JP2002230006A - Method for analyzing free descriptive answer, method for extracting keyword from free descriptive document, and method for supporting analysis of the document - Google Patents

Abstract

PROBLEM TO BE SOLVED: To construct a keyword data base by using an actual free descriptive answer text, and to analyze the free descriptive answer text based on the keyword data base. SOLUTION: This method comprises a process for generating the elements of keywords by extracting the same character string included in the answers of not less than two answerers from a plurality of free descriptive answers, a process for constructing a keyword data base by erasing the same overlapped expression from the elements of the keywords, and erasing any unnecessary character string which can not be any keyword, a process for comparing the plurality of free descriptive answers with the data base, and for, when any character string including each keyword is present in each free descriptive answer, counting it as 'the presence of reaction', and a process for relating the attribute of a person whose reaction is counted with the keyword.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for analyzing a free description answer such as a response to a questionnaire and a method for extracting a keyword from a large number of free description documents. In the present specification, the "free answer" includes not only documents that exist in text format from the beginning,
Also includes the resulting document data. The "freely described answer" includes, for example, electronically documented information obtained from a consumer or a respondent via voice at a call center.

[0002]

2. Description of the Related Art With the spread of e-mail, opportunities for obtaining responses to questionnaire surveys from customers and opinions of customers as electronic documents are increasing. Usually, a document sent as an electronic document is a so-called free description answer sentence, and the analysis of the obtained free description answer sentence such as a questionnaire result has conventionally been performed manually. However, manually analyzing hundreds or thousands of documents by hand is extremely cumbersome and time consuming.

Recently, automatic analysis of Japanese sentences by computers has been advanced, and electronic documents are searched by keywords. However, the answers and general opinions of the questionnaire survey are free description answers, and even if a database of keywords is created in advance, it is not possible to completely correspond to all possible free description answers. Further, even in a questionnaire survey of the same question, the answer content may change significantly depending on the survey time and the like, and keywords may not be narrowed down unless the answer content is actually analyzed.

[0004]

SUMMARY OF THE INVENTION According to the present invention, a keyword database is constructed by using an actual free description answer, and the free description answer is analyzed based on the keyword database and the analysis is supported. The purpose is to do so. Another object of the present invention is to extract keywords from a large number of free description sentences.

[0005]

The technical means adopted by the present invention to achieve the object is to extract the same character string included in the answers of two or more respondents from a large number of freely described answers, Generating a keyword database, deleting redundant identical expressions from the keyword components, deleting unnecessary character strings that cannot be keywords, and constructing a keyword database. And, if there is a character string including each keyword in each free description answer, counting as "reaction", and associating the attribute of the person whose response was counted with the keyword. is there.

[0006] Another technical means adopted by the present invention is a step of extracting the same character string included in the answers of two or more respondents from a large number of freely described answers to generate a prime of a keyword.
A step of deleting the same character string that is duplicated from the element of the keyword, deleting unnecessary character strings that cannot be a keyword, and extracting keyword candidates, comparing a keyword database stored in advance with the keyword candidates,
If the keyword candidate is a character string that does not exist in the database, updating the database by adding it as a new keyword and constructing a new database; and comparing the large number of freely described answers with the new database. If there is a character string including each keyword of the new database in each free description answer, counting is performed as “reacted”, and a step of associating the attribute of the person whose response is counted with the keyword. is there. This technique is, in one preferred embodiment, employed for subsequent analyzes of the same question. That is, the existing keyword database is constructed by the method described in claims 1 and 17. However, a preliminary keyword database may be created in advance, and this may be used as an existing keyword database. For example, when there is a certain question, a thesaurus of keywords (important words) that can be expected as answers may be prepared, and this may be used as a keyword database. In one preferred embodiment, each vocabulary constituting the thesaurus is provided with an identification symbol (for example, a numeral), and when the obtained keyword includes any of the vocabulary constituting the thesaurus, the keyword is included. Assigning an identification symbol of the vocabulary to the keyword, and classifying the obtained keyword based on the assigned identification symbol. In this way, the extracted keywords can be automatically classified using the identification symbol as a key.

[0007] The two or more answers basically mean the answers of two or more different respondents, but the two or more answers also include a duplicate of the same person's answer. .
For example, when the number of answers is small, it may be useful to extract the element of the keyword from the answer of the same person.

The keyword extraction method employed by the present invention includes a step of generating a keyword prime and a step of extracting a keyword from the keyword prime. Extracting a longest matching expression in a character string between free description documents, and repeating the process until the extracted character string reaches a predetermined number of characters, thereby obtaining at least two or more documents from a large number of free description documents. Is extracted to generate the prime of the keyword by extracting the same character string included in. In the step of extracting the keyword, the same character string that is duplicated from the element of the keyword is deleted, and an unnecessary character string that cannot be a keyword is deleted to extract a keyword candidate.

The free description document includes one or more words, one or more phrases, one or more sentences,
And any one of these arbitrary combinations.

Still another means adopted by the present invention has a step of extracting an expression that matches the longest in a character string between a large number of freely described documents, and the character string to be extracted in the step has a predetermined number of characters. This is a free description document analysis support method that includes extracting the same character string included in at least two or more free description documents and generating keyword candidates by repeating the above steps. How to use, use, analyze, further process, etc. the keyword candidates extracted by the present support method are not limited. For example, a manual operation may be included in the arrangement of the obtained keyword candidates.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention relates to a method for analyzing a free description answer using a computer. The open-ended answer is electronic document data, and in one preferred embodiment, is an electronic document obtained by electronic mail from the subject,
The method of generating the electronic document data is not limited, and includes, for example, electronic documents of handwritten documents and voices. Such an electronic document is stored in a storage device of a computer and analyzed by the analysis method according to the present invention. In one preferred embodiment, the present invention is used for analyzing a large number of answers to a certain question obtained from a large number of subjects.

[Acquisition of free description] In response to the question "Lunch was the most delicious this summer", answers were obtained by electronic mail from many subjects by e-mail. Each subject is associated with the attribute of the subject, and this answer includes gender and age as attributes of the subject, but the attributes are not limited to these. The association between the subject and the attribute may be such that the gender and age are entered at the time of answering, or if the answer is obtained from the member, the gender and age previously stored in the computer may be used. . For example, in the latter example, a member number is input at the time of answering, and attributes such as gender and age are associated with the answer using the member number as a key. FIG. 1 shows a part of an answer in a free description sentence. Age and gender are associated as attributes together with the free description answer.

[Generation of Keyword Prime] In an answer in a free description sentence, the same expression (character string) of a plurality of persons is detected and stored as the keyword prime. In other words, character strings existing in two or more answers are used as keywords. For example, if the answer "Hamburger and fried lunch" is picked up first, then the remaining answer about the presence of the same character string as "Hamburger and fried lunch" is first or last sentence, or Scan from both sides.
First, the same character string as “Hamburger and fried lunch” is compared with other answers, and then the character string is sequentially reduced. Therefore, the string "hamburg"
When it shrank, "Hamburg lunch" was included in other answers
If the character string “hamburg” exists in a plurality of answers, it is picked up as a keyword element. When the character string is shortened to "Fried fried lunch", if there is a character string "Fried fried lunch" in other answers, the character string "Fried fried lunch" is present in multiple answers, It is picked up as a keyword element. Alternatively, when the character string is shortened to “Fried chicken” and there is a character string “Fried chicken lunch” in other answers, the character string “Fried chicken” exists in multiple answers, so the keyword Will be picked up as Further, when the character string is "Lunch box", if the other answer includes the character string "Lunch box of tempura", the string "Lunch box" is picked up as a keyword element. The shortest string is 2
Characters or three characters. As is apparent from the above description, in the analysis method according to the present invention, a plurality of keywords may be generated from one answer. FIG. 2 shows an excerpt of a keyword element. The numbers to the left of the string are simply serial numbers.

[Arrangement of Element of Keyword] The element group of the keyword is rearranged in the order of Iueo. FIG. 3 shows the rearranged files. Further, a file in which the same expression is deleted from the rearranged file is created. This is shown in FIG. For example, in FIG. 3, there are a plurality of character strings “onigiri”, but by deleting the duplicate character strings, in FIG. 4, the character string “onigiri” becomes one. Certain reordering of the keyword groups is not essential, but is useful for deleting identical expressions.

The method according to the present invention picks up a character string that is common to a plurality of answers regardless of the meaning of the character string. Therefore, the keyword element includes a character string that cannot be a keyword. Therefore, a step of deleting a character string that cannot be a keyword from a keyword element is required. The first of the elementary character strings of the keywords is a long note (-), a repellent sound (n),
Also, if the beginning is Kanji, Hiragana, Katakana, or a symbol other than the alphabet (for example,

[,], <,>, ",", ",", [,],! , #,
＄,%, &, =,? Etc.), these are automatically deleted. This work can be performed manually from the input means on the display device of the computer, but it is desirable to perform the work automatically by mechanical processing from the viewpoint of labor saving.

Further, when a keyword that is not appropriate is included, the keyword is automatically or manually deleted. For example, the character string “bento” included in FIG. 4 is not appropriate as a keyword, and should be deleted from the keyword. Such an unnecessary keyword may be deleted by creating and storing an unnecessary word database, so that a character string in the unnecessary word database is automatically dropped from the keyword source. Alternatively, it may be manually deleted from the keywords. These steps may be performed at the same time in “database construction” described below.

[Construction of Database] The arranged keywords themselves constitute a database, and grouping these keywords is useful for later analysis. FIG. 5 shows an excerpt of the classified keywords. Grouping is performed by a method of classifying a common character string into the same class (character strings having a predetermined matching rate may be classified into the same class by determining a matching rate of character strings), or a synonym or meaning having a synonym or meaning. For example, similar words are classified into one kind. For example,
By combining the items including the character string “Maku no Uchi bento”, “Maku no Uchi bento”, “Japanese style curtain inner lunch”, and “Deluxe curtain inner lunch” are grouped into a “Maku no Uchi bento” group (class). Alternatively, the characters including the character string “Kalbi” are put together to make “Kalbee bento,“ Gyu-Kalbi-don ”, and“ Stamina-Kalbi-don ”into the“ Kalbi bento ”group (class).

A dictionary of synonyms or similar words is stored in advance, and "Kalbi" and "Kalbee" are synonymous, and "Karaage", "Karaage", "Karaage", and "Karaage" are Synonymous, or "spaghetti" and "peperoncino" may be classified by automatically determining that they are similar. Note that the present invention does not exclude that a manual operation is added to the classification operation. For example, the character strings “do not buy”, “particularly none”, “particularly none”, “separately”, and “do not eat” may be manually classified as a “do not buy” group.

When analyzing the second and subsequent free description answer sentences for the same theme, a "keyword element" is created by the above-described method and arranged. The arranged “keyword candidates” are compared with a database already constructed. When the same character string as the “keyword candidate” exists in the “database”, it is not necessary to newly add the “keyword candidate” to the database. If the "keyword candidate" is different from the character string in the existing "database", it is determined whether to include the keyword in an existing group or create a separate group. The determination is performed by determining a matching rate between the keyword candidate and the character string of the existing keyword, and includes a keyword having a predetermined matching rate in the same class as the compared keyword. Produces a class of Note that such a determination may be made manually.

[Analysis of Freely Descended Answer] An analysis data sheet is created based on the grouping of keywords. The analysis data sheet is a sheet composed of a keyword (character string) at the head of each group. A plurality of keywords exist hierarchically at each head keyword. As shown in FIG. 6, "010 Calbee lunch" includes "beef rib lunch", "Calbee bowl", and "Calbi bowl".
Compare this analysis data sheet with the free text answer,
If there is an answer that includes the same character string as the keyword, all are counted as "Responsive". FIG. 7 is a diagram exemplifying a file created at the time of the free description answer analysis,
A list of responses to “010 Calbee lunch” and a list of responses to “030 Act Uchino bento” are shown. In the analysis method of the present invention, one answer may respond to a plurality of keywords, and the analysis result has the same meaning as "one or more answers possible". The analysis data sheet is useful for organizing the analysis results later.
The “reaction” may be counted by comparing the database with the free description answer, and may be arranged for each class later.

FIG. 8 is a table displaying the analysis results. The number of reactions for each keyword in the analysis data sheet is represented by gender (male and female) and age. Layer (boy / girl, 20's, 30)
Teens and forties). FIG. 9 is a diagram further illustrating a positional relationship between elements based on principal component analysis.

[Another Example of Freely-Descripted Answer] FIG. 10 shows an excerpt of a freely-described answer related to the “automobile purchase process”. The number on the left is a member number, and the attribute of the member (subject) who answered is extracted using the member number as a key, and is associated with the answer. "The process of purchasing a car"
Generally tend to be longer than the answer to the above-mentioned "Delicious Lunch Box", but the analysis method according to the present invention is not affected by the length of the free description answer.

An example of the analysis when the free description answer is long will be described. The character string from the beginning of the sentence to the first punctuation mark or the punctuation mark, and the character string from the punctuation mark to the next punctuation mark or the punctuation mark is set as the longest character string, and the character string is sequentially reduced therefrom (that is, the punctuation marks do not form a character string). For example, if there is a free description answer such as "Searching for cars that match the price range and rough conditions on the Internet. Car models / manufacturers will be narrowed down, contact after narrowing down to some extent, estimate and then test drive the actual car" Free-form answers are divided into three categories: "Searching for cars that match the price range and rough conditions on the Internet", "Narrowing down vehicle models / manufacturers", and "Contacting them after narrowing down to some extent and estimating and then testing actual cars" To divide. Regarding "Searching for a car that meets the price range and rough conditions on the Internet", first, check whether the same character string as "Search for a car that matches the price range and rough conditions on the Internet" exists in other open-ended answers. As a result, while gradually shortening the character string, it is determined whether or not the character string matches with the character string in another free description answer. For example, when the character string is shortened to "on the Internet" and another free description answer includes the character string "search on the Internet", "on the Internet" is extracted as a keyword element. In the same process, the keywords are extracted for "car type / manufacturer is narrowed down" and "test after a certain squeeze to contact and estimate the actual vehicle". After extracting the element of the keyword, the free description answer is analyzed in the same steps as in the above-described embodiment.

[Classification of Keywords Using Thesaurus] Classification of keywords using a thesaurus will be described.
FIG. 11 is an example of a thesaurus in which important words collected from a survey on labor issues are classified. The method of creating the thesaurus is not limited. For example, the thesaurus is created by extracting a keyword using the method according to the present invention and manually classifying the extracted keyword. FIG. 12 is a diagram showing a keyword extracted by using the method according to the present invention collated with the thesaurus shown in FIG. 11. In the case where a vocabulary in the thesaurus is found, up to three keywords are included. Candidates are assigned. A column of three numbers from the left indicates a code number. For example, for the keyword “sales performance”, “sales (4
Two vocabularies of “1)” and “graduation (11)” are assigned. A value of 0 in the second and third columns indicates that there is no corresponding second or third candidate. By the way, the number to the left of the keyword is the number of hits. Then
The keywords classified into the same code number in the thesaurus are automatically collected and the keywords are classified. For example, keywords to which the code number (11) is assigned include “its performance”, “points”, “business results”, “basic salary + actual results”, “performance”, and “business results”. Classify as one type. When a final analysis data sheet is created, manual fine adjustment may be performed.

[0025]

According to the present invention, the free description answer can be automatically analyzed or the analysis can be supported, so that the analysis can be labor-saving and the analysis time can be reduced as compared with the conventional manual operation. Can be drastically improved. According to the present invention, keywords are picked up by comparing each free description answer sentence (raw data) and extracting the same expression (character string) without decomposing the free description answer into morphemes. The means is simple and a keyword database can be constructed from up-to-date keywords. Furthermore, since the free description answer sentence is searched by the keyword composed of the expression in the actual free description answer sentence, the free description answer sentence including the keyword can be reliably counted.

[Brief description of the drawings]

FIG. 1 is a diagram showing an example of a free description answer.

FIG. 2 is a diagram illustrating a part of a keyword extracted as the same expression of a plurality of persons;

FIG. 3 is a diagram showing a part of the arrangement of FIG.

FIG. 4 is a diagram showing a part of the configuration shown in FIG. 4 from which redundant expressions have been deleted;

FIG. 5 is a diagram showing a part of a keyword obtained by deleting unnecessary words from those shown in FIG. 5 and classifying keywords;

FIG. 6 is a diagram showing a data sheet for analysis.

FIG. 7 is a diagram showing an example of a file created when analyzing a freely described answer using a keyword database.

FIG. 8 is a table showing the relationship between the number of reactions and attributes for each keyword.

FIG. 9 is a diagram showing a positional relationship of elements by principal component analysis.

FIG. 10 is a diagram illustrating another free description answer sentence.

FIG. 11 is a diagram showing a thesaurus for classifying keywords.

FIG. 12 is a diagram obtained by comparing the obtained keyword with a thesaurus.

Claims (22)

[Claims] (A) extracting the same character string included in two or more answers from a large number of freely described answers to generate a prime of the keyword; and (b) repeating the same expression from the prime of the keyword. And a step of constructing a keyword database by deleting unnecessary character strings that cannot be keywords, and (c) comparing the large number of free description answers with the database, and including each keyword in each free description answer If there is a character string, count as "Responsive",
(D) associating the attribute of the person whose reaction has been counted with the keyword and analyzing the free-form answer. 2. The method according to claim 1, wherein the two or more answers are answers of two or more different respondents. 3. The method according to claim 1, wherein the two or more answers include a copy of the same person's answer. 4. The analysis of a free description answer according to claim 1, wherein the step (a) includes extracting an expression that matches the longest character string between the free description answers. Law. 5. The free description according to claim 4, wherein the step of extracting the expression that matches the longest in the character string between the free description answers is repeated until the extracted character string reaches a predetermined number of characters. How to parse the answer. 6. A method according to claim 5, wherein the predetermined number of characters is two or three. 7. A method according to claim 4, wherein the character string from the beginning of the sentence to the first punctuation mark or reading point and the character string from the punctuation mark to the next punctuation mark or reading point is the longest character string, and the character string is sequentially reduced therefrom. Analysis method of open-ended answer which is the feature. 8. The method according to claim 1, wherein in the step (b), the step of deleting an unnecessary character string deletes a character string whose head is any one of a prompt sound, a long note, a sound repellent sound, and a symbol. A method of analyzing a free-form answer, which includes: 9. An analysis of a free description answer according to claim 1, wherein in step (b), the step of deleting unnecessary words is performed by comparison with a previously stored unnecessary word database. Law. 10. The method of analyzing a free description answer according to claim 1, wherein the step (b) further includes a step of classifying a keyword. 11. A method according to claim 10, further comprising: classifying keywords including the same character string into the same class. 12. A free description answer analysis method according to claim 10, further comprising: classifying keywords having similar meanings into the same class. 13. A method according to claim 12, wherein keywords are classified by a previously stored similar word database. 14. The free description answer analysis method according to claim 1, wherein the free description answer sentence and the attribute information of the respondent are associated in advance. 15. A step of (a) extracting the same character string included in two or more answers from a large number of freely described answers to generate a prime of a keyword; and (b) an identical character duplicated from the prime of the keyword. (C) comparing a keyword database stored in advance with the keyword candidate, and deleting the keyword candidate and deleting the unnecessary character string that cannot be a keyword. If the character string does not exist in the list, a step of adding a new keyword and updating the database to construct a new database; and (d) comparing the large number of free description answers with the new database, and If there is a character string including each keyword of the new database in the descriptive answer, counting as "reaction"; Analysis of open-ended answers with a step, the associating the counted human attributes and the keyword. 16. A method according to claim 15, wherein keywords in the existing database are classified, and when a new keyword is added, it is determined whether to include the keyword in an existing class or create another class. Analysis method of open-ended answer including. 17. The keyword database according to claim 15, wherein the keyword database is a thesaurus, and each vocabulary constituting the thesaurus is provided with an identification symbol, and the obtained keywords constitute the thesaurus. If any of the vocabulary words is included, an analysis method of a free description answer including a step of assigning an identification symbol of the vocabulary to the keyword and classifying the keyword obtained based on the assigned identification symbol. 18. A method comprising the steps of: (a) extracting a longest-matching expression in a character string between a large number of freely described documents, and repeating this step until the extracted character string has a predetermined number of characters. Extracting the same character string included in two or more free description documents and generating a keyword prime;
(B) a step of deleting the same character string that is duplicated from the element of the keyword, and a step of extracting an unnecessary character string that cannot be a keyword and extracting a keyword candidate. 19. The free description document according to claim 18, wherein the free description document includes one or more words, one or more phrases, one or more sentences, and any combination thereof. A keyword extraction method characterized in that: 20. The keyword extracting method according to claim 18, wherein the predetermined number of characters is two or three. 21. The method according to claim 18, wherein
A keyword extraction method characterized in that a character string from the beginning of a sentence to the first punctuation mark or a punctuation mark or a character string from a punctuation mark to the next punctuation mark or a reading point is the longest character string, and the character string is sequentially reduced therefrom. 22. A method of extracting a longest matching expression in a character string between a large number of freely described documents, and repeating this step until the extracted character string has a predetermined number of characters, so that at least two or more characters are extracted. A free description document analysis method including extracting the same character string contained in a free description document and generating keyword candidates.