JP2009075747A - Similar sentence retrieval system and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To enable efficient retrieval of an example including a specific expression from a database, when the specific expression noted by a user is included in a retrieval target sentence. <P>SOLUTION: A keyword designation unit 24 receives designation of a keyword in the retrieval target sentence received with a retrieval target input unit 23. A language analysis unit 25 analyzes role information in syntax/semantics of each word comprising the retrieval target sentence. A retrieval expression generation unit 27 generates a retrieval expression including a keyword and role information in syntax/semantics of the keyword among the pieces of role information in syntax/semantics of respective words analyzed with the language analysis unit 25. A database retrieval unit 28 retrieves examples including the keyword from an example database 21 by using the retrieval expression, and retrieves an example having matching role information in syntax/semantics from them. An output unit 29 outputs the retrieved example to a display device 17. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a similar sentence search system and program for searching from an example database using similar sentences of input search target sentences as examples.

  The similar sentence search system searches a similar sentence similar to the search target sentence from the example database with respect to the input search target sentence, and the example search of such a similar sentence is referred to during information search or translation. Used in example searches.

  In the similar sentence search, the similarity between the search target sentence and each sentence registered in the example database is calculated, and a sentence having a similarity higher than a threshold set by the user is detected as a search result. There are various methods for calculating the similarity, but the basic method is based on the character strings of words constituting the sentence and the matching in units of words. In addition to this, synonyms are registered in the dictionary in advance, it is considered that the words that are related to the synonyms match, and similar sentences that are close as topics even if they do not contain the same word as a character string. Some of them can be detected (see, for example, Patent Document 1). Further, in calculating similarity, there is one that can detect similar sentences with closer semantic contents in consideration of matching of sentence syntax and semantic relationship (for example, see Patent Document 2).

  In such a similar sentence search system, when the focus is on a specific phrase in the search target sentence rather than the similarity as a whole sentence, and the purpose is to detect a similar sentence including the phrase, the detection result is There are cases where it does not meet the user's wishes. This is because the conventional search system treats the weight of each component of the sentence equally when calculating the similarity. That is, even if a specific phrase is included, if the ratio of the phrase is low with respect to the entire sentence, the similarity as a whole sentence is low.

  In order to solve such a problem, the following three search methods have been proposed in which the weight is changed according to the portion constituting the sentence when calculating the similarity.

(1) In the first search method, for both the example and the search target sentence, each information is assigned to a place in the sentence that can be replaced, deleted, or added grammatically and semantically, and the replaceable word The degree of similarity is calculated by comparing each other, deleting unnecessary portions, and adding insufficient portions (see, for example, Patent Document 3). This makes it possible to calculate the degree of similarity with more emphasis on the outline of the sentence. In this first search method, it is necessary to give information to both the example and the search target sentence in advance, but replacement, deletion, and addition are semi-automatically by comparing the sentences. Done. For example, regarding the designation of a part that can be deleted, the following two examples A1 and A2 are compared, and it is automatically determined that the modifier “precious” can be deleted.
Example A1: Nakayama scored with a free kick.
Example A2: And 30 minutes later, Nakata scored a valuable score with PK.

(2) In the second search method, a “sentence fragment” is created by extracting the relationship between the subject and the verb and the verb and the object for both the example and the search target sentence, and each sentence fragment is similar. Make a decision. A sentence having a complicated structure is subjected to similarity determination for each sentence fragment, so that it is possible to detect an example including a part having a high degree of similarity to a search target sentence even though it is a part of the sentence (for example, patents) Reference 4). For example, when there is a search target sentence X1 and examples A3 and A4 as shown below, the similarity as a whole search sentence is not so high for either example A3 or A4, but the example A3 is a search target sentence. Since the similarity between the first half and the second half in example A4 is high, both examples A3 and A4 are detected as similar sentences.
Search target sentence X1: “The total number of applications is the highest of 41.50 million, and there are nearly 30,000 participating schools.”
Example A3: “The total number of applications reached a record high of about 41.50 million.”
Example A4: “The number of participating schools in Japan and overseas is close to 30,000.”

(3) In the third search method, the constituent elements of the sentence are grouped into chunks of meanings, the weight of the chunk having the case information constituting the skeleton of the syntax such as the main case and the purpose case is increased, and the chunk is The similarity is calculated by adding a weight to the word that is the center of (see, for example, Patent Document 5). For example, the search target sentence is "Australian famous general Ferriman has won a gold medal for the 400m women's final". In the example book, "September 25, 2000, 4pm Beijing time, 400m final for women" In the battle, if there is a sentence saying "Australian veteran Ferriman won a gold medal", take the wording such as "Women's 400 meter final", "Australian vice versa Ferriman", "Gold medal", "I won" The weight of the important components above increases. This makes it possible to calculate the degree of similarity with more emphasis on the outline of the sentence.

With these three search methods, sentences that are similar to the search target sentence or similar to the whole sentence but not so high, but part of the sentence that has high similarity to the search target sentence Can be detected.
Japanese Patent Application Laid-Open No. 11-110395 JP 2000-242650 A JP 2001-357065 A JP-A-2005-190185 JP 2006-65387 A

However, the sentence components emphasized in the three search methods described above do not necessarily match the part that the user is paying attention to. Even with these search methods, the user pays attention. The detection result of a similar sentence including a specific wording cannot satisfy the user's request. For example, a case where the examples A5, A6, and A7 are stored in the example database for the following search target sentence X2 will be described as an example.
Search target sentence X2: 40 patients and 20 healthy adults were randomly assigned to a group using or not using XX medicine.
Example A5: 100 patients with △ Δ disease who can be operated were randomly assigned to a group in which surgery was performed after 3 cycles of administration of XX drug before surgery and a group in which surgery was performed alone.
Example A6: For example, half of 60 factories are randomly assigned to treatment groups and half to control groups.
Example A7: Children were randomly assigned to a chemotherapy group and a standard follow-up care group.

  The first search method has a problem that even if the user pays attention to the modifier, it is determined that the modifier can be deleted and is not taken into account when calculating the similarity. For example, the search target sentence X2 is a sentence related to medical statistics, and “the assignment method is random” in this sentence is an important point. Nevertheless, in the first search method, there is a difference in similarity between a sentence including an important keyword and a sentence not including it because the word “random” is a modifier (adjective verb). Does not occur. As a result, a sentence including “random” is not detected preferentially, and even if it is detected, it takes time to distinguish a desired sentence from a sentence that is not.

  Similarly, the second search method described above has a problem that “random” is not considered in the similarity calculation. In addition to this, the indirect object of “assign” in the search target sentence X2 is “the group that uses XX medicine” and “the group that does not use”. Compared with “the group that performed” and “the surgery alone group”, since the ratio of words other than “group” is high, the similarity between the two becomes low. This is due to the fact that in the second search method, the unit for comparing the degree of similarity is made fine, but the weights are still not given to the portions that the user pays attention to. In other words, what the user wants to detect is a sentence in which two words including “group”, “group” and “group”, which are part of a compound word, are in a parallel relationship. Since it is arbitrary, the user's search purpose cannot be satisfied without giving weight to this relationship itself. Therefore, there is a possibility that the examples A6 and A7 are not searched.

  On the other hand, in the third search method described above, weights are added to the central word of the meaning chunk, so in the above example, weights are added to the “group”. However, since the detection criterion is the similarity of the whole sentence, the similarity between the example 2 and the search target sentence is low. Further, since the modifier “random” does not constitute a case element, there still remains a problem that the weight of similarity is not taken into account.

  An object of the present invention is to provide a similar sentence search system and program capable of efficiently searching an example including a specific phrase from a database when there is a specific phrase noted by the user in the search target sentence. With the goal.

  A similar sentence search system according to the present invention includes a storage device storing a similar sentence search program and an example database, an input device for inputting a search target sentence for an example search of a similar sentence and inputting information necessary for an operation, In a similar sentence search system comprising a display device that displays a search target sentence and an example searched from the example database, and a CPU that executes the similar sentence search program, the search target sentence input from the input device is A search target input unit that accepts, a keyword specification unit that receives specification of a keyword in a search target sentence received by the search target input unit, and a syntax of each word constituting the search target sentence received by the search target input unit The language analysis unit that analyzes semantic role information, and the syntax and semantic role information of each word analyzed by the language analysis unit A search expression generation unit that generates a search expression including syntactic and semantic role information of the keyword and the keyword, and an example including the keyword using the search expression generated by the search expression generation unit A database search unit that searches for an example in which the syntactic and semantic role information matches, and an output unit that outputs the example searched in the database search unit to the display device. Features.

  ADVANTAGE OF THE INVENTION According to this invention, when there exists a specific phrase which a user pays attention in a search object sentence, the example containing the specific phrase can be searched efficiently from a database.

  FIG. 1 is a configuration diagram of a similar sentence search system according to an embodiment of the present invention. The similar sentence search system 11 searches and outputs a similar sentence of an input search target sentence from an example database as an example. For example, a software program such as a similar sentence search program is installed in a general computer, A software program is executed by the CPU 12. In the similar sentence search system 11, a CPU 12, a ROM (Read Only Memory) 13, and a RAM (Random Access Memory) 14 are connected via a bus 15. An input device 16, a display device 17, and a storage device 19 are connected to the bus 15.

  The storage device 19 stores a similar sentence search program 20 and an example database 21 and a synonym dictionary 22. The similar sentence search program 20 includes a search target sentence input unit 23, a keyword specification unit 24, a language analysis unit 25, a search control unit 26, a search expression generation unit 27, a database search unit 28, and an output unit 29.

  The CPU 12 reads out and executes a boot program for starting the machine translation device 11 from the ROM 13 based on an input signal from the input device 16, and further reads an operating system (not shown) stored in the storage device 19. The CPU 12 controls each device based on the input signal of the input device 16, reads out the similar sentence search program 20, the example database 21, and the synonym dictionary 22 stored in the storage device 19 and loads them into the RAM 14. At the same time, a similar sentence search process to be described later is realized based on the command of the program read from the RAM 14.

  The input device 16 is an input means for inputting data and commands such as character data and file data of a search target sentence in a similar sentence example search. Usually, the input device 16 is a pointing device such as a keyboard or a mouse / touch panel, voice recognition or character recognition. This function is realized by an external storage medium reading device such as a CD drive or a network input device. The display device 17 is an output means such as data input from the input device 16 and a similar sentence example search result, and a CRT display device or a liquid crystal display device is used.

  FIG. 2 is a functional block diagram of the CPU 12 of the similar sentence search system 11 according to the embodiment of the present invention. Each functional block in the CPU 12 shown in FIG. 2 includes each unit constituting the similar sentence search program 20, that is, a search target sentence input unit 23, a keyword designation unit 24, a language analysis unit 25, a search control unit 26, and a search expression generation unit. 27 corresponds to the database search unit 28 and the output unit 29.

  The search target sentence input unit 23 receives an input of a sentence (search target sentence) to be searched for similar sentences. For example, the search target sentence is directly input by the user himself / herself from a keyboard or the like which is the input device 16, or is input from a disk drive via a storage medium. The search target sentence input unit 23 receives a search target sentence input from the input device 16.

  The keyword specifying unit 24 receives a word (keyword) specified by the user in the search target sentence. For example, the keyword is specified by a drag operation using a mouse or the like as the input device 16. As will be described in detail later, the keyword is used to generate a search expression. If only one keyword is specified, it will always be included in the detected similar sentence, and if multiple keywords are specified, Specifies the condition that the priority order is always included in the detected similar sentence.

  The language analysis unit 25 acquires the role information on the syntax and meaning of each word constituting the sentence by performing language analysis of the sentence. The role information on the syntax and meaning is, for example, information such as part of speech, case, and parallel expression of each word constituting the sentence. Specifically, by performing morphological analysis, syntactic analysis, dependency analysis, and semantic analysis, which are analysis methods generally used in natural language processing, role information on the syntax and semantics of each word constituting a sentence To get. The analysis by the language analysis unit 25 is performed on both the input search target sentence and the examples stored in the example database 21.

  In the example database 21, a plurality of examples including words constituting a sentence and syntax / semantic role information of each word in the context are registered. The example is a similar sentence similar to the search target sentence. Registration of the example in the example database 21 is performed, for example, via a storage medium inserted in a disk drive that is the input device 16. Here, in each example, it is assumed that syntax / semantic role information obtained by using the language analysis unit 25 is associated with each word constituting a sentence in advance.

  The search control unit 26 controls a search expression generation unit 27, a database search unit 28, and an output unit 29, which will be described later.

  The search expression generation unit 27 is controlled by the search control unit 26 and uses the keyword specified by the keyword specifying unit 24, each word constituting the search target sentence, and its syntactic / semantic role information to generate a search expression. Is to be generated. That is, among the syntax / semantic role information of each word analyzed by the language analysis unit 25, a search expression including the keyword syntax / semantic role information and the keyword is generated. As a result, a search expression including not only the keyword specified by the user but also the role information in the syntax and meaning is created. Details of the search expression will be described later.

  The database search unit 28 is controlled by the search control unit 26 and searches for an example registered in the example database 21. Specifically, the example and the role information on the syntax and semantics associated with the keyword are compared between the example registered in the example database 21 and the search formula, and the example that matches or is similar is searched.

  In the synonym dictionary 22, a word having a synonym relation is registered in association with a specific word. The synonym dictionary 22 is referred to when comparing the search expression with the example registered in the example database 21, and the words having the synonym relation are regarded as matching. The user can specify whether to perform a search with reference to the synonym dictionary 22.

  The output unit 29 is controlled by the search control unit 26 and outputs the search result of the example registered in the example database 21. For example, the search result is output by displaying on the display device 13. The output is not limited to the output to the display device 13. Instead of the output from the display device 13, the search result may be output by voice or a printing device in addition to the output by the display device 13. Good.

  Next, search processing contents in the similar sentence search system 11 according to the embodiment of the present invention will be described. FIG. 3 is a flowchart showing the search processing contents in the similar sentence search system 11 according to the embodiment of the present invention.

  When a search process is performed using the similar sentence search system 11 according to the embodiment of the present invention, first, a search target sentence is input from the user via the input device 16. The input of the search character string of the search target sentence is accepted (S11). Since the keyword in the search target sentence is specified by the user after the search target sentence is input, the keyword specifying unit 24 accepts the specification of the keyword when the keyword is specified (S12).

  When the keyword is accepted, the language analysis unit 25 performs language analysis of the search target sentence (S13). Thereby, the syntax / semantic role information of each word constituting the search target sentence is acquired, and the syntax / semantic role information is given to each word. Then, the search expression generation unit 27 generates a search expression from the specified keyword and the syntactic / semantic role information associated with each keyword (S14).

  If the search expression is generated, the database search unit 28 uses the keyword and the syntactic / semantic role information associated with the keyword between the example registered in the example database 21 and the search expression. Are matched or similar (S15). Then, it is determined whether or not there is an example whose similarity is equal to or higher than the threshold set by the user, that is, whether or not a search result is obtained (S16), and the similarity is equal to or higher than the threshold set by the user. If there is an example that becomes, the example is extracted from the example database 21 and used as a search result (S17). If there are a plurality of search results, the search results are sorted in descending order of similarity. On the other hand, if there is no example data whose similarity is equal to or greater than the threshold set by the user, the search result is that there is no corresponding example (S18). When the search by the database search unit 28 is completed, the output unit 29 outputs the search result (S19). In this way, a series of similar sentence search processing ends.

  Now, “40 patients and 20 healthy adults are randomly assigned to a group that uses XX drugs and a group that does not use XX drugs” (search target sentence X2) as a search target sentence. Further, it is assumed that “assignment”, “group” (two places), and “random” are designated as keywords, and that all of these keywords are included (required) in the search result.

  The input of the search target sentence X2 and the specification of the keyword are accepted by the search target sentence input unit 23 and the keyword specifying unit 24 (S11, S12). Then, the language analysis unit 25 performs language analysis of the search target sentence X2 (S13). FIG. 4 is an explanatory diagram showing a language analysis result in the language analysis unit 25.

  As shown in FIG. 4, a tree structure of each word constituting the search target sentence X2 is obtained, and the independent words (nodes) constituting the search target sentence X2 and the syntactic and semantic relationships (links) between the independent words ) Becomes clear. In this case, the keyword “Random” is modified by the keyword “Random” as an adverb, and two noun phrases including the keyword “Group” are connected as “Nice case (indirect object)”. It becomes clear that In other words, the role information on the syntax and semantics of the keyword becomes clear. In FIG. 4, for the sake of brevity, links relating to words other than keywords are described in a form that is appropriately omitted.

  When the language analysis of the search target sentence X2 is completed, the search expression generation unit 27 uses the keywords “assignment”, “group”, “random” specified in step S12, and the syntax and meaning acquired in step S13. A search formula is generated from the above role information (S14).

  FIG. 5 is an explanatory diagram of a search expression including role information on the syntax and meaning of keywords and keywords. FIG. 5A shows a search expression expressed in a tree structure, and FIG. 5B shows a search expression expressed in a character string. Note that the role information is attached to the keyword word in FIG. 5A, and the letter “(required)” in FIG. 5B is that the specified keyword is “required”. Indicates that there is.

  When such a search expression is generated, the database search unit 28 searches the example data registered in the example database 21 using the search expression (S15). Specifically, first, a sentence including the keyword “essential” is searched from the examples registered in the example database 21. Next, the primary search results are narrowed down by matching the nodes and links of the tree structure with the search results (primary search results). Note that, as a matching technique based on such a tree structure, for example, a technique disclosed in Japanese Patent Laid-Open No. 2005-208825 by the present applicant is used.

  FIG. 6 is an explanatory diagram of an example of an example searched by the database search unit 28. In FIG. 6, in the case of the above-mentioned Example A5, “100 patients with △ Δ disease who can be operated on were randomly assigned to the group in which surgery was performed after 3 cycles of administration of the XX drug and the surgery alone group”. Is shown.

  That is, as shown in FIG. 6, the example A5 has the same tree structure as the search expression of the search target sentence X2 shown in FIG. Therefore, the example A5 having the same tree structure as the search expression is detected. When the search by the database search unit 28 is completed, the output unit 29 outputs the search result (S17). In this example, since all the keywords are “required”, only those having a completely matched tree structure are detected. However, when a keyword that is not “required” is specified, or when the synonym dictionary 22 is used. The example data whose similarity is equal to or greater than the specified threshold value is detected. Specific examples of these will be described later.

  Here, in the database search unit 28, the search target sentence and the search expression are handled as character string data. For example, the above-described search target sentence X2 “40 patients and 20 healthy adults were randomly assigned to a group that uses XX medicine and a group that does not use XX medicine” and the search formula in FIG. Is called.

Search target sentence: Allocation (fukushi_random kan-moku_group (rentai-shu_used (choku-moku_drug) heiretu_group (hitei_syouryaku used)) choku-moku_patient (shu_40 heiretu_adult (shu_healthy) (shu_20 people))) fukushi_ each)
Search formula: Allocation @@ (fukushi _random @@ kan-moku_group @@ (heiretu_group @@))
In this example, parentheses indicate a connection between words, and a character string written in Roman characters indicates syntax / semantic role information. The symbol “@@” indicates that the keyword is “essential”. These examples and the search formulas are judged to match because the part of “assignment (fukushi_random kan-moku_group (heiretu_group))” matches.

  The contents of the search described above will be described more specifically. The following five examples A5 to A9 are registered in the example database 21, and a case where the search target sentence X2 is searched by the following three search methods 1, 2, and 3 will be described.

(Search target text)
Search target sentence X2: 40 patients and 20 healthy adults were randomly assigned to a group using or not using XX medicine.
(Example database)
Example A5: 100 patients with △ Δ disease who can be operated were randomly assigned to a group in which surgery was performed after 3 cycles of administration of XX drug before surgery and a group in which surgery was performed alone.
Example A6: For example, half of 60 factories are randomly assigned to treatment groups and half to control groups.
Example A7: Children are randomly assigned to a chemotherapy group and a standard follow-up care group.
Example A8: This bias can occur at any point in time, starting with target selection, assigning to target and intervention groups, assessing exposure status and results, and analyzing data.
Example A9: It can be said that it is important to randomly assign patients to two groups, a treatment group and a control group, using an envelope method or the like, and to perform a comparative test with an appropriate sample size.

(Search method 1)
This is a case where all keywords are designated as “required”. That is, “assignment”, “group”, and “random” are designated as keywords, and there is no difference in the priority order of the keywords, and all are included in the search result. Then, the search formula is as follows.
"Assignment (required) + (adverb_random (required)) + (indirect purpose_ (group (required) _parallel_group (required)))"
As a result of the search, examples A5, A6, and A7 including essential keywords are first detected from the five examples A5 to A9 in the example database 21. Next, from this primary search result, the search results are narrowed down to example A5 and example A6 having the same tree structure as the search expression. In this way, by performing a search that considers not only keyword matching but also syntactic and semantic role information in the keyword sentence, the usage of the keyword in the search target sentence is narrowed down to examples that are closer. An efficient search can be performed.

(Search method 2)
This is a case where the priority order of keywords and the condition of whether or not to be included in the detected similar sentence are specified. Assume that “patient” is specified as a keyword in addition to the specified keyword, and a condition is specified that the priority is lower than the other four keywords and it is not necessarily included in the detection result. Under this condition, the example A6 does not include “patient”, but “patient” is not a keyword of “essential”, so the example A6 is a detection target, and has a lower similarity than the example A5 including “patient”. . In this way, by specifying a condition when specifying a keyword, it is possible to perform a search that focuses on a portion having a high degree of importance in a wording of interest while a high degree of search coverage.

(Search method 3)
This is a case where the use of the synonym dictionary 22 is designated. If it is specified that the synonym dictionary 22 is used, words having synonym relations are regarded as matching. In the synonym dictionary 22, if “random” and “random” are associated with words having a synonym relationship, “random” and “random” in Example A9 are regarded as synonyms, It is also detected during the primary search. As described above, when the synonym dictionary 22 is used, similar phrases can be detected even if the keyword and the character string do not completely match.

  As described above, in the similar sentence search system according to the embodiment of the present invention, since the search is performed in consideration of the keyword and the role information on the syntax and semantics of the keyword, the user pays attention in the search target sentence. When there is a specific wording, an example including the specific wording can be efficiently retrieved from the example database 21. In addition, by specifying a condition when specifying a keyword, it is possible to perform a search that focuses on a portion having a high degree of importance in the wording of interest while increasing the search coverage. Furthermore, by using the synonym dictionary 22, even if the keyword and the character string do not completely match, it is possible to detect similar phrases.

The block diagram of the similar sentence search system concerning embodiment of this invention. The functional block diagram of CPU of the similar sentence search system concerning embodiment of this invention. The flowchart which shows the search processing content in the similar sentence search system concerning embodiment of this invention. Explanatory drawing which shows an example of the language analysis result in the language analysis part in embodiment of this invention. Explanatory drawing of an example of the search formula produced | generated by the search formula production | generation part in embodiment of this invention Explanatory drawing of an example of the example searched by the database search part in embodiment of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 11 ... Similar sentence search system, 12 ... CPU, 13 ... ROM, 14 ... RAM, 15 ... Bus, 16 ... Input device, 17 ... Display device, 19 ... Memory | storage device, 20 ... Similar sentence search program, 21 ... Example database, DESCRIPTION OF SYMBOLS 22 ... Synonym dictionary, 23 ... Search target sentence input part, 24 ... Keyword specification part, 25 ... Language analysis part, 26 ... Search control part, 27 ... Search formula production | generation part, 28 ... Database search part, 29 ... Output part

Claims (4)

It is searched from the storage device storing the similar sentence search program and the example database, the input device for inputting the search target sentence for the example search of the similar sentence and the information necessary for the operation, and the search target sentence and the example database. In a similar sentence search system comprising a display device that displays an example and a CPU that executes the similar sentence search program, a search target input unit that receives a search target sentence input from the input device, and the search target A language specification unit that receives specification of a keyword in a search target sentence received by the input unit, and a language analysis that analyzes syntax / semantic role information of each word constituting the search target sentence received by the search target input unit And syntactic / semantic role information of the keyword among the syntactic / semantic role information of each word analyzed by the language analysis unit And a search expression generation unit that generates a search expression including the keyword, and an example including the keyword is searched from the example database using the search expression generated by the search expression generation unit, and the syntax / meaning A similar sentence search system, comprising: a database search unit that searches for an example that matches the above role information; and an output unit that outputs an example searched by the database search unit to the display device. The keyword specifying unit, when receiving a plurality of keyword designations from the input device, also accepts a priority order of the accepted keywords and a condition whether or not to be included in a search result. The similar sentence search system described. A synonym dictionary in which a synonym-related word is registered in association with a specific word is provided, and when there is a request from the input device to use the synonym dictionary for the keyword, the database search unit The similar sentence search system according to claim 1 or 2, wherein an example including a word having a synonym relation is also searched. Searched from the storage device storing the similar sentence search program and the example database, the input device for inputting the search target sentence of the example of the similar sentence and the information necessary for the operation, and the search target sentence and the example database In a similar sentence search system comprising a display device that displays an example and a CPU that executes the similar sentence search program, a function for receiving a search target sentence input from the input device and a search that is received by a computer A function that accepts keyword specification in the target sentence, a function that analyzes syntax / semantic role information of each word constituting the accepted search target sentence, and a syntax / semantic role information of each analyzed word A function for generating a search expression including syntax and semantic role information of the specified keyword and the keyword, and a generated search A function for searching an example including the keyword using an expression from the example database, searching for an example in which the syntactic and semantic role information matches, and a function for outputting the searched example to the display device; A program to realize

Images