JP6934621B2 - Methods, equipment, and programs - Google Patents

Description

The present disclosure relates to a technique for translating input sentences.

In recent years, when translating an input sentence, research has been actively conducted in which a multifaceted translation result is presented to a user rather than simply presenting the machine translation result of the input sentence.

For example, Patent Document 1 generates a plurality of paraphrase sentences in which an input text sentence is paraphrased by another expression having the same content, machine translates the generated paraphrase sentences, and translates the generated paraphrase sentences into translation reliability. Based on this, candidates for translation target paraphrase sentences are extracted, and a technique for identifying translation target paraphrase sentences from the extracted translation target paraphrase sentences is disclosed.

In Patent Document 2, in order to supplement the uncertainty of machine translation, an example sentence having an expression close to the input original sentence is searched, a parallel translation text of the target language corresponding to the searched example sentence is acquired, and the acquired parallel translation text is input. Disclose the technology to display together with the machine translation result of.

Japanese Unexamined Patent Publication No. 2012-1599969 Patent No. 5103718

However, the above-mentioned conventional technique has a problem that the translation reliability cannot be expected unless the knowledge space provided in the translator is enhanced, and therefore needs to be further improved.

The method according to one aspect of the present disclosure is a method of providing a translated text.
Obtain the first sentence written in the first language to be translated via the user's terminal,
It is determined whether or not the first sentence is included in the database containing a plurality of pairs of the sentence described in the first language and the bilingual sentence described in the second language.
When it is determined that the first sentence is not included in the database, a plurality of second sentences in which one or more words constituting the first sentence are replaced based on a predetermined rule are generated.
The degree of syntactic matching between the plurality of second sentences and the plurality of sentences described in the first language included in the database is calculated.
One or more third sentences described in the first language included in the database whose calculated degree of matching is equal to or more than the threshold value are extracted.
In the database, one or more fourth sentences described in the second language, which are parallel translations of the one or more third sentences, are displayed on the user's terminal as a translation reference of the first sentence. ..

According to the present disclosure, it is possible to present a translation result useful for a user without enhancing the knowledge space in order to generate a translation of an input sentence or a similar sentence with high reliability.

It is a block diagram which shows the structure of the translation support apparatus which is an example of the apparatus which concerns on embodiment of this disclosure. It is a figure which shows an example of the translation support device when it is composed of a mobile information terminal, and the translation support device when it is composed of a stationary computer. It is a figure which shows an example of the translation support apparatus 1 in the case of being configured by a cloud system. It is a block diagram which shows the detailed structure of the paraphrase generation part. It is a figure which shows an example of an input paraphrase sentence and an example sentence. It is a table which summarized the specific example in this embodiment. It is a figure which shows an example of the output image which an output part displays. It is a flowchart which shows an example of the processing of the translation support apparatus which concerns on embodiment of this disclosure. It is a flowchart which shows an example of the details of the process of S5 of FIG. It is a flowchart which shows an example of the details of the process of S6 of FIG.

(Background to one aspect of this disclosure)
In order to improve the translation quality of the translator, the input sentence is paraphrased, multiple input paraphrases are generated, the translated sentences of the generated multiple input paraphrases are presented, and the translated sentences of the presented multiple input paraphrases are presented. A technique has been proposed in which the user selects the most suitable translation from among them (Patent Document 1).

Further, a technique has been proposed in which an example translation result of a sentence similar to an input sentence or a sentence having a partial match is presented together with the machine translation result (Patent Document 2).

However, the technique of Patent Document 1 cannot present a highly accurate translated sentence when there is no knowledge data in the vicinity of the input sentence and the input paraphrase sentence in the knowledge space used by the translator to generate the translated sentence. There's a problem.

Further, the technique of Patent Document 2 complements the uncertainty of machine translation of the input sentence unless the example parallel translation database owned by the translator contains an example sentence similar to the input sentence or a partially matching example sentence. There is a problem that it is not possible to present an example translation that can be obtained.

As described above, the techniques of Patent Documents 1 and 2 have a problem that the translation reliability cannot be expected unless the knowledge space is strengthened. Further, in the techniques of Patent Documents 1 and 2, even if the knowledge space is enhanced, improvement in translation reliability cannot be expected when an input sentence having an expression outside the range of the enhanced knowledge space is input. In addition, the enhancement of the knowledge space has a problem in terms of cost effectiveness.

The present disclosure provides a technique for presenting a translation useful for a user without enhancing the knowledge space in order to generate a translation of an input sentence or a similar sentence with high reliability.

The method according to one aspect of the present disclosure is a method of providing a translated text.
Obtain the first sentence written in the first language to be translated via the user's terminal,
It is determined whether or not the first sentence is included in the database containing a plurality of pairs of the sentence described in the first language and the bilingual sentence described in the second language.
When it is determined that the first sentence is not included in the database, a plurality of second sentences in which one or more words constituting the first sentence are replaced based on a predetermined rule are generated.
The degree of syntactic matching between the plurality of second sentences and the plurality of sentences described in the first language included in the database is calculated.
One or more third sentences described in the first language included in the database whose calculated degree of matching is equal to or more than the threshold value are extracted.
In the database, one or more fourth sentences described in the second language, which are parallel translations of the one or more third sentences, are displayed on the user's terminal as a translation reference of the first sentence. ..

This aspect does not merely present a translated sentence of the paraphrase sentence of the first sentence to be translated. That is, in this embodiment, the degree of syntactic matching is equal to or greater than the threshold value for a plurality of second sentences in which the first sentence is replaced by a predetermined rule from among a plurality of sentences written in the first language stored in the database. One or more sentences that are are extracted as the third sentence. Then, in this aspect, the fourth sentence, which is a parallel translation of the extracted third sentence, is displayed as a parallel translation reference.

Here, although the content is different from the first sentence, it is better to present the fourth sentence, which is an example parallel translation of the third sentence whose sentence structure is the same as or similar to that of the first sentence. Rather than presenting, it is more likely that a translation result that is useful to the user can be presented.

That is, when such a fourth sentence is presented, the knowledge space used when generating the translated sentence is widely used, and a translation result useful for the user can be obtained. Since this aspect pays attention to this point, it is possible to present a translation result useful for the user.

Further, since this aspect does not require that a translated sentence of the first sentence or a similar sentence be generated with high reliability, it is necessary to use a knowledge space having abundant knowledge data that can meet the request. There is no. Therefore, this aspect can present translation results that are useful to the user without enhancing the knowledge space.

In the above aspect, one or more second sentences out of the plurality of second sentences are machine-translated into the second language to generate one or more fifth sentences.
At least one of the one or more fourth sentences and the one or more fifth sentences may be displayed on the terminal of the user.

In this embodiment, one or more fifth sentences, which are translations of one or more second sentences out of a plurality of second sentences in which the first sentence is paraphrased according to a predetermined rule, are generated, and the fourth sentence and the fifth sentence are combined. At least one of the above is presented. Therefore, by presenting the fifth sentence, various translation results are presented, and the possibility of presenting useful translation results for the user is increased.

In the above aspect, the degree of agreement may be calculated based on a first index indicating the text similarity between the plurality of second sentences and the plurality of sentences included in the database.

According to this aspect, a sentence having the same or similar text similarity to the plurality of second sentences is extracted as the third sentence from the plurality of sentences stored in the database. Therefore, it is possible to prevent a sentence unrelated to the plurality of second sentences from being extracted as the third sentence while using the knowledge space extensively.

In the above aspect, the degree of coincidence is a sentence having a sentence structure that matches or is similar to the plurality of second sentences among the plurality of sentences included in the database, and the degree of text similarity with the first sentence. It may be calculated based on the second index which shows a larger value as the sentence has a smaller value.

According to this aspect, among a plurality of sentences included in the database, example sentences that have the same or similar sentence structure to the second sentence but have different contents from the first sentence are extracted. .. Therefore, various third sentences can be extracted and the knowledge space can be widely used.

In the above aspect, the degree of agreement may be calculated based on a third index indicating the similarity of the sentence structure between the plurality of second sentences and the plurality of sentences included in the database.

According to this aspect, a sentence having a sentence structure that matches or is similar to a plurality of second sentences is extracted as the third sentence. Therefore, it is possible to prevent the extraction of the third sentence, which is less relevant to the plurality of second sentences, while using the knowledge space over a wide range.

In the above aspect, the degree of matching may be calculated based on a fourth index showing a larger value as the number of matching parts of speech increases in the plurality of second sentences and the plurality of sentences included in the database. good.

According to this aspect, a sentence having a large number of matching part of speech for a plurality of second sentences is extracted as the third sentence. Therefore, it is possible to prevent sentences having low relevance to a plurality of second sentences from being extracted while using the knowledge space extensively.

Further, in this aspect, the evaluation value may be calculated by combining at least two of the first to fourth indexes. As a result, a wider variety of third sentences can be extracted, and various translation results can be presented. As a result, it is possible to extract a third sentence in which the translation result is a hint for translation of the first sentence, although the similarity in the first language is low with respect to the first sentence.

In the above aspect, the first index may show a larger value as the second sentence has more replacement points.

According to this aspect, since the first index becomes larger as the second sentence has many replacement parts, the third sentence similar to the second sentence with many replacement parts can be extracted, and more diverse translation results can be presented.

In the above aspect, the one or more second sentences may be extracted from the plurality of second sentences based on the text similarity between the plurality of second sentences and the one or more third sentences. ..

According to this aspect, a similar second sentence is extracted from a plurality of second sentences with respect to the third sentence extracted based on the degree of agreement, so that while using the knowledge space extensively, It is possible to prevent the second sentence unrelated to the first sentence from being extracted.

In the above aspect, the predetermined rule may be a first paraphrase rule in which the first word included in the element piece constituting the first sentence is paraphrased by the second word having a context-like relationship.

According to this aspect, a second sentence is generated in which the first word constituting the first sentence is paraphrased by the second word having a context-like relationship. Therefore, it is possible to generate a variety of second sentences as compared with the case of adopting the mode of generating the second sentence having the same meaning with respect to the first sentence. As a result, it is possible to generate a second sentence in which the translation result is a hint for translation of the first sentence, although the similarity in the first language is low with respect to the first sentence.

In the above aspect, the predetermined rule may be a second paraphrase rule that paraphrases the first word contained in the element piece constituting the first sentence into the second word having a co-occurrence relationship.

According to this aspect, a second sentence is generated in which the first word constituting the first sentence is paraphrased by the second word having a co-occurrence relationship. Therefore, it is possible to generate a variety of second sentences as compared with the case of adopting the mode of generating the second sentence having the same meaning with respect to the first sentence. As a result, it is possible to generate a second sentence in which the translation result is a hint for translation of the first sentence, although the similarity in the first language is low with respect to the first sentence.

In the above aspect, the predetermined rule may be a third paraphrase rule that paraphrases the first word contained in the element piece constituting the first sentence into the second word having an implication relationship.

According to this aspect, a second sentence is generated in which the first word constituting the first sentence is paraphrased by the second word having an implication relationship. Therefore, it is possible to generate a variety of second sentences as compared with the case of adopting the mode of generating the second sentence having the same meaning with respect to the first sentence. As a result, it is possible to generate a second sentence in which the translation result is a hint for translation of the input sentence, although the similarity in the first language is low with respect to the first sentence.

In the above aspect, the predetermined rule may be a fourth paraphrase rule in which the first word included in the element piece constituting the first sentence is paraphrased into the second word having a higher-lower relationship.

According to this aspect, a second sentence is generated in which the first word constituting the first sentence is paraphrased by the second word having a higher-lower relationship. Therefore, it is possible to generate a variety of second sentences as compared with the case of adopting a mode in which a second sentence having the same meaning and content is simply generated for the first sentence. As a result, it is possible to generate a second sentence in which the translation result or the like is a hint for the translation of the first sentence, although the similarity in the first language is low with respect to the first sentence.

In the above aspect, the one or more fourth sentences may be presented with a paraphrase part for the first sentence distinguished from other parts.

According to this aspect, the user can easily recognize the paraphrase part for the first sentence.

(Embodiment)
FIG. 1 is a block diagram showing a configuration of a translation support device 1 which is an example of the device according to the embodiment of the present disclosure. The translation support device 1 is a device that translates an input sentence written in the first language into a second language. As the first language, for example, languages such as Japanese, English, French, and German can be adopted. As the second language, a language different from the first language can be adopted. In the following explanation, Japanese is adopted as the first language and English is adopted as the second language, which is an example.

The translation support device 1 includes an input unit 2, an example match determination unit 3, an example parallel translation DB (database) 4 (an example of a database), a paraphrase generation unit 5, an extraction unit 6, a machine translation unit 7, and a reliability imparting unit 8. And an output unit 9 (an example of a presentation unit). In FIG. 1, the translation support device 1 is composed of, for example, a computer including a CPU, a ROM, and a RAM. The input unit 2 is composed of, for example, an input device such as a touch panel or an input device such as a keyboard and a mouse. The example match determination unit 3, the paraphrase generation unit 5, the extraction unit 6, the machine translation unit 7, and the reliability imparting unit 8 are realized, for example, by executing a program in which the CPU functions the computer as the translation support device 1. It may be realized by a dedicated hardware circuit. The program may be provided by downloading over a network or may be recorded and provided on a computer-readable non-temporary recording medium. Further, the example bilingual translation DB4 may be composed of a storage device (memory). Further, the output unit 9 may be composed of a display device or a speaker.

The translation support device 1 may be composed of a mobile information terminal such as a smartphone or a tablet terminal, or may be composed of a stationary computer.

FIG. 2 is a diagram showing an example of a translation support device 1 when configured with a portable information terminal and a translation support device 1 when configured with a stationary computer. In the left figure of FIG. 2, the translation support device 1 is composed of a mobile information terminal such as a smartphone or a tablet terminal. In the right figure of FIG. 2, the translation support device 1 is composed of a stationary computer. In these cases, each component shown in FIG. 1 is aggregated in a personal digital assistant or a stationary computer.

Alternatively, the translation support device 1 may be configured by a cloud system. FIG. 3 is a diagram showing an example of the translation support device 1 when configured with a cloud system. The cloud system is composed of a server SV1 and one or a plurality of terminals TE1. The server SV1 and the terminal TE1 are communicably connected via a network NT such as the Internet. The server SV1 is a cloud server composed of one or a plurality of computers. The terminal TE1 may be composed of a mobile information terminal such as a smart phone or a tablet terminal, or may be composed of a stationary computer.

In this case, the input unit 2 and the output unit 9 shown in FIG. 1 are composed of the terminal TE1 owned by the user. Further, the example match determination unit 3, the example parallel translation DB 4, the paraphrase generation unit 5, the extraction unit 6, the machine translation unit 7, and the reliability imparting unit 8 shown in FIG. 1 are composed of the server SV1. That is, the translation support function is implemented in the server SV1, and the terminal TE1 provides a user interface.

The reference is returned to FIG. The input unit 2 acquires an input sentence (an example of the first sentence) written in the first language to be translated. The input sentence is a sentence input by the user and is a sentence written in the first language.

The example matching determination unit 3 determines whether or not an example sentence matching the input sentence acquired by the input unit 2 is stored in the example bilingual translation DB 4. Then, if the example sentence matching the input sentence is stored in the example parallel translation DB 4, the example matching determination unit 3 outputs the matching example sentence and the example parallel translation including the example sentence to the output unit 9. Here, the example match determination unit 3 may determine, for example, that the example sentence and the input sentence match when they completely match. On the other hand, if the input sentence acquired by the input unit 2 is not stored in the example parallel translation DB 4, the example match determination unit 3 outputs the input sentence to the paraphrase sentence generation unit 5 and the machine translation unit 7.

The example bilingual translation DB4 is a database that stores one or more example bilingual translations in which an example sentence written in a first language and an example bilingual translation described in a second language are associated with each other. The example bilingual translation DB4 is an example of a database including a plurality of pairs of sentences written in the first language and bilingual sentences written in the second language. Specifically, the example bilingual translation DB4 is, for example, a database in which one record is assigned to one example bilingual translation, and includes an example sentence field and an example bilingual translation field. The example sentence is stored in the example sentence field, and the translated sentence corresponding to the example sentence is stored in the example parallel translation field. An example sentence is a sentence that has been used, and an example parallel translation is a translated sentence of an example sentence that has been translated. The translation reliability of the example sentence and the example parallel translation is, for example, 100%.

When the paraphrase sentence generation unit 5 determines that the input sentence is not stored in the example bilingual translation DB 4 by the example match determination unit 3, the paraphrase sentence generation unit 5 divides the input sentence acquired by the input unit 2 into a plurality of elements and a plurality of pieces. Multiple input paraphrases (plural second sentences) by paraphrasing (replacing) one or more of the elements with other expressions in the first language using a predetermined paraphrase rule (an example of a predetermined rule). An example) is generated.

Here, as a method of dividing the input sentence into a plurality of elementary pieces, for example, a method of dividing the input sentence into words is adopted. However, the present embodiment is not limited to this, and a method of dividing the input sentence by part of speech may be adopted, or a method of dividing the input sentence by a predetermined number of characters (for example, 2 characters or 3 characters) is adopted. Alternatively, a method of separating input sentences by phrase may be adopted, a method of separating input sentences by meaning class may be adopted, or a method of separating input sentences by morpheme may be adopted. good.

Further, as the paraphrase rule, the following first to fourth paraphrase rules can be adopted.

The first paraphrase rule is a rule that paraphrases the first word included in the element piece constituting the input sentence with the second word having a context-like relationship. Here, the context-similar relationship refers to the relationship between words that are similar in context, and for example, the relationship between words registered in the context-similar word database of the ALAGIN language resource can be adopted. For example, words that have a similar context to "Lupin III" include "Detective Conan" and "Space Battleship Yamato". In addition, "Brahms", "Schumann", "Mendelssohn" and the like correspond to words having a context-like relationship with "Tchaikovsky". In the context-similar word database of ALAGIN language resources, it is judged that "Lupin III" and "Lupin III" have a context-similar relationship, but in this embodiment, the contents are too close to each other, so the context. It may be excluded from the similarity.

The second paraphrase rule is a rule for paraphrasing the first word contained in the element piece constituting the input sentence into the second word having a co-occurrence relationship. Here, the co-occurrence relationship refers to the relationship between words that frequently appear in the same document, and corresponds to, for example, the relationship between words registered in the word co-occurrence frequency database of the ALAGIN language resource. For example, as words co-occurring with "overseas travel", "domestic travel", "cheap airline ticket", "tour", "airline ticket", and "travel" correspond in descending order of DICE coefficient. In addition, as words having a co-occurrence relationship of "Christmas", "New Year", "Birthday", "Santa", "Winter", and "Year-end" correspond in descending order of DICE coefficient. The DICE coefficient is an index that quantifies the similarity and co-occurrence between words.

The third paraphrase rule is a rule for paraphrasing the first word contained in the element piece constituting the input sentence into the second word having an implication relationship. Here, the implication relationship refers to a relationship in which the first word implies the second word, and for example, the relationship between words registered in the implication relation database of the ALAGIN language resource is applicable. The fact that the first word implies the second word means that if the situation represented by the first word is established, the situation represented by the second word is also established at the same time or before that. For example, "heat" for "tin", "detoxify" for "detox", "prowl" for "silver bra", "die" for "apoptosis", "die", " "Break" corresponds to "squid", and "drunk" corresponds to "drink". As for the implication relationship, the upper-lower relationship may be established, but the upper-lower relationship may not be established, such as "tinning" and "heating".

The fourth paraphrase rule is a rule for paraphrasing the first word contained in the element piece constituting the input sentence into the second word having a higher-lower relationship. Here, the hypernym-lower relationship refers to, for example, the relationship between words registered in the hypernym hierarchy database of the ALAGIN language resource. If the first word refers to something more general, more generic, and more abstract, including the second word, then the first word is superior to the second word.

FIG. 4 is a block diagram showing a detailed configuration of the paraphrase generation unit 5. The paraphrase sentence generation unit 5 includes a paraphrase DB storage unit 51 that stores a paraphrase DB (database), a paraphrase candidate generation unit 52, and a paraphrase sentence identification unit 53. The paraphrase DB is a database in which a word in the first language and a second word in which the first word is expressed by another expression in the first language are associated with each other.

In the present embodiment, the paraphrase DB storage unit 51 stores the context-like word DB511, the co-occurrence relation DB512, the implication relation DB513, and the superordinate relation DB514. Hereinafter, when the context-like word DB511, the co-occurrence relation DB512, the implication relation DB513, and the upper-lower relation DB514 are not particularly distinguished, they are described as paraphrase DB. The context-similar word DB511 is a database for paraphrasing an input sentence according to the first paraphrase rule, and is a database in which words having a context-similar relationship are stored in association with each other in advance. Here, as the context-like word DB511, for example, a context-like word database of ALAGIN language resources can be adopted.

The co-occurrence relationship DB512 is a database for paraphrasing an input sentence according to the second paraphrase rule, and is a database for storing words having a co-occurrence relationship in advance in association with each other. Here, as the co-occurrence relationship DB512, for example, a word co-occurrence frequency database of ALAGIN language resources can be adopted.

The implication relationship DB 513 is a database for paraphrasing an input sentence according to the third paraphrase rule, and is a database for storing words having an implication relationship in advance in association with each other. Here, as the implication relation DB 513, for example, an implication relation database of the ALAGIN language resource can be adopted.

The upper-lower relationship DB 514 is a database for paraphrasing an input sentence according to the fourth paraphrase rule, and is a database for storing words having a higher-lower relationship in advance in association with each other. Here, as the superordinate / subordinate relationship DB514, for example, a hypernym hierarchical database of ALAGIN language resources can be adopted.

The paraphrase candidate generation unit 52 paraphrases the input sentence according to each of the first to fourth paraphrase rules by referring to the paraphrase DB, and generates the input paraphrase sentence. Here, if the input sentence B1 "I want to make a taxi to Kadoma" is input, the paraphrase candidate generation unit 52 divides the input sentence B1 into words such as "Kadoma / to / taxi / to / want". do. Then, the paraphrase candidate generation unit 52 refers to each of the context-like word DB511, the co-occurrence relation DB512, the implication relation DB513, and the superordinate relation DB514, and inputs an input sentence in each of the first to fourth paraphrase rules. In other words, generate at least four input paraphrases.

Here, the paraphrase candidate generation unit 52 paraphrases one word to one when generating an input paraphrase sentence using the i (i = 1 to 4) paraphrase rule among the first to fourth paraphrase rules. An input paraphrase may be generated, or a word at a plurality of places may be paraphrased to generate one input paraphrase. Further, the paraphrase candidate generation unit 52 may generate a plurality of input paraphrase sentences having different numbers of word paraphrases when paraphrasing the input sentence using the i-paraphrase rule.

For example, the paraphrase candidate generation unit 52 randomly identifies one word from the divided input sentences, and if the same word as the identified one word is registered in the paraphrase DB, the one word is stored in the paraphrase DB. You can paraphrase it with another registered word that can be paraphrased. On the other hand, if the same word as the specified one word is not registered in the paraphrase DB, the paraphrase candidate generation unit 52 randomly identifies and identifies another one word other than the one word from the input sentence. If the same word as another one word is registered in the paraphrase DB, the other one word may be paraphrased with another paraphrasable word registered in the paraphrase DB. The paraphrase candidate generation unit 52 may repeat such a process to generate one or more input paraphrase sentences paraphrased according to the i-paraphrase rule.

For example, in the context-similar word DB511, if "bus" and "truck" are registered as words having a context-similar relationship with "taxi", one word is randomly selected from "bus" and "truck". You may decide and paraphrase "taxi" with the one word, or you may paraphrase "taxi" with the word most similar to "taxi".

The paraphrase candidate generation unit 52 may add additional data indicating the paraphrase location to the generated input paraphrase sentence and output it to the paraphrase sentence identification unit 53.

The paraphrase sentence identification unit 53 extracts a word-like sentence from the input paraphrase sentence and outputs it to the extraction unit 6. Here, the paraphrase identification unit 53 calculates the appearance probability of the input paraphrase using, for example, the N-gram language model, and outputs the input paraphrase whose appearance probability is equal to or greater than the reference value to the extraction unit 6. .. The N-gram language model is a probabilistic language model that models the "word-likeness" that humans would use as a probability. For example, if there is a sentence B2 that says "Today's supper is curry" and a sentence B3 that "Today's supper is baseball", it can be said that sentence B2 is more plausible than sentence B3. In this case, in the N-gram language model, the appearance probability of sentence B2 is higher than the appearance probability of sentence B3. Here, as the reference value, a value obtained empirically, which is a value determined to be an unnatural sentence when the appearance probability is further lowered, can be adopted. In addition, the paraphrase sentence identification unit 53 outputs to the extraction unit 6 including the additional data indicating the paraphrase location in the input paraphrase sentence to be output.

The reference is returned to FIG. The extraction unit 6 calculates a comprehensive evaluation value (an example of the degree of agreement) indicating the relationship between the input paraphrase sentence output from the paraphrase sentence generation unit 5 and the example sentence stored in the example bilingual translation DB4, and the calculated total. Based on the evaluation value, one or more example sentences (an example of the third sentence) are extracted from the example parallel translation DB4. Further, the extraction unit 6 extracts one or more input paraphrase sentences similar to the extracted example sentences from the input paraphrase sentences output from the paraphrase sentence generation unit 5. Hereinafter, the extracted input paraphrase sentence will be described as a "paraphrase extract sentence" (an example of one or more second sentences). Note that the relevance means that the input paraphrase sentence and the example sentence have a certain syntactical relationship.

Here, the extraction unit 6 calculates the comprehensive evaluation value of each input paraphrase sentence and each example sentence by using the following indexes A1 to A4.

The index A1 (an example of the third index) is an index showing the similarity of the sentence structure between each input paraphrase sentence and each example sentence. FIG. 5 is a diagram showing an example of an input paraphrase sentence and an example sentence.

With reference to FIG. 5, for example, as input paraphrases of input sentence B1 "I want to take a taxi to Kadoma", input paraphrase C1 "I want to take a taxi to Kadoma" and input paraphrase C2 "I want to use a bus to Kadoma" It is assumed that it is generated by the paraphrase generation unit 5.

Further, it is assumed that the example sentence D1 "I want to talk to the fullest" and the example sentence D2 "I want to go by train to Kyobashi" are stored in the example bilingual translation DB4.

First, the extraction unit 6 separates the input paraphrases C1 and C2 by clauses or words, analyzes the sentence structure of the input paraphrases C1 and C2, and generates a syntax tree. Here, in the example of the input paraphrase sentence C1, the phrase "to Kadoma" and the phrase "to taxi" are both related to the phrase "Noritai". Therefore, a tree containing two edges E11 and E12 that connect the node N11 corresponding to the clause "Kadoma" and the node N12 corresponding to the clause "Taxi" to the node N13 corresponding to the clause "Noritai", respectively. Structure T1 has been generated.

The generation of such a tree structure can be realized by using, for example, a parsing tool "KNP". Further, the analysis of the part of speech of the words constituting the sentence can be realized by using, for example, "juman" which is a morphological analysis tool. Therefore, the extraction unit 6 may use "KNP" and "juman" to generate the tree structure of the sentence and extract the part of speech of the words constituting the sentence.

In the example of the input paraphrase sentence C2, the phrase "to Kadoma" and the phrase "bus" are both related to the phrase "want to use". Therefore, a tree including two edges E21 and E22 that connect the node N21 corresponding to the clause "Kadoma" and the node N22 corresponding to the clause "Taxi" to the node N23 corresponding to the clause "I want to use", respectively. Structure T2 is generated.

In the example of example sentence D1, the phrase "to the fullest" and the phrase "to talk" are both related to the phrase "Noritai". Therefore, a tree containing two edges E31 and E32 that connect the node N31 corresponding to the clause "Tokonto" and the node N32 corresponding to the clause "Taxi" to the node N33 corresponding to the clause "Noritai", respectively. Structure T3 is generated.

In the example sentence D2, the phrase "to Kyobashi" and the phrase "by train" are both related to the phrase "I want to go". Therefore, a tree including two edges E41 and E42 that connect the node N41 corresponding to the clause "to Kyobashi" and the node N42 corresponding to the clause "by train" to the node N43 corresponding to the clause "I want to go", respectively. Structure T4 is generated.

In this way, the extraction unit 6 analyzes the tree structure of the input paraphrase sentence and the example sentence. Then, the extraction unit 6 may calculate the similarity of the tree structure between the input paraphrase sentence and the example sentence as the index A1 by using, for example, a tree matching method. The tree structure of the example sentence may be stored in advance in the example bilingual translation DB4. In the present embodiment, the index A1 takes a numerical value of 0 to 100%, and the higher the degree of matching of the tree structures, the larger the value.

In the example of FIG. 5, the tree structures T1 to T4 of the input paraphrases C1 and C2 and the example sentences D1 and D2 all have the same structure. Therefore, the extraction unit 6 calculates the index A1 for the example sentences D1 and D2 of the input paraphrase sentence C1 as 100%, respectively. Further, the extraction unit 6 also calculates the index A1 for the example sentences D1 and D2 of the input paraphrase sentence C2 as 100%, respectively.

In FIG. 5, the description in parentheses such as <uninflected word> and <phrase: action> is added for convenience to explain the part of speech of the corresponding phrase, and is actually used in the calculation of the index A1. It's not a thing.

The index A2 (an example of the fourth index) is an index showing a larger value as the number of matching parts of speech in each input paraphrase sentence and each example sentence increases.

In the present embodiment, the extraction unit 6 uses the phrase sentences located at the same position in the input paraphrase sentence and the example sentence having the same sentence structure (in the input paraphrase sentence and the example sentence in which the index A1 is 100%). The index A2 is calculated from the number of matching parts of speech.

In the example of FIG. 5, the extraction unit 6 calculates the index A2 so that in the input paraphrase sentence and the example sentence having the same sentence structure, the value becomes larger as the part of speech of the phrase located at the same place matches with the noun. .. Hereinafter, "phrases located in the same place" will be described as "corresponding clauses". Further, the "noun phrase" means a "phrase containing a noun". For example, the phrase "Kadoma" is composed of the word "Kadoma" and the word "Kadoma", but since the word "Kadoma" is a noun, "Kadoma" is a noun phrase.

In detail, the index A2 is defined by the following equation (1).

Index A2 = (1-α / β) × 100 (%) (1)
α: Number of corresponding clauses that are not nouns β: Total number of noun clauses in the input paraphrase sentence In the input paraphrase sentence C1 shown in FIG. .. In addition, the phrase "Tokonto" in the example sentence D1 corresponding to the phrase "Kadoma" in the input paraphrase C1 is an adverb, and the phrase "To the story" in the example sentence D1 corresponding to the phrase "Taxi" in the input paraphrase C1. Is a noun. Therefore, in the input paraphrase sentence C1 and the example sentence D1, β = 2 and α = 1, and the index A2 becomes 50%.

In addition, the phrase "to Kyobashi" in the example sentence D2 corresponding to the phrase "to Kadoma" in the input paraphrase C1 is a noun, and the phrase "by train" in the example sentence D2 corresponding to the phrase "to taxi" in the input paraphrase C1. Is a noun. Therefore, in the input paraphrase sentence C1 and the example sentence D2, β = 2 and α = 0, and the index A2 becomes 100%. Similarly, the indexes A2 of the input paraphrase sentence C2 and the example sentences D1 and D2 are 50% and 100%, respectively.

Note that α in equation (1) may be a number in which the corresponding clauses have different categories. Here, the category refers to a type to which a noun word belongs, such as a place name, traffic, or an abstraction. In FIG. 5, "Kadoma" and "Kyobashi" are applicable as place names, "taxi" and "bus" are applicable as transportation, and "story" is applicable as an abstraction.

When this aspect is adopted, for example, the phrase "taxi" in the input paraphrase C1 and the phrase "talk" in the example sentence D1 are both noun phrases, but the former category is "traffic" and the latter category. Since the category is "abstract", α is incremented by 1 and the index A2 is smaller than when the mode that does not consider the category is adopted.

Here, the extraction unit 6 calculates the index A2 for the input paraphrase sentence and the example sentence having the same sentence structure, but the present disclosure is not limited to this, and the presence or absence of the matching sentence structure is not considered. That is, the index A2 may be calculated independently of the index A1. Further, although the index A2 is calculated based on the number of matching noun clauses, the index A2 may be calculated based on the number of matching part of speech.

For example, suppose that there is an input paraphrase sentence C1X composed of "phrase C11 / clause C12 / clause C13 / clause C14". Further, it is assumed that there is an example sentence D1X composed of "phrase D11 / clause D12 / clause D13". In addition, "/" indicates a break of a phrase. In this case, the extraction unit 6 extracts the clauses located in the same order from the beginning in the input paraphrase sentence C1X and the example sentence D1X as corresponding clauses, and is based on the number of matching part of speech between the extracted clauses. The index A2 may be calculated.

For example, the extraction unit 6 extracts three phrase pairs of "clause C11" and "clause D11", "clause C12" and "clause D12", and "clause C13" and "clause D13" as corresponding clauses. do. Note that "clause C14" is excluded from the extraction target because there is no clause corresponding to the example sentence D1X. Then, the extraction unit 6 may calculate the index A2 using the equation (1) with the total number of clause pairs whose part of speech does not match as α and the number of clauses extracted from the input paraphrase C1X as β.

The index A3 (an example of the second index) is an index showing a larger value as the text similarity between the example sentence and the input sentence whose sentence structure matches the input paraphrase sentence is smaller.

In detail, the index A3 is defined by the equation (2).

Index A3 = 100-Text similarity between input sentences and example sentences (2)
First, the extraction unit 6 extracts an example sentence whose sentence structure matches the input paraphrase sentence from the example parallel translation DB 4. Then, the extraction unit 6 calculates the text similarity between the extracted example sentence and the input sentence, and calculates the index A3 for each example sentence so that the smaller the calculated text similarity is, the larger the value is.

When the sentence structures match, it means that the tree structures match as described above, that is, the index A1 is 100%. The text similarity indicates how much the contents of sentences such as the expression and the character face of the sentence match, and is calculated using, for example, the similla_text function of the PHP language that calculates the similarity between two character strings. NS.

For example, suppose that the input paraphrase C1 "I want to take a taxi to Kadoma" is generated as the input paraphrase for the input sentence B1 "I want to take a taxi to Kadoma". In this case, the extraction unit 6 extracts an example sentence having the same sentence structure as the input paraphrase sentence C1 from the parallel translation example DB4. Here, example sentence D2 "I want to go to Kyobashi by train", example sentence D3 "I want to use a car to Moriguchi", example sentence D4 "I want to go to Tokyo by Shinkansen", and example sentence D5 "I want to talk to the whole story" It is assumed that the four example sentences of are extracted.

In this case, the extraction unit 6 calculates the respective indexes A3 of the input sentence B1 and the four example sentences D2 to D5 by using the equation (2).

In the above description, when calculating the index A3, the extraction unit 6 extracts an example sentence in which the input paraphrase sentence and the sentence structure match, but the present disclosure is not limited to this, and the input paraphrase sentence and the sentence structure are similar. Example sentences may be extracted. Here, the fact that the sentence structures are similar corresponds to, for example, the case where the index A1 is equal to or more than the reference value. As the reference value, a value larger than at least 50% such as 50%, 60%, 70%, 80%, and 90% can be adopted.

The index A4 (an example of the first index) is an index indicating the text similarity between the input paraphrase sentence and the example sentence. The text similarity is the same as the text similarity used when calculating the index A3.

In the present embodiment, the extraction unit 6 extracts an example sentence whose sentence structure matches the input paraphrase sentence, that is, an example sentence in which the index A1 is 100% from the bilingual example DB4, and extracts the extracted example sentence and the input paraphrase sentence. Each text similarity with and is calculated as an index A4.

For example, assuming that the above input paraphrase sentence C1 "I want to take a taxi to Monshin" is generated, the extraction unit 6 extracts an example sentence whose sentence structure matches that of the input paraphrase sentence C3 from the bilingual example DB4. Here, the four example sentences D2 to D5 described in the index A3 are extracted. In this case, the extraction unit 6 may calculate the text similarity between the input paraphrase sentence C3 and the example sentences D2 to D5 as the index A4.

The extraction unit 6 may calculate the value of the index A4 larger as the input paraphrase sentence has more paraphrase points. For example, the extraction unit 6 may calculate the final index A4 by multiplying the text similarity by the paraphrase rate. As the paraphrase rate, for example, the ratio of the number of paraphrased characters to the total number of characters in the input paraphrase sentence can be adopted.

Then, the extraction unit 6 calculates, for example, the product of the indexes A1 to A4 of each example sentence as the comprehensive evaluation value of each example sentence. Then, the extraction unit 6 extracts n (integer of 1 or more) example sentences in descending order of the comprehensive evaluation value.

In addition, the extraction unit 6 may extract an example sentence whose comprehensive evaluation value is larger than a reference value (an example of a threshold value) from each example sentence. Alternatively, the extraction unit 6 extracts from each example sentence an example sentence whose overall evaluation value is larger than the reference value, and if the number of extracted example sentences is n or more, n example sentences in descending order of the overall evaluation value. May be extracted.

Then, the extraction unit 6 extracts the example translations of the extracted n example sentences from the example translation DB4 and outputs them to the output unit 9.

When the process of extracting the example sentences is completed, the extraction unit 6 performs a process of extracting n paraphrase extraction sentences similar to the extracted n example sentences. Here, the extraction unit 6 extracts the input paraphrase sentence having the maximum index A4 for each of the extracted n example sentences, thereby extracting as n paraphrase extraction sentences.

For example, assuming that four input paraphrases C1 to C4 are output from the paraphrase generation unit 5 and two example sentences D1 and D2 are extracted from the comprehensive evaluation value, the extraction unit 6 uses the example sentences D1 and D2. For each, the index A4 of each of the input paraphrases C1 to C4 is calculated. Then, the extraction unit 6 extracts the input paraphrase sentence having the maximum index A4 in each of the example sentences D1 and D2 as the paraphrase extraction sentence.

In the above description, the extraction unit 6 has calculated the comprehensive evaluation value using all of the indicators A1 to A4, but the present disclosure is not limited to this, and the extraction unit 6 uses at least one of the indicators A1 to A4. The comprehensive evaluation value may be calculated. Further, the extraction unit 6 has adopted the product of the indicators A1 to A4 as the comprehensive evaluation value, but the present disclosure is not limited to this, and the extraction unit 6 comprehensively evaluates the average value and the weighted average value of the indicators A1 to A4. It may be adopted as a value.

The machine translation unit 7 generates n paraphrase translation sentences (an example of the fifth sentence) by machine translating each of the n paraphrase extraction sentences output from the extraction unit 6 into a second language. Here, the machine translation unit 7 performs machine translation by using some kind of translation engine. For example, the machine translation unit 7 may use the translation engine provided on the web site, or may use the translation application software provided by the translation support device 1 itself. Further, the machine translation unit 7 machine-translates the input sentence output from the example match determination unit 3 to generate the input translation sentence.

The reliability imparting unit 8 calculates the translation reliability of n paraphrase translations generated by the machine translation unit 7. Here, the reliability imparting unit 8 may calculate the translation reliability from the degree of coincidence with the corresponding paraphrase extract sentence when the paraphrase translation sentence is back-translated from the second language to the first language. In addition, the reliability imparting unit 8 also calculates the translation reliability of the input translated sentence.

The output unit 9 displays n example parallel translations (an example of the fourth sentence) extracted by the extraction unit 6. Further, the output unit 9 displays the n paraphrase translations generated by the machine translation unit 7 and the corresponding n input paraphrases together with the translation reliability. Further, the output unit 9 displays the input sentence and the input translation sentence together with the translation reliability.

When the output unit 9 determines that the input sentence matches any of the example sentences stored in the example parallel translation DB 4, the example match determination unit 3 displays the example parallel translation output from the example match determination unit 3. Just do it.

FIG. 7 is a diagram showing an example of the output image 500 displayed by the output unit 9. The output image 500 includes an input sentence display field 510 and a translation support information display field 520. The input sentence display field 510 displays the input sentence 511 "I want to take a taxi to Kadoma" and the input translation sentence 512 "I want to taxi to kadoma" which is the machine translation result of the input sentence 511 side by side. Further, in the input sentence display field 510, a reliability display field 513 for displaying the translation reliability of the input translation sentence 512 is also displayed. Here, since the degree of agreement between the reverse translation result of the input translated sentence 512 and the input sentence 511 was 70%, "70%" is displayed in the reliability display column 513.

The translation support information display field 520 is a field for displaying an example sentence or the like related to the input sentence 511. Here, since two example sentences are extracted by the extraction unit 6 as example sentences for the input sentence 511, two translation support information display fields 530 and 540 corresponding to the two example sentences 531a and 541a are displayed. Further, since the example sentence 531a had a higher overall evaluation value than the example sentence 541a, the translation support information display field 530 corresponding to the example sentence 531a was higher than the translation support information display field 540 corresponding to the example sentence 541a. It is displayed on the upper side.

The translation support information display field 530 includes a reference example display field 531 with the heading "Reference example 1" and a reference translation display field 532 with the heading "Reference translation 1".

In the reference example display column 531, the example sentence 531a "I want to go to Kyobashi by train" and the corresponding example translation 531b "I want to go train to kyobashi" are displayed side by side. There is.

In the reference translation display column 532, there is a paraphrase extraction sentence 532a "I want to go by train to Kadoma" with the maximum text similarity (index A4) with respect to the example sentence 531a, and a paraphrase translation sentence 532b "I want to go" corresponding to it. "by train to kadoma" is displayed side by side.

Further, in the translation support information display column 530, a reliability display column 533 indicating the translation reliability of the paraphrase translation sentence 532b is displayed. Here, since the degree of agreement between the reverse translation result of the paraphrase translation sentence 532b and the paraphrase extract sentence 532a was 95%, it is displayed as "95%".

Further, in the paraphrase extraction sentence 532a, the part "I want to go by train" is underlined, and the paraphrase part for the input sentence 511 is displayed so as to be distinguishable from other parts. Further, also in the paraphrase translation sentence 532b, the part of "I want to go by train" is underlined, and the translation result of the paraphrase part is displayed so as to be distinguishable from the translation result of other parts.

As a result, the user can recognize at a glance the paraphrase portion for the input sentence 511 in the paraphrase extract sentence 532a and the paraphrase translation sentence 532b.

Similar to the translation support information display field 530, the translation support information display field 540 also displays the reference example display field 541 and the reference translation display field 542.

In the reference example display column 541, the example sentence 541a "I want to use the car to Moriguchi" and the corresponding example translation 541b "I want to take a car to moriguchi" are displayed. ..

In the reference translation display column 542, there is a paraphrase extraction sentence 542a "I want to use the bus to Kadoma" with the maximum text similarity (index A4) with respect to the example sentence 541a, and a paraphrase translation sentence 542b "I want to take" corresponding to it. "the bus to kadoma" is displayed side by side.

In the paraphrase extraction sentence 542a, the paraphrase part for the input sentence 511 is "I want to use the bus", so that part is underlined. Further, in the paraphrase translation sentence 542b, the translation portion "I want to take the bus" corresponding to the paraphrase portion is underlined. Further, since the degree of agreement between the reverse translation result of the paraphrase translation sentence 542b and the paraphrase extract sentence was 90%, "90%" is displayed in the reliability display column.

In this way, since the output image 500 is displayed on the upper side of the translation support information display field 520 including the example sentences having a high overall evaluation value, the user can use the translation support including the highly important example parallel translations and paraphrase translations. Information can be recognized at a glance.

In the example of FIG. 7, two translation support information display fields 520 are shown, but this is an example. If three or more example sentences are extracted by the extraction unit 6, the output image 500 is The translation support information display field 520 including three or more example sentences may be displayed. In this case as well, the translation support information display field 520 may be displayed so that the example sentence having the higher overall evaluation value is located on the upper side.

Further, in the example of FIG. 7, the paraphrase portion (character string) is highlighted by using an underline, but the present disclosure is not limited to this, and a marker is attached to the background of the paraphrase portion to highlight the paraphrase portion. May be adopted, the color of the characters in the paraphrased part may be changed to the color of the characters in the non-paraphrased part and highlighted, or the paraphrased part may be highlighted in bold. Alternatively, a mode in which these modes are combined may be adopted. Further, in the present disclosure, the paraphrased portion may not be highlighted and the non-paraphrased portion may be highlighted.

Further, in the example of FIG. 7, the example sentence and the example translation are not particularly highlighted, but the present disclosure is not limited to this, and the part (character) of the example sentence and the example translation corresponding to the paraphrase extraction sentence is not limited to this. Column) may be highlighted.

Next, a specific example of the processing of the extraction unit 6 in the translation support device 1 will be described. Here, it is assumed that the input sentence (I) "I want to take a taxi to Kadoma" is input, and the following three input paraphrase sentences are generated by the paraphrase sentence generation unit 5. In this example, it is assumed that (A) to (C) all have the same sentence structure, that is, the same tree structure.

(A) "I want to go to Kadoma by train"
(B) "I want to use the bus to Kadoma"
(C) "I want to take a taxi to Kadoma"
Further, the following four example sentences (1) to (4) having the same sentence structure with respect to the above input paraphrase sentences (A) to (C) are extracted from the example bilingual translation DB4 when the index A1 is 100%. Suppose it was done. In this specific example, the extraction unit 6 extracts an example sentence having the same sentence structure as the input paraphrase sentence from the example parallel translation DB4, and calculates indexes A2 to A4 for the extracted example sentence.

(1) "I want to go to Kyobashi by train"
(2) "I want to use a car to Moriguchi"
(3) "Go to Tokyo by Shinkansen"
(4) "I want to talk to you thoroughly"
Next, the extraction unit 6 calculates the index A2 using the above formula (1) for each of the example sentences (1) to (4). In this specific example, since the input paraphrase sentences (A) to (C) have the same sentence structure, the input paraphrase sentences (A) are represented, and the example sentences (1) to (4) and the input paraphrase sentences ( Calculate the index A2 with A).

Since the total number of noun clauses in the input paraphrase (A) is "to Kyobashi" and "by train", β = 2.

Further, in the example sentences (1) to (3), α = 0 and the index A2 = 100% because the number of corresponding phrases that are not nouns is 0 with respect to the input paraphrase sentence (A). On the other hand, in the example sentence (4), the noun phrase "story" corresponding to the noun phrase "train" in the input paraphrase sentence (A) is the noun, but the noun phrase "Kadoma" in the input paraphrase sentence (A) The phrase "to the fullest" corresponding to "" is not a noun. Therefore, in the example sentence (4), the number of corresponding phrases that are not nouns is one with respect to the input paraphrase sentence (A). Therefore, in the example sentence (4), α = 1 and the index A2 = (1-1 / 2) × 100 = 50%. Therefore, as shown in Table H1 of FIG. 6, the indexes A2 of the example sentences (1) to (4) are "100%", "100%", "100%", and "50%", respectively. There is. FIG. 6 is Table H1 summarizing specific examples in the present embodiment. In Table H1, the indexes A1 to A4 for the example sentences (1) to (4) are calculated.

Next, the extraction unit 6 calculates the index A3 of the example sentences (1) to (4) and the input sentence (I) by using the above formula (2). Here, as shown in Table H1, the indicators A3 of the example sentences (1) to (4) are "36.8%", "41.1%", "61.8%", and "55. It was calculated as "8%". As a result, among the example sentences (1) to (4), the example sentence (3) "Go to Tokyo by Shinkansen" has the lowest text similarity to the input sentence (I) "I want to take a taxi to Kadoma". That is, it can be seen that the meaning and content are the farthest.

In this specific example, the index A3 of the input paraphrases (B) and (C) has the same value as the index A3 of the input paraphrase (A). This is because, for the input paraphrase sentences (B) and (C), the example sentences (1) to (4) are used when calculating the index A3.

In this way, by extracting a large example sentence of the index A3, it is possible to extract an example sentence having a sentence structure similar to that of the input sentence but having a different meaning and content. As a result, various example translations can be presented to the user.

Next, the extraction unit 6 calculates the indexes A4 of the input paraphrase sentences (A) to (C) and the example sentences (1) to (4), respectively. In this specific example, 3 × 4 = 12 indexes A4 are calculated, and their respective values are as shown in Table H1.

Next, the extraction unit 6 calculates the comprehensive evaluation value “%” of the example sentences (1) to (4) from the index A1 × index A2 × index A3 × index A4. In this specific example, higher (larger) comprehensive evaluation values are obtained in the order of example sentences (1) to (4). In this specific example, the example sentences (1) to (4) have the same sentence structure as the input paraphrase sentences (A) to (C), so that the index A1 of the example sentences (1) to (4) is All are said to be 100%.

Next, the extraction unit 6 extracts the top n example sentences in descending order of the overall evaluation value, and extracts n example translations including the extracted n example sentences from the example translation DB4. For example, if n = 2, the extraction unit 6 extracts two example translations including the example sentences (1) and (2).

Next, the extraction unit 6 extracts the input paraphrase sentence having the maximum index A4 (text similarity) as the paraphrase extraction sentence in the extracted example sentence. Here, since the example sentences (1) and (2) are extracted, the input paraphrase sentence (A) having the maximum index A4 in the example sentence (1) and the maximum index A4 in the example sentence (2) are the maximum. The input paraphrase sentence (B) of is extracted as a paraphrase extract sentence.

Next, the flowchart of the translation support device 1 will be described. FIG. 8 is a flowchart showing an example of processing of the translation support device 1 according to the embodiment of the present disclosure.

First, the input unit 2 receives an operation from the user and acquires an input sentence (S1). Here, for example, it is assumed that the input sentence (I) "Take a taxi to Kadoma" is acquired.

Next, the example match determination unit 3 determines whether or not an example sentence matching the input sentence (I) is stored in the example parallel translation DB 4 (S2). Here, if there is an example sentence matching the input sentence (I) in the example parallel translation DB4 (YES in S2), the example match determination unit 3 extracts the example parallel translation of the matching example sentence from the example parallel translation DB4 and outputs the output unit. 9 displays the extracted example parallel translation (S3).

On the other hand, if the example sentence matching the input sentence (I) is not stored in the example bilingual translation DB4 (NO in S2), the process proceeds to S4.

In S4, the paraphrase sentence generation unit 5 generates a plurality of input paraphrase sentences by paraphrasing the input sentence (I) using the above-mentioned first to fourth paraphrase rules (S4). As a result, for example, the above-mentioned input paraphrase sentences (A) to (C) are generated.

Next, the extraction unit 6 calculates the above-mentioned comprehensive evaluation value by comparing the input paraphrase sentences (A) to (C) with the example sentences stored in the example parallel translation DB4, and the calculated comprehensive evaluation value is obtained. By extracting n example sentences in descending order, n example translations are extracted (S5). As a result, for example, the above-mentioned two example sentences (1) and (2) and the example bilingual translation including them are extracted.

Next, the extraction unit 6 extracts n paraphrases from the input paraphrases generated in S4 by extracting the input paraphrases having the maximum text similarity for each of the n example sentences extracted in S5. Extract the sentence (S6). As a result, for example, the above-mentioned two input paraphrase sentences (A) and (B) are extracted as paraphrase extraction sentences.

Next, the machine translation unit 7 machine-translates the n paraphrase extract sentences extracted in S6 to generate n paraphrase translation sentences and machine translates the input sentence acquired in S1. Then, an input translated sentence is generated (S7). As a result, for example, the above-mentioned two input paraphrase sentences (A) and (B) paraphrase translation sentences and input translation sentences are generated.

Next, the reliability imparting unit 8 calculates the translation reliability of the input translation sentence and n paraphrase translation sentences generated in S7 (S8). Next, the output unit 9 displays the output image 500 of the translation result including the example translation extracted in S5, the input translation sentence and the paraphrase translation sentence generated in S7, and the translation reliability calculated in S8. (S9).

FIG. 9 is a flowchart showing an example of details of the process of S5 of FIG. In the loop L5, the processing of S501 is performed for each of the pair of the input paraphrase sentence of 1 out of all the input paraphrase sentences generated in S4 and the example parallel translation of 1 of all the example translations stored in the example translation DB4. It is a loop that repeats. The loop L5 ends when the processing of S501 is executed for all the input paraphrases generated in S4 and all the example translations stored in the example translation DB4.

In S501, the extraction unit 6 calculates the indexes A1 to A4 for the set of the input paraphrase sentence of 1 and the example sentence of 1. Further, in S501, the extraction unit 6 calculates the comprehensive evaluation value for the set of the calculated indexes A1 to A4 to 1.

Assuming that the input paraphrases (A) to (C) are generated in S4, first, for the input paraphrase (A), the comprehensive evaluation value of each of the input paraphrases (A) and all the example sentences stored in the example parallel translation DB4 is calculated, and then In addition, the comprehensive evaluation value of each of the input paraphrase sentences (B) and all the example sentences stored in the example bilingual translation DB4 is calculated, and then the input paraphrase sentence (C) is all stored in the example bilingual translation DB4. Each comprehensive evaluation value with the example sentence is calculated.

In S502, the extraction unit 6 extracts n example sentences having the highest overall evaluation value and n example translations corresponding to the n example sentences.

FIG. 10 is a flowchart showing an example of details of the process of S6 of FIG. The loop L61 is a loop executed for each example sentence (i) of one of the n example sentences extracted in S5. i is an index that identifies one of the n example sentences, and is an integer of 1 or more and n or less. The end condition of the loop L61 is that the processing for n example sentences is completed, that is, i = n. The loop L62 is a loop in which the processes of S601 to S602 are repeated for each pair of the example sentence (i) of 1 and the all-input paraphrase sentence in the loop of 1 of the loop L61. The termination condition of the loop L62 is that the processing of S601 to S602 is terminated for each of the example sentence (i) of 1 and all the input paraphrase sentences.

In S601, the extraction unit 6 calculates the index A4 of the example sentence (i) of 1 and the input paraphrase sentence of 1 of all the input paraphrase sentences. Next, if the calculated index A4 is the largest of the example sentences (i) of 1 (YES in S601), the extraction unit 6 holds the input paraphrase sentence as the paraphrase extraction sentence (i) in the memory ( S602).

On the other hand, if the calculated index A4 is not the maximum among the example sentences (i) of 1, the processing of S602 is not performed and the loop L62 is continued. By repeating the loop L62, the input paraphrase sentence (i) having the maximum index A4 is determined from all the input paraphrase sentences for the example sentence (i) of 1. Then, the loop L61 extracts n input paraphrase sentences (i) having the maximum index A4 for n example sentences (i).

For example, it is assumed that the input paraphrase sentences (A) to (C) are generated in S4 and the example sentences (1) to (4) are extracted in S5. In this case, first, regarding the example sentence (1), the input paraphrase sentence having the maximum index A4 is extracted as the paraphrase extraction sentence (1) from the input paraphrase sentences (A) to (C), and then the example sentence ( Regarding 2), four paraphrase extraction sentences are extracted so that the input paraphrase sentence having the maximum index A4 is extracted as the paraphrase extraction sentence (2) from the input paraphrase sentences (A) to (C). NS.

As described above, according to the present embodiment, the translated sentence of the input paraphrase sentence is not simply presented. That is, in the present embodiment, among the example sentences stored in the example bilingual translation DB4, n example sentences whose comprehensive evaluation value for the input paraphrase sentence is equal to or more than the reference value are extracted, and the extracted n example sentences are extracted. N similar input paraphrases are extracted. Then, n paraphrase translations obtained by machine-translating the extracted n input paraphrases and n example parallel translations of the extracted n example sentences are presented.

As a result, the knowledge space used when generating a translated sentence of an input sentence or a similar sentence thereof is widely used, and a translation result useful for the user can be presented.

Further, since the present embodiment does not require that a translated sentence of an input sentence or a similar sentence is generated with high reliability, a knowledge space having a wide range of abundant knowledge data that can meet the request is provided. There is no need to use it. Therefore, this embodiment can present a translation result useful for the user without enhancing the knowledge space.

Further, in the present embodiment, since the translated sentence of the input paraphrase sentence similar to the extracted example sentence is presented, it is possible to prevent the translated result of the input paraphrase sentence which is less related to the input sentence is presented. ..

The following aspects can be adopted in the present disclosure.

(1) In the above embodiment, the output unit 9 displays the translation result of the example parallel translation and the input paraphrase sentence using the image as shown in the output image 500, but the present disclosure is not limited to this, and the output is not limited to this. The unit 9 may output the content included in the output image 500 by voice. In this case, the output unit 9 is composed of a speaker.

(2) The output image 500 shown in FIG. 7 is an example, and in the present disclosure, any item shown in FIG. 7 may be omitted from the output image 500. For example, in the translation support information display field 530, the reference example display field 531 may be omitted, or the reference translation display field 532 may be omitted.

(3) In the output image 500 shown in FIG. 7, when all the translation support information display fields 520 cannot be displayed at once, the output unit 9 may scroll the output image 500. As a result, when the display area of the display device is small, it is possible to prevent the user from being unable to view all the translation support information display fields 520.

(4) In the output image 500 shown in FIG. 7, the paraphrase extract sentence (one or more second sentences) and the paraphrase translation sentence (fifth sentence) may not be displayed.

This disclosure is useful for the technical field of providing an automatic translation service because it is possible to present a translated sentence useful to the user without enhancing the knowledge space.

A1, A2, A3, A4 Index B1 Input sentence C1, C2, C3 Input paraphrase sentence D1, D2, D3, D4, D5 Example sentence 511 Context-like word DB
512 Co-occurrence relational database
513 Implications DB
514 Upper-lower relational database
1 Translation support device 2 Input unit 3 Example match judgment unit 4 Example translation DB
5 Paraphrase generation unit 6 Extraction unit 7 Machine translation unit 8 Reliability addition unit 9 Output unit 51 Paraphrase DB storage unit 52 Paraphrase candidate generation unit 53 Paraphrase sentence identification unit 500 Output image 510 Input sentence display field 520, 530, 540 Translation support Information display field

Claims (13)

It ’s a way to provide a translation,
Obtain the first sentence written in the first language to be translated via the user's terminal,
It is determined whether or not the first sentence is included in the database containing a plurality of pairs of the sentence described in the first language and the bilingual sentence described in the second language.
When it is determined that the first sentence is not included in the database, a plurality of second sentences in which one or more words constituting the first sentence are replaced based on a predetermined rule are generated.
The degree of syntactic matching between the plurality of second sentences and the plurality of sentences described in the first language included in the database is calculated.
One or more third sentences described in the first language included in the database whose calculated degree of matching is equal to or more than the threshold value are extracted.
In the database, one or more fourth sentences described in the second language, which is a bilingual sentence of the one or more third sentences, are extracted.
One or more of the second sentences of the plurality of second sentences are machine-translated into the second language to generate one or more fifth sentences.
At least one of the one or more fourth sentences and the one or more fifth sentences is displayed on the user's terminal as a parallel translation reference of the first sentence.
Method. The degree of agreement is calculated based on a first index indicating the text similarity between the plurality of second sentences and the plurality of sentences included in the database.
The method according to claim 1. The degree of coincidence is a sentence having a sentence structure that matches or is similar to the plurality of second sentences among a plurality of sentences included in the database, and a sentence having a smaller text similarity with the first sentence. Calculated based on a second index that shows a large value,
The method according to claim 1 or 2. The degree of agreement is calculated based on a third index indicating the similarity of the sentence structure between the plurality of second sentences and the plurality of sentences included in the database.
The method according to any one of claims 1 to 3. The degree of matching is calculated based on a fourth index showing a larger value as the number of matching parts of speech increases in the plurality of second sentences and the plurality of sentences included in the database.
The method according to any one of claims 1 to 4. The first index shows a larger value as the second sentence has more replacement points.
The method according to claim 2. The one or more second sentences are extracted from the plurality of second sentences based on the text similarity between the plurality of second sentences and the one or more third sentences.
The method according to claim 1. The predetermined rule is a first paraphrase rule in which the first word contained in the element piece constituting the first sentence is paraphrased by the second word having a context-like relationship.
The first paraphrase rule is defined by a context synonym database in which a predetermined word is associated with a word having a context-like relationship with the predetermined word.
When the first word is registered by referring to the context synonym database, the word associated with the first word in the context synonym database is specified as the second word.
The method according to any one of claims 1 to 7. The predetermined rule is a second paraphrase rule that paraphrases the first word contained in the element piece constituting the first sentence into the second word having a co-occurrence relationship.
The second paraphrase rule is defined by a co-occurrence relational database in which a predetermined word and a word having a co-occurrence relationship with the predetermined word are associated with each other.
When the first word is registered by referring to the co-occurrence relational database, the word associated with the first word in the co-occurrence relational database is specified as the second word.
The method according to any one of claims 1 to 8. The predetermined rule is a third paraphrase rule that paraphrases the first word contained in the element piece constituting the first sentence into the second word having an implication relationship.
The third paraphrase rule is defined by an implication relation database in which a predetermined word is associated with a word having an implication relationship with the predetermined word.
When the first word is registered with reference to the implication relation database, the word associated with the first word in the implication relation database is specified as the second word.
The method according to any one of claims 1 to 9. The predetermined rule is a fourth paraphrase rule that paraphrases the first word contained in the element piece constituting the first sentence into the second word having a higher-lower relationship.
The fourth paraphrase rule is defined by an upper-lower relational database in which a predetermined word is associated with a word having a higher-lower relationship with the predetermined word.
When the first word is registered by referring to the upper-lower relational database, the word associated with the first word in the upper-lower relational database is specified as the second word.
The method according to any one of claims 1 to 10. A device that provides translated text
A database including a plurality of pairs of the described translated sentence in the description sentence and the second language in a first language,
An input unit that acquires the first sentence written in the first language to be translated via the user's terminal, and
When the first sentence is not included in the database, a paraphrase sentence generation unit that generates a plurality of second sentences in which one or more words constituting the first sentence are replaced based on a predetermined rule, and a paraphrase sentence generation unit.
The degree of syntactic matching between the plurality of second sentences and the plurality of sentences described in the first language included in the database is calculated, and the degree of matching is included in the database having a threshold value or more. An extraction unit that extracts one or more third sentences described in the first language,
In the database, one or more fourth sentences described in the second language, which is a bilingual sentence of the one or more third sentences, are extracted, and one or more second sentences out of the plurality of second sentences are described. Machine translation into a second language is performed to generate one or more fifth sentences, and at least one of the one or more fourth sentences and the one or more fifth sentences is used as a parallel translation reference for the first sentence. A device including a presentation unit to be displayed on the user's terminal. A program for causing a computer to execute the method according to any one of claims 1 to 11.

Images