JP2008021270A - Data conversion apparatus and its method, database management apparatus and its method, and database search system and its method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To realize retrieval with retrieval conditions specified by semantic-related notation between sentence elements such as words and phrases and concepts prescribed by the semantic-relation. <P>SOLUTION: A management server 30 which manages database of documents of a prescribed language includes a syntax analysis part 32 which executes syntax analysis of documents using document data 60, grammar information, and vocabulary information of the languages; an index making part 34 which outputs indexes showing each description position and hierarchy depth of tags, words and attributes of the tags by associating tags granted to the words etc. equivalent to the predicate in a sentence and tags granted phrases etc. that become paragraphs with attributes, while granting prescribed tags in which attributes corresponding to the sentences in the documents, and phrases and words included in the sentences based on the result of the syntax analysis; and an index holding part 35 which holds outputted indexes for every document data 60. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a data conversion apparatus and method for converting data indicating a sentence in a predetermined language into structured data to which a grammatical semantic annotation is added, and a database storing a plurality of document data of documents in a predetermined language. The present invention relates to a database management apparatus and method for managing, and a database search system and method for searching a database in which a plurality of document data of documents in a predetermined language are stored.

Conventionally, XML (Extensible Markup Language) has been known as a markup language for describing the meaning and structure of documents, etc., and document databases that store document data structured using XML have gradually spread. I am doing. Moreover, XPath (refer nonpatent literature 1), XQuery (refer nonpatent literature 2), etc. are known as a query language used for the search of such an XML database.
XML Path Language (XPath) Version 1.0 W3C Recommendation 16 November 1999, Internet <http://www.w3.org/TR/xpath> XQuery 1.0: An XML Query Language W3C Candidate Recommendation 8 June 2006, Internet <http://www.w3.org/TR/xquery/>

  By the way, in recent years, a large-scale database of documents such as literatures and academic papers is being developed in various fields, and accordingly, there is an increasing need for a technology that enables a large-scale document database to be efficiently searched. In this case, even if a general keyword search is applied to academic papers, etc., only documents containing all of the search keywords are output as search results, so that the desired document can be accessed accurately and quickly. To achieve this, for example, a relation concept, which is a concept defined by a semantic relation between sentence elements (words and phrases), such as a correspondence between a predicate in a document and its term, can be set as a search condition. Should. However, the semantic relationships between sentence elements are often extremely complicated and complicated, and when a document is expressed in XML including the semantic relations between sentence elements, the data structure is extremely complex. Therefore, even if a query language such as XPath or XQuery is used, it is extremely difficult to quickly and accurately perform a search using the above relational concept as a search condition.

  Therefore, the data conversion apparatus and method according to the present invention are structured to enable notation of a semantic relationship between sentence elements such as words and phrases and a search using a concept defined by the semantic relationship as a search condition. The purpose is to provide data. Further, the database management apparatus and method and the database search system and method according to the present invention can search a database in which a plurality of document data are stored using a concept defined by a semantic relationship between sentence elements such as words and phrases as a search condition. The purpose is to do.

  The data conversion apparatus and method, the database management apparatus and method, and the database search system and method according to the present invention employ the following means in order to achieve at least a part of the above object.

The data conversion device according to the present invention comprises:
A data conversion device that converts data indicating a sentence in a predetermined language into structured data including grammatical semantic annotations,
A syntax analysis means for performing syntax analysis of the sentence using the data and grammatical information and word information of the predetermined language;
Based on the result of the parsing by the parsing means, a predetermined tag in which a corresponding attribute is described is assigned to a phrase and a word included in the sentence, and a phrase or word corresponding to a predicate in the sentence After associating a given tag with a tag given to a phrase or word as a term of the predicate by the attribute, the description position and the hierarchical depth of each of the tag, the word, and the tag attribute Structured data output means for outputting the indicated index as the structured data;
Is provided.

  In this data conversion device, when data indicating a sentence in a predetermined language is converted into structured data, first, syntax analysis of the sentence using the data indicating the sentence and the grammatical information and word information of the predetermined language is performed. Executed. Based on the result of the parsing, a phrase and a word included in the sentence are given a predetermined tag in which a corresponding attribute (attribute name, attribute value if present) is described, A tag attached to a phrase or word corresponding to a predicate in the sentence is associated with a tag attached to a phrase or word serving as a term of the predicate using an attribute, and finally, the tag, word, and tag An index indicating the description position and hierarchical depth of each attribute is output as structured data. As a result, the structured data, ie, the index, shows the substantial address of each of the tag, word, and tag attribute, so the predicate and its terms are complex in the sentence structured using the tag. Even if it is confusing, if you refer to the structured data that is an index, the sentence syntax tree defined by the correspondence between the start tag and the end tag, the tag assigned to the word corresponding to the predicate, and the It is possible to easily specify a predicate term structure in a sentence defined by a correspondence relationship with a tag attached to a phrase or the like that becomes a predicate term. Therefore, according to this data conversion apparatus, the notation of the predicate term relationship, which is a semantic relationship between the predicate (predicate) and the term in the sentence, and the concept defined by such a semantic relationship are expressed. It is possible to realize structured data that enables a search as a search condition.

  In the data conversion apparatus according to the present invention, the syntax analysis means may be a parser that executes syntax analysis based on HPSG to acquire the syntax tree and predicate term structure of the sentence. As a result, the sentence can be parsed more accurately, and the association between the predicate and the term in the index can be made more appropriate.

  Furthermore, the data conversion apparatus according to the present invention determines, based on predetermined term information, whether or not each phrase or word included in the sentence is bundled with one substantive meaning. In addition, when the phrase or word is bundled with the substantive meaning, it may further comprise term recognition means for acquiring the substantive meaning, and the structured data output means includes the acquired substantive meaning. The attribute value to be indicated may be described in a predetermined attribute of the corresponding tag. As a result, in the index, attribute values indicating the same substantive meaning are described in tags corresponding to phrases or words whose surface expression is different in spite of showing substantially the same meaning (concept). Therefore, if structured data that is the index is used, it is possible to easily perform a search for phrases and words having the same substantial meaning, or sentences including the same.

The data conversion method according to the present invention includes:
A data conversion method for converting data indicating a sentence in a predetermined language into structured data including grammatical semantic annotations,
(A) performing a syntax analysis of the sentence using the data and grammatical information and word information of the predetermined language;
(B) Based on the result of the syntax analysis in step (a), a predetermined tag describing a corresponding attribute is assigned to a phrase and a word included in the sentence, and a phrase corresponding to a predicate in the sentence Alternatively, the tag assigned to the word and the phrase or tag assigned to the word as a term of the predicate are associated with each other by the attribute, and the description position and the hierarchy of each of the tag, the word, and the tag attribute Outputting an index indicating depth as the structured data;
Is included.

  By referring to structured data that is an index obtained by this method, the syntax tree of the sentence defined by the correspondence between the start tag and the end tag, the tag attached to the word corresponding to the predicate, and the predicate It becomes possible to easily specify the predicate term structure in the sentence defined by the correspondence relationship with the tag given to the phrase or the like that becomes the term. Therefore, according to this method, a notation of a predicate term relationship that is a semantic relationship between a predicate and a term in a sentence, and a search using a concept defined by such a semantic relationship as a search condition, Structured data can be realized.

The database management device according to the present invention is:
A database management apparatus that manages a database that stores a plurality of document data of documents in a predetermined language,
Syntax analysis means for executing syntax analysis of the document using the document data and grammatical information and word information of the predetermined language;
Based on the result of the parsing by the parsing means, a predetermined tag describing attributes corresponding to sentences in the document and phrases and words included in the sentence is given, and predicates in the sentence are given. Description of each of the tag, the word, and the attribute of the tag after associating the tag attached to the corresponding phrase or word with the tag attached to the phrase or word as the term of the predicate by the attribute Index output means for outputting an index indicating a position and a hierarchy depth;
Index holding means for holding the index output by the index output means for each document data;
Is provided.

  In this database management apparatus, for example, when document data of a new document is accumulated in the database, document parsing using the document data, grammatical information and word information of a predetermined language is executed. Based on the result of the parsing, a sentence in the document and a phrase and a word included in the sentence are given predetermined tags in which corresponding attributes are described, and correspond to predicates in the sentence. A tag assigned to a phrase or word and a tag assigned to the phrase or word that is the term of the predicate are associated using attributes, and the description position and hierarchy of each of the tag, word, and tag attributes An index indicating the depth is output as structured data and is stored in association with the new document. In this way, if an index as described above is held for each document data, a syntax tree or a predicate of a sentence in the document specified by the correspondence between the start tag and the end tag can be referred to by referring to the index. It is possible to easily specify a predicate term structure in a sentence defined by a correspondence relationship between a tag assigned to a word corresponding to と and a tag assigned to a phrase or the like serving as a term of the predicate. Therefore, according to this database management apparatus, it is possible to search a database in which a plurality of document data is stored using a concept defined by a semantic relationship between sentence elements such as words and phrases as a search condition.

The database management method according to the present invention includes:
A database management method for managing a database storing a plurality of document data of documents in a predetermined language,
(A) performing syntax analysis of the document using the document data and grammatical information and word information of the predetermined language;
(B) Based on the result of the parsing in step (a), a predetermined tag describing attributes corresponding to sentences in the document and phrases and words included in the sentence is given, The tag assigned to the phrase or word corresponding to the predicate of and the tag assigned to the phrase or word serving as the term of the predicate are associated by the attribute, and the attribute of the tag, the word and the tag Outputting an index indicating each description position and hierarchy depth and holding each document data; and
Is included.

  If an index such as that described above is held for each document data as in this method, the syntax tree of sentences in the document specified by the correspondence between the start tag and the end tag can be obtained by referring to the index. The predicate term structure in the sentence defined by the correspondence between the tag assigned to the word corresponding to the predicate and the tag assigned to the phrase or the like serving as the term of the predicate can be easily specified. Therefore, according to this method, it is possible to search a database in which a plurality of document data are accumulated, using a concept defined by a semantic relationship between sentence elements such as words and phrases as a search condition.

A database search system according to the present invention includes:
A database search system for searching a database storing a plurality of document data of documents in a predetermined language,
The sentence in the document, a predetermined tag assigned to each of the phrase and the word included in the sentence, the description position and the hierarchical depth of each attribute described in the word and the tag, and the sentence An index created for each piece of document data is stored so that a tag assigned to a phrase or word corresponding to the predicate of and a tag assigned to a phrase or word serving as a term of the predicate are associated by the attribute. Index holding means;
Accepting means capable of accepting a plurality of search words and a predicate term relationship between the search words as a search condition;
Extraction means for extracting a tag region including all the accepted search words based on the index held by the index holding means;
Determining means for determining, based on the index, whether or not the matching corresponding to the accepted predicate term relationship is made between the tags corresponding to the search word included in the tag area extracted by the extracting means. When,
Search result output means for outputting a search result based on a determination result by the determination means;
Is provided.

  This database search system includes index holding means for holding an index created for each document data. The index for each document data includes a sentence in the corresponding document, a phrase and a word included in the sentence, respectively. A tag assigned to a phrase or word corresponding to a predicate in a sentence and a description tag and a word corresponding to a predicate in the sentence, and a predetermined tag, a word, and a description of each attribute described in the tag. And a tag attached to a phrase or word. Therefore, referring to the index, the sentence syntax tree in the document defined by the correspondence between the start tag and the end tag, the tag assigned to the word corresponding to the predicate, the phrase that is the term of the predicate, etc. It is possible to easily specify the predicate term structure in the sentence defined by the correspondence relationship with the tag assigned to. Based on this, in this database search system, when a plurality of search words and a predicate term relationship between the search words are accepted as search conditions, first, a tag area including all the accepted search words based on an index Is extracted. Further, the search included in the extracted tag region based on the attribute that associates the tag attached to the index, that is, the tag assigned to the word corresponding to the predicate, and the tag attached to the phrase or the like of the term of the predicate. It is determined whether or not the matching corresponding to the predicate term relationship accepted as the search condition is made between the tags corresponding to the word, and the search result is output based on the determination result. As a result, it is possible to search a database in which a plurality of document data is stored using a concept defined by a semantic relationship between sentence elements such as words and phrases as a search condition.

  In the index, the predetermined attribute of the tag corresponding to the phrase or word corresponding to the predicate in the sentence and the predetermined attribute of the tag corresponding to the phrase or word serving as the term of the predicate are the same as the attribute value. A value may be set, and the determination unit may determine whether there is a tag pair corresponding to the search word in which the same value is described as the attribute value based on the index. Good. As a result, a more accurate search result can be obtained quickly.

  In this case, the database search system according to the present invention may further include appearance frequency storage means for storing the appearance frequency in the database for each word, and the determination means includes the search words that are in a predicate term relationship with each other. The value of the variable is acquired using the attribute value described in the tag corresponding to one search word with a lower appearance frequency as a variable, and the attribute value that matches the acquired variable value is set. It may be determined whether or not a tag corresponding to the other search word exists. In this way, if the attribute value described in the tag corresponding to one of the search words having a smaller appearance frequency is used as a variable, the number of variable values to be acquired is reduced, and the corresponding value should be determined. Since the number of tags corresponding to the other search word can be reduced, it is possible to efficiently determine in a short time whether or not there is a tag pair corresponding to the search word in which the same value is described as the attribute value. Become.

  Further, the database search system according to the present invention is a search query setting means for setting a search query using as a variable an attribute value that correlates the predicate term relationship based on the accepted search word and the previous description term term relationship. The extraction means identifies a maximum search area that is a maximum search target area that appears in the set search query and includes a start tag for a maximum search area that includes all of the accepted search words and An end tag is extracted, and the determination unit executes a query excluding the variable from at least one partial query including any of the search words obtained by decomposing the set search query and the variable, It is determined whether or not there exists the maximum search area that matches the search query assigned the value of the variable returned as an execution result. Good.

  The index may be created in advance based on a result of syntax analysis of the document using the document data and grammatical information and word information of the predetermined language. As a result, it is possible to reduce the time required from the input of the search condition to the output of the search result by reducing the actual search processing load.

The database search method according to the present invention includes:
A database search method for searching a database storing a plurality of document data of documents in a predetermined language,
(A) indicates a sentence in the document, a predetermined tag assigned to each of a phrase and a word included in the sentence, a description position and a hierarchical depth of each attribute described in the word and the tag, and An index created for each document data so that a tag attached to a phrase or word corresponding to a predicate in the sentence and a tag attached to a phrase or word serving as a term of the predicate are associated by the attribute. Creating and maintaining
(B) receiving a plurality of search words and a predicate term relationship between the search words as a search condition;
(C) extracting a tag region including all the accepted search words based on the index;
(D) Based on the index, whether or not the correspondence corresponding to the accepted predicate term relationship is made between the tags corresponding to the search word included in the tag area extracted in step (c). A determining step;
(E) outputting a search result based on the determination result in step (d);
Is included.

  According to this method, it is possible to search a database in which a plurality of document data are accumulated using, as a search condition, a concept defined by a semantic relationship between sentence elements such as words and phrases.

  Next, the best mode for carrying out the present invention will be described with reference to examples.

  FIG. 1 is a schematic configuration diagram of a database search system 20 according to an embodiment of the present invention. As shown in the figure, the database search system 20 of the embodiment mainly manages a data storage device 21 that stores a huge amount of data in a database and a database on the data storage device 21 and searches the database. The management server 30 is configured to be enabled. The database of the embodiment is constructed as a document database in which a large number of document data 60 of documents such as documents and academic papers in a predetermined field written in a predetermined language (for example, English) is accumulated. The management server 30 includes a server CPU, ROM, RAM, system bus, various interfaces, storage devices, external storage devices, etc. (not shown). The management server 30 is connected to the management server 30 via a network such as the Internet, for example. It can be accessed from the terminal 50. As shown in FIG. 1, the management server 30 inputs a document by cooperation of one or both of hardware such as a server CPU, ROM, RAM, various interfaces, and a storage device and various programs installed in advance. A unit 31, a syntax analysis unit 32, a term recognition unit 33, an index creation unit 34, an index holding unit 35, an input reception unit 36, a search processing unit (search engine) 37, and a result output unit 38 are constructed as functional blocks.

  The document input unit 31 receives input of document data 60 indicating a document in a predetermined language described in a predetermined data format (for example, text format), stores the received document data 60 in the data storage device 21, and the document data 60 is given to the syntax analysis unit 32 and the term recognition unit 33. The syntax analysis unit 32 uses the document data 60 from the document input unit 31 and the grammatical information (grammar rules) and word information (dictionary) of the language of the document shown in the document data 60 to parse the document shown in the document data 60. It is something to execute. The syntax analysis unit 32 according to the embodiment is configured as a parser that executes syntax analysis based on HPSG (Head-Driven Phrase Structure Grammar) and acquires a syntax tree and a predicate term structure of a sentence (sentence) in the document. . Grammar information and word information used for syntax analysis by the syntax analysis unit 32 are prepared in advance and stored in a predetermined storage device (not shown). The term recognition unit 33 has a phrase or word for each phrase or word in the document indicated by the document data 60 from the document input unit 31 based on predetermined term information stored in a predetermined storage device (not shown). It is determined whether or not the phrase is enclosed in one entity meaning, and when the phrase or word is enclosed in the one entity meaning, the entity meaning is acquired. The index creation unit 34 creates an index to be used when searching the database on the data storage device 21 based on the result of the syntax analysis by the syntax analysis unit 32 and the result of the term recognition by the term recognition unit 33, and stores the index. The unit 35 holds the index created by the index creating unit 34 in a state in which the index is associated with the document data 60 on the database. The input receiving unit 36 receives search conditions input from the terminal 50 or the like connected to the management server 30 via the network 40. In the embodiment, as a search condition, a plurality of search words and a predicate term relationship between the search words can be input. Then, the search processing unit 37 sets a search query based on the search condition (search word and predicate term relationship) received by the input receiving unit 36 and executes the set search query. The result output unit 38 A search result is output based on the processing result of the processing unit 37.

  Next, an index creation process for creating an index used for database search using the database search system 20 of the embodiment will be described. Further, following the description of the index creation process, the database search process on the data storage device 21 using the index in the database search system 20 of the embodiment will be described.

[Index generation processing]
FIG. 2 is a flowchart illustrating an example of an index creation process executed in the management server 30 included in the database search system 20 according to the embodiment. The index creation processing of FIG. 2 is executed when new document data 60 is accumulated in a database on the data storage device 21, for example.

  When the input of new document data 60 is received by the document input unit 31 of the management server 30 (step S100), the document input unit 31 stores the received document data 60 in the data storage device 21 and the document data 60. Is given to the syntax analysis unit 32 and the term recognition unit 33. The syntax analysis unit 32 that has received the document data 60 from the document input unit 31 executes a syntax analysis of the document using the document data 60 and the grammatical information and word information of the target language, and similarly the term that has received the document data 60. The recognition unit 33 executes term recognition for each word in the document indicated by the document data 60 from the document input unit 31 based on the term information prepared in advance in parallel with the syntax analysis processing by the syntax analysis unit 32. (Step S110).

  Here, with reference to FIGS. 3 to 6, the document syntax analysis processing by the syntax analysis unit 32 of the embodiment in step S110 will be described. First, for example, considering the English sentence “Mary slaps John”, it can be said that this English sentence shows a concept (relationship concept) composed of two relations [Mary, subject, slap] and [John, object, slap]. As shown in FIG. 3, this English sentence has three nodes “Mary”, “John”, and “slaps”, and the words “Mary” and “slaps” are at the edge labeled “subject”. In addition to being connected, the words “John” and “slaps” can be represented by a directed graph with a label connected by the edge of the label “object”. In such a graph, not only each node but also an edge connecting the nodes shows very important information in terms of grammar. In other words, the English sentence “John slaps Mary” shows a completely different relationship concept even though it contains the same three words as the previous English sentence “Mary slaps John”, so this English sentence is different from the graph of FIG. It will be represented by a graph. On the other hand, the English sentence “John is slapped by Mary” has the same meaning as the above “Mary slaps John” although the character string is different on the surface layer, and a graph approximated to the graph of FIG. 3 is used. Can be expressed. Therefore, in order to execute a search using such a relation concept (predicate term structure) as a search condition, the above-described edge must not be ignored as in the existing document search method for determining only the presence or absence of a search word. Therefore, a graph showing the predicate term relationship between the sentence elements (the relationship between the semantic subject, predicate, and object) is required.

  For this reason, in the embodiment, in order to obtain a graph showing a predicate term relationship between sentence elements, that is, a syntax tree of a sentence in a document and a predicate term structure, the parsing unit 32 performs syntax analysis based on HPSG as described above. It is configured as a parser to execute. In this case, the grammatical information of the syntax analysis unit 32 is constructed by aggregating so-called triplet and case restrictions, and restrictions based on the learning result of “ambiguity” in a natural language sentence. Further, the word information of the syntax analysis unit 32 is created as a list of data (vocabulary items) indicating the syntactic properties for each word. As shown in FIGS. 4 and 5, each word list includes four items “HEAD”, “SUBJ”, “COMPS”, and “SEM”, where “HEAD” is a part of speech, “SUBJ” "Can be a syntactic subject (a word or phrase that can be located to the left of the predicate)," COMPS "can be a syntactic object (a word or phrase that can be to the right of the predicate)," SEM " Represents a semantic structure (predicate term structure). Taking the English sentence “Mary slaps John” as an example, for the words “Mary” and “John”, “prernn” indicating a proper noun is described in the item “HEAD”, and “Mary” in the item “SEM”. Alternatively, the character string “John” is described. Also, for the word “slap (s)”, the verb “verb” is displayed in the item “HEAD” and the item “SUBJ” is displayed in the item “SUBJ” to indicate that the verb is a noun for the subject and the object. “HEAD (= part of speech) noun (= noun)” is described, and “HEAD noun” is described in the item “COMPS”. Further, “SEM 1” is described in the item “SUBJ”, “SEM 2” is described in the item “COMPS”, and “slap (1, 2)” is described in the item “SEM”. Here, the numbers “1” and “2” are variables for defining (mapping) the correspondence between the syntactic subject, the object and the semantic subject, and the object, and in the item “SEM” The notation in parentheses following the character string “slap” indicates “(a variable indicating a semantic subject, a variable indicating a semantic object)”. As a result, as shown in FIG. 4, for the word “slap (s)”, a value 1 is described in “SEM” of the item “SUBJ” of the list and a value 2 is stored in “SEM” of the item “COMPS”. And the item “SEM” contains slap (1, 2) ”so that the syntactic subject becomes the semantic subject and the syntactic object becomes the semantic object. As can be seen from the list, as can be seen from Fig. 5, for the word "slapped", the syntactic subject becomes the semantic object and the syntactic object becomes the semantic subject. You can know from the list.

  Then, the syntax analysis unit 32 acquires a vocabulary item for each sentence word in the document based on the word information as described above, and follows the “noun phrase (NP) as illustrated in FIG. ) ”,“ Verb Phrase (VP) ”,“ Sentence ”, etc. In addition to showing the hierarchical structure of elements in the sentence (see the solid line in the figure), the predicate (word“ slaps ”) and its terms, meaning A structure tree indicating a predicate term relationship that is a relationship with the above subject (word “Mary”) and object (word “John”) is acquired. For example, when an English sentence “Mary slaps John” is given to the syntax analysis unit 32, as shown in FIG. 4, the word “slaps” takes one noun (noun) on the right side from the word information (list). The list of words “John” (HEAD) describes “prernn”, so when “slaps” and “John” are adjacent, these two It is recognized that the words combine to form the phrase “slaps John”. Furthermore, the phrase “slaps John” is a verb phrase that can take a noun on the left side. If the word “Mary” (proprietary noun) is on the left side, the English sentence “Mary slaps John” A syntax tree is obtained. Further, when adjacent phrases have a property that can be combined with each other, a larger phrase is recognized. As for the English sentence “Mary slaps John”, the syntactic subject / object of “slap” is recognized as a semantic subject / object as it is, but as shown in FIG. 5, “John is slapped”. For the English sentence “by Mary”, the syntactic subject and object are recognized as the semantic object and subject, respectively. This allows for syntactic variations in various sentences (for example, passive form “John is slapped by Mary”, relative pronoun “John who Mary slaps”, etc.), even if the surface expression is different. 4 and FIG. 5, it is recognized that the relationship between the semantic subject, object, etc., that is, the predicate term structure is substantially the same. The syntax analysis unit 32 performs such processing recursively to acquire a syntax tree indicating the predicate term relationship for each sentence in the document, and the index creation unit 34 as data in a predetermined mode indicating the result of the syntax analysis. Output to. When a single document language is handled on the database, a syntax analysis algorithm, grammar information, and word information corresponding to the language may be prepared in advance. When handling, what corresponds to each language may be prepared as a parsing algorithm, grammatical information, and word information. As an example of a parser that performs such parsing based on HPSG, “Enju”, which was developed to obtain the syntax tree and predicate term structure of an English sentence, is known. (See http://www-tsujii.is.su-tokyo.ac.jp/enju/index.html>.)

  Subsequently, term recognition processing by the term recognition unit 33 in step S110 will be described. First, consider the English sentence “Mary hits John” for the example sentence “Mary slaps John”. Although these two English sentences have different surface expressions, the word “slaps” and the word “hits” have substantially the same meaning as each other. It shows the content (relationship concept). Therefore, in order to be able to search for phrases or words having the same substantial meaning, or sentences containing the same, it is necessary to consider the substantial meaning of such words. For this reason, in the management server 30 of the embodiment, a term recognition unit 33 that executes term recognition processing using term information relating to the substantial meaning of each phrase or word is constructed in the management server 30. In the embodiment, as the term information of the term recognition unit 33, ontologies relating to synonyms of verbs and nounized expressions are prepared based on the fact that the verb expression plays an important role in the notation of the sentence predicate term structure. The ontology related to this verb assigns a unique ID (hereinafter referred to as “ontology ID”) to each of the substantial meanings (representing synonyms) that appear when the verbs are classified according to their substantial meanings. It is created in advance as a database in which ontology IDs are associated with synonymous verbs, verbal noun expressions, etc., and stored in a predetermined storage device (not shown). In addition, in the embodiment, as the term information of the term recognition unit 33, the term (technical term) in the field of the document (for example, the field of biology, medicine, etc.) that is the object of the database is obtained for each substantive meaning in the real world. An ontology in which a unique ID is assigned to each of the substantial meanings that appear when classified and a synonym is associated with each ontology ID is prepared. A part of ontology in such a specific field is illustrated in FIG. This figure shows one entry in the ontology related to genes / proteins. In this entry, the ID (ontology ID), synonym (Name, Synonym), and that of the gene called CRP are observed. Information such as species (Species), proteins encoded by the gene (Product), and links to external databases (External links) are collected. As can be seen from FIG. 7, the gene represented by this ID is sometimes expressed as “MGC88244”, “PTX1”, etc., but the same ontology is entered in the entry of the gene (character string) showing the same substantial meaning. By assigning IDs, it is possible to extract words and phrases having the same substantial meaning even if there is a difference in surface expression. The term recognizing unit 33 refers to the term information and determines whether each phrase or word in the document indicated in the document data 60 is bundled with a certain substantive meaning. Judgment is made, and when the phrase or word is bundled with the one substantial meaning, the substantial meaning (ontology ID) is acquired. Then, the term recognition unit 33 outputs the correspondence relationship between the ontology ID and the phrase or the word to the index creation unit 34 as data in a predetermined mode indicating the result of term recognition.

  The index creation unit 34 that has received the result of the syntax analysis by the syntax analysis unit 32 and the result of the term recognition by the term recognition unit 33 is a data storage device based on the result of the syntax analysis and the result of term recognition by the term recognition unit 33. An index used when searching the database on 21 is created and output (step S120). That is, the index creation unit 34 structures the document data 60 with tags based on the results of syntax analysis and term recognition, and outputs that creates and outputs an index based on the results of the structure with the structured module. Modules (both not shown), and the results of parsing and term recognition are first given to the structured module. Based on the result of syntax analysis by the syntax analysis unit 32, the structured module of the index creation unit 34 has attributes corresponding to sentences in the document indicated by the document data 60 and phrases and words included in the sentence. The tag assigned to the phrase or word corresponding to the predicate in the sentence and the tag assigned to the phrase or word as the term of the predicate are associated using the above attribute. To do. FIG. 8 illustrates the types of tags used for structuring such a document and the attributes described therein, and FIG. 9 illustrates the contents of the attributes described in the tags. FIG. 10 shows an output example when the English text “Pretty Mary slaps John” structured by the structured module is output as it is, for example, in XML format.

As shown in FIGS. 8 to 10, when a document is structured by the structuring module, each sentence in the document is surrounded by a start tag <Sentence> and an end tag </ Sentence>. Each phrase in the sentence is surrounded by a <phrase> start tag and </ phrase> end tag. Each <phrase> tag has an id attribute indicating the tag ID and a syntactic category. The cat attribute to be shown is described. An attribute value corresponding to each tag attribute is set. For example, the id attribute of each <phrase> tag has an attribute value of p1, p2,..., P (n),... PN (where “n” and “N” are positive positive numbers) in the order of appearance. Set in the cat attribute, the sentence (S), noun phrase (NP), adjective phrase (ADJP), verb phrase (VP) in the sentence where the phrase enclosed by the tag appears in the “so that” syntax, for example ) Etc., attribute values (S, NP, ADJP, VP, etc.) are set. Further, each word is surrounded by a start tag <word> and an end tag </ word>, and each <word> tag describes a pos attribute indicating the part of speech of the corresponding word. In this case as well, attribute values corresponding to the tag attributes are set. For example, attribute values indicating parts of speech such as adjectives, proper nouns, and verbs are set in the pos attribute of each <word> tag. The <word> tag corresponding to the word that is a verb describes a rel_type attribute. In this rel_type attribute, an ontology ID based on the result of term recognition by the term recognition unit 33, that is, an ontology ID of the actual meaning of the verb. (In the example of FIG. 10, “i789” given to the words “slap” and “hit”) is set. In the <word> tag, an arg (X) attribute (X = 1, 2,...) For identifying a predicate term relationship between statement elements acquired by syntax analysis is described. In an embodiment, the arg (X) attribute is, for example,
arg1: The main part of the verb
arg2: Direct target part of the verb
arg3: It is determined to specify the relationship such as the indirect objective part of the verb. The arg1 attribute contains the ID of the <phrase> tag assigned to the phrase that is the main part of the corresponding verb, and the arg2 attribute contains , The ID of the <phrase> tag assigned to the phrase that is the direct target part of the corresponding verb, and the arg3 attribute is the ID of the <phrase> tag assigned to the phrase that is the indirect target part of the corresponding verb Are set respectively. For example, in the example of FIG. 10, the attribute value of the arg1 attribute of the <word> tag assigned to the verb “slaps” is “p2”, but the attribute value p2 is assigned to the noun phrase “Pretty Mary”. This matches the ID of the <phrase> tag, and is defined by the <phrase> tag and the </ phrase> tag attached to the noun phrase “Pretty Mary” as shown by the broken line in FIG. The region and the region specified by the <word> tag and </ word> tag attached to the word “slaps” are associated, and the main part (arg1) of the verb “slaps” is the noun phrase “Pretty Mary” It will be expressed as. In the example of FIG. 10, the attribute value of the arg2 attribute of the <word> tag assigned to the verb “slaps” is “p8”, but the attribute value p8 surrounds the word “John” <phrase> As shown by the broken line in FIG. 10, the region defined by the <word> tag and the </ word> tag attached to the word “slaps” and “John” are matched. The <phrase> tag corresponding to the word and the area specified by the </ phrase> tag are associated, and the target part (arg2) of the verb “slaps” is expressed as the noun (phrase) “John” It will be. That is, in the embodiment, the same attribute value is set in the arg (X) attribute of the tag corresponding to the word corresponding to the predicate in the sentence and the id attribute of the <phrase> tag corresponding to the phrase that is the term of the predicate. Is done. In this way, by associating a word that is a predicate and a phrase that is a term thereof, it is possible to more appropriately express a semantic relationship between sentence elements in a sentence. Further, as a result of the term recognition by the term recognition unit 33, a <word> tag of a word recognized as a term (technical term) is surrounded by a start tag <entity_name> and an end tag </ entity_name>. In the entity_name> tag, a type attribute indicating the type of term and an ontology attribute indicating the ontology ID of the term are described. In the type attribute, an attribute value indicating the type of term is described. The value of the ontology ID acquired from is described as the attribute value. For example, in the example of FIG. 10, the <entity_name> tag is assigned to the words “Mary” and “John”, the type attribute is “person”, and the “mary” ontology attribute is “i1234”. "I3456" is set as the attribute value for the ontology attribute of "John".

  Here, even a very simple English sentence such as “Pretty Mary slaps John” as described above, when it is structured by XML, the data is as shown in FIG. In particular, the semantic relationships between sentence elements in academic papers and the like are often extremely complicated and complicated. The structure becomes extremely complicated. Therefore, it is extremely difficult to quickly and accurately perform a search using a relation concept as a search condition based on structured data as shown in FIG. 10 by the structured module even using a query language such as XPath or Xqury. Become. Therefore, in the management server 30 of the embodiment, the result of structuring the document by the structuring module is given to the output module in a predetermined manner, and the tag, word, and tag attributes (attribute name, exist) are given to the output module. In the case where there is, an index indicating each description position and hierarchical depth of the attribute value) is output as structured data. Then, when the index created and output by the index creating unit 34 is held by the index holding unit 35 in a state associated with the document data 60 input in step S100 (step S130), this processing is temporarily ended. Will do.

  An example of the index output by the output module of the index creation unit 34 is shown in FIG. This figure shows an index corresponding to the output example of FIG. 10, that is, an index corresponding to the English sentence “Pretty Mary slaps John”. As shown in FIG. 11, each tag, the attribute name and attribute value of each tag, and each word are grouped as 1 (description position): 0 (hierarchy depth), 2: 1. The description position from the top and the hierarchy depth are given. At this time, the tag attributes and attribute values are indexed together with the corresponding tag names. The hierarchy depth is basically based on the depth of each tag, attribute, etc. in the output example of FIG. 10 with respect to the <phrase> tag and its attributes, but in the embodiment, <Sentence For the> tag, <word> tag and their respective attributes, the <entity_name> tag, its attributes, words, etc., the layer depth = 0. Thereby, in an Example, hierarchy depth will show the number of the area | regions enclosed with the tag of the same name as that containing a certain tag area | region. Such an index as structured data treats a sentence (document) as a set of tag areas instead of a so-called tree structure, and provides addresses of tags, words, and tag attributes. Thus, referring to the description position of the index and the hierarchy depth, even if an area defined by a tag having the same name exists outside the area defined by a tag surrounding a word, for example, the area described later By using an algebraic framework, it is possible to easily specify from the smallest tag region surrounding the word to the largest tag region. Further, by referring to the index, it is possible to easily specify the tag correspondence based on the predicate term structure in the sentence as shown by the broken line in FIG.

  As described above, in the management server 30 including the document input unit 31, the syntax analysis unit 32, the term recognition unit 33, and the index creation unit 34 as a data conversion device, for example, new document data indicating a document in a predetermined language When 60 is stored in the database on the data storage device 21, first, sentence syntax analysis using the document data 60 inputted in step S100 and the grammar information and word information of the predetermined language is executed (step S100). Step S110). Subsequently, based on the result of the parsing, a phrase and a word included in the sentence in the document are given a predetermined tag in which a corresponding attribute is described, and a phrase corresponding to the predicate in the sentence or The tag attached to the word is associated with the phrase or the tag attached to the word as the term of the predicate using the attribute, and finally, the description position and the hierarchy depth of each of the tag, the word, and the tag attribute An index indicating that is output as structured data (step S120). The output index is held by the index holding unit 35 in a state associated with the document data 60 (step S130). In the index created in this way, the substantial address of each of the tag, the word, and the tag attribute is indicated as described above. Therefore, referring to the index, the correspondence between the start tag and the end tag is used. Easily create a predicate term structure in a sentence that is defined by the correspondence tree between a tag attached to a syntactic tree of the specified sentence or a tag assigned to a word corresponding to a predicate and a phrase that is a term of the predicate. It becomes possible to specify. Therefore, according to the management server 30 of the embodiment, the notation of the predicate term relationship that is a semantic relationship between the predicate (predicate) included in the sentence in the document and the term, and such a semantic relationship. It is possible to realize an index as structured data that enables a search using a concept defined by the above as a search condition. If such an index is held for each document data 60, the syntax tree of the sentence in the document and the predicate term structure in the predicate sentence can be easily specified by referring to the index. It is possible to search a database in which a plurality of data 60 are stored using a concept defined by a semantic relationship between sentence elements such as words and phrases as a search condition.

  In addition, the syntax analysis unit 32 included in the management server 30 of the embodiment is configured as a parser that executes syntax analysis based on HPSG and acquires the syntax tree and the predicate term structure of the sentence. By parsing, it is possible to make the association between the predicate and the term in the index more appropriate. Further, the term recognition unit 33 included in the management server 30 according to the embodiment is a unit that includes each phrase or word included in the sentence based on predetermined term information in a single substantial meaning. It is determined whether or not there is a phrase, and when the phrase or word is bundled with the one substantive meaning, the substantive meaning is acquired, and the index creation unit 34 is acquired by the term recognition unit 33. The attribute value (ontology ID value) indicating the actual meaning is described in the rel_type attribute and ontolgy attribute of the tag in the index. As a result, in the index, attribute values indicating the same substantive meaning are described in tags corresponding to phrases or words whose surface expression is different in spite of showing substantially the same meaning (concept). Therefore, if the index held by the index holding unit 35 is used, it is possible to easily search for phrases and words having the same substantial meaning, or sentences containing the same.

In the index creation processing of FIG. 2, the management server 30 of the embodiment is applied to an existing database, the document data on the database is individually taken into the management server 30, and an index is created for each document data. It may be executed when. In any case, it is preferable that the index is created in advance based on the document parsing result using the document data 60 and the grammatical information and word information of a predetermined language prior to the search request. As a result, it is possible to reduce the time required from the input of the search condition to the output of the search result by reducing the actual search processing load.
[Search processing]

  Next, a database search process on the data storage device 21 using an index in the database search system 20 of the embodiment will be described.

  FIG. 12 is a flowchart illustrating an example of search processing executed by the management server 30 according to the embodiment. The search process shown in the figure is executed when a search request is made by the user at the terminal 50. In the embodiment, when the management server 30 is accessed from the terminal 50, a search screen as shown in FIG. 13 is displayed on the display screen of the terminal 50. This search screen includes a “subject” box in which a word or phrase corresponding to the subject or main part can be entered, a “verb” box in which a verb or verb phrase corresponding to the predicate or predicate can be entered, an object or An “object” box is provided in which a word or phrase corresponding to the target part can be input. Therefore, if the user inputs a word or phrase as a desired search word in at least one of the “subject” box, the “verb” box, and the “object” box, the user requests a database search on the data storage device 21. be able to. In addition, if a word or phrase as a search word is entered in at least one of the “verb” box, the “subject” box, and the “object” box, the concept or sentence element defined by the semantic relationship between the sentence elements In the embodiment, the search result for the search request is output in sentence units by default, but “Advanced Search” on the search screen. If you click the tab and make detailed settings, you can change the settings so that the search results are output, for example, in document units, and the “Keyword Search” tab is displayed on the search screen as shown in FIG. Click to display a search screen (not shown) for more general keyword search on the display screen, and click the “GCL Search” tab to search the search query described later. Search screen that allows direct input (not shown) is displayed on the display screen.

  In the search screen of FIG. 13, for example, when a search word “Mary” is input in the “subject” box and a search word “slaps” is input in the “verb” box, a signal indicating that is sent from the terminal 50 to the management server 30. Is received by the input receiving unit 36 as a search condition (step S200). The input reception unit 36 gives the search processing unit 37 the received search conditions, that is, the search words “Mary” and “slaps”, and the predicate term relationship between the search words “the subject of“ slaps ”is“ Mary ””. The search processing unit 37 sets a search query based on the received search condition “sentence whose subject is the search word“ Mary ”and predicate is the search word“ slaps ”” (step S210). The search query in this case is as shown in the following formula (1). For example, when the search word “Mary” is input in the “subject” box, the search word “slaps” is input in the “verb” box, and the search word “John” is input in the “object” box, step S210 is performed. A search query like the following formula (2) is set.

[sentence] >> (([[word arg1 = "$ s"] >>"slaps")& ([phrase id = "$ s"]>"Mary"))… (1)
[sentence] >> (([word arg1 = "$ s" arg2 = "$ o"] >>"slaps")& ([phrase id = "$ s"]>"Mary"))& ([phrase id = "$ o"]>"John"))… (2)

  Here, how to set a search query in step S210 will be described. The search process in the database search system 20 according to the embodiment is to obtain a region that meets the search condition using the region algebra framework based on the index held in the index holding unit 35. In the embodiment, the inclusion relation, the order relation, etc. between the areas are designated using the seven operators (>>, >>, <<, <<, &, |,-) shown in FIG. Among these seven operators, characteristic ones are two operators “>” and “<”. In other words, operators “<<” and “>>” that specify inclusion relations between regions in a general domain algebra framework are basically tags with the same name in a region surrounded by tags with a certain name. This is used when searching for the minimum tag region that forms a nested structure under the assumption that no exists. For example, when a query “[phrase] >>“ Mary ”” is executed on the index of FIG. 11, it is defined by <phrase> and </ phrase> tags and includes the word “Mary” (enclosed). ) Only the area corresponding to the <phrase> tag in which the attribute value p5 is described in the id attribute is extracted as the minimum area. On the other hand, the two operators ">" and "<" are in the area enclosed by a tag with a certain name, such as "<phrase> <phrase> A </ phrase> </ phrase>". This tag is introduced to extract all of the tag areas when a tag with the same name exists in For example, when a query “[phrase]>“ Mary ”” is executed on the index of FIG. 11, the area corresponding to the <phrase> tag in which the attribute value p1 is described in the id attribute and the attribute value in the id attribute An area corresponding to the <phrase> tag in which p2 is described and an area corresponding to the <phrase> tag in which the attribute value p5 is described in the id attribute are extracted. In any case, these processes can be quickly executed by referring to the description position and the hierarchy depth indicated in the index. However, processing using “>” or “<” requires a little time compared to processing using “>>” or “<<”. , “>” And “<” are used only for tags that have been found to have a tag with the same name in an area enclosed by a tag with a certain name, such as a <phrase> tag. And preferred.

Further, in order to search for tag areas that do not have a nested structure but have a corresponding relationship, it is necessary to refer to attribute values that are spaced apart from each other on the index. In order to cope with this, in the embodiment, as shown in the equations (1) and (2), the variable $ x (“x” has “s” as the subject and “o” as the object. Is introduced into the search query so that “the attribute value of the attribute A1 of a certain tag T1 is the same as the value of the attribute A2 of another tag T2” can be expressed. For example, in the case of formula (1),
“([Word arg1 =“ $ s ”] >> slaps) & ([phrase id =” $ s ”]>“ Mary ”)”
The part including the variable $ s expresses that the attribute value of arg1 corresponding to “slaps” is the same as the attribute value of the id attribute of phrase including “Mary”. In the embodiment, the variable $ x is basically targeted for the attribute values p1, p2,..., P (n) described in the id attribute and the arg (X) attribute and indicating the predicate term relationship between sentence elements in the sentence. As many as the predicate term relations accepted as the search conditions appear (two variables appear in the expression (2)). After introducing the operators “>” and “<” and the variable $ x as described above, each sentence (document) is treated as a set of tag areas rather than a so-called tree structure, and each of the tag, word, and tag attributes By referring to the index that provides the address, it is possible to execute a quick and highly accurate search while using the concept defined by the semantic relationship between sentence elements as a search condition, as described below. Become.

  When a search query is set in step S210, the search processing unit 37 first identifies the maximum search area that is the maximum search target area that appears in the set search query (step S220). The maximum search area here is the widest area among areas defined by tags such as “sentence”, “phrase”, and “word” that can be included in the search query. When the search query requests “sentence” as a search result as in Expression (1), the maximum search area appearing in the search query is the “sentence” area as in the search result. When the search query is a request for a search query such as “a phrase“ xx included in the sentence “...” ”using an operator“ << ”or the like, the maximum search area corresponds to the search result. This is not the “phrase” area, but the “sentence” area. When the maximum search area is identified in this way, the search processing unit 37 refers to all the indexes in the index holding unit 35 and includes the maximum search area (in this case, the “sentence” area) including all the search words accepted in step S200. A process of extracting all start tags and end tags is executed (step S230), and it is further determined whether or not there is a maximum search area including all search words (step S240). When it is determined in step S240 that the maximum search area including all search words does not exist, the number of hits is zero, so the search processing unit 37 finally indicates that the number of hits is zero. A search result is generated (step S380), and such a search result is transmitted from the result output unit 38 to the terminal 50 where the search request is executed (step S390), and the search process is temporarily ended. . In this case, on the display screen of the terminal 50, for example, a search result such as “No sentence corresponding to“ Mary ”“ slaps ”was found” is displayed.

  On the other hand, when at least one maximum search area including all search words exists in step S240, for example, when the “Sentence” area shown in the index of FIG. If it is determined, the search processing unit 37 acquires the appearance frequency (number of appearances) in each database of the search word accepted in step S200 (step S250). In the embodiment, the appearance frequency is obtained in advance for all words appearing on the database and stored in a predetermined storage device of the data storage device 21 or the management server 30. In step S250, the search processing unit 37 The appearance frequency of each search word accepted in S200 is read from the storage device. Then, the search processing unit 37 sets a partial query for determining the variable $ x included in the search query set in step S210 based on the acquired appearance frequency of each search word (step S260). In step S260, the search processing unit 37 sets a partial query as follows. In this case, first, after the search query set in step S210 is decomposed into parts for each variable $ x, the search word corresponding to one search word having a smaller appearance frequency among search words having a predicate term relationship with each other is handled. Set the part as a partial query. For example, when the search query is shown in the formula (1) and the appearance frequency of the word “Mary” is smaller than the appearance frequency of the word “slaps”, the search query is expressed by the following formulas (3) and (4) ), And what is shown in Expression (4) corresponding to the word “Mary” having a low appearance frequency is set as a partial query. Further, when the search query is shown in the formula (2), the search query is broken down into three parts shown in the following formulas (4) to (6), for example, the word “Mary” and the word “John”. When the appearance frequency is smaller than the appearance frequency of the word “slaps”, the following expressions (4) and (5) are set as partial queries. That is, in step S260, partial queries are set as many as the number of predicate term relationships accepted as search conditions.

[word arg1 = “$ s”] >> “slaps”… (3)
[phrase id = “$ s”]> “Mary”… (4)
[phrase id = “$ o”]> “John”… (5)
[word arg1 = "$ s" arg2 = "$ o"] >>"slaps"… (6)

  When the partial query is set in this way, the search processing unit 37 sets the maximum search area to be processed first among the maximum search areas extracted in step S230 in accordance with predetermined constraints (step S270), and then step Of the partial queries set in S260, a partial query including a search word with a lower appearance frequency is set as an execution partial query (step S280). Then, a query excluding variables from the partial query set in step S280 is executed, and all variable values returned as execution results of the query are acquired and held (step S290). That is, if the search query is shown in Expression (1) and the appearance frequency of the word “Mary” is smaller than the appearance frequency of the word “slaps”, the variable $ s is removed from Expression (4) in Step S280. When the query [phrase]> “Mary” is executed, and the “Sentence” area shown in the index of FIG. 11 is the target maximum search area, for example, p1, p2, and p5 are used as variable values from the index. Is acquired and retained. Next, the search processing unit 37 determines whether all the partial queries set in step S260 have been processed as described above (step S300). If all the partial queries have not been processed, all of them are processed. Steps S270 and S280 are repeated until the above process is completed. That is, when the search query is shown in Expression (2) and the appearance frequency of the word “Mary” is smaller than the appearance frequency of the word “John”, first, for example, the variable $ s is removed from Expression (4). After all the variable values returned as execution results are executed and retained, the query [phrase]> “John” excluding the variable $ o from Expression (5) is executed as the execution result All returned variable values are acquired and held (values p1, p6, and p8 are returned from the index of FIG. 11). When all partial queries are processed, the search processing unit 37 acquires all combinations of variable values obtained for each partial query set in step S260 (step S310). For example, when the search query is as shown in the formula (1), the variable values (p1, p2, p5, etc.) returned as the query execution result obtained by removing the variable $ s from the formula (4) are obtained in step S310. It becomes. When the search query is as shown in the formula (2), the variable values (p1, p2, p5, etc.) returned as the query execution result obtained by removing the variable $ s from the formula (4), and the formula (5) The combination with the variable values (p1, p6, p8, etc.) returned as the execution result of the query excluding the variable $ o from is the acquisition result in step S310. That is, when there are two partial queries, if the variable value obtained based on one partial query is set to px and the variable value obtained based on the other partial query is set to py, (px, py ) Are obtained in step S310.

  After the process of step S310, the search processing unit 37 sets the first set of variable values according to a predetermined constraint from the acquired combinations of variable values (step S320), and sets the variable values of the set set in step S210. Is substituted into the variable $ x corresponding to the search query set in (Step S330). Then, the search processing unit 37 determines whether or not the target maximum search area set in step S270 satisfies the search query after the variable value substitution (step S340). That effect is held in a predetermined storage area as an intermediate result (step S350). After the process of step S350, or when a negative determination is made in step S340, the search processing unit 37 determines whether or not all variable value pairs have been processed (step S360), and step S320 until an affirmative determination is made. To S350 are repeatedly executed. The processing up to this point will be described by taking as an example the case where the “Sentence” area shown in the index of FIG. 11 is the target maximum search area. If the search query is as shown in Expression (1), in step S310 Since a set of variable values is acquired as p1, p2, and p5, first, the value p1 is assigned to the variable $ s in the expression (1) (step S330), and the “Sentence” area is assigned after the value p1 is assigned. It is determined whether or not Expression (1) is satisfied (step S340). In this case, as can be seen from FIG. 11, since “word arg1 =“ p1 ”” does not exist, a negative determination is made in step S340, and then the value p2 is assigned to the variable $ s in equation (1). (Step S330), it is determined whether or not the “Sentence” region satisfies Expression (1) after the value p2 is substituted (Step S340). In this case, as can be seen from FIG. 11, since “word arg1 =“ p2 ”” exists, an affirmative determination is made in step S340, and the <Sentence> tag extracted from the index of FIG. The Sentence> tag is held as an intermediate result (step S350). If the search query is as shown in equation (2), a set of variable values is acquired as (p1, p1), (p1, 6),... (P2, p8),. First, the value p1 is substituted into the variables $ s and $ o of the equation (2) (step S330), and whether or not the “Sentence” area satisfies the equation (2) after the substitution of the value p1. A determination is made (step S340). In this case, as can be seen from FIG. 11, since “word arg1 =“ p1 ”” and “word arg2 =“ p1 ”” do not exist, a negative determination is made in step S340. Thereafter, the processes in steps S330 and S340 are repeated. In step S330, when the value p2 is substituted for the variable $ s in the equation (2) and the value p8 is substituted for the variable $ o (step S330), “ It is determined that word arg1 = "p2" "and" word arg2 = "p8" "exist, and the <Sentence> tag and </ Sentence> tag extracted from the index of FIG. 11 are held as intermediate results (Step S350). If the search query of the formula (2) is set in step S210 and the appearance frequency of the search word “slaps” is the minimum among the three search words, the above-described operation is performed in step S280. The query of formula (6) is set as the partial query to be executed first, and when the query excluding the variable $ x is executed (step S290), all the variable values are acquired at that stage. become. Therefore, in this case, it is not necessary to process any partial query of formula (4) or (5). If the search query of the formula (2) is set in step S210 and the appearance frequency of the search word “slaps” is the second smallest among the three search words, in step S280. A query based on the above formula (4) or (5) is executed first, and then a query based on the above formula (6) is executed. In this case, when the query based on the formula (4) or (5) is first executed, the variable value corresponding to either the variable $ s or $ o is acquired. When executing a query based on, it is only necessary to acquire the other variable value that has not been acquired.

  After the process of step S360, the search processing unit 37 determines whether or not all the maximum search areas extracted in step S230 have been processed (step S370), and if all the maximum search areas have not been processed. The processes from step S270 to S360 are repeated until the process is completed. When an affirmative determination is made in step S370, the search processing unit 37 creates a final search result based on the intermediate result held in the storage area (step S380). In this case, if no intermediate result is held in the storage area, a final search result indicating that the number of hits is zero is created in step S380. Then, the final search result is transmitted from the result output unit 38 to the terminal 50 where the search request has been executed (step S390), and the search process is temporarily ended.

  As described above, the database search system 20 according to the embodiment includes the index holding unit 35 that holds an index created for each document data 60. Then, as described above, the index for each document data 60 is a predetermined tag assigned to each sentence in the corresponding document, a phrase and a word included in the sentence, and an attribute described in the word and the tag. Indicating the description position and depth of each of the tags, and associating the tag attached to the phrase or word corresponding to the predicate in the sentence with the tag attached to the phrase or word serving as the term of the predicate It is. Therefore, referring to the index, the sentence syntax tree in the document defined by the correspondence between the start tag and the end tag, the tag attached to the word corresponding to the predicate and the phrase that is the term of the predicate, etc. It is possible to easily specify the predicate term structure in the sentence defined by the correspondence relationship with the assigned tag. Based on this, in the database search system 20 of the embodiment, when a plurality of search words and a predicate term relationship between the search words are received as search conditions (step S200), the search words received based on the index Are extracted (step S230). More specifically, based on the search word received in step S200 and the predicate term relationship, a search query with the attribute value p (n) associating the predicate term relationship as a variable $ x is set (step S210), after the maximum search area that is the maximum search target area appearing in the set search query is identified (step S220), the start tag of the maximum search area including all the search words accepted in step S200, and An end tag is extracted (step S230). The address of the attribute value p (n) of the id attribute or arg (x) attribute that associates the index, that is, the tag attached to the word corresponding to the predicate with the tag attached to the phrase that is the term of the predicate Based on the above, it is determined whether or not the correspondence corresponding to the predicate term relationship accepted in step S200 is made between the tags corresponding to the search words included in the extracted tag region (maximum search region). (Step S340). More specifically, a query is executed by removing the variable $ x from at least one partial query including any of the search words obtained by decomposing the search query set in step S200 and the variable $ x (step S200). S290) After determining whether or not there is a maximum search area that matches the search query substituted with the variable value p (n) returned as the execution result (steps S320 to S340), the result of the determination is Based on this, a search result is output (steps S380 and S390). As a result, according to the database search system 20 of the embodiment, the database is a database on the data storage device 21 in which a plurality of document data 60 is stored. While searching, it becomes possible to quickly obtain a more accurate search result. Further, in the database search system 20 of the embodiment, the attribute value described in the tag corresponding to one search word having a smaller appearance frequency among search words having a predicate term relationship is used as a variable $ x, and the variable The value p (n) is acquired, and it is determined whether or not there is a tag corresponding to the other search word in which the attribute value matching the acquired variable value p (n) is set (steps S320 to S320). S360). That is, in the above-described embodiment, the query including one search word having a smaller appearance frequency among search words having a predicate term relationship among the queries obtained by decomposing the search query set in step S210. Is set as a partial query (step S260), and an attribute value described in a tag corresponding to one search word having a smaller appearance frequency is obtained as a variable value based on the partial query (step S290). As described above, if the attribute value described in the tag corresponding to one search word having a smaller appearance frequency is used as a variable, the number of variable values acquired in step S290 is reduced, and the determination is made accordingly. The number of tags corresponding to the other search word to be targeted can be reduced. As a result, it is possible to efficiently determine in a short time whether or not there is a tag pair corresponding to a search word in which the same value is described as an attribute value.

In the search process of FIG. 12, after obtaining variable values for each partial query based on all partial queries set in step S260 (step S290), a combination of variable values is obtained (step S310). ) Substituting a set of variable values into the search query (step S330), it is determined whether or not the target maximum search area satisfies the search query after the variable substitution (step S340), but this is limited to this. It is not a thing. That is, in order to improve the efficiency of the search process and execute a quicker search, the following search process as a modified example according to the algorithm illustrated in FIG. 15 may be executed. In this case, for example, when two partial queries are set in step S260, one partial query corresponding to a search word with a smaller appearance frequency (first search) in the target maximum search area set in step S270. The query obtained by removing the variable $ x from (partial query 1) is executed, and only one variable value is acquired first. Next, a query obtained by removing $ x from the other partial query (second partial query) is executed to sequentially obtain variable values, and the one variable value and the second partial query based on the first partial query are acquired. It is determined whether or not the target maximum search area satisfies a search query substituted with a variable value based on. If a negative determination is made here, the second variable value is acquired again based on the first partial query. Then, a query obtained by removing $ x from the second partial query is executed again to sequentially obtain variable values, and the second variable value based on the first partial query and a variable based on the second partial query are obtained. It is determined whether the target maximum search area satisfies the search query with the value substituted. Thereafter, it is determined whether or not there is a maximum search area that matches the search condition by sequentially repeating these processes. According to the search processing of such a modification,
It is possible to achieve a quicker search by appropriately reducing the tag area to be searched.

  The present invention is useful in the field of information processing.

It is a schematic block diagram which shows the database search system 20 which concerns on one Example of this invention. It is a flowchart which shows an example of the index creation process performed in the management server 30 contained in the database search system 20 of an Example. It is explanatory drawing which shows an example of the directed graph with a label expressing the semantic relationship between sentence elements. It is explanatory drawing for demonstrating the procedure of the syntax analysis using the word information prepared with respect to the syntax analysis part 32 of the management server 30 of an Example, and the said word information. It is explanatory drawing for demonstrating the procedure of the syntax analysis using the word information prepared with respect to the syntax analysis part 32 of the management server 30 of an Example, and the said word information. It is explanatory drawing which shows an example of the syntax tree obtained by parsing a sentence. It is explanatory drawing which illustrates ontology in a specific field. It is a chart which illustrates the kind of tag used for structuring a document, and the attribute described in it at the time of creation of the index in management server 30 of an example. FIG. 8 is a chart illustrating contents of attributes described in tags illustrated in FIG. 7. FIG. It is explanatory drawing which shows the output example at the time of outputting the English text structured by the index preparation part 34 of the management server 30 of an Example in XML format. It is explanatory drawing which shows an example of the index produced by the index production part 34 of an Example. It is a flowchart which shows an example of the search process performed by the management server 30 of an Example. It is explanatory drawing which shows an example of the search screen displayed on the display screen of the terminal 50 connected to the management server 30 of an Example. It is a chart which illustrates the operator used at the time of the search process in the database search system 20 of an Example. It is explanatory drawing which shows an example of the algorithm for implement | achieving the modification of the search process in the database search system 20 of an Example.

Explanation of symbols

  20 database search system, 21 data storage device, 30 management server, 31 document input unit, 32 syntax analysis unit, 33 term recognition unit, 34 index creation unit, 35 index holding unit, 36 input reception unit, 37 search processing unit, 38 Result output unit, 40 networks, 50 terminals, 60 document data.

Claims (12)

A data conversion device that converts data indicating a sentence in a predetermined language into structured data including grammatical semantic annotations,
A syntax analysis means for performing syntax analysis of the sentence using the data and grammatical information and word information of the predetermined language;
Based on the result of the parsing by the parsing means, a predetermined tag in which a corresponding attribute is described is assigned to a phrase and a word included in the sentence, and a phrase or word corresponding to a predicate in the sentence After associating a given tag with a tag given to a phrase or word as a term of the predicate by the attribute, the description position and the hierarchical depth of each of the tag, the word, and the tag attribute Structured data output means for outputting the indicated index as the structured data;
A data conversion device comprising:   The data conversion apparatus according to claim 1, wherein the syntax analysis unit is a parser that executes syntax analysis based on HPSG to acquire a syntax tree and a predicate term structure of the sentence. The data change device according to claim 1 or 2,
For each phrase or word included in the sentence, it is determined whether or not the phrase or word is enclosed in one substantial meaning based on predetermined term information, and the phrase or word has the substantial meaning Further comprising a term recognition means for acquiring the substantive meaning when
The structured data output means is a data conversion device that describes an attribute value indicating the acquired substantive meaning in a predetermined attribute of a corresponding tag. A data conversion method for converting data indicating a sentence in a predetermined language into structured data including grammatical semantic annotations,
(A) performing a syntax analysis of the sentence using the data and grammatical information and word information of the predetermined language;
(B) Based on the result of the syntax analysis in step (a), a predetermined tag describing a corresponding attribute is assigned to a phrase and a word included in the sentence, and a phrase corresponding to a predicate in the sentence Alternatively, the tag assigned to the word and the phrase or tag assigned to the word as a term of the predicate are associated with each other by the attribute, and the description position and the hierarchy of each of the tag, the word, and the tag attribute Outputting an index indicating depth as the structured data;
Data conversion method including A database management apparatus that manages a database that stores a plurality of document data of documents in a predetermined language,
Syntax analysis means for executing syntax analysis of the document using the document data and grammatical information and word information of the predetermined language;
Based on the result of the parsing by the parsing means, a predetermined tag describing attributes corresponding to sentences in the document and phrases and words included in the sentence is given, and predicates in the sentence are given. Description of each of the tag, the word, and the attribute of the tag after associating the tag attached to the corresponding phrase or word with the tag attached to the phrase or word as the term of the predicate by the attribute Index output means for outputting an index indicating a position and a hierarchy depth;
Index holding means for holding the index output by the index output means for each document data;
A database management device comprising: A database management method for managing a database storing a plurality of document data of documents in a predetermined language,
(A) performing syntax analysis of the document using the document data and grammatical information and word information of the predetermined language;
(B) Based on the result of the parsing in step (a), a predetermined tag describing attributes corresponding to sentences in the document and phrases and words included in the sentence is given, The tag assigned to the phrase or word corresponding to the predicate of and the tag assigned to the phrase or word serving as the term of the predicate are associated by the attribute, and the attribute of the tag, the word and the tag Outputting an index indicating each description position and hierarchy depth and holding each document data; and
Database management method including. A database search system for searching a database storing a plurality of document data of documents in a predetermined language,
The sentence in the document, a predetermined tag assigned to each of the phrase and the word included in the sentence, the description position and the hierarchical depth of each attribute described in the word and the tag, and the sentence An index created for each piece of document data is stored so that a tag assigned to a phrase or word corresponding to the predicate of and a tag assigned to a phrase or word serving as a term of the predicate are associated by the attribute. Index holding means;
Accepting means capable of accepting a plurality of search words and a predicate term relationship between the search words as a search condition;
Extraction means for extracting a tag region including all the accepted search words based on the index held by the index holding means;
Determining means for determining, based on the index, whether or not the matching corresponding to the accepted predicate term relationship is made between the tags corresponding to the search word included in the tag area extracted by the extracting means. When,
Search result output means for outputting a search result based on a determination result by the determination means;
A database search system comprising: The database search system according to claim 7, wherein
In the index, a predetermined attribute of a tag corresponding to a phrase or word corresponding to a predicate in the sentence and a predetermined attribute of a tag corresponding to a phrase or word serving as a term of the predicate have the same value as an attribute value. Is set,
The database search system for determining whether or not there is a tag pair corresponding to the search word in which the same value is described as the attribute value based on the index. The database search system according to claim 8, wherein
An appearance frequency storage means for storing the appearance frequency in the database for each word;
The determination means obtains the value of the variable using the attribute value described in the tag corresponding to one search word having a smaller appearance frequency among the search words having a predicate term relationship as a variable, A database search system that determines whether or not a tag corresponding to the other search word in which the attribute value that matches the value of the acquired variable is set exists. The database search system according to claim 9, wherein
Search query setting means for setting a search query using a variable as an attribute value that associates the predicate term relationship based on the accepted search word and the predescription term term relationship,
The extraction means identifies a maximum search area that is a maximum search target area that appears in the set search query and extracts a start tag and an end tag of the maximum search area including all the accepted search words,
The determination means executes a query excluding the variable from at least one partial query including any of the search words obtained by decomposing the set search query and the variable, and is returned as an execution result. A database search system that determines whether or not the maximum search area that matches the search query into which the value of the variable is substituted exists.   11. The database search system according to claim 7, wherein the index is created in advance based on a result of syntax analysis of the document using the document data and grammatical information and word information of the predetermined language. . A database search method for searching a database storing a plurality of document data of documents in a predetermined language,
(A) indicates a sentence in the document, a predetermined tag assigned to each of a phrase and a word included in the sentence, a description position and a hierarchical depth of each attribute described in the word and the tag, and An index created for each document data so that a tag attached to a phrase or word corresponding to a predicate in the sentence and a tag attached to a phrase or word serving as a term of the predicate are associated by the attribute. Creating and maintaining
(B) receiving a plurality of search words and a predicate term relationship between the search words as a search condition;
(C) extracting a tag region including all the accepted search words based on the index;
(D) Based on the index, whether or not the correspondence corresponding to the accepted predicate term relationship is made between the tags corresponding to the search word included in the tag area extracted in step (c). A determining step;
(E) outputting a search result based on the determination result in step (d);
Database search method including.