JP6433937B2 - Keyword evaluation device, similarity evaluation device, search device, evaluation method, search method, and program - Google Patents

Description

  The present invention relates to a keyword evaluation device, a similarity evaluation device, a search device, an evaluation method, a search method, and a program that use speech or text as input.

  As an information search system, there is known a system that performs a search process such as keyword matching on a query input by a user and searches for a document that matches the query. However, in the case of search processing using keyword matching, the keyword of the query must match the keyword in the document completely, and there is a problem that the recall rate (Recall) of the search is reduced. Sometimes. Thus, there is a case where query expansion is performed to expand and increase keywords included in a query and match a query input by a user with a wider range of documents (Patent Document 1).

  Moreover, concept search is known as a search method other than keyword matching. In concept search, a keyword is expressed by an N-dimensional vector of continuous values, and the center of gravity of the N-dimensional vector is regarded as a query vector. Similarly, the document vector is also expressed by the centroid of the keyword vector in the document, the similarity between the query vector and the document vector is calculated, and the search result is output in descending order of the similarity, thereby searching for a document that matches the query. Run. Unlike keyword matching, conceptual search has the advantage that documents related to topics close to the query can be searched even if the keywords do not match completely (Patent Document 2).

JP 2010-123036 A JP 2010-182041 A

  However, keyword matching and concept searching assume that a query is considered as a set of keywords (bag-of-words). That is, when a query is input in a natural language format, a keyword or the like including a topic close to the query is searched by extracting the keyword from the natural language by morphological analysis or keyword extraction.

  Therefore, in the process of generating a keyword set by extracting keywords from the natural language, a situation may occur in which the word order or syntax information originally possessed by the natural language is lost.

  For example, it is assumed that two queries expressed in a natural language, such as “mail cannot be sent” and “some mail cannot be sent” are input in the FAQ search. Although the meanings represented by these two queries are different, when two queries are converted to keyword sets, both result in a keyword set that includes the same elements of “email, send, cannot” and the difference between the two queries cannot be distinguished. End up. Thus, when performing information retrieval in a natural language, it is sometimes necessary to consider not only the keywords included in the sentence but also the context around the keywords.

  The present invention has been made in view of the above circumstances, a keyword evaluation device capable of accurately evaluating the importance of a keyword included in an input sentence expressed in a natural language, a keyword evaluation method, and The purpose is to provide a program. Also, a similarity evaluation device, a search device, and a similarity that can evaluate the similarity between an input sentence expressed in a natural language and a sentence to be compared, and accurately search for a sentence similar to the input sentence. It is an object to provide an evaluation method, a search method, and a program.

  In order to achieve the above object, the keyword evaluation apparatus according to the present invention is similar to the first keyword extracted from the input first sentence and the first keyword in the input second sentence. A calculation for calculating at least two similarities among the similarity between keywords, the similarity between clauses including the keyword, and the similarity between clauses including the keyword based on the second keyword as a keyword And an evaluation unit that evaluates the importance of the first keyword higher as the absolute value of the change value of the at least two similarities calculated by the calculation unit is smaller.

  The similarity evaluation apparatus according to the present invention is included in the first sentence in the combination having the highest similarity among the combinations of the word included in the input first sentence and the word included in the input second sentence. The first keyword as the first word, the second keyword as the word contained in the second sentence, the similarity between the keywords, the similarity between the clauses including the keyword, and the similarity between the clauses including the keyword And the first sentence and the second sentence are more similar as the absolute value of the change value of the at least two similarity degrees calculated by the calculation part is smaller. And an evaluation unit that evaluates that

  The search device according to the present invention includes, for each of a plurality of search target sentences prepared in advance, a keyword vector that represents each keyword included in the search target sentence, a phrase vector that represents a phrase including a keyword for each keyword, and A storage unit for storing a dependency vector representing a dependency relationship including a dependency destination of the clause including the keyword for each keyword, and a keyword included in the input query statement and the search for each of the plurality of search target statements The keyword included in the query sentence in the combination with the highest similarity among the keywords included in the target sentence is the first keyword, the keyword included in the search target sentence is the second keyword, and the keyword vector Similarity between keywords based on the same phrase, including phrases based on phrase vectors A calculation unit that calculates at least two similarities for each of the plurality of search target sentences among the similarities of the phrases and the similarities between the dependencies of the clauses including keywords based on the dependency vectors, and for each of the plurality of search target sentences The evaluation unit that evaluates that the query sentence and the search target sentence are similar to each other as the absolute value of the change value of the at least two similarities calculated by the calculation unit is smaller, and the evaluation unit And a search unit that searches for a search target sentence similar to the query sentence based on the evaluation result.

In the keyword evaluation method according to the present invention, a computer uses a first keyword extracted from a first sentence input via an input unit, and a sentence input via the input unit together with the first sentence. A phrase including a similarity between keywords based on a second keyword that is similar to the first keyword in the second sentence stored in the storage device in association with the first sentence, and a phrase including the keyword The step of calculating at least two similarities among the similarities between each other and the similarities between clauses including keywords, and the smaller the absolute value of the change values of the at least two similarities, the smaller the first step of displaying evaluation results were evaluated as severity of keywords increases the display device, for execution.

Evaluation of the similarity of the sentence according to the present invention, the computer, there in the words included in the first sentence inputted through the input unit, which is input through the input unit together with the first sentence sentence The word included in the first sentence in the combination having the highest similarity among the combinations included in the second sentence stored in the storage device in association with the first sentence is the first keyword, The word contained in the second sentence is set as the second keyword, and at least two similarities are calculated among the similarity between the keywords, the similarity between the clauses including the keyword, and the similarity between the clauses including the keyword. And the step of evaluating that the first sentence and the second sentence are more similar as the absolute value of the change value of the at least two similarities is smaller, and displaying the evaluation result on a display device. Execute .

In the sentence search method according to the present invention, for each of a plurality of search target sentences prepared in advance, a computer represents a keyword vector representing each keyword included in the search target sentence and a phrase including a keyword for each keyword. A dependency vector that represents a dependency relationship including a phrase vector and a dependency destination of the clause including a keyword for each keyword is generated, and the keyword vector, the phrase vector, and the dependency generated for each of the plurality of search target sentences A step of associating vectors and storing them in a storage device, and for each of the plurality of search target sentences, a combination of a keyword included in the query sentence input via the input unit and a keyword included in the search target sentence The keyword included in the query statement in the combination having the highest similarity is the first keyword. The keyword included in the search target sentence is the second keyword, the similarity between the keywords based on the keyword vector, the similarity between the phrases including the keyword based on the phrase vector, and the phrase relationship including the keyword based on the dependency vector At least two similarities among the similarities between the receivers are calculated for each of the plurality of search target sentences, and the calculated at least two similarities are associated with each of the plurality of search target sentences and stored in the storage device. And, for each of the plurality of search target sentences, evaluating that the query sentence and the search target sentence are more similar as the absolute value of the calculated change value of the at least two similarities is smaller ; based on the evaluation, it searches the search subject sentence similar to the query statement, displayed on the display device the search subject sentence similar to the query statement That step, to run.

  The program of the keyword evaluation device according to the present invention is a program for causing a computer to function as each part of the keyword evaluation device.

  The program of the similarity evaluation apparatus according to the present invention is a program for causing a computer to function as each part of the similarity evaluation apparatus.

  The program of the search device according to the present invention is a program for causing a computer to function as each part of the search device.

As described above, according to the keyword evaluation device, the keyword evaluation method, and the program of the present invention, it is possible to accurately evaluate the importance of keywords included in an input sentence expressed in a natural language. Is obtained.
Further, according to the similarity evaluation device, search device, similarity evaluation method, search method, and program of the present invention, the similarity between an input sentence expressed in a natural language and a sentence to be compared is evaluated. As a result, it is possible to retrieve a sentence similar to the input sentence with high accuracy.

It is the schematic which shows the structural example of the similarity evaluation apparatus which concerns on 1st Embodiment. It is a figure explaining an example of the execution result of dependency analysis. It is a flowchart which shows an example of the process of the similarity evaluation process routine in a similarity evaluation apparatus. It is a figure which shows an example of the calculation process of the score at the time of performing a similarity evaluation process routine. It is the schematic which shows the structural example of the keyword evaluation apparatus which concerns on 2nd Embodiment. It is a flowchart which shows an example of the process of the keyword evaluation process routine in a keyword evaluation apparatus. It is the schematic which shows the structural example of the search device which concerns on 3rd Embodiment. It is a flowchart which shows an example of a process of the search process routine in a search device.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In the following description, the same reference numerals are given to the components or processes having the same functions throughout the drawings, and the repeated description is omitted as appropriate.

<First Embodiment>
In the first embodiment, a similarity evaluation apparatus 100 that inputs two sentences of a first sentence and a second sentence described in a natural language and outputs the score by quantifying the similarity between the two sentences will be described.

<System configuration>
FIG. 1 is a diagram illustrating a system configuration example of the similarity evaluation apparatus 100. As shown in FIG. 1, the similarity evaluation apparatus 100 is composed of a computer including a CPU, a RAM, and a ROM that stores a program for executing a similarity evaluation processing routine described later, and is functionally Is configured as follows.

  The similarity evaluation device 100 includes an input unit 10, a calculation unit 20, a storage unit 30, and an output unit 40. The computing unit 20 includes a sentence analyzing unit 21, a vector generating unit 22, a calculating unit 23, and an evaluating unit 24, and the evaluating unit 24 includes a similarity change rate calculating unit 25 and a similarity evaluating unit 26.

  The first sentence and the second sentence input to the input unit 10 are output to the sentence analysis unit 21. The sentence analysis unit 21 performs language processing such as keyword extraction and phrase dependency analysis on each of the first sentence and the second sentence, and analyzes the language structure of the first sentence and the second sentence.

  The vector generation unit 22 receives the analysis result of the language structure of the first sentence and the second sentence performed by the sentence analysis unit 21 and inputs, for each sentence, a different expression unit such as a keyword, a phrase, and a dependency of the phrase. Generate a concept vector.

  Then, the calculation unit 23 uses the concept vector for each expression unit generated by the vector generation unit 22 as an input, and calculates the similarity between keywords, phrases, and phrases.

  The evaluation unit 24 receives the similarity between the keywords calculated by the calculation unit 23, the phrases, and the dependency of the phrases, and calculates the similarity change rate by the similarity change rate calculation unit 25. Based on the degree change rate, the similarity evaluation unit 26 digitizes the final similarity between the first sentence and the second sentence and evaluates the score as a score.

  The output unit 40 outputs the score evaluated by the evaluation unit 24 and notifies the similarity between the first sentence and the second sentence.

  In the following, the granularity of each expression unit, such as keywords, clauses, and dependency of clauses, may be described. Compared to the keyword, the phrase has a coarser granularity of the expression unit, and the range including the clause dependency than the phrase has a coarser granularity of the expression unit, so the keyword granularity is the finest and the clause dependency The grain size is the coarsest.

  Next, the processing content of the calculating part 20 is demonstrated in detail.

<Sentence Analysis Department>
The sentence analysis unit 21 performs language processing on the first sentence and the second sentence, and performs dependency analysis and keyword extraction.

  For each sentence of the first sentence and the second sentence, the sentence analysis unit 21 uses, for example, a dependency analyzer, the morpheme in the sentence, the part of speech of each morpheme, the phrase information (the number of phrases, the morpheme included in the phrase) ), Information on language structures such as dependency relations between phrases is acquired. The dependency analyzer used in the sentence analysis unit 21 is not limited, and any dependency analyzer may be used as long as it can acquire information on the language structure shown above as an example. .

  FIG. 2 is a diagram illustrating an example of the execution result of the dependency analysis in the sentence analysis unit 21. For example, when “mail cannot be sent” is received as the second sentence, “mail is” is clause 1, “send” “Cannot” is phrase 2, and phrase 1 is related to phrase 2. Each clause is classified by morpheme and the part of speech of each morpheme is shown.

  Furthermore, the sentence analysis unit 21 extracts keywords from each sentence according to the keyword extraction rule based on the result of dependency analysis. In addition, there is no restriction | limiting in particular in the keyword extraction rule used in the sentence analysis part 21, It can prescribe | regulate freely according to the evaluation objective of the similarity of sentences. In general, it is preferable to extract words that are independent words such as nouns and verbs from sentences as keywords.

  The sentence analysis unit 21 manages information related to the language structure acquired from the first sentence and the second sentence and the extracted keywords in association with each sentence.

<Vector generator>
The vector generation unit 22 creates a concept vector for evaluating the similarity between sentences based on the analysis results of the first sentence and the second sentence in the sentence analysis unit 21.

  First, the vector generation unit 22 generates a keyword vector that is a conceptual vector of a keyword included in each sentence. Further, the vector generation unit 22 generates a phrase vector and a dependency vector, which are concept vectors of the phrase and the dependency, by synthesizing the keyword vectors.

  In order to generate a keyword vector, it is necessary to generate a concept vector model in advance. The concept vector model can be generated using a concept vector generation method by preparing a document set for generating a concept vector.

  There are no particular restrictions on the concept vector generation document set used in the vector generation unit 22 and the concept vector generation method. As the document set for generating concept vectors, for example, a page set of Wikipedia (registered trademark) whose contents overlap with a set of sentences that may be input as the first sentence (input sentence set) may be used. For example, an arbitrary document set such as a Web page set included in a search result when a Web search is performed using a keyword extracted from an input sentence set can be used.

  As for the concept vector generation method, for example, an arbitrary concept vector generation model such as a latent semantic indexing (LSI) using singular value decomposition, a topic model, and a model using a neural network can be used. . At this time, in the concept vector generation model used in the vector generation unit 22, not only the keywords included in the sentence but also all morpheme vectors such as particles are generated.

  Then, the vector generation unit 22 synthesizes the keyword vector of the keyword included in the phrase and the morpheme vector other than the keyword included in the phrase, and determines the phrase vector of the phrase including the keyword as the first sentence and the second sentence. Generate for each sentence.

  The vector synthesis method is not limited, and a vector synthesis method using a neural network such as Recursive AutoEncoder (RAE) may be used in addition to the synthesis method based on the centroid vector.

  Specifically, when using the synthesis method based on the centroid vector, the vector generation unit 22 calculates the centroid of all morpheme vectors including the keyword included in the phrase, and the vector representing the centroid is used as the phrase including the keyword. The phrase vector of. Also, when using the RAE combining method, the vector generation unit 22 first combines the vectors of the two morphemes that appear first and second from the beginning of the clause including the keyword, and the combined morpheme vectors are obtained. This time, it combines with the vector of the morpheme that appears third. Thereafter, the vector generation unit 22 repeats the process of sequentially synthesizing the synthesized morpheme vector and the next morpheme vector according to the word order of the morpheme until all the morpheme vectors included in the phrase are synthesized, The finally synthesized morpheme vector is used as the phrase vector of the phrase including the keyword.

  Next, the vector generation unit 22 generates a dependency vector representing the dependency relationship of the phrase. Here, the dependency relationship between clauses represents a dependency relationship between two clauses. Therefore, the vector generation unit 22 generates a dependency vector by synthesizing the phrase vectors of the two clauses at the source and destination of the clause. The method for synthesizing the phrase vector is not limited as in the case of the phrase vector generation, and any vector synthesis method such as a vector synthesis method using a neural network such as RAE is used in addition to a synthesis method using a centroid vector. be able to.

  In the case of clause dependency in Japanese, since it is a unidirectional dependency relationship, there is one clause that is a dependency destination in clauses other than the clause corresponding to the predicate.

  For example, in the sentence “I bought a red hat in Yokohama”, the phrase related to the phrase “red” becomes “hat”, so the vector generation unit 22 includes a phrase vector corresponding to the phrase “red”, A dependency vector corresponding to “red hat” is generated by synthesizing the phrase vector corresponding to the clause “cap”.

  Further, since the clauses related to the clauses “in Yokohama” and the clause “hat” are both “buyed”, the vector generation unit 22 determines the dependency vectors corresponding to “bought in Yokohama” respectively. , A dependency vector corresponding to “I bought a hat” is generated.

  For the clause “Bought” corresponding to the predicate, there is no clause as a dependency destination. Therefore, the vector generation unit 22 uses the clause vector corresponding to the clause “Bought” as a dependency vector.

  As a result of the above processing, the vector generation unit 22 determines, for each sentence of the first sentence and the second sentence, a keyword vector group, a phrase vector group of a phrase including the keyword, and a dependency vector representing a dependency relation of the phrase including the keyword. Create a group and associate it with each sentence.

<Calculation section>
The calculation unit 23 calculates the similarity between keywords, the similarity between clauses, and the similarity between clauses based on the keyword vector, phrase vector, and dependency vector generated by the vector generation unit 22. . At this time, the calculation result of the similarity differs depending on which of the first sentence and the second sentence is used as a reference sentence for which the similarity is to be evaluated. The similarity is calculated using the standard sentence.

  First, the calculation unit 23 selects one keyword included in the first sentence serving as a reference. Note that the keywords included in the first sentence have already been extracted from the first sentence by the sentence analysis unit 21. Then, the calculation unit 23 calculates the similarity between the keyword selected from the first sentence (the keyword of interest) and all the keywords included in the second sentence using the keyword vector corresponding to each keyword, Of the keywords included in the two sentences, a keyword (corresponding keyword) having the highest similarity to the keyword of interest is selected.

  Note that the calculation method of the similarity between keywords used in the calculation unit 23 is any calculation as long as the value indicating the calculated similarity is normalized so as to take a value in the range of 0 to 1. For example, a cosine distance or the like can be used. When the value indicating the similarity is “0”, it indicates that the keywords are not similar to each other. The larger the value indicating the similarity, the higher the similarity between the keywords. When the value indicating the similarity is “1”, the similarity between the keywords is the maximum.

  Next, the calculation unit 23 associates each phrase with the similarity between the phrase including the keyword of interest selected from the first sentence and the phrase of the second sentence including the corresponding keyword having the highest similarity with the keyword of interest. For example, the same calculation method as that used when calculating the similarity between keywords is used.

  When there are a plurality of clauses including the corresponding keyword in the second sentence, the calculation unit 23 calculates the similarity between each of the clauses including the keyword of interest and all the clauses including the corresponding keyword. Select a combination of clauses with the largest degree value.

  Then, the calculation unit 23 uses the dependency vectors corresponding to each dependency relationship, for example, to determine the similarity between the dependency relationship of the clause including the keyword of interest and the dependency relationship of the clause including the corresponding keyword. The calculation method is the same as that used when calculating the similarity between keywords.

  That is, for the attention keyword selected from the first sentence, the calculation unit 23 calculates the similarity between the keywords with the corresponding keyword included in the second sentence, the phrase including the attention keyword and the phrase including the corresponding keyword. Three types of similarity are calculated: the similarity between clauses, and the dependency between clauses including the keyword of interest and the dependency between clauses including the corresponding keyword.

  Then, the calculation unit 23 sequentially selects each keyword included in the first sentence as a keyword of interest, and calculates the above three types of similarity for each keyword included in the first sentence. To do.

<Evaluation Department>
The evaluation unit 24 evaluates the similarity between the first sentence and the second sentence based on the three types of similarities calculated by the calculation unit 23 for each keyword included in the first sentence.

  Specifically, first, the similarity change rate calculation unit 25 calculates the similarity change rates for the three types of similarity.

  Here, the similarity change rate is a scale indicating a change in similarity when the granularity of the similarity determination unit between the first sentence and the second sentence is coarsened from the keyword to the phrase and from the clause to the dependency relation of the phrase. It is.

  For example, as the similarity between phrases decreases with respect to the similarity between keywords, the first sentence and the second sentence are similar at the keyword level, but are similar in phrase units including the periphery of the keyword. Will show that not. That is, when viewed in a coarser unit than the keyword level, it indicates that the phrases are not similar to each other as represented by the similarity between the keywords.

  In other words, with respect to the similarity between keywords, the first sentence and the second sentence show the similarity at the keyword level as the degree of decrease in similarity between phrases expanded to the phrase including the keyword decreases. This indicates that the phrase unit including the surroundings is maintained.

  Similarly, as the similarity between clauses decreases with respect to the similarity between clauses, the first and second sentences are more similar at the clause level, but include the dependency relationship between clauses. This means that the unit is not similar. That is, when viewed in a coarser unit than a phrase, it indicates that the dependency relationship between the phrases is not similar as the degree of similarity between the phrases is expressed.

  In other words, the lower the degree of decrease in similarity between clauses relative to the similarity between clauses, the more the first sentence and the second sentence have the similarity at the phrase level to the dependency relation between phrases. It shows that the unit is included.

  Summarizing the above, it can be said that a sentence having a lower degree of decrease in similarity even if the degrees of similarity are compared with a coarser granularity has a higher similarity between the sentences.

  Therefore, the similarity change rate is such that the similarity between the first sentence and the second sentence decreases as the similarity determination unit becomes rough, for example, so as to be compatible with the above-described similarity change state. The absolute value of the similarity change rate is set to be large. Specifically, the similarity change rate calculation unit 25 calculates the similarity change rate for all keywords included in the first sentence using the equations (1) and (2).

Here, d _ws (word ₁ ) represents the degree of similarity change when the similarity determination unit is coarsened from a keyword to a phrase for an arbitrary keyword word ₁ included in the first sentence. The word ₂ is a corresponding keyword of the second sentence corresponding to the keyword word ₁ , and sim (word ₁ , word ₂ ) indicates the similarity between the keywords. Further, seg ₁ indicates a clause including the keyword word ₁ and seg ₂ indicates a clause including the keyword word ₂ . Therefore, sim (seg ₁ , seg ₂ ) indicates the similarity between corresponding phrases.

D _sd (word ₁ ) is the rate of change of similarity for the case of arbitrary keyword word ₁ included in the first sentence, when the similarity determination unit is coarsened from the clause containing keyword word ₁ to the dependency relationship of the clause. Show. dep ₁ indicates the dependency relationship of the phrase including the keyword word ₁ , and dep ₂ indicates the dependency relationship of the phrase including the keyword word ₂ . Therefore, sim (dep ₁ , dep ₂ ) indicates the similarity between the corresponding clauses.

In equations (1) and (2), the similarity when the similarity determination unit is fine is subtracted from the similarity when the similarity determination unit is coarser, but the similarity when the similarity determination unit is finer is subtracted. D _ws (word ₁ ) and d _sd (word ₁ ) may be calculated by subtracting the similarity when the similarity determination unit is coarse.

  The similarity evaluation unit 26 uses the similarity between the keywords calculated by the calculation unit 23 and the similarity change rate calculated by the similarity change rate calculation unit 25 to determine the similarity between the first sentence and the second sentence. Is expressed as a score.

Specifically, the similarity evaluation unit 26 calculates a score SIM (S ₁ , S ₂ ) indicating the degree of similarity between the first sentence and the second sentence using the expression (3).

Here, S ₁ represents the first sentence and S ₂ represents the second sentence. w represents a keyword included in the first sentence S ₁ , and N represents the number of keywords included in the first sentence S ₁ . Further, w ′ _s2 represents the corresponding keyword of the second sentence S ₂ having the highest similarity with the keyword w. As can be seen from the equation (3), the score SIM (S ₁ , S ₂ ) is normalized to take a value in the range of 0 to 1, and the score SIM (S ₁ , S ₂ ) is set to “1”. The closer it is, the more similar the first sentence S ₁ and the second sentence S ₂ are.

  As described above, the calculation unit 20 calculates the similarity between the first sentence and the second sentence.

<Operation of similarity evaluation device>
Next, the operation of the similarity evaluation device 100 according to the first embodiment will be described.

  When the similarity evaluation apparatus 100 receives the first sentence and the second sentence described in a natural language by the input unit 10, the similarity evaluation apparatus 100 stores the received first sentence and second sentence in the storage unit 30, for example. And the similarity evaluation apparatus 100 performs the similarity evaluation processing routine shown in FIG. 3 by CPU.

  First, in step S100, dependency analysis is performed on the first sentence and the second sentence using a dependency analyzer, and information on the language structure acquired from each sentence is acquired. Then, keywords are extracted from each sentence of the first sentence and the second sentence based on the information on the language structure, and the acquired information on the language structure and the keyword are stored in the storage unit 30 in association with each sentence. To do.

  In step S102, with reference to the language structure information and keywords corresponding to each sentence of the first sentence and the second sentence acquired in step S100, each sentence is included based on a predetermined concept vector model. A keyword vector is generated for each of the keywords to be displayed.

  Then, by referring to the phrase information included in the information related to the language structure, by synthesizing the keyword vector of the keyword included in the phrase and the morpheme vector other than the keyword included in the phrase, the first sentence and the second sentence For each sentence, generate each phrase vector of the phrase containing the keyword.

  Furthermore, by referring to the dependency relations of the clauses included in the information related to the language structure and synthesizing the phrase vectors having the dependency relationships, the dependency of the clauses for each sentence of the first sentence and the second sentence is obtained. Generate a dependency vector for each of the relationships.

  The generated keyword vector, phrase vector, and dependency vector are stored in the storage unit 30 in association with each of the first sentence and the second sentence.

  In step S104, one keyword selected from the storage unit 30 is selected as the keyword of interest that has not been selected in step S100 and associated with the first sentence.

  In step S106, a keyword vector corresponding to the attention keyword selected in step S104 is acquired from the storage unit 30. Then, the cosine distance between the keyword vector corresponding to the attention keyword and all the keyword vectors associated with the second sentence is calculated, and the second sentence having the shortest cosine distance with the keyword vector corresponding to the attention keyword. The keyword represented by the keyword vector associated with is acquired as a corresponding keyword. At this time, the cosine distance between the attention vector and the corresponding vector is stored in the storage unit 30 as the similarity between the keywords.

  In step S108, referring to the information on the language structure acquired in step S100, the first sentence phrase including the attention keyword selected in step S104, and the second sentence phrase including the corresponding keyword acquired in step S106; To get.

  Then, the cosine distance between the phrase vector corresponding to the phrase of the first sentence including the keyword of interest and the phrase vector corresponding to the phrase of the second sentence including the corresponding keyword, generated in step S102, is used as the similarity between the phrases. The calculated degree of similarity between clauses is stored in the storage unit 30.

  Needless to say, the similarity between phrases may be expressed by an index other than the cosine distance.

  In step S110, referring to the information on the language structure acquired in step S100, the dependency relationship of the first sentence including the attention keyword selected in step S104 and the second sentence including the corresponding keyword acquired in step S106. Get the dependency relationship of the phrase.

  A dependency vector corresponding to the dependency relationship of the first sentence including the attention keyword, and a dependency vector corresponding to the dependency relationship of the second sentence including the corresponding keyword, generated in step S102; The cosine distance is calculated as the similarity between clause dependencies, and the calculated similarity between clause dependencies is stored in the storage unit 30.

  Needless to say, the similarity between clauses may be expressed by an index other than the cosine distance.

  Through the processing in steps S104 to S110, the similarity between keywords between the keyword of interest and the corresponding keyword of the second sentence most similar to the keyword of interest, the similarity of clauses between the phrase including the keyword of interest and the phrase including the corresponding keyword , And the degree of similarity between the dependency of the clauses between the dependency relationship of the clause including the keyword of interest and the dependency relationship of the clause including the corresponding keyword.

  In step S112, it is determined whether or not all the keywords included in the first sentence have been selected in step S104. If there are keywords that have not been selected in step S104, the process proceeds to step S104. Each keyword included in the first sentence is repeatedly selected by selecting an unselected keyword included in the first sentence as a keyword of interest until there is no unselected keyword among the keywords included in the first sentence. On the other hand, the similarity between keywords, the similarity between clauses, and the similarity between clause dependencies are calculated.

  On the other hand, if the determination process in step S112 is affirmative, that is, if all keywords included in the first sentence are selected in step S104, the process proceeds to step S114.

In step S114, the similarity sim (word ₁ , word ₂ ) between keywords calculated in step S106 is the keyword word ₁ included in the first sentence and the corresponding keyword word ₂ of the second sentence corresponding to the keyword. Based on the similarity sim (seg ₁ , seg ₂ ) between clauses corresponding to each keyword calculated in step S108, the similarity change rate d _ws (word ₁ ) according to the above equation ( ₁ ). Is calculated for each keyword word ₁ included in the first sentence.

Also, the similarity sim (seg ₁ , seg ₂ ) between clauses calculated in step S108 and the similarity sim between the dependency of the clauses corresponding to each of the clauses seg ₁ and seg ₂ calculated in step S110. Based on (dep ₁ , dep ₂ ), the similarity change rate d _sd (word ₁ ) is calculated for each keyword word ₁ included in the first sentence, according to the above equation (2).

In step S116, the similarity sim (word ₁ , word ₂ ) between the keywords calculated in step S 106, the similarity change rates d _ws (word ₁ ) and d _sd (word ₁ ) calculated in step S 114, Based on the above, the score SIM (S ₁ , S ₂ ) indicating the degree of similarity between the first sentence S ₁ and the second sentence S ₂ is calculated according to the above equation (3). The calculated score SIM (S ₁ , S ₂ ) is stored in the storage unit 30, and the output unit 40 gives a score SIM (S of the first sentence S ₁ and the second sentence S ₂ to a display device such as a display, for example. ₁ , S ₂ ) is output.

In equation (3), the similarity change rates d _ws (word ₁ ) and d _sd (word ₁ ) are used to calculate the score SIM (S ₁ , S ₂ ). For example, the similarity change rate d _ws The score SIM (S ₁ , S ₂ ) may be calculated using at least one of (word ₁ ) and d _sd (word ₁ ).

<Execution result of similarity evaluation device>
FIG. 4 shows the score SIM (S in the similarity evaluation apparatus 100 according to the first embodiment when “I want to change PW” is the first sentence and “I can't log in after changing my password” is the second sentence. FIG. 2 is a diagram illustrating an example of a calculation process of ₁ , S ₂ ).

  In this case, for example, “PW” and “change” are extracted as keywords of the first sentence, and “password”, “change”, and “login” are extracted as keywords of the second sentence, for example. “Change” is a standard notation of “if changed”.

  When the similarity between the keywords was calculated, the keyword of the second sentence having the highest similarity with “PW” was “password”, and the similarity was 0.90. In addition, the keyword of the second sentence having the highest similarity with “change” is “change”, and the similarity is 0.95.

  When the similarity between clauses is calculated, the similarity between clauses of the phrase “PW” including the keyword “PW” of the first sentence and the phrase “password” including the keyword “password” of the second sentence is It was 0.75. In addition, the similarity between phrases of the phrase “change” including the keyword “change” in the first sentence and the phrase “if changed” including the keyword “change” in the second sentence was 0.32.

  Further, when the similarity between clause dependencies is calculated, the dependency relationship “change PW” including the dependency destination of the phrase “PW” of the first sentence and the clause “password” of the second sentence are included. The degree of similarity with the dependency relationship “if the password is changed” including the dependency point is 0.15. Also, the dependency relationship “I want to change” including the dependency of the phrase “change” in the first sentence and the dependency relationship “including the dependency of the phrase“ if changed ”in the second sentence“ I cannot log in if I change. The degree of similarity was 0.04.

Furthermore, the similarity change rate d _ws (PW) for the keyword “PW” is “−0.15” from the equation (1), and the similarity change rate d _sd (PW) is “−0.60” from the equation (2). Become. The similarity change rate d _ws (change) for the keyword “change” is “−0.63” from the equation (1), and the similarity change rate d _sd (change) is “−0.28” from the equation (2). Become.

  Therefore, the score SIM indicating the degree of similarity between the first sentence "I want to change PW" and the second sentence "I can't log in if I change my password" (I want to change my PW, I can't log in if I change my password) From equation (3), it is “0.59”.

  As described above, when the similarity evaluation apparatus 100 according to the first embodiment evaluates the similarity between two sentences described in a natural language, not only the comparison of similarities between keywords included in each sentence, The similarity between two sentences is evaluated in consideration of word order and syntactic similarity such as the phrase including the keyword and the meaning of the dependency expression of the phrase.

  Therefore, the similarity evaluation apparatus 100 is more accurate in comparing the similarity to the sentence to be compared with the conventional similarity evaluation apparatus that evaluates the similarity of each sentence only by the similarity between the keywords included in the sentence. Can be evaluated well.

  For example, when comparing the first sentence “I can no longer send mail” with the second sentence A “I cannot send mail” and the second sentence B “I cannot send mail”, only the keyword “mail” Since both the second sentence A and the second sentence B include the keyword “mail”, there is a difference between the similarity between the first sentence and the second sentence A, and the similarity between the first sentence and the second sentence B. There is no.

  Further, even when the similarity determination unit is extended to the phrase, the phrase “e-mail” of the first sentence is included in both the second sentence A and the second sentence B, so the first sentence and the second sentence A And the similarity between the first sentence and the second sentence B is not different.

  However, when the similarity determination unit is expanded to the dependency relation of clauses, the second sentence A is “unable to send mail” and the second sentence B is different from the first sentence “no longer able to send mail”. Is "I have an email". Therefore, it turns out that the 1st sentence is more similar to the 2nd sentence A than the 2nd sentence B, The evaluation value is output as a score.

  In the similarity evaluation apparatus 100 according to the first embodiment, as an example, the input unit 10 has been described to receive the first sentence and the second sentence as text. However, for example, the input unit 10 receives voices corresponding to the first sentence and the second sentence, and performs the known voice recognition for converting the voices into texts on the received voices, thereby converting the first texts into texts. The sentence and the second sentence may be acquired.

  In this case, since speech can be used as the input interface of the similarity evaluation apparatus 100, it is not necessary to preliminarily make the content for evaluating the similarity text. Therefore, the operability of the similarity evaluation device 100 can be improved as compared with the case where text is received as input.

Second Embodiment
In the first embodiment, sentences having a low degree of decrease in the degree of similarity change calculated based on the similarity between keywords, the similarity between clauses including a keyword, and the similarity between clauses including a keyword It was explained that the similarity of each sentence was high, in other words, this is similar even if the similarity is compared with the coarser similarity determination unit granularity among the keywords included in the sentence. It is indicated that a keyword having a smaller degree of decrease is an important keyword having a larger influence on the determination of sentence similarity.

  Accordingly, in the second embodiment, a keyword evaluation apparatus 200 that inputs two similar sentences, a first sentence and a second sentence, and evaluates the importance of each keyword included in the first sentence will be described.

<System configuration example>
FIG. 5 is a diagram illustrating a system configuration example of the keyword evaluation device 200. The difference between the system configuration of the keyword evaluation device 200 of FIG. 5 and the system configuration example of the similarity evaluation device 100 of FIG. 1 according to the first embodiment is that the similarity evaluation unit 26 is replaced with the keyword importance evaluation unit 26A. Accordingly, the evaluation unit 24 is replaced with an evaluation unit 24A.

  The other keyword evaluation apparatus 200 has the same configuration as the system configuration example of the similarity evaluation apparatus 100.

  The keyword importance degree evaluation unit 26A determines, for each keyword included in the first sentence, the similarity between the keywords calculated by the calculation unit 23 and the similarity change rate calculated by the similarity change rate calculation unit 25. Calculate keyword importance.

When determining the similarity with the second sentence S ₂ , the importance SIM (S _{1, w} , S ₂ ) of the keyword w included in the first sentence S ₁ is calculated by, for example, Expression (4).

Here, w ′ _s2 represents the corresponding keyword of the second sentence S ₂ having the highest similarity with the keyword w included in the first sentence S ₁ as described in the expression (3). Also, as can be seen from the equation (4), the importance SIM (S _{1, w} , S ₂ ) is normalized to take a value in the range of 0 to 1, and the importance SIM (S _{1, w} , The closer S ₂ ) is to “1”, the more important the keyword w is.

<Operation of keyword evaluation device>
When the keyword evaluation device 200 receives a similar first sentence and second sentence by the input unit 10, the keyword evaluation apparatus 200 stores the received first sentence and second sentence in the storage unit 30, for example. Then, the keyword evaluation device 200 executes a keyword evaluation processing routine shown in FIG.

  The keyword evaluation processing routine shown in FIG. 6 is different from the similarity evaluation processing routine of the similarity evaluation apparatus 100 according to the first embodiment shown in FIG. 3 in that step S116 is replaced with step S118. Other processes are the same as the similarity evaluation processing routine of the similarity evaluation apparatus 100. Therefore, the process of step S118 will be described below.

In step S118, based on the similarity sim (word ₁ , word ₂ ) between the keywords calculated in step S106 and the similarity change rates d _ws (word ₁ ) and d _sd (word ₁ ) calculated in step S114. Thus, according to the above equation (4), in determining the similarity of the first sentence to the second sentence, the importance w SIM (S _{1, w} , S ₂ ) of the keyword w indicating the degree of influence on the determination result is determined as the keyword w Calculate every time.

The calculated importance SIM (S _{1, w} , S ₂ ) is stored in the storage unit 30, and the importance SIM (S _{1, w} , S ₂ ) for each keyword w is displayed on the display device such as a display by the output unit 40. Is output.

The formula (4) for calculating the importance SIM (S _{1, w} , S ₂ ) of the keyword w is an example, and the absolute value of the change value between the similarity between keywords and the similarity between clauses is The smaller the value, the higher the importance SIM (S _{1, w} , S ₂ ), or the smaller the absolute value of the change between the similarity between clauses and the similarity between clauses, the more important If the degree SIM (S _{1, w} , S ₂ ) can be highly evaluated, the importance SIM (S _{1, w} , S ₂ ) can be calculated using another evaluation formula instead of the formula (4). Good.

Thus, according to the keyword evaluation device 200 according to the second embodiment, the importance SIM (S of the keyword w becomes smaller as the absolute values of the at least two similarity change rates d _ws (w) and d _sd (w) are smaller. _{1, w} , S ₂ ) are highly appreciated.

  Therefore, for example, in a search system for searching for a sentence by inputting a keyword, if a keyword with higher importance obtained by the keyword evaluation apparatus 200 is preferentially input, the target sentence can be searched with high accuracy. Can do.

  Note that the keyword evaluation device 200 receives the speech corresponding to the first sentence and the second sentence and performs known speech recognition for converting the speech into text, as in the similarity evaluation device 100 according to the first embodiment. Thus, the first sentence and the second sentence converted into text may be acquired.

<Third Embodiment>
In the similarity evaluation device 100 according to the first embodiment and the keyword evaluation device 200 according to the second embodiment, the similarity change rates d _ws (w) and d _sd (w) at different granularities of sentences are calculated, Based on the calculated similarity change rates d _ws (w) and d _sd (w), the importance of keywords used to determine the similarity between sentences or the similarity between sentences was evaluated.

In the third embodiment, similarly to the similarity evaluation device 100 according to the first embodiment and the keyword evaluation device 200 according to the second embodiment, the similarity change rates d _ws (w) and d _sd (w) are calculated. The search device 300 that searches for a sentence most similar to the first sentence from a plurality of sentences based on the calculated similarity change rates d _ws (w) and d _sd (w) will be described.

<System configuration example>
FIG. 7 is a diagram illustrating a system configuration example of the search device 300. The system configuration example of the search device 300 in FIG. 7 is different from the system configuration example of the similarity evaluation device 100 in FIG. 1 according to the first embodiment, in that the similarity evaluation unit 26 is replaced with a query sentence similarity evaluation unit 26B. Accordingly, the evaluation unit 24 is replaced with the evaluation unit 24B. Further, a search unit 27 is added to the search device 300, and a search target sentence DB 30A is built in the storage unit 30 in advance.

  Other configurations of the search device 300 are the same as the system configuration example of the similarity evaluation device 100.

  In the search target sentence DB 30A, it is assumed that a plurality of search target sentences are stored in advance in association with keyword vectors, phrase vectors, and dependency vectors generated in the same manner as in the first embodiment.

  Here, as the keyword vector, phrase vector, and dependency vector associated with each search target sentence, for example, the output result of the vector generation unit 22 when the search target sentence is input to the search apparatus 300 can be used.

  The search device 300 searches for a search target sentence most similar to the content of a query sentence described in a natural language from a plurality of search target sentences stored in advance in the search target sentence DB 30A.

  Thus, in the search device 300, since a plurality of search target sentences are stored in the search target sentence DB 30A in advance, the similarity evaluation apparatus 100 according to the first embodiment and the keyword evaluation apparatus 200 according to the second embodiment. Unlike the above, only the query sentence is input to the input unit 10.

  Therefore, the sentence analysis unit 21 performs dependency analysis on the query sentence received from the input unit 10 by a method similar to that of the similarity evaluation apparatus 100, and based on the result of the dependency analysis, the keyword is calculated from the query sentence. To extract.

  Then, the vector generation unit 22 generates a keyword vector, a phrase vector, and a dependency vector based on the keyword extracted from the query sentence by the sentence analysis unit 21 by the same method as the similarity evaluation device 100.

  The calculation unit 23 also includes a keyword vector, a phrase vector, and a dependency vector corresponding to the query sentence generated by the vector generation unit 22, and a keyword vector and a phrase vector corresponding to the search target sentence stored in the search target sentence DB 30A. And the dependency vectors, the similarity between keywords, the similarity between phrases, and the similarity between phrases are calculated by the same method as the similarity evaluation apparatus 100.

Similar to the similarity evaluation unit 26 of the similarity evaluation apparatus 100 according to the first embodiment, the query sentence similarity evaluation unit 26B and the similarity between keywords calculated by the calculation unit 23 and the similarity change rate calculation unit 25 On the basis of the similarity change rate calculated in (5), the score SIM (S ₁ , S ₂ ) of the expression (3) is calculated for each query target sentence S ₂ for the query sentence S ₁ . Then, the query statement similarity degree evaluation unit 26B, for example score SIM (S _1, S ₂₎ is the most "1" becomes close to the search subject sentence S _2, the search subject sentence S ₂ that are similar to the contents of the query statement S ₁ Evaluate as

  However, in the case of a device that searches for a search target sentence corresponding to the content of a query sentence, such as the search device 300, the similarity between keywords, the similarity between phrases, and the similarity between clause dependencies are high. However, an appropriate search target sentence corresponding to the content of the query sentence is not necessarily searched.

  For example, a keyword that frequently appears in a sentence tends to represent the subject of the sentence, whereas a keyword that frequently appears in a plurality of sentences is not an important keyword.

  Therefore, in addition to the similarity between keywords, the similarity between clauses, and the similarity between clauses, it is preferable to further consider the weight of the keyword in the sentence.

  There are known methods such as the Term Frequency-Inverse Document Frequency (TF-IDF) method or the BM25 method as the keyword weight calculation method in the sentence. The query sentence similarity evaluation unit 26B has such a known method. A keyword weight calculation method can be applied.

Therefore, the query sentence similarity evaluation unit 26B takes a score Score (S) indicating the degree of similarity between the query sentence S ₁ and the search target sentence S ₂ in consideration of the keyword weights calculated using a known keyword weight calculation method. ₁ , S ₂ ) is calculated using, for example, equation (5).

Here, SIM (S _{1, w} , S ₂ ) focuses only on the importance of the keyword w included in the query sentence S ₁ represented by the equation (4), that is, only the keyword w included in the query sentence S _1. Is the similarity between the query sentence S ₁ and the search target sentence S ₂ . Weight represents a weight value. Therefore, weight (argmax _ws2 (sim (w, w _s2 ))) is a weight value of the keyword w _{s2 in} the search target sentence S ₂ having the highest similarity with the keyword w included in the query sentence S ₁ .

The search unit 27 uses the score Score (S ₁ , S ₂ ) calculated by the query sentence similarity evaluation unit 26B, for example, a search target sentence S satisfying a predetermined condition with respect to the score Score (S ₁ , S ₂ ). ₂ is searched from the search target sentence DB 30 </ b> A and output to the output unit 40.

  When the search device 300 is connected to a network such as the Internet, for example, the search target sentence DB 30A is stored in an external device such as a storage device connected to the network, and the search device 300 is stored in the external device. You may make it refer to sentence DB30A.

<Operation of search device>
When the input unit 10 receives a query sentence written in a natural language, the search device 300 stores the received query sentence in the storage unit 30, for example. Then, the search device 300 executes a search processing routine shown in FIG. 8 by the CPU.

  The difference between the search processing routine shown in FIG. 8 and the similarity evaluation processing routine of the similarity evaluation apparatus 100 according to the first embodiment shown in FIG. 3 is that step S107 is added instead of step S106. . In the search processing routine, steps S103, S120, and S122 are newly added. Other processes are the same as the similarity evaluation processing routine of the similarity evaluation apparatus 100. Therefore, the search processing routine will be described below with a focus on processing different from the similarity evaluation processing routine.

  In step S103, one unselected search target sentence is selected from a plurality of search target sentences stored in advance in the search target sentence DB 30A.

  In step S107, a corresponding keyword that is most similar to the attention keyword selected in step S104 is extracted from the keywords associated with the search target sentence selected in step S103. The similarity between the keywords may be calculated by, for example, the cosine distance of each keyword vector corresponding to the keyword of interest and the corresponding keyword, as in step S106 in FIG. 3, and the calculated cosine distance is calculated based on the similarity between the keywords. It is stored in the storage unit 30 as the degree.

  Thereafter, in steps S108 and S110, the similarity between the clause including the keyword of interest and the clause including the corresponding keyword, and the similarity of the dependency of the clause including the keyword of interest and the dependency of the clause including the corresponding keyword are calculated. In step S112, it is determined whether the processes in steps S104 to S112 have been executed for all the keywords included in the query sentence.

In step S114, the similarity sim (word ₁ , word ₂ ) between the keywords word ₁ included in the query sentence calculated in step S107 and the corresponding keyword word ₂ of the search target sentence corresponding to the keyword is calculated. Based on the similarity sim (seg ₁ , seg ₂ ) between clauses corresponding to each keyword calculated in step S108, the similarity change rate d _ws (word ₁ ) Is calculated for each keyword included in the query statement.

Also, the similarity sim (seg ₁ , seg ₂ ) between clauses calculated in step S108 and the similarity sim between the dependency of the clauses corresponding to each of the clauses seg ₁ and seg ₂ calculated in step S110. Based on (dep ₁ , dep ₂ ), the similarity change rate d _sd (word ₁ ) is calculated for each keyword included in the query sentence according to the above equation (2).

Next, in step S116, the similarity sim (word ₁ , word ₂ ) between the keywords calculated in step S107 and the similarity change rates d _ws (word ₁ ) and d _sd (word ₁ ) calculated in step S114 based on the above (4) according to formula similarity SIM (S ₁ to the query statement S ₁ in the case of focusing only on the keyword w included in the query statement S ₁ and the search subject sentence S _{2, w,} S ₂ ), and calculates for each of the query statement S ₁ keyword.

Further, the similarity SIM (S _{1, w} , S ₂ ) for each keyword w calculated in this step and the keyword w in the search target sentence S ₂ having the highest similarity with the keyword w included in the query sentence S ₁ Based on the weight value of _s2 , the score Score (S ₁ , S ₂ ) is calculated according to the above equation (5). The calculated score Score (S ₁ , S ₂ ) is associated with the search target sentence S ₂ and stored in, for example, the search target sentence DB 30A.

Note that the weight value of the keyword w _s2 may be calculated using a known weight calculation method such as the TF-IDF method as described above.

Here, as an example, the similarity between the query sentence and the search target sentence is calculated according to equation (5), but the method of calculating the similarity between the query sentence and the search target sentence is not limited to this. For example, the similarity may be calculated according to the expression (3), and the sum of the importance SIM (S _{1, w} , S ₂ ) for each keyword in the query sentence calculated according to the expression (4) is used as the query sentence. It is good also as a score which shows the similarity degree with a search object sentence.

  In step S120, it is determined whether or not all search target sentences included in the search target sentence DB 30A are selected in step S103. If there is a search target sentence that has not been selected in step S103, the process proceeds to step S103. To do.

Then, there is no unselected search target sentence in the search target sentences included in the search target sentence DB 30A, and the search target sentence is repeatedly selected in step S103 until the determination process in step S120 becomes affirmative determination, thereby step S116. The score Score (S ₁ , S ₂ ) of each search target sentence S ₂ for the query sentence S ₁ is calculated.

  On the other hand, if the determination process in step S120 is affirmative, the process proceeds to step S122.

In step S122, the score Score (S ₁ , S ₂ ) associated with each retrieval target sentence is referred to, and the score Score (S ₁ , S ₂ ) equal to or higher than a predetermined threshold is associated with the retrieval target sentence. Is retrieved from the retrieval target sentence DB 30A. And the output part 40 outputs the search object sentence acquired by step S122 to display apparatuses, such as a display, for example.

Note that the search target sentence acquired in step S122 is not limited to the search target sentence associated with a score Score (S ₁ , S ₂ ) equal to or higher than a predetermined threshold. For example, a predetermined number of search target sentences may be searched from the search target sentence DB 30A in order from the highest score Score (S ₁ , S ₂ ).

  In search device 300, voice corresponding to the query sentence is received by input unit 10, and the query sentence converted into text is obtained by performing known voice recognition for converting the voice into text for the received voice. You may make it do.

  In this case, since the voice received by the search device 300 can be used for the search as it is, the operability of the search device 300 can be improved compared to the case of receiving text as input.

As described above, the search device 300 according to the third embodiment searches for a search target sentence similar to the content of the query sentence from the search target sentence DB 30 </ b> A stored in the storage unit 30 in advance. In this case, when the query device receives the query sentence, the search device 300 acquires information on the language structure using the dependency analyzer only for the query sentence, and also includes a keyword vector corresponding to the keyword included in the query sentence, A phrase vector and a dependency vector are generated. That is, the keyword vector, phrase vector, and dependency vector related to the search target sentence are stored in the search target sentence DB 30A in advance, so that each time a query sentence is received, various vectors related to the search target sentence are generated. The score Score (S ₁ , S ₂ ) can be calculated at high speed.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  For example, although the similarity evaluation device 100, the keyword evaluation device 200, and the search device 300 described above have a computer system therein, the “computer system” is a case where a WWW system is used. It also includes the homepage provision environment (or display environment).

  In the specification of the present application, the embodiment in which the program is preinstalled in the ROM has been described. However, the program may be provided by being stored in a computer-readable recording medium.

DESCRIPTION OF SYMBOLS 10 ... Input part 20 ... Operation part 21 ... Sentence analysis part 22 ... Vector generation part 23 ... Calculation part 24 (24A, 24B) ... Evaluation part 25 ... Similarity change Rate calculation unit 26 ... similarity evaluation unit 26A ... keyword importance evaluation unit 26B ... query sentence similarity evaluation unit 27 ... search unit 30 ... storage unit 40 ... output unit 100 ..Similarity evaluation device 200 ... Keyword evaluation device 300 ... Search device
d _sd , d _ws ... similarity change rate 30A ... search target sentence DB

Claims (12)

Similarity between keywords based on the first keyword extracted from the input first sentence and the second keyword that is similar to the first keyword in the input second sentence A calculation unit that calculates at least two similarities among similarities between clauses including a keyword and similarities between clauses including a keyword;
An evaluation unit that evaluates the importance of the first keyword higher as the absolute value of the change value of the at least two similarities calculated by the calculation unit is smaller;
Keyword evaluation device including The evaluation unit evaluates the importance of the first keyword higher as the absolute value of the change value between the similarity between the keywords calculated by the calculation unit and the similarity between the phrases is smaller, or The importance of the first keyword is evaluated higher as the absolute value of the change value between the similarity between the clauses calculated by the calculation unit and the similarity between the dependency of the clauses is smaller. The keyword evaluation device described. The word included in the first sentence in the combination having the highest similarity among the combinations of the word included in the input first sentence and the word included in the input second sentence is the first keyword, the first The word included in the two sentences is set as the second keyword, and at least two similarities are calculated among the similarity between the keywords, the similarity between the clauses including the keyword, and the similarity between the dependencies of the clause including the keyword. A calculation unit;
An evaluation unit that evaluates that the first sentence and the second sentence are similar as the absolute value of the change value of the at least two similarities calculated by the calculation unit is smaller;
Similarity evaluation apparatus including In the evaluation unit, the first sentence and the second sentence are more similar as the absolute value of the change value between the similarity between the keywords calculated by the calculation unit and the similarity between the phrases is smaller. The first sentence and the second sentence are smaller as the absolute value of the change value between the similarity between the clauses calculated by the calculation unit and the similarity between the dependency of the phrases is smaller. The degree-of-similarity evaluation apparatus according to claim 3. For each word included in the input first sentence, the calculation unit has the highest similarity between the word included in the first sentence and the word included in the first sentence. The word included in the second sentence is the second keyword, the similarity between the keywords, the similarity between clauses including the keyword, and the similarity between clauses including the keyword are calculated,
The evaluation unit changes, for each word included in the first sentence, a similarity between the first keyword and the second keyword, a similarity between the keywords, and a similarity between the phrases. Based on the average value of the score calculated based on the absolute value of the value and the absolute value of the change value of the similarity between the clauses and the similarity between the clauses, the first sentence and the The similarity evaluation apparatus according to claim 4, wherein it is evaluated whether the second sentence is similar. For each of a plurality of search target sentences prepared in advance, a keyword vector representing each keyword included in the search target sentence, a phrase vector representing a phrase including a keyword for each keyword, and a keyword for each keyword A storage unit for storing a dependency vector representing a dependency relationship including a clause destination;
For each of the plurality of search target sentences, a keyword included in the query sentence in a combination having the highest similarity among the combinations of the keyword included in the input query sentence and the keyword included in the search target sentence is first. The keyword included in the search target sentence is a second keyword, and the similarity between the keywords based on the keyword vector, the similarity between the phrases including the keyword based on the phrase vector, and the keyword based on the dependency vector are included. A calculation unit that calculates at least two similarities for each of the plurality of search target sentences among the similarities between the dependency of the clauses;
Evaluation that evaluates that the query sentence and the search target sentence are more similar as the absolute value of the change value of the at least two similarities calculated by the calculation unit is smaller for each of the plurality of search target sentences. And
A search unit that searches for a search target sentence similar to the query sentence, based on an evaluation result by the evaluation unit;
Search device including A computer is a first keyword extracted from a first sentence input through an input unit, and a sentence input through the input unit together with the first sentence, and is associated with the first sentence Based on the second keyword that is similar to the first keyword in the second sentence stored in the storage device, the similarity between the keywords, the similarity between the phrases including the keyword, and the phrase including the keyword calculating at least two similarity of the dependency similarity between,
Displaying the evaluation result evaluated so that the importance of the first keyword is higher as the absolute value of the change value of the at least two similarities is smaller ;
Keyword evaluation method to execute . A computer includes a word included in a first sentence input via an input unit, and a sentence input via the input unit together with the first sentence, and is associated with the first sentence in a storage device Among the combinations with the words included in the stored second sentence, the word included in the first sentence in the combination having the highest similarity is the first keyword, and the word included in the second sentence is the second keyword. Calculating a similarity between at least two of a similarity between keywords, a similarity between clauses including the keyword, and a similarity between clauses including the keyword ;
A step of evaluating that the first sentence and the second sentence are similar to each other as the absolute value of the change value of the at least two similarities is smaller, and displaying the evaluation result on a display device;
To evaluate the similarity of statements that execute . For each of a plurality of search target sentences prepared in advance by a computer, a keyword vector representing each keyword included in the search target sentence, a phrase vector representing a phrase including a keyword for each keyword, and a keyword for each keyword A dependency vector representing a dependency relationship including the dependency destination of the clause including the symbol is generated, the keyword vector, the phrase vector, and the dependency vector generated for each of the plurality of search target sentences are associated and stored in the storage device . Steps,
For each of the plurality of search target sentences, included in the query sentence in the combination having the highest similarity among the combinations of the keyword included in the query sentence input via the input unit and the keyword included in the search target sentence The first keyword, the keyword included in the search target sentence as the second keyword, the similarity between keywords based on the keyword vector, the similarity between phrases including the keyword based on the phrase vector, and the dependency vector And calculating at least two similarities for each of the plurality of search target sentences among the similarities of the dependency of the clause including the keyword based on the keywords, and corresponding the calculated at least two similarities for each of the plurality of search target sentences And storing in the storage device;
For each of the plurality of search target sentences, a step of evaluating that the query sentence and the search target sentence are more similar as the absolute value of the calculated change value of the at least two similarities is smaller ;
Searching a search target sentence similar to the query sentence based on the evaluation, and displaying a search target sentence similar to the query sentence on a display device;
How to search for statements that execute   The program for functioning a computer as each part of the keyword evaluation apparatus of Claim 1 or Claim 2.   The program for functioning a computer as each part of the similarity evaluation apparatus of any one of Claims 3-5.   The program for functioning a computer as each part of the search device of Claim 6.