JP2003162542A - Information retrieval device and patent information retrieval device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an information retrieval device for retrieving a document similar to a document for which retrieval is requested, and to provide a patent information retrieval device particularly suitable when a document of retrieval target is a patent document. <P>SOLUTION: The information retrieval device has a keyword automatic extraction part 2 that extracts by the semantic role of a keyword included in the document for which retrieval is requested, and a retrieval part 4 that retrieves a document similar to the document for which retrieval is requested from a database 6 storing retrieval target documents by a retrieval conditional expression based on the semantic role of the keyword extracted with the keyword automatic extraction part 2. <P>COPYRIGHT: (C)2003,JPO

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 for inputting a sentence described in natural language as a retrieval request and outputting a document similar to the retrieval request as a retrieval result, and particularly to a retrieval object. The present invention relates to a patent information search device suitable when a document is a patent document.

[0002]

2. Description of the Related Art In order to perform a search, a conventional information search apparatus holds a document in a database in which a certain keyword appears as index information, obtains a search result for a search request using the index information, The method of providing it to the user was taken.

However, in order to create the index information, it is necessary to obtain information such as statistical information of keywords from the documents in the database, which requires a lot of time. Therefore, a large amount of labor is required to operate the search device, and there is a problem that a document newly registered in the database cannot be searched until the index information is reconstructed.

A conventional patent information retrieval device is a patent electronic library (I) provided by the JPO on the Internet.
It is common to employ a keyword search such as PDL). In this case, since the semantic relationship of the keywords included in the search request cannot be expressed in the search program, the means indicated by the keywords in the search target document and the purpose of use of the substance in the case of the substance invention and It was difficult to perform a search considering the usage method. for that reason,
Even when the search keyword included in the search target document is used for another purpose different from the purpose, there is a problem that the document including the search keyword is output to the search result.

Further, by extracting structure information represented by a case frame or the like using the structure information of the sentence and matching the structure extracted from the sentence appearing in the retrieval target document at the structure level, Attempts have also been made to perform searches that take into consideration the purpose and usage of keywords, but for that purpose it was necessary to equip the search engine with an advanced mechanism for performing structure-level matching.

[0006]

Therefore, the present invention is
In view of the problems of the conventional search device as described above, a keyword automatic extraction unit that extracts keywords from a search request document by their semantic roles and a search means by keyword search based on the semantic roles of the keywords are adopted. However, an object of the present invention is to provide an information search device for searching a document similar to the search request document, and also to provide a patent information search device suitable when the document to be searched is a patent document in particular. Has an aim.

Further, by using a massively parallel computer, there is provided an information retrieval device and a patent information retrieval device capable of retrieving similar documents with high accuracy without indexing retrieval target documents in a database. Is intended.

[0008]

In order to achieve the above object, the invention according to claim 1 inputs a sentence described in natural language as a search request document and sets a document similar to the search request as a search result. An information retrieval device for outputting, comprising a keyword automatic extraction unit for extracting a keyword included in the retrieval request document for each semantic role of the keyword, and a semantic role of the keyword extracted by the keyword automatic extraction unit. And a search unit that searches a document similar to the search request document from a database that stores search target documents according to a search condition expression based on the search condition document.

According to a second aspect of the present invention, in addition to the configuration of the first aspect, the search unit sets a plurality of different search condition expressions on a plurality of different processors of a massively parallel computer, and the search target document is set. Is simultaneously and concurrently searched for in full text with the plurality of different search condition expressions, and a result matching the plurality of search condition expressions is output.

According to a third aspect of the present invention, the keyword automatic extraction unit of the information retrieval apparatus according to the first aspect statistically processes the surface expression of the language of the retrieval request document, the morpheme information and the retrieval target document. It is characterized in that the information is used to automatically extract and classify keywords according to the semantic role in the search request document of the noun representing the subject of the invention and the means for realizing the object of the invention.

According to a fourth aspect of the present invention, in addition to the configuration of the third aspect, the keyword automatic extraction unit uses the morpheme information forming a keyword and statistical information obtained by statistically processing the search target document, It is characterized in that the boundary of a keyword consisting of a noun representing an object of the invention used for narrowing down search target documents is determined. Here, the noun is a broad concept that includes not only a "noun" in a narrow sense but also a "compound noun".

According to a fifth aspect of the invention, in addition to the structure of the third aspect, the keyword automatic extraction unit uses the morpheme information forming the keyword and the statistical information obtained by statistically processing the search target document to perform a search. It is characterized in that the boundaries of keywords that represent means for achieving the object of the invention used for scoring results are determined.

According to a sixth aspect of the invention, in addition to the configuration of the first aspect, the search unit narrows down the search target document with a search keyword consisting of a noun representing the object of the invention, and the narrowed down search is performed. A feature of the present invention is that a search result by a search keyword representing a means for realizing the object of the invention for a target document is ranked based on scoring.

According to a seventh aspect of the present invention, in addition to the configuration of the sixth aspect, the search unit represents a keyword for searching that represents means for realizing the object of the invention by statistical information obtained by statistically processing the document to be searched. It is characterized in that scoring based on the co-occurrence of the above and automatic scoring based on the single appearance of the search keyword are performed.

The invention described in claim 8 is the invention according to claims 3 to 7.
In addition to the configuration described in any one of,
A plurality of different search condition expressions are set on a plurality of different processors of a massively parallel computer, the full-text search of the search target document is performed in parallel with the plurality of different search condition expressions, and the plurality of search condition expressions are matched. It is characterized by outputting the result.

[0016]

BEST MODE FOR CARRYING OUT THE INVENTION A patent information retrieval device according to an embodiment of the present invention will be described in detail below with reference to the drawings.

FIG. 1 is a block diagram of a patent information search device according to an embodiment of the present invention.

Considering that the invention is a technical idea for providing a specific means for achieving a specific object, when analyzing a patent document, the invention described therein is grasped by the object and the means. This is very important.

That is, a document having a noun (including a compound noun) that is the same as or similar to a noun (including a compound noun) representing the "object" of the invention is used for the purpose of searching for a similar patent document in the patent information search device. It is considered that this can be achieved by searching documents that have the same or similar "means" for realizing the object of the invention and "specific substance name" in the case of substance invention. To that end, the main subject of the invention is
Represents a keyword consisting of nouns (including compound nouns) (hereinafter referred to as "target" keyword), "means" for realizing the object of the invention, and "specific substance name" in the case of a substance invention. It has a function to classify and extract keywords (hereinafter referred to as "means" keywords) according to their semantic roles in the document, and narrows down search target documents using the keywords extracted according to roles and "means" keywords. It is desirable to have a means for scoring and ranking.

Further, the "target" keyword and "means" keyword have morphological information and the statistical information obtained by statistically processing the search target document has a function capable of determining an appropriate keyword boundary, and the search target document at the time of search. It is desirable to have a means for automatically determining scoring based on the co-occurrence of the "means" keyword and scoring based on the single occurrence using the statistical information obtained by statistically processing.

In FIG. 1, the input document analysis unit 1 is composed of a morpheme analysis unit 11 and a division unit 12 into sections or clauses. The morpheme analysis unit 11 divides the search request document into morpheme units and acquires the POS information of the morpheme. Also,
The division unit 12 into clauses or clauses uses the morpheme information obtained by the morpheme analysis unit 11 to obtain information on clauses and clause breaks from the search request document.

The bunsetsu demarcation position is a bunsetsu demarcation position immediately after a particle, a conjunction or the like, or immediately after a continuation of a particle or a connection particle. A generally known method such as setting the section delimitation position immediately after is used.

In FIG. 1, the keyword automatic extraction unit 2
Is a keyword candidate extraction unit 21, a “target” keyword selection unit 22, a “target” keyword boundary determination unit 23,
It is composed of a "means" keyword selection unit 24 and a "means" keyword boundary determination unit 25.

In the keyword candidate extraction unit 21,
As keyword candidates, for example, morphemes predicted to represent the characteristics of the document such as nouns, compound nouns, and unknown words not registered in the morphological analysis system are designated, and the keyword candidates are extracted.

The compound noun is created, for example, by processing a series of nouns into a compound noun. However, it is necessary to save the morpheme information that composes the compound noun for the boundary determination processing of the extracted keyword.

Further, after extracting the keyword candidates, a word or the like to be excluded from the keywords may be retained as an exclusion list, and a function of excluding the keyword candidates appearing in the exclusion list may be added. This exclusion list includes
In addition to the word that is an exclusion condition that exactly matches the keyword candidate, a word that is an exclusion target may be included if it partially matches.

The "target" keyword selection unit 22 selects "target" keyword candidates from the keyword candidates extracted by the keyword candidate extraction unit 21. For example, in the case of a claim in which the input document is described in a so-called Jepson format, which is "characterized in that", the patent application of the invention filed before "before" A general description of the "subject" is described. Furthermore, the specific part of the “object” of the invention to be applied for a patent is limited as a subject in the part after “in”, and a specific method of realizing the “object” of the subject invention is described. ing. Therefore, among the keyword candidates extracted by the keyword candidate extraction unit 21, “leave”
A keyword candidate consisting of all the keyword candidates appearing in the front part of and the nouns (including compound nouns) representing the subject explained in the part after the "leave" is extracted as the "target" keyword candidate.

[0028] The subject explained after "Aki" is
For example, it can be extracted by focusing on particles, and the particles "ha", case particles "ga", case particles "as", and case particles "wa" (however, Just before the preceding "", the keyword candidate that appears immediately before the above morpheme in the priority order of case particle + verb "consisting of" is 1
Just extract one. However, among the extracted “target” keyword candidates, the “target” keyword candidates having a high ratio of the number of documents that include each keyword in the search target documents with respect to the total number of search target documents have a low ability to narrow down the search target documents. , Excluded from "target" keyword candidates.

In order to obtain the number of documents including keywords in the documents to be searched, search condition expressions corresponding to the respective keyword candidates are created, and these search condition expressions are provided in a massively parallel manner having a plurality of different processors. It is set on a plurality of different processors of the computer 5 at the same time, the document group stored in the database 6 in which the document to be searched is stored is subjected to full-text search by a plurality of search condition expressions, and the result matching the search condition expression is used. be able to. Here, the number of results that match each search condition expression is the number of documents in which each keyword candidate is included in the search target documents. Since the statistical information is updated by performing a full-text search every time the processing of the keyword extracting unit 2 is performed, the accuracy of the statistical information can be maintained.

Here, the massively parallel computer 5 incorporates several thousands to tens of thousands of processors (hereinafter collectively referred to as a pipeline), so that a plurality of different search condition expressions are simultaneously set in this pipeline. It is possible. Then, by operating a large number of these processors at the same time, a plurality of different search condition expressions and database 6
Perform a full-text search that matches search target documents saved in. As a result of matching, if a document that matches the search condition expression is found, it has a function of assuming that the document has been searched.

The massively parallel computer 5 is a full-text search engine (for example, FDF (registered trademark) manufactured by Paracel).
A device such as a 4T TextFinder) is desirable, but a device such as a workstation having a function and performance equivalent to this may be used.

The "target" keyword boundary determining unit 23 determines the boundary of the "target" keyword of a compound word such as a compound noun among the "target" keyword candidates extracted by the "target" keyword selecting unit 22. . The morphemes that compose the compound words are divided, and the co-occurrence is set at the same time when the boundaries of the keywords having the appropriate search target document narrowing ability are determined. Boundary determination and co-occurrence settings are performed using a compound word, a co-occurrence of divided morphemes, and a ratio of the number of documents each of which is included in the search target document to the total number of search target documents for each of the divided morphemes. It is performed by co-occurrence of the divided morphemes and selection of the divided morphemes. The procedure for determining the boundary of the “target” keyword is shown in the flowchart of FIG. 2, and details will be described later. Note that the “target” keyword candidates that are not the target of boundary determination are directly selected as the “target” keyword.

The "means" keyword selection unit 24 selects "means" keyword candidates from the keyword candidates output from the keyword candidate extraction unit 21. For example, in the case of a claim in which the input document is described in the so-called Jepson format, which is characterized in that "in ~,", the invention to be applied for patent in the portion after "preceding" The specific part of the "object" is limited as a subject, and the "object" of the subject invention is specifically realized, that is, "means" for realizing the substance or the "specific" used in the case of the invention. The substance name is described. Therefore, the keyword candidates appearing in the portion after "leave" are extracted as "means" keyword candidates. When extracting "means" keyword candidates, delete the keyword candidate immediately before the case particle "no", which is considered to be a complementary modification of other "means" keyword candidates in order to improve the precision. Alternatively, in order to delete more modified contents, the “means” keyword candidates may be limited to the keyword candidates that appear at the end of the section.

The "means" keyword boundary determining unit 25 determines the boundaries of the "means" keywords of compound words such as compound nouns among the "means" keyword candidates extracted by the "means" keyword selecting unit 24. . Since the "means" keyword is used for matching after the search target is limited by the "target" keyword, the search target documents have already been narrowed down, and it is important to rank more similar documents from the score. . Therefore, by using a boundary determination method, an appropriate delimiter position of the keyword is obtained so that more similar documents can be scored and ranked by the keyword "means" of the compound word.

For example, a keyword in which it is clear that "rubbery" is used to describe "sheet" such as "rubbery sheet" appears only in "sheet" as a "means" keyword. May be However, if the ratio of the number of documents that include "sheets" in the search target documents to the total number of search target documents is high, "sheet" is a very general term in the search target documents and the characteristics of the means, etc. Since it can be said that the ability to perform is low, “rubber” that characterizes the “sheet” may be used as the “means” keyword as the keyword having the ability to perform more characterization. In addition, by obtaining the frequency of adjacency of each morpheme that composes a compound word, the strength of the morpheme connection is determined, and the boundary is determined based on the strength of the connection. Techniques may be used.

Further, since the "means" keyword that does not appear in the document to be searched does not play the role of the search keyword, even if the co-occurrence of the morphemes forming the "means" keyword is used for the search, etc. Good. Note that the “means” keyword candidates that are not subject to boundary determination are directly selected as the “means” keyword.

In FIG. 1, the search condition expression creating unit 3 has a search condition expression creating unit 31. The search condition formula creating unit 31 creates a search condition formula using the “target” keyword and the “means” keyword extracted by the keyword automatic extraction unit 2. The “target” keyword is applied to the creation of a search condition expression that narrows down the search target documents. For example, all (and combination) or a part (combination of and combination or or combination) of the “target” keyword is applied. Create a search condition expression to search for documents that appear.

The "means" keyword is applied to the scoring ranking for the search target documents narrowed down by the "target" keyword, and basically, when any of the keywords appears, a score may be given (or combination. ), For example, in a “means” keyword in which the ratio of the number of documents containing each keyword in the search target document to the total number of search target documents is high, the “means” keyword that appears immediately before and immediately after in the search request document It is also possible to increase the relevance ratio by modifying the search condition by using a process such as creating a search condition expression so as to give a score when it co-occurs with (immediately before and after the keyword and combination). . Further, in any of the search condition expressions for narrowing down the search target documents and scoring ranking, synonyms or synonyms for the keyword may be added to the search condition expression using a thesaurus or the like.

In FIG. 1, the search unit 4 has a search execution unit 41. The search execution unit 41 has a function capable of performing two types of search methods, that is, a Boolean search, a score addition by keyword appearance, and a search by score ranking such as similarity score calculation by a vector space method.

The Boolean search and the search based on the scoring ranking may be realized by one search engine or may be realized by different search engines. In the search execution unit 41, first, the search target documents are narrowed down by the “target” keyword by the Boolean search, and the search results are scored by the “means” keyword by the scored ranking. Also, FDF made by Paracel as a search engine
With (registered trademark) 4T TextFinder,
Since it is possible to describe the search target documents by the “target” keyword and the scoring ranking by the “means” keyword in one search condition expression, the search can be performed at the same time. Is possible.

FIG. 2 is a flow chart showing the procedure for determining the boundary of the "target" keyword.

Here, the idf value is a compound word, the co-occurrence of divided morphemes, and the ratio of the number of documents each of which is included in the search target document to the total number of search target documents for each of the divided morphemes is made an inverse number, and log It is a value obtained by taking a value of ₂ , and the larger the value is, the higher the narrowing-down capability of the search target document is.

Therefore, first, the compound word, the co-occurrence of the divided morphemes, and the idf value of each of the divided morphemes are obtained (step S1). If there is a divided morpheme whose idf value is higher than the set threshold value (step S2), and if there is a divided morpheme whose idf value is higher than the set threshold value, all morphemes exceeding the threshold value are targeted. It is adopted as a keyword (step S4). In step S2, there is no divided morpheme exceeding the threshold, and it is determined whether or not a co-occurrence of the divided morphemes whose idf values are higher than the set threshold exists (step S3), and the division whose idf value is higher than the threshold is determined. When the co-occurrence of the morpheme that has been performed exists, the co-occurrence of the divided morpheme is set as the “target” keyword (step S
4). If the other divided morphemes and their co-occurrence are not the “target” keyword, the compound word is set as the “target” keyword (step S4). Step S1 above
The process from to step S4 is performed for all "target" keyword candidates (step S5). The thresholds of the idf value in step S2 and step S3 may be set to different values.

Table 1 shows an example in which the idf threshold of step S2 is set to 5 and the idf threshold of step S3 is set to 4 with respect to the boundary determination program for the "target" keyword of FIG.

[0045]

[Table 1]

In the example shown in Table 1, "thermal recording transfer material" and "back coat layer" are selected as "target" keyword candidates.
When "heating means" is entered, co-occurrence of "transfer material", "back" and "coat layer" (co-occurrence of divided morphemes)
Is extracted as a “target” keyword, and further “heating means” (compound word) is extracted as a “target” keyword.

The "thermosensitive transfer material" is divided into "heat sensitive", "recording" and "transfer material" by the morpheme division. Next, in step S1, "heat sensitive",
Idf values for "recording" and "transfer material" and idf when "heat sensitive", "recording", and "transfer material" co-occur
Find the value. In step S2, when the idf values of “heat sensitive”, “recording”, and “transfer material” are compared with 5, which is the threshold value of idf, “transfer material” having an idf value of 5.10 is regarded as the “target” keyword. Selected, Step S
It will move to 5.

The "back coat layer" is divided into "back" and "coat layer". Next, even if the idf value is obtained and compared with 5 in which the idf values of "back" and "coat layer" are set as idf thresholds in step S2, the idf value of both "back" and "coat layer" is lower than 5. It becomes a value. Therefore, the process proceeds to step S3, and the idf value at the time of co-occurrence of “back” and “coat layer” is set to i at the time of co-occurrence.
Compare with 4 as the df threshold. As a result, idf of co-occurrence of "back" and "coat layer" (co-occurrence of divided morphemes)
Since the value is 4.11, which exceeds the idf threshold, the co-occurrence of “back” and “coat layer” (co-occurrence of divided morphemes) is selected as the “target” keyword, and the process proceeds to step S5. .

The "heating means" is divided into "heating" and "means", but the idf values of both "heating" and "means" are below the idf threshold value, and the "heating" and "means" are the same. The co-occurrence idf value is also below the idf threshold. Therefore, the "heating means" (compound word) before the division is the "target"
The keyword is selected, and the process proceeds to step S5.

[0050]

As described above, according to the invention described in claim 1, the keyword automatic extraction unit for extracting each keyword included in the search request document according to the semantic role, and the keyword automatic extraction unit. The search unit searches for a document similar to the search request document from the database storing the search target document according to the search condition expression based on the semantic role of the selected keyword. Since it is not necessary to make a selection or set a search condition expression, the search work can be efficiently performed. Further, since keywords for each semantic role are used as search keywords, unimportant words are not used for search, so that the matching rate can be further increased. Further, since it is not necessary to reconstruct the index information each time the data in the database is updated, the index information can be searched at any time.

Further, unlike the prior art, since a similar document can be searched only by a search using a keyword without performing a complicated process such as matching at a structure level, an advanced and complicated mechanism can be used as a search engine. Since it is not necessary to provide it, a high-speed search can be realized.

According to the second aspect of the present invention, the search unit sets a plurality of different search condition expressions on a plurality of different processors of the massively parallel computer, and the search target document is set by the plurality of different search condition expressions. Since full-text search is performed simultaneously and in parallel, and the results that match multiple search condition expressions are output, indexing of search target documents in the database is not required, so operation is performed at low cost in addition to the effect of claim 1. It enables high-speed information retrieval.

According to the third aspect of the invention, the keyword automatic extraction unit represents the object of the invention by using the surface expression of the language of the retrieval request document, the morpheme information and the statistical information obtained by statistically processing the retrieval target document. In addition to the effect of claim 1, in addition to the effect of claim 1, the relevance ratio in the search of patent documents can be improved because the keywords are automatically extracted and classified according to the meaning role in the search request document of the noun and the means for achieving the object of the invention. It can be increased.

According to the invention described in claim 4, the keyword automatic extraction unit uses the morpheme information that constitutes the keyword and the statistical information obtained by statistically processing the search target document to narrow down the search target document. Since the boundary of the keyword consisting of the noun representing the object is determined, in addition to the effect of claim 3, the optimum search keyword is selected when collecting patent documents having the same or similar purpose as the invention of the search request document. You can decide.

According to the fifth aspect of the present invention, the keyword automatic extraction unit uses the morpheme information that constitutes the keyword and the statistical information obtained by statistically processing the search target document to score the search results. Since the boundary of the keyword representing the means for realizing the object is determined, it is most suitable for collecting the patent documents having the same or similar means as the means for realizing the object of the invention of the search request document in addition to the effect of claim 3. Search keywords can be determined.

According to the sixth aspect of the present invention, the search unit narrows down the search target documents with the search keyword consisting of a noun representing the object of the invention, and realizes the object of the invention for the narrowed down search target documents. Since the search results by means of the search keyword representing the means are ranked based on the scoring, after performing the primary search for collecting the patent documents having the same or similar noun as the noun representing the subject of the invention of the search request document, Since the secondary search can be performed by collecting the patent documents having the same or similar means as the means for realizing the object of the invention of the search request document, the number of search documents can be adjusted according to the use and the search can be performed. You can search for patent documents that meet your requirements with a high relevance rate.

According to the invention described in claim 7, the search unit uses the statistical information obtained by statistically processing the document to be searched for scoring by co-occurrence of a search keyword representing means for realizing the object of the invention and the search. Since the scoring is automatically determined by the single appearance of the search keyword, the statistical information with the search target document as the original data is reflected in the search condition expression, and in addition to the effect of claim 6, the object of the invention of the search request document is It is possible to search for a patent document having a means that is the same as or similar to the means for realizing it with a high matching rate.

According to the invention described in claim 8, the search section sets a plurality of different search condition expressions on a plurality of different processors of the massively parallel computer, and the search target document is set by the plurality of different search condition expressions. In addition to the effect of any one of claims 3 to 7, since the full-text search is performed simultaneously in parallel and the results matching the plurality of search condition expressions are output, indexing of the search target document in the database is unnecessary. Thus, high-speed patent information search that can be operated at low cost becomes possible.

[Brief description of drawings]

FIG. 1 is a block diagram of a patent information search device according to an embodiment of the present invention.

FIG. 2 is a flowchart showing a procedure for determining a boundary of a “target” keyword.

[Explanation of symbols]

1 Input document analysis unit 2 Keyword automatic extraction unit 3 Search condition formula creation section 4 Search Department 5 Massively parallel computers 6 database 11 Morphological analyzer Division into 12 clauses or clauses 21 Keyword candidate extraction unit 22 "Target" Keyword Selection Section 23 "Target" Keyword Boundary Determiner 24 "Means" Keyword Selection Section 25 "Means" Keyword Boundary Determiner

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Mamoru Matsuo             Mitsubishi Spa, 792 Kamimachiya, Kamakura City, Kanagawa Prefecture             Software Co., Ltd. in Kamakura Division F-term (reference) 5B075 NK31 QS01 QS05

Claims (8)

[Claims] 1. An information search apparatus for inputting a sentence written in natural language as a search request document and outputting a document similar to a search request as a search result, wherein the keyword included in the search request document is A keyword automatic extraction unit for extracting by semantic role, and a search request document from a database storing search target documents by a search condition expression based on the semantic role of the keyword extracted by the keyword automatic extraction unit An information retrieval apparatus comprising: a retrieval unit that retrieves similar documents. 2. The search unit sets a plurality of different search condition expressions on a plurality of different processors of a massively parallel computer,
The information retrieval apparatus according to claim 1, wherein the retrieval target document is simultaneously and concurrently full-text searched by the plurality of different retrieval condition expressions, and a result matching the plurality of retrieval condition expressions is output. 3. The keyword automatic extraction unit of the information search device according to claim 1, using the surface expression of the language of the search request document, the morpheme information, and the statistical information obtained by statistically processing the search target document. A patent information retrieval device characterized by automatically extracting keywords and classifying roles according to semantic roles in the search request document of a noun representing an object and means for realizing the object of the invention. 4. The keyword automatic extraction unit is composed of a noun representing an object of the invention used for narrowing down the search target document by using morpheme information forming a keyword and statistical information obtained by statistically processing the search target document. The patent information search device according to claim 3, wherein a boundary between keywords is determined. 5. The automatic keyword extraction unit represents means for realizing an object of the invention used for scoring a search result by using morpheme information forming a keyword and statistical information obtained by statistically processing the search target document. The patent information search device according to claim 3, wherein a boundary between keywords is determined. 6. The information retrieval apparatus according to claim 1, wherein the retrieval unit narrows down the retrieval target documents with a retrieval keyword composed of a noun representing an object of the invention, and the retrieval target documents narrowed down have an object of the invention. A patent information retrieving apparatus, which performs a ranking based on a scoring of a retrieval result by a retrieval keyword representing a means for realizing. 7. The scoring by co-occurrence of a search keyword, which represents means for realizing the object of the invention, and score by the single appearance of the search keyword, based on statistical information obtained by statistically processing the search target document. The patent information search device according to claim 6, wherein automatic determination of attachment is performed. 8. The search unit sets a plurality of different search condition expressions on a plurality of different processors of a massively parallel computer,
8. The full-text search of the document to be searched by the plurality of different search condition expressions simultaneously and in parallel, and a result that matches the plurality of search condition expressions is output. Patent information retrieval device described.