JP7324577B2 - Text processing method and text processing device - Google Patents

Description

Article 30, Paragraph 2 of the Patent Act applies [Exhibition name] 2nd AI/Artificial Intelligence EXPO [Date] April 4 to 6, 2018 [Venue] Tokyo Big Sight [Posting address] https:// /content-tokyo2018. tems-system. com/eguide/jp/AI/details? id=868

Application of Article 30, Paragraph 2 of the Patent Law [Seminar name] SOLIZE Innovations seminar [Date] June 15, 2018 [Venue] TKP Shinagawa Conference Center 6F Banquet Hall 6G [Posting address] https://www. soly-group. com/event/2018/index. html

The present invention relates to a computer-based text processing method and apparatus, and more particularly to a technology for evaluating the degree of relevance between a relevant part described in text data and a topic of interest from input text data.

Many technologies have been proposed for retrieving a desired document from among a large number of electronic documents. The simplest one is a technique of inputting a word to be searched and extracting and displaying documents containing the word. Searches on the Internet are also performed by extracting and presenting web data including search terms from a huge amount of data automatically collected by search robots.

It is one of the great effects of computer text processing methods to extract documents containing search terms from an extremely large amount of documents in a short time. There is a need for a method of evaluating not only whether a search term is included, but also how relevant it is.

For example, Japanese Laid-Open Patent Publication No. 2002-301000 proposes a material retrieval device or the like for retrieving books/materials that match a user's interest based on a written document image. The apparatus includes text extracting means for performing character recognition processing on a document image including writing to extract text data, writing extracting means for extracting the type and position of writing, and a first characteristic word of the material to be retrieved and its importance. calculating the importance of the second feature word using a storage means for storing a search index including the index, a feature word extraction means for extracting the second feature word from the text data, and the type and position of the writing; , a configuration comprising feature data creating means for creating the feature word data of text data and relevance calculating means for calculating the relevance between the search index and the feature word data.

Patent Literature 2 discloses a method of extracting representative opinions from a set of texts on a specific topic. That is, in a representative comment extraction method for extracting a representative comment from a text set containing a plurality of comments, a key word extraction step of extracting key words that indicate the characteristics of the text set from the text set, and a plurality of comments, It is disclosed that representative comments are extracted by a clustering step of clustering based on important words and a representative comment extraction step of extracting representative comments from each cluster divided by clustering.

As a related technique, Patent Document 3 proposes a method of automatically learning feature words corresponding to product categories. A system for automatically learning characteristic words corresponding to each product category, comprising means for obtaining a maximum of 1000 web texts by a search engine using the category name as a query, and maximum 1000 obtained using the category name as a query. means for extracting feature word candidates from the Web text of; means for calculating a degree of relevance between the extracted feature word candidates and categories; as a feature word corresponding to the category, and a feature word database that associates and stores the feature word corresponding to the category.

JP 2015-179385 A JP 2013-15971 A Japanese Unexamined Patent Application Publication No. 2010-9307

In the conventional technology described above, it is possible to display characteristic words and key words related to user interests and representative opinions in text data. It is not possible to ascertain whether or not it is particularly important.

SUMMARY OF THE INVENTION The present invention was created in view of the above-mentioned problems of the prior art, and it is an object of the present invention to provide a technique for evaluating the degree of relevance between a relevant part described in text data and a topic of interest. and

In order to solve the above problems, the present invention provides the following text processing method and apparatus.
First, according to the first embodiment, with respect to topics, which are items obtained by classifying topics of interest from input first text data, the degree of relevance between related parts described in the text data and the topic is evaluated. A method for processing text by a computer is provided. In the method,
(S1) An input means inputs a plurality of second text data groups, a plurality of predefined topics, and related character string information describing the topics in the second text data group associated in advance for learning. data input step for learning to input as data,
(S2) a feature word extraction step in which the feature word extraction means extracts feature words from the related character string information;
(S3) After a learning step in which the aspect data creating means creates aspect data, which is information in which the feature words extracted for each topic are associated, and stores the aspect data in the storage means,
(S4) a text data input step in which the input means inputs the first text data;
(S5) a feature word search step in which the feature word search means refers to the aspect data and searches for a feature word contained in the first text data for at least one topic;
(S6) The output means performs a relevance evaluation step having each step of an output step of outputting, as the relevance, a value based on the search result of the feature word for each topic.

According to the second embodiment, in the feature word extraction step of the text processing method, when the feature word extraction means extracts the feature word from the related character string information, based on the weight information defined for the feature word: A weight value is set by a predetermined arithmetic expression in the aspect data, and the weight value is stored in the aspect data for the feature word extracted for each topic. In the feature word search step, the search condition is determined based on the weight value.

According to the third embodiment, the weight value defined for the feature word may be a value related to at least one of the appearance frequency and the co-occurrence frequency calculated when the feature word extracting means extracts the feature word. .

According to the fourth embodiment, in the above-mentioned output step, a calculated value relating to the number of a plurality of feature words, which are search results for each topic, or the sum of their weight values can be output as the above-mentioned degree of relevance.

According to the fifth embodiment, the feature word extracted for each topic and the distance value related to at least one of the distances before and after the appearance of the feature word are stored in the aspect data, and in the feature word search step , a search condition for the same type or different type of feature words included in the first text data can be determined based on at least the distance value.

According to the sixth embodiment, in the feature word search step, the feature word search means searches for the feature word based on the aspect data, reads the distance value of each extracted feature word, and It is possible to output a value calculated by a predetermined operation from the weight value of the feature word in , as the degree of relevance.

According to the seventh embodiment, in the aspect data creating step, the aspect data creating means defines hierarchical information for each feature word extracted for each topic, and the hierarchical information includes a weight value and a distance A pattern of value combinations can be defined, and in the feature word search step, a feature word search condition can be determined according to the pattern.

According to the eighth embodiment, in a configuration in which topics are classified into two or more levels of concept information from superordinate concepts to subordinate concepts so that readers can easily understand the content contained in the text data, the above aspect data is created. In the step, the aspect data creating means can automatically define the hierarchical information according to the conceptual information, and associate each of the feature words for each topic with the conceptual information.

According to the ninth embodiment, in the above-described output step, the output means may display the extracted feature words and their degrees of association in a predetermined graph.

According to the tenth aspect, in the output step, the output means may display a predetermined graph according to the position of the line of the extracted characteristic word in the configuration for displaying the first text data. can.

According to the eleventh embodiment, in the above output step, the output means displays the first text data, in which the entire sentence containing the feature word, or a predetermined range of sentences near the feature word, or the feature The presentation within the range of word distance values can be varied.

According to the twelfth aspect, in the above output step, for each feature word belonging to the concept information, the output means outputs the entire sentence containing the feature word, or a predetermined range of sentences near the feature word, or the feature It is also possible to change the display mode within the range of the distance value of the words so that each sentence including each concept from a higher concept to a lower concept can be displayed so as to be distinguished.

According to a thirteenth embodiment, a text processing device can also be provided.
That is, it is a text processing apparatus using a computer that evaluates the degree of relevance between a related part described in text data and a topic, which is an item obtained by classifying a topic of interest from input first text data. a plurality of second text data groups, a plurality of predefined topics, and related character string information describing the topics in the pre-associated second text data group as learning data. Aspect data input means, feature word extraction means for extracting feature words from related character string information, and aspect data that is information in which the feature words extracted for each topic are associated with each other are created and stored in a storage means. text data input means for inputting first text data; feature word search means for searching for feature words contained in the first text data for at least one topic by referring to aspect data; and an output means for outputting a value based on the search result of the feature word as the degree of relevance.

According to the present invention, it is possible to provide a technology for evaluating the degree of relevance between a related part described in text data and a topic of interest.

1 is an overall view of a text processing device (1) in the present invention; FIG. 1 is a flow chart of a text processing method according to the present invention; FIG. 4 is an explanatory diagram of a method of calculating a degree of association according to the present invention; FIG. 4 is an explanatory diagram of a method of calculating a degree of association according to the present invention; It is a first screen display example of the present invention. It is a second screen display example of the present invention. It is the 3rd example of a screen display of this invention. It is a fourth screen display example of the present invention.

An embodiment of the present invention will be described below with reference to the drawings. The present invention is not limited to the following examples, and can be carried out as appropriate within the scope of the claims.
FIG. 1 is an overall view of a text processing device (1) in the present invention. This device (1) can be implemented by a known personal computer, and can also be implemented in a server device such as a web server device. The details of these devices are well known, so descriptions thereof will be omitted.
FIG. 2 is a flow chart of the text processing method according to the present invention.

The text processing method of the present invention is roughly divided into two steps. A learning step in which machine learning is performed in advance, and a relevance evaluation step in which the degree of relevance is evaluated for related parts in the input text using the aspect data created by the learning step. In this embodiment, the learning process and the relevance evaluation process are described as being continuous, but if aspect data is created and stored, only the relevance evaluation process can be implemented and executed.

In the learning process, first, the input processing section (101) in the CPU (10), which is the input means, inputs learning data (201) stored in the storage section (20) such as a hard disk. (Learning data input step: S1)
The learning data (201) is a combination of a plurality of second text data groups, a plurality of predefined topics, and related character string information describing the topics in the pre-associated second text data group. consists of

A topic is an item that classifies a topic of interest. For example, if it is a technical document, it can correspond to headings classified step by step from major to minor, or general main points included in technical documents. may be manually extracted. Information from news sites, newspapers, magazines, etc. is categorized by field such as "international," "economy," and "society." It may be classified into subordinate concepts such as . Alternatively, they may be classified according to themes such as "disparity problem" and "housing problem".

Furthermore, as a feature of the present invention, topics are classified into two or more stages of conceptual information from superordinate concepts to subordinate concepts so that readers can easily understand the content contained in the text data, and this conceptual information will be described later. It is also possible to reflect it in the search conditions for searching for characteristic words.

The second text data to be input is text data to be used for learning, and relates to a document including at least the above topic. good.

The related character string information is a pre-defined related character string describing the topic in the second text data group, and in principle is defined by manually judging the relevance.
For example, if a large amount of past news text data is input as the second text data group, and the topic "International" includes "United States,""Europe," and "China," the articles about "United States" are manually selected. is the related character string information. The related character string information in this case is the entire article.

A feature word extraction unit (102) extracts feature words from related character string information using the input learning data (201). (Characteristic word extraction step: S2)
A feature word is a phrase that characterizes that a topic is described in the text data. If the above-mentioned "United States" is the topic, for example, articles on other topics such as "United States", "New York", and "President Trump" Words and phrases that can be distinguished from

Feature word extraction methods are well known in the field of language processing, and TF-IDF is a technique often used for feature word extraction. TF is word appearance frequency, and IDF is inverse document frequency. On the premise that words with high appearance frequency are important, how many articles in all articles the feature word appears by considering inverse document frequency Introduce a scale to express IDF is the value obtained by adding 1 to the logarithm of the value obtained by dividing the total number of articles by the number of articles in which a certain phrase appears. case becomes smaller. The value obtained by multiplying this IDF by TF is the weight value of the word.

Other methods for extracting feature words include SVM (Tsutomu Hirao, Hideki Isozaki, Eisaku Maeda, Yuji Matsumoto: Extracting Important Sentences Using Support Vector Machine, Transactions of Information Processing Society of Japan. Vol. 44, No. 8, pp. .2230-2243 (2003)), KeyGraph (Internet URL: http://iit.kke.co.jp/keygraph/ searched on September 10, 2018), etc. Various methods are known, and are known in the present invention. method can be used as appropriate.
It is also a well-known technique to morphologically analyze text data and divide it into predetermined morphemes in order to extract characteristic words.

According to the feature word extraction step (S2), the related character string information can be associated with the feature words in the second text data group, so that the related character string information can be associated with the corresponding topic data and feature words. In the present invention, this associated information is called aspect data (202), and is stored in the storage section (20) by the aspect data creation section (103). (Aspect data creation step: S3)

The simplest data as the aspect data (202) is that a characteristic word is defined for each topic, and unlike the prior art, the present invention is characterized by associating aspect data with various topics. . Therefore, as a minimum configuration, only feature words may be defined in the aspect data.
In this embodiment, in addition to the feature words for each topic, the aspect data stores the weight values together with the feature words.

As the weight value, a value related to at least one of the appearance frequency and the co-occurrence frequency calculated when the feature word extraction means extracts the feature word, such as the weight value in TF-IDF described above, can be used.
Alternatively, a weight value can be set by a predetermined arithmetic expression based on the weight information defined for the feature word, and the weight value for the feature word extracted for each topic can be stored in the aspect data.

Furthermore, the aspect data can also store a feature word extracted for each topic and a distance value (referred to as a margin) relating to at least one of the distances before and after the appearance of the feature word. Here, the margin indicates the distance at which it is judged that the feature word and other feature words appear at the same time, and is an important parameter for calculating the degree of association in the present invention. Table 1 shows an example of aspect data including distance values.

The meaning of the above table is that the text whose degree of relevance is to be evaluated and the weight of the characteristic word related to the topic "United States", the weight value of "United States" is 1, which is small, while the other characteristic word, that is, "United States" The margin for judging whether it co-occurred with "New York" and "President Trump" was set wide at 2,500 characters in the front and 2,500 characters in the back. On the other hand, if "President Trump" is included, it is considered more likely to be associated with the United States and is given a weight value of 10. In this case, the front and rear margins are narrowly set to 250 in order to specify the range more.

The definition of the distance value may be defined mechanically in association with the weight value, such as 2500 when the weight value is 1, 1000 when the weight value is 5, etc., or may be defined manually while considering characteristic words. You may

The above processing completes the learning process, and as a result, aspect data (202) is generated.
In the relevance evaluation step, first text data (203) to be evaluated is input from the input processing unit (101) and recorded in the storage unit (20). (Text data input step: S4)

Next, the feature word search unit (104) refers to the aspect data (202) and searches for feature words contained in the first text data (203) for at least one topic. (Characteristic word search step: S5)
Specifically, the feature words stored in the aspect data (202) are searched for each topic, and the degree of relevance is evaluated, for example, according to the number of feature words. A simple method is to output the parts where the feature words appear repeatedly within a narrower range than a certain threshold value as the relevant parts. can also

In the present embodiment, a more complicated process is performed, and a weight value is set for the feature word for each topic. FIG. 3 is an explanatory diagram of a method of calculating the degree of association according to the present invention.
In this drawing, the horizontal axis is from the beginning to the end of the first text data (203), and the left represents the beginning and the right represents the ending character position. The vertical column of each characteristic word indicates the weight value.

For example, the feature word "America" in the first row is located at the 3000th character, and a graph with a weight of 1 is drawn in the range of 500 to 5500 characters as 2500 characters before and after the margin. Similarly, graphs are drawn in the range of 5500 to 10500 characters in the 8000th character and in the range of 11500 to 16500 characters in the 14000th character.
In this embodiment, search results based on the same rule are OR-operated. In other words, the feature word "United States" in the second row appears in the 4500th and 8000th characters. Instead, by taking the OR, the graph with a weight of 1 is obtained from 2000 characters to 10500 characters.

The feature word "New York" appears at the 4000th and 14000th characters, has a front-back width of 2000 characters, and has a weight of 3. "President Trump" appears at the 4750th character, has a front-back width of 500 characters, and has a weight of 5.
Then, if an AND operation is performed on these weight values, a graph such as the degree of relevance shown in the lower row is obtained. This graph visually shows the relevant parts in the text data, where the position of a indicates no relation to the topic "United States", and the positions of b and c indicate that the relation is high.

In the present invention, in order to calculate the degree of relevance, the concept of the margin is created along with the weight value, and calculation is performed using the margin, thereby providing a completely new method of calculating the degree of relevance. As a calculation method, it is preferable to perform an OR operation for the same rule (same feature word) and an AND operation for different rules as described above, but the same feature word may be used for the AND operation, or any other calculation may be performed. A value can also be used.

As an example of the aspect data (202), an example in which weight values and distance values are defined for a feature word consisting of one word or morpheme is shown, but the feature word may be a plurality of word strings. For example, a word string of "New York City, USA" may be used as a feature word.

Also, one rule may be the case where two words co-occur within a predetermined range. For example, rule 1 that word A and word B or word C co-occur within 50 characters is expressed as A(B C)<50. Within 50 characters, for example, if A is fish and B is bird, the space between h of fish and b of bird should be within 50 characters. Assuming that the characteristic degree of this rule 1 is 1, a graph with a weight of 1 is drawn from f of fish to d of bird as shown in FIG. 4(a).

Unlike the above embodiment, FIG. 4B of this embodiment shows an example in which the weight value changes within the range of the front and rear margins. As described above, the weight value is not constant, and may be changed so as to decrease as the distance from the feature word increases.

Then, as shown in FIG. 4C, in the case of the graph of the co-occurrence of A and B defined in rule 1 and the graph of the co-occurrence of A and C, since they are the same rule, OR is performed.
On the other hand, as shown in FIG. 4(d), rule 2 defines X(YZ)<100, and if there is a graph of the co-occurrence of X and Y, a graph of the co-occurrence of A and B and Take the AND of the graph by the co-occurrence of X and Y.
As described above, the characteristic word of the present invention can include a plurality of word strings and a combination of words co-occurring within a predetermined range.

A display processing unit (105) displays a value based on a search result of characteristic words as a degree of relevance by distinguishing topics from the monitor (30). As for the display method, it is preferable to display in a graph form as shown in FIG.
FIG. 5 shows a first screen display example. A topic column (40) is arranged on the left side of the screen, and the user selects a topic to be displayed using a keyboard (31), mouse, or the like. In the figure, the major item "Production Strategy and Base Strategy" and the minor item "Movement of Domestic Return" are selected, and the calculated value of the weight value is displayed in a line graph in the score column (41) on the right. It is The score column (41) has the beginning of the text at the top and the end at the bottom.

Looking at the score column (41), the position of the high score on the horizontal axis indicates the related part with the high degree of relevance, and it is possible to easily grasp the position of the description related to the topic in the entire sentence. can. By specifying the line graph in the score column with a mouse or the like, the description part may be browsed.

Locations exceeding the threshold value in the score column (41) are highlighted in the document list display (42) in the right column. By selecting a topic from the topic column (40), the user can understand at a glance where and how much description related to the topic is described.
The range to be displayed as a marker may be an entire sentence including a word or phrase whose weight value calculation result exceeds a predetermined threshold value, or a predetermined range of sentences in the vicinity of the word or phrase.

FIG. 6 shows a second screen display example. In this embodiment, when a topic is selected from the topic field (40) on the left side, related parts of a plurality of documents can be displayed in the related part display field (43) on the right side. For example, if a document such as a white paper issued every year is designated for a plurality of years and a topic is selected, a portion related to one topic is extracted from each year's white paper. By arranging only the portion exceeding the calculation result of the predetermined weight value as shown in the figure, it is possible to easily compare and contrast descriptions on the same topic in a plurality of documents.

FIG. 7 shows an example of displaying a web site article on the web browser and displaying a bar graph indicating the degree of relevance next to the article. The bar graph corresponds to the position of the line of the article, for example, it is configured so that it is easy to understand in which position in the article the information related to the topic specified by the user by entering it in the search window is described.
Such a display method can be displayed in a small area on the side of the web browser in a visually easy-to-understand manner, and is suitable for application to search service screens, for example.

The display examples according to the present invention have been shown above, but in the relevance output step (S6), it is not always necessary to display the results. It may be configured to The degree of association may be a value obtained by calculating a weight value as described above, or may be information only indicating whether or not there is an association.
The form of output may also be a configuration in which the information on the degree of association is temporarily stored in a memory (not shown) or the like, or transmitted to another computer through the network via the communication section (32).

As another embodiment of the present invention, in the aspect data creation step (S3), the aspect data creation unit (103) can define hierarchical information for each feature word extracted for each topic.
For example, if the topic is "President Trump" as shown in Table 2, "Country" → "Region" → Topics and conceptual information are classified from the higher concept to the lower concept. , the topic is at level 3.

By classifying the feature words with conceptual information that is easy for the reader to understand, it is possible to divide the newspaper article into stages, from a wide range of parts describing "America", to specific regions, and further to parts of desired topics. be able to comprehend.

When such classification is performed, when a desired topic (50) is selected as shown in FIG. A medium density marker display (52) and a portion of the selected topic (50) can be displayed with a dark marker display (53).
As for the display range of the marker, for each feature word belonging to the concept information, the entire sentence including the feature word, the sentences in a predetermined range near the feature word, or the display mode within the margin range of the feature word can be changed. can be done. By displaying the sentences containing each concept from the superordinate concept to the subordinate concept so that they can be distinguished, the reader can visually understand the degree of relevance in the article.

The above hierarchical information can be further used to define weight values and distance values in the aspect data creation step (S3).
That is, the weight value and the initial value of the margin are determined for each layer. In the example of Table 2, layer 1 has a weight value of 1 and a leading and trailing margin of 2500 characters, layer 2 has a weight value of 5 and a leading and trailing margin of 1000 characters, layer 3 has a weight value of 10 and a leading and trailing margin of 250 characters. is.

By defining in this way, after the feature words are extracted, they are classified into conceptual information that is easy for the reader to understand, and at the same time the hierarchy is determined, so the weight values and margins are appropriately set. That is, it is possible to define a combination pattern of the weight value and the distance value in the hierarchical information, and determine the search condition for the feature word according to the pattern in the feature word search step (S5).

As conceptual information, it is also possible to refer to databases such as dictionary data and thesaurus for countries, regions, and topics, and automatically classify low-level concepts from high-level concepts. By matching with these databases after the feature words are extracted, for example, "America", "New York", and "President Trump" are country, region, and proper nouns, so they are classified into broader concept, middle concept, and lower concept. It is possible to classify concepts, define corresponding hierarchical information, and even define weight values and distance values.

As described above, the present invention can pattern the conceptual information that is easy for the reader to understand and the set of values when the computer calculates the degree of relevance. It is possible to obtain results that do not cause any discomfort to the user.

1 text processing device 10 CPU
101 input processing unit 102 feature word extraction unit 103 aspect data creation unit 104 feature word search unit 105 display processing unit 20 storage unit 201 learning data 202 aspect data 203 text data 30 monitor 31 keyboard 32 communication unit

Claims (10)

From the input first text data,
A text processing method by a computer for evaluating the degree of relevance between a related part described in text data and the topic for a topic, which is an item obtained by classifying a topic of interest, comprising:
An input means inputs a plurality of second text data groups, a plurality of predefined topics, and related character string information described about the topics in the second text data group associated in advance as learning data. training data input step,
a feature word extraction step in which the feature word extraction means extracts a feature word from the related character string information;
After the learning step having each step of an aspect data creation step, the aspect data creation means creates aspect data, which is information in which the feature words extracted for each topic are associated, and stores the aspect data in the storage means;
a text data input step in which the input means inputs the first text data;
a feature word search step in which a feature word search means searches for a feature word contained in the first text data for at least one topic by referring to the aspect data;
In a configuration in which the output means performs a relevance evaluation step having each step of an output step in which the relevance is output as the relevance based on the search result of the feature word for each topic,
The aspect data stores a weight value for the feature word extracted for each topic, a feature word extracted for each topic, and a distance value relating to at least one of the distances before and after the appearance of the feature word,
In the feature word extraction step, when the feature word extraction means extracts the feature word from the related character string information, the weight value is set by a predetermined arithmetic expression based on the weight information defined in the feature word. death,
In the feature word search step, a search condition is determined based on the weight value, and a search condition for feature words of the same type or different type included in the first text data is determined based at least on the distance value. ,
The feature word search means searches for a feature word based on the aspect data, reads the distance value of each extracted feature word, and performs a predetermined calculation from the weight value of the feature word within the range of the distance value. A text processing method characterized by outputting a calculated value as the degree of relevance. The weight value defined for the feature word is a value related to at least one of the appearance frequency and the co-occurrence frequency calculated when the feature word extraction means extracts the feature word,
A text processing method according to claim 1. 2. The text processing method according to claim 1, wherein, in said output step, a calculated value relating to the number of a plurality of feature words that are search results for each topic or the sum of their weight values is output as said degree of relevance. In the aspect data creating step, the aspect data creating means
Defining hierarchical information for each of the feature words extracted for each topic,
The hierarchical information defines patterns of combinations of the weight values and the distance values,
4. The text processing method according to any one of claims 1 to 3, wherein in said characteristic word search step, search conditions for characteristic words are determined according to said pattern. In a configuration in which the topic is classified into two or more stages of conceptual information from a superordinate concept to a subordinate concept so that the reader can easily understand the contents included in the text data,
In the aspect data creating step, the aspect data creating means
automatically defining the hierarchical information according to the conceptual information;
5. The text processing method according to claim 4, wherein each of the feature words for each topic is associated with the conceptual information. In the output step, the output means
6. A text processing method according to any one of claims 1 to 5, wherein the degree of association is displayed in a predetermined graph together with the extracted feature words. In the output step, the output means
7. The text processing method according to any one of claims 1 to 6, wherein in the configuration for displaying the first text data, a predetermined graph is displayed according to the position of the line of the extracted characteristic word. In the output step, the output means
In the configuration for displaying the first text data, changing the display mode of the entire sentence including the characteristic word, the sentence of a predetermined range near the characteristic word, or the range of the distance value of the characteristic word. Item 8. A text processing method according to any one of Items 1 to 7. In the output step, the output means
For each feature word belonging to the concept information, the entire sentence including the feature word, the sentence in a predetermined range near the feature word, or the display mode within the range of the distance value of the feature word is changed, and the superordinate concept is changed. to display each sentence containing each concept of subordinate concepts so that they can be distinguished from
9. A text processing method as claimed in claim 8, comprising the arrangement of claim 5 . A computer-based text processing apparatus for evaluating the degree of relevance between a related part described in text data and a topic, which is an item obtained by classifying a topic of interest from input first text data, and the topic. ,
Learning in which a plurality of second text data groups, a plurality of predefined topics, and related character string information described about the topics in the second text data group associated in advance are input as learning data. data entry means for
a feature word extracting means for extracting a feature word from the related character string information;
aspect data creation means for creating aspect data, which is information in which characteristic words extracted for each topic are associated, and storing the aspect data in a storage means;
text data input means for inputting first text data;
feature word search means for searching for feature words included in the first text data for at least one topic by referring to the aspect data;
and an output means for distinguishing between topics and outputting a value based on the search result of the feature word as the degree of relevance,
The aspect data stores a weight value for the feature word extracted for each topic, a feature word extracted for each topic, and a distance value relating to at least one of the distances before and after the appearance of the feature word,
when the feature word extracting means extracts the feature word from the related character string information, a weight value is set by a predetermined arithmetic expression based on the weight information defined for the feature word;
The feature word search means determines a search condition based on the weight value, and
The feature word search means determines a search condition for the same type or different type of feature word included in the first text data based on at least the distance value, and searches for the feature word based on the aspect data; , reading out the distance value of each extracted feature word, and outputting a value calculated by a predetermined operation from the weight value of the feature word within the range of the distance value as the degree of relevance. Device.

Images