JP2003296345A - Input template preparation knowledge extraction method from text db - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To drastically reduce the load of manual template description, and to extract detail template knowledge whose extraction used to be impossible by a normal data mining method. <P>SOLUTION: The morpheme analysis and modification analysis of sentences in a DB are carried out, and the candidates of a typical description pattern called generalized case expression is generated. Thus, it is possible to automatically generate template candidates having proper fineness by executing the frequency analysis of words or generalized case expression in a remark field appearing in records having a discrimination field fulfilling a prescribed condition. <P>COPYRIGHT: (C)2004,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a DB that stores a record including a finding field that stores a set of sentences that describe a fact and a judgment field that stores a sentence set that describes some judgment about the depicted fact. In the above, the present invention relates to a technique for extracting a template for supporting the input of a finding field for those having a judgment field satisfying a specific condition. The DB that stores the text-type data as described above is fundamental to various businesses, and in order to utilize the large amount of accumulated data as intellectual assets, the formal consistency and homogeneity of the described contents must be maintained. There is a need for improvement, and the importance of means for supporting user input using templates to realize this has been strongly recognized. In order to generate such a template, the present invention extracts knowledge typifying the contents of the DB,
The present invention relates to a technique for generating a template that guides user input.

An example of the above-mentioned DB is a "radiographic report" in which a doctor takes a picture of a part of the body of a patient using various imaging devices and writes findings and diagnoses on the obtained image.
There is data called. A large amount of data has already been accumulated in the radiology report database in hospitals that have become more electronic, but there is a strong demand for an input support system to reduce the labor of doctors and improve the quality of report content. In these respects, it is typical as an application target of the present invention. Therefore, in the following, the conventional technique and the present invention will be described by using the DB of the "interpretation report" as an example.

[0003]

2. Description of the Related Art As a conventional technique, there is a method of first interviewing an expert (a doctor in the case of an image interpretation report) with knowledge for creating a template for an input support system. The problem with this method is that the human burden is extremely heavy, which has been a bottleneck in system development.

As an alternative method for reducing the human burden, it is conceivable to apply a general data mining method and derive a "correlation rule". Here, in data mining, each element that appears in the DB is called an item, a set of items in processing is called a transaction, and a set of any number of items is called an item set. For β, item sets X and Y that are sparse with respect to each other, the ratio of transactions that include X∪Y in all transactions is α or more, and the ratio of transactions that include Y in transactions that include X is β or more.
The pair of X and Y is represented by the symbol X⇒Y, which is called the association rule. The basic idea is to define that "X and Y are related" as "there is a strong tendency for X and Y to be included in the same transaction." The ratio of transactions that include X∪Y in all transactions is called X⇒Y support, and the ratio of transactions that include X in X is called the confidence of X⇒Y. A method has been developed to obtain an association rule of α or higher and confidence of β or higher (Agrawa
l, A., Imielinski, T., and Swami, A. (1993).
Mining association rules between sets of items in
large databases. In Proceedings of the ACM SIGMOD
Conferences on Management of Data, pp.94-105).

[0005]

The above basic framework records, for example, supermarket sales.
In a typical case such as a "basket analysis" in which "a group of items X and a group of items Y that have a strong tendency to sell together" are checked in the DB, it is immediately applicable. However, there is a problem, for example, in extracting knowledge for supporting input creation of findings from the DB of "radiographic report". First, in a typical example like a supermarket "basket analysis", each item is an individual item sold and a transaction is "purchased together", the definition of which is self-explanatory.

On the other hand, in the case of the problem that we deal with, when "gastric cancer (is)" is described in the judgment field, for example, in the findings field, "irregularities in the gastric mucosa are observed in UGI". There are many situations. Even though it is natural to use one unit of transaction as one record, it is not possible to consider the finding "(gastric cancer)" and the sentence expression "irregularities in the stomach mucous membrane in UGI" as unit items. It is not always obvious from the general theory of data mining. (For example, you could just define a "word" as an item). Furthermore,
Even if it is determined that the sentence is regarded as a unit item as described above, merely deriving a large number of sentence pairs is not sufficient for extracting the desired "knowledge for template generation". In order to extract "knowledge for template generation", it is necessary to collectively capture sentences with common content,
For that purpose, we perform morphological analysis and crossword analysis using a dictionary that describes parts of speech, case slots (in the case of verbs), meaning codes, etc. It is necessary to recognize the "dependencies between words" that are included and to extract the major dependencies that characterize the finding fields that satisfy the conditions, that is, the typical descriptive patterns described above, from among them. After extracting the typical descriptive pattern, the variation of the modifiers that modify the noun phrase appearing in the typical descriptive pattern may be captured, and the process may be presented as a list. Here, once the typical description pattern is acquired, if each sentence is treated as a description pattern as it is, it will be different due to the difference in word order and small modifiers.
This is because sentences that have essentially the same content are treated as different sentences, and the appearance frequency of each description pattern is extremely low, making it difficult to find a characteristic pattern. Clearly, the above cannot be achieved by the conventional framework.

The problem to be solved by the present invention is to significantly reduce the burden of manual template description and to use it for input template description that cannot be extracted by the normal data mining method as described above. It is to extract the knowledge that can be formed.

[0008]

[Means for Solving the Problems] Since the purpose is knowledge extraction for creating an input support template for a DB that stores data in a format having a finding field and a corresponding judgment field, Then, prepare a list of "words for which the user wants to input the finding as a template". This may be specified by the user, or may be a list of words automatically determined to be particularly frequent in the determination field. If we find a set of records that include the words in this list (the name of the disease in the case of a radiology report), and extract the typical descriptive patterns that are specific enough to characterize their findings fields and that do not have unnecessary detail at the same time. Good.

For this purpose, first, it is necessary to recognize the sentence contents in the above-mentioned set of finding fields by absorbing differences in wording and word appearance order. In order to "ignore differences in wording and word appearance order", each sentence is expressed as a word dependency centered on the verb VP at the end of the sentence. That is,
The VP can take grammatically possible case slots {C_1,…, C_VP
(M)}, the noun phrase NP corresponding to each case slot C_i
_i is certified, and VP and noun phrase pair {VP, NP_1,…, NP_V
(M)}. Here, the noun phrase corresponding to C_i is
When it is a long one that includes adjectives and adnominal modifiers, the main part excluding these modifiers, that is, the last noun (single or compound noun) is NP_i. Thus obtained {V
Hereinafter, P, NP_1, ..., NP_V (m)} is referred to as a case expression.
Actually, the case may be omitted, so in that case NP
Replace _I with a particular identical symbol, called a wildcard. In addition, we obtained {VP, NP_1,…, NP_V
(M)} is generated by replacing any non-wildcard of {NP_1, ..., NP_V (m)} with a wildcard. Since the expression in which the elements of the case slot are not specified is more general, they will be referred to as generalized case expressions below.

All generalized case expressions are generated from each sentence in the finding field in the DB, and in addition to the words in the DB, the generalized case expressions are subjected to frequency analysis in the DB. Of the generalized case expressions that appear in the finding field in the record that has a judgment field that satisfies a specific condition, the most specific one that satisfies the predetermined "high frequency" condition,
That is, the generalized case expression with the fewest wildcard noun phrases is extracted. As a result, a generalized case expression corresponding to “a phrase having specificity enough to characterize the finding field but not having unnecessary detail” can be extracted. It can be said that this is the typical description pattern in the "finding field corresponding to the judgment field satisfying the specific condition" described above. Once the typical description patterns are found, for the selected ones, the wildcards are first reduced to concrete noun phrases, which are presented to the user for selection. Furthermore, regarding the selected noun phrase, the modifier phrases are collected and presented, and the user is allowed to select.

The above procedure is summarized as follows. (1) word recognition in sentences and dependency relation extraction processing between words, (2) case expression extraction centering on sentence end verbs and its generalization processing, (3) words and generalized case expressions Creating an index file that includes (4) “Words that represent judgments and for which you want to input the findings for that template”
(5) Using the index file, a typical descriptive pattern of a generalized case expression that frequently appears specifically in a finding record that has a judgment field (in the case of an image interpretation report, a disease name) having a specific condition using the index file. Process of presenting the generalized case expressions extracted in (6) and (5) to the user and allowing the user to select an arbitrary number of elements from them (7) and (6) For the personalized expression, the process of restoring and presenting the noun phrase that was a wildcard in the generalized personality expression, and allowing the user to select any number of the elements, (8) and (7) For noun phrases, processing that restores and presents the modified phrase that was modified in the original sentence, and allows the user to select an arbitrary element from them. Here, (4) is a process different from other processes,
It is not necessary when the user gives a word, but when automatically extracting the target word, the result of the morphological analysis and the frequency information of the word are necessary, so it must be placed immediately before (5).

[0012]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 shows an example of a system configuration for realizing the knowledge extraction method described above. A storage device 101 stores document data and various program modules using a hard disk or the like. It is also used as a work area for program modules.

In the following, 1010 is a control module for managing the exchange of data between the user interface and other modules, 1011 is a document data DB for accumulating documents for knowledge extraction, and 1012 is a morphological analysis and a catching analysis. A dictionary for doing, part-of-speech and case slots for words
(For verbs), the meaning code, etc. are described. 101
Reference numeral 3 denotes a morphological analysis module, which identifies words that form a document. In the case of Japanese, it is divided into words and part-of-speech, and in the case of English, processing such as prototype reduction is performed. We do not specify these methods. In both languages, various systems have been released for commercial and research purposes. By morphological analysis, for example, the sentence "recognizing irregularities of mucous membrane in the pyloric part" is like "recognizing pyloric part / in / mucosa / of irregularity /"
It is broken down into the sequence of words that make up the sentence. Reference numeral 1014 is a hangup analysis module, which analyzes a word in a sentence and a dependency relationship between the words. "Pyle part / on / mucosa / of / irregularity / recognize"
Are analyzed as [⇒ recognized in pylorus], [⇒ recognized irregularity], [⇒ irregularity in mucous membrane]. By these, the procedure (1) described in the means for solving the problem is realized. 1
The analysis results of 013 and 1014 are the shared data recording area 10
Recorded in 26 and used at 1015, 1016, 1017.

Reference numeral 1015 is a module for extracting the "case expression" from the result of 1014, and paying attention to the verb at the end of the sentence, generating expressions such as {accept, pylorus (d), irregularity (wo)}. To do. Here, the “admit” rating is “irregular mucosa”
However, in order to leave a more general and essential part, the main part is reduced to “irregularity”, and the modified part “mucosal” is recorded in the shared data recording area 1027. 1016
From the output of 1015, {Acknowledge, pylorus
(D), irregularity (wo)}, {accept, * (d), irregularity (wo)}, {acknowledge, pylorus (d), * (wo)}, {accept, * (d), * (wo )}
Generates a generalized case expression. The reason why all generalized case expressions are generated is that it is not known in advance how much detailed description level is appropriate as a template for describing a finding field having a judgment field that satisfies a specific condition. This is because, in consideration of the expression of the level, the one having the most biased appearance is selected by calculating the “frequent specificity”. The important thing is that template input improves input efficiency, so even if you can not cover all the input by force, you can assist the input in the majority of cases, and leave the input that cannot be covered to free input. do it.

For example, {accept, * (d), * (wo)} almost always appears in the finding field, so if this is used as a template, the user has to specify all but the verb. Less useful than simple templates. Similarly, all cases are buried {acknowledge,
Case patterns such as the pylorus (d) and irregularity (wo)} are too detailed to be typical patterns and are generally infrequent. It is not preferable to present a low-frequency one as a typical expression as a template input menu because many low-frequency ones are presented. Generally suitable for use as templates is a generalized case expression with a moderate generality such that the peculiarity of occurrence is in between. For example, in the finding field of a record that has “stomach cancer” in the judgment field, {accept, * (d), irregular
(Wo)} becomes a generalized case expression with a moderate generality that is intermediate between the two. The frequent uniqueness of each generalized case expression is calculated by the frequent uniqueness calculation module 1018, and is recorded in 1026 together with the original case expression and the frequent uniqueness. 1015 and 1016 realize the procedure (2) described in the means for solving the problem.

Word / generalized case expression recording module 101
7 is the result of the morphological analysis recorded in 1026 and 10
Based on the generalized case expression generation result generated in 16, the number of times each word or generalized case expression appears in which field of which record is recorded. That is, an index file is created. By 1017, the procedure (3) described in the means for solving the problem is realized.

The frequent uniqueness calculation module 1018 determines whether or not a word or a generalized case expression appears specifically and frequently in a set satisfying a specific condition. There are several possible ways to do this. The simplest one is to simply return the frequency except for all the case slots that are filled, and a slightly more complicated one is disclosed in "Method of selecting characteristic words using probability: P2000-354407". It is also conceivable to use a stochastic measure as described above. That is, the frequency of the word or generalized case expression whose frequency specificity is to be measured is K in the entire document set, the frequency of the word or generalized case expression in the set satisfying the specific condition is k, and the word and general case When the total frequency of personalized expressions in all document sets is N and the number of words in the set satisfying the specific condition is n, the weight of the word is “marked in N balls. When there are K balls, the probability that more than k balls will be included when any n balls are taken out from this ”p (N, K, n, k)
Corresponding to (see Mathematical Expression 1) and reversing the sign of the logarithmic value of the probability value is taken as the frequent uniqueness of the word or generalized case expression whose frequent uniqueness is to be measured. It has been found that this scale has a good property that, unlike mere frequency, it can exclude high-frequency unnecessary words such as "do", and at the same time, it does not cause overestimation of low-frequency elements (Hisamitsu, T. and Niwa,
Y. (2001) Topic-Word Selection by Combinatorial Pr
obability, Proceedingsof NLPRS2001, pp.289-296).
p (N, K, n, k) is given by the following equation. The scale used is not specified here.

[0018]

[Equation 1] The judgment field analysis module 1019 analyzes the judgment field, refers to the data of 1017, and determines 1018.
Characteristic words in the judgment field are extracted by using the specificity calculated in step. This may be fully automatic or may be a method of presenting candidates to the user for selection. If the user has previously listed up “a word that represents a judgment, and a word for which the user wants to input a finding for that” as a template, the judgment field analysis module 1019 is not used,
The list is recorded directly in the target word recording module 1020. This control is performed by the control module 1010. The target word recording module records a list of "words that represent judgments and words for which the user wants to input a finding therefor as a template" obtained at 1020 or 1019 or designated by the user. If the user does not provide the list manually, 1019 and 1020 implement the procedure (4) mentioned in the means for solving the problem.

Typical description pattern extraction module 1021
For each word recorded in 1020, a record that includes it in the judgment field is extracted, and the data of 1017 is referred from the finding field included in them, and 1
Using the frequent uniqueness calculated in 018, a generalized case expression that is uniquely frequent is extracted as a typical description pattern and recorded in the shared data recording area 1026 together with the word of the target word recording module 1020 that is the source of extraction. To do. 10
21 realizes the procedure (5) as a means for solving the problem.

Typical description pattern presentation module 1022
Presents the generalized case expression recorded in the shared data recording area 1026 to the user, allows the user to select any number of the elements, and selects the selected element from the shared data recording area 1
Record at 026. The method of presenting the generalized case expression to the user is not specified, but, for example, for each "word that represents a judgment and for which the user wants to input the finding for that template", the verbs with high specific frequency are displayed in order. There is a method. By 1022, the procedure (6) described in the means for solving the problem is realized.

The noun phrase presentation module 1023 presents the generalized case expression recorded in the shared data recording area 1026 to the user, allows the user to select any number of the elements, and selects the selected element from the shared data recording area. Record at 1026. The method of presenting the generalized case expression to the user is not specified. By 1023, the procedure (7) described in the means for solving the problem is realized.

The modifier presentation module 1023 presents the noun replaced by the wildcard for the generalized case expression selected by the user, which is recorded in the shared data recording area 1026, and any of the elements To the user and select the selected element from the shared data recording area 102.
Record at 6. The method of presenting the noun phrase to the user is not specified. By 1023, the procedure (7) described in the means for solving the problem is realized.

The modifier presentation module 1024, for each generalized case expression selected by the user and the nouns (groups) in the case slot, recorded in the shared data recording area 1026, provides a phrase for modifying it. Extracted from the output record of 1014, made a list, and let the user select any number of the elements, and select the selected element as a template record.
Record in DB1027. By 1024, the procedure (8) described in the means for solving the problem is realized.

FIG. 2 shows an example of a plurality of records including "stomach cancer", "esophageal cancer", "colon cancer", etc. in the judgment field.
201 shows one record as a unit.

FIG. 3 shows appearance frequencies and appearance field information of words and generalized case information extracted from the record group shown in FIG.

FIG. 4 shows a state in which a typical descriptive pattern that is the basis of a template for inputting “stomach cancer” is presented in the judgment field using the information shown in FIG. 3 and selection is requested. . 401 is a presentation window, 402
Is the selection subwindow.

FIG. 5 shows a state in which noun candidates to be adopted in the two cases are presented and selection is requested, assuming that “accept irregularity in (...)” is selected in FIG. Reference numeral 501 is a presentation window, and 502 is a selection subwindow.

FIG. 6 shows a state in which, when “on front wall” is selected in FIG. 5, candidate modifier phrases to be used as a template menu are presented among the phrases that modify “irregularity” and selection is requested. Reference numeral 501 is a presentation window, and 502 is a selection subwindow. Here, 502 indicates "to
The modifiers that appear throughout the "Accept irregularities" are shown, but it is also possible to limit this to those that appear in the "Accept irregularities on the front wall" pattern.

[0029]

According to the knowledge extraction method described in the present invention, it is possible to generate a finer template than at the time of using a general data mining method without incurring human cost.

[Brief description of drawings]

FIG. 1 is a system configuration diagram for realizing a knowledge extraction method of the present invention.

FIG. 2 is an example of a plurality of records.

FIG. 3 is a word extracted from the record group shown in FIG.
And information on generalized personality information.

FIG. 4 is a state in which a typical description pattern generated using the information shown in FIG. 3 is presented and selection is requested.

5 is a state in which candidates to be adopted for a noun phrase that has become a wildcard are presented among the typical descriptive patterns selected in FIG. 4 and selection is requested.

FIG. 6 is a state in which candidates for a menu of modified phrases of noun phrases are presented to the template in which noun phrases are selected and detailed in FIG. 5, and selection is requested.

[Explanation of symbols]

101: Storage device 1011: Document data DB 1012: Dictionary 1013: Morphological analysis module 1014: Dependency analysis module 1015: Case expression extraction module 1016: Generalized case expression generation module 1017: Word / generalized case expression recording module 1018: Frequent occurrence Specificity calculation module 1019: Judgment field analysis module 1020: Target word recording module 1021: Typical description pattern extraction module 1022: Typical description pattern presentation module 1023: Noun phrase presentation module 1024: Modifier presentation module 1025: Template record DB 1026: Shared Data recording area 1027: Work area 201: Word extracted from the record group shown in one record 301: 201 which is a unit of DB, and generalization case information 401: Typical description pattern Display window 402: typically described pattern selection sub-window 501: noun candidate presentation window 502 employing the template: 501 selection sub-window 601: qualifiers candidate presentation window 602 employing the template: 601 of the selected sub-window.

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Masakazu Fujio             1-280, Higashi Koikekubo, Kokubunji, Tokyo             Central Research Laboratory, Hitachi, Ltd. (72) Inventor Kaichi Kaji             1-14-1 Kanda, Uchida, Chiyoda-ku, Tokyo             Inside the Hitachi Medical Co. F term (reference) 5B075 ND03 NK32 NR02 PP13 QM05                       UU28                 5B091 AA15 CA02 CA05 CC04 CC16                       DA06

Claims (4)

[Claims] 1. A DB storing a record containing a finding field for storing a set of sentences describing facts and a judgment field for storing a set of sentences describing some judgments about the depicted facts. A record that satisfies a specified condition is extracted, and a pattern of a sentence description that is unique and frequent, called a typical description pattern, is extracted from the finding field in the extracted record set. A knowledge extraction method characterized in that it is presented and can be selected by a user, and the selected typical description pattern is recorded in association with a condition regarding a judgment field used for record extraction. 2. The knowledge extraction method according to claim 1, wherein the DB
For the verb phrase VP in the sentence that appears inside, the case that the VP takes {C_
1,…, C_VP (m)}, the noun phrase NP_ corresponding to the case C_i
generate i, a set {VP, NP_1, ..., NP_VP (m)} of the recognized noun phrase and VP, which i is recognized and is called a case expression, NP
A set of generalized case expressions {VP, NP_1 ', in which any number of _i is replaced with a specific identical symbol called a wildcard
, NP_VP (m) '} is generated, and the generalized case expression is used as a typical description pattern. 3. In presenting a generalized case expression extracted as a typical description pattern by using the knowledge extraction method according to claim 2, a generalized case expression having the verb phrase as an element for each same verb phrase A knowledge extraction method characterized by presenting the noun phrases appearing in, and allowing the user to select any of the noun phrases. 4. When presenting a generalized case expression extracted as a typical description pattern using the knowledge extraction method according to claim 2, a noun phrase that modifies the noun phrase appearing therein is extracted from the DB and presented. A knowledge extraction method characterized in that a user can select any of the modifiers.