JP2006344011A - Multilingual document retrieving device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To precisely perform multilingual document retrieval for performing document retrieval from retrieval conditions described in different languages by removing or reducing any influence due to the performance characteristics of a translation means. <P>SOLUTION: Retrieval conditions in first language are input from an input means 4, and respective translation devices 1a, 1b and 1c translate the retrieval conditions in a first language into those in a second language which is the same as that of a retrieval object document, and an integration means 2 searches for the word vectors of translated three translation results(retrieval conditions), and integrates those three word vectors into one word vector. The word vectors are preliminarily added to each retrieval object document, and a retrieval means 3 outputs the pertinent document as a retrieval result from an output means 5 based on the fact that the integrated word vectors of the retrieval conditions and the object document word vectors have a prescribed similar relation. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a technique for performing a document search using a search condition described in a language different from that of a search target document, and in particular, performs a translation process for making the search target document and the search condition the same language by a plurality of translation means. Regarding technology.

When searching for documents that match the search conditions, various types of search conditions such as logical expressions combined with search terms (keywords) and search query sentences are used as the search conditions.
Such a document search is usually based on the premise that the language describing the search condition and the language describing the document to be searched are the same. Such requests are increasing rapidly.

(1) I want to search for a document in another language (second language) that matches the search condition described in a certain language (first language) semantically and contentally.
(2) A search condition described in a certain language (first language) is desired to be searched across documents including documents described in another language (second language).
A search technique for responding to such a request is called “multilingual information retrieval (CLIR)”.

In order to respond to such a request, a technique has been proposed in which a search query is translated into another language and then a document search is performed (see Patent Document 1).
However, in the method of performing the search after simply translating the search question into another language, the following problems arise, for example, depending on the performance of the search depending on the quality of the bilingual dictionary and the translation device. .

(1) The translation device is still incomplete technology, and there are strengths and weaknesses regarding the description contents and syntax of sentences depending on individual algorithms and translation dictionaries to be used. At present, there is no universal translation device that covers everything.
(2) The correspondence between words in different languages is not one-to-one, and the words have ambiguity. If this is translated incorrectly, there is a problem that the recall of the search is reduced.
(3) In general, there are a plurality of words representing one concept. If such a notation or expression blur is not dealt with, the recall rate of the search will decrease.
(4) When expanding from one word to a plurality of words based on all ambiguities, there is a problem that the relevance rate of the search is lowered due to inappropriately expanded words.

As a technique that can deal with such problems, assuming that there is a directory (translated directory found on the Internet) that maintains the translation relationship between languages, first, a directory in the first language similar to the search query is created. A method to achieve high-precision search by limiting the documents to be searched by searching and specifying directories in other languages that have correspondence as search ranges and using the correspondence of translated directories Has been proposed (see Patent Document 2).
Also,
Assuming the existence of a bilingual document group (associated with which document the document is a translation of), the first language document is searched for the first language search question, and the search result document A method has been proposed in which a document in the second language associated with is acquired and the search is performed again using the document as a search question (see Patent Document 3).

As techniques applicable to the present invention described below, the following are known.
A technique is known in which a natural language word is not used as an axis of a multidimensional space, but a mathematically extracted component such as Latent Semantic Indexing (LSI) is used as an axis (see Non-Patent Document 1). ).
Techniques relating to weighting of vector space models and vector elements are known (see Non-Patent Document 2).
A search method based on a probability model is known (see Non-Patent Document 3).
JP-A-5-324719 JP 2002-189745 A Japanese Patent Laid-Open No. 2003-196309 S. Deerwester, ST Dumais, GW Furnas, TK Landauer, R. Harshman: Indexing by latent semantic analysis, Journal of the American Society of Information Science, Vol. 41, No. 6, pp. 391-407, 1990. G. Salton and MJ McGill: Introduction to modern information retrieval, McGraw-Hill, 1983. Tokunaga Takenobu: Information Retrieval and Language Processing, Language and Calculation 5, The University of Tokyo Press, 1999.

The methods described in Patent Documents 2 and 3 are good examples for realizing highly accurate multilingual information retrieval (CLIR) by using the translation relationship between directories and documents.
However, directory and document translations rarely exist in practice, and are limited to a few official documents that many people view. Most office documents do not have translations (if any), and it is the case that multilingual documents are mixed. Therefore, it must be said that the existence of translation relations between documents and directories is too strict as a premise for realizing CLIR.

  On the other hand, although the technique described in Patent Document 1 does not assume the existence of a translation relationship between documents and directories, as described above, the appropriateness of the translation itself varies depending on the algorithm of the translation device and the translation dictionary. There is a problem such as a decrease in the relevance rate of the search.

  The present invention has been made in view of the above-described conventional circumstances, and is desired by a user without assuming the existence of a translation relationship between documents and directories and without depending on the performance of the translation device or dictionary as much as possible. An object of the present invention is to realize a highly accurate multilingual document search that enables acquisition of a document to be executed.

The main technical idea of the invention is that a plurality of translations of at least one of a search condition (search query, keyword logical expression, etc.) described in the first language and a search target document described in the second language. A multilingual document search is realized by translating the description into the same language using a device (including a dictionary) and performing a search using these search conditions and the target document.
In an example of a typical aspect of the present invention, a search condition described in a first language is translated into a second language using a plurality of translation devices, and the second language is translated based on a plurality of translation results obtained. A search is performed on the target document group, but not the search condition but the target document group is translated into the first language by using a plurality of translation devices, or the search condition and the target document group are used by a plurality of translation devices. It is also possible to translate into a third language.

In the present invention, by using a plurality of translation devices, the “feature” of each translation device is absorbed, and a highly accurate multilingual document search is realized.
Here, “Kise” of a translation device means a field that is good at translation processing, a field that is not good at translation processing, a syntax that is good at or not good at, vocabulary, description contents of sentences, description format (such as syntax and style), etc. This refers to the performance characteristics of translation devices that depend on
For example, the word “artificial life” in English is translated as “artificial life” in Japanese as a general expression, or it is interpreted as “artificial life” in terms of technical terms. There is a difference in ambiguity such as translating the English word “bank” into “bank” or “bank”. Such performance characteristics depend on the translation dictionary and translation algorithm provided in the translation apparatus.

  The present invention relates to a multilingual document search apparatus that embodies the main technical idea as described above, a program that configures such a multilingual document search apparatus by a computer, and a multilingual document search apparatus that executes the multilingual document search apparatus. This is realized as a language document search method.

  In one aspect of the present invention, the multilingual document search apparatus includes a plurality of translation units, an integration unit, and a search unit, and the plurality of translation units translate at least one of the search condition and the target document, The search condition and the target document are translated into a description in the same language, and the integration unit converts a plurality of translation results from each translation device into feature quantities based on the words included in the search condition or the target document. The search means is integrated with each feature quantity, and the search means searches for the target document from the search condition based on the similarity relation of the integrated feature quantity, so that the second is obtained based on the search condition described in the first language. A document that matches the search condition is searched from among a plurality of target documents written in a language.

  In this way, by integrating the translation results from a plurality of translation devices and performing a search process, it is possible to eliminate or reduce the influence of the performance characteristics of each translation device. Based search processing is realized.

Here, the search condition is generally input to the multilingual document search apparatus in the form of a logical combination expression of keywords or a search question sentence, but in the present invention, other than the logical combination expression of the search question sentence, etc. The various forms may be used.
In addition, as the integration of feature amounts, an average method for obtaining an average value of a plurality of feature amounts, a weighted average method for obtaining a weighted average value of a plurality of feature amounts, a majority method for performing a majority decision operation among a plurality of feature amounts, Various known methods for making a plurality of quantities into a single quantity according to desired conditions, such as a best selection method for selecting the best one from a plurality of feature quantities, can be employed.
The search processing based on the feature quantity and the similarity relation of the feature quantity is, for example, a vector space model (non-patent document) which is a technique for extracting words from sentences according to importance and performing similarity searches by regarding sentences as word sets. Reference 2) can be used.

When the vector space model is used in this way, the search condition or the feature amount of the target document is determined based on a plurality of types of words (for example, the contents of the sentence) set in advance among the words included in the search condition or the target document. A characteristic word that is assumed to represent a typical characteristic), and characterizes the importance of the word in the search condition or target document (for example, the frequency of occurrence of the word in the sentence, the sentence content of the word) It is determined that the similarity relationship between the feature amounts is similar as the cosine between the word vectors is smaller (the matching degree of the search condition is higher).
In the present invention, various known methods can be used as the search processing based on the feature quantity and the similarity relation of the feature quantity. For example, a probability model known in Non-Patent Document 3 can be used.

  In a more specific aspect of the present invention, the multilingual document search apparatus includes a plurality of translation units, an integration unit, and a search unit, and a target document holding unit includes a plurality of target documents. A plurality of translation means translates the search condition into a second language description, respectively, and the integration means includes a plurality of translation results by each translation device. Are converted into feature quantities based on the words to be integrated and integrated into the feature quantities of the search conditions, and the search means includes the feature quantities of the integrated search conditions and the feature quantities of the target document held in the target document holding means. A document that matches the search condition from among a plurality of target documents described in the second language based on the search condition described in the first language by searching the target document from the search condition based on the similarity relationship of Search for.

  In this manner, by integrating the translation results (translated search conditions) obtained by the translation means translating the search conditions into the second language descriptions, the individual translation means The influence of performance characteristics can be eliminated or reduced, and a search process based on an appropriate translation result as compared with the prior art is realized.

  In another more specific aspect of the present invention, the multilingual document search apparatus includes a plurality of translation units, an integration unit, a characterization unit, and a search unit, and the plurality of translation units include a plurality of target documents. Are translated into descriptions in the first language, and the integration means converts the plurality of translation results for each target document by each translation means into feature quantities based on the words that they contain, and And the characterization means converts the search condition into a feature quantity based on the word that it contains, and the search means integrates the feature quantity for each target document and the feature quantity of the search condition. By searching for the target document from the search condition based on the similarity relationship with, the search condition is matched from the plurality of target documents described in the second language based on the search condition described in the first language. Search for documents.

  In this way, a plurality of translation means integrates a plurality of translation results (translated target documents) obtained by translating each target document into a description in the first language, thereby performing search processing. The influence of the performance characteristics of the translation means can be eliminated or reduced, and a search process based on a translation result that is more appropriate than the conventional one is realized.

  In yet another more specific aspect of the present invention, the multilingual document search apparatus includes a plurality of translation units, an integration unit, a characterization unit, and a search unit, and the plurality of translation units include a search condition and Each of the plurality of target documents is translated into a description in the third language, and the integrating means converts the plurality of translation results of the search conditions into feature quantities based on the words included in the search conditions to obtain the feature quantities of the search conditions. Integrated, converted multiple feature results for each target document into feature values based on the words they contain, integrated into feature values for each target document, and search means integrated search conditions A plurality of targets described in the second language based on the search condition described in the first language by searching the target document from the search condition based on the similarity relationship between the feature amount of the target document and the feature amount of the target document Search for documents that meet the search criteria from the documents.

  In this manner, a plurality of translation results obtained by translating the search condition and the target document into a common third language description by a plurality of translation means (translated target documents and translated plural documents) By integrating the search conditions), it is possible to eliminate or reduce the influence of the performance characteristics of the individual translation means, and a search process based on an appropriate translation result can be realized as compared with the prior art. .

  According to the present invention, since the document search is performed from the search conditions described in different languages based on the result of translation by a plurality of translation means, the influence due to the performance characteristics of the translation means can be eliminated or reduced. High-level multilingual document search can be realized.

The present invention will be specifically described based on examples.
FIG. 1 shows the configuration of the main part of a multilingual document search apparatus according to an example of the present invention.
This multilingual document search apparatus is integrated with three translation apparatuses 1a, 1b, 1c, an integration means 2 for obtaining and integrating word vectors as feature quantities from the translation results of the translation apparatuses 1a, 1b, 1c. There is provided search means 3 for searching for a document that matches the search condition based on the similar relationship between the feature quantities (word vectors).

Note that the exemplary multilingual document search apparatus includes three translation apparatuses. However, in the present invention, any aspect including two or more translation apparatuses may be used.
The main functions of the multilingual document search apparatus according to the present invention exemplified here (translation apparatus, integration means, search means, characterization means, etc.) are configured by executing the program according to the present invention on a computer. ing.

As an embodiment of the present invention, three types of multilingual document search processes as outlined in FIG. 2 will be described below.
As a first embodiment, a search question sentence (search condition) q described in a first language is input from the input means 4, and the search question sentence q into which each translation device 1a, 1b, 1c is input is searched. Each of them is translated into a description qh in the same second language as the document, and the integration means 2 obtains word vectors from the three translation results qh output from the respective translation devices, and these three word vectors are converted to 1 by a method described later. Integrate into one word vector. On the other hand, a word vector (feature amount) is calculated and added in advance to each document D to be searched, and the word vector of the search condition integrated with the search means 3 and the target document word vector are predetermined. A target document having a similar relationship is specified from the target document set D, and the specified document is output from the output unit 5 as a search result.

  In the second embodiment, all the search target documents (search target document set) D described in the second language are input from the input unit 4, and the translation devices 1a, 1b, and 1c are input. Each target document is translated into a description Dh of the same first language as the search condition q, and the integration unit 2 obtains a word vector from each of the three translation results qh output from each translation device. It is integrated into one word vector by a method as described later, and one word vector is obtained for each target document. On the other hand, a word vector is calculated for the input search condition q, and a target document in which the word vector of the search condition calculated by the search means 3 and the target document word vector have a predetermined similar relationship is included in the target document set Dh. And the specified document is output from the output means 5 as a search result.

  As a third embodiment, a search question sentence q in the first language is input from the input means 4, and the search question sentence q into which each translation device 1a, 1b, 1c is input is the first and second languages. Is translated into a third language description qh different from, and the integration means 2 obtains word vectors from the three translation results qh output from the respective translation devices, and integrates these three word vectors into one word vector. To do. Also, all D search target documents in the second language are input from the input means 4, and each target document input by each of the translation devices 1a, 1b, 1c is translated into a third language description Dh, respectively, and integrated. The means 2 obtains word vectors from the three translation results qh output from each translation device, and integrates the three word vectors into one word vector to obtain one word vector for each target document. Then, the search unit 3 specifies a target document in which the word vector of the search condition and the word vector of the target document have a predetermined similarity relationship from the target document set Dh, and the specified document is output from the output unit 5 as a search result. Output.

FIG. 3 shows a more specific configuration of the multilingual document search apparatus corresponding to the first embodiment.
In the multilingual document search apparatus according to the present embodiment, a plurality of documents to be searched are stored in the target document database 6 with feature quantities (word vectors) based on words included in each document added.
For example, as shown in FIG. 4, the target document database 6 stores each document in association with a document ID, a document name, a word vector, and a pointer to actual data of the document. Note that the search target document may be document data arranged on a network other than those collected in a database format.

  Further, for example, the word vector of a document indicates, for a plurality of preset characteristic words, a word corresponding to the characteristic word among the words included in the document, and correspondingly indicates the content of the document. This is a word vector whose element value is a degree (weight: for example, appearance frequency, content feature), and three feature words a, b, and c are weights wa, wb, and wc, respectively, as illustrated in FIG. The word vector V of the existing document is expressed as a three-dimensional vector (wa, wb, wc).

The same applies to the search condition and the word vector of the translation result. For a plurality of preset feature words, the word corresponding to the feature word among the words included in the search condition and the translation result is executed. This is a word vector whose element value is a degree indicating the content of the document (weight: for example, appearance frequency, content characteristic).
Note that the word vectors are determined to be similar as the cosine value between the word vectors decreases, but these word vectors have the same row (feature word). For example, the word vector calculated from the translation result of the search condition and the word vector calculated from the target document are vectors having the same feature word as a row.

  In the multilingual document search apparatus according to the present embodiment, as shown in FIG. 5, the search condition (search question sentence) q in the first language is input from the input means 4, and each translation apparatus 1a, 1b, 1c is input. The retrieved query sentences are translated into the same second language as the document to be retrieved, and the integration means 2 uses the three translation results qhA, qhB, and qhC for the search question sentences as a reference for the words included in each translation result. The word vectors VA, VB, VC are obtained, these three word vectors are integrated into one word vector V by a method described later, and the search means 3 searches the target document database 6 to search for the words of the integrated search condition. A document in which the vector and the target document word vector have a predetermined similarity (cosine) is specified, and the specified document is output from the output unit 5 as a search result (for example, a list of document names).

  Here, since the search question sentence translated into the second language is actually used in the search process as a word set, it is not necessarily complete as the second language sentence (that is, human-readable and natural Text)). However, imperfections in the usage and tense of particles are allowed (does not affect search performance), but correct translation (translation) for semantically important words such as function words. Preferably it is done.

Also, each translated search question sentence (document in other embodiments) is extracted and weighted in each search question sentence (document in other embodiments) of the feature word. As a method of determining the weights, various known methods such as word appearance frequency (TF), reciprocal number of word document frequency (IDF), and combinations thereof (TF-IDF) can be adopted. The degree to which the word characterizes the content in the search question sentence or document in which the word exists may be determined as a weight.
In addition, although the following description advances a case as TF-IDF utilized most frequently, please refer to the said nonpatent literature 2 for the detail of weighting.

In addition, as a vector integration method for integrating a plurality of word vectors generated by translation and generating one word vector used in a search, for example, any one of an average method, a weighted average method, a majority method, and a best selection method It can be employed depending on the purpose of use.
These methods will be described using an example in which n word vectors V1, V2,... Vn are generated by translation and integrated into one word vector V.

In the “average method”, the centroid of the word vectors V1, V2,. Specifically, the word vectors V1, V2,... Vn are all added and divided by n is the integrated word vector V.
The average method uses the performance characteristics of each translation system evenly. If the premise is that “average gives good results”, it is expected that multilingual information retrieval will give good results.
When the faithful translation of the search query sentence q exists in qh1, qh2,... Qhn, the search performance may be degraded compared to using only it (because other bad translation sentences are also used). However, overall (ie, on average) the average method can be expected to give good results.

In the “weighted average method”, the word vectors V1, V2,... Vn are weighted and averaged with specific gravity α1, α2,. Formally, it can be expressed as (α1 × V1 + α2 × V2 +... + Αn × Vn) / (α1 + α2 +... + Αn).
The specific gravity of each vector can be considered as the reliability of each translation system. How to determine the specific gravity may be learned from the training data. As training data, a search query sentence in the first language may be input, and a document set in the second language that matches it may be given. Furthermore, it is desirable to determine the specific gravity of each translation system according to the characteristics of the search question sentence in the first language.
The weighted average method is basically the same as the average method, except that a good translation system is emphasized more and a bad translation system is neglected. However, since weighting for this translation system greatly affects the performance in search, learning of weighting from training data is important.

In the “majority voting method”, if the number of elements other than “0” in the word vectors V1, V2,... Vn is more than half of each vector row, the sum is used as an element, otherwise “0” is used as an element. The integrated word vector V to be
The majority voting method literally determines the elements for each individual row of vectors (ie words in the second language). It is expected that the results will be more conspicuous than the average method, and the fluctuations in search performance (good and bad) are expected to be larger than the average method.

In the “best selection method”, the best one of a plurality of translated search question sentences is selected, and a word vector generated therefrom is adopted as an integrated vector V.
Some examples of how to determine the quality of translation results (search question text) are shown.
First, it is conceivable that each search query sentence in the second language is back-translated into the first language, the degree of similarity with the original search question sentence is calculated, and the one with the highest degree of similarity is regarded as the best. It is not always necessary to provide a plurality of reverse conversion translation systems, and all search query sentences may be reverse translated using different single translation systems. As a method for calculating the similarity of the search question sentence, the following method for calculating the similarity between documents may be used.

Next, a method of learning which translation system is the best from the training data according to the nature of the search question sentence can be considered.
Since the best selection method uses the best translation result, it is expected to give good results in the search, but it cannot be expected to absorb the fluctuation in the expression of the translation result due to the habit of the translation system. . Also, the search performance varies depending on what is considered the best.

In addition, the process of performing a similar document search based on the vector space model using the word vector V integrated from the translation results of a plurality of search question sentences can be performed as follows.
In the vector space model, a search target document is also expressed as a point on the same vector space. That is, it is necessary to generate a corresponding word vector for each document in the target document set prior to the search. The similarity between documents (or questions and documents) is defined as the cosine of a vector. Details of the vector space model are publicly known, and see Non-Patent Document 2.

FIG. 7 shows a more specific configuration of the multilingual document search apparatus corresponding to the second embodiment.
In the multilingual document search apparatus of this embodiment, all the search target documents in the second language are input from the input unit 4 from the target document database 6, and the target documents to which the translation apparatuses 1a, 1b, and 1c are input are respectively input. Each is translated into the same first language as the search condition, and the integrating means 2 obtains word vectors from the three translation results for each target document as described above, and these three word vectors are converted into 1 as described above. One word vector is integrated for each target document.

On the other hand, the search condition of the first language is input from the input means 41 (which may be shared with the input means 4) to the integration means 2, and the integration means 2 uses the words included in the search condition as a reference. The calculated feature amount (word vector) is calculated. In this example, the integration unit 2 is shared by the characterization unit that calculates the word vector with respect to the search condition. However, the characterization unit may be provided separately.
Then, it is determined whether the word vector of the search condition calculated by the search means 3 and the word vector of the target document have a predetermined similarity, and the target document that matches the search condition (has the predetermined similarity) is targeted. The document is identified from the document database 6, and the identified document is output from the output means 5 as a search result.

In this embodiment, when searching using a vector space model, as in the first embodiment, the document does not necessarily have to be translated as a complete sentence, and correspondence between words can be taken as accurately as possible. What is necessary is just to translate as a thing.
Compared with the first embodiment for translating search conditions, this embodiment needs to translate the document to be searched into the first language prior to the search, and it can be said that the overhead is high. There is an advantage that it is not necessary to translate the search condition. Therefore, this embodiment is effective when the search target document set does not change and the speed at the time of search is important.

FIG. 8 shows a more specific configuration of the multilingual document search apparatus corresponding to the third embodiment.
In the multilingual document search apparatus according to the present embodiment, the search query sentence in the first language is input from the input unit 4, and the search query sentences input by the translation devices 1a, 1b, and 1c are respectively translated into the third language. The integration unit 2 obtains word vectors from the three translation results output from each translation device, and integrates these three word vectors into one word vector.

Further, all the search target documents in the second language are input from the target document database 6 by the input means 4, and the respective target documents input by the translation devices 1a, 1b, and 1c are respectively translated into the third language and integrated. The conversion unit 2 obtains word vectors from the three translation results output from the respective translation devices, and integrates the three word vectors into one word vector to obtain one word vector for each target document.
Then, the search unit 3 specifies a target document in which the word vector of the search condition and the word vector of the target document have a predetermined similarity relationship from the target document database 6 and outputs the specified document from the output unit 5 as a search result. To do.

As an advantage of performing multilingual document search through an intermediate language (third language) as in the present embodiment, not only in the first language and the second language, but also in the case of handling documents in many different languages. All of the search through the third language has an advantage that the algorithm becomes simple. For example, for 3 types of languages, if you want to obtain documents of 3 types of languages that match the search conditions in any language, if you do not assume the existence of an intermediate language, all languages It is necessary to translate the combination of the above, but assuming that there is an intermediate language, each language need only be translated with the intermediate language.
Here, the intermediate language does not necessarily need to be a natural language such as Japanese or English. Instead of using a natural language word as an axis of a multidimensional space, a non-patent document 1 describes a Latent Semantic Indexing ( A component that is mathematically extracted (such as LSI) may be used as an axis.

By translating search conditions and target documents with multiple translation systems as described above, the same words are converted into multiple words, so translation blurring can be absorbed. Because it is converted, it is possible to emphasize the importance of stable words (such as technical terms) that have no ambiguity.
In addition, as a search model for regarding a document as a word set, a search method based on a probability model as described in Non-Patent Document 3 may be used in addition to a vector space model.

In the above embodiment, a word vector is generated after a search condition or target document is translated using a plurality of translation systems. However, a plurality of languages after translation after the search condition or target document is made a word vector. It is also possible to convert to a word vector. However, what is needed at this time is not a translation system but a bilingual dictionary for mapping words, and the processing performed there is not a translation but a dictionary lookup.
It is also possible to expand the word into a synonym after generating the word vector and before performing the document search, and by performing a “query expansion” that is commonly used in information search, It can also absorb the blur of the notation.

It is a figure which shows the principal part structure of the multilingual document search device based on one Example of this invention. It is a figure explaining the multilingual document search process based on the Example of this invention. It is a figure which shows the structure of the multilingual document search apparatus based on 1st Example of this invention. It is a figure which shows the data structure of the object document which concerns on an example of this invention. It is a figure explaining the multilingual document search process which concerns on 1st Example of this invention. It is a figure explaining the vector space which concerns on an example of this invention. It is a figure which shows the structure of the multilingual document search apparatus based on 2nd Example of this invention. It is a figure which shows the structure of the multilingual document search apparatus based on 3rd Example of this invention.

Explanation of symbols

1a, 1b, 1c: translation device, 2: integration means (characterizing means),
3: Search means 4, 41: Input means
5: output means, 6: target document database,

Claims (12)

A multilingual document search device that searches a document that matches a search condition from a plurality of target documents described in a second language based on a search condition described in a first language,
A plurality of translation means for translating at least one of the search condition and the target document, and translating the search condition and the target document into a description in the same language;
An integration means for converting a plurality of translation results into a feature value based on a word included in the translation result and integrating the feature value for each search condition or target document;
A search means for searching for a target document from a search condition based on the similarity relationship of integrated feature values;
A multilingual document search device characterized by comprising: A multilingual document search device that searches a document that matches a search condition from a plurality of target documents described in a second language based on a search condition described in a first language,
A target document holding means for holding a plurality of target documents in association with feature quantities based on words included in the target documents;
A plurality of translation means for translating the search condition into a description in the second language;
An integration means for converting a plurality of translation results into a feature value based on a word included in the translation result and integrating the feature value in a search condition;
A search means for searching for a target document from a search condition based on a similarity relationship between the feature quantity of the integrated search condition and the feature quantity of the target document held in the target document holding means;
A multilingual document search device characterized by comprising: A multilingual document search device that searches a document that matches a search condition from a plurality of target documents described in a second language based on a search condition described in a first language,
A plurality of translation means for translating a plurality of target documents respectively into a description in a first language;
An integration means for converting a plurality of translation results for each target document into a feature value based on a word included in the target document, and integrating the feature value for each target document;
Characterization means for converting a search condition into a feature value based on a word included in the search condition;
A search means for searching for a target document from a search condition based on a similarity relationship between the feature quantity of each integrated target document and the feature quantity of the search condition;
A multilingual document search device characterized by comprising: A multilingual document search device that searches a document that matches a search condition from a plurality of target documents described in a second language based on a search condition described in a first language,
A plurality of translation means for translating the search condition and the plurality of target documents, respectively, into a third language description;
Converts multiple translation results of search conditions into feature quantities based on the words they contain, integrates them into the feature quantities of the search conditions, and also includes words that contain multiple translation results for each target document An integration means for converting into a feature value as a reference and integrating it into a feature value for each target document;
A search means for searching for the target document from the search condition based on the similarity relationship between the feature quantity of the integrated search condition and the feature quantity of the target document;
A multilingual document search device characterized by comprising: The multilingual document search device according to any one of claims 1 to 4,
The feature amount of the search condition or target document includes a plurality of preset words among the words included in the search condition or target document, and the importance of the word in the search condition or target document. Is a word vector with element value
The multilingual document search apparatus, wherein the similarity relation of the feature quantities is determined based on a cosine between word vectors. A computer program for configuring a multilingual document search apparatus that searches a document that matches a search condition from a plurality of target documents described in a second language based on the search condition described in the first language. And
A plurality of translation means for translating the input search condition into a description in the second language;
An integration means for converting a plurality of translation results into a feature value based on a word included in the translation result and integrating the feature value in a search condition;
Similarity relationship between the feature value of the integrated search condition and the feature value of the target document with reference to the target document holding means that holds a plurality of target documents in association with the feature values based on the words that they contain A search means for searching for a target document from a search condition based on
A multilingual document search program comprising: A computer program for configuring a multilingual document search apparatus that searches a document that matches a search condition from a plurality of target documents described in a second language based on the search condition described in the first language. And
A plurality of translation means for translating a plurality of target documents respectively into a description in a first language;
An integration means for converting a plurality of translation results for each target document into a feature value based on a word included in the target document, and integrating the feature value for each target document;
Characterization means for converting a search condition into a feature value based on a word included in the search condition;
A search means for searching for a target document from a search condition based on a similarity relationship between the feature quantity of each integrated target document and the feature quantity of the search condition;
A multilingual document search program comprising: A multi-language document search apparatus and a computer program configured to search a document that matches a search condition from a plurality of target documents described in a second language based on a search condition described in a first language. ,
A plurality of translation means for translating the search condition and the plurality of target documents, respectively, into a third language description;
Converts multiple translation results of search conditions into feature quantities based on the words they contain, integrates them into the feature quantities of the search conditions, and also includes words that contain multiple translation results for each target document An integration means for converting into a feature value as a reference and integrating it into a feature value for each target document;
A search means for searching for the target document from the search condition based on the similarity relationship between the feature quantity of the integrated search condition and the feature quantity of the target document;
A multilingual document search program comprising: A method for searching a document that matches a search condition from a plurality of target documents described in a second language based on the search condition described in a first language,
The multilingual document search device translates at least one of the search condition and the target document by a plurality of translation means, translates the search condition and the target document into a description in the same language, and includes a plurality of translation results. The feature amount based on the word to be used and integrated into the feature amount for each search condition or target document, and then the target document is searched from the search condition based on the similarity relationship of the integrated feature amount. A multilingual document search method as a feature. A method for searching a document that matches a search condition from a plurality of target documents described in a second language based on the search condition described in a first language,
The multilingual document search device translates the search condition into a description in the second language by a plurality of translation means, converts the plurality of translation results into feature quantities based on the words included therein, and After integrating into a quantity, refer to the target document holding means that holds a plurality of target documents in association with the feature quantity based on the word that they contain, and the feature quantity and target document in the integrated search condition A multilingual document search method characterized in that a target document is searched from a search condition based on a similarity relationship with a feature amount of a document. A method for searching a document that matches a search condition from a plurality of target documents described in a second language based on the search condition described in a first language,
A multilingual document search device translates a plurality of target documents into a description of a first language by a plurality of translation means, and converts a plurality of translation results for each target document into feature quantities based on words included in them. Are integrated into the feature quantity for each target document, and after the search condition is converted into a feature quantity based on a word included in the search condition, the integrated feature quantity for each target document and the feature quantity of the search condition are A multilingual document search method characterized in that a target document is searched from a search condition based on the similarity relationship. A method for searching a document that matches a search condition from a plurality of target documents described in a second language based on the search condition described in a first language,
A multilingual document search apparatus translates a search condition and a plurality of target documents into a description in a third language by a plurality of translation units, and uses a plurality of translation results of the search condition as a feature amount based on a word included in them. Converted and integrated into the feature value of the search condition, and also converted multiple translation results for each target document into feature values based on the words they contain, and integrated them into the feature values for each target document Then, a multilingual document search method characterized in that the target document is searched from the search condition based on the similar relationship between the feature quantity of the integrated search condition and the feature quantity of the target document.

Images