JP6817556B2 - Similar sentence generation method, similar sentence generation program, similar sentence generator and similar sentence generation system - Google Patents

Description

The present disclosure relates to a similar sentence generation method for generating a similar sentence from an original sentence, a similar sentence generation program, a similar sentence generation device, and a similar sentence generation system including the similar sentence generation device.

In recent years, machine translation that translates a sentence in the first language into a sentence in a second language different from the first language has been researched and developed, and in order to improve the performance of such machine translation, many example sentences that can be used for translation have been studied and developed. You will need a bilingual corpus that collects. Therefore, one or a plurality of similar sentences (paraphrase sentences) similar to the original sentence are generated from one original sentence.

For example, in Patent Document 1, in order to transform a sentence in a predetermined pattern and determine whether or not it is a suitable transformation, an evaluation value is calculated using an evaluation function, and a unified language conversion process is selected to select the expression having the highest evaluation value. The system is disclosed.

Further, Patent Document 2 discloses a natural language processing method in which points related to activity are set as morphemes, the points are increased or decreased, and information is extracted from text based on the increased or decreased points.

Further, in Patent Document 3, a new post-paraphrase example is generated based on the pre-paraphrase example and the post-paraphrase example specified by the user, and the paraphrase sentence created by applying the difference to the analyzed sentence is output. Document processing equipment is disclosed.

Japanese Patent No. 3932350 Japanese Unexamined Patent Publication No. 2005-339043 Japanese Patent No. 5060539

However, in order to improve the performance of machine translation, it is preferable that there are more example sentences available for translation, and further improvement is required to generate similar sentences that can be used as example sentences.

The present disclosure solves the above-mentioned conventional problems, and can reduce the search cost for the database of the language model, and can identify similar sentences with high accuracy. Similar sentence generation method, similar sentence generation program, similar sentence. It is an object of the present invention to provide a generator and a similar sentence generation system.

The uniform method of the present disclosure is a method of generating a similar sentence from the original sentence, in which the first sentence is input and one or more of the plurality of words constituting the first sentence have the same meaning as the first word. The second word / phrase of is extracted from the first database, the first database associates the word with a synonym of the word / phrase contained in the first database, and the one or more second words obtained based on the second database. The N-gram value is calculated based on the context-dependent value corresponding to the phrase, and the second database associates the phrase with the context-dependent value corresponding to the phrase contained in the second database, and the context-dependent value. Indicates the degree to which the meaning of the phrase contained in the second database depends on the context, and one or more second sentences in which the first phrase is replaced with the one or more second phrase in the first sentence. In, one or more consecutive third words and phrases including the second word and phrase corresponding to the N-gram value are extracted, and the appearance frequency of the one or more third words and phrases in the third database is calculated. The third database associates words and phrases with the appearance frequency of words and phrases contained in the third database in the third database, determines whether the calculated appearance frequency is equal to or higher than the threshold value, and the calculated appearance frequency is calculated. When it is determined that the threshold value is equal to or higher than the threshold value, the one or more second sentences are adopted as similar sentences to the first sentence and output to an external device.

According to the present disclosure, the search cost for the database of the language model can be reduced, and similar sentences can be identified with high accuracy.

It is a block diagram which shows an example of the structure of the similar sentence generator in Embodiment 1 of this disclosure. It is a figure which shows an example of the data structure of the replacement candidate dictionary shown in FIG. It is a figure which shows an example of the data structure of the context dependency rate dictionary shown in FIG. It is a figure which shows an example of the data structure of the language model database shown in FIG. It is a flowchart which shows an example of the similar sentence generation processing by the similar sentence generation apparatus shown in FIG. It is a block diagram which shows an example of the structure of the similar sentence generation system in Embodiment 2 of this disclosure. It is a flowchart which shows an example of the similar sentence generation processing including the feedback data update process of the similar sentence generation system shown in FIG.

(Knowledge on which this disclosure was based)
As described above, in order to improve the performance of machine translation, it is preferable that there are more example sentences available for translation, and it is desired to automatically expand the amount of sentences based on a small amount of parallel translation corpus by generating similar sentences using phrase replacement. ing. When generating a similar sentence using this phrase replacement, there are cases where the quality of the replacement depends on the context in the selection of similar candidate sentences including the expression (phrase) to be replaced.

For this reason, we would like to enable learning and reflection of cases while considering context dependence by dynamically selecting replacement rules based on the language model, but how similar is how to efficiently select similar candidate sentences. It is important to select the candidate sentences.

For example, when expanding the bilingual corpus and generating similar candidate sentences by replacement (paraphrase), as paraphrase rules, "I can't speak" is (1) "I can't speak", (2) "I can't speak", (3). When replaced with either "secret", applying the above paraphrase rule to the sentence "I can't speak English", "I can't speak English", "I can't speak English", and "I can't speak English" Three similar candidate sentences of "It's a secret" are generated.

In this case, from the context, "I can't speak English" and "I can't speak English" can be adopted as similar sentences, but "English is a secret" is not an appropriate expression for Japanese, so it is adopted as a similar sentence. It cannot be done and will be rejected. In this way, even if the same paraphrase rule is applied, there are cases where similar candidate sentences can be adopted as similar sentences and cases where they cannot be adopted, depending on the context.

Conventional methods for distinguishing between adopted sentences that can be adopted as similar sentences and rejected sentences that cannot be adopted are similarities in distributed expression models using word vectors and sentence vectors, and language models (for example, N-gram language model). Judgment was made based on the frequency of appearance of. Specifically, by increasing the identification target area (search range) of the language model (for example, increasing N of N-gram), it is determined whether or not it exists as an expression, and a paraphrase rule that depends on the context. It was decided to adopt and reject the (replacement rule).

Evaluation was also performed by modeling the fluency of sentences using a language model. For example, let the language model be the N-gram language model, increase the score of translations and phrases that use expressions that are more contained in the database of the N-gram language model, and lower the score of those that are not included much. There is a method called. By applying this method, the score of similar candidate sentences is calculated, and "good sentences" (adopted sentences that can be adopted as similar sentences) or "bad sentences" (rejected sentences that cannot be adopted as similar sentences) are identified by threshold processing. Was.

However, when the identification target area is increased, the data amount and the calculation amount increase and the data distribution becomes sparse. Therefore, in order to search all the replacement candidates from the large identification target area, the data amount and the calculation amount increase. To do. For example, 2-gram has about 80 million entries, but 5-gram has about 800 million entries, and if N of N-gram is increased, there is a problem that the amount of data and the amount of calculation increase dramatically. ..

In order to solve the above problem, in the present disclosure, for example, a context-dependent value storage unit that stores a plurality of replacement candidate character strings in association with a context-dependent value indicating the degree to which the replacement candidate character string depends on the context is provided. Whether to refer to a language model that includes words before and after the replacement candidate character string in a similar sentence whose replacement quality changes depending on the context, depending on whether the replacement quality changes depending on the context. Decide whether or not.

That is, the search range (identification target area) for the language model database is determined according to the context-dependent value, and the language model database is searched using the determined search range, so that the context-dependent value is considered to be high. Only the candidate character string is identified in a larger search area, and the replacement candidate character string having a low context-dependent value is identified in a small search area in order to balance the search cost and the identification accuracy.

In addition, in the conventional method of generating similar sentences, distributed expressions and expressions that are not included in the language model cannot be identified in the first place and are rejected. For example, if there is no sentence containing the phrase "it is a secret" in the training data, there is a problem that similar candidate sentences including "it is a secret" cannot be identified and are rejected.

In order to solve the above problems, in the present disclosure, for example, when a context-sensitive replacement candidate character string is input by an external input (for example, feedback from a user or a predetermined device), a language model is used. Update the database and context-sensitive value storage section of. In addition, when a new sentence expression is input, the context-dependent value of the word in the context-dependent value storage unit is changed according to the expression, and an N-gram or the like including the new sentence expression is partially constructed. Reflect the new sentence expression in the language model. By adding the correct data in this way, the frequency of appearance of the language model including the words before and after the replacement character string is adjusted, and the context-dependent value storage unit itself is updated according to the external input.

As described above, the identification accuracy can be improved by feeding back external knowledge and new knowledge and updating the database of the language model. As a result, similar candidate sentences can be identified at low cost and with high accuracy, and highly efficient and autonomous similar candidate sentences can be identified by updating and responding to expressions that do not exist in the N-gram model database. It can be carried out.

Based on the above findings, the inventors of the present application have completed the present disclosure as a result of diligent studies on how to generate a similar sentence from the original text.

The method according to one aspect of the present disclosure is a method of generating a similar sentence from the original sentence, in which the first sentence is input and has the same meaning as the first word / phrase among the plurality of words / phrases constituting the first sentence. The above second words and phrases are extracted from the first database, the first database associates the words and phrases with synonyms of the words and phrases contained in the first database, and the one or more first words obtained based on the second database. The N-gram value is calculated based on the context-dependent value corresponding to the two words, and the second database associates the word with the context-dependent value corresponding to the phrase contained in the second database, and the context-dependent The value indicates the degree to which the meaning of the phrase contained in the second database depends on the context, and the first phrase is replaced with the one or more second phrase in the first sentence. In the sentence, one or more consecutive third words including the second word corresponding to the N-gram value are extracted, and the appearance frequency of the one or more third words in the third database is calculated. , The third database associates words and phrases with the frequency of appearance of words and phrases contained in the third database in the third database, determines whether the calculated frequency of appearance is equal to or greater than the threshold value, and determines the frequency of appearance calculated. When is determined to be equal to or greater than the threshold value, the second sentence of one or more is adopted as a similar sentence of the first sentence and output to an external device.

With such a configuration, the first sentence is input, and one or more second words having the same meaning as the first word among the plurality of words constituting the first sentence are extracted from the first database, and the first database is used. The N-gram value is calculated based on the context-dependent values corresponding to one or more second words obtained based on the second database by associating the words with the synonyms of the words contained in the first database. The 2nd database associates words with the context-dependent values corresponding to the words contained in the 2nd database, and the context-dependent values indicate the degree to which the meanings of the words contained in the 2nd database depend on the context. In one or more second sentences in which the first word is replaced with one or more second words in one sentence, one or more consecutive third words including the number of second words corresponding to the N-gram value are extracted. Then, for one or more third words, the frequency of appearance in the third database is calculated, and the third database associates the words with the frequency of appearance of the words included in the third database in the third database, and the calculated frequency of appearance is calculated. If it is determined whether the sentence is equal to or higher than the threshold and the calculated frequency of appearance is higher than the threshold, one or more second sentences are adopted as similar sentences to the first sentence and output to an external device. , Only the second phrase with a high context-dependent value can be identified in a large search area, and the second phrase with a low context-dependent value can be identified in a small search area, with respect to the third database, which is a database of language models. The search cost can be reduced, and similar sentences can be identified with high accuracy.

The first sentence is written in a first language, the first sentence is included in a bilingual corpus, and the bilingual corpus has a plurality of pairs of a sentence written in the first language and a bilingual sentence written in a second language. If it is determined that the calculated frequency of appearance is equal to or higher than the threshold value, the one or more second sentences may be added to the bilingual corpus as similar sentences to the first sentence.

With such a configuration, similar sentences can be added to the bilingual corpus.

The third database includes a database of the N-gram language model, determines N of the N-gram language model to i (a positive integer) according to the context-dependent value, and collates the third database. Thereby, the appearance frequency of the i-gram including the second phrase is obtained, and based on the appearance frequency of the i-gram including the second phrase, the one or more second sentences are regarded as similar sentences to the first sentence. It may be decided whether or not to adopt it.

With such a configuration, N of the N-gram language model is determined to be i (a positive integer) according to the context-sensitive value, and by collating the database of the N-gram language model, i- including the second phrase is included. Since the frequency of appearance of the gram is obtained and it is determined whether or not one or more second sentences are adopted as similar sentences to the first sentence based on the obtained frequency of appearance, the larger the context-sensitive value, the larger i. By setting i smaller as the context-sensitive value is smaller, the appearance of an i-gram containing the second phrase having a large context-sensitive value by using a wide identification target area for the second phrase having a high context-sensitive value. The frequency can be calculated with high accuracy, and the frequency of occurrence of i-gram including the second phrase with a small context-sensitive value can be reduced at low cost by using a narrow identification target area for the second phrase with low context-sensitiveness. It can be obtained with high accuracy, and similar sentences can be identified efficiently and with high accuracy.

Generated based on the one or more second sentences determined to be adopted as similar sentences to the first sentence and the translated sentence obtained by translating the first sentence that generated the one or more second sentences in a second language. Using the translated translation model, a predetermined translation target sentence is translated to create a translation result sentence, the translation result sentence is evaluated, and based on the evaluation result of the translation result sentence, the language of the translation target sentence and / Or, feedback information including linguistic information regarding the language of the translation result sentence and evaluation information for the linguistic information may be generated.

With such a configuration, a translation model generated based on one or more second sentences determined to be adopted and a translated sentence obtained by translating the first sentence that generated one or more second sentences in a second language. Use to translate a predetermined translation target sentence to create a translation result sentence, evaluate the created translation result sentence, and based on the evaluation result of this translation result sentence, the language of the translation target sentence and / or the translation result sentence. Since the feedback information including the linguistic information about the language and the evaluation information for this linguistic information is generated, it is possible to autonomously generate the feedback information for learning and reflecting the case considering the context dependence. ..

At least one of the first database, the second database, and the third database may be updated using the feedback information.

With such a configuration, at least one of the first database, the second database, and the third database is updated by using the feedback information including the linguistic information and the evaluation information. Therefore, a case in which context dependence is considered. Can be reflected in at least one of the first database, the second database, and the third database, and is highly efficient and can handle expressions that do not exist in the first database, the second database, and the third database before the update. It is possible to autonomously identify similar sentences.

If the feedback information contains the second phrase that is context sensitive, the second database and the third database may be updated.

With such a configuration, when the feedback information contains a second word having a context dependency, the second database and the third database are updated. Therefore, the second database and the third database are examples in which the context dependency is taken into consideration. It is possible to identify similar sentences with high efficiency and autonomously in consideration of context dependence.

When the feedback information includes a new sentence expression, the context-dependent value of the second database may be changed according to the sentence expression.

With such a configuration, when the feedback information includes a new sentence expression, the context-dependent value of the second database is changed according to the new sentence expression, so that it is highly efficient and autonomous enough to support the new sentence expression. It is possible to identify similar sentences.

If the feedback information includes a new sentence expression, the third database may be updated to include the sentence expression.

With such a configuration, when the feedback information includes a new sentence expression, the third database is updated to include the new sentence expression, so that it is possible to cope with a new sentence expression that does not exist in the third database before the update. Efficient and autonomous identification of similar sentences can be performed.

Further, the present disclosure can be realized not only as a similar sentence generation method for executing the above-mentioned characteristic processing, but also a computer for causing a computer to execute the characteristic processing included in such a similar sentence generation method. It can also be realized as a program. It can also be realized as a similar sentence generation device having a characteristic configuration corresponding to a characteristic process executed by the similar sentence generation method. Therefore, the same effect as the above-mentioned similar sentence generation method can be obtained in the following other aspects as well.

The program according to another aspect of the present disclosure is a program for operating a database as a device for generating a similar sentence from the original sentence, and the first sentence is input to the computer to form the first sentence. One or more second words having the same meaning as the first word among a plurality of words are extracted from the first database, and the first database associates the words with synonyms of the words included in the first database. The N-gram value is calculated based on the context-dependent value corresponding to the one or more second words and phrases obtained based on the second database, and the second database is divided into words and phrases and words and phrases included in the second database. Corresponding to the corresponding context-dependent value, the context-dependent value indicates the degree to which the meaning of the phrase contained in the second database depends on the context, and in the first sentence, the first phrase is the one. In one or more second sentences replaced with the above second words, one or more consecutive third words including the number of the second words corresponding to the N-gram value are extracted, and the one or more words are extracted. For the third word, the frequency of appearance in the third database is calculated, the third database associates the word with the frequency of appearance of the phrase included in the third database in the third database, and the calculated frequency of appearance is calculated. It is determined whether the sentence is equal to or higher than the threshold value, and if it is determined that the calculated frequency of appearance is equal to or higher than the threshold value, the second sentence of one or more is adopted as a similar sentence of the first sentence, and the external device is used. Output, execute the process.

The device according to another aspect of the present disclosure is a device that generates a similar sentence from the original sentence, and includes an input unit for inputting the first sentence and the first word / phrase among a plurality of words / phrases constituting the first sentence. The second word / phrase extraction unit that extracts one or more second words / phrases having the same meaning from the first database, the first database associates the words with the synonyms of the words / phrases contained in the first database, and the second database The calculation unit that calculates the N-gram value based on the context-dependent value corresponding to the one or more second words and phrases obtained based on the above, and the second database includes words and phrases and words and phrases included in the second database. Corresponding to the corresponding context-dependent value, the context-dependent value indicates the degree to which the meaning of the phrase contained in the second database depends on the context, and in the first sentence, the first phrase is the one. In one or more second sentences replaced by the above second words, a third word extraction unit that extracts one or more consecutive third words including the second words in a number corresponding to the N-gram value. And the calculation unit that calculates the appearance frequency of the one or more third words and phrases in the third database, and the third database corresponds to the words and phrases and the appearance frequency of the words and phrases included in the third database in the third database. In addition, a determination unit for determining whether the calculated appearance frequency is equal to or higher than the threshold, and when it is determined that the calculated appearance frequency is equal to or higher than the threshold, the first or more second sentences are referred to as the first sentence. It is adopted as a sentence similar to the sentence and has an output unit that outputs it to an external device.

The system according to another aspect of the present disclosure is a system that generates a similar sentence from the original sentence, and is the above-mentioned device and the one or more second sentence determined to be adopted as the similar sentence of the first sentence by the above-mentioned device. Using a translation model generated based on the sentence and the translated sentence obtained by translating the first sentence that generated one or more second sentences in the second language, a predetermined translation target sentence is translated and the translation result is obtained. Based on the translation unit that creates the sentence, the evaluation unit that evaluates the translation result sentence created by the translation unit, and the evaluation result of the evaluation unit, the language and / or the translation result sentence of the translation target sentence It includes a generation unit that generates feedback information including linguistic information about the language and evaluation information for the linguistic information.

With such a configuration, the same effect as the above-mentioned similar sentence generation method can be obtained, and one or more second sentences determined to be adopted as similar sentences of the first sentence and the one or more second sentences. Using the translation model generated based on the translated sentence obtained by translating the first sentence generated in the second language, the predetermined translation target sentence is translated to create a translation result sentence, and the created translation result sentence is created. Is evaluated, and based on this evaluation result, feedback information including the language of the translation target sentence and / or the language of the translation result sentence and the evaluation information for the linguistic information is generated. It is possible to realize a similar sentence generation system that can autonomously generate feedback information for learning and reflecting the considered cases, and autonomously learn and reflect the cases considering the context dependence.

Needless to say, the above-mentioned computer program can be distributed via a computer-readable non-temporary recording medium such as a CD-ROM or a communication network such as the Internet.

Further, a system in which some of the components of the similar sentence generator or the similar sentence generation system according to the embodiment of the present disclosure and other components are distributed to a plurality of computers may be configured.

It should be noted that all of the embodiments described below are for showing a specific example of the present disclosure. The numerical values, shapes, components, steps, order of steps, etc. shown in the following embodiments are examples, and are not intended to limit the present disclosure. Further, among the components in the following embodiments, the components not described in the independent claims indicating the highest level concept are described as arbitrary components. In addition, each content can be combined in all the embodiments.

Hereinafter, each embodiment of the present disclosure will be described with reference to the drawings.

(Embodiment 1)
FIG. 1 is a block diagram showing an example of the configuration of the similar sentence generator according to the first embodiment of the present disclosure. The similar sentence generation device 1 shown in FIG. 1 generates a similar sentence from the replacement target sentence (original sentence). The similar sentence generation device 1 includes a replacement target sentence input unit 10, a replacement candidate extraction unit 11, a context dependency rate matching unit 12, a context dependency determination unit 13, a language model matching unit 14, a replacement determination unit 15, and a replacement result output unit 16. , A replacement candidate dictionary 21, a context-sensitive dictionary 22, and a language model database 23.

The replacement target sentence input unit 10 receives a predetermined operation input by the user, and outputs the replacement target sentence (first sentence) input by the user to the replacement candidate extraction unit 11. For example, the replacement target sentence "I can't speak English, so please use Japanese" is input to the replacement target sentence input unit 10. The language of the similar sentence generated by the similar sentence generator 1 is not particularly limited to Japanese, and is another language such as English, Chinese, Korean, French, German, Italian, and Portuguese. May be good.

The replacement candidate dictionary 21 is a replacement candidate storage unit that stores replacement cases in clauses / words / morphemes as a dictionary, and is one or a plurality of replacement candidate character strings that are replacement candidates for the replacement target portion to be replaced from the replacement target sentence. Is memorized in advance. The replacement candidate dictionary 21 is an example of a first database in which words and phrases are associated with synonyms of words and phrases included in the replacement candidate dictionary 21.

FIG. 2 is a diagram showing an example of the data structure of the replacement candidate dictionary 21 shown in FIG. As shown in FIG. 2, the replacement candidate dictionary 21 stores the replacement target portion (word and phrase) and the replacement candidate character string (synonyms of the word and phrase) in association with each other. For example, replacement candidate character strings such as "this is" and "this is" are stored in association with "this is" of the replacement target part, and "speak" is associated with "cannot speak" of the replacement target part. Replacement candidate character strings such as "No", "I can't speak", and "It's a secret" are stored.

The replacement candidate extraction unit 11 replaces a replacement candidate character string (one or more second words) having the same meaning as the replacement target part (first word) among a plurality of words and phrases constituting the replacement target sentence (first sentence). Extract from the candidate dictionary 21. Specifically, the replacement candidate extraction unit 11 divides the replacement target sentence input from the replacement target sentence input unit 10 into units such as clauses / words / morphemes, and replaces the divided clauses / words / morphemes. A part is determined, a replaceable character string pair (replacement candidate character string) stored in association with the replacement target part is searched from the replacement candidate dictionary 21, and one or more replacement candidate character strings are extracted and replaced. It is output to the context dependency rate collation unit 12 together with the target sentence. For example, when the replacement target part is "cannot speak", the replacement candidate extraction unit 11 extracts replacement candidate character strings such as "cannot speak", "cannot speak", and "secret" from the replacement candidate dictionary 21. To do. The method of dividing the sentence to be replaced is not particularly limited to the above example, and various known methods can be used.

The context-sensitive dictionary 22 stores context-sensitive values indicating applicability (context-dependency) when replaced with a phrase / word / morpheme, etc. as a dictionary of a pair of a phrase / word / morpheme, etc. and a numerical value. Dependent value storage unit. Specifically, the context dependency rate dictionary 22 stores in advance a plurality of data pairs in which the replacement candidate character string and the context dependency rate pc indicating the degree to which the replacement candidate character string depends on the context are associated with each other. .. The context-dependent rate dictionary 22 is an example of a second database in which a phrase is associated with a context-dependent value corresponding to a phrase included in the context-dependent rate dictionary 22, and the context-dependent value is included in the context-dependent rate dictionary 22. Indicates the degree to which the meaning of a phrase depends on the context.

FIG. 3 is a diagram showing an example of the data structure of the context-sensitive dictionary 22 shown in FIG. As shown in FIG. 3, in the context-sensitive grammar dictionary 22, for example, pc = 0.35 for the replacement candidate character string "desu", pc = 0.05 for "da ga", and "I can't speak". PC = 0.25 for "", pc = 0.01 for "cannot speak", pc = 0.75 for "secret", etc. are stored in advance in the context-sensitive grammar dictionary 22.

Here, the context dependency rate pc is, for example, a value expressing the probability that a similar candidate sentence using the replacement candidate character string is rejected in the range of 0 to 1 because the replacement candidate character string depends on the context. .. The context-dependent value is not particularly limited to the above-mentioned context-dependent rate pc, and various changes can be made. Other numerical values indicating the degree to which the replacement candidate character string depends on the context can be used, or the replacement candidate character string can be used. You may classify the degree of context dependence (for example, classify the context dependence into large, medium, small, etc. classes) and remember which class it belongs to.

The context dependency rate matching unit 12 searches the context dependency rate pc of the replacement candidate character string from the context dependency rate dictionary 22, extracts the context dependency rate pc stored in association with the replacement candidate character string, and extracts it. The context dependency rate pc is output to the context dependency determination unit 13 together with the replacement target sentence. For example, as the context dependency rate pc, 0.25 is extracted when the replacement candidate character string is "cannot speak", 0.01 is extracted when "cannot speak", and 0.75 is extracted when "it is secret". ..

The context dependency determination unit 13 calculates the N-gram value based on the context dependency value corresponding to the replacement candidate character string (one or more second words) obtained based on the context dependency rate dictionary 22. Specifically, the context dependency determination unit 13 determines the identification target area of the language model database 23 to be referred to for determining the similar candidate sentence including the replacement candidate character string from the value of the context dependency rate pc. The determination result is output to the language model collation unit 14 together with the replacement target sentence.

Here, in the present embodiment, the N-gram language model database is used as the language model database 23, and the language model database 23 contains data in a table format in which language information and its appearance frequency are associated with each other. Is remembered. The language model database 23 is an example of a third database in which words and phrases appearing in the language model database 23 included in the language model database 23 are associated with each other.

FIG. 4 is a diagram showing an example of the data structure of the language model database 23 shown in FIG. As shown in FIG. 4, in the language model database 23, for example, "234,567,890" is used for "English" and "12,345" is used for "English" as the language information and its appearance frequency. , 670 ”,“ 22,222,220 ”for“ English ”, and“ 999,001 ”for“ I like English ”are stored in advance in the language model database 23 in a table format. ing. Further, based on this appearance frequency, for example, the appearance probability can be obtained.

The information stored in the language model database 23 is not particularly limited to the above example, and can be any content as long as it is a table in which language information and values corresponding to its appearance frequency and the like are associated with each other. May be good. Further, the language model of the language model database 23 is not particularly limited to the above N-gram language model, and other language models may be used.

When the language model database 23 is a database of the N-gram language model, the context dependency determination unit 13 determines N (N-gram value) of the N-gram language model of the language model database 23 according to the context dependency rate pc. Is determined to i (a positive integer). Specifically, for example, the context dependency determination unit 13 classifies the context dependency rate pc into four classes, 0 ≦ pc ≦ 0.25 for class 1, and 0.25 <pc ≦ 0.5 for class 2. , 0.5 <pc ≦ 0.75 is class 3, 0.75 <pc ≦ 1 is class 4, and N (positive integer) of N-gram is set as N = 4 in class 1 and N = in class 2. 5. In class 3, N = 6 is determined, and in class 4, N = 7 is determined.

For example, in the case of the replacement candidate character string "I can't speak", the context dependency rate pc is 0.25, which belongs to class 1, and the context dependency determination unit 13 uses class 1 as the identification target area of the language model database 23. Determine the equivalent N-gram or N = 4. The criterion for determining the identification target area is not particularly limited to the above example, and various changes can be made, and the identification target area may be directly mathematically expressed using the context dependency rate pc. For example, when N = floor (k-log2 (pc)) (where k is a constant) and the replacement candidate character string is "cannot speak", the context dependency rate pc is 0.25 and the constant k = 6 In the case of, N = 4.

Further, the language model database 23 is not limited to the N-gram language model, and may be a database based on other language resources. For example, it may be a language model described in a distributed representation such as a real value or a vector, and can be constructed by combining any existing method or existing data. In either case, the range to be searched for the database as the identification target area can be defined by an arbitrary variable, and the arbitrary variable can be determined according to the context dependency rate pc.

In the language model collation unit 14, the replacement sentence (one or more second sentences) in which the replacement target part (first word) is replaced with the replacement candidate character string (one or more second words) in the replacement target sentence (first sentence). ), A continuous N-gram (one or more third words) including a number of replacement candidate character strings (second words) corresponding to the N-gram value is extracted, and an N-gram (one or more third words) is extracted. The language model database 23 is calculated for words).

That is, the language model collation unit 14 searches and extracts the identification target data corresponding to the identification target area determined by the context dependency determination unit 13 from the language model database 23, thereby collating with the replacement candidate character string. Then, pair data of linguistic information consisting of phrases / words / morphological elements related to the replacement candidate character string and a value corresponding to the appearance frequency or appearance probability of the linguistic information is generated, and the replacement judgment unit is generated together with the replacement target sentence. Output to 15.

Specifically, the language model collation unit 14 uses the value of N given as the size of the identification target area referred from the context dependency determination unit 13, and N-gram (for example, replacement candidate) from the language model database 23. When the character string belongs to class 1, the appearance frequency or appearance probability of 4-gram) is acquired, and the collated replacement candidate character string and the acquired appearance frequency or appearance probability are output to the replacement determination unit 15.

The replacement determination unit 15 corresponds to the linguistic information composed of phrases / words / morphemes, etc. obtained from the language model collation unit 14 by using pair data of the appearance frequency or the appearance probability of the linguistic information. It is determined whether to apply the replacement candidate character string to the replacement target sentence or to reject it, and the replacement result is output to the replacement result output unit 16 together with the replacement target sentence.

As an example of the above determination method, the replacement determination unit 15 determines whether the calculated appearance frequency is equal to or greater than the threshold value. Specifically, when the value of the frequency of appearance of the j-th language information (j is an arbitrary integer) is nj and the predetermined threshold value is Th, the replacement determination unit 15 sets nj for all js. If it is> Th, it is determined that the replacement candidate character string is applied to the replacement target sentence, and in other cases, it is determined to be rejected.

For example, 4-gram is used as the N-gram, and for the replacement candidate character string "I can't speak", the language information and the frequency of appearance of that language information are "I can't speak English". "51,550" is "1,720" for "I can't speak English", "530" for "I can't speak Japan", and "I can't speak Japanese" When "3,220" is acquired and Th = 500, the appearance frequency is equal to or higher than the threshold value Th for all of j = 1 to 4, and the replacement candidate character string "I can't speak" is judged to be applicable. Will be done.

The method of determining the application or rejection of the replacement candidate character string to the replacement target sentence is not particularly limited to the above example, and various changes can be made, and the application or rejection is determined according to the distribution of nj. Or (for example, reject the bottom 3% of the appearance frequency of 4-gram), decide application or rejection depending on whether j with nj = 0 exists, or arbitrarily using nj. The application or rejection may be decided according to the value calculated from the formula of.

When it is determined that the calculated appearance frequency is equal to or higher than the threshold value, the replacement result output unit 16 replaces the replacement sentence (one or more second sentences) generated by the replacement candidate character string determined to be applicable. Adopted as a similar sentence to the sentence (first sentence) and output to an external device. Specifically, the replacement result output unit 16 replaces the replacement target part of the replacement target sentence with the replacement candidate character string determined to be applied by the replacement determination unit 15, based on the replacement result, and the replacement candidate character determined to be applicable. The replacement sentence (sentence after replacement) generated by the column is adopted as a similar sentence, and the generated similar sentence is output to an external device (not shown) or the like.

Further, the replacement target sentence (first sentence) is described in the first language (for example, Japanese), the replacement target sentence (first sentence) is included in the bilingual corpus, and the bilingual corpus is described in the first language. The replacement result output unit 16 includes a plurality of pairs of a sentence and a bilingual sentence written in a second language (for example, English), and determines that the sentence is applicable when it is determined that the calculated frequency of appearance is equal to or higher than the threshold value. The replacement sentence (one or more second sentences) generated by the replacement candidate character string may be added to the bilingual corpus as a similar sentence to the replacement target sentence (first sentence).

As described above, the configuration of the similar sentence generation device 1 is not particularly limited to an example in which dedicated hardware is used for each function, and the CPU (Central Processing Unit), ROM (Read Only Memory), and RAM (Random) are used. One or more computers or servers (information processing devices) equipped with an Access Memory) and an auxiliary storage device install a similar sentence generation program for executing the above processing so as to function as a similar sentence generation device. It may be configured as. Further, the replacement candidate dictionary 21, the context dependency rate dictionary 22, and the language model database 23 are not particularly limited to the example provided inside the similar sentence generation device 1, and the replacement candidate dictionary 21, the context dependency rate dictionary 22 is provided on an external server or the like. And the language model database 23 may be provided so that the similar sentence generator 1 acquires necessary information via a predetermined network. The same applies to other embodiments in this regard.

Next, the similar sentence generation process by the similar sentence generation device 1 configured as described above will be described in detail. FIG. 5 is a flowchart showing an example of the similar sentence generation process by the similar sentence generation device 1 shown in FIG. In the following processing, the application / rejection of the replacement candidate character string is determined using the appearance frequency, but the present invention is not particularly limited to this example, and for example, the appearance probability or the like may be used. The same applies to other embodiments in this regard.

First, in step S11, the replacement target sentence input unit 10 accepts the input of the replacement target sentence (original sentence) by the user, and outputs the input replacement target sentence to the replacement candidate extraction unit 11.

Next, in step S12, the replacement candidate extraction unit 11 divides the replacement target sentence into units such as clauses / words / morphemes, determines the replacement target portion from the divided clauses / words / morphemes, and the replacement candidate dictionary. The replacement candidate character string stored in association with the replacement target portion is extracted from 21 and output to the context dependency rate collation unit 12 together with the replacement target sentence.

Next, in step S13, the context dependency rate collation unit 12 collates the context dependency rate dictionary 22, extracts the context dependency rate pc of the replacement candidate character string, and outputs the context dependency rate pc together with the replacement target sentence to the context dependency determination unit 13. To do.

Next, in step S14, the context dependency determination unit 13 refers from the context dependency by determining N of the N-gram of the language model database 23 from the value of the context dependency rate pc of the replacement candidate character string. The language model length is determined, and the determined value of N is output to the language model collation unit 14 together with the replacement target sentence.

For example, the replacement target sentence is "I can't speak English, so please use Japanese", the replacement candidate string is "I can't speak", and the replacement candidate sentence is "I can't speak English." In the case of "Please use Japanese", the context dependency determination unit 13 determines N = 4 as N of the N-gram of the language model database 23.

Next, in step S15, the language model collation unit 14 uses the value of N given as the size of the identification target area referred to by the context dependency determination unit 13, and the appearance frequency of N-gram from the language model database 23. Is acquired, and the collated replacement candidate character string and the acquired appearance frequency are output to the replacement determination unit 15 together with the replacement target sentence.

For example, when replacing "I can't speak" in the above replacement target sentence with "I can't speak", the language model collation unit 14 has a surrounding 4-gram (for example, "ha") including the replaced phrase "I can't speak". "I can't speak English", "I can't speak English", "I can't speak Japanese", "I can't speak Japanese"), collate with the language model database 23, and check each 4-gram. Frequency of appearance (for example, 51,550 for "I can't speak English", 1,720 for "I can't speak English", 530 for "I can't speak Japan", "I can't speak Japanese" 3,220).

Next, in step S16, the replacement determination unit 15 acquires the N-gram including the replacement candidate character string from the language model collation unit 14 and its appearance frequency, and calculates the score of the replacement candidate character string.

Next, in step S17, the replacement determination unit 15 applies the replacement candidate character string to the replacement target sentence by determining whether or not the score (appearance frequency) of the replacement candidate character string is equal to or higher than a predetermined threshold value Th. It is determined whether to reject or reject, and this determination result is output to the replacement result output unit 16 together with the replacement target sentence.

If it is determined in step S17 that the score (appearance frequency) of the replacement candidate character string is less than the predetermined threshold Th, the replacement result output unit 16 rejects the replacement candidate character string and ends the process in step S20. ..

On the other hand, when it is determined in step S17 that the score (appearance frequency) of the replacement candidate character string is equal to or higher than the predetermined threshold value Th, the replacement result output unit 16 replaces the replacement candidate character string with the replacement target sentence in step S18. Create a replacement statement that applies to the target part and replaces the replacement target part of the replacement target sentence with the replacement candidate character string.

Next, in step S19, the replacement result output unit 16 outputs the replacement sentence generated by the replacement candidate character string determined to be applied as a similar sentence, and ends the process.

By the above processing, in the present embodiment, N of the N-gram language model is determined according to the context-sensitive rate pc, N is larger as the context-sensitive rate pc is larger, and N is smaller as the context-sensitive rate pc is smaller. It is set. Further, by collating the language model database 23 with the determined N, the appearance frequency of the N-gram including the replacement candidate character string is obtained, and the replacement candidate character string is generated based on the obtained appearance frequency. Since it is determined whether or not to adopt the replacement sentence as a similar sentence, the frequency of occurrence of the i-gram including the replacement candidate character string having a large context dependency rate pc is obtained with high accuracy by using a wide identification target area. In addition, the frequency of appearance of the N-gram including the replacement candidate character string having a small context dependence rate pc can be obtained at low cost and with high accuracy by using the narrow identification target area. As a result, the search cost for the language model database 23 can be reduced, and similar sentences can be identified with high accuracy.

(Embodiment 2)
FIG. 6 is a block diagram showing an example of the configuration of the similar sentence generation system according to the second embodiment of the present disclosure. The similar sentence generation system shown in FIG. 6 includes a similar sentence generation device 1a and a translation device 2.

The similar sentence generation device 1a includes a replacement target sentence input unit 10a, a replacement candidate extraction unit 11, a context dependency rate matching unit 12, a context dependency determination unit 13, a language model matching unit 14, a replacement determination unit 15, and a replacement result output unit 16. , A data update unit 17, a replacement candidate dictionary 21, a context-sensitive dictionary 22, and a language model database 23. The translation device 2 includes a translation corpus generation unit 31, a translation model generation unit 32, a translated sentence input unit 33, a machine translation unit 34, a translation result sentence output unit 35, a translation result evaluation unit 36, and a feedback data generation unit 37. ..

The similar sentence generation device 1a generates a similar sentence from the replacement target sentence (original sentence), and outputs the similar sentence or the like determined to be adopted to the translation device 2. The translation device 2 arbitrarily uses a translation model generated based on a similar sentence determined to be adopted by the similar sentence generator 1a and a translated sentence obtained by translating the original sentence that generated the similar sentence in a predetermined language. Translate the translation target sentence of the above to create a translation result sentence, and based on the evaluation result of the created translation result sentence, linguistic information regarding at least one of the language of the translation target sentence and the language of the translation result sentence, and this language information Feedback information including the evaluation information is generated and fed back to the similar sentence generator 1a. The similar sentence generator 1a updates at least one data of the replacement candidate dictionary 21, the context-sensitive dictionary 22, and the language model database 23 based on the feedback information.

Here, the difference between the similar sentence generating device 1a shown in FIG. 6 and the similar sentence generating device 1 shown in FIG. 1 is that the data update unit that updates the data of the replacement candidate dictionary 21, the context dependency rate dictionary 22, and the language model database 23. 17 is added, and the replacement target sentence input unit 10a outputs the translated sentence of the input replacement target sentence (original sentence) to the translation device 2 in addition to the input of the replacement target sentence, and the other points are the same. Therefore, the same parts are designated by the same reference numerals, and detailed description thereof will be omitted.

The replacement target sentence input unit 10a receives a predetermined operation input by the user, outputs the replacement target sentence input by the user to the replacement candidate extraction unit 11, and then outputs the replacement result output unit from the replacement candidate extraction unit 11 for the replacement target sentence. The processing up to 16 is the same as the processing from the replacement candidate extraction unit 11 to the replacement result output unit 16 shown in FIG. 1, and the replacement result output unit 16 is a replacement candidate character string determined to be applied by the replacement determination unit 15. The replacement sentence (similar sentence) generated by is output to the bilingual corpus generation unit 31.

Further, the replacement target sentence input unit 10a receives a predetermined operation input by the user, and translates the original sentence generated by the replacement sentence into a predetermined language, that is, a translation sentence (original sentence) of the replacement target sentence. The bilingual text corresponding to) is output to the bilingual corpus generation unit 31. For example, when the above replacement sentence is created in Japanese (original language sentence) and the translation device 2 performs Japanese-English translation, the above translation sentence is created in English (target language sentence). The original language sentence and the target language sentence are not particularly limited to the above example, and when the similar sentence generator 1a generates a similar sentence in English, English may be the original language sentence and Japanese may be the target language sentence. , Also may be other languages such as Chinese, Korean, French, German, Italian, Portuguese.

The translation corpus generation unit 31 associates the replacement sentence output from the replacement result output unit 16 with the translation sentence of the replacement target sentence output from the replacement target sentence input unit 10a, generates a new translation corpus, and generates a translation model. Output to the generation unit 32. The method for generating the bilingual corpus is not particularly limited to the above example, and a new bilingual corpus may be added to the already created bilingual corpus, and various known methods can be used.

The translation model generation unit 32 generates a translation model by predetermined learning using the new translation corpus generated by the translation corpus generation unit 31, and outputs the translation model to the machine translation unit 34. Since various known methods can be used as the method for generating the translation model, detailed description thereof will be omitted.

The translated sentence input unit 33 receives a predetermined operation input by the user, and outputs the translation target sentence (original language sentence) input by the user to the machine translation unit 34. The machine translation unit 34 translates the translation target sentence using the translation model generated by the translation model generation unit 32, and outputs the translation result sentence (target language sentence) to the translation result sentence output unit 35 together with the translation target sentence. .. The translation result sentence output unit 35 outputs the translation result sentence together with the translation target sentence to the translation result evaluation unit 36 as the translation result.

The translation result evaluation unit 36 evaluates the translation accuracy and quality of the translation result sentence (target language sentence) output from the translation result sentence output unit 35. Here, as the evaluation method of the translation result evaluation unit 36, the evaluation may be performed by a mechanical numerical index, or the manual evaluation result may be input to the translation result evaluation unit 36. The translation result evaluation unit 36 outputs the evaluation information such as the evaluation value or the evaluation category to the feedback data generation unit 37 as the evaluation result in association with the translation result sentence (target language sentence) and / or the translation target sentence (original language sentence). To do.

The feedback data generation unit 37 generates feedback data to be fed back to the similar sentence generation device 1a as feedback information based on the evaluation result output from the translation result evaluation unit 36, and outputs the feedback data to the data update unit 17. Here, the feedback data is pair data of arbitrary language information on the original language and / or target language side and evaluation information of a value or a state related to the language information. As this feedback data, various data can be used, and the following data can be used.

For example, when the translation result is bad, the translation result sentence (target language sentence) is corrected by the user or a predetermined translation result sentence correction device, and a better translation sentence is input, so that the input translation sentence and the original translation sentence are used. The pair data of the language information of the pair with the translation target sentence (original language sentence) and the evaluation information of the state (bad) of the translation result may be used as feedback data.

In addition, when a better translation result sentence can be obtained by correcting the translation target sentence (original language sentence) by the user or a predetermined translation target sentence correction device and inputting a translation target sentence having a different expression to the same effect. , The language information of the pair of the original translation target sentence (original language sentence) and the translation target sentence (original language sentence) with good translation result, and the evaluation information of the state of the translation result (good / bad binary value). The pair data may be used as feedback data.

In addition, one or more sentences close to the translation target sentence (original language sentence) are extracted from the bilingual corpus, and an evaluation value (for example, good) of whether or not the sentence is broken as the original language by the user or a predetermined translation sentence evaluation device is good. (2 values of / bad) is obtained, an evaluation value is given to the extracted sentence close to the original language sentence, and the pair data of this evaluation value and the language information indicating the sentence close to the original language sentence is used as feedback data. May be good.

Further, the machine translation unit 34 creates a plurality of translation result sentences, selects a more appropriate translation result sentence from the translation result sentences by the user or a predetermined translation sentence evaluation device, and selects the selected translation result sentence and the translation result sentence is not selected. The pair data of the language information paired with the translation result sentence and the evaluation information indicating the selection result of these translation result sentences may be used as feedback data.

The data update unit 17 is based on the feedback data (pair data of the language information and the evaluation information of the value or the state related to the language information) generated by the feedback data generation unit 37, the replacement candidate dictionary 21, the context dependency rate dictionary 22, and the language. Update the contents of at least one of the model databases 23.

Further, when the feedback data includes a replacement candidate character string having a context dependency, the data update unit 17 updates the context dependency rate dictionary 22 and the language model database 23. Further, when the feedback data includes a new sentence expression, the data update unit 17 changes the value of the context dependency rate of the context dependency rate dictionary 22 according to the sentence expression, and the language model so as to include the new sentence expression. The N-gram of the database 23 is partially constructed, and the language model database 23 is updated.

Further, the data update unit 17 includes information on the original language side in the language information, and the information registered in the replacement candidate dictionary 21, the context dependency rate dictionary 22, or the language model database 23 in the language information. Is included, the corresponding information in the replacement candidate dictionary 21, the context dependency dictionary 22 or the language model database 23 may be updated or added, depending on the value or state evaluation information of the corresponding feedback data. Delete it.

For example, when the language information of the original language side having the positive (positive) value or the evaluation information of the state is fed back, the data update unit 17 adds a predetermined weight to the appearance frequency to increase the value of the appearance frequency. The value including the language information in the language model database 23 is changed in the positive direction. On the other hand, when the language information on the original language side having the evaluation information of the negative value or the state is fed back, the data update unit 17 updates the context dependency rate in the direction of increasing the context dependence rate. By doing so, the value including the language information of the context-sensitive rate dictionary 22 is changed in the negative direction.

In addition, the language information of the original translation target sentence (original language sentence) with poor translation result and the translation target sentence (original language sentence) with good translation result, and the evaluation information of each translation result status (bad / good). When the pair data of the above is fed back and the difference of the translation target sentence corresponding to the good state with respect to the original translation target sentence in the bad state is not registered in the replacement candidate dictionary 21, the data update unit 17 performs the difference corresponding to the good state. Is registered in the replacement candidate dictionary 21.

In addition, when the pair data of the language information of the translation target sentence (original language sentence) whose translation result was bad and the evaluation information of the translation result state (bad) was fed back, the data update unit 17 had a bad translation result. The replacement candidate character string of the translation target sentence is deleted from the replacement candidate dictionary 21.

The configuration of the similar sentence generation device 1a and the translation device 2 is not particularly limited to the example of configuring each function with dedicated hardware as described above, and includes a CPU, ROM, RAM, auxiliary storage device, and the like1. A unit or a plurality of computers or servers (information processing devices) may be configured to install a program for executing the above processing and function as a similar sentence generator or a translation device.

Next, the similar sentence generation process including the feedback data update process by the similar sentence generation system configured as described above will be described in detail. FIG. 7 is a flowchart showing an example of the similar sentence generation process including the feedback data update process of the similar sentence generation system shown in FIG. Of the processes shown in FIG. 7, the same processes as those shown in FIG. 5 are designated by the same reference numerals, and detailed description thereof will be omitted.

First, as a similar sentence generation process by the similar sentence generation device 1a, in step S11a, the replacement target sentence input unit 10a accepts the input of the bilingual sentence and the replacement target sentence corresponding to the original sentence by the user, and converts the bilingual sentence into the bilingual corpus generation unit. It is output to 31 and the replacement target sentence is output to the replacement candidate extraction unit 11. The timing of outputting the bilingual sentence to the bilingual corpus generation unit 31 is not particularly limited to the above example, and the replacement target sentence input unit 10a outputs the bilingual sentence to the bilingual corpus generation unit 31 at the time of processing in step S17. It may be.

Next, in steps S12 to S17, the same processing as in steps S12 to S17 shown in FIG. 5 is executed, and in step S17, it is determined that the score (appearance frequency) of the replacement candidate character string is less than the predetermined threshold value Th. In the case, in step S20, the replacement result output unit 16 rejects the replacement candidate character string and ends the process.

On the other hand, when it is determined in step S17 that the score (appearance frequency) of the replacement candidate character string is equal to or higher than a predetermined threshold value Th, in step S18, the same process as in step S18 shown in FIG. 5 is executed, and then the step In S19, the replacement result output unit 16 outputs the replacement sentence (replacement sentence) generated by the replacement candidate character string determined to be applied by the replacement determination unit 15 to the parallel translation corpus generation unit 31, and is used by the similar sentence generation device 1a. The similar sentence generation process ends.

Next, as a feedback data update process by the translation device 2 and the similar sentence generation device 1a, in step S21, the translation corpus generation unit 31 receives the replacement sentence output from the replacement result output unit 16 and the replacement target sentence input unit 10a. It associates with the output bilingual sentence, generates a new bilingual corpus, and outputs it to the translation model generation unit 32.

Next, in step S22, the translation model generation unit 32 uses the new translation corpus generated by the translation corpus generation unit 31 to generate a translation model by learning and outputs it to the machine translation unit 34.

Next, in step S23, the translated sentence input unit 33 accepts the input of the translation target sentence by the user, and outputs any translation target sentence that the user desires to translate to the machine translation unit 34.

Next, in step S24, the machine translation unit 34 translates the translation target sentence into the translation result sentence by the translation model generated by the translation model generation unit 32, and translates the translation result sentence together with the translation target sentence in the translation result sentence output unit 35. Output to.

Next, in step S25, the translation result sentence output unit 35 outputs the translation result sentence together with the translation target sentence to the translation result evaluation unit 36.

Next, in step S26, the translation result evaluation unit 36 evaluates the translation accuracy and quality of the translation result sentence output from the translation result sentence output unit 35, and the evaluation result includes an evaluation value or an evaluation category. The information is associated with the translation result sentence and output to the feedback data generation unit 37.

Next, in step S27, the feedback data generation unit 37 generates feedback data from the evaluation result output from the translation result evaluation unit 36 and outputs it to the data update unit 17.

Finally, in step S28, the data update unit 17 updates the database contents of at least one of the replacement candidate dictionary 21, the context dependency rate dictionary 22, and the language model database 23 based on the feedback data generated by the feedback data generation unit 37. Then, the feedback data update process is terminated.

By the above processing, in the present embodiment, the translation result sentence obtained by translating the predetermined translation target sentence is evaluated by using the translation model generated based on the replacement sentence determined to be adopted and the bilingual sentence to the original sentence. However, based on this evaluation result, feedback information including the language of the translation target sentence and / or the language of the translation result sentence and the evaluation information for this language information is generated, so context dependence is taken into consideration. Feedback data for learning and reflecting the case in the similar sentence generator 1a can be autonomously generated.

Further, in the present embodiment, the replacement candidate dictionary 21, the context dependency rate dictionary 22, and the language model database 23 are updated by using the feedback data including the language information and the evaluation information, so that the context dependency is taken into consideration. Cases can be reflected in the replacement candidate dictionary 21, the context dependency dictionary 22 and the language model database 23, and even in new sentence expressions that do not exist in the replacement candidate dictionary 21 before update, the context dependency dictionary 22 and the language model database 23. It is possible to identify similar sentences with high efficiency and autonomously.

The present disclosure can reduce the search cost for the database of the language model and can identify similar sentences with high accuracy. Therefore, a similar sentence generation method, a similar sentence generation program, and a similar sentence generation to generate a similar sentence from the original sentence. It is useful for an apparatus and a similar sentence generation system including the similar sentence generation apparatus.

1, 1a Similar sentence generator 2 Translation device 10, 10a Replacement target sentence input section 11 Replacement candidate extraction section 12 Context dependency rate matching section 13 Context dependency judgment section 14 Language model matching section 15 Replacement judgment section 16 Replacement result output section 17 Data update section 21 Replacement candidate dictionary 22 Context dependency rate dictionary 23 Language model database 31 Bilingual corpus generation section 32 Translation model generation section 33 Translated sentence input section 34 Machine translation section 35 Translation result sentence output section 36 Translation result evaluation section 37 Feedback data Generator

Claims (7)

A method in a device that generates similar sentences from the original text.
Enter the first sentence,
Of the plurality of words and phrases constituting the first sentence, one or more second words and phrases having the same meaning as the first word and phrase are extracted from the first database, and the first database is the words and phrases and words and phrases included in the first database. Correspond with synonyms of
The N-gram value is calculated based on the context-sensitive value corresponding to the one or more second words and phrases obtained based on the second database, and the second database is divided into words and phrases and words and phrases contained in the second database. Corresponding to the corresponding context-dependent value, the context-dependent value indicates the degree to which the meaning of the phrase contained in the second database depends on the context.
In one or more second sentences in which the first word is replaced with the one or more second words in the first sentence, one or more consecutive words including the second word corresponding to the N-gram value. Extract the third phrase of
Wherein for one or more of the third word, calculates the appearance frequency in the third database, the third database, and phrases, the frequency of occurrence of the word before Symbol third database association,
It is determined whether the calculated frequency of appearance is equal to or higher than the threshold value.
When it is determined that the calculated appearance frequency is equal to or higher than the threshold value, the one or more second sentences are adopted as similar sentences to the first sentence and output to an external device.
Method. The first sentence is written in the first language and
The first sentence is included in a bilingual corpus, and the bilingual corpus includes a plurality of pairs of a sentence written in a first language and a bilingual sentence written in a second language.
When it is determined that the calculated appearance frequency is equal to or higher than the threshold value, the one or more second sentences are added to the bilingual corpus as similar sentences to the first sentence.
The method according to claim 1. The third database includes a database of N-gram language models.
The N of the N-gram language model is determined to be i (a positive integer) according to the context-sensitive value.
By collating the third database, the frequency of occurrence of i-gram including the second phrase was determined.
Based on the frequency of appearance of the i-gram including the second phrase, it is determined whether or not the one or more second sentences are adopted as similar sentences to the first sentence.
The method according to claim 1 or 2. Generated based on the one or more second sentences determined to be adopted as similar sentences to the first sentence, and the translated sentence obtained by translating the first sentence that generated the one or more second sentences in a second language. Using the translated model, the specified translation target sentence is translated to create a translation result sentence.
Evaluate the translation result sentence and
Based on the evaluation result of the translation result sentence, the language and / or language clauses of the translation result sentence translated sentence, and language information including at least one of words and morphemes, said against the language information Generate feedback information, including evaluation information that indicates whether the text to be translated is broken as the language .
The method according to any one of claims 1 to 3. A program for operating a computer as a device for generating similar sentences from the original text.
On the computer
Enter the first sentence,
Of the plurality of words and phrases constituting the first sentence, one or more second words and phrases having the same meaning as the first word and phrase are extracted from the first database, and the first database is the words and phrases and words and phrases included in the first database. Correspond with synonyms of
The N-gram value is calculated based on the context-sensitive value corresponding to the one or more second words and phrases obtained based on the second database, and the second database is divided into words and phrases and words and phrases contained in the second database. Corresponding to the corresponding context-dependent value, the context-dependent value indicates the degree to which the meaning of the phrase contained in the second database depends on the context.
In one or more second sentences in which the first word is replaced with the one or more second words in the first sentence, one or more consecutive words including the second word corresponding to the N-gram value. Extract the third phrase of
Wherein for one or more of the third word, calculates the appearance frequency in the third database, the third database, and phrases, the frequency of occurrence of the word before Symbol third database association,
It is determined whether the calculated frequency of appearance is equal to or higher than the threshold value.
When it is determined that the calculated appearance frequency is equal to or higher than the threshold value, the one or more second sentences are adopted as similar sentences to the first sentence and output to an external device.
A program that executes processing. A device that generates similar sentences from the original text
The input part where the first sentence is input and
A second phrase extraction unit that extracts one or more second phrases having the same meaning as the first phrase from the plurality of phrases constituting the first sentence from the first database, and the first database are the phrases and the first phrase. Corresponds to synonyms of words and phrases contained in the database
A calculation unit that calculates an N-gram value based on a context-sensitive value corresponding to the one or more second words and phrases obtained based on the second database, and the second database are included in the words and phrases and the second database. Corresponding to the context-dependent value corresponding to the phrase, the context-dependent value indicates the degree to which the meaning of the phrase contained in the second database depends on the context.
In one or more second sentences in which the first word is replaced with the one or more second words in the first sentence, one or more consecutive words including the second word corresponding to the N-gram value. The third word extraction part that extracts the third word of
Wherein for one or more of the third word, a calculation unit for calculating an appearance frequency in the third database, the third database, and phrases, the frequency of occurrence of the word before Symbol third database association,
A determination unit for determining whether the calculated appearance frequency is equal to or higher than the threshold value,
When it is determined that the calculated appearance frequency is equal to or higher than the threshold value, the one or more second sentences are adopted as similar sentences to the first sentence, and an output unit for outputting to an external device is provided.
apparatus. A system that generates similar sentences from the original text
The device according to claim 6 and
The one or more second sentences determined to be adopted as similar sentences to the first sentence by the apparatus, and the translated sentence obtained by translating the first sentence that generated the one or more second sentences into a second language. A translation department that translates a predetermined translation target sentence and creates a translation result sentence using the translation model generated based on it,
An evaluation unit that evaluates the translation result sentence created by the translation unit,
Based on the evaluation result of the evaluation unit, the language of the sentence to be translated and / or language clauses of the translation result sentence, and language information including at least one of words and morphemes, the translation against the language information It is provided with a generation unit that generates feedback information including evaluation information indicating whether or not the language of the target sentence is broken .
system.

Images