JP6303148B2 - Document feature extraction device, document feature extraction method, document classification device, document classification method, document search device, document search method, computer program, and recording medium on which computer program is recorded - Google Patents

Description

  The present invention relates to a document feature extraction apparatus, a document feature extraction method, and a document feature extraction method for extracting features of a target document when classifying or searching documents using machine learning. The present invention relates to a document classification device, a document classification method, a document retrieval device, a document retrieval method, a computer program, and a recording medium on which the computer program is recorded. The “document” as used in the present invention is not limited by its length or its purpose, but includes a simple sentence or a search expression, and is not limited to a document consisting of a character string. Other data, for example, audio, video, image, and other data.

  In machine learning, cases are often expressed using vectors ("feature vectors" or "feature vectors") whose elements are the features of the target document. In that case, the similarity between cases and the distribution of cases are given by the similarity between the feature vectors representing each case and the distribution of the feature vectors.

  For example, in document classification, an example to be classified is an individual document, and 0, 1, or a plurality of document categories in a given set of document categories are assigned to each document. Document classification is performed. At this time, the feature vector corresponding to each document is based on the presence / absence (0/1) of the word included in the document, the frequency of the word in the document, or the word weight (evaluation) calculated from the frequency and the importance of the word. Value) etc. as a vector. One example of a word evaluation value often used in practice is tfidf. tfidf is calculated based on two indices, tf (frequency of appearance of words in the document) and idf (inverse document frequency). There are several methods for calculating tfidf, such as tf-idf-cf, tfc, ltc, atc, ntc, nn, bnn, lnn, nnn, and the like, for example, can be obtained by the following formula.

Here, n _{i, j} is the appearance frequency of the word t _{i in} the document d _j , | D | is the total number of documents, and | {d: d∋t _i } | is the number of documents including the word t _i .
Note that the term “word” is used here, but as described above, the “document” in the present invention is not only a document consisting of a character string but also other data such as audio, video, and the like. "Word" in the present invention is not limited to the meaning of a word in natural language, but has some meaning that is a superordinate concept. It means a collection of data with

The vector elements for the word t _i of the document vector in the document d _j are as follows.

  As an example of a document classification method using tfidf as a word weight, there is a supervised learning method described in Non-Patent Document 1. In document classification by supervised learning, a learning process is performed on a learning device using learning data including a set of documents to which a document category is assigned, and then a document category is assigned to an arbitrary document. In the first step of learning, each document is converted to a set of feature-feature value pairs. In the method described in Non-Patent Document 1, an SVM (Support Vector Machine) is used as a learning device, and “and”, “or”, “the”, etc. among words appearing three or more times in the learning data as features. Words other than the words excluded from the search target, called stop words, are used. As the feature value, an index tfc, which is a kind of tfidf, is used. In Non-Patent Document 1, as a result of comparing classification accuracy of five types of machine learners including SVM in a document classification experiment using a Reuters corpus, SVM shows the highest accuracy, and the KNN method (k-nearest neighbor classifier) Was reported to have the next highest accuracy.

  In addition, it has been reported that SVM and KNN have achieved high accuracy in document classification using machine learning. In addition, in these machine learnings, many reports have been made on a feature selection method for selecting a word to be used as a feature from a huge number of words. However, it has not been fully clarified yet what kind of character string should be used to obtain high accuracy and how to obtain high accuracy by obtaining a feature value. .

  For example, when performing Japanese document classification, it is necessary to first divide a sentence into words before selecting a feature, but there is a feature mismatch between the learning data and the test data due to differences in word extraction There is. As an example, as shown in Table 1, let us consider a case where there is a document including a character string “medical science department” in the learning data and “health medical science department” in the test data. Here, it is assumed that both documents are closely related to each other. If this is the case, the words “medical science department” in the learning data are “medical department” + “faculty”, and “health medical science department” in the test data are “health care” + “science” + “department”. If divided, there will be no common word between the two character strings, and the relevance between the learning data and the test data will be estimated lower than the actual amount.

  In this way, when performing Japanese document classification, the conventional method in which words are extracted by deterministic analysis and the extracted words are used as features (or feature candidates) as they are, the relationship between documents is determined. There was a problem that it could not be grasped sufficiently.

  Long words in document classification, information retrieval, and various other language processing systems, characterized by taking a document as a set of words and inferring the characteristics and identities of the document from the characteristics and identities of the words in the document It can be said that there is a tendency to express the characteristics of the document as well. For example, when trying to measure the closeness of the contents of documents A and B, or the closeness of the contents of search B and the contents of document B in the search target document group, the word “color” is added to document A / search formula A and document B. When the “image processing apparatus” appears in common, the words “color”, “image”, “information”, “processing”, and “apparatus” appear in document A / search formula A and document B, respectively. It can be estimated that the contents of the document A / search formula A and the document B are closer than the case where Therefore, when emphasizing the identification ability as a feature, it can be said that a longer term is extracted more effectively as a feature in these language processing systems.

  However, if the word “color image information processing apparatus” appears only in document A / search formula A and the word “image information processing apparatus” appears only in document B, “color image information processing apparatus” and “image information processing apparatus” "Is cut out as a word, an important common element between document A / search formula A and document B is lost. Therefore, in order to avoid missing, that is, to improve the completeness, “image”, “information”, “processing”, and “apparatus” are extracted as words in each of the document A / retrieval formula A and document B, etc. It can be said that a shorter extraction works more effectively as a feature. In the prior art, there has been a problem that these trade-off relationships cannot be effectively resolved.

  As a conventional method for replacing or reinforcing a word-based feature, a document classification method using a word n-gram as a feature has been proposed. Patent Documents 1 and 2 propose a document classification method that extracts feature information including a word n-gram and makes it a vector feature. However, it does not show what word n-gram is extracted from the document. Also in other conventional techniques, when a word n-gram is used as a feature, it is specified in the form of “word n-gram (n = 1, 2, 3,...)”. However, in the word n-gram in which the words are simply arranged, there is a great difference in the identification ability as a feature. There is also a problem that the number of features increases explosively as the value of n increases. Thus, in order to extract a word n-gram that is effective as a feature, the conventional method of simply limiting the value of n is insufficient.

  Patent Documents 3 and 4 describe that a part of a sentence is combined as an n-gram of a phrase and used as a feature of the sentence by combining adjacent phrases from sentences composed of phrases. 4 also has the same problems as those described in Patent Documents 1 and 2.

JP 2014-0667154 A JP 2014-123286 A JP 2013-171550 A JP, 2015-011426, A Japanese Patent No. 4478042 T. Joachims, Text categorization with support vector machines: learning with many relevant features, in: Proceedings of 10th European Conference on Machine Learning, (ECML), Chemnitz, Germany, pp. 137-142, 1998 Yang, Yiming, and Jan O. Pedersen. "A comparative study on feature selection in text categorization." ICML. Vol. 97. 1997 Information Retrieval (Z39.50): Application Service Definition and Protocol Specification, ANSI / NISO Z39.50-2003, http://www.loc.gov/z3950/agency/Z39-50-2003.pdf

  Provided are a document feature extraction device and a document feature extraction method for extracting features of a target document, which can obtain high accuracy when classifying or searching documents using machine learning. In addition, a document classification device, a document classification method, a document search device, and a document search method using such features are also provided. Here, the “document” includes not only a document consisting of a character string but also data composed of other data, for example, audio, video, image, and other data.

  If importance is attached to the discrimination ability as a feature described in the prior art, it is better to extract words longer. If there is no miss, that is, if you want to improve completeness, extract words shorter. The following two methods are proposed as means for adjusting the trade-off relationship that the one that has worked effectively as a feature.

  First, it is a technique of using a word string in a specific area as a feature. For example, consider the case where the word “color image information processing apparatus” appears only in document A / search formula A, and the word “image information processing apparatus” appears only in document B. First, the words “color”, “image”, “information”, “processing”, “device”, and the like are extracted from the document A, and the word strings “color image information processing device” “image information processing device” “information processing” are concatenated. "Device", "Image", "Information", "Processing", "Device", etc. are obtained. Next, the words “image”, “information”, “processing”, “device”, and the like are extracted from the document B, and the word strings “image information processing device” “information processing device” “image” “information” “ Get processing, equipment, etc. Here, when the obtained word strings are compared with each other, the same word strings exist, and by detecting them, it is determined whether or not the document A / search formula A and the document B are similar.

  Note that the “word string” here refers to the word string obtained by concatenating the words “color”, “image”, “information”, “processing”, and “device” extracted from the document A in the above example. As included in “information”, “processing”, and “apparatus”, not only a concatenation of two or more adjacent words but also a word before the concatenation is included.

  Furthermore, in addition to this, a method is also proposed in which a plurality of word strings exist together in a specific region, that is, a co-occurrence is used. For example, when the document A is a patent document and the technical field of the technology described in the patent document is to be known, a word string “the present invention” and “natural language processing” are included in a sentence of the document A. In the case where both word strings exist, compared to the case where the word string “this invention” exists in one sentence of document A and the word string “natural language processing” exists in another sentence of document A, The possibility that the technical field of the technology described in the document A is natural language processing is higher.

  Here, the description has been made on the assumption that a plurality of word strings exist together, that is, co-occurrence. However, for example, the word string α is a secondary word string, and the word string α and the word string If β co-occurs, it means one meaning, and if the word string α is present but the word string β does not exist, the other meaning is usually considered. In order to include such a case, the meaning of the word “co-occurrence” is not a simple co-occurrence, but one of the word string α and the word string β, which has the word string α but does not have the word string β. Think of it as an extension to exist only. An example of this “co-occurrence” will be described in detail in Example 3.

  By using such a method, both discrimination ability and completeness can be achieved, and as a result, high classification accuracy and search accuracy can be achieved.

3 is a flowchart of the first embodiment. In Example 1, it is a figure which shows how the character string used as the object of a process is processed. It is a figure which shows the production | generation process of the word string performed by step S104 of Example 1. FIG. It is a figure which shows the parsing tree produced in Example 2. FIG. 10 is a flowchart of Example 2. 10 is a flowchart of Example 3. It is a figure which shows the co-occurrence process of the word string performed by step S605 of Example 3. FIG. It is a block diagram of the document classification device of Example 5. 10 is a flowchart when document classification is performed in the document classification device according to the fifth exemplary embodiment. 10 is a flowchart for creating feature statistics used in the document classification apparatus according to the fifth embodiment. 10 is a flowchart for creating a classification model used in the document classification apparatus according to the fifth embodiment. It is a block diagram of the document search apparatus of Example 6. FIG. 16 is a flowchart for performing a document search in the document search apparatus according to the sixth embodiment. 14 is a flowchart for creating feature statistics used in the document search apparatus according to the sixth embodiment. 15 is a flowchart for processing a search expression in the document search device according to the sixth embodiment.

  Embodiments for carrying out the present invention will be described in detail with reference to the drawings.

What is described as Example 1 is a technique of using a word string in a specific area as a feature.
FIG. 1 is a flowchart of the first embodiment. FIG. 2 is a diagram illustrating how a character string to be processed is processed in the first embodiment.

First, in the first embodiment, in a reading step S101 in FIG. 1, a document that is a feature extraction target is read.
Next, in the morpheme analysis step S102, the read document is subjected to morpheme analysis using knowledge of the target language grammar and a dictionary, and divided into words. Note that the unit of division in this morpheme analysis step S102 is not necessarily limited to words, and other units do not necessarily have a problem, so the word here means to the extent that it can also be called “first unit” This means that the morphological analysis step S102 can also be called the first boundary division step.

In the boundary detection step S103, the phrase boundary is extracted from the analyzed document.
Here, this extracted phrase boundary is for use when creating a word string in the next step S104, and by shortening the length of the target range for creating the word string as much as possible, In order to prevent the number of combinations of word strings extracted in the next word string extraction step S104 from becoming very large, what is extracted in the boundary detection step S103 is a word that crosses the boundary. It may be a boundary that does not exist, and can be replaced by another boundary such as a sentence boundary or a punctuation boundary instead of a phrase boundary. Furthermore, depending on the language to be processed, a boundary other than a clause or sentence may be used as long as the word does not have a word that crosses the boundary.

  In addition, considering that the process of the first embodiment is applied to the entire document to extract features, the boundary extracted in the boundary detection step S103 at a certain part of the document is incorrect, and Even if the result of the word string extraction step S104 performed around the boundary is not necessarily accurate due to the division of the compound word / compound word, the important description in the document is not included in the document. Since it usually appears more than once, and a single mistake does not necessarily have a large weight as viewed from the whole document, the boundary detection step S103 satisfies the premise that there is no word that crosses the boundary. On the other hand, even if the boundary is delimited by a certain number of characters, it does not necessarily have a big impact on the final result when sorting and searching. No.

  That is, the phrase here means that it can be called a unit longer than a word, that is, a “second unit longer than the first unit”, and has a meaning that the boundary detection step S103 can also be called a second boundary division step. It is.

  In rare cases, there may be a semantic relationship between the words before and after the phrase boundary. Therefore, if the second unit is longer than the phrase, higher classification accuracy and search accuracy may be achieved. .

The next word string extraction step S104 is a step for performing the most important processing in the first embodiment, and extracts a word string from a specific area in the character string.
Referring to FIG. 2, the word string extraction step S104 shows how to process a document whose feature is to be extracted and a character string included in the search expression. In addition, using FIG. 2, as a premise of the word string extraction step S104, the morpheme analysis step S102 and the boundary detection step S103 that perform the process before the word string extraction step S104 are also described. To do.

First, in a reading step S101, a character string “improve ground performance in the rolling control device of the ground work machine” is read as a character string to be processed.
Next, in morpheme analysis step S102, morpheme analysis is performed, and the input character string is divided into words. In the morpheme analysis in the morpheme analysis step S102 of the first embodiment, since the processing of the word string extraction step S104 is performed later, a combined word / compound word such as “color image information processing apparatus” or “ground work machine” is collected. There is no need to divide words as words. As shown in FIG. 2, “ground work machine” is “to” “ground” “work” “machine”, and “rolling control device” is “rolling” “control” “device” As there is no problem even if it is divided finely, the knowledge of the target language grammar and the dictionary required for this morphological analysis should be simplified, and the processing time required for morphological analysis should be shortened. Can do.

Then, in the boundary detection step S103, the phrase boundary is extracted. In FIG. 2, the extracted phrase boundary is indicated by “/”.
In the word string extraction step S104, the region surrounded by the phrase boundary extracted in the boundary detection step S103 is processed, and within that region, the adjacent words among the words obtained by the morpheme analysis in the morpheme analysis step S102. What is to be connected is extracted as a word string, and the whole word is output as a word string together with the word before connection. For example, as shown in FIG. 2, “ground work machine” is extracted as a phrase, and “pair”, “ground”, “work”, “machine”, “no” are extracted as words from these, so that they are adjacent to each other. As a concatenation of words, “ground” “ground work” “ground work machine” “ground work machine” “ground work” “ground work machine” “ground work machine” “work machine” “work machine” “ A word string “machine” is generated, and the words “pair”, “ground”, “work”, “machine”, and “no” before connection are all output as a word string. FIG. 3 schematically shows this connection process. FIG. 3 shows a word string that is output when the word string generation processing shown in the word string extraction step S104 is performed on the phrase when there is a phrase composed of four words in the order of αβγδ. ing.

Since such a word string generation process is performed, almost all of the words that may be included in the target document are detected although they are in a duplicated form.
As already described, the word here means to the extent that it can also be called a “first unit”, and the morphological analysis step S102 may be divided into units other than words. The extraction step S104 also has a meaning that can be called a column extraction step.

  In addition, in order to perform such connection processing by providing a word string extraction step S104, in the morpheme analysis in the morpheme analysis step S102, compound words / compound words such as “color image information processing apparatus” and “ground work machine” are used. It is not necessary to divide the words as a single word, simplify the knowledge of the target language grammar and the dictionary required for the morphological analysis in the morphological analysis step S102, and further reduce the processing time required for the morphological analysis. Thus, the processing of the first embodiment as a whole can be shortened.

In the feature output step S105, the word string extracted in the word string extraction step S104 as shown in FIG. 2 is output as the document feature.
By using the word string generated in the first embodiment as a document feature and performing search and classification, both discrimination ability and completeness can be achieved, and as a result, high classification accuracy and search accuracy can be achieved.

Note that Patent Document 5 describes a method of extracting words in a specific area in a form that allows duplication, and it can be said that it is similar to the first embodiment when only the output is observed.
However, the morpheme lattice 80 generated by the morpheme lattice generation unit 10 described in Patent Document 5 is “East”, “Kyo”, “Miyako”, as described in Paragraph 0045 of FIGS. Morphological analysis is performed on the assumption that “Tokyo”, “Kyoto”, etc. all exist as words in the dictionary for morphological analysis. On the other hand, in the morphological analysis performed in the morphological analysis step S102 of the first embodiment, only “East”, “Kyo” and “Miya” exist as words in the dictionary for morphological analysis, and “Tokyo” and “Kyoto” are words. Even if it does not exist in the dictionary for morphological analysis, “East”, “Kyo”, “Tokyo”, “Tokyo”, “Kyoto”, “Tokyo”, etc. are obtained as word strings, and finally obtained Even if things can be said to be similar, it can be said that the way to find them is different.

The second embodiment is an example in which the first embodiment is applied to English. In the second embodiment, in addition to simply applying the first embodiment to English, processing adapted to English grammar is added.
4 and 5 are diagrams illustrating English text processing according to the second embodiment.

FIG. 4 shows a parse tree obtained by parsing an input sentence. In FIG. 4, NP ^[2] , V ^[4] , NP ^[5] , NP ^[7] , V ^[9] , NP ^[10] , P ^[12] , NP ^[13] "(Congress)", "(sent)", and "(president)" next to these indicate the head that is the main part of the phrase.

FIG. 5 is a flowchart corresponding to FIG. 1 of the first embodiment.
As shown in FIG. 4, the second embodiment will be described with reference to an example of inputting a sentence “Congress sent the president an $ 18.4 billion fiscal 1990 Treasury and Postal Service bill providing $ 5.5 billion for the Internal Revenue Service”.

  First, the input document is parsed and divided into units of words, phrases, and sentences, and a parse tree is created. This process is performed in reading step S501, syntax analysis step S502, and boundary detection step S503. As in Example 1, the word means “first unit”, the phrase means “second unit longer than the first unit”, and the sentence means “first unit”. This means that it can be called a “third unit longer than either the unit or the second unit”. Further, the boundary detection step S503 has a meaning that can be called a second boundary division step.

  As described for the boundary detection step S103 of the first embodiment, depending on the target language, the boundary extracted in the boundary detection step S503 may be a boundary other than a phrase or a sentence. While satisfying the premise that there is no word that crosses the boundary, it does not necessarily delimit the boundary for every fixed number of characters, nor does it necessarily have a great influence on the final result when performing classification or search.

  At the same time, all words in the sentence are extracted by the parsing step S502. In the example of FIG. 4, "Congress", "sent", "the", "president", "an", "$", "18.4", "billion", "fiscal", "1990", "Treasury", "and", "Postal", "Service", "bill", "providing", "$", "5.5", "billion", "for", "the", "Internal", "Revenue", "Service "Will be extracted.

  Then, in the phrase processing step S504, based on the outputs of the syntax analysis step S502 and the boundary detection step S503, the noun phrase (NP), verb phrase (VP), adjective phrase (AP), adverb phrase (AdvP), Classify into preposition phrases (PP), conjunction phrases (ConjP), quantifier phrases (QP), etc., and extract the head of each phrase. The two types of processing performed in the phrase processing step S504 are not necessarily essential as will be described below. In addition, the processing from step S502 to step S504 may be performed in one step by using a syntax analyzer based on context-free grammar with head information.

Next, in the main character string extraction step S505, among the extracted main characters (head), the leaf node of the parse tree, that is, the one given to the internal node having a node having only one word as a child node, FIG. Is given to NP ^[2] , V ^[4] , NP ^[5] , NP ^[7] , V ^[9] , NP ^[10] , P ^[12] , and NP ^[13]. The heads are concatenated and extracted as a main character string such as “Congress sent president bill providing billion for Service”. Depending on the grammatical theory, in the preposition phrase (PP) consisting of the preposition (P) and the noun phrase (NP), the main word of the preposition phrase (PP) is the preposition (P). The noun phrase (NP) was treated as the main word.

Further, in the word string extraction step S506, a word string is generated as follows for each phrase of the input sentence and the main word string generated in the main word string extraction step S505.
First, for each noun phrase (NP) in the input sentence, adjacent words in the noun phrase (NP) are sequentially connected to generate a word string. For example, from the noun phrase (NP) "the Internal Revenue Service" to "the", "Internal", "Revenue", "Service", "the Internal", "the Internal Revenue", "the Internal Revenue Service" , "Internal Revenue", "Internal Revenue Service", "Revenue Service". In this example, adjacent words in a noun phrase (NP) are concatenated sequentially to generate a word string.Of course, regardless of the type, adjacent words are concatenated sequentially to form a word string. May be generated. In that case, the classification of phrases in the phrase processing step S504 described above is not necessary.

  Then, adjacent words in the main word string are sequentially connected to generate a word string. That is, from the main remarks such as “Congress sent president bill providing billion for Service”, “Congress”, “Congress sent”, “Congress sent president”, “Congress sent president bill”, “Congress sent president bill providing”, “ Congress sent president bill providing billion "," Congress sent president bill providing billion for "," Congress sent president bill providing billion for Service "," sent "," sent president "," sent president bill "," sent president bill providing " , "sent president bill providing billion", "sent president bill providing billion for", "sent president bill providing billion for Service", "president", "president bill", "president bill providing" ... Generate a column.

  As a result, from one sentence, "Congress sent the president an $ 18.4 billion fiscal 1990 Treasury and Postal Service bill providing $ 5.5 billion for the Internal Revenue Service", from the word string and main word string generated from the noun phrase (NP) Since two word strings, the generated word strings, are generated, duplication is eliminated as necessary, and the word strings for the sentence are collectively collected.

  Of course, the generation of a word string from this main word string may be omitted, and in that case, the main word (head) extraction performed in the phrase processing step S504, the main word string extraction step S505, and the word string extraction step S506 are performed. It is not necessary to generate a word string from the main word string and eliminate duplicates.

Finally, in the feature output step S507, the word string extracted in the word string extraction step S506 is output as the document feature.
In addition, although Example 2 demonstrated English as an example, of course, it is also possible to apply to languages other than English including Japanese.

  What is described as the third embodiment is that, with respect to the word string extracted as a feature, as described in the first and second embodiments, a plurality of word strings exist mainly in a specific area, that is, This is a technique that uses co-occurrence.

  Note that the term “co-occurrence” is used here, but a typical example assumed in the third embodiment is that a plurality of word strings exist together, that is, co-occurrence. However, the present invention is not necessarily limited thereto. In the following description of the third embodiment, this co-occurrence will be mainly described. As described above, for example, the word string α is a secondary word string, and the word string α and the word string β are In the case of co-occurrence, there is a case in which one meaning is present, but if the word string α is present but the word string β does not exist, the other meaning is usually considered. The meaning is not considered to be simple co-occurrence but is expanded to include the word string α but the word string β does not exist, or only one of the word α and the word β exists.

  In addition, “co-occurrence” here is a kind of proximity calculation, and the application of proximity calculation to a simple word is described in, for example, “3.7.2 Proximity” on pages 97 to 98 of Non-Patent Document 3. As is well known. However, in Example 3, a synergistic effect can be obtained by applying “co-occurrence” to a word string instead of a word.

FIG. 6 is a flowchart of the third embodiment.
First, in the third embodiment, in step S601 of FIG. 6, a document that is a target of feature extraction is read. Next, in step S602, the read document is read using grammar knowledge or a dictionary of the target language. Perform morphological analysis as a character string and divide into words. This reading step S601 is the same as the reading step S101 of the first embodiment. As in the first embodiment, the unit of division in the morphological analysis step S602 is not necessarily limited to words, and other units are not necessarily problematic, so the word here is also referred to as “first unit”. The morpheme analysis step S602 also means the first boundary division step.

  Next, in a boundary detection step S603, a phrase boundary and a sentence boundary are extracted from the analyzed document. The extracted phrase boundary has a meaning that can be called a “second unit longer than the first unit” as in the first embodiment, and is used in the word string extraction step S604. In contrast, the extracted sentence boundary is used as a range for examining the co-occurrence relationship in the word string co-occurrence detection step S605. Similar to the description of the sentence boundary in the boundary detection step S503 of the second embodiment, this sentence boundary is also a unit longer than a word or phrase, that is, “a third unit longer than either the first unit or the second unit”. It can be called as well. That is, this means that the boundary detection step S603 can also be called a second boundary division step.

  As described in the boundary detection step S103 of the first embodiment and the boundary detection step S503 of the second embodiment, the boundary extracted in the boundary detection step S603 may be a boundary other than a clause or sentence depending on the target language. Moreover, even if the boundary detection step S603 satisfies the premise that there is no word having a shape that crosses the boundary, the final result when performing classification or search even if the boundary is delimited by a certain number of characters, It does not necessarily have a big influence.

The word string extraction step S604 corresponds to the word string extraction step S104 of the first embodiment and performs the same operation, so that the description thereof is omitted.
The word string co-occurrence detection step S605 is the step of performing the most important processing in the third embodiment, and detects that a plurality of word strings co-occur in a specific area in the character string. As already described, the word here means to the extent that it can also be called a “first unit”, and the morpheme analysis step S602 may be divided into units other than words. The co-occurrence detection step also means that it can be called a column co-occurrence detection step.

  Referring to FIG. 7, for the co-occurrence to be detected by the word string co-occurrence detection step S605, as described above, only one of the word string α and the word string β is present, although the word string α is present but the word string β is not present. Will be explained including the extension to the existence of

  As an example of “co-occurrence”, FIG. 7A shows that when a sentence to be processed is composed of three word strings “αβγ”, “word string“ α ”and word string“ β ”are both Exists. When the existence of the word string “α” and the word string “β” has a logical product (AND) relationship, the word string “α” exists but the word string “β” does not exist. When the presence of the word string “α” and the absence of the word string “β” have a logical product (AND) relationship, either or both of the word string “α” and the word string “β” exist. To do. ”And the word string“ β ”have a logical sum (OR) relationship, only one of the word string“ α ”and the word string“ β ”exists. ”And the word string“ β ”have an exclusive OR (ExOR) relationship. The output of the word string co-occurrence detection step S605 only outputs whether or not there is a combination of word strings to be detected, and the finally obtained value is either “0” or “1”. It has become. Therefore, as shown in FIG. 7B, even in the case where the sentence to be processed includes the same word string a plurality of times such as “αβγαβ”, “word string“ α ”and word string“ β ” Both exist. "Will output" 1 ". The reason for processing in this way is that “αβγαβ” is a sentence to be processed, and “a word string“ α ”and a word string“ β ”exist together. ”As the first word string“ α ”, the second word string“ β ”and the fifth word string“ β ”are both appropriate This is because it is difficult to determine whether both are appropriate as co-occurrence word strings, but on the other hand, it correctly reflects the fact that multiple co-occurrence occurred with the same word string combination in one sentence. There is also a problem that it becomes impossible. Therefore, in this case, “2” in which the number of closer combinations is detected and “4” in which the number of all combinations is detected may be output.

  In the actual processing of the word string co-occurrence detection step S605 in the case of outputting either “0” or “1”, the word string to be detected, for example, the word string “α” and the word string “β” First, it is detected whether or not it exists in a specific area in the character string to be processed, for example, `` αβγ '', and the detection result is logical product (AND), logical sum (OR), negated (NOT), a logical operation (Boolean operation) represented by exclusive OR (ExOR) is performed.

  The “co-occurrence” detected in the word string co-occurrence detection step S605 does not occur only between the two word strings, but “word string“ α ”, word string“ β ”, word string shown in FIG. Both “γ” exist. It also includes “co-occurrence” between three or more words, as typified by.

  Finally, in the feature output step S606, the word string extracted in the word string extraction step S604 and the “co-occurrence” detected in the word string co-occurrence detection step S605 are output together as the document feature. In this way, handling the co-occurrence of the word string itself and the word string as a feature is hereinafter referred to as “composite feature”.

  The presence of a plurality of word strings close to each other can be considered to express a more limited meaning than the presence of a single word string. Therefore, the word string co-occurrence detection step S605 is performed. Using the presence or absence of co-occurrence of the word string to be output as a feature can be said to be more accurate than using the presence or absence of a simple word string as a feature. Then, by using the presence or absence of word string co-occurrence as shown in FIG. 7 generated in Example 3 as a feature and performing a search and classification, both discrimination ability and completeness can be achieved, resulting in high classification accuracy. And search accuracy can be achieved.

  In the first and second embodiments, the word string as shown in FIG. 2 is used as a feature, and in the third embodiment, the word string co-occurrence as shown in FIG. 7 is used as a feature. As described above, it was possible to achieve both classification ability and completeness, and as a result, high classification accuracy and search accuracy could be achieved. Based on this, in the fourth embodiment, as described in the third embodiment, when the word sequence is used as a feature and the co-occurrence of the word sequence is used as a feature to make a “composite feature”, how tfidf An example of whether to use will be described.

Expressions (1) to (4) of the background art show an example of how to obtain tfidf for a word and the definition of a document vector based on the example. As described in the first and second embodiments, the word string is a concatenation of words. Therefore, when tfidf is used for the word string, each extracted word string is directly used in the equations (1) to (4). It can be treated as the word t _i .

On the other hand, the part using the word string co-occurrence shown in the third embodiment as a feature is as follows.
First, a word string α and a word string β are considered as word strings to be co-occurred. Basically, the concept of a word string pair P in which the word string α and the word string β are combined is considered, and the co-occurrence of the word string pair P is expressed by the word t in the expressions (1) to (4) as in the word string. _Treat as _i . However, since there are a plurality of types of co-occurrence for the word string α and the word string β as shown in FIG. 7, the actual number of co-occurrence of the word string pair P is as follows. The number of combinations of the word string α and the word string β to be detected for co-occurrence is multiplied by the number of co-occurrence types to be detected.

  As described in the third embodiment, “co-occurrence” does not occur only between two word strings, but “word string“ α ”, word string“ β ”, word string“ γ ”shown in FIG. "Exists together. The word sequence pair P also includes not only a combination of two word sequences but also three word sequences, because “co-occurrence” between three or more words as represented by It is a concept including combinations of four word strings and the like.

Eventually, when the word string as shown in the first and second embodiments and the word string co-occurrence described in the third embodiment are used together as features, the word string co-occurrence of the word string and the word string pair P Each is treated as the word t _i in the equations (1) to (4), tfidf is obtained, and finally a document vector is obtained.

As a fifth embodiment, an example of a document classification apparatus that uses a document vector obtained from the complex feature shown in the fourth embodiment will be described with reference to FIGS.
FIG. 8 is a block diagram of the document classification apparatus according to the fifth embodiment.

The input document receiving unit 801 is a part that receives an input of a document to be classified.
In the document analysis unit 802 and the appearance frequency generation unit 803, the generation of the word string and the co-occurrence processing of the word string described in the first and second embodiments are performed on the document to be classified to obtain a complex feature.

The document vector generation unit 804 obtains a document vector as described in the fourth embodiment for the document to be classified, and uses idf stored in the feature statistics storage unit 811.
The feature statistic generation document storage unit 807 stores a document for obtaining the idf.

  The document analysis unit 808 and the appearance frequency generation unit 809 generate the word strings and share the word strings described in the first and second embodiments with respect to the document for obtaining the idf stored in the feature statistics generation document storage unit 807. Perform complex processing to obtain complex features.

  The feature statistic calculation unit 810 calculates the total number of documents | D | stored in the feature statistic generation document storage unit 807 and idf from the documents stored in the feature statistic generation document storage unit 807, and stores them in the feature statistic storage unit 811. Remember.

The learning document storage unit 812 stores a learning document for generating a rule for classifying the document.
The document analysis unit 813 and the appearance frequency generation unit 814 are the word strings described in the first and second embodiments for the learning document for generating the rules for classifying the document stored in the learning document storage unit 812. The document vector generation unit 815 obtains the document vector as described in the fourth embodiment for the learning document by performing the generation of the word sequence and the co-occurrence processing of the word string, and the feature statistics storage The idf stored in the unit 811 is used.

  The learning data generation unit 816, the machine learning unit 817, the learning result storage unit 818, and the classification model unit 819 create and store a classification model by processing the learning document using machine learning. As an algorithm used for the machine learning, an algorithm such as SVM or CRF (Conditional Random Fields) that can be used without assuming the independence of the features is preferable, but other learning algorithms may be used.

  The category ranking unit 805 and the category assigning unit 806 classify the document vector for the document to be classified, which is obtained by the document vector generation unit 804, into categories using the classification model stored in the classification model unit 819. Output.

  FIG. 9 is a flowchart for performing document classification in the document classification apparatus according to the fifth embodiment. The input document reception unit 801, the document analysis unit 802, the appearance frequency generation unit 803, and the document vector generation unit illustrated in FIG. In step 804, the category ranking unit 805 and the category assigning unit 806 are executed.

  A morpheme analysis step S902, a boundary detection step S903, a word string extraction step S904, a word string co-occurrence detection step S905, and a feature output step S906 are performed on the classification target document read in the classification target document reading step S901 according to the first embodiment. 2, the details of which are the reading step S <b> 601, the morpheme analysis step S <b> 602, the boundary detection step S <b> 603, the word string extraction step S <b> 604, the word string co-occurrence detection step S <b> 605, and the features described in the third embodiment. Since this is the same as the output step S606, the description is omitted.

  In the appearance frequency generation step S907, the feature statistics reading step S908, and the document vector generation step S909, the idf read from the feature statistics storage unit 811 in the feature statistics reading step S908 is used to obtain the tfidf value as described in the fourth embodiment, Based on this, a document vector of a document to be classified is generated.

The classification model is read in the document model reading step S910, and the document to be classified is classified in the document classification execution step S911 based on the classification model and the document vector.
FIG. 10 illustrates feature statistics stored in the feature statistics storage unit 811 of the document classification apparatus according to the fifth embodiment. The feature statistics generation document storage unit 807, the document analysis unit 808, the appearance frequency generation unit 809, and the feature statistics calculation unit 810 It is a flowchart which shows how it calculates | requires.

  The morphological analysis step S1002, the boundary detection step S1003, the word string extraction step S1004, the word string co-occurrence detection step S1005, and the feature output step S1006 are feature statistics read from the feature statistics generation document storage unit 807 in the feature statistics document reading step S1001. Is a part that performs the processing described in the first and second embodiments, and details thereof are a reading step S601, a morpheme analyzing step S602, a boundary detecting step S603, a word string described in the third embodiment. Since this is the same as extraction step S604, word string co-occurrence detection step S605, and feature output step S606, description thereof will be omitted.

  The feature statistics generation step S1007 is a step for obtaining idf as described in the fourth embodiment, and the obtained idf is written in the feature statistics storage unit 811 in the feature statistics write step S1008.

  FIG. 11 is a flowchart for creating a classification model used in the document classification apparatus according to the fifth embodiment. The learning document storage unit 812, the document analysis unit 813, the appearance frequency generation unit 814, and the document shown in FIG. It is executed by the vector generation unit 815, the learning data generation unit 816, the machine learning unit 817, the learning result storage unit 818, and the classification model unit 819.

  The morpheme analysis step S1102, the boundary detection step S1103, the word string extraction step S1104, the word string co-occurrence detection step S1105, and the feature output step S1106 are read from the learning document storage unit 812 into the learning document read step S1101. This is a part for performing the processing described in the first and second embodiments, and details thereof are the reading step S601, the morphological analysis step S602, the boundary detection step S603, the word string extraction step S604, and the word string described in the third embodiment. Since this is the same as the origin detection step S605 and the feature output step S606, the description thereof is omitted.

  In appearance frequency generation step S1107, feature statistics reading step S1108, and document vector generation step S1109, similar to appearance frequency generation step S907, feature statistics reading step S908, and document vector generation step S909, feature statistics reading step from feature statistics storage unit 811 Using the idf read in S1108, the tfidf value is obtained as described in the fourth embodiment, and a document vector of the learning document is generated based on the tfidf value.

  In machine learning step S1110 and classification model creation step S1111, a classification model is created by performing machine learning using the generated document vector of the learning document.

  In the document classification device according to the fifth embodiment, by using the complex feature based on the co-occurrence of the word strings shown in the first and second embodiments and the word strings shown in the third embodiment, both the identification ability and the completeness can be achieved. As a result, high classification accuracy can be achieved.

As a sixth embodiment, an example of a document search apparatus that uses a document vector obtained from the complex feature shown in the fourth embodiment will be described with reference to FIGS.
FIG. 12 is a block diagram of the document search apparatus according to the sixth embodiment.

The search target document storage unit 1201 stores a search target document.
In the document analysis unit 1202 and the appearance frequency generation unit 1203, a complex feature is obtained by performing the word string generation and the word string co-occurrence processing described in the first and second embodiments for the search target document. The document vector generation unit 1204 obtains a document vector as described in the fourth embodiment for the document to be searched, and uses idf stored in the feature statistics storage unit 1211.

  A feature statistics generation document storage unit 1207, a document analysis unit 1208, an appearance frequency generation unit 1209, and a feature statistics calculation unit 1210 are the feature statistics generation document storage unit 807, document analysis unit 808, and appearance frequency generation unit 809 of the fifth embodiment. , Which is the same as the feature statistic calculation unit 810, and the total number of documents | D | stored in the feature statistic generation document storage unit 1207, idf from among the documents stored in the feature statistic generation document storage unit 1207 The process of obtaining and storing in the feature statistics storage unit 1211 is performed.

In the search expression input unit 1212, a search expression for searching is input. In this case, a natural sentence is assumed as a search expression.
The search expression analysis unit 1213 and the appearance frequency generation unit 1214 perform the word string generation and word string co-occurrence processing described in the first and second embodiments for the search expression input in the search expression input unit 1212. The search vector generation unit 1215 obtains a document vector as described in the fourth embodiment as a search vector for the search formula based on those results, and uses idf stored in the feature statistics storage unit 1211.

  In the search unit 1205 and the search result output unit 1206, a similar search target document is obtained by comparing the document vector for the search target document obtained by the document vector generation unit 1204 with the search vector obtained by the search vector generation unit 1215. It ranks and outputs.

  FIG. 13 is a flowchart for performing a document search in the document search apparatus according to the sixth embodiment. The search target document storage unit 1201, the document analysis unit 1202, the appearance frequency generation unit 1203, and the document vector generation shown in FIG. This is executed by the unit 1204, the search unit 1205, and the search result output unit 1206.

  The morpheme analysis step S1302, the boundary detection step S1303, the word string extraction step S1304, the word string co-occurrence detection step S1305, and the feature output step S1306 are performed on the search target document read in the search target document reading step S1301 according to the first embodiment. 2, the details of which are the reading step S <b> 601, the morpheme analysis step S <b> 602, the boundary detection step S <b> 603, the word string extraction step S <b> 604, the word string co-occurrence detection step S <b> 605, and the features described in the third embodiment. Since this is the same as the output step S606, the description is omitted.

  In the appearance frequency generation step S1307, the feature statistics reading step S1308, and the document vector generation step S1309, using the idf read from the feature statistics storage unit 1211 in the feature statistics reading step S908, the tfidf value is obtained as described in the fourth embodiment, Based on this, a document vector of a search target document is generated.

  In the search vector reading step S1310, the search vector is read, and based on the search vector and the document vector, in the search execution step S1311, a document that matches or is similar to the search vector is selected from the search target documents, and ranked. Output.

  FIG. 14 illustrates feature statistics stored in the feature statistics storage unit 1211 of the document search apparatus according to the sixth embodiment. The feature statistics generation document storage unit 1207, the document analysis unit 1208, the appearance frequency generation unit 1209, and the feature statistics calculation unit 1210 Although it is a flowchart which shows how it calculates | requires, feature statistical document reading step S1401, morpheme analysis step S1402, boundary detection step S1403, word sequence extraction step S1404, word sequence co-occurrence detection step S1405, feature output step S1406, feature The statistical generation step S1407 and the feature statistical writing step S1408 are the feature statistical document reading step S1001, the morpheme analyzing step S1002, the boundary detecting step S1003, the word string extracting step S1004, and the word string co-occurrence shown in FIG. Detection step S10 5, feature output step S1006, feature statistics generation step S1007, since and performs exactly the same processing as the feature statistics writing step S1008 description thereof will be omitted.

  FIG. 15 is a flowchart showing how the document search apparatus according to the sixth embodiment analyzes a search expression and obtains a search vector. The search expression input unit 1212 and the search expression analysis unit 1213 shown in FIG. This is executed by the appearance frequency generation unit 1214 and the search vector generation unit 1215.

  The morphological analysis step S1502, the boundary detection step S1503, the word string extraction step S1504, the feature output step S1506, and the word string co-occurrence detection step S1505 are described in the first and second embodiments with respect to the search expression read in the search expression reading step S1501. The details are read out step S601, morpheme analysis step S602, boundary detection step S603, word string extraction step S604, word string co-occurrence detection step S605, and feature output step S606 described in the third embodiment. Since this is the same as that described above, the description thereof is omitted.

  In appearance frequency generation step S1507, feature statistics reading step S1508, and search vector generation step S1509, similar to appearance frequency generation step S1307, feature statistics reading step S1308, and document vector generation step S1309, feature statistics reading step from the feature statistics storage unit 1211 Using the idf read in S1508, the tfidf value is obtained as described in the fourth embodiment, and a search vector is generated based on the tfidf value.

  In the document search apparatus according to the sixth embodiment, by using the complex feature based on the co-occurrence of the word strings shown in the first and second embodiments and the word strings shown in the third embodiment, it is possible to achieve both discrimination ability and completeness. As a result, high search accuracy can be achieved.

  The data feature extraction device, data classification device, data classification method, and data search device according to the embodiments of the present invention have been described with some examples. The above-described embodiments are merely examples for facilitating understanding of the present invention, and are not intended to limit the present invention. Although these embodiments have been described with reference to a document made up of character strings, as already mentioned, the “document” in the present invention is not only a document made up of character strings but also other data such as audio. The word “word” in the present invention is not limited to the meaning of the word in the natural language but has a certain meaning that is a superordinate concept thereof. Since this means a collection of data, it is obvious that the present invention can be applied to a document composed of data other than character strings with a necessary portion modified in accordance with the characteristics of the data. For example, as a boundary to be extracted in the first and second embodiments, a boundary of a phoneme or a syllable is selected instead of a phrase or a sentence boundary, and a phoneme or a syllable is selected instead of a word. It is self-evident that a document consisting of speech can be targeted instead of a document consisting of columns. In addition, image area separation and scene change detection are used as algorithms for extracting boundaries in Embodiments 1 and 2, and image areas and scenes that are separated instead of words are used. It is obvious that the present invention can be applied to a document composed of a still image or a moving image instead of a document composed of a character string.

  In addition, the present invention can be changed and improved without departing from the gist thereof, and it is needless to say that the present invention includes equivalents thereof.

S101 Reading step S102 Morphological analysis step S103 Boundary detection step S104 Word string extraction step S105 Feature output step S501 Reading step S502 Syntax analysis step S503 Boundary detection step S504 Phrase processing step S505 Main dictionary extraction step S506 Word string extraction step S507 Feature output step S60 Feature output step S60 Reading step S602 Morphological analysis step S603 Boundary detection step S604 Word string extraction step S605 Word string co-occurrence detection step S606 Feature output step 801 Input document reception unit 802 Document analysis unit 803 Appearance frequency generation unit 804 Document vector generation unit 805 Category ranking unit 806 Category assigning unit 807 Feature statistics generation document storage unit 808 Document analysis unit 809 Appearance frequency Unit 810 feature statistical calculation unit 811 feature statistical storage unit 812 learning document storage unit 813 document analysis unit 814 appearance frequency generation unit 815 document vector generation unit 816 learning data generation unit 817 machine learning unit 818 learning result storage unit 819 classification model unit S901 Classification target document reading step S902 Morphological analysis step S903 Boundary detection step S904 Word string extraction step S905 Word string co-occurrence detection step S906 Feature output step S907 Appearance frequency generation step S908 Feature statistics reading step S909 Document vector generation step S910 Document model reading step S911 Document classification execution step S1001 Feature statistical document reading step S1002 Morphological analysis step S1003 Boundary detection step S1004 Word string extraction step S10 5 Word Sequence Co-occurrence Detection Step S1006 Feature Output Step S1007 Feature Statistics Generation Step S1008 Feature Statistics Write Step S1101 Learning Document Reading Step S1102 Morphological Analysis Step S1103 Boundary Detection Step S1104 Word Sequence Extraction Step S1105 Word Sequence Co-occurrence Detection Step S1106 Feature Output step S1107 Appearance frequency generation step S1108 Feature statistics reading step S1109 Document vector generation step S1110 Machine learning step S1111 Classification model creation step 1201 Search target document storage unit 1202 Document analysis unit 1203 Appearance frequency generation unit 1204 Document vector generation unit 1205 Search unit 1206 Search result output unit 1207 Feature statistics generation document storage unit 1208 Document analysis unit 1209 Appearance frequency generation unit 121 Feature statistics calculation unit 1211 Feature statistics storage unit 1212 Search formula input unit 1213 Search formula analysis unit 1214 Appearance frequency generation unit 1215 Search vector generation unit S1301 Search target document reading step S1302 Morphological analysis step S1303 Boundary detection step S1304 Word string extraction step S1305 Word Sequence co-occurrence detection step S1306 Feature output step S1307 Appearance frequency generation step S1308 Feature statistics read step S1309 Document vector generation step S1310 Search vector read step S1311 Search execution step S1401 Feature statistics document read step S1402 Morphological analysis step S1403 Boundary detection step S1404 Word string Extraction step S1405 Word string co-occurrence detection step S1406 Feature output step S1407 Feature group Generation step S1408 Feature statistical writing step S1501 Retrieval expression reading step S1502 Morphological analysis step S1503 Boundary detection step S1504 Word sequence extraction step S1505 Word sequence co-occurrence detection step S1506 Feature output step S1507 Appearance frequency generation step S1508 Feature statistics reading step S1509 Search vector Generation step

Claims (30)

An input means for inputting a document;
First boundary dividing means for analyzing and dividing the document into first units;
A second boundary dividing unit that divides a document divided into first units by the first boundary dividing unit into second units longer than the first unit;
Within each range of the second units divided by the second boundary dividing means, adjacent first units are connected to form a column, and column extracting means for extracting the first feature together with the first unit before connection. When,
Feature output means for outputting the first feature as a document feature;
A document feature extraction apparatus characterized by comprising:   The column extraction means determines whether or not to connect adjacent first units within each range for each second unit divided by the second boundary dividing means according to the type of each second unit. The document feature extraction apparatus according to claim 1, wherein the document feature extraction apparatus determines the document feature.   The second boundary dividing unit adds the document divided for each first unit by the first boundary dividing unit for each second unit longer than the first unit, and adds a third unit longer than the second unit. 3. The document feature extraction apparatus according to claim 1, wherein the document feature extraction unit is also divided into units. Extracting the main first unit within each range for each of the second units, and connecting the extracted main first units within the range of the third unit and outputting them as a main word sequence Having means,
The column extraction means further generates a second column by concatenating adjacent first units included in the main word column, and combines the first column and the second column as a first feature. Generating,
The document feature extraction apparatus according to claim 3, wherein:   5. The document feature extraction apparatus according to claim 4, wherein the column extraction unit generates a first feature by eliminating an overlap between the first column and the second column. 6. Detecting whether or not there is a first feature having a different characteristic with respect to the first feature output by the column extracting unit within the respective ranges of the third units divided by the second boundary dividing unit. Column co-occurrence detecting means for performing a logical operation on whether or not the first features having different features are within the same third unit range, and outputting a result of the logical operation as a second feature. Have
The feature output means outputs the first feature and the second feature together as a document feature;
The document feature extraction apparatus according to any one of claims 3 to 5.   The logical operation according to claim 6, wherein the logical operation includes at least one of logical product (AND), logical sum (OR), negation (NOT), and exclusive logical sum (ExOR). Document feature extraction device.   The document feature extraction apparatus according to claim 1, wherein the document is a document composed of a character string described in a natural language.   The document feature extraction apparatus according to claim 8, wherein the first unit is a word.   The document feature extraction apparatus according to claim 8 or 9, wherein the second unit is divided into phrases, phrases, or a certain number of characters.   The document feature extraction apparatus according to claim 1, wherein the document is a document including audio, video, or an image.   The document feature extraction apparatus according to claim 1, wherein a tfidf value is used as the feature value. Document classification apparatus, which comprises using the identity of the document extracted by the document feature extraction apparatus according to any one of claims 1 to 12. Document retrieval system, which comprises using the identity of the document extracted by the document feature extraction apparatus according to any one of claims 1 to 12. An input step for entering the document;
A first boundary splitting step that parses the document and splits it into first units;
A second boundary dividing step of dividing the document divided into first units by the first boundary dividing step into second units longer than the first unit;
Column extraction step of extracting adjacent first units as a column within the respective ranges of the second units divided by the second boundary division step and extracting the first feature together with the first unit before connection. When,
A feature output step of outputting the first feature as a document feature;
A document feature extraction method characterized by comprising: In the column extraction step, whether or not to connect adjacent first units within the respective ranges of the second units divided by the second boundary dividing step depends on the type of each of the second units. The document feature extraction method according to claim 15 , wherein the document feature extraction is performed.   The second boundary dividing step adds the document divided for each first unit by the first boundary dividing step for each second unit longer than the first unit, and adds a third longer than the second unit. The document feature extraction method according to claim 15 or 16, wherein the unit is also divided into units. Extracting the main first unit within each range for each of the second units, and connecting the extracted main first units within the range of the third unit and outputting them as a main word sequence Has steps,
The column extraction step further generates a second column by concatenating adjacent first units included in the main word column, and combines the first column and the second column as a first feature. Generating,
The document feature extraction method according to claim 17, wherein:   19. The document feature extraction method according to claim 18, wherein the column extraction step generates the first feature by eliminating the overlap between the first column and the second column. Detecting whether or not there is a first feature having a different characteristic with respect to the first feature output by the column extraction step within the respective ranges of the third units divided by the second boundary division step. A column co-occurrence detecting step of performing a logical operation on whether or not the first features having different features are within the same third unit range and outputting a result of the logical operation as a second feature. Have
The feature output step outputs the first feature and the second feature together as a document feature;
The document feature extraction method according to any one of claims 17 to 19.   21. The logical operation according to claim 20, wherein the logical operation includes at least one of logical product (AND), logical sum (OR), negation (NOT), and exclusive logical sum (ExOR). Document feature extraction method.   The document feature extraction method according to any one of claims 15 to 21, wherein the document is a document composed of a character string described in a natural language. The document feature extraction method according to claim 22 , wherein the first unit is a word.   24. The document feature extraction method according to claim 22 or 23, wherein the second unit is a section, a phrase, or a predetermined number of characters.   The document feature extraction method according to any one of claims 15 to 21, wherein the document is a document including audio, video, or an image.   The document feature extraction method according to any one of claims 15 to 25, wherein a tfidf value is used as the feature value. 27. A document classification method using the document features extracted by the document feature extraction method according to any one of claims 15 to 26. 27. A document search method using a document feature extracted by the document feature extraction method according to any one of claims 15 to 26.   A program for executing the method according to any one of claims 15 to 28.   A recording medium on which a program for executing the method according to any one of claims 15 to 28 is recorded.