JP6955963B2 - Search device, similarity calculation method, and program - Google Patents

Description

The present invention relates to a search device, a similarity calculation method, and a program.

In Patent Document 1, the difference between the two sentences is extracted by converting each of the two sentences into a feature vector, calculating the similarity, and then excluding the words having the same dependency relationship using syntactic analysis. The technology to be used is disclosed.

Japanese Patent No. 5367099

When comparing two sentences, when the similarity between the feature vectors of the sentences is calculated as described in Patent Document 1, the similarity is high when words similar to each other appear in common in the two sentences. Become. However, words that are similar to each other do not always play the same role in the two sentences. For example, in the first sentence "A boy has a small dictionary" and the second sentence "A small boy has a dictionary", the word "small" modifies "dictionary" in the first sentence. On the other hand, in the second sentence, the word "small" modifies "boy". On the other hand, since the words appearing in the two sentences match, the similarity of the feature vectors of the sentences is highly evaluated.
An object of the present invention is to provide a search device for calculating the similarity, a similarity calculation method, and a program in consideration of the roles of words appearing in two sentences.

According to the first aspect of the present invention, the similarity identification device (100) includes a feature amount specifying unit (105) that specifies a feature amount for each word constituting the first sentence and the second sentence, and the first sentence. The role specifying unit (106) for specifying the grammatical role of each of the words constituting the sentence 1 and the second sentence, the feature quantities of the words constituting the first sentence, and the second sentence. Of the words constituting the first sentence, the inter-word similarity specifying unit (108) for specifying the inter-word similarity, which is the similarity with the feature amount related to the word having a common role with the word constituting the first sentence, It is provided with an inter-sentence similarity specifying unit (109) that specifies the inter-sentence similarity, which is the similarity between the first sentence and the second sentence, based on the inter-word similarity for each word. As a result, the similarity identification device can identify the similarity by taking into account the roles of the words appearing in the two sentences.

According to the second aspect of the present invention, the inter-sentence similarity specifying unit in the similarity identifying device according to the first aspect is a load of the inter-word similarity using a weighting coefficient according to the role of the word. By calculating the sum, the similarity between sentences may be specified. Thereby, the similarity identification device can identify the similarity between two sentences in consideration of the importance in the sentence for each word constituting the sentence.

According to the third aspect of the present invention, the interword similarity specifying unit in the similarity identifying device according to the first or second aspect includes words having the same role as the words constituting the first sentence. If it is not in the second sentence, the word-to-word similarity may be specified as a predetermined penalty value. As a result, the similarity identification device can calculate the inter-sentence similarity low in the comparison between sentences having different structures.

According to the fourth aspect of the present invention, the feature amount in the similarity specifying device according to any one of the first to third aspects may be a vector having a dimension number smaller than the number of words. As a result, the similarity identification device can calculate the similarity between words having different notations.

According to the fifth aspect of the present invention, the role specifying unit in the similarity specifying device according to any one of the first to fourth aspects identifies the role for each word by the phrase structure analysis process. You can.

According to the sixth aspect of the present invention, the similarity identification method includes a step in which the computer specifies the feature amount for each word constituting the first sentence and the second sentence, and the computer uses the first sentence. The step of identifying the sentence and the grammatical role of each word constituting the second sentence, the computer constructs the feature quantity of the words constituting the first sentence, and the second sentence. The step of specifying the inter-word similarity, which is the similarity with the feature amount of the word having the same role as the word constituting the first sentence among the words, and the inter-word similarity for each word by the computer. It includes a step of specifying the inter-sentence similarity, which is the similarity between the first sentence and the second sentence based on the degree.

According to a seventh aspect of the present invention, the program tells a computer a step of specifying a feature amount for each word constituting the first sentence and the second sentence, and the first sentence and the second sentence. The step of specifying the grammatical role of each word for each of the words constituting the first sentence, the feature amount of the words constituting the first sentence, and the first sentence of the words constituting the second sentence. The first sentence and the first sentence based on the step of specifying the inter-word similarity, which is the similarity with the feature quantity of the word having the same role as the word constituting the word, and the inter-word similarity for each word. The step of specifying the inter-sentence similarity, which is the similarity with the second sentence, is executed.

According to at least one of the above aspects, the similarity identification device can calculate the similarity by taking into account the roles of the words appearing in the two sentences.

It is a schematic block diagram which shows the structure of the search apparatus which concerns on 1st Embodiment. It is a figure which shows the example of the method of specifying the grammatical function by phrase structure analysis. It is a figure which shows the example of the method of specifying a grammatical function by a dependency analysis. It is a figure which shows the example of the method of specifying the corresponding part of two sentences. It is a flowchart which shows the operation of the search apparatus which concerns on 1st Embodiment. It is a schematic block diagram which shows the structure of the search apparatus which concerns on 2nd Embodiment. It is a schematic block diagram which shows the structure of the computer which concerns on at least one Embodiment.

<First Embodiment>
<< Configuration of search device >>
FIG. 1 is a schematic block diagram showing a configuration of a search device according to the first embodiment.
The search device 100 according to the first embodiment receives the input of a sentence and searches a plurality of sentences similar to the input sentence. The search device 100 is an example of a similarity specifying device that specifies the similarity between two sentences.
The search device 100 includes a conversion model generation unit 101, a conversion model storage unit 102, a sentence storage unit 103, a sentence acquisition unit 104, a feature amount specifying unit 105, a grammar function specifying unit 106, a corresponding part specifying unit 107, and a word-to-word similarity identification unit. A unit 108, a sentence-to-sentence similarity specifying unit 109, and a search result output unit 110 are provided.

The conversion model generation unit 101 generates a sentence vector having the number of dimensions of the vocabulary number from each of the plurality of learning sentences, and based on the sentence vector, generates a conversion model for generating a feature vector representing the feature of the word. Generate. The word is not limited to a word and may include a compound word or a phrase. The learning sentence may be a search target sentence stored in the sentence storage unit 103, or may be another sentence. When the search target sentence is used as the learning sentence, the conversion model generation unit 101 can create a conversion model specialized for the expression used in the search target sentence. Further, when a sentence other than the search target sentence is included in the learning sentence, the conversion model generation unit 101 can create a conversion model corresponding to an expression not used in the search target sentence.

For example, the transformation model generation unit 101 can generate a transformation model using an autoencoder. An autoencoder is a neural network in which the number of nodes in the input layer and the output layer is equal, and the number of nodes in the intermediate layer is smaller than that in the input layer and the output layer. The transformation model generation unit 101 inputs each sentence vector to the autoencoder, and trains the autoencoder so that the input and the output become equal. Then, the transformation model generation unit 101 generates a transformation model using the output of the intermediate layer as a feature vector by taking out the input layer and the intermediate layer of the autoencoder. For example, Word2Vec (https://code.google.com/archive/p/word2vec/) can be used to create a conversion model using an autoencoder.
Further, the transformation model generation unit 101 may create a transformation model based on tf (Term Frequency) -idf (Inverse Document Frequency), latent semantic analysis, and principal component analysis.

The transformation model storage unit 102 stores the transformation model generated by the transformation model generation unit 101.

The sentence storage unit 103 stores a plurality of search target sentences.
The sentence acquisition unit 104 acquires the query sentence input by the user and the search target sentence stored in the sentence storage unit 103.

The feature amount specifying unit 105 generates a feature vector for each of a plurality of words constituting the sentence acquired by the sentence acquisition unit 104, using the conversion model stored in the conversion model storage unit 102. For example, the feature amount specifying unit 105 can generate a feature vector by the following processing. The feature amount specifying unit 105 divides the sentence acquired by the sentence acquisition unit 104 into a plurality of words. Morphological analysis can be used to divide sentences. The feature amount specifying unit 105 generates a word vector having a number of dimensions equal to the number of vocabularies from each of the divided words. The feature amount specifying unit 105 obtains a feature vector by inputting a word vector into the transformation model.

The grammar function specifying unit 106 parses the sentence acquired by the sentence acquisition unit 104 and specifies the grammar function which is the role of each of the plurality of words constituting the sentence. The grammatical function refers to the classification of the elements that make up a sentence (for example, words, phrases, clauses) according to the relationship with other elements. Examples of grammatical functions include noun phrase (NP: noun phrase), verb phrase (VP: verb phrase) and adjective phrase division (phrase structure), subject, predicate, object, complement and modifier division (dependency). Relationship), classification of part words, etc. Examples of parsing include phrase structure analysis and dependency analysis. The grammar function specifying unit 106 is an example of a role specifying unit. In other embodiments, the grammatical role of each word constituting the sentence may be information other than the grammatical function.

FIG. 2 is a diagram showing an example of a method of specifying a grammatical function by phrase structure analysis.
For example, the grammar function specifying unit 106 analyzes the sentence acquired by the sentence acquisition unit 104 in a phrase structure, and generates a phrase structure tree having a plurality of words constituting the sentence as elements. A phrase structure tree is tree structure data in which a sentence is a root node, a clause or a phrase is an internal node, and each of a plurality of words is a leaf node. By the phrase structure analysis, each node is tagged to indicate the grammatical function related to the phrase structure of the node. The grammar function specifying unit 106 specifies, for each leaf node of the phrase structure tree, the permutation of all the tags attached to the path connecting the leaf node and the root node as the grammar function of the word related to the leaf node.

FIG. 3 is a diagram showing an example of a method of specifying a grammatical function by dependency analysis.
Further, for example, the grammar function specifying unit 106 receives and analyzes the sentence acquired by the sentence acquisition unit 104, and generates a dependent structure tree having a plurality of words constituting the sentence as elements. The dependency structure tree is tree structure data in which each word is a node and the node of the source word (depender) is a child node of the node of the dependency word (dependee). The dependent structure tree may be tagged with a dependency relationship on the branches connecting the nodes. The grammar function specifying unit 106 specifies the dependency (for example, the grammar function of the dependency specified by the phrase structure analysis) or the dependency relationship of each word as the grammar function based on the dependency structure tree. Examples of dependency relationships include det, which indicates the assignment of determiners, nsubj, which indicates the subject of a predicate, dobj, which indicates an object of a predicate, and amod, which indicates a modifier that modifies a noun.

FIG. 4 is a diagram showing an example of a method of specifying a corresponding portion between two sentences.
Corresponding part specifying unit 107 identifies a pair of a word of a query sentence and a word of a search target sentence and having a common grammatical function. That is, the corresponding portion specifying unit 107 specifies a word pair for each word in the query sentence by specifying a word having a common grammatical function with the word from the search target sentence. For example, as shown in FIG. 4, when the query statement is "A boy has a small dictionary" and the search target sentence is "A small boy has a dictionary", the corresponding part identification unit 107 is as follows. Identify word pairs. The word "A" in the query sentence has a grammatical function that the phrase structure is "S (sentence) -NP-DT (determiner)" and the dependency relationship is the word "det". Corresponding part specifying unit 107 identifies a word having a grammatical function that the phrase structure is "S-NP-DT" and the dependency relationship is "det" from the search target sentence. As shown in FIG. 4, the query sentence “A” has a grammatical function common to that of the search target sentence “A”. Corresponding part identification unit 107 executes this for each word, and as shown in FIG. 4, the pair of the query sentence “A” and the search target sentence “A”, the query sentence “boy” and the search target sentence. “Boy” pair, query statement “has” and search target statement “has” pair, query statement “a” and search target statement “a” pair, query statement “dictionary” and search target statement "Dictionary" pairs of, are extracted as word pairs. Further, as shown in FIG. 4, the corresponding portion specifying unit 107 does not have a word having the same grammatical function as the query sentence “small” in the search target sentence, and has the same grammatical function as the search target sentence “small”. Identify that the word you want does not exist in the query statement.

The inter-word similarity specifying unit 108 specifies the inter-word similarity, which is the similarity between feature vectors, for each pair specified by the corresponding location specifying unit 107. Examples of inter-word similarity include cosine similarity of feature vectors, Euclidean distance, Levenshtein distance, and the like. The inter-word similarity specifying unit 108 sets a predetermined penalty value as the inter-word similarity for words for which the corresponding location specifying unit 107 does not specify a pair. The penalty value is set to, for example, a value of 0 or less.

The inter-sentence similarity specifying unit 109 specifies the inter-sentence similarity, which is the similarity between the query sentence and the search target sentence, based on the inter-word similarity of each word pair. For example, the inter-sentence similarity specifying unit 109 uses the average value or the sum of the inter-word similarity as the inter-sentence similarity.

The search result output unit 110 outputs a search target sentence having a high degree of inter-sentence similarity with the query sentence as a search result. For example, the search result output unit 110 outputs a search target sentence having the highest inter-sentence similarity with the query sentence as a search result. In another embodiment, the search result output unit 110 may output, for example, a plurality of search target sentences whose inter-sentence similarity is equal to or greater than a predetermined threshold value, or a plurality of sentences arranged in descending order of inter-sentence similarity. The search target sentence of may be output. The search result is output, for example, by displaying it on a display, recording it on a storage medium, transmitting it to an external device, or the like.

<< Operation of search device >>
The conversion model generation unit 101 of the search device 100 generates a conversion model from a plurality of learning sentences in advance and records it in the conversion model storage unit 102 before executing the sentence search process.

FIG. 5 is a flowchart showing the operation of the search device according to the first embodiment.
The search device 100 accepts input of a query sentence from the user. The statement acquisition unit 104 acquires the input query statement (step S1). Next, the feature amount specifying unit 105 generates a feature vector for each of the plurality of words constituting the query sentence by using the conversion model stored in the conversion model storage unit 102 (step S2). Next, the grammar function specifying unit 106 parses the query sentence and specifies the grammar function of each of the plurality of words constituting the query sentence (step S3).

Next, the search device 100 selects the search target sentences to be stored by the sentence storage unit 103 one by one, and executes the processes of steps S5 to S11 shown below (step S4).
The sentence acquisition unit 104 acquires the search target sentence selected from the sentence storage unit 103 (step S5). Next, the feature quantity specifying unit 105 generates a feature vector for each of the plurality of words constituting the search target sentence by using the conversion model stored in the conversion model storage unit 102 (step S6). Next, the grammar function specifying unit 106 parses the search target sentence and specifies the grammatical function of each of the plurality of words constituting the search target sentence (step S7).

Next, the corresponding part specifying unit 107 is a pair of words having a common grammatical function based on the grammatical function of the word of the query sentence specified in step S3 and the grammatical function of the word of the search target sentence specified in step S7. Is specified (step S8). The word-to-word similarity specifying unit 108 specifies the word-to-word similarity for each pair specified by the corresponding location specifying unit 107 (step S9). Further, the inter-word similarity specifying unit 108 sets a predetermined penalty value as the inter-word similarity for words for which the corresponding location specifying unit 107 has not specified a pair (step S10). The inter-sentence similarity specifying unit 109 specifies the inter-sentence similarity between the query sentence and the search target sentence by calculating the average value of the inter-word similarity specified in steps S9 and S10 (step S11).

When the inter-sentence similarity is calculated for all the search target sentences stored in the sentence storage unit 103, the search result output unit 110 outputs the search target sentence having the highest inter-sentence similarity with the query sentence as a search result. (Step S12).

《Action / Effect》
As described above, the search device 100 according to the first embodiment uses the feature vectors of the words that make up the query sentence and the words that have the same grammatical function as the words that make up the query sentence among the words that make up the search target sentence. The inter-word similarity is specified by comparing with the feature vector, and the inter-sentence similarity is specified from each inter-word similarity. As a result, the search device 100 can search for similar sentences by taking into account the roles of the words appearing in the two sentences.

Here, a comparison between the query statement “A boy has a small dictionary” and the search target statement “A small boy has a dictionary” will be described as an example. In the comparison relating to the simple vector space model, the similarity is calculated without distinguishing between the word "small" in the query sentence and the word "small" in the search target sentence, so that the similarity is highly evaluated. On the other hand, in the search device 100 according to the first embodiment, as shown in FIG. 4, the word “small” in the query sentence and the word “small” in the search target sentence are evaluated separately because they have different grammatical functions. Will be done. As a result, the search device 100 according to the first embodiment can evaluate the inter-sentence similarity lower than that of the example of the simple vector space model. As described above, the search device 100 according to the first embodiment can calculate the similarity by distinguishing the words having the same notation but having different roles in the sentence.

Further, the search device 100 according to the first embodiment sets a predetermined penalty value for the inter-word similarity when there is no word in the other sentence that has the same grammatical function as the word constituting one sentence. As a result, the search device 100 can calculate the inter-sentence similarity to be low when comparing sentences having different structures.

Further, the feature vector according to the first embodiment is a vector having a number of dimensions smaller than the number of vocabularies. That is, the feature vector is a vector obtained by reducing the dimension of the word vector, which is a vector having the same number of dimensions as the number of vocabularies. As a result, the search device 100 can calculate the degree of similarity between words having different notations.

<Second embodiment>
The inter-sentence similarity specifying unit 109 of the search device 100 according to the first embodiment calculates the inter-sentence similarity based on the average value or the sum of the inter-sentence similarity of each word pair.
On the other hand, the importance of the meaning of a sentence differs depending on the words that make up the sentence. For example, the presence or absence of the article "a" may vary depending on whether the word is singular or plural, but the presence or absence of the article in a sentence often has no significant meaning. On the other hand, the presence or absence of the adverb "not" often has an important meaning in a sentence. Therefore, the search device 100 according to the second embodiment calculates the inter-sentence similarity based on the weighted average or the sum of loads of the inter-word similarity. Since the weighted average is obtained by dividing the weighted sum by the number of elements, the search device 100 can use the weighted average of the word-to-word similarity to specify the sentence-to-sentence similarity by calculating the weighted sum of the word-to-word similarity. Includes the inter-sentence similarity.

<< Configuration of search device >>
FIG. 6 is a schematic block diagram showing the configuration of the search device according to the second embodiment.
The search device 100 according to the second embodiment further includes a coefficient specifying unit 111 in addition to the configuration of the first embodiment.
The coefficient specifying unit 111 specifies a weighting coefficient according to the grammatical function of the word for each word pair specified by the corresponding location specifying unit 107. The weighting coefficient may be determined, for example, by the part of speech of the word, by the classification of the noun phrase and the verb phrase, or by the part of speech of the word involved.

The inter-sentence similarity specifying unit 109 multiplies each of the inter-word similarity of each word pair specified by the inter-word similarity specifying unit 108 by a weighting coefficient related to the word pair specified by the coefficient specifying unit 111. , Calculate the inter-sentence similarity by finding the average value.

As described above, according to the second embodiment, the search device 100 can specify the degree of similarity between the two sentences in consideration of the importance in the sentence for each word constituting the sentence.

<Other Embodiments>
Although one embodiment has been described in detail with reference to the drawings, the specific configuration is not limited to the above, and various design changes and the like can be made.
For example, the search device 100 according to the above-described embodiment identifies a pair of words having the same grammatical function, but the present invention is not limited to this. For example, the search device 100 according to another embodiment may ignore differences in word parts of speech and conjugations, or may specify word pairs whose grammatical functions partially match at a certain rate. Pairs whose grammatical function similarity is equal to or greater than a predetermined threshold may be specified. A pair of words with matching grammatical functions, a pair of words with partial matching grammatical functions, and a pair of words with similarities of grammatical functions above a predetermined threshold are examples of word pairs with common grammatical functions. Is.

Further, the search device 100 according to the above-described embodiment generates a feature vector and parses each search target sentence, but the present invention is not limited to this. For example, in the search device 100 according to another embodiment, the feature vector and the grammatical function of each word included in the sentence may be stored in advance in the sentence storage unit 103 in association with the search target sentence. That is, in another embodiment, the feature vector is generated and the syntax is analyzed for the query sentence, and the feature vector and the grammatical function of the search target sentence may be read from the sentence storage unit 103 at the time of the search.

Further, the search device 100 according to the above-described embodiment specifies the grammatical function of a word by using both the phrase structure analysis result and the dependency analysis result, but the present invention is not limited to this. For example, the search device 100 according to another embodiment may specify the grammatical function using only the phrase structure analysis result, or may specify the grammatical function using only the dependency analysis result.

Further, the search device 100 according to the above-described embodiment calculates the inter-word similarity for each word constituting the sentence, but is not limited to this in other embodiments. For example, the search device 100 according to another embodiment may calculate the inter-word similarity for each phrase constituting the sentence.

Further, in the above-described embodiment, the similarity specifying device is applied to the search device 100 that searches for a sentence similar to the query sentence from a plurality of sentences, but the other embodiments are not limited to this. For example, the similarity identification device according to another embodiment may calculate the similarity between two input sentences.

<Computer configuration>
FIG. 7 is a schematic block diagram showing the configuration of a computer according to at least one embodiment.
The computer 90 includes a CPU 91, a main storage device 92, an auxiliary storage device 93, and an interface 94.
The above-mentioned search device 100 is mounted on the computer 90. The operation of each processing unit described above is stored in the auxiliary storage device 93 in the form of a program. The CPU 91 reads a program from the auxiliary storage device 93, expands it to the main storage device 92, and executes the above processing according to the program.

Examples of the auxiliary storage device 93 include HDD (Hard Disk Drive), SSD (Solid State Drive), magnetic disk, optical magnetic disk, CD-ROM (Compact Disc Read Only Memory), and DVD-ROM (Digital Versatile Disc Read Only). Memory), semiconductor memory and the like. The auxiliary storage device 93 may be internal media directly connected to the bus of computer 90, or external media connected to computer 90 via an interface 94 or a communication line. When this program is distributed to the computer 90 via a communication line, the distributed computer 90 may expand the program to the main storage device 92 and execute the above processing. In at least one embodiment, the auxiliary storage device 93 is a non-temporary tangible storage medium.

Further, the program may be for realizing a part of the above-mentioned functions. Further, the program may be a so-called difference file (difference program) that realizes the above-mentioned function in combination with another program already stored in the auxiliary storage device 93.

100 Search device 101 Conversion model generation unit 102 Conversion model storage unit 103 Sentence storage unit 104 Sentence acquisition unit 105 Feature quantity specification unit 106 Grammar function specification unit 107 Corresponding part specification unit 108 Word-to-word similarity specification unit 109 Sentence-to-sentence similarity specification unit 110 Search result output unit 111 Coefficient identification unit

Claims (6)

A feature quantity specifying unit that specifies a feature quantity for each word that constitutes the first sentence and the second sentence,
A role specifying unit that specifies the grammatical role of each word that constitutes the first sentence and the second sentence, and
The degree of similarity between the feature quantity of the words constituting the first sentence and the feature quantity of the words constituting the second sentence and having a role common to the words constituting the first sentence. A word-to-word similarity specification part that specifies a certain word-to-word similarity,
It is provided with an inter-sentence similarity specifying unit that specifies the inter-sentence similarity, which is the similarity between the first sentence and the second sentence based on the inter-word similarity for each word.
When the second sentence does not have a word having the same role as the word constituting the first sentence, the inter-word similarity specifying unit has inter-word similarity related to the word constituting the first sentence. A similarity identification device that sets the degree to a predetermined penalty value. The similarity degree according to claim 1, wherein the sentence-to-sentence similarity specifying unit specifies the sentence-to-sentence similarity by calculating the weight sum of the word-to-word similarity using a weighting coefficient according to the role of the word. Specific device. The similarity identification device according to claim 1 or 2, wherein the feature quantity is a vector having a number of dimensions smaller than the number of vocabularies. The similarity specifying device according to any one of claims 1 to 3, wherein the role specifying unit specifies a role for each word based on the result of parsing processing. A step in which the computer identifies the features for each word that composes the first sentence and the second sentence.
A step in which the computer identifies the grammatical role of each word constituting the first sentence and the second sentence.
The computer has a feature amount of a word that constitutes the first sentence and a feature amount of a word that has the same role as the word that constitutes the first sentence among the words that constitute the second sentence. Steps to identify word-to-word similarity, which is the degree of similarity,
When the computer does not have a word having the same role as the word constituting the first sentence in the second sentence, a predetermined penalty is given to the inter-word similarity related to the word constituting the first sentence. Steps to value and
A method for specifying similarity, which comprises a step in which the computer specifies inter-sentence similarity, which is the degree of similarity between the first sentence and the second sentence, based on the inter-word similarity for each word. On the computer
Steps to identify features for each word that makes up the first and second sentences,
A step of identifying the grammatical role of each word constituting the first sentence and the second sentence, and
It is the degree of similarity between the feature amount of the word constituting the first sentence and the feature amount related to the word having the same role as the word constituting the first sentence among the words constituting the second sentence. Steps to identify word-to-word similarity and
When the second sentence does not have a word having the same role as the word constituting the first sentence, the step of setting the inter-word similarity related to the word constituting the first sentence to a predetermined penalty value. When,
A program for executing a step of specifying the inter-sentence similarity, which is the similarity between the first sentence and the second sentence, based on the inter-word similarity for each word.

Images