JP2002230021A - Information retrieval device and method, and storage medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To easily and quickly perform information retrieval intended by a user. SOLUTION: A word is extracted from a retrieval query designated by a user, and modification analysis is operated (S41), and polysemy elimination is tried based on the result of the modification analysis and a co-occurrence data base (S42), and when the polysemy is eliminated, the processing is immediately moved to a S45, and in the other case, polysemy elimination based on user profile information is operated, and then the processing is moved to the S45. Then, extension processing such as near-synonym development is operated by using a query extension dictionary (S45), and the document vector generation processing of the retrieval query is operated (S46).

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an information retrieval apparatus, an information retrieval method, and a storage medium.
The present invention relates to an information search device that performs information search according to a “query”, an information search method using the information search device, and a storage medium that stores a search procedure for information search and a data structure for performing information search.

[0002]

2. Description of the Related Art With the recent development of computers and communication networks (networks), the accumulation of a large number of electronic documents in a database has been progressing. There is an increasing demand for information retrieval for retrieving document data.

Conventionally, this type of information search is based on a search method premised on matching the notation with a query, such as a keyword search or a full-text search. A method for searching for a similar document has been proposed.

[0004] As a similar document search method, a document is mapped to points in an n-dimensional vector space, and the similarity between the documents is determined by the distance between them.
Alternatively, a vector space model method for calculating the similarity between a query and a document is already known (for example, “Kumamoto, Shimada,
Kato: Application of Concept-based Information Retrieval, IEICE Technical Report, Vol.
AI98-63, pp.9-16, 1999 ”).

[0005] However, in such a vector space model system, particularly when a similar document is searched using a short sentence as a query, a document not desired by the user is often searched.

That is, since a character string used in a query has ambiguity, the query is not interpreted in a meaning desired by the user, or the meaning of the character string used in the query is incompletely interpreted. Therefore, the meaning of the query is not sufficiently supplemented, and therefore, in the vector space model method described above, the search result often does not meet the user's intention.

Therefore, as a measure for avoiding such a search of a document not intended by the user as much as possible, a technique has been proposed in which, when a query is expanded with synonyms, synonyms having low co-occurrence between words are excluded from expansion targets. (See, for example,
No. 45274; hereinafter, referred to as “first related art”).

In the first prior art, a synonym having a low co-occurrence is removed from a search target of a document search.
Documents in which synonyms unrelated to the document desired by the user appear are excluded from search targets.

Another conventional technique is to record the access status of each user for each document, create a preference vector based on the access status, and shift the query vector so as to approach the preference vector. (Japanese Patent Application Laid-Open No. H11-53394; hereinafter, referred to as "second conventional technique").

In the second prior art, the search result is made closer to the user's preference by shifting the vector of the query so as to be closer to the preference vector, thereby making it possible to retrieve a document reflecting the user's preference. I have to.

[0011]

However, in the first prior art, when synonym expansion is performed to expand the meaning of a query, only synonyms with low co-occurrence are excluded. The meaning of each word in is not specified,
Therefore, the ambiguity of each word remains unresolved, and a document that the user does not want is often displayed as a search result.

That is, in the first prior art, for example,
If the user enters a query such as "Please tell me what form types are available," many documents are searched, including documents related to forms and baseball pitching forms. Displayed and output in the retained state.

[0013] However, a user who is engaged in the design of financial forms needs documents related to the types of forms, and usually does not need documents related to baseball pitching postures, swimming styles, and the like. On the other hand, a user who enjoys sports often desires documents relating to baseball pitching postures and swimming techniques, and usually does not desire documents relating to types of forms.

That is, in the first prior art, although a synonym having low co-occurrence is excluded, a search is performed while retaining the polysemy of the query. Is displayed and output, which makes it difficult to obtain a desired search result easily and quickly.

In the second prior art, the meaning of the entire query is forcibly shifted in the direction of the user's past preferences, that is, in the direction of the preference history. When examining a form, it is necessary to consider not just a "form type" but a character string "throwing form type" in order to avoid misunderstanding the meaning interpretation on the system. There was a problem that it became worse.

Moreover, in the second prior art, since the query uniformly approaches the preference vector, even if a character string of "the type of pitch form" is input, the form designer is instructed. However, there is a risk that a document related to a form may be searched, and as a result, a search result that is not desired by the user may be displayed.

As described above, in the first and second prior arts, the search processing is merely performed based on the description contents of the query and the user's preference, and the ambiguity of the query is not taken into consideration. There is also a problem that the search result that is not desired by the user is often obtained because the supplement of the meaning content is not sufficiently performed.

The present invention has been made in view of such a problem, and provides an information search apparatus, an information search method, and a storage medium that can easily and quickly perform information search according to a user's intention. The purpose is to:

[0019]

In order to achieve the above object, an information retrieval apparatus according to the present invention comprises a search condition input means for inputting a search condition, and a method for retrieving words from the search condition input by the search condition input means. Morphological analysis means to be extracted, co-occurrence relation storage means for storing co-occurrence relations between a plurality of words in association with meanings, co-occurrence information stored in the co-occurrence relation storage means, and analysis results of the morphological analysis means A first ambiguity eliminating means for extracting a co-occurrence relation of the search condition based on the first and second ambiguities, and performing an information search on the basis of the sense of ambiguity eliminated by the first ambiguity eliminating means. User profile storage means for storing user profile information expressing user preferences, and user profile information stored in the user profile storage means. Second ambiguity eliminating means for eliminating the ambiguity of the search condition based on the file information and the analysis result of the morphological analysis means, wherein the first ambiguity eliminating means comprises the second ambiguity eliminating means. It is characterized by giving priority over means.

Further, the information search method according to the present invention includes a search condition input step of inputting search conditions, a morphological analysis step of extracting words from the search conditions input in the search condition input step, and A first polysemy resolving step of extracting a co-occurrence relation of the retrieval condition and resolving polysemy based on co-occurrence information stored in association with co-occurrence relation with word meaning and the analysis result in the morphological analysis step When,
An information search step of performing an information search based on the meaning of the ambiguity eliminated in the first polysemy elimination step, further comprising: user profile information expressing user preferences; A second ambiguity eliminating step for eliminating the ambiguity of the search condition based on the analysis result, wherein the first ambiguity eliminating step has priority over the second ambiguity eliminating step. .

Further, the storage medium according to the present invention is characterized in that:
A second disambiguation procedure for resolving the disambiguation of the search condition is stored based on the user profile information expressing the user's preference and the analysis result of the dependency analysis step, which is prioritized over the disambiguation procedure. In addition, a semantic extension procedure for performing an extension process on the semantics whose ambiguity has been eliminated in the second ambiguity eliminating procedure is stored.

The other features of the present invention will be apparent from the following description of embodiments of the invention.

[0023]

Next, an embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of a document search device as an information search device according to the present invention. The document search device includes an input device 1 including a keyboard and a mouse, and a CRT. , A hard disk (HD) 3 storing predetermined data described later, and a flexible disk (F).
D), a removable disk drive 4 for accessing an external storage medium such as a CD (compact disk) and a DVD (digital video disk), and a communication device such as a modem or a LAN controller for exchanging data with the outside via a communication line. 5, a read-only fixed memory (ROM) 6 in which a predetermined control program described later is stored, and a writable random access memory (RAM) 7 for temporarily storing various data and used as a work area
And a central processing unit (CPU) 9 connected to each of the above components via a bus 8 and controlling the entire apparatus.

The HD 3 has a word vector dictionary 3a storing meaning vectors of words on a meaning basis, a document database 3b to be searched, a user profile 3c recording user preferences, and co-occurrence among dependency words. A co-occurrence database 3d in which relationships are stored, and a query expansion dictionary 3e that performs synonym expansion and related expansion of queries are stored.

In this embodiment, the control program that is processed by the CPU 9 is stored in the ROM 6. However, the control program is stored in the HD 3 instead of the ROM 6, and loaded from the HD 3 onto the RAM 7 and executed. Alternatively, the program may be stored in an external storage medium such as an FD, loaded on the RAM 7 via the removable disk drive 4, and executed.

FIG. 2 is a diagram showing an example of a display screen of the display device 2. The display screen includes a query display section 2a for displaying a query (search condition) and a search for displaying a search result based on the query. And a result display section 2b.

The query display section 2a specifically displays search conditions input by the user via the input device 1, for example, natural sentences (“form type”, “want to know form type”, “variation of pitch form” ), A list of a plurality of keywords (such as “form, type”), or an existing document specified by the user (such as “document number 267”).

The search result display section 2b displays, as a search result, a document ID for identifying each document, a document title corresponding to the document ID, and the similarity of the document ID to the query.

FIG. 3 is a format diagram of the word vector dictionary 3a stored in the HD 3.

The word vector dictionary 3a is a set of semantic vectors (feature lists for each semantic classification) indicating the meaning of each word, and each dimension (1, 2, 3,...) Represents a semantic classification. I have.

That is, in the word vector dictionary 3a, a specific meaning is assigned to each word (1, 2, 3,...), And to what extent the meaning classification of each dimension is implied for each word. That is, the feature amounts of the meaning vectors are written in a matrix.

For example, dimension 3 indicates a semantic classification of “space / sky”, dimension 4 indicates a semantic classification of “trading / trading”, and dimension 7 indicates a semantic classification of “gesture / action”. 7 is given a specific meaning called “form (form)”. In the word vector dictionary 3a, since the feature amount of the dimension 3 meaning vector in the word 7 is "0", the word "form (form)" has no meaning of "space / sky" at all. You can see that.

In the case of the word 7, the feature amount of dimension 4 is “2”.
“1” is relatively large compared to other feature amounts, and the feature amount of dimension 7 in word 7 is relatively small as “1”, which means that “form (form)” is “transaction / trading”. Is large, but the contribution of "gesture / motion" is small.

The word 8 has the meaning of "form (posture)". In the word 8, the dimension 4 has a feature amount of "0" and the dimension 7 has a feature amount of "23". Big. This means that "form (posture)" has no meaning of "transaction / buy / sell", but "gesture / action"
Means that the contribution degree is large.

As described above, by the word vector dictionary 3a,
It is possible to recognize the degree of contribution of each word according to meaning.

FIG. 4 is a format diagram of the document database 3b stored in the HD 3, in which the feature amount of the document vector is written.

It is determined that the meaning of the document is determined by what word is used in the document, and the meaning of each document is calculated by adding the meaning vectors of the words constituting the document. You. Therefore, the dimension of the calculated vector becomes the same as the dimension of the meaning vector of the word vector dictionary 3a, and expresses a specific meaning classification. Then, the vector obtained by the addition is normalized on the basis of “1”, and the normalized vector is stored in the document database 3b as a feature amount of the document vector.

As is apparent from FIG. 4, for example, when the document ID is “6949”, the feature quantity of dimension 4 is “0.009” and the feature quantity of dimension 7 is “0.42”.
5 ", and the document ID is" 6953 ".
Is "0.362", and the feature of dimension 7 is "0.008". As a result, the sentence having the document ID “6949” includes the semantic classification of “gesture / hand gesture” to some extent, but hardly includes the semantic classification of “transaction / trading”. It can be seen that the sentence "6953" includes the semantic classification of "transaction / trade" to some extent, but hardly includes the semantic classification of "gesture / action".

FIG. 5 is a format diagram of the user profile 3c stored in the HD 3.

The user profile 3c has the same dimension as the meaning vector of the word vector dictionary 3a, and is updated every time a user accesses a document file.

That is, in the initial state, the profile is set to "0", but when the user accesses a specific document, the document vector of the document is calculated, and the calculated value is added to the cumulative profile. And after a new cumulative profile is obtained,
The profile is normalized based on “1”, and the normalized profile is updated.

FIG. 5A shows, for example, a profile of a user who is interested in sports, and the feature amount of dimension 7 (gesture, motion) is relatively large at "0.186". This indicates that the user refers to many documents having a semantic classification of “gesture / action”.

On the other hand, FIG. 5B shows a profile of a user who is deeply interested in the window service, and has a dimension 7 (gesture,
The characteristic amount of the operation (operation) is “0.000”, but the characteristic amount of dimension 4 (trade / trade) shows a large numerical value of “0.329”. This indicates that the user refers to many documents related to “trading / trading”.

FIG. 6 is a format diagram of the co-occurrence database 3 d stored in the HD 3, in which a related word, a received word,
And three pieces of co-occurrence information of particles intervening between them are stored. If the particle does not exist, "null"
Is written.

In the present embodiment, after performing a morphological analysis on a character string in an input query, a dependency analysis is performed to extract dependency words, received words, and particle information, and the co-occurrence database 3d is referred to. A collation process is performed between these related words, received words, and particle information. If the co-occurrence database 3d has character strings corresponding to the above-mentioned related words, received words, and particle information, each word is stored in the co-occurrence database 3d.
The meaning is interpreted as described in d.

For example, when the phrase “throwing form” is extracted as “throw / form” by dependency analysis, the meaning of this form is interpreted as the meaning of “posture” according to the co-occurrence database 3d. When the query is "input information in a form", the word "form / ni / input" is extracted by the dependency analysis, and a collation process is performed between the word and the particle and the co-occurrence database 3d. And as a result this "form"
Is interpreted as a meaning of “form”.

FIG. 7 is a format diagram of the query expansion dictionary 3e stored in the HD 3. FIG. 7A shows a synonym dictionary, and FIG. 7B shows a related word dictionary.

That is, the synonym dictionary stores synonyms to be developed for the entry word. For example, the headword “form (posture)” is a synonym for “posture, shape,
"Style, pose" is stored, and "form (form), form" is stored as a synonym in the entry word "form (form)".

Usually, synonyms have a synonymous concept or a lower concept with respect to the headword. The synonym dictionary is used to expand each word in the query and extend the meaning of each word to an allowable range. Conventionally, there was a synonym that was expanded for the “form” described in the query, so “posture” and “form” were mixed and expanded, but in the present embodiment, the synonym dictionary is Since it is held at the base, "posture" and "form" are not mixed and developed. Note that synonyms also store semantic information, and can be processed on a synonym basis even after synonym expansion.

The related word dictionary stores related words to be developed for the starting word. For example, the starting word "form (posture)" is related to "sports,
Analysis, improvement ”is stored, and the starting word“ form (form) ”is stored.
Is related to "purchase, sign up, transfer, remittance"
Is stored.

It should be noted that the related word differs from the synonym described above in that there is no higher-order or lower-order relationship with the originating word. The related word dictionary is used to expand each word in the query to enhance the meaning of the entire query to some extent.

The document retrieval apparatus thus configured operates according to various inputs from the input device 1, an input signal from the input device 1 is supplied to the CPU 9, and the CPU 9 stores the input signal in the ROM 6. Read the control program
Various controls are performed according to the control program.

FIG. 8 is a flowchart showing an embodiment of a processing procedure of the document search method executed by the document search apparatus. The program is executed by the CPU 9.

After performing initialization processing such as initialization of various parameters and lighting of an initial screen in step S1, step S1 is performed.
2 waits for an operation input from the input device 1, and in the following step S3, the content of the input operation is determined.

That is, the search procedure of this document search method is as follows.
The process is roughly divided into three processes: a registration process to a document database, a user profile update process, and a search execution process according to a query. Therefore, the user selects one of these three processes according to the search stage. Perform input operation.

When the registration processing to the document database 3b is instructed, the process proceeds to step S4 to execute the document registration process. When the update process of the user profile 3c is instructed, the process proceeds to step S5 to perform the profile updating process. Is executed, and when the search execution process is instructed, the process proceeds to step S6 to execute the search execution process.
The process proceeds to step S7, where the processing result of each of the above processes is developed into a display pattern and output, and the process returns to step S2.

FIG. 9 is a flowchart showing the procedure of the document registration process executed in step S4 (FIG. 8). In order to execute a search process described later, a document vector is registered in the document database 3b.

In step S11, morphological analysis is performed on the input query to extract words, and then dependency analysis is performed. In the subsequent step S11, the dependency word and the received word analyzed by the dependency analysis are collated with the co-occurrence database 3d, and a character string in which the related word and the received word are paired is stored in the co-occurrence database 3d. Specifies the meaning of a word.

For words whose meaning cannot be specified, the word vectors of all the meanings having the notation are weighted and added.

Next, in step S13, a document vector generation process is performed. That is, step S11 and step S
The word vector dictionary 3a is searched from the word extracted in step 12 and the specified meaning, and the feature amount of the meaning vector is calculated, and the feature amount of the document vector characterizing the document is generated from the sum. That is, the document vector represents the meaning expressed by the document as described above, and is generated by adding the feature amount of the meaning vector written in the word vector dictionary 3a for each word.

Then, in step S14, a registration process to the document database 3b is performed, and the process returns to the main routine (FIG. 8). That is, the content of the document and the feature amount of the document vector obtained in step S13 are stored in the document database 3b.
And updates the index of the document database 3b.

FIG. 10 is a flowchart showing the procedure of the profile update process executed in step S5 (FIG. 8). When a specific document file is accessed by an instruction from the user, for example, it is registered in the document database 3b. This is executed when a file is read from or written to an external storage medium such as an FD that is not used for personal use, or when a Web page is accessed via the Internet.

In step S21, document data is obtained.
Next, in step S22, after extracting words by morphological analysis, dependency analysis is performed. Next, step S23
Then, as described above, the co-occurrence database 3c is referred to, and when the set of the related word and the received word is stored in the co-occurrence database 3c, the meaning of the word is specified.

Next, in step S24, as described above,
The word vector dictionary 3a is searched from the word extracted in step S21 and step S22 and the specified meaning to generate a meaning vector, and then a document vector is generated. In step S25, the generated document vector is added to the cumulative profile, and in step S26, “1” is added.
Is normalized, and a normalized profile is created.

After updating the user profile in this manner, in step S27, the original processing of each file (for example, file reference processing, writing processing, etc.) is performed, and thereafter, the process returns to the main routine.

FIG. 11 is a flowchart showing the procedure of the search execution process executed in step S6 (FIG. 8).

In step S31, a search query input process is executed, and the user inputs a query using a natural sentence, a plurality of keywords, or designation of an existing document, and obtains a text string of the query according to the input content. For example, when an existing document is specified as a query, the document is accessed, the format is changed to an appropriate format, the contents of the document are converted to a text file, and the text string is obtained.

Next, in step S32, a query vector generation process is performed based on the text string.

FIG. 12 is a flowchart of the query vector generation processing routine executed in step S32.

That is, in step S41, words are extracted from the search query specified by the user, morphological analysis is performed using the morphological analysis dictionary, and dependency analysis is further performed. In the following step S42, all the results of the dependency analysis are collated with the co-occurrence database 3d, and if the analyzed combination of the dependency word and the received word is stored in the co-occurrence database 3d, the meaning of the word is specified. .

Next, in step S43, it is determined whether or not the ambiguity of all the words in the query has been resolved. If the ambiguity has been resolved, the process immediately proceeds to step S45. If not, the process proceeds to step S44. Then, the disambiguation is performed based on the user profile.
Go to 5.

More specifically, in step S44, the cosine measure SD (X, Q) of the word vector X indicating all the meanings of the words described in the query that has not been disambiguated and the normalized user profile vector Q Is calculated, and the cosine measure SD (X, Q) is calculated as the similarity. That is, the word vector X may be selected from only one meaning or may be plural. Generally, the word vector X is represented by an n-dimensional (x1 to xn) vector as shown in Expression (1). Is done. Similarly, the normalized user profile vector Q is represented by an n-dimensional (q1 to qn) vector as shown in Expression (2), and the cosine measure SD (X, Q) is obtained by calculating the inner product of both vectors by the absolute value of both vectors. The value is divided by the product of the values. However, since the word vector X and the normalized user profile vector Q are normalized based on “1”, the cosine measure SD (X, Q) corresponds to the inner product, and therefore, the cosine measure SD (X, Q) Is the sum of the products of the same-dimensional features of both vectors, as shown in equation (3).

[0074]

(Equation 1)

Thus, the cosine measure SD (X, Q),
That is, by calculating the similarity, selecting a meaning that is similar to a certain threshold or more and excluding a meaning that is interpreted as irrelevant,
Eliminate ambiguity by user profile.

As described above, in the step S42, the ambiguity is eliminated by the user profile for the word that could not be eliminated by the co-occurrence database. What should be emphasized here is that polysemy elimination is performed by co-occurrence information prior to polysemy elimination by the user profile.

Then, in step S45, the query is expanded using the query expansion dictionary 3e. That is, according to the user's instruction, variation processing such as “only synonym expansion”, “synonym expansion + related word expansion”, and “only related word expansion” is performed.

Next, in step S46, document vector generation processing of a search query is performed. In other words, the word vector dictionary 3a is searched from the word extracted in the processing up to this point and the specified meaning, the feature amount for each dimension is calculated for each word, and a document vector is generated from the sum thereof, and the routine shown in FIG. Return to

Next, in step S33 of FIG. 11, the cosine measure SD (X ', Q'), that is, the query vector obtained in step S32, Q 'and the document vector X' of the document database 3b to be searched, Calculate the similarity,
Store in RAM7.

FIG. 13 is a flowchart of the similarity generation processing routine executed in step S33.

That is, in step S51, the count value N of the counter for designating the document to be searched in the document database 3b is set to the initial value 1, and the subsequent step S51
At 52, the document database 3b is searched, and the document vector X 'of the Nth (N = 1 in the first loop) document is read. At step S53, the Nth document vector X' and the query vector Q 'of the search query are converted. A similarity is calculated based on the similarity. That is, in step S53, the search query is regarded as one document to obtain a query vector Q ', and the cosine measure SD (X', X ',) of the document vector X' of the document on the search target document database 3b and the query vector Q 'is obtained.
Q ′) and calculate the similarity (formula (4) below)
(6)).

[0082]

(Equation 2)

For example, when the notation “form” is interpreted according to the two profiles shown in FIGS. 5A and 5B, according to the profile shown in FIG. 5A, the feature of dimension 7 is large. In the words 7 and 8 in FIG. 3, the inner product of the word 8 (“form (posture)”) having a large dimension 7 is increased, and the meaning of “posture” is adopted.

On the other hand, according to the profile of FIG. 5B, since the feature of dimension 4 is large, the word 7 and the word 8 of FIG.
(Form) ”and the meaning of“ form ”is adopted.

Next, in step S54, the calculated similarity is stored in the RAM 7, and in the following step S55, it is determined whether or not there is a remaining document in the document to be searched in the document database 3b. While the process returns to the routine of No. 11, if it exists, the count value N of the counter is incremented by "1" in step S56, the process returns to step S52, and the above process is repeated.

In the present embodiment, the similarity is stored in RAM
7, but may be registered in the document database 3b.

Next, returning to the routine of FIG. 11, in step S34, the similarity degree for each document obtained in step S33 is ordered with reference to the RAM 7. In step S35, the documents ordered in step S34 are listed as search results and displayed on the display device 2.
At this time, the value of the similarity registered in step S33 is also displayed.

As described above, according to the present embodiment, by analyzing a query according to the co-occurrence database 3d,
The meaning of the query can be interpreted accurately, whereby the ambiguity of the query can be eliminated with high accuracy. If the ambiguity can not be eliminated based on the co-occurrence database 3d, the user profile information As a result, the sense of ambiguity can be resolved according to the user's past preferences, so that a search result with a higher search accuracy that is more suited to the user's intention can be obtained easily and quickly.

FIG. 14 is a block diagram showing a second embodiment of a document search apparatus as an information search apparatus according to the present invention. In the second embodiment, a case where a query is not extended is described. And the query expansion dictionary is omitted.

FIG. 15 is a flowchart of a query vector generation processing routine according to the second embodiment. In step S61, words are extracted from a user-specified search query, and morphological analysis is performed using a morphological analysis dictionary. Perform the dependency analysis. In the following step S62, all the results of the dependency analysis are compared with the co-occurrence database 3d, and the meaning of each meaning is determined according to whether or not a set of the analyzed dependency word and the received word is stored in the co-occurrence database 3d. The likelihood (word sense likelihood) is calculated.

Next, in step S63, similar to the first embodiment, the similarity based on the cosine measure is calculated, and the word meaning likelihood is calculated according to the similarity. And step S6
2. The word likelihoods calculated in S63 are summed up, and finally the most likely word meaning is selected. At this time, by increasing the weight of the semantic likelihood in S62 as the weight of the semantic likelihood, the disambiguation by co-occurrence information is prioritized over the disambiguation by the profile.

In step S64, document vector generation processing of the search query is performed. That is, the word vector dictionary 3a is searched from the word extracted in the processing up to this point and the specified meaning, a feature amount for each dimension of each word is calculated, and a document vector is generated from the sum thereof.

As described above, this embodiment shows an example in which query expansion is not performed, and shows another embodiment in which polysemy resolution based on a co-occurrence database is prioritized over polysemy resolution based on user profile information. I have.

The present invention is not limited to the above embodiment.

In the above embodiment, similar document retrieval has been described as a document retrieval method, but the present invention can be applied to other retrieval methods. For example, in a conventional full-text search system, a query may be expanded into a synonym, but according to the present invention, since the query is analyzed on a synonym basis, it is more easily required by a synonym dictionary based on the semantics as shown in FIG. Since only similar synonyms can be expanded, and expanded words are limited to synonyms relating to a specific meaning, search noise can be reduced.

In the above embodiment, the user profile is updated every time a file is accessed. However, only the history information of the accessed file is recorded, and the profile is collectively updated every predetermined period. Is preferably updated. In this case, the file can be accessed comfortably without extra processing time for each file access.
The profile can be updated collectively at midnight when a normal computer is not used, and in this case, the influence of the update of the profile on the normal user layer in processing time becomes smaller.

Although the user profile is configured from the history of accessed files, the user may directly create the user profile.

[0098] In some cases, it is difficult for a user to directly create user profile information from the beginning. Therefore, a certain type of guideline may be created and a profile may be created simply by responding to system questions. Good.

Further, in the above-described embodiment, the user profile is configured so that each individual has different information. However, in a company organization, etc., it is conceivable that each individual group has a profile. The profile information is updated by file access by another member in the group.

In the above-described embodiment, as a method for resolving polysemy of a meaning, a method using a co-occurrence database,
Although two types of methods based on the user profile have been described, the disambiguation means is not limited to these two types. For example, a meaning list of each word may be displayed, and the user may select a desired meaning from the meaning list,
The meaning may be interpreted according to the context. That is, for many queries issued by the user, a context vector is prepared, and the context vector is updated for each query input, so that even if an important word is omitted in a new query input, a search as intended by the user is performed. It is possible to carry out such a method, and such a polysemy elimination method may be mixed.

[0101]

As described above in detail, according to the present invention, in a similar document search using a vector space model, a search query is analyzed according to co-occurrence information and ambiguity is resolved, so that the query can be interpreted accurately, and Information retrieval with high retrieval accuracy according to the intention can be performed.

Also, since polysemy in the search query is resolved by referring to the user profile information expressing the user's past preference, the query can be interpreted more in accordance with the user's preference, and the user's intention can be interpreted. Information retrieval with high retrieval accuracy can be performed. In addition, since disambiguation by co-occurrence information is prioritized over disambiguation by user profile information, it is possible to perform information retrieval with higher retrieval accuracy in accordance with the intention of the user.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment (first embodiment) of a document search device as an information search device according to the present invention.

FIG. 2 is a diagram illustrating an example of a display screen of a display device.

FIG. 3 is a format diagram of a word vector dictionary.

FIG. 4 is a format diagram of a document database.

FIG. 5 is a format diagram of a user profile.

FIG. 6 is a format diagram of a co-occurrence database.

FIG. 7 is a format diagram of a query expansion dictionary.

FIG. 8 is a flowchart of a main routine showing a search procedure of a document search method as an information search method according to the present invention.

FIG. 9 is a flowchart of a document registration processing routine.

FIG. 10 is a flowchart of a profile update processing routine.

FIG. 11 is a flowchart of a search execution processing routine.

FIG. 12 is a flowchart of a query vector generation processing routine.

FIG. 13 is a flowchart of a similarity generation processing routine.

FIG. 14 is a block diagram showing a second embodiment of a document search device as an information search device according to the present invention.

FIG. 15 is a flowchart of a query vector generation processing routine according to the second embodiment.

[Explanation of symbols]

 1 input device 3 HD 9 CPU

Claims (27)

[Claims] A search condition input unit for inputting a search condition; a morphological analysis unit for extracting a word from the search condition input by the search condition input unit; A co-occurrence relation storage means for storing;
A first polysemy removing means for extracting a co-occurrence relation of the search condition based on the co-occurrence information stored in the co-occurrence relation storage means and the analysis result of the morphological analysis means and eliminating polysemy;
An information search device for performing an information search based on the meaning of which the ambiguity has been eliminated by the first polysemy elimination device. 2. The information retrieval apparatus according to claim 1, further comprising a meaning extension means for performing an extension process on a meaning whose meaning has been eliminated by said first ambiguity eliminating means. 3. The meaning expansion means includes at least synonym expansion means for expanding the meanings into synonyms, related word expansion means for expanding related words, and combination expansion means combining these. The information retrieval device according to claim 2. 4. A user profile storage means for storing user profile information expressing user preferences,
A second ambiguity removing unit for removing ambiguity of the search condition based on the user profile information stored in the user profile storage unit and an analysis result of the morphological analysis unit; The cancellation means is the second
4. The information retrieval apparatus according to claim 1, wherein the information retrieval apparatus has priority over the polysemy elimination means. 5. The information retrieval apparatus according to claim 4, further comprising a meaning extension means for performing an extension process on a meaning whose ambiguity has been eliminated by said second polysemy elimination means. 6. The synonym expansion means includes at least synonym expansion means for expanding synonyms of meanings, related word expansion means for expanding related words, and combination expansion means combining these. The information retrieval device according to claim 5. 7. A document vector generating means for generating a document vector based on the meaning of which the ambiguity has been eliminated by the first ambiguity eliminating means, and the document vector generated by the document vector generating means is registered as a database. 7. The information search device according to claim 1, further comprising a registration unit that performs the registration. 8. A document vector generating means for generating a document vector based on the meaning of which the ambiguity has been eliminated by the first ambiguity eliminating means, and a user profile storing user profile information expressing user preferences. 8. A user profile information updating unit for updating a storage content of a storage unit based on a document vector generated by the document vector generating unit. Information retrieval device. 9. Search condition input means for inputting search conditions,
A morphological analysis unit for extracting a word from the search condition input by the search condition input unit, a user profile storage unit for storing user profile information expressing a user's preference, and the user profile information and the morphological analysis unit. Second ambiguity eliminating means for eliminating the ambiguity of the search condition based on the analysis result, and information retrieval means for performing information retrieval based on the meaning of the ambiguity eliminated by the second ambiguity eliminating means An information retrieval device comprising: 10. A search condition inputting step of inputting a search condition, a morphological analysis step of extracting a word from the search condition input in the search condition inputting step, and associating a co-occurrence relationship between a plurality of words with a meaning. A first polysemy elimination step for extracting a co-occurrence relation of the search condition based on the stored co-occurrence information and the analysis result in the morphological analysis step to eliminate polysemy, and the first polysemy elimination An information search step of performing an information search based on a meaning that has been eliminated in the step. 11. The information retrieval method according to claim 10, further comprising a meaning extension step of performing an extension process on the meaning whose ambiguity has been eliminated in said first ambiguity eliminating step. 12. The semantic expansion step includes at least a synonym expansion step of expanding a meaning into a synonym, a related word expansion step of expanding a related word, and a combination expansion step in which these are combined. The information retrieval method according to claim 11. 13. The method according to claim 1, further comprising a second polysemy resolving step for resolving polysemy of the search condition based on user profile information expressing user preference and an analysis result of the morphological analysis step. 13. The information search method according to claim 10, wherein the disambiguation step has priority over the second disambiguation step. 14. The information retrieval method according to claim 13, further comprising a semantic extension step of performing an extension process on the semantics whose ambiguity has been eliminated in the second ambiguity eliminating step. 15. The semantic expansion step includes at least a synonym expansion step of expanding the meaning into a synonym, a related word expansion step of expanding the meaning into a related word, and a combination expansion step combining these. The information retrieval device according to claim 14. 16. A document vector generating step of generating a document vector based on the meaning of which the ambiguity has been eliminated in the first ambiguity eliminating step, and registration for registering the generated document vector as a database in a registration unit. 16. The information search method according to claim 10, further comprising the steps of: 17. A document vector generating step of generating a document vector based on a meaning removed from the sense of ambiguity in the first ambiguity removing step, and a user profile storing user profile information expressing a user's preference. 17. The information retrieval method according to claim 10, further comprising a user profile information updating step of updating contents stored in a storage unit based on the generated document vector. 18. A search condition input step of inputting search conditions, a morphological analysis step of extracting words from the search conditions input by the search condition input means, and user profile information expressing user preferences are stored. A second ambiguity eliminating step for eliminating the ambiguity of the search condition based on the user profile information and the analysis result of the morphological analysis step, and a word having the ambiguity eliminated by the second ambiguity eliminating step. An information search step of performing an information search based on the information search method. 19. A computer-readable storage medium storing an information search procedure for performing an information search based on a search condition input from an input device, wherein the morphological analysis extracts a word from the input search condition. And extracting a co-occurrence relationship of the search condition based on co-occurrence information stored in association with a co-occurrence relationship between a plurality of words and a meaning and an analysis result of the morphological analysis procedure to eliminate polysemy. A computer-readable storage storing a first ambiguity eliminating procedure and an information retrieval procedure for performing information retrieval based on the meaning of the ambiguity eliminated by the first ambiguity eliminating procedure. Medium. 20. The storage medium according to claim 19, wherein a semantic extension procedure for performing an extension process on the semantics whose ambiguity has been eliminated in the first ambiguity eliminating procedure is stored. 21. A method of eliminating the ambiguity of the search condition based on user profile information expressing user's preference and an analysis result of the dependency analysis step, which is prioritized over the first ambiguity eliminating procedure. 21. The storage medium according to claim 19, wherein the second ambiguity resolution procedure is stored. 22. The storage medium according to claim 21, wherein a semantic extension procedure for performing an extension process on the semantics whose ambiguity has been eliminated in the second ambiguity eliminating procedure is stored. 23. A document vector generating procedure for generating a document vector based on the meaning of which the ambiguity has been eliminated in the first ambiguity eliminating procedure, and registration for registering the generated document vector as a database in a registration unit. 23. The storage medium according to claim 19, wherein the procedure is stored. 24. A document vector generating procedure for generating a document vector based on the meaning of which the ambiguity has been eliminated in the first ambiguity eliminating procedure, and a user profile storing user profile information expressing user preferences. 24. A user profile information updating step of updating the storage content of the storage unit based on the generated document vector.
The storage medium according to any one of the above. 25. A co-occurrence relationship between a word vector dictionary in which word meaning vectors are stored on a semantic basis, a document database to be searched, a user profile in which user preferences are recorded, and dependency words are stored. A computer-readable storage medium having a data structure including a co-occurrence database. 26. The storage medium according to claim 25, wherein an expansion dictionary for expanding a search condition is stored. 27. A computer-readable storage medium storing an information search procedure for performing an information search based on a search condition input from an input device, wherein the morphological analysis extracts a word from the input search condition. A second ambiguity eliminating procedure for eliminating the ambiguity of the search condition based on user profile information in which user profile information expressing user preference is stored and an analysis result of the morphological analysis procedure; A computer-readable storage medium storing an information search procedure for performing an information search based on a meaning that has been resolved by the second polysemy resolution procedure.