 can be obtained.

  In the first embodiment, the number of appearances is counted even when the classification keyword candidate and the comparison keyword are partially matched, but may be counted only when they are completely matched.

  A keyword that satisfies the predetermined condition for the number of appearances of the conditional keyword and the number of appearances of the main character thus determined is determined as a classification keyword. The predetermined condition is, for example, a condition shown in the above expressions (1) and (2).

  In the above formulas (1) and (2), when α = 0.8 and th = 1, the value of formula (1) for each keyword is as follows, and “only adult” is , Extracted as a classification keyword.

f (Adult only) = 0.8 * 1 + 0.2 * 1 = 1.0
f (small child) = 0.8 * 1 = 0.8
f (umbrella) = 0.2 * 1 = 0.2
f (adult) = 0.2 * 1 = 0.2

  In the first embodiment, for easy explanation, classification keywords are obtained from three answer documents. Originally, a large number of answer documents are input to obtain classification keywords. For this reason, depending on the value of the threshold value th in Equation (2), many keywords may be determined as classification keywords. In this case, the classification keyword may be determined by adding a condition based on the number of cases, such as the top number of values of the expression (1).

  The classification keyword determination unit 15 sends the classification keyword, the answer sentence ID including the keyword, and the sentence number information to the text output unit 40 in combination.

  The text output means 40 receives the classification keyword output by the classification keyword determination means 15 and classifies and outputs the answer document for each classification keyword. Since the process of extracting keywords that are classification keywords is performed in units of sentences, an example in which output is performed in units of sentences will be described.

  Based on the keyword received as input, the answer sentence ID including this keyword, and the information capital of the sentence number, the corresponding sentence is retrieved and acquired from the question / answer document accumulating means 30. Generate a set of A classification keyword and a set of sentences including the keyword may be output at a time. First, only the classification keyword may be displayed, and a set of sentences including the classification keyword selected by the user may be displayed.

  That is, the answer document classification apparatus 100 is an answer document classification apparatus that classifies the answer document for the question document for each group including a classification keyword that is a keyword representing the group.

  The condition pattern accumulating unit 11 is an example of a condition pattern accumulating unit that is configured by a combination of a word and a part of speech and that accumulates a condition pattern that is a pattern indicating a predetermined condition. The condition pattern determination unit 12 is an example of a condition pattern determination unit that determines whether a sentence included in the input text set includes a condition pattern stored in the condition pattern storage unit. The conditional keyword extracting unit 14 is an example of a conditional keyword extracting unit that extracts a conditional keyword that is an array of words arranged in front of the conditional pattern if the conditional pattern is included. The main keyword extracting means 14 is an example of a main keyword extracting means for extracting a main keyword, which is a word included in a phrase that becomes the main sentence, if the condition pattern is not included.

  In addition, the classification keyword determination unit 15 determines, for each keyword included in the conditional keyword and the main keyword, the number of appearances that is the number of times that the focused keyword appears as the conditional keyword and the number of times that the keyword appears as the main keyword. A classification keyword determining unit that counts a certain number of appearances and determines that the keyword of interest is a classification keyword when the number of appearances of the conditional keyword and the number of appearances of the main keyword satisfy a predetermined condition; It is an example.

  FIG. 11 is a block diagram showing an answer document classification apparatus 200 that is Embodiment 2 of the present invention.

  The answer document classification apparatus 200 accumulates the classification keyword candidates extracted by the conditional keyword extraction unit 16 and the main keyword extraction unit 17 in the classification keyword candidate accumulation unit 18, and the search result according to the search request from the user. The classification keyword candidates in the sentence having the answer sentence ID included in the are acquired from the classification keyword storage means 18, the condition keyword appearance count and the main keyword appearance count of each classification keyword candidate are aggregated, and the classification keyword is determined. Device.

  That is, the answer document classification device 200 is connected to the text input unit 21 and the text output unit 40, and the condition pattern storage unit 11, the condition pattern determination unit 12, the condition keyword extraction unit 16, and the main keyword extraction unit 17 are connected. And classified keyword candidate accumulating means 18 and classified keyword determining means 19. In addition, the same code | symbol is attached | subjected to the component same as the component in Example 1. FIG.

  The text input means 21 periodically acquires unprocessed answer documents stored in the question / answer document storage means 30 and sends them to the condition pattern determination means 12.

  The condition pattern determination unit 12 and the condition pattern storage unit 11 are the same as those in the first embodiment, and thus the description thereof is omitted.

  The condition keyword extraction unit 16 receives the sentence of the morpheme analysis result including the condition pattern and the part that matches the condition pattern from the condition pattern determination unit 12, and selects the keyword that exists in front of the part that matches the condition pattern as the condition keyword. Are extracted and stored in the storage device.

  In the first embodiment, an answer sentence ID and a sentence number are associated with the extracted condition keyword and passed to the classification keyword determining unit 15. In the second embodiment, a symbol indicating the condition keyword is included in the input morphological analysis result. And is recorded in the classification keyword supplement storage means 18. As described above, the reason for adding the symbol indicating the condition keyword to the input morpheme analysis result is to clearly indicate the condition keyword.

  FIG. 12 is a diagram illustrating an accumulation example of the classification keyword accumulation unit 18 according to the second embodiment.

  Symbols 122 and 123 indicating condition keywords are symbols indicating that the classification keyword candidates stored in the classification keyword candidate storage unit 18 are condition keywords, and a symbol “X” is added to the end of the classification keyword candidate line. ing. If the symbol “X” is given across a plurality of lines, it indicates that a phrase is extracted instead of a word.

  The main keyword extracting unit 17 receives a sentence determined not to include the condition pattern from the condition pattern determining unit 12, performs dependency analysis, and based on the result of the analysis, whether there is a main phrase in the sentence. If it exists, the keywords (words, nouns, noun phrases) included in the main phrase are extracted as main keywords and stored in the storage device.

  In the first embodiment, an answer sentence ID and a sentence number are associated with the extracted main keyword and passed to the classification keyword determining unit 15. In the second embodiment, a symbol indicating the main keyword is included in the input morphological analysis result. And is recorded in the classification keyword candidate accumulating means 18. Symbols 121 and 124 indicating the main keyword are symbols indicating that the classification keyword stored in the classification keyword candidate storage unit 18 is a main keyword, and a symbol Y is given to the end of the classification keyword candidate line.

  In the text input means 21, the condition pattern determination means 12, the condition keyword extraction means 16, and the main keyword extraction means 17 process each time the processing is periodically executed, and the classification keyword candidate storage means 18 includes the classification keyword candidates. Accumulate sentences.

  Next, when the user inputs a search keyword to the keyword search means 50, the keyword search means 50 searches the question / answer document storage means 30 and a set of answer documents that answer the question document including the search keyword. To get. A set of answer text IDs of the acquired answer document set is passed to the classification keyword determination means 19.

  When the classification keyword determination unit 19 receives the answer sentence ID from the keyword search unit 50 to be processed, the classification keyword determination unit 19 acquires a sentence with the same answer sentence ID from the classification keyword candidate accumulation unit 18. Condition keywords or main keywords are acquired from each sentence, and main keyword information consisting of a set of condition keyword information 91 consisting of a set of “condition keyword answer sentence ID sentence number” and a set of “main keyword answer text ID sentence number”. 92 sets are acquired, and the classification keyword determination process shown in FIG. 10 is executed to determine the classification keyword. As a result of the determination, the classification keyword, the answer sentence ID including the keyword, and the sentence number information are combined and sent to the text output means 40.

  In the first embodiment, the processing time is required to add the time of the morphological analysis process, the dependency process, and the classification keyword determination process until the classification keyword is output after the answer sentence ID is given from the text input unit 20. is there. Since these processes are processes proportional to the number of sentences, they are faster than the clustering process described in the prior art (proportional to the square of the number of sentences).

  In the second embodiment, the classification keyword candidate accumulating unit 18 extracts and accumulates the classification keyword candidates in advance, so that the classification keyword can be output in only the classification keyword determination processing time, which is very fast.

  The classification keyword candidate accumulating unit 18 adds an analysis result in which information indicating that it is a condition keyword or information indicating that it is a main keyword is added to each sentence of the answer document, and an answer sentence ID that uniquely indicates the answer document. This is an example of classified keyword candidate accumulation means for accumulating classified keyword candidates associated with.

  When a plurality of response sentence IDs are given, the classification keyword determination unit 19 acquires a set of analysis results of the answer document corresponding to the response sentence ID from the classification keyword complement accumulation unit, and adds the analysis result set to the set of analysis results. For each keyword included in the condition keyword and the main keyword included, the number of appearances, which is the number of times the keyword of interest appears as the condition keyword, and the number of appearances, the number of appearances as the main keyword, are tabulated. This is an example of a classification keyword determination unit that determines that the keyword of interest is a classification keyword when the number of appearances of the conditional keyword and the number of appearances of the main keyword satisfy a predetermined condition.

  Further, in the above embodiment, the conditional keyword extracting means adds a phrase obtained by adding zero or more phrases related to the phrase including the condition pattern to the phrase excluding the condition pattern from the phrase including the condition pattern. As a means of extraction. The main keyword extracting means is means for extracting, as the conditional keyword, a phrase obtained by adding zero or more phrases related to the main phrase to the noun phrase included in the main phrase. And the said classification keyword determination means is a means to determine the said classification keyword using the conditions which become easy to determine as the said classification keyword, so that there are many appearance frequency as the said conditional keyword.

  In the above embodiment, if the means is replaced with a process, the above embodiment can be grasped as a method invention. That is, the above embodiment is a pattern composed of combinations of words and parts of speech in an answer document classification method for classifying answer documents for a question document into groups each including a classification keyword that is a keyword representing the group. A condition pattern determination is made as to whether or not a sentence included in the input text set includes a condition pattern stored in a condition pattern storage means in which a condition pattern that is a pattern indicating a predetermined condition is stored. The condition keyword extracting means extracts a condition keyword that is an array of words arranged in front of the condition pattern if the condition pattern is determined and stored in the storage device and the condition pattern is included. If the condition keyword extraction step stored in the storage device and the condition pattern are not included, The main keyword that is a word included in the phrase to be rated is extracted by the main keyword extracting means and stored in the storage device, and the condition keyword and the keyword included in the main keyword are focused on. The number of appearances, which is the number of appearances of the keyword as the conditional keyword, and the number of appearances, which is the number of appearances of the main keyword, are counted, and the number of appearances of the conditional keyword and the number of appearances of the main keyword are predetermined. When the above condition is satisfied, this is an example of an answer document classification method including a classification keyword determination step in which a classification keyword determination unit determines that the keyword of interest is a classification keyword and stores it in a storage device.

  In the above embodiment, the answer document classification method for classifying the answer document for the question document into groups each including a classification keyword that is a keyword representing the group is a pattern configured by a combination of words and parts of speech. The condition pattern determination determines whether the sentence included in the answer document set includes the condition pattern stored in the condition pattern storage means that stores the condition pattern that is a pattern indicating a predetermined condition. The condition keyword extracting means extracts a condition keyword that is an array of words arranged in front of the condition pattern if the condition pattern is determined and stored in the storage device and the condition pattern is included. If the condition keyword extraction process stored in the storage device and the above condition pattern are not included, The main keyword that is a word included in the phrase is extracted by the main keyword extracting means and stored in the storage device, and information indicating that it is a condition keyword for each sentence of the answer document or the main keyword The classification keyword candidate accumulating unit accumulates the classification keyword candidate in which the analysis result to which the information indicating that the information is added and the answer sentence ID uniquely indicating the answer document are associated is stored in the storage device When a plurality of keyword candidate accumulation steps and answer sentence IDs are given, a set of the analysis results of the answer document corresponding to the answer sentence ID is acquired from the classification keyword supplement accumulation step, and the set of analysis results is obtained. Number of times the keyword of interest appears as the above condition keyword for each included condition keyword or for each main keyword When the number of appearances and the number of appearances, which is the number of appearances as the main keyword, are counted, and the condition keyword and the number of appearances of the main keyword satisfy a predetermined condition, the keyword of interest This is an example of an answer document classification method including a classification keyword determination step in which classification keyword determination is determined to be a classification keyword and stored in a storage device.

  Further, the above embodiment can be grasped as a program invention. That is, the said Example is an example of the program which makes a computer function as each means which comprises the reply document classification | category apparatus of any one of Claims 1-5.

100 ... Reply document classification device,
11 ... Condition pattern storage means,
12 ... Condition pattern determination means,
13: Conditional keyword extraction means,
14 ... Main keyword extraction means,
15: Classification keyword determination means,
20 ... Text input means,
30: Question / answer document storage means,
40 ... text output means,
200 ... Reply document classification device,
16 ... Conditional keyword extraction means,
17 ... main keyword extraction means,
18: Classification keyword candidate storage means,
19: Classification keyword determination means,
50: Keyword search means.

Claims (8)

In an answer document classification device that classifies answer documents for a question document into groups each including a classification keyword that is a keyword representing the group,
A condition pattern accumulating means for accumulating a condition pattern which is a pattern composed of a combination of a word and a part of speech and which indicates a predetermined condition;
Condition pattern determination means for determining whether a sentence included in the input text set includes a condition pattern stored in the condition pattern storage means;
Condition keyword extraction means for extracting a condition keyword that is an array of words arranged in front of the condition pattern if the condition pattern is included;
If it does not include the condition pattern, a main keyword extracting means for extracting a main keyword that is a word included in the main clause;
For each keyword included in the conditional keyword and the main keyword, the number of appearances, which is the number of times that the keyword of interest appears as the conditional keyword, and the number of appearances, which is the number of appearances as the main keyword, are tabulated. Classification keyword determination means for determining that the keyword of interest is a classification keyword when the number of appearances of a conditional keyword and the number of appearances of the main keyword satisfy a predetermined condition;
An answer document classification device characterized by comprising: In an answer document classification device that classifies answer documents for a question document into groups each including a classification keyword that is a keyword representing the group,
A condition pattern accumulating means for accumulating a condition pattern which is a pattern composed of a combination of a word and a part of speech and which indicates a predetermined condition;
Condition pattern determination means for determining whether or not a sentence included in the answer document set includes a condition pattern stored in the condition pattern storage means;
Condition keyword extraction means for extracting a condition keyword that is an array of words arranged in front of the condition pattern if the condition pattern is included;
If it does not include the condition pattern, a main keyword extracting means for extracting a main keyword, which is a word included in the main clause;
Classification keyword candidates in which an analysis result in which information indicating that it is a condition keyword or information indicating that it is a main keyword is added to each answer document sentence and an answer sentence ID that uniquely indicates the answer document Classification keyword candidate accumulation means for accumulating;
When a plurality of answer sentence IDs are given, a set of analysis results of the answer document corresponding to the answer sentence ID is acquired from the classification keyword supplement accumulation means, and the condition keyword included in the set of analysis results and For each keyword included in the main keyword, the number of appearances, which is the number of times that the keyword of interest appears as the condition keyword, and the number of appearances, which is the number of appearances as the main keyword, are tabulated, and the number of appearances of the condition keyword And a classification keyword determination unit that determines that the keyword of interest is a classification keyword when the number of appearances of the main keyword satisfies a predetermined condition;
An answer document classification device characterized by comprising: In claim 1 or claim 2,
The conditional keyword extraction means is a means for extracting, as a conditional keyword, a phrase obtained by adding zero or more phrases related to a phrase including a conditional pattern to a phrase excluding the conditional pattern from the phrase including the conditional pattern. An answer document classification device characterized by In claim 1 or claim 2,
The prominent keyword extracting means is a means for extracting, as the conditional keyword, a phrase obtained by adding zero or more phrases related to a prominent phrase to a noun phrase included in the prominent phrase. Answer document classification device. In any one of Claims 1-4,
An answer document classification device, wherein the classification keyword determination means is a means for determining the classification keyword using a condition that is more likely to be determined as the classification keyword as the number of appearances as the conditional keyword increases. In an answer document classification method for classifying answer documents for a question document into groups each including a classification keyword that is a keyword representing the group,
The input text set includes a condition pattern stored in a condition pattern storage means that is a pattern composed of a combination of words and parts of speech and that stores a condition pattern that is a pattern indicating a predetermined condition. A condition pattern determination means for determining whether or not the sentence contains a sentence, and storing it in a storage device;
A condition keyword extraction step in which the condition keyword extraction means extracts a condition keyword that is an array of words arranged in front of the condition pattern and stores the condition keyword in a storage device;
If it does not include the condition pattern, a main keyword extraction means that the main keyword extracting means extracts a main keyword that is a word included in the main clause and stores it in a storage device;
For each keyword included in the conditional keyword and the main keyword, the number of appearances, which is the number of times that the keyword of interest appears as the conditional keyword, and the number of appearances, which is the number of appearances as the main keyword, are tabulated. When the number of appearances of the condition keyword and the number of appearances of the main keyword satisfy a predetermined condition, the classification keyword determination unit determines that the keyword of interest is a classification keyword, and stores the classification keyword in the storage device A judging step;
An answer document classification method characterized by comprising: In an answer document classification method for classifying answer documents for a question document into groups each including a classification keyword that is a keyword representing the group,
The answer document set includes condition patterns stored in the condition pattern storage means that are configured by combinations of words and parts of speech and that store condition patterns that are patterns indicating predetermined conditions. A condition pattern determination means for determining whether or not the sentence contains a sentence, and storing it in a storage device;
A condition keyword extraction step in which the condition keyword extraction means extracts a condition keyword that is an array of words arranged in front of the condition pattern and stores the condition keyword in a storage device;
If it does not include the condition pattern, a main keyword extraction means that the main keyword extracting means extracts a main keyword that is a word included in the main clause and stores it in a storage device;
Classification keyword candidates in which an analysis result in which information indicating that it is a condition keyword or information indicating that it is a main keyword is added to each answer document sentence and an answer sentence ID that uniquely indicates the answer document And a classified keyword candidate accumulating step in which the classified keyword candidate accumulating means accumulates and stores in the storage device;
When a plurality of response text IDs are given, a set of the analysis results of the response document corresponding to the response text ID is acquired from the classification keyword supplement accumulation step, and the condition keyword included in the analysis result set For each keyword included in the main keyword, the number of appearances, which is the number of times that the keyword of interest appears as the condition keyword, and the number of appearances, which is the number of appearances as the main keyword, are tabulated, and the appearance of the condition keyword A classification keyword determination step in which the classification keyword determination determines that the keyword of interest is a classification keyword when the number of times and the number of appearances of the main keyword satisfy the predetermined condition;
An answer document classification method characterized by comprising:   6. A program for causing a computer to function as each means constituting the answer document classification device according to claim 1.

Images