JP2007304950A - Document processing device and document processing method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve information extraction accuracy from a document file. <P>SOLUTION: A word of a learning corpus 200 is classified into one of a plurality of classes. The document processing device 100 holds the identity of the word in the learning corpus 200 as class identity information in each class in a class identity holding section 170. The document processing device 100 extracts a word from a document 210 to be inspected before processing, calculates the identity of the word and adaptation of the class identity information in the document 210 to be inspected before processing for each of the plurality of classes, adjusts the adaptation calculated for a predetermined class, and specifies the class corresponding to the word extracted based on the adaptation for each class. The specified class name is added as a tag, thereby creating a document 212 to be inspected after processing. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to an information extraction technique from a document file, and more particularly to a specific expression extraction technique.

With the spread of computers and the development of network technology, the exchange of electronic information via networks has become popular. As a result, many of the business processes that have been conventionally performed on a paper basis are being replaced by network-based processes. Advances in digitalization and network technology have drastically reduced information acquisition costs. In this situation, information extraction technology that extracts central information from individual documents, especially named entities such as names, company names, dates, and job titles from documents. Attention has been focused on the technique for extracting named expressions.
JP 2006-048536 A

  Of various terms included in a document file, a term corresponding to a so-called unique expression is often important information that characterizes the content of a document. However, in the current situation where new specific expressions are constantly created, it is not easy to determine which words correspond to specific expressions and which words do not correspond to specific expressions.

  An object of the present invention is to provide a technique for improving the accuracy of extracting information from a document, in particular, the accuracy of extracting a specific expression.

One embodiment of the present invention is a document processing apparatus.
This device classifies words into a plurality of classes, and holds the word features of a predetermined corpus as class feature information for each class. This apparatus extracts a word from a document to be examined, calculates the feature of the word in the document to be examined and the suitability of class feature information for each of a plurality of classes, and calculates the suitability calculated for a predetermined class. After the adjustment, the class corresponding to the extracted word is specified based on the fitness for each class.

  The feature here is information indicating an appearance mode of a word in a document. For example, it can be said that a word with “san” in the back or a term with “Mr.” or “dear” in front is likely to be a word indicating a person's name. By comparing the feature of the word extracted from the document to be examined with the tendency of the feature of the word belonging to a certain class (category), it is specified to which class the extracted word corresponds. At this time, the accuracy of information extraction is improved by adjusting the fitness of a predetermined class, for example, the class to which the most words belong in the corpus.

  It should be noted that any combination of the above-described constituent elements and a conversion of the expression of the present invention between a method, a system, a recording medium, a computer program, etc. are also effective as an aspect of the present invention.

  According to the present invention, it is possible to improve information extraction accuracy from a document.

FIG. 1 is a schematic diagram for explaining an outline of processing by the document processing apparatus 100.
The processing of the document processing apparatus 100 can be divided into two stages of “learning processing” and “inspection processing”. The outline of each processing process for each stage is as follows. The “document file” in the present embodiment will be described as a structured document file structured by a predetermined tag set such as XML (eXtensible Markup Language).

1. Learning Process The learning process is a process from generating class feature information from the learning corpus 200 to registering it in the class feature holding unit 170. The learning corpus 200 in this embodiment is a collection of a large number of document files.
A “class” is set for a huge number of words included in the learning corpus 200. A class is a concept indicating a category of words. For example, a class “place name” is set for the word “Shibuya” in a document file in the learning corpus 200. Such setting is usually performed manually. More specifically, a tag <place name> is set in a name space “class” for the word “Shibuya” in a predetermined document file in the learning corpus 200. Classes may be set as attributes instead of tags. In addition to place names, there are several types of classes for classifying words that correspond to specific expressions such as personal names, dates, organization names, job titles, and numerical expressions. The document processing apparatus 100 extracts words from such a learning corpus 200.

  In the learning corpus 200, some words are explicitly designated by some class, and some words are not explicitly designated by the class as not corresponding to the specific expression. Usually, the latter is overwhelmingly larger than the former. Words that do not correspond to such a specific expression are classified as “N class”. On the other hand, a class explicitly specified for a word corresponding to a specific expression is called a “unique class”. In this embodiment, to simplify the explanation, only three classes of “person name”, “place name”, and “date / time” are assumed as unique classes, and all other words are classified into N classes. . In other words, all the words included in the learning corpus 200 are classified into any one of four classes, ie, three unique classes of “person name”, “place name”, and “date and time” and one N class. Become.

  Next, the document processing apparatus 100 extracts “feature” for each word collected from the learning corpus 200. The feature indicates an appearance mode of a word in a document file. For example, it is assumed that the word “Shibuya” is used in the context of “Meeting at Shibuya Station” in the learning corpus 200. Further, it is assumed that words classified into another “place name” class “Daikanyama” are used in the learning corpus 200 in the context of “Accident at Daikanyama Station”. The point common to the two words belonging to the “place name” class is that the word “station” appears behind. In other words, when the word “station” appears in the document file, it is assumed that the word located immediately before is likely to be the “place name” class. As another example, when the word “san” is added behind “Mr. Shibuya is ...”, there is a high possibility that the class is “person name”. Thus, the features for each class are registered in the class feature holding unit 170 as class feature information from the words, classes, and features of the learning corpus 200. For example, in the “place name” class, the probability of “station” following 30% of the corresponding words, and in the “person name” class, the probability of “station” following the corresponding word is 1%. Each time, the feature of the word to which it belongs is registered as class feature information.

2. Inspection Process The inspection process is a process for specifying a class for words in an inspection target document for which “class” is not explicitly specified. The “inspection document” is a document file that should newly specify a class of words in the document by the inspection processing of the document processing apparatus 100.

  First, the document processing apparatus 100 acquires a pre-processing inspection target document 210 that is an inspection target document for which “class” is not specified. The document processing apparatus 100 sequentially extracts words from the pre-processing inspection target document 210. Further, the feature of each word in the pre-processing inspection target document 210 is also detected.

  Next, the document processing apparatus 100 compares the word feature of the pre-processing inspection target document 210 with the class feature information in the class feature holding unit 170. For example, for the word A in the pre-processing inspection target document 210, the feature of the word A is compared with the class feature information of the “person name” class, and the similarity is calculated as the “fitness”. For word A, in addition, four types of fitness are calculated: fitness for the “place name” class, fitness for the “date and time” class, and fitness for the “N” class. A specific method for calculating the fitness will be described later.

The document processing apparatus 100 classifies the class having the highest fitness among the four types of fitness as the class of the word A. Here, it is assumed that the word A is classified into the “place name” class. In that case, the document processing apparatus 100 gives a <place name> tag to the word A of the inspection target document 210 before processing. On the other hand, if the word A is “N” class, no tag is given. In this way, the document processing apparatus 100 identifies classes from the features and the class feature information for all the words included in the pre-processing inspection target document 210. The document processing apparatus 100 outputs the processed inspection target document 212. The processed inspection target document 212 is a document file to which a tag indicating a unique class (hereinafter referred to as “unique tag”) is assigned to the pre-processing inspection target document 210. In the figure, a document file indicated by diagonal lines is a document file to which a unique tag is assigned.
As an example of using the processed inspection target document 212, information required from the pre-processing inspection target document 210 may be searched using the unique tag of the processed inspection target document 212 as a search key. For example, a document structure including a “place name” class, a “date and time” class, and a “person name” class is used as a search pattern, such as a sentence “Meeting with Mr. Kato at 16:00 in Shibuya”. Among the information included in the, information of the pattern “when, where, with whom” can be searched efficiently.

  The above processing process is also a common process in a typical named entity extraction process using the Bayes algorithm. However, in the case of the processing process as described above, even if the word group in the pre-processing inspection target document 210 is a word that does not actually correspond to a specific expression (hereinafter referred to as a “non-unique word”), document processing is performed. In many cases, a unique class is identified by the device 100 as a word corresponding to a unique expression (hereinafter referred to as a “unique word”). In this embodiment, a problem in such a unique expression extraction algorithm is solved by adding a new process to the above process. A specific method will be described in detail with reference to FIG. 5 and subsequent drawings. First, the configuration of the document processing apparatus 100 will be described first.

FIG. 2 is a functional block diagram of the document processing apparatus 100.
Each block shown here can be realized in hardware by an element such as a CPU of a computer or a mechanical device, and in software it is realized by a computer program or the like. Draw functional blocks. Therefore, those skilled in the art will understand that these functional blocks can be realized in various forms by a combination of hardware and software.

The document processing apparatus 100 includes a user interface processing unit 110, a data processing unit 120, and a data holding unit 130.
The user interface processing unit 110 is in charge of processing related to the entire user interface such as input processing from the user and information display for the user. In the present embodiment, the user interface processing unit 110 will be described as providing the user interface service of the document processing apparatus 100. As another example, the user may operate the document processing apparatus 100 via the Internet. In this case, a communication unit (not shown) receives operation instruction information from the user terminal, and transmits processing result information executed based on the operation instruction to the user terminal.

  The data processing unit 120 executes various data processing based on the data acquired from the user interface processing unit 110. The data processing unit 120 also serves as an interface between the user interface processing unit 110 and the data holding unit 130. The data holding unit 130 stores various data such as setting data prepared in advance and data received from the data processing unit 120. The data holding unit 130 includes a class feature holding unit 170 for holding class feature information.

  The user interface processing unit 110 includes an input unit 112 and a display unit 114. The input unit 112 receives an input operation from the user. The display unit 114 displays various information to the user. The input unit 112 includes a document acquisition unit 116 for acquiring a document file such as an inspection target document.

  The data processing unit 120 includes an inspection unit 122 and an adjustment coefficient specifying unit 134. The inspection unit 122 inspects the document file and identifies a class for words in the document file. The inspection unit 122 includes a word extraction unit 124, a fitness level calculation unit 126, a fitness level adjustment unit 128, and a class classification unit 132. The word extraction unit 124 extracts words from the document file and detects their features. The goodness-of-fit calculation unit 126 calculates the similarity between the class feature information of the class feature holding unit 170 and the extracted feature of the word based on an index of goodness of fit. The calculation method of the fitness may be arbitrarily set, but as an example, it may be calculated by the following calculation method.

  Suppose that the word immediately after the word A in the document to be inspected is “station” (hereinafter, the word to be inspected for classification as the word A is “inspection word”). I will call it). According to the class feature information obtained from the learning corpus 200, it is assumed that the word “station” appears immediately after 200 words (20%) among the 1000 words belonging to the place name class. Moreover, 20 words (1%) out of 2000 words belonging to the personal name class, 10 words (2%) out of 500 words belonging to the date class, and 1500 words (15%) out of 10000 words belonging to the N class. Suppose there was. At this time, the fitness of the word A with respect to each of the place name class, personal name class, date / time class, and N class is 20 points, 1 point, 2 points, and 15 points. In this case, it is assumed that the inspection target word A is highly likely to be a word belonging to the “place name” class based on its features. The feature is not limited to the word immediately after the inspection target word, but may be various words appearing around the inspection target word in the document file. As the definition of the feature, a known parameter in the machine learning field may be used. However, in this embodiment, to simplify the explanation, the word immediately after the inspection target word is treated as the feature of the inspection target word.

  The class classification unit 132 specifies a class of words in the document file based on the fitness. In the case of the above example, the class of the word A is a place name class having the highest matching score of 20 points. The fitness level adjustment unit 128 adjusts the fitness level calculated by the fitness level calculation unit 126. The adjustment processing by the matching level adjustment unit 128 solves the above-described problem that “there is often a case where a non-unique word is actually identified as a unique word by the document processing apparatus 100”. Using the adjustment coefficient T specified by the adjustment coefficient specifying unit 134, the fitness level adjustment unit 128 adjusts the fitness level by multiplying the fitness level of the inspection target word with respect to the N class by (2-T). Further details will be described later with reference to FIG.

  The adjustment coefficient specifying unit 134 specifies the adjustment coefficient T. The adjustment coefficient specifying unit 134 includes a temporary optimal coefficient specifying unit 136. The provisional optimum coefficient specifying unit 136 specifies the provisional optimum coefficient S necessary for obtaining the adjustment coefficient T. The provisional optimum coefficient will be described in detail with reference to FIG.

FIG. 3 is a flowchart showing the inspection process when the adjustment process is not executed. Here, the process in the case where a specific word included in a specific inspection target document is set as the inspection target word is shown. Therefore, the flowchart shown in the figure is sequentially executed for each word in the document to be examined.
First, the word extraction unit 124 sequentially extracts inspection target words from the inspection target document acquired by the document acquisition unit 116 (S10). The word extraction unit 124 detects the feature of the inspection target word (S12). Here, a word appearing immediately after the inspection target word is detected as a feature.

  The fitness level calculation unit 126 first selects a personal name class from the four types of classes (S14). The goodness-of-fit calculation unit 126 calculates the goodness of the test target word with respect to the personal name class based on the class feature information about the personal name class of the class feature holding unit 170 and the feature of the test target word (S16). The processes of S14 and S16 are executed for each class other than the personal name class. When the calculation of the fitness is not completed for all classes (N in S18), the process returns to S14 and the next class is selected. When completed (Y in S18), the class classification unit 132 specifies the class of the inspection target word based on the four types of matching degrees (S20). If the identified class is not N class (N in S22), in other words, if it is any unique class, the class classification unit 132 inserts a corresponding unique tag for the inspection target word (S24). If it is N class (Y in S22), the processing is ended as it is. By executing such processing for all the words in the inspection target document, the class of each word in the inspection target document is specified.

In order to index the accuracy of such inspection processing, three indexes called precision, recall, and F value are used. In this embodiment, these indexes are used to measure the degree of matching between “actual unique words” and “words identified as unique words”. It is assumed that N ₁ words are included in the document to be examined, and N _{2 of} them are specific expression words. Then, it is assumed that the document processing apparatus 100 detects N ₃ words as unique words from the N ₁ words in the inspection target document. At this time, if N ₄ = N ₂ ∩N ₃ , that is, N ₄ is the number of unique words correctly detected by the document processing apparatus 100 among the actual unique words, each index is
Precision V = N ₄ / N ₃
Reproducibility W = N ₄ / N ₂
F value = 2 · V · W / (V + W)
As required. Such a formula for calculating the F value is a harmonic average of the precision V and the recall W.

  The relevance rate V is the ratio of words that are actually unique words among the words that are identified as unique words by the class classification unit 132, and is an index that indicates how misdetection has occurred. The recall rate W is a ratio of words detected as unique words by the class classification unit 132 among actual unique words, and is an index indicating how many unique words are detected. It is desirable that both the precision V and the recall W are close to 1.0, but usually they are in a trade-off relationship. The F value is an index value used to evaluate the entire system based on both of these indices. The larger the F value, the better the unique expression detection performance as a system.

FIG. 4 is a graph showing the precision and recall when the adjustment process is not executed.
The horizontal axis indicates the data size of the learning corpus 200. The vertical axis shows the precision and recall values. The graph shown in the figure shows the results of experiments conducted by the inventor on a predetermined learning corpus 200. As can be seen from this graph, as the size of the learning corpus 200 is larger, in other words, as the class feature information is enriched, both the recall rate and the matching rate gradually increase. However, the increase in the size of the learning corpus 200 increases the work of designating classes for words in the learning corpus 200. Needless to say, the increase in the data amount of the class feature storage unit 170 puts pressure on the memory of the document processing apparatus 100. According to this graph, it can be seen that even if the size of the learning corpus 200 is 10 times or 100 times, the recall rate and the matching rate are not greatly improved. In particular, there is a problem that the precision is low. Low precision means a lot of false detections. That is, many non-unique words identified as unique words are generated. In order to make it difficult for the inventor to be identified as a unique word, in other words, the inventor reduces the adjustment coefficient T in order to facilitate identification as a non-unique word, so that the matching degree of the word to be examined with respect to the N class is increased. It was thought that the F value could be improved by adjusting.

FIG. 5 is a graph showing the precision and the recall when the adjustment coefficient T is changed.
The horizontal axis represents the adjustment coefficient T, and the vertical axis represents the precision and recall values. Hereinafter, as the “degree of conformity with respect to the N class”, the degree of conformity before adjustment is referred to as the degree of conformance X before adjustment, and the degree of conformity after adjustment is referred to as the degree of conformance Y after adjustment.
Post-adjustment fitness Y = adjustment fitness X × (2−adjustment coefficient T)
It is. Therefore, the adjustment coefficient T = 1.0 corresponds to the case where there is no adjustment. Also, the smaller the adjustment coefficient, the greater the post-adjustment fitness Y. That is, the inspection target word is easily classified into N classes.

When the adjustment coefficient T is reduced, the reproduction rate is lowered, but the matching rate is increased. As the adjustment coefficient T becomes smaller, the adjusted fitness Y for the N class increases. Since the number N _{3 of} words classified as unique words is narrowed down, the precision (V = N ₄ / N ₃ ) increases. On the other hand, as the number N ₃ of words classified as unique words decreases, the number N _{4 of} unique words correctly detected by the document processing apparatus 100 among the actual unique words also decreases. N ₄ / N ₂ ) will decrease. The reverse is true when the adjustment coefficient T is increased. Therefore, by setting the value of the adjustment coefficient T so that the F value is maximized, the classification accuracy of the detection target word by the document processing apparatus 100 can be improved.
Next, an algorithm for automatically obtaining such an optimal adjustment coefficient T based on the learning corpus 200 will be described. The document processing apparatus 100 according to the present exemplary embodiment specifies the adjustment coefficient T at the learning process stage.

FIG. 6 is a flowchart showing a process of calculating the fitness based on a plurality of types of adjustment coefficients in the learning process.
In order to specify the optimum adjustment coefficient T, the learning corpus 200 is divided into K groups (K is an integer of 2 or more). For example, if 1000 document files are included in the learning corpus 200 and are divided into 10 groups, 100 documents may be grouped. Of the K groups, the i-th group (1 ≦ i ≦ K) is referred to as a test group, and the other groups are referred to as corpus groups. For example, when the first group is set as the inspection group, a group in which the second to Kth groups are collected becomes a corpus group. The figure shows the processing executed for a specific inspection target word A in the i-th group when the i-th group of the K groups is the inspection group. Therefore, the process shown in FIG. 9 is also executed for other inspection target words included in the i-th group. The same applies to groups other than the i-th group.

  First, the word extraction unit 124 extracts the inspection target word A from the i-th group that is the inspection group (S30). The word extraction unit 124 detects the feature of the inspection target word A (S32). The fitness level calculation unit 126 selects a class from four types of classes (S34). The goodness-of-fit calculation unit 126 applies class feature information obtained by limiting to a corpus group composed of groups other than the i-th group, that is, the remaining corpus excluding the i-th group from the learning corpus 200 and features of the inspection target word A. Based on this, the degree of suitability of the inspection target word A for the class selected in S34 is calculated (S36). If the calculation of the fitness is not completed for all classes (N in S38), the process returns to S34 and the next class is selected. When it is completed (Y in S38), the fitness adjustment unit 128 sets 1.0 to the provisional coefficient R (S40). The provisional coefficient is a provisional coefficient that is variable within a predetermined range, for example, a range of 0.5 to 1.0, before determining the value of the adjustment coefficient.

  If the provisional coefficient R is larger than 0.5 (Y in S42), the fitness adjustment unit 128 adjusts the fitness of the inspection target word A with respect to the N class by the provisional coefficient R (S44). The adjustment method is as described above. The class classification unit 132 specifies the class of the inspection target word A based on the four types of matching degrees (S46). The fitness adjustment unit 128 subtracts a predetermined value, for example, 0.01 from the provisional coefficient (S48). If the provisional coefficient R is larger than 0.5, the processes from S44 to S48 are re-executed by the new provisional coefficient R. When the provisional coefficient R is 0.5 or less (N in S42), the process for the inspection target word A ends. By such processing, the class of one inspection target word A in the i-th group is specified based on various provisional coefficients. For example, the inspection target word A may be specified as a personal class when the provisional coefficient is 0.8, and may be specified as the N class when the provisional coefficient is 0.6. All groups are sequentially selected as inspection groups, and all words included in each group are sequentially selected as inspection target words. As a result, for all words in the learning corpus 200, the class in each case of the provisional coefficients 0.5 to 1.0 is specified.

FIG. 7 is a flowchart showing a process for obtaining the provisional optimum coefficient S after the process shown in FIG. 6 is completed.
Here, the process for specifying the provisional coefficient as the provisional optimum coefficient S that is optimum for the i-th group of the K groups, in other words, the F value that is the highest is shown. The process shown in the figure is also executed for each group other than the i-th group. As a result, K provisional optimum coefficients S are obtained for K groups.

  First, the adjustment coefficient specifying unit 134 sets 1.0 to the provisional coefficient R (S50). If the provisional coefficient R is greater than 0.5 (Y in S52), the provisional optimum coefficient specifying unit 136 calculates the relevance ratio for the words included in the i-th group (S54). That is, the ratio of words that are actually unique words among the words specified as unique words in the word group included in the i-th group is calculated as the relevance rate. Next, the provisional optimum coefficient specifying unit 136 calculates a recall rate for words included in the i-th group (S56), and calculates an F value based on the matching rate and the recall rate (S58). The provisional optimum coefficient specifying unit 136 subtracts 0.01 from the provisional coefficient R (S60). When the provisional coefficient R is 0.5 or less (N in S52), the provisional optimum coefficient specifying unit 136 determines the provisional coefficient R when the F value is maximum among the F values for each provisional coefficient as the provisional optimum coefficient S. (S62). Thus, for the i-th group in the learning corpus 200, the provisional optimum coefficient S, which is an adjustment coefficient when the F value is maximized, is specified.

  Similarly, K temporary optimal coefficients are obtained for each of the K groups. The adjustment coefficient calculation unit 138 sets the average value of the K provisional optimum coefficients as the adjustment coefficient T. According to such a processing method, a suitable adjustment coefficient corresponding to the learning corpus 200 can be found by obtaining a suitable F value using the learning corpus 200 in which the correct class for the word has already been set. it can. The adjustment coefficient T is not limited to the average value, and may be obtained by a predetermined arithmetic expression using K temporary optimum coefficients as variables. For example, of the K temporary optimum coefficients S, the temporary optimum coefficient when the F value is the highest may be used as the adjustment coefficient T as it is. Assuming that K = 2, the provisional optimum coefficient S in the first group is 0.6 and the F value is 0.8, and the provisional optimum coefficient S in the second group is 0.65 and the F value is 0. When it is .75, the adjustment coefficient T is 0.6. This is because the F value of 0.8 is larger than the other F value of 0.75.

  In the present embodiment, the provisional coefficient when the F value is maximized is the provisional optimum coefficient. However, the degree of coincidence between the original class of the words included in the examination group and the identified class is not limited to the F value. The provisional coefficient when the degree of coincidence becomes the maximum may be used as the provisional optimum coefficient S based on an arbitrary index indicating the above. For example, the provisional optimal coefficient S may be the provisional coefficient when the number of words that match the original class and the identified class among the words in the examination group is maximized. As an example, the provisional coefficient when the number of words actually designated as the person name class among the words designated as the person name class in the examination group may be the provisional optimum coefficient.

  In the present embodiment, the learning corpus 200 is classified into K groups, and each group is sequentially set as a check group, and K provisional optimum coefficients S are obtained. However, the learning corpus 200 is divided into two. Thus, one temporary optimum coefficient S may be obtained with one inspection group and the other as a corpus group. In this case, the obtained provisional optimum coefficient S may be used as an adjustment coefficient as it is. You may test | inspect by all the combinations of each group which divided | segmented the learning corpus 200, but the test | inspection by all the combinations is not an essential condition of this invention. For example, when the learning corpus 200 is divided into three groups, group A, group B, and group C, (inspection group, corpus group) has three types (A, B + C), (B, A + C), and (C, A + B). The combination of is considered. However, only two combinations of these three combinations may be inspected.

FIG. 8 is a flowchart showing the inspection process in the case where the adjustment process is executed. Also here, the processing contents in which a word included in a certain inspection target document is an inspection target word are shown. Therefore, the flowchart shown in the figure is executed for each word in the document to be examined.
The processing from S70 to S78 is the same as the processing from S10 to S18 in FIG. The fitness level adjustment unit 128 adjusts the fitness level for the N class using the adjustment coefficient determined by the adjustment coefficient specifying unit 134 (S80). The class classification unit 132 identifies the class of the inspection target word with reference to the adjusted fitness (S82). The processing content after S84 is the same as the processing content after S22 in FIG. That is, in the inspection process, a process of multiplying the fitness for the N class by (2−adjustment coefficient T) is actually added.

FIG. 9 is a graph showing the verification result of the adjustment coefficient specifying algorithm shown in this embodiment.
When the inventor executed a learning process for the predetermined learning corpus 200, the adjustment coefficient specifying unit 134 was calculated as an adjustment coefficient T = 0.912. After this, for the document to be inspected whose class to be specified in advance is known, when the adjustment factor T is changed to various values, the precision, recall, and F value are actually calculated. A graph as shown was formed. Actually, the F value is a maximum value in the vicinity of the adjustment coefficient T = 0.900. Although there is a slight deviation from the adjustment coefficient T = 0.912 specified by the adjustment coefficient specifying unit 134, the optimum adjustment coefficient T has been successfully obtained from the data of the learning corpus 200 with considerably high accuracy. In particular, it was confirmed that the F value can be remarkably improved as compared with the case where the adjustment process is not performed, that is, when the adjustment coefficient T = 1.0.

The present invention has been described above based on the present embodiment.
According to the adjustment process by the document processing apparatus 100, the extraction accuracy can be improved in the information extraction from the document file based on the Bayes algorithm. Specifically, the extraction accuracy of unique words from the inspection target document can be increased. In this embodiment, the main purpose is to improve the extraction accuracy of specific words by roughly classifying them into specific classes and N classes. However, a predetermined class A and other classes are arbitrarily set, and an F value is set. By obtaining, the extraction accuracy of words corresponding to class A can be increased. Not only the N class, but the class in which the most words are classified in the learning corpus 200 among a plurality of classes may be the adjustment target. However, in the case of the normal learning corpus 200, it is considered that the class into which the most words are classified is often the N class.

  Further, the algorithm for automatically specifying the adjustment coefficient T based on the learning corpus 200 has been described. For this reason, the document processing apparatus 100 can automatically obtain the adjustment coefficient that maximizes the extraction accuracy at the stage of the learning process.

  By generating the processed inspection target document 212 from the pre-processing inspection target document 210, it is easy to detect what class words are used in an arbitrary inspection target document. For example, a sentence that includes words in the order of personal name class, place name class, place name class, such as "I traveled from Kyoto to Nara", is relatively likely to be a sentence indicating the movement of people. . In this way, it becomes easy to extract information related to a predetermined event from the class and how it is arranged.

  Note that if the conformity is defined as log (conformance rate) (the base is 10), the conformance rate is a value between 0 and 1, and therefore the conformity is always a negative value. In this case, it may be calculated as post-adjustment fitness Y = adjustment fitness X × adjustment coefficient T That is, by multiplying the pre-adjustment fitness level X by an adjustment coefficient T of 1.0 or less, the absolute value of the negative pre-adjustment fitness level X becomes smaller. It becomes larger than degree X.

  The present invention has been described based on the embodiments. This embodiment is an exemplification, and it will be understood by those skilled in the art that various modifications can be made to combinations of the respective constituent elements and processing processes, and such modifications are also within the scope of the present invention. is there.

  It should also be understood by those skilled in the art that the functions to be fulfilled by the constituent elements described in the claims are realized by a single function block shown in the present embodiment or a combination thereof.

It is a schematic diagram for demonstrating the outline | summary of the process by a document processing apparatus. It is a functional block diagram of a document processing apparatus. It is a flowchart which shows the test | inspection process process in the case of not performing an adjustment process. It is a graph which shows a relevance rate and a reproducibility in the case of not performing adjustment processing. It is a graph which shows a relevance rate and a recall when changing the adjustment coefficient T. It is a flowchart which shows the process in which a fitness is calculated based on multiple types of adjustment coefficients in a learning process. 7 is a flowchart showing a process for obtaining a provisional optimum coefficient S after the process shown in FIG. 6 is completed. It is a flowchart which shows the test | inspection process process in the case of performing adjustment processing. It is a graph which shows the verification result of the adjustment coefficient specific algorithm shown in the present Example.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 100 Document processing apparatus, 110 User interface processing part, 112 Input part, 114 Display part, 116 Document acquisition part, 120 Data processing part, 122 Inspection part, 124 Word extraction part, 126 Conformity calculation part, 128 Conformance adjustment part, 130 data holding unit, 132 class classification unit, 134 adjustment coefficient specifying unit, 136 provisional optimum coefficient specifying unit, 170 class feature holding unit, 200 learning corpus, 210 pre-processing inspection target document, 212 processed inspection target document

Claims (14)

A class feature holding unit that classifies words into a plurality of classes and holds the word features in a predetermined corpus as class feature information for each class;
A document acquisition unit for acquiring a document to be inspected;
A word extraction unit for extracting words from the inspection target document;
A fitness calculation unit for calculating the fitness of the extracted word feature and class feature information in each of the plurality of classes in the document to be examined;
A fitness adjustment unit that adjusts the fitness calculated for a predetermined class of the plurality of classes;
A class classification unit that identifies a class corresponding to the extracted word based on the degree of matching of the extracted word with respect to each class;
A document processing apparatus comprising:   The document processing apparatus according to claim 1, wherein the predetermined corpus is a document group in which a class is designated in advance for a word in the document.   The document processing apparatus according to claim 2, wherein the predetermined class is a class in which words not explicitly specified in the predetermined corpus are classified.   The document processing apparatus according to claim 1, wherein the predetermined class is a class in which most words are classified in the predetermined corpus.   The fitness level adjustment unit adjusts the fitness level of the extracted word with respect to the predetermined class by executing a predetermined calculation using an adjustment coefficient as a variable with respect to the fitness level with respect to the predetermined class. 5. The document processing apparatus according to claim 1, wherein the document processing apparatus is characterized in that: An adjustment coefficient specifying unit for specifying an adjustment coefficient value;
When identifying the adjustment factor,
The goodness-of-fit calculation unit divides the predetermined corpus into a corpus group and a check group, and sets the word features in the check group and the goodness of class feature information required for the corpus group for each of the plurality of classes. Calculated,
The fitness level adjustment unit adjusts the fitness level of the words included in the test group with respect to the predetermined class by a provisional coefficient that is a provisional adjustment coefficient,
The class classification unit specifies a class of words included in the examination group after the degree of conformity to the predetermined class is adjusted by a provisional coefficient;
The adjustment coefficient specifying unit specifies an adjustment coefficient from a plurality of types of provisional coefficients according to a degree of coincidence between a specified class and an original class in a word group included in a test group. 5. The document processing apparatus according to 5.   The adjustment coefficient specifying unit increases the accuracy of class classification by comparing words belonging to a class other than the predetermined class and words specified to belong to a class other than the predetermined class among words included in the inspection group. 7. The document processing apparatus according to claim 6, wherein the provisional coefficient when the accuracy of classification is maximized among the plurality of types of provisional coefficients is specified as an adjustment coefficient.   The document processing apparatus according to claim 7, wherein the adjustment coefficient specifying unit calculates an F value (F measure) obtained from a precision and a recall as a classification accuracy.   The adjustment coefficient specifying unit uses, as the adjustment coefficient, a provisional coefficient when the number of words whose original class matches the class specified in the word group included in the examination group is the maximum among the plurality of types of provisional coefficients. The document processing apparatus according to claim 6, wherein the document processing apparatus is specified.   The adjustment coefficient specifying unit divides the predetermined corpus into a predetermined number of groups of 2 or more, one of which is a test group and the other is a corpus group, so that the adjustment coefficient specifying unit is specified in a word group included in the test group. The predetermined number of provisional coefficients obtained by specifying a provisional optimum coefficient from a plurality of types of provisional coefficients according to the degree of coincidence between the original class and the original class, and setting each of the predetermined number of groups as an inspection group 10. The document processing apparatus according to claim 6, wherein the adjustment coefficient is determined based on the optimum coefficient.   The document processing apparatus according to claim 10, wherein the adjustment coefficient specifying unit uses an average value of the predetermined number of provisional optimum coefficients as an adjustment coefficient.   The adjustment coefficient specifying unit determines, as the adjustment coefficient, the temporary optimal coefficient when the degree of coincidence between the original class and the class specified in the word group included in the inspection group is the maximum among the predetermined number of temporary optimal coefficients. The document processing apparatus according to claim 10. Obtaining a document to be inspected;
Extracting words from the inspection target document;
Class features information and class feature information of the extracted word in the document to be examined with reference to class feature information obtained by classifying words into a plurality of classes and classifying word features in a predetermined corpus for each class Calculating a goodness of fit for each of the plurality of classes;
Adjusting the fitness calculated for a predetermined class of the plurality of classes;
Identifying a class corresponding to the extracted word based on a degree of fitness for each class of the extracted word;
A document processing method comprising: A function that classifies words into a plurality of classes and holds the word features of a given corpus as class feature information for each class,
A function to obtain a document to be inspected;
A function of extracting words from the inspection target document;
A function for calculating the degree of suitability of the extracted word feature and class feature information in each of the plurality of classes in the inspection target document;
A function of adjusting the degree of fitness calculated for a predetermined class of the plurality of classes;
A function for identifying a class corresponding to the extracted word based on the degree of matching of the extracted word with respect to each class;
A document processing program characterized by causing a computer to exhibit the above.

Images