JP5915274B2 - Information search method, program, and information search apparatus - Google Patents

Description

  The present technology relates to a text document search technology.

  In a search using a well-known keyword, a query in keyword units is used, so the relationship between keywords cannot be included in the search condition. For example, when the user wants to search for “technology for reducing environmental load by reducing heat release”, the keyword “heat release, environmental load, reduction” is input. However, this keyword query includes ambiguity. In this case, the meanings of both “reducing heat release” and “reducing environmental load” are included, but the meaning is not limited to one. Also, the relationship between heat release and environmental load is not specified.

Thus, the keyword search in which the ambiguity is included in the query has the following problems.
(1) The search is not performed as intended by the user.
(2) An unintended document is searched even if it contains a keyword.

  In particular, the problem (2) includes a drawback that it is troublesome to find out the part intended by the user from the extracted document. For example, even when a sentence including all keywords is detected by searching with three keywords, the sentence may include an unintended relationship. In other words, even if a portion that hits a keyword is output, it may not always be the information that the user wants. Therefore, the user spends time in determining to extract useful sentences.

Japanese Patent Laid-Open No. 05-197758

“Natural Language Understanding” edited by Junzo Tanaka and Junichi Sakurai (Ohm, 1988) Steven Bird, Ewan Klein, Edward Loper Masato Sugawara, Takahiro Nakayama, Takaaki Mizu "Introduction to Natural Language Processing" O'Reilly Japan, 2010. WEB site "Japanese Natural Language Processing with Python" (http://nltk.googlecode.com/svn/trunk/doc/book-jp/ch12.html)

  An object of the present technology is, in one aspect, to extract a sentence intended by a user who intends to search.

  The information search method according to one aspect of the present technology is as follows: (A) Search for a semantic minimum unit composed of two words and each element of the semantic relationship between the words, or one word and each element of the role of the one word. (B) A search target including a minimum semantic unit that matches the minimum semantic unit of the search key from the data storage unit storing the minimum semantic unit included in the search target sentence for each search target sentence. The process includes specifying a sentence and outputting a search result for the specified search target sentence.

  The sentence intended by the user can be extracted.

FIG. 1 is a diagram illustrating an example of a module configuration of the information search apparatus. FIG. 2A is a diagram illustrating an example of an effective graph. FIG. 2B is a diagram illustrating an example of the smallest semantic unit. FIG. 3A is a diagram illustrating an example of the position of the original text. FIG. 3B is a diagram illustrating an example of a search index. FIG. 4 is a diagram showing an example of the flow of search target document preparation processing. FIG. 5 is a diagram illustrating an example of a search target document. FIG. 6 is a diagram illustrating an example of a search target document. FIG. 7 is a diagram illustrating an example of a search target document. FIG. 8 is a diagram illustrating an example of a search target document. FIG. 9 is a diagram illustrating an example of the evaluation value table. FIG. 10A is a diagram illustrating an example of a flow of search processing. FIG. 10B is a diagram illustrating an example of a flow of search processing. FIG. 11 is a diagram illustrating an example of a query and a minimum semantic unit. FIG. 12 is a functional block diagram of a computer.

  FIG. 1 shows an example of the module configuration of the information search apparatus. The information search apparatus includes a search target document database 11, a search index 13, an evaluation value table 15, an input unit 31, a semantic analysis unit 33, a semantic minimum unit generation unit 35, a search unit 37, an evaluation value calculation unit 39, and a ranking unit 41. And an output unit 43. The input unit 31 receives a natural language sentence search query by a user operation. The semantic analysis unit 33 performs semantic analysis of the search query accepted by the input unit 31. In this process, a natural language sentence is analyzed, and a relationship between concepts (word meanings) in the sentence is expressed by a directed graph. The semantic minimum unit generator 35 generates a semantic minimum unit from the directed graph output by the semantic analyzer 33. The semantic minimum unit will be described later.

  The search index 13 is a database that stores a minimum semantic unit of each sentence included in a search target document. In the preparatory process performed in advance, semantic analysis is performed on a plurality of search target documents, a minimum semantic unit is generated for each sentence in the documents, and a search index 13 is generated. Further, in the preparation process, the document frequency of each meaning minimum unit included in the search index 13 is calculated. The document frequency is the number of search target documents in which the smallest semantic unit appears. In the preparation process, the idf value of each semantic minimum unit is further calculated using these document frequencies. The idf value for each semantic minimum unit is stored in the evaluation value table 15.

  The idf value of each minimum semantic unit is log (total number of documents / number of documents including the minimum semantic unit). The idf value is larger when the number of search target documents including the smallest semantic unit is smaller. The idf value is an example of an evaluation value in the smallest semantic unit. The evaluation value of the smallest semantic unit may be another value. It is desirable that the evaluation value indicates the usefulness of the smallest semantic unit.

  The search unit 37 searches the search index 13 on the condition of the smallest semantic unit generated from the search query. For example, a plurality of semantic minimum units are used in the OR condition. A document corresponding to the matched minimum semantic unit is selected from the search index 13. The evaluation value calculation unit 39 refers to the evaluation value table 15 and the search index 13 and calculates the evaluation value of the document including each sentence extracted based on the matched meaning minimum unit. The ranking unit 41 ranks each extracted document. That is, the ranking unit 41 sorts the document evaluation values calculated by the evaluation value calculation unit 39 using the sort key. The output unit 43 outputs the ranked result. The output form is, for example, display, printing, or transmission. The extracted documents are arranged in a useful order and presented to the user. For example, the extracted sentence is displayed.

  Next, a directed graph and a minimum semantic unit will be described. The minimum semantic unit represents a partial structure of a directed graph obtained as a result of semantic analysis. The directed graph includes nodes and arcs. FIG. 2A shows an example of an effective graph, and FIG. 2B shows an example of a minimum semantic unit. The original text is "Taro gave a book to Hanako." The directed graph in the figure is the result of semantic analysis of this original text.

For directed graphs, the following documents are helpful.
・ Takenaka Tanaka and Junichi Sakurai “Natural Language Understanding” (Ohm, 1988)
・ Steven Bird, Ewan Klein, Edward Loper Masato Sugawara, Takahiro Nakayama, Takaaki Mizu “Introduction to Natural Language Processing” O'Reilly Japan, 2010.
・ Web site "Japanese natural language processing with Python" (http://nltk.googlecode.com/svn/trunk/doc/book-jp/ch12.html)

  The node represents the concept (meaning) of a word in the input sentence. “Give”, “Book”, “Taro”, and “Hanako” are examples of nodes. A symbol representing the concept (referred to as a concept symbol) is added to each node. “GIVE”, “BOOK”, “TARO”, and “HANAKO” are examples of conceptual symbols.

  An arc represents the relationship between nodes or the role of nodes. If the arc is between two nodes, the arc represents the relationship between the two nodes. For example, an arc drawn from the “GIVE” node to the “BOOK” node in the figure is named “target”. This display means that “BOOK” is an object of “GIVE”. On the other hand, an arc having no end node represents the role of the start node. For example, one of the arcs whose starting point is “GIVE” and whose end point is not a node in the figure is named “past”. “GIVE” means a past role.

In the generation of the smallest semantic unit, an arc is extracted from the directed graph and processed as follows.
(1) When the arc connects two nodes, (the starting point node where the arc comes out, the end point node where the arc goes, and the name of the arc) are output as the minimum semantic unit for each arc.
(2) If there is no origin node where the arc is generated, (NIL, end node toward which the arc is directed, arc name) is output as the smallest semantic unit.
(3) If there is no end node to which the arc is headed, (starting node from which the arc occurs, “NIL”, arc name) is output as the smallest semantic unit.

  Thus, the minimum semantic unit represents the relationship between two concepts in a sentence or the role of a concept. By searching the database using this as a search key, a search reflecting the intention of the searcher embedded in the natural language sentence can be performed.

  Since the minimum semantic unit is based on the partial structure of the directed graph, the search based on the matching of the minimum semantic unit is more flexible than the search based on the matching of the directed graph. In addition, the idf value of the minimum semantic unit included in the search target document is prepared in advance, the idf value of the minimum semantic unit matched is specified, and the idf value is used to determine the document including the sentence about the matched minimum semantic unit. An evaluation value can be calculated. The evaluation value of the document is useful for ranking.

  Hereinafter, an example of a processing flow will be described. A feature of this process is that semantic analysis is performed on each query and each sentence included in each search target document, each minimum semantic unit is obtained, and the minimum semantic unit is searched as a search key. Furthermore, ranking is performed by calculating an evaluation value of the extracted document using the idf value of the smallest semantic unit.

  In this embodiment, a query of a natural language sentence is accepted and a minimum semantic unit is automatically generated from the natural language sentence. However, the minimum semantic unit can be specified by receiving a query of the minimum semantic unit.

  In this embodiment, it is assumed that a patent database is searched with a query sentence “having means for registering a word corrected by a translator in a dictionary”. The purpose of this example is to retrieve patent specifications that are closely related to the query statement.

  Prior to the search stage, a database of search indexes 13 is generated in advance. FIG. 3A shows an example of the position of the original text, and FIG. 3B shows an example of a search index. The search index 13 includes, for each minimum semantic unit, the content of the minimum semantic unit, the document ID of the original document, the sentence ID of the original sentence, the position of the start node, the character string length of the start node, and the position of the end node. And a record that associates the character string length of the end node. The position of the start point node is the start position in the sentence of the character string corresponding to the node. The same applies to the position of the end node.

  FIG. 4 is a diagram showing an example of the flow of search target document preparation processing. This process is executed by a preparation unit (not shown). For each search target document included in the search target document database 11 (S101), the preparation unit divides the document into a plurality of sentences (S103). For each sentence (S105), the preparation unit performs semantic analysis, generates a minimum semantic unit, and adds a new record to the search index (S107).

  5 to 8 are diagrams illustrating examples of search target documents. FIG. 5 shows a sentence m included in the search target document A “The machine translation system according to the present invention has registration means for registering a word corrected by a translator or a corrected or created translation sentence in the dictionary and the example sentence database, respectively. Is shown as an example. Further, FIG. 5 shows an example of the minimum semantic unit generated by the semantic analysis of the sentence m. The meaning represented by the smallest semantic unit in the figure is for reference only. The generated minimum semantic unit is stored in association with the original document A and the sentence m. This association information is used, for example, for calculating an evaluation value. FIG. 6 similarly shows the sentence n included in the search target document B, the meaning minimum unit, and the meaning represented by the meaning minimum unit. FIG. 7 similarly shows the sentence p included in the search target document C and the meaning represented by the minimum semantic unit and the minimum semantic unit. FIG. 8 similarly shows the sentence q included in the search target document D, and the meaning represented by the smallest semantic unit and the smallest semantic unit.

  Returning to the description of the processing in FIG. 4, when the preparation unit completes the generation of the smallest semantic unit and the registration of the search index (S109 and S111), it calculates an idf value for each generated smallest semantic unit, and evaluates the evaluation value. It stores in the table 15 (S113). In the figure, df indicates the document frequency. FIG. 9 is an example of an evaluation value table. The minimum semantic unit and the idf value are associated with each other.

  Next, document search will be described. 10A and 10B are diagrams illustrating an example of a flow of search processing. The input unit 31 accepts a natural language sentence query (S121). For example, the query is input from a character input device or received via a network. When a query including a plurality of sentences is accepted, the query is divided into sentences (S123). This dividing process is executed by the input unit 31 or a dividing unit (not shown). For each sentence (S125), the semantic analyzer 33 performs semantic analysis, and the semantic minimum unit generator 35 generates a semantic minimum unit. The semantic analysis unit 33 adds the generated minimum semantic unit to the set of search keys (S127). The above-described processes of S125 and S127 are repeated for all sentences in the query (S129).

  FIG. 11 is a diagram illustrating an example of a query and a minimum semantic unit. An example of the original query “has means for registering a word corrected by a translator in a dictionary” is shown. Further, FIG. 11 shows an example of the minimum semantic unit generated by the semantic analysis of the query sentence. The meaning represented by the smallest semantic unit in the figure is for reference only.

  The semantic analysis unit 33 specifies a search key including one or more minimum semantic units (S131). The search unit 37 performs the following processing for each semantic minimum unit (S133). The search unit 37 extracts a matching minimum semantic unit from the search index 13. The matching of the smallest semantic unit is determined here based on whether or not it is a perfect match. That is, it is determined that they match only when the start point node, end point node, and arc names included in the smallest semantic unit are the same. In this example, all of the sentence m, the sentence n, the sentence p, and the sentence q in FIGS. 5 to 8 match. Then, the semantic analysis unit 33 reads information (document ID and sentence ID) for identifying a sentence corresponding to the matched minimum semantic unit, and temporarily stores it as data about the matched sentence (S135). When the search unit 37 has processed all the smallest semantic units in the search key (S137), the process proceeds to document evaluation processing.

  First, the evaluation value calculation unit 39 sets the evaluation values of all documents to 0 (S139). Then, the evaluation value calculation unit 39 repeats the following processing for each matched sentence (S141). The evaluation value calculation unit 39 calculates the evaluation value of the sentence (S143). Then, the evaluation value calculation unit 39 adds the evaluation value of the sentence to the evaluation value of the document including the sentence (S145). The evaluation value calculation unit 39 processes all of the matched sentences, thereby obtaining the evaluation value of the document (S147). The evaluation value of a document is the sum of evaluation values of sentences included in the document. The ranking unit 41 sorts the documents using the document evaluation values as sort keys (S149). Documents are sorted in descending order of evaluation. Then, the output unit 43 outputs the rearranged document data (S149). For example, the output unit 43 displays a matched sentence included in the document. In this case, the ranking of the document is indirectly the ranking of the sentence. The output unit 43 acquires a document corresponding to the document ID from the search target document database 11 and acquires a sentence corresponding to the matched sentence ID from the search target document database 11.

  It is also effective to sort the extracted sentences using the evaluation values of the extracted sentences as sort keys and display the extracted sentences in the sort order.

  The calculation of the sentence evaluation value will be described. The evaluation value of the sentence is expressed by the following expression.

Evaluation value (Sn) of sentence n = [set of minimum semantic units of query (K1, K2,..., Ki,...)] ((Idf value of Ki appearing in sentence n × number of occurrences of Ki in sentence n)) Sum] x M squared

However, M = number of K types that appear simultaneously in sentence n

  M helps to evaluate that it covers the whole query. Further, the use of the square value of M increases the degree of evaluation.

  The number of occurrences of Ki in the sentence n is the number of minimum semantic units that match the minimum semantic unit specified as the search key among the minimum semantic units included in one search target sentence. M is the number of types of minimum semantic units that match one of the minimum semantic units included in one search target sentence and specified as a search key.

  A specific evaluation value calculation process will be described. First, the evaluation value calculation unit 39 acquires the idf value of the minimum semantic unit from the evaluation value table 15 when the minimum semantic unit of the query appears in the sentence to be calculated. The evaluation value calculation unit 39 further counts the number of records including the sentence ID and the minimum semantic unit from the search index 13 to obtain the number of appearances described above. Then, the evaluation value calculation unit 39 calculates the product of the idf value and the number of appearances. The evaluation value calculation unit 39 calculates the sum of the products obtained in all the smallest semantic units that have appeared. Then, the evaluation value calculation unit 39 counts the number of different semantic minimum units that appear in the sentence to be calculated, and obtains M. Further, the evaluation value calculation unit 39 obtains the square of M and multiplies the calculated sum of products. In this way, the evaluation value calculation unit 39 calculates the evaluation value of the sentence.

  The evaluation value calculation unit 39 may select a minimum semantic unit for obtaining a product and a minimum semantic unit for counting to M under specific conditions. For example, when the idf value is less than the threshold value, the evaluation value calculation unit 39 may perform processing so that the product related to the smallest semantic unit is not added to the sum. Similarly, a minimum meaning unit whose idf value is less than the threshold value may not be included in the M count.

  Thus, the evaluation value of a sentence is the sum of the evaluation values in the sentence of the smallest semantic unit (or part thereof) that appears in the sentence and the smallest semantic unit (or part thereof) that appears in the sentence for the query. It is expressed as the product of the degree of coverage.

  The evaluation value of the document is expressed by the following formula.

Document evaluation value (D) = total of evaluation value (Sn) of sentence n

As shown in FIG. 10B, the evaluation value calculation unit 39 calculates the evaluation value of the document by adding the evaluation values of the sentences included in the document.

An example of calculation of an evaluation value is shown. In order to simplify the explanation, it is assumed that the idf value of all the semantic minimum units is 2.0. For example, the sentence m matches six minimum semantic units. The number of occurrences of each semantic minimum unit in the sentence m is one. The number of minimum semantic units appearing in the sentence m is six. Therefore, the evaluation value of sentence m is
(2.0 * 1 + 2.0 * 1 + 2.0 * 1 + 2.0 * 1 + 2.0 * 1 + 2.0 * 1) * (6 ^ 2) = 432.0
It becomes. The evaluation value of sentence n is
(2.0 * 1 + 2.0 * 1) * (2 ^ 2) = 16.0
It becomes. The evaluation value of sentence q is
(2.0 * 1 + 2.0 * 1 + 2.0 * 1) * (3 ^ 2) = 54.0
It becomes. The evaluation value of sentence p is
(2.0 * 1) * (2 ^ 1) = 4.0
It becomes. As described above, as the sentence including more minimum semantic units related to the search query becomes larger, M ² becomes larger and is highly evaluated.

  Next, the evaluation value of the document is calculated. The evaluation value of document A is the sum of the evaluation values of all sentences included in document A. Assume that document A consists of two sentences, sentence i and sentence m. If the evaluation value of sentence i is 18.0, the evaluation value of document A is 18.0 + 432.0 = 450.0.

The evaluation values of Document B, Document C, and Document D are calculated in the same manner, and are assumed to be 106.0, 253.0, and 90.0, respectively. Rank documents in descending order by evaluation value. The result is as follows.
Search result 1st place: Document A (Evaluation value = 450.0)
Search result 2nd place: Document C (Evaluation value = 253.0)
Search result 3rd place: Document B (Evaluation value = 106.0)
Search result 4th place: Document D (Evaluation value = 90.0)

  In this example, even if the expressions are different or different expressions are included between words, the documents including the information having the same meaning are ranked higher.

  As shown below, sentences with the same meaning match even if the expressions are different. Even if two words in a sentence are separated by another word string, the original meaning is not lost, and matching is successful even in such a case.

  For example, the expression “translator corrected” in the sentence m matches the expression “corrected by translator” in the query. The smallest semantic unit is both (correction, translator, actor). The smallest semantic unit derived from the expression “register word or correction or... In the dictionary and example sentence database” of sentence m is (registration, word, object). Therefore, in this case, it matches the query. The smallest semantic unit derived from the expression “word may be registered in dictionary 3” of sentence q is (registration, word, object). In this case, “word” is not the subject of “registration”, but is interpreted as an object. Therefore, in this case, it matches the query.

  For example, when the user selects a document, the output unit 43 may highlight a sentence with a high evaluation value included in the document. Since the evaluation value for each sentence is calculated, such a display is also possible.

This example has the following effects, for example.
1. You can search by specifying the meaning of a word and its relationship.
2. The extracted documents can be ranked and output in descending order of relation to the query.

Note that the document evaluation value calculation method includes, for example, the following variations.
1. Total of evaluation values (Sn) of sentence n: calculation method 2 described above The sentences included in document D are sorted in order of evaluation value, and the highest m evaluation values from the highest evaluation value are totaled.
3. Only the evaluation values of sentences with a certain value k or more are totaled.
4). The sum of the evaluation values x of the sentence numbers _{i to i + h is} summed up (Σ_i (x _i · x _{i + 1} ... X _{i + h} )).
This value is useful for evaluating the synergistic value of multiple sentences.
5. 4. above. As a variation of the above, the average value of the evaluation values of the sentence numbers i to i + h is summed up to the power h.
This value is useful for stressing and evaluating the value density of the entire document.
6). 4. above. As variations of the above, regarding the evaluation values of sentence numbers i to i + h, those having an evaluation value of k or less are excluded from multiplication, and the evaluation values of i to i + h are totaled.
This value helps prevent overestimation due to the accumulation of less important information.

Further, the sentence evaluation value calculation method includes, for example, the following variations.
1. Evaluation value (Sn) of sentence n = [(id, value of Ki appearing in sentence n × number of occurrences of Ki in sentence n) of the set of smallest semantic units of the query (K1, K2,... Ki,...) Total] X M squared, where M = number of K appearing simultaneously in sentence n: calculation method 2 above. Above 1. In the set of minimum semantic units, the minimum semantic unit that meets the specific condition (s) is excluded from the calculation. For example, conditions such that the arc is “center”, the arc is “past”, the idf value is a minimum semantic unit with a threshold value or less, and NIL is included.
3. Above 1. In regard to “M = number of semantic minimum units K appearing simultaneously in sentence n”, the minimum semantic units that meet a specific condition (one or more) are not counted.
4). Above 1. Variation of “M-squared”: Power of M (log (M))
5. Above 1. 5. Replace “Meaning unit” with “Separated smallest unit”. Above 1. 6. Replace “Meaning unit” with “Meaning unit and morpheme”. Above 1. 7. Replace “Meaning unit” with “Meaning unit divided into smallest unit”. Above 1. 8. Replace “Meaning unit” with “Meaning unit divided by morpheme”. Above 1. 9. Replace “Meaning unit” with “Separated smallest unit and morpheme”. 2. To 9. Any combination of

  The above-described “divided minimum semantic unit” will be described. The divided meaning minimum unit is a unit obtained by dividing the meaning minimum unit (node 1, node 2, arc) into (node 1, arc) and (arc, node 2). However, when the arc is NIL, it is assumed that (node 1, node 2, arc).

  For example, the minimum semantic units divided by the minimum semantic unit (GIVE, BOOK, target) are (GIVE, target) and (target, BOOK). Further, the smallest semantic unit divided by the smallest semantic unit (GIVE, NIL, predicate) is (GIVE, NIL, predicate).

  A search using “divided minimum semantic unit” corresponds to a matching search for a predetermined portion using “minimum semantic unit”. The predetermined part means a combination of predetermined elements such as (node 1, arc) and (arc, node 2) among the elements of node 1, node 2, and arc constituting the minimum semantic unit.

  The function of the information retrieval apparatus may be realized by a plurality of computers instead of a single computer.

  Further, the processing flow is an example, and may be changed as long as the processing result does not change. The functional block division of FIG. 1 is an example, and may not match the actual program module configuration.

  The information retrieval apparatus described above is a computer apparatus, and as shown in FIG. 12, a memory 1501, a CPU (Central Processing Unit) 1503, a hard disk drive (HDD: Hard Disk Drive) 1505, and a display apparatus 1509. A display control unit 1507 connected to, a drive device 1513 for a removable disk 1511, an input device 1515, and a communication control unit 1517 for connecting to a network are connected by a bus 1519. An operating system (OS) and an application program for executing the processing in this embodiment are stored in the HDD 1505, and are read from the HDD 1505 to the memory 1501 when executed by the CPU 1503. The CPU 1503 controls the display control unit 1507, the communication control unit 1517, and the drive device 1513 according to the processing content of the application program, and performs a predetermined operation. Further, data in the middle of processing is mainly stored in the memory 1501, but may be stored in the HDD 1505. In the embodiment of the present technology, an application program for performing the above-described processing is stored in a computer-readable removable disk 1511 and distributed, and is installed from the drive device 1513 into the HDD 1505. The HDD 1505 may be installed via a network such as the Internet and the communication control unit 1517. Such a computer apparatus realizes various functions as described above by organically cooperating hardware such as the CPU 1503 and the memory 1501 described above with programs such as the OS and application programs. .

  The embodiments of the present technology described above are summarized as follows.

  In the information search method according to the present embodiment, (A) a semantic minimum unit consisting of two words and each element of the semantic relationship between the two words or one word and each element of the role of the one word, The first specifying process specified as a search key, and (B) for each search target sentence, matches the minimum semantic unit of the search key from the data storage unit storing the minimum semantic unit included in the search target sentence. Alternatively, a second specifying process for specifying a search target sentence including a partially matching minimum meaning unit and (C) an output process for outputting a search result for the specified search target sentence. In this way, a search result according to the user's intention that cannot be extracted by keyword search is obtained. That is, semantic elements other than words are also subject to matching.

  The data storage unit may store the identification information of the search target sentence in association with the smallest semantic unit. In this case, the output process may acquire (c1) a search target sentence corresponding to the identification information of the specified search target sentence from a database storing the search target sentence. In this way, the search index can be reduced, the search processing load can be reduced, and the extracted sentence can be reliably output.

  The second specifying process is (b1) a process for determining whether all elements included in the minimum semantic unit of the search key are the same as all elements included in the minimum semantic unit included in the search target sentence. It may be included. In this way, the concept based on the smallest semantic unit is completely secured, so that the search accuracy is improved.

  The second specifying process (b2) determines whether some elements included in the minimum semantic unit of the search key are the same as some elements included in the minimum semantic unit included in the search target sentence. Processing may be included. In this way, information of similar categories can be extracted using the concept based on the smallest semantic unit.

  Furthermore, (D) a natural language sentence search query is accepted, (E) a natural language sentence is semantically analyzed, and (F) a semantic minimum unit serving as a search key is generated from the semantic analysis result of the natural language sentence. It may be. In this way, it is possible to efficiently extract a search result reflecting the searcher's intention embedded in the natural language sentence.

  In addition, when a plurality of semantic minimum units serving as search keys are specified in the first specifying process, the second specifying process includes a search including a minimum semantic unit that matches at least one of the plurality of semantic minimum units. You may make it include the process which extracts an object sentence. In this way, it is possible to flexibly search for partially strongly related sentences other than sentences that match all the smallest semantic units.

  Further, (G) the evaluation value of the minimum semantic unit that matches the minimum semantic unit of the search key is acquired from the evaluation value storage unit in which the evaluation value of the minimum semantic unit included in the search target sentence is stored. The evaluation value of the search target sentence specified based on the evaluation value may be calculated. In this way, it is possible to give priority to a sentence including a highly meaningful minimum unit. For example, rare information is easily extracted.

  In addition, the calculation process (g1) out of the minimum semantic units included in one search target sentence, based on the number of minimum semantic units that match the minimum semantic unit specified as the search key, You may make it include the process which calculates an evaluation value. In this way, prioritization focusing on the smallest semantic unit repeated in the same sentence can be performed. For example, a sentence specialized in a specific theme is easily extracted.

  In addition, the calculation process (g2) is based on the number of types of the minimum semantic unit that matches one of the minimum semantic units included in one search target sentence and is specified as the search key. You may make it include the process which calculates an evaluation value. In this way, prioritization focusing on the degree of matching of the entire query can be performed. For example, it helps to find sentences with the same tone.

  In the calculation process, (g3) The sentence evaluation value may be calculated according to the following equation.

Sentence evaluation value = [Of the search key semantic minimum units (K1, K2,... Ki,...), The same as the semantic minimum unit included in the sentence or the same as the minimum semantic unit included in the sentence. The sum of the evaluation value of the minimum semantic unit Ki satisfying a predetermined condition × the number of occurrences of the minimum semantic unit Ki in the sentence] × (according to the number of types of the minimum semantic unit Ki that matches the minimum semantic unit included in the sentence) Value)

  In this way, it is possible to perform weighting that comprehensively takes into account the evaluation of the minimum semantic unit for the entire search target document and the appearance mode in the sentence of the minimum semantic unit.

  Furthermore, (I) based on the evaluation value of the specified search target sentence, the evaluation value of the search target document including the specified search target sentence is calculated, and (J) the evaluation value of the search target document is used as a sort key. The search target documents including the specified search target sentences may be ordered. In this case, the output process may include (c1) a process of outputting search results according to the order of the search target documents. In this way, when reading desired information from the entire document including not only the extracted sentence but also the nearby sentence, a highly useful document can be prioritized.

  (I1) The evaluation value of the search target document may be the sum of the evaluation values of the specified search target sentences. In this way, it is possible to extract a document having a high degree of fitness as a whole.

  A program for causing a computer to perform the processing according to the above method can be created. The program can be a computer-readable storage medium such as a flexible disk, a CD-ROM, a magneto-optical disk, a semiconductor memory, a hard disk, or the like. It is stored in a storage device. The intermediate processing result is temporarily stored in a storage device such as a main memory.

  The following supplementary notes are further disclosed with respect to the embodiments including the above examples.

(Appendix 1)
A first specifying process for specifying, as a search key, a semantic minimum unit comprising two words and each element of a semantic relationship between the two words or one word and each element of the role of the one word;
A second target for specifying a search target sentence including a minimum semantic unit that matches the minimum semantic unit of the search key in a data storage unit storing the minimum semantic unit included in the search target sentence for each search target sentence Specific processing,
An information search method executed by a computer, including output processing for outputting a search result for the specified search target sentence.

(Appendix 2)
The data storage unit stores identification information of the search target sentence in association with the minimum semantic unit,
The information search method according to claim 1, wherein the output process includes a process of acquiring a search target sentence corresponding to the identification information of the specified search target sentence from a database storing the search target sentence.

(Appendix 3)
The second specifying process includes a process of determining whether all elements included in the minimum semantic unit of the search key are the same as all elements included in the minimum semantic unit included in the search target sentence. Including the information retrieval method according to appendix 1 or 2;

(Appendix 4)
The second specifying process determines whether a part of elements included in the minimum semantic unit of the search key is the same as a part of elements included in the minimum semantic unit included in the search target sentence. The information search method according to appendix 1 or 2, including processing.

(Appendix 5)
Accept natural language search queries,
Semantic analysis of the natural language sentence,
The information search method according to any one of appendices 1 to 4, wherein the computer further executes a process of generating the minimum semantic unit serving as the search key from a semantic analysis result of the natural language sentence.

(Appendix 6)
When a plurality of minimum semantic units serving as the search keys are specified in the first specifying process,
The information search according to any one of appendices 1 to 5, wherein the second specifying process includes a process of extracting the search target sentence including the minimum semantic unit that matches at least one of the plurality of minimum semantic units. Method.

(Appendix 7)
From the evaluation value storage unit storing the evaluation value of the minimum semantic unit included in the search target sentence, the evaluation value of the minimum semantic unit that matches the minimum semantic unit of the search key is acquired, and the evaluation value The information search method according to any one of supplementary notes 1 to 6, wherein the computer further executes a process of calculating an evaluation value of the search target sentence specified based on the search target sentence.

(Appendix 8)
The calculation process is based on the number of the minimum semantic units that match the minimum semantic unit specified as the search key among the minimum semantic units included in one search target sentence. The information search method according to appendix 7, which is a process of calculating an evaluation value of a sentence.

(Appendix 9)
The calculation process is based on the number of types of the minimum semantic unit that matches one of the minimum semantic units included in one search target sentence and is specified as the search key. The information search method according to appendix 7, which is a process of calculating an evaluation value.

(Appendix 10)
The calculation process is a process of calculating an evaluation value of a sentence according to the following formula: evaluation value of the sentence = [the minimum semantic unit (K1, K2,... Ki,...) Of the search key. (Evaluation value of the minimum semantic unit Ki that matches the minimum semantic unit included in the sentence or that matches the minimum semantic unit included in the sentence and satisfies a predetermined condition × the minimum semantic unit in the sentence) Sum of the number of occurrences of Ki) × (value corresponding to the number of types of minimum semantic unit Ki matching the minimum semantic unit included in the sentence)
The information search method according to appendix 7.

(Appendix 11)
Based on the evaluation value of the specified search target sentence, the evaluation value of the search target document including the specified search target sentence is calculated,
The computer further executes a process of ordering search target documents including the specified search target sentence using the evaluation value of the search target document as a sort key,
The information search method according to any one of appendices 1 to 10, wherein the output process includes a process of outputting a search result according to an order of the search target document.

(Appendix 12)
The information search method according to claim 11, wherein the evaluation value of the search target document is a sum of evaluation values of the specified search target sentences.

(Appendix 13)
A first specifying process for specifying, as a search key, a semantic minimum unit comprising two words and each element of a semantic relationship between the two words or one word and each element of the role of the one word;
A second target for specifying a search target sentence including a minimum semantic unit that matches the minimum semantic unit of the search key in a data storage unit storing the minimum semantic unit included in the search target sentence for each search target sentence Specific processing,
A program for causing a computer to execute an output process including outputting a search result for the specified search target sentence.

(Appendix 14)
A first specifying unit that specifies, as a search key, a semantic minimum unit composed of two words and each element of a semantic relationship between the two words or one word and each element of the role of the one word;
A second target for specifying a search target sentence including a minimum semantic unit that matches the minimum semantic unit of the search key in a data storage unit storing the minimum semantic unit included in the search target sentence for each search target sentence A specific part,
An output unit for outputting a search result for the specified search target sentence;
An information retrieval apparatus having

11 Search Target Document Database 13 Search Index 15 Evaluation Value Table 31 Input Unit 33 Semantic Analysis Unit 35 Meaning Minimum Unit Generation Unit 37 Search Unit 39 Evaluation Value Calculation Unit 41 Ranking Unit 43 Output Unit 1501 Memory 1503 CPU (Central Processing Unit)
1505 Hard Disk Drive (HDD)
1507 Display control unit 1509 Display device 1511 Removable disk 1513 Drive device 1515 Input device 1517 Communication control unit 1519 Bus

Claims (9)

A first specifying process for specifying, as a search key, a semantic minimum unit comprising two words and each element of a semantic relationship between the two words or one word and each element of the role of the one word;
A second target for specifying a search target sentence including a minimum semantic unit that matches the minimum semantic unit of the search key in a data storage unit storing the minimum semantic unit included in the search target sentence for each search target sentence Specific processing,
An evaluation value of the minimum semantic unit that matches the minimum semantic unit of the search key is acquired from an evaluation value storage unit in which an evaluation value for each minimum semantic unit included in the search target sentence is stored, and the evaluation value A calculation process for calculating an evaluation value of the specified search target sentence based on
An output process for outputting a search result for the specified search target sentence based on an evaluation value of the search target sentence, and an information search method executed by a computer ,
The calculation process is a process for calculating an evaluation value of a sentence according to the following formula.
Evaluation value of the sentence = [Of the minimum semantic units (K1, K2,... Ki,...) Of the search key (matches the minimum semantic unit included in the sentence or is included in the sentence. The sum of the evaluation value of the minimum semantic unit Ki that satisfies the predetermined minimum unit and satisfying a predetermined condition × the number of occurrences of the minimum semantic unit Ki in the sentence] × (the minimum semantic unit included in the sentence) The value according to the number of types of matching semantic minimum unit Ki)
Information retrieval method . A first specifying process for specifying, as a search key, a semantic minimum unit comprising two words and each element of a semantic relationship between the two words or one word and each element of the role of the one word;
A second target for specifying a search target sentence including a minimum semantic unit that matches the minimum semantic unit of the search key in a data storage unit storing the minimum semantic unit included in the search target sentence for each search target sentence Specific processing,
An evaluation value of the minimum semantic unit that matches the minimum semantic unit of the search key is acquired from an evaluation value storage unit in which an evaluation value for each minimum semantic unit included in the search target sentence is stored, and the evaluation value A calculation process for calculating an evaluation value of the specified search target sentence based on
A process for calculating an evaluation value of a search target document including the specified search target sentence based on the specified evaluation value of the search target sentence;
A process for ordering search target documents including the specified search target sentence using the evaluation value of the search target document as a sort key;
According to the order of based rather the target document on the evaluation value of the search subject sentence includes an output process of outputting the search result for the specified the search subject sentence, information search method executed by a computer. The information search method according to claim 2 , wherein the evaluation value of the search target document is a sum of evaluation values of the specified search target sentences. A first specifying process for specifying, as a search key, a semantic minimum unit comprising two words and each element of a semantic relationship between the two words or one word and each element of the role of the one word;
A second target for specifying a search target sentence including a minimum semantic unit that matches the minimum semantic unit of the search key in a data storage unit storing the minimum semantic unit included in the search target sentence for each search target sentence Specific processing,
An evaluation value of the minimum semantic unit that matches the minimum semantic unit of the search key is acquired from an evaluation value storage unit in which an evaluation value for each minimum semantic unit included in the search target sentence is stored, and the evaluation value A calculation process for calculating an evaluation value of the specified search target sentence based on
Output processing for outputting a search result for the specified search target sentence based on the evaluation value of the search target sentence
The A program to be executed by a computer,
The calculation process is a process for calculating an evaluation value of a sentence according to the following formula.
Evaluation value of the sentence = [Of the minimum semantic units (K1, K2,... Ki,...) Of the search key (matches the minimum semantic unit included in the sentence or is included in the sentence. The sum of the evaluation value of the minimum semantic unit Ki that satisfies the predetermined minimum unit and satisfying a predetermined condition × the number of occurrences of the minimum semantic unit Ki in the sentence] × (the minimum semantic unit included in the sentence) The value according to the number of types of matching semantic minimum unit Ki)
Program . A first specifying process for specifying, as a search key, a semantic minimum unit comprising two words and each element of a semantic relationship between the two words or one word and each element of the role of the one word;
A second target for specifying a search target sentence including a minimum semantic unit that matches the minimum semantic unit of the search key in a data storage unit storing the minimum semantic unit included in the search target sentence for each search target sentence Specific processing,
An evaluation value of the minimum semantic unit that matches the minimum semantic unit of the search key is acquired from an evaluation value storage unit in which an evaluation value for each minimum semantic unit included in the search target sentence is stored, and the evaluation value A calculation process for calculating an evaluation value of the specified search target sentence based on
A process for calculating an evaluation value of a search target document including the specified search target sentence based on the specified evaluation value of the search target sentence;
A process for ordering search target documents including the specified search target sentence using the evaluation value of the search target document as a sort key;
The search according to the order of based rather the target document on the evaluation value of the subject statement, the output process of outputting the search result for the specified the search subject sentence
Program for executing the the computer. The program according to claim 5 , wherein the evaluation value of the search target document is a sum of evaluation values of the specified search target sentences. A first specifying unit that specifies, as a search key, a semantic minimum unit composed of two words and each element of a semantic relationship between the two words or one word and each element of the role of the one word;
A second target for specifying a search target sentence including a minimum semantic unit that matches the minimum semantic unit of the search key in a data storage unit storing the minimum semantic unit included in the search target sentence for each search target sentence A specific part,
An evaluation value of the minimum semantic unit that matches the minimum semantic unit of the search key is acquired from an evaluation value storage unit in which an evaluation value for each minimum semantic unit included in the search target sentence is stored, and the evaluation value A calculation unit that calculates an evaluation value of the specified sentence to be searched based on
Based on the evaluation value of the search target sentence, an output unit that outputs a search result for the specified search target sentence;
I have a,
The calculation unit calculates an evaluation value of a sentence according to the following formula:
Evaluation value of the sentence = [Of the minimum semantic units (K1, K2,... Ki,...) Of the search key (matches the minimum semantic unit included in the sentence or is included in the sentence. The sum of the evaluation value of the minimum semantic unit Ki that satisfies the predetermined minimum unit and satisfying a predetermined condition × the number of occurrences of the minimum semantic unit Ki in the sentence] × (the minimum semantic unit included in the sentence) The value according to the number of types of matching semantic minimum unit Ki)
Information retrieval device. A first specifying unit that specifies, as a search key, a semantic minimum unit composed of two words and each element of a semantic relationship between the two words or one word and each element of the role of the one word;
A second target for specifying a search target sentence including a minimum semantic unit that matches the minimum semantic unit of the search key in a data storage unit storing the minimum semantic unit included in the search target sentence for each search target sentence A specific part,
An evaluation value of the minimum semantic unit that matches the minimum semantic unit of the search key is acquired from an evaluation value storage unit in which an evaluation value for each minimum semantic unit included in the search target sentence is stored, and the evaluation value A calculation unit that calculates an evaluation value of the specified sentence to be searched based on
A calculation unit that calculates an evaluation value of a search target document including the specified search target sentence based on the specified evaluation value of the search target sentence;
An ordering unit that orders the search target documents including the specified search target sentence using the evaluation value of the search target document as a sort key;
According to the order of based rather the target document on the evaluation value of the search subject sentence, and an output unit for outputting a search result for the specified the search subject sentence,
An information retrieval apparatus having The information search apparatus according to claim 8 , wherein the evaluation value of the search target document is a sum of evaluation values of the specified search target sentences.

Images