JP2019082931A - Retrieval device, similarity calculation method, and program - Google Patents

Abstract

To calculate the degree of similarity taking the roles of words appearing in two sentences into consideration.SOLUTION: A feature quantity specification unit specifies a feature quantity for each of words constituting first and second sentences. A role specification unit specifies a grammatical role of each of the words constituting the first and second sentences. An inter-word similarity specification unit specifies an inter-word similarity that is the similarity between the feature quantity of words constituting the first sentence and a feature quantity relating to words having a role common to the words constituting the first sentence among the words constituting the second sentence. An inter-sentence similarity specification unit specifies an inter-sentence similarity that is the similarity between the first and second sentence on the basis of the inter-word similarities between words.SELECTED DRAWING: Figure 4

Description

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

  In Patent Document 1, two sentences are respectively converted into feature vectors to calculate the degree of similarity, and then, by using syntactic analysis, the difference between the two sentences is extracted by excluding words having the same dependency relation. Technology is disclosed.

Patent No. 5367099 gazette

When comparing the two sentences, if 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 between the two sentences. Become. However, words that are similar to one another 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", in the first sentence, the word "small" modifies "dictionary" 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, a similarity calculation method, and a program that calculate similarity by taking into consideration the role of words appearing in two sentences.

  According to the first aspect of the present invention, the similarity specifying device (100) comprises: a feature identifying unit (105) for identifying a feature according to words constituting the first sentence and the second sentence; A role specification unit (106) for specifying a grammatical role of each of the words constituting the first sentence and the second sentence; feature quantities of the words constituting the first sentence; and the second sentence An inter-word similarity specifying unit (108) for specifying an inter-word similarity, which is a similarity between feature words related to a word having a role common to the word constituting the first sentence among words constituting the word; An inter-situment similarity specifying unit (109) for specifying an inter-sentence similarity which is a similarity between the first sentence and the second sentence based on the inter-word similarity according to the words. Thereby, the similarity specifying device can specify the similarity in consideration of the role of the words appearing in the two sentences.

  According to a second aspect of the present invention, in the similarity specifying device according to the first aspect, the inter-word similarity specifying unit is configured to calculate the weight of the inter-word similarity using a weighting factor according to the role of the word. Inter-sentence similarity may be specified by calculating the sum. Thus, the similarity specifying device can specify the similarity between the two sentences in consideration of the importance of each word constituting the sentence.

  According to a third aspect of the present invention, in the similarity specifying device according to the first or second aspect, the inter-word similarity specifying unit has a word having the same role as the word constituting the first sentence. In the case where the second sentence is not present, the inter-word similarity may be specified to a predetermined penalty value. Thus, the similarity specifying device can calculate the similarity between sentences low in 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. Thus, the similarity specifying device can calculate the similarity between words having different notations.

  According to a 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 specifies a role according to the word by phrase structure analysis processing. You may

  According to a sixth aspect of the present invention, there is provided a similarity degree specifying method comprising the steps of: identifying a feature amount according to words constituting a first sentence and a second sentence; and the first sentence and the second sentence Identifying a grammatical role of each of the words constituting the word, feature quantities of the words constituting the first sentence, and the first sentence among the words constituting the second sentence Identifying an inter-word similarity, which is a similarity between the word and the feature associated with the word having the same role, and the first sentence and the second sentence based on the inter-word similarity for each word And (d) identifying an inter-sentence similarity that is a similarity to

  According to a seventh aspect of the present invention, a program includes the steps of: specifying, on a computer, a feature amount for each word constituting a first sentence and a second sentence, the first sentence and the second sentence Identifying the grammatical role of each word according to the words constituting the word, the feature amount of the words constituting the first sentence, and the first sentence among the words constituting the second sentence Specifying an inter-word similarity, which is a similarity between the word forming the word and a feature associated with the word having the same role, and based on the inter-word similarity between the words, the first sentence and the Specifying an inter-sentence similarity that is a similarity to two sentences.

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

It is a schematic block diagram showing composition of a search device concerning a 1st embodiment. It is a figure which shows the example of the identification method of the grammar function by phrase structure analysis. It is a figure which shows the example of the identification method of the grammar function by dependency analysis. It is a figure which shows the example of the identification method of the corresponding part of two sentences. It is a flowchart which shows operation | movement of the search device which concerns on 1st Embodiment. It is a schematic block diagram which shows the structure of the search device which concerns on 2nd Embodiment. It is a schematic block diagram showing composition of a computer concerning at least one embodiment.

First Embodiment
<< Configuration of Search Device >>
FIG. 1 is a schematic block diagram showing the configuration of a search device according to the first embodiment.
The search device 100 according to the first embodiment receives an input of a sentence, and searches a plurality of sentences that are similar to the input sentence. The search device 100 is an example of a similarity specifying device that specifies the similarity of 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 specification unit 105, a grammar function specification unit 106, a corresponding part specification unit 107, inter-word similarity specification A section 108, an inter-sentence similarity specification section 109, and a search result output section 110 are provided.

  The conversion model generation unit 101 generates a sentence vector having the dimensionality of the vocabulary number from each of the plurality of learning sentences, and a conversion model for generating a feature vector representing the feature of the word based on the sentence vector. Generate The words are not limited to words but may include compound words and phrases. The learning sentence may be a search target sentence stored in the sentence storage unit 103 or may be another sentence. When a search target sentence is used as a 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 conversion model generation unit 101 can generate a conversion model using an auto encoder. The auto encoder is a neural network in which the number of nodes in the input layer and the number in the output layer are equal and the number of nodes in the middle layer is smaller than that in the input layer and the output layer. The conversion model generation unit 101 inputs each sentence vector to the auto encoder, and learns the auto encoder so that the input and the output become equal. Then, the conversion model generation unit 101 generates a conversion model having the output of the intermediate layer as a feature vector by extracting the input layer and the intermediate layer of the auto encoder. For example, Word2Vec (https://code.google.com/archive/p/word2vec/) can be used to create a conversion model using an auto encoder.
Further, the conversion model generation unit 101 may create a conversion model based on tf (Term Frequency) -idf (Inverse Document Frequency), latent semantic analysis, and principal component analysis.

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

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

  The feature amount specifying unit 105 generates a feature vector using the conversion model stored in the conversion model storage unit 102 for each of a plurality of words constituting the sentence acquired by the sentence acquisition unit 104. For example, the feature amount specifying unit 105 can generate a feature vector by the following process. The feature amount identifying unit 105 divides the sentence acquired by the sentence acquiring 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 dimension number equal to the vocabulary number from each of the divided words. The feature amount specifying unit 105 obtains a feature vector by inputting a word vector to the conversion model.

  The grammar function specification unit 106 parses the sentence acquired by the sentence acquisition unit 104, and specifies a grammar function that is the role of each of a plurality of words constituting the sentence. The grammatical function is a classification based on the relation of elements (eg, words, phrases, clauses) constituting a sentence to other elements. Examples of grammatical functions include noun phrases (NP), verb phrases (VP: verb phrase) and adjective phrases (phrase structure), subjects, predicates, objects, complements and modifiers (dependencies) Relationship), classification of parts of speech, etc. Examples of syntactic analysis include phrase structure analysis and dependency analysis. The grammar function specification unit 106 is an example of a role specification unit. In another embodiment, 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 grammar function by phrase structure analysis.
For example, the grammar function identification unit 106 analyzes the phrase structure of the sentence acquired by the sentence acquisition unit 104, and generates a phrase structure tree having a plurality of words constituting the sentence as elements. The phrase structure tree is tree structure data having a sentence as a root node, a clause or a phrase as an internal node, and a plurality of words as leaf nodes. By the phrase structure analysis, each node is tagged with a grammar 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, a permutation of all tags attached to a route connecting the leaf node and the root node as a grammar function of a word related to the leaf node.

FIG. 3 is a diagram showing an example of a method of specifying a grammar function by dependency analysis.
Further, for example, the grammar function identification unit 106 performs dependency analysis on the sentence acquired by the sentence acquisition unit 104, and generates a dependency 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 a node of a dependency source word (depender) is a child node of a dependency destination node (dependee). In the dependency structure tree, the branches connecting nodes may be tagged with a dependency relation. The grammar function specification unit 106 specifies, as a grammar function, a dependency destination of each word (for example, a grammar function of a dependency destination identified by phrase structure analysis) or a dependency relationship based on the dependency structure tree. Examples of the dependency relationship include det indicating assignment of a determinate, nsubj indicating a subject related to a predicate, dobj indicating an object related to a predicate, and amod indicating a modifier modifying a noun.

FIG. 4 is a diagram showing an example of a method of specifying corresponding portions of two sentences.
The corresponding part specifying unit 107 specifies a pair of words of the query sentence and words of the search target sentence, which have a common grammatical function. That is, for each word of the query sentence, the corresponding part specifying unit 107 specifies a word having a common grammatical function with the word from the search target sentence, thereby specifying the word pair. For example, as shown in FIG. 4, when the query statement is “A boy has a small dictionary” and the search target statement is “A small boy has a dictionary”, the corresponding part identifying unit 107 Identify word pairs. The word “A” which is the word of 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”. The corresponding part specifying unit 107 specifies, from the search target sentence, a word having a grammatical function that the phrase structure is “S-NP-DT” and the dependency relation is “det”. As shown in FIG. 4, the query sentence "A" has a grammatical function in common with the search target sentence "A". The corresponding part specifying unit 107 executes this for each word, and as shown in FIG. 4, a pair of “A” in the query sentence and “A” in the search target sentence, “boy” in the query sentence and the search target sentence Pair of "boy", pair of "has" of query sentence and search target sentence "has", pair of query sentence "a" and search target sentence "a", query sentence "dictionary" and search target sentence The "dictionary" pairs of are extracted as word pairs respectively. Further, as shown in FIG. 4, the corresponding part specifying unit 107 does not have a word having a grammatical function in common with “small” in the query sentence in the search target sentence, and has a common grammatical function with “small” in the search target sentence. Specify that the target word is not present in the query statement.

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

  The inter-statement similarity specification unit 109 specifies inter-state 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 sets the average value or the sum of inter-word similarity as the inter-sentence similarity.

  The search result output unit 110 outputs a search target sentence having a high degree of similarity between the query sentence and the sentence as a search result. For example, the search result output unit 110 outputs a search target sentence having the highest degree of similarity between the query sentence and the 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-statement similarity is equal to or greater than a predetermined threshold, or a plurality of search target sentences arranged in descending order of inter-state similarity. The search target sentence of may be output. The output of the search result is performed, for example, by display on a display, recording on a storage medium, transmission to an external device, or the like.

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

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

Next, the search device 100 selects the search target sentences stored in the sentence storage unit 103 one by one, and executes the processing from step S5 to step S11 described 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 amount specifying unit 105 generates a feature vector using the conversion model stored in the conversion model storage unit 102 for each of a plurality of words constituting the search target sentence (step S6). Next, the grammar function identification unit 106 parses the search target sentence, and specifies the grammar function of each of a plurality of words constituting the search target sentence (step S7).

  Next, 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, the corresponding part specification unit 107 pairs words having common grammatical functions. Are identified (step S8). The inter-word similarity identification unit 108 identifies inter-word similarity for each pair identified by the corresponding part identification unit 107 (step S9). The inter-word similarity specifying unit 108 sets a predetermined penalty value to the inter-word similarity as to the word for which the corresponding part specifying unit 107 has not specified a pair (step S10). The inter-statement similarity specifying unit 109 specifies an inter-state similarity between the query sentence and the search target sentence by calculating the average value of the inter-word similarity specified in step S9 and step S10 (step S11).

  When the inter-statement similarity is calculated for all 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-state similarity with the query sentence as the search result. (Step S12).

<< Operation / Effect >>
As described above, the search device 100 according to the first embodiment uses the feature vector of the word forming the query sentence and the word having the grammatical function common to the word forming the query sentence among the words forming the search target sentence. Inter-word similarity is identified by comparing with such a feature vector, and inter-sentence similarity is identified from each inter-word similarity. Thus, the search device 100 can search for similar sentences in consideration of the role of words appearing in two sentences.

  Here, the comparison between the query sentence “A boy has a small dictionary” and the search target sentence “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 evaluated high because the similarity is calculated without distinction between the word “small” of the query sentence and the word “small” of the search target sentence. On the other hand, as shown in FIG. 4, the search device 100 according to the first embodiment distinguishes and evaluates the word "small" of the query sentence and the word "small" of the search target sentence because the grammatical functions are different. Be done. Thereby, the search device 100 according to the first embodiment can evaluate the similarity between sentences lower than 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 in the same written word if the roles in the sentences are different.

  In addition, the search device 100 according to the first embodiment sets a predetermined penalty value as the inter-word similarity, when there is no word in the other sentence in which the word constituting the one sentence and the grammatical function are common. Accordingly, the search device 100 can calculate the inter-sentence similarity low in comparison between sentences having different structures.

  Also, the feature vector according to the first embodiment is a vector with a dimension number smaller than the number of words. That is, the feature vector is a vector obtained by reducing the dimension of the word vector, which is a vector of the number of dimensions equal to the vocabulary number. Thus, the search device 100 can calculate the degree of similarity between words having different notations.

Second Embodiment
The inter-text similarity specifying unit 109 of the search device 100 according to the first embodiment calculates the inter-text similarity based on the average value or the sum of the inter-word similarity of each pair of words.
On the other hand, the importance of the meaning of a sentence is different for each word constituting the sentence. For example, the presence or absence of the article "a" may change depending on whether the word is a singular form or a plural form, 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 sentences. Therefore, the search device 100 according to the second embodiment calculates the inter-sentence similarity based on the weighted average or the weighted sum of inter-word similarity. Note that, since the weighted average is obtained by dividing the weighted sum by the number of elements, specifying the inter-symbol similarity by calculating the weighted sum of the inter-word similarity means that the search device 100 calculates the weighted average of inter-word similarity. Including making inter-statement similarity.

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

  The inter-speech similarity specifying unit 109 multiplies each of the inter-word similarity of each pair of words specified by the inter-word similarity specifying unit 108 by the weighting factor relating to the word pair specified by the coefficient specifying unit 111. By calculating an average value, inter-statement similarity is calculated.

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

Other Embodiments
As mentioned above, although one embodiment was described in detail with reference to drawings, a concrete configuration is not restricted to the above-mentioned thing, It is possible to do various design changes etc.
For example, although search device 100 concerning an embodiment mentioned above specifies a word pair in which a grammar function corresponds, it is not restricted to this. For example, the search device 100 according to another embodiment may ignore the difference in part of speech or inflected form of a word, or may specify a pair of words that partially match in grammatical function, It is also possible to specify a pair whose grammatical function similarity is equal to or higher than a predetermined threshold. It should be noted that a pair of words whose grammatical functions match, a pair of words whose partial grammatical functions match, and a pair of words whose grammatical function similarity is greater than or equal to a predetermined threshold are all examples of word pairs having common grammatical functions. It is.

  Moreover, although the search device 100 which concerns on embodiment mentioned above performs the production | generation and syntactic analysis of a feature vector about each search object sentence, it is not restricted to this. For example, the search device 100 according to another embodiment may store the feature vector of each word included in the sentence and the grammatical function in advance in the sentence storage unit 103 in association with the search target sentence. That is, in another embodiment, a feature vector may be generated and syntactic analysis is performed on 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.

  Moreover, although the search device 100 which concerns on embodiment mentioned above specifies the grammatical function of a word using both a phrase structure analysis result and a dependency analysis result, it is not restricted to this. For example, the search device 100 according to another embodiment may specify the grammar function using only the phrase structure analysis result, or may specify the grammar function using only the dependency analysis result.

  In addition, although the search device 100 according to the above-described embodiment calculates the inter-word similarity for each word constituting a sentence, the present invention is not limited to this. For example, the search device 100 according to another embodiment may calculate the inter-word similarity for each phrase constituting a sentence.

  In the above-described embodiment, the similarity specifying device is applied to the search device 100 that searches for a sentence similar to a query sentence among a plurality of sentences, but the embodiment is not limited to this. For example, the similarity specifying 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 92, an auxiliary storage 93, and an interface 94.
The above-described search device 100 is implemented in 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 in the main storage device 92, and executes the above processing according to the program.

  Examples of the auxiliary storage device 93 include a hard disk drive (HDD), a solid state drive (SSD), a magnetic disk, a magneto-optical disk, a compact disc read only memory (CD-ROM), and a digital versatile disc read only (DVD-ROM) Memory, semiconductor memory, and the like. The auxiliary storage device 93 may be internal media directly connected to the bus of the computer 90, or may be external media connected to the computer 90 via the interface 94 or a communication line. When the program is distributed to the computer 90 by a communication line, the computer 90 that has received the distribution may deploy the program in the main storage device 92 and execute the above processing. In at least one embodiment, secondary storage 93 is a non-transitory tangible storage medium.

  Further, the program may be for realizing a part of the functions described above. Furthermore, the program may be a so-called difference file (difference program) that realizes the above-described function in combination with other programs 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 correspondence location specification unit 108 inter-word similarity specification unit 109 inter-sentence similarity specification unit 110 Search result output unit 111 Coefficient identification unit

Claims (7)

A feature amount identification unit that identifies a feature amount according to words constituting the first sentence and the second sentence;
A role identification unit that identifies grammatical roles of the words that make up the first sentence and the second sentence;
The similarity between the feature quantity of the word that constitutes the first sentence and the feature quantity according to the word of the role in common with the word that constitutes the first sentence among the words that constitute the second sentence An inter-word similarity specifying unit for specifying an inter-word similarity;
An inter-sentence similarity specification unit which specifies inter-sentence similarity, which is the similarity between the first sentence and the second sentence, based on the inter-word similarity according to the words. The inter-sentence similarity specifying unit specifies an inter-sentence similarity by calculating a weighted sum of the inter-word similarity using a weighting factor according to the role of the words. Specific device. The inter-word similarity specifying unit specifies an inter-word similarity to a predetermined penalty value when the second sentence does not have a word having the same role as that of the words constituting the first sentence. Or the similarity identification apparatus of Claim 2. The similarity specifying device according to any one of claims 1 to 3, wherein the feature amount is a vector of a dimension number smaller than a vocabulary number. The similarity specifying device according to any one of claims 1 to 4, wherein the role specifying unit specifies a role for each of the words based on a syntactic analysis processing result. Identifying a feature amount according to words constituting the first sentence and the second sentence;
Identifying a grammatical role for each of the words that make up the first sentence and the second sentence;
The similarity between the feature of the word constituting the first sentence and the feature of the word having the same role as the word constituting the first sentence among the words constituting the second sentence Identifying inter-word similarity;
Identifying an inter-sentence similarity that is the similarity between the first sentence and the second sentence based on the inter-word similarity according to the words. On the computer
Identifying a feature amount according to words constituting the first sentence and the second sentence;
Identifying a grammatical role for each of the words that make up the first sentence and the second sentence;
The similarity between the feature of the word constituting the first sentence and the feature of the word having the same role as the word constituting the first sentence among the words constituting the second sentence Identifying inter-word similarity;
A program for causing an inter-sentence similarity, which is a similarity between the first sentence and the second sentence, to be identified based on the inter-word similarity according to the words.

Images