JP7442324B2 - Machine translation devices and programs - Google Patents

Description

The present invention relates to a machine translation device and a program.

In various fields and applications, there is an increasingly strong need to improve the accuracy of machine translation processing. In neural machine translation, when machine learning is performed using training data that combines data from different domains or data that contains noise, translation accuracy may decrease. It is important to avoid deterioration in translation accuracy due to such factors.

For example, in Non-Patent Document 1 and Non-Patent Document 5, after machine learning is performed using learning data that combines data from different domains, learning data consisting only of data from the domain that you actually want to translate is used. A method of relearning has been proposed.

Furthermore, Non-Patent Document 2, Non-Patent Document 3, and Non-Patent Document 4 propose a method of performing machine learning by adding a domain tag representing the characteristics of a domain. This method can be considered as a method of domain adaptation.

Rico Sennrich, Barry Haddow, Alexandra Birch, Improving Neural Machine Translation Models with Monolingual Data, In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, p.86-96, Association for Computational Linguistics, 2016. Catherine Kobus, Josep Crego, Jean Senellart, Domain Control for Neural Machine Translation, In Proceedings of the International Conference Recent Advances in Natural Language Processing, RANLP 2017, p. 372-378, Varna, Bulgaria, September, 2017. Alexandre Berard, Ioan Calapodescu, Marc Dymetman, Claude Roux, Jean-Luc Meunier, Vassilina Nikoulina, Machine Translation of Restaurant Reviews: New Corpus for Domain Adaptation and Robustness, In Proceedings of the 3rd Workshop on Neural Generation and Translation (WNGT 2019), p. 168-176, Hong Kong, November, Association for Computational Linguistics, 2019. Alexandre Berard, Ioan Calapodescu, Claude Roux, Naver Labs Europe's Systems for the WMT19 Machine Translation Robustness Task, In Proceedings of the Fourth Conference on Machine Translation (Volume 2: Shared Task Papers, Day 1), p. 526-532, Florence, Italy, August, Association for Computational Linguistics, 2019. Yonatan Belinkov, Yonatan Bisk, Synthetic and Natural Noise Both Break Neural Machine Translation, In 6th International Conference on Learning Representations, Conference Track Proceedings, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018.

However, all of the above conventional methods have problems. The method described in Non-Patent Document 1 and Non-Patent Document 5 (learning with data that combines different domains and then re-learning with data of only a specific domain) is a method that strengthens the knowledge of a specific domain through re-learning. This is what we aim to do. However, there is a problem in that the learning results for only a specific domain and the learning results for different domains are not well separated, leading to a decrease in translation accuracy.

In addition, in the methods described in Non-patent Document 2, Non-patent Document 3, and Non-patent Document 4 (learning by adding domain tags that represent the characteristics of the domain), depending on the corpus used as learning data, appropriate tags may be There is a problem that it cannot be attached. In addition, although the characteristics of the corpus used are reflected in the learning results, there is a problem that the characteristics of the corpus cannot be expressed with tags.

The present invention was made based on the above-mentioned problem recognition, and aims to provide a machine translation device and a program that can improve the efficiency of machine learning and improve translation accuracy.

[1] In order to solve the above problems, a machine translation device according to one aspect of the present invention includes a translation target sentence written in a source language, a source language feature tag representing a feature of the sentence in the source language, and a target sentence. A target language side feature tag representing the feature of a sentence on the language side, and a translation written in the target language corresponding to the translation target sentence, the source language side feature tag, and the target language side feature tag. A translation model that holds information on the relationship between the translation target sentence, the source language feature tag, the target language feature tag, and the translation result sentence, and a translation model that is written in the source language so as to output the result sentence. A sentence pair consisting of a source language side sentence and a target language side sentence written in the target language is provided with a source language side feature tag according to the type of corpus to which the source language side sentence belongs, and the target language side sentence. a learning data tag assigning unit that assigns a target language side feature tag according to the type of corpus to which the training data tag attaches; The present invention includes a learning processing unit that supplies a set of sentence pairs to the translation model as learning data for use in machine learning processing.

[2] Further, one aspect of the present invention is the machine translation device according to [1] above, wherein the translation model is configured using a neural network.

[3] In addition, in the above machine translation device, one aspect of the present invention is to provide a translation target sentence written in the source language with a source language feature tag representing a feature of the sentence in the source language and a target language feature tag. a target language side feature tag representing a desired sentence feature of the sentence, and supply the translation target sentence to which the source language side feature tag and the target language side feature tag are attached to the translation model for translation processing. The present invention further includes a translation target sentence tag adding unit.

[4] In addition, one aspect of the present invention is that in the above machine translation device, the translation model further inputs a noise-related feature tag regarding noise between the sentence in the source language and the sentence in the target language. The translation target sentence, the source language side feature tag, the noise related feature tag, the target language side feature tag, and the translated result sentence are configured to output the translated result sentence corresponding to the noise related feature tag as well. The learning data tagging unit further adds the predetermined noise-related feature tag to the sentence pair, and the learning processing unit further The set of sentence pairs to which the noise-related feature tag has been added is supplied to the translation model as the learning data.

[5] Further, one aspect of the present invention is that in the machine translation device according to [4] above, a source language side feature tag representing a feature of the sentence on the source language side is added to the translation target sentence written in the source language; A target language feature tag representing a feature of a desired sentence in the target language and the predetermined noise-related feature tag are added, and the source language feature tag, the target language feature tag, and the noise-related feature tag are added. The present invention includes a translation target sentence tag adding unit that supplies the translation target sentence to the translation model for translation processing.

[6] The machine translation device according to one aspect of the present invention also includes a translation target sentence written in a source language, a source language feature tag representing the feature of the sentence in the source language, and a feature of the sentence in the target language. A target language side feature tag representing , is input, and a translation result sentence written in the target language is outputted in correspondence with the translation target sentence, the source language side feature tag, and the target language side feature tag. A translation model that holds information on the relationship between the translation target sentence, the source language feature tag, the target language feature tag, and the translation result sentence; A source language feature tag representing the feature of a sentence in the language and a target language feature tag representing a desired feature of the target language are added, and the source language feature tag and the target language feature tag are combined. a translation target sentence tag adding unit that supplies the added translation target sentence to the translation model for translation processing, and the translation model has been subjected to machine learning in advance using learning data. be.

[7] Further, one aspect of the present invention is the machine translation device according to [6] above, wherein the translation model is configured using a neural network.

[8] Further, one aspect of the present invention is the machine translation device according to [6] or [7] above, in which the translation model further includes noise related to noise between sentences in the source language and sentences in the target language. A noise-related feature tag is input, and the translation target sentence, the source language side feature tag, the noise-related feature tag, and the target language side are configured to output the translation result sentence corresponding to the noise-related feature tag as well. The device retains information on the relationship between a feature tag and the translation result sentence, and the translation target sentence tag adding unit further adds the noise-related feature tag to the translation target sentence, and further adds the noise-related feature tag to the translation target sentence. The translation target sentence, which has also been given a feature tag, is supplied to the translation model for translation processing.

[9] Further, one aspect of the present invention is a program for causing a computer to function as the machine translation device according to any one of [1] to [8] above.

According to the present invention, a machine translation device performs machine learning while grasping the characteristics of a corpus used as learning data. Thereby, the machine translation device accumulates knowledge while distinguishing between knowledge that should be strongly incorporated and other knowledge according to the characteristics of the corpus. This improves the translation accuracy of the machine translation device.

FIG. 1 is a block diagram showing a schematic functional configuration of a machine translation device according to an embodiment of the present invention. FIG. 2 is a schematic diagram showing the configuration of learning data used for learning a translation model according to the embodiment. FIG. 2 is a schematic diagram showing the configuration of learning data including tag information for performing machine learning of a translation model according to the embodiment. FIG. 2 is a schematic diagram illustrating a configuration of input data for translation processing including tag information for the machine translation device according to the embodiment to perform translation processing. This is a first example of tags added according to the embodiment, and is a table showing the relationship between examples of tags representing features on the source language side and the type of corpus associated with each tag. This is a first example of tags added according to the embodiment, and is a table showing the relationship between examples of tags representing characteristics of the target language and the type of corpus associated with each tag. It is a table listing examples of tag combinations in the first example of tags used in the embodiment. This is a second example of tags added according to the embodiment, and is a table showing the relationship between examples of tags representing features on the source language side and the type of corpus associated with each tag. This is a second example of tags added according to the embodiment, and is a table showing the relationship between examples of tags representing characteristics of the target language and the type of corpus associated with each tag. It is a table listing examples of tag combinations in a second example of tags used in the same embodiment. FIG. 7 is a schematic diagram illustrating an example of a tag that a learning data tag assigning unit assigns to a bilingual sentence pair input for learning data in the same embodiment. FIG. 2 is a schematic diagram showing a configuration of learning data including tag information (including feature tags related to noise) for performing machine learning of a translation model according to the embodiment. FIG. 2 is a schematic diagram showing a configuration of input data for translation processing including tag information (including feature tags related to noise) for the machine translation device according to the embodiment to perform translation processing. This is a third example of tags added according to the embodiment, and is a table showing the relationship between examples of tags representing features on the source language side and the type of corpus associated with each tag. This is a third example of tags added according to the embodiment, and is a table showing the relationship between examples of tags representing characteristics of the target language and the type of corpus associated with each tag. This is a third example of tags assigned according to the same embodiment, and is a table showing examples of tags related to noise. It is a table listing examples of tag combinations in a third example of tags used in the same embodiment.

Next, embodiments of the present invention will be described. The machine translation device according to the present embodiment learns the relationship between a translation target sentence and a translation result sentence according to the tag by adding a tag to learning data. The tag includes a source language feature tag to which a source language sentence of the learning data belongs, and a target language feature tag to which a target language sentence of the learning data belongs. These feature tags are determined based on the type of corpus to which each sentence belongs. Furthermore, the tags may include tags related to so-called noise features in the correspondence relationship between the source language side and the target language side (noise-related feature tags). In this embodiment, the translation model is realized using, for example, a neural network. In this way, by using the tagged learning data, the translation model advances learning while classifying the learning data according to its characteristics. This can be expected to prevent unnecessary knowledge from being mixed in. Therefore, the machine translation device of this embodiment can be expected to improve translation accuracy.

FIG. 1 is a block diagram showing a schematic functional configuration of a machine translation device according to this embodiment. As illustrated, the machine translation device 1 includes a corpus acquisition unit 10, a learning data tagging unit 20, a learning processing unit 30, a translation target sentence input unit 40, a translation target sentence tagging unit 50, It is configured to include a translation model 60 and a translation result sentence output unit 70. Each of these functional units can be realized by, for example, a computer and a program. Furthermore, each functional section has a storage means, if necessary. The storage means is, for example, variables on the program or memory allocated by executing the program. Furthermore, if necessary, nonvolatile storage means such as a magnetic hard disk device or a solid state drive (SSD) may be used. Furthermore, at least some of the functions of each functional unit may be realized as a dedicated electronic circuit instead of a program. The functions of each functional section will be explained next.

In this embodiment, the source language (original language) is the language in which the sentence to be translated (the sentence before translation) is written. Further, the target language (target language) is the language in which the translated sentence is written. Specifically, the source language and the target language may each be any language. Each of the source and target languages may be, for example, Japanese, English, German, French, Russian, Spanish, Italian, Chinese, Korean, Arabic, or other languages.

The corpus acquisition unit 10 acquires corpus data from the outside. The corpus acquired by the corpus acquisition unit 10 includes a source language corpus and a target language corpus. The sentences included in the corpus acquired by the corpus acquisition unit 10 are sentences used as learning data for the translation model 60. A method of using sentences included in the corpus as learning data will be explained later with reference to FIG. 2 and the like.

The learning data tagging unit 20 adds tag information to data for machine learning (a pair of a source language side sentence and a target language side sentence). In other words, the learning data tagging unit 20 determines that the source language side sentence belongs to a sentence pair consisting of a source language side sentence written in the source language and a target language side sentence written in the target language. A source language feature tag corresponding to the type of corpus and a target language feature tag corresponding to the type of corpus to which the target language sentence belongs are added.

The learning data tag adding unit 20 may further add a tag representing information regarding noise. That is, in this case, the learning data tag adding unit 20 further adds the predetermined noise-related feature tag to the sentence pair. The learning data tag adding unit 20 may acquire information on the type of noise-related feature tag to be added here from outside.

The learning processing unit 30 supplies learning data for the translation model 60 to perform machine learning, and causes the translation model 60 to perform the learning process. In other words, the learning processing unit 30 translates the set of sentence pairs to which the learning data tagging unit 20 has added the source language feature tag and the target language feature tag as learning data for use in machine learning processing. Supply to model 60. The learning processing unit 30 may further supply the set of sentence pairs to which the noise-related feature tag has been added to the translation model as the learning data.

The translation target sentence input unit 40 is for acquiring a translation target sentence. The translation target sentence input unit 40 receives a translation target sentence from, for example, an external device or reads it from a storage medium. The translation target sentence input unit 40 passes the translation target sentence to the translation target sentence tag adding unit 50.

The translation target sentence tag adding unit 50 adds a tag to the translation target sentence passed from the translation target sentence input unit 40. In other words, the translation target sentence tag adding unit 50 adds a source language feature tag representing the feature of the sentence in the source language and a desired sentence feature in the target language to the target sentence written in the source language. The sentence to be translated, to which the source language feature tag and the target language feature tag have been added, is supplied to the translation model 60 for translation processing.

Note that the translation target sentence tag adding unit 50 may further add tag information regarding noise. In other words, in this case, the translation target sentence tag adding unit 50 adds a source language feature tag representing the feature of the sentence in the source language and a desired sentence in the target language to the translation target sentence written in the source language. a target language side feature tag representing a feature of the target language and the predetermined noise-related feature tag, and the source language side feature tag, the target language side feature tag, and the noise-related feature tag are added. is supplied to the translation model 60 for translation processing.

Translation model 60 is a model for translation. The translation model 60 uses machine learning techniques. The translation model 60 is realized using, for example, a neural network. In this case, translation model 60 is configured to perform neural machine translation (NMT). However, other machine learning techniques may be used as the translation model 60. Translation processing using machine learning itself can be realized using existing technology, but in this embodiment, tag information is used to improve learning efficiency and translation accuracy. Specifically, in this embodiment, the translation model 60 generates a translation result in association with an input tag. In other words, the translation model 60 includes a sentence to be translated written in the source language, a source language feature tag representing the feature of the sentence in the source language, and a target language feature tag representing the feature of the sentence in the target language. is input, the translation target sentence and the source language are set so that a translation result sentence written in the target language is output corresponding to the translation target sentence, the source language side feature tag, and the target language side feature tag. Information about the relationship between the language-side feature tag, the target language-side feature tag, and the translation result sentence is held. The translation model 60 acquires learning data when performing machine learning. Further, after the machine learning of the translation model 60 is completed, when performing translation processing, a sentence to be translated is acquired.

The translation model 60 may further handle tags representing characteristics related to noise. That is, in this case, the translation model 60 further receives as input a noise-related feature tag related to noise between a sentence in the source language and a sentence in the target language, and also translates the noise-related feature tag corresponding to the noise-related feature tag. In order to output a result sentence, information on the relationship between the translation target sentence, the source language feature tag, the noise-related feature tag, the target language feature tag, and the translation result sentence is held. In this case, the translation model 60 outputs a translated sentence based also on the noise-related feature tag during translation processing.

Note that when the translation model 60 is realized using, for example, a neural network, the input/output data is converted into data in an appropriate format (for example, data in a format such as a one-hot vector) and processed. . For learning a neural network, an existing method, error backpropagation, can be used.

The translation result sentence output unit 70 outputs the translation result sentence generated by the translation model 60 to the outside.

FIG. 2 is a schematic diagram showing the structure of learning data used for learning the translation model 60. As shown in the figure, the learning data is a set of bilingual sentence pairs consisting of sentences in the source language and sentences in the target language. In the figure, only bilingual sentence pairs 1 to 4 are shown, but in reality, it is preferable to prepare a large number of sentence pairs (as an example, about several hundred thousand pairs). In this training data, all sentences in the source language belong to a single corpus. All sentences in the target language also belong to a single corpus. The source language side corpus to which source language sentences belong and the target language side corpus to which target language sentences belong have information on their respective corpus types. The information on the type of corpus is information that is the basis of the tag assigned by the learning data tag assignment unit 20. In other words, the tag given to learning data is determined by the type of corpus. The process of adding tags to learning data will be further explained in FIG. 3 and subsequent figures.

Note that multiple sets of training data (learning data in which a set of sentences on the source language side belongs to a single corpus and a set of sentences on the target language side belongs to a single corpus) having the configuration shown in Figure 2 were used. The translation model 60 can be trained.

FIG. 3 is a schematic diagram showing the structure of learning data for performing machine learning of the translation model 60 of the machine translation device 1. The learning data includes data on the input side to the translation model 60 and data on the output side. The learning processing unit 30 supplies the input side data of this learning data to the translation model 60, and also compares the data output from the translation model 60 as a result with the output side data of this learning data. The translation model 60 adjusts internal parameters based on the difference. In other words, the output data in the learning data is used as correct data. As shown in the figure, the data on the input side of the learning data is configured to include a source language side feature tag, a target language side feature tag, and a source language side sentence. In addition, when processing using a computer, the source language side feature tag, the target language side feature tag, and the source language side sentence may be separated by appropriate separator data. On the other hand, the data on the output side of the learning data is the target language side sentence. Here, the source language feature tag depends on the type of source language corpus (see FIG. 2) to which the illustrated source language sentence belonged. That is, this source language feature tag is data representing the feature of the source language corpus. Further, the target language feature tag depends on the type of target language corpus (see FIG. 2) to which the illustrated target language sentence belonged. That is, this target language side feature tag is data representing the characteristics of the target language side corpus. As described above, the learning data tag adding unit 20 adds the source language side feature tag and the target language side feature tag to the learning data.

By performing machine learning using the training data shown in FIG. 3, the translation model 60 is able to differentiate between source language sentences and target language sentences in a manner that depends on source language feature tags and target language feature tags. Learn the relationship with side sentences. In other words, by performing such machine learning, the trained translation model 60 expresses the relationship between the source language side sentence and the target language side sentence depending on the source language side feature tag and the target language side feature tag. Become something.

The data shown in FIG. 3 corresponds to a pair of bilingual sentences for machine learning. When actually performing machine learning on the translation model 60, a large amount of data having the configuration shown in FIG. 3 is prepared.

FIG. 4 is a schematic diagram showing the structure of input data for translation processing for translation processing by the translation model 60 of the machine translation device 1. As shown in the figure, the input data for translation processing is configured to include a source language side feature tag, a target language side feature tag, and a source language side sentence. In addition, as in the case of learning data (Figure 3), when processing using a computer, the above source language side feature tags, target language side feature tags, and source language side sentences are changed as appropriate. It may be separated by separator data. In the input data for translation processing, the source language feature tag is data representing the feature of the corpus to which the source language sentence that is the target of translation processing belongs. Alternatively, if the source language side sentence is not obtained from a corpus, the source language side feature tag is data of tags appropriately selected to represent the characteristics of the source language side sentence. Further, in the input data for translation processing, the target language side feature tag is data of a tag appropriately selected to represent the characteristics of a desired target language side sentence (translation result sentence). As described above, the translation target sentence tag adding unit 50 adds the above-described source language side feature tag and target language side feature tag to the translation target sentence.

According to the above configuration, the translation model 60 uses learning data to store knowledge associated with tag information. Further, the translation model 60 performs translation processing based on tag information.

Below, examples of tags added by the learning data tagging unit 20 and the translation target sentence tagging unit 50 will be described.

[Example of tag to be added (first example)]
FIG. 5 is a table showing the relationship between examples of source language feature tags and the type of source language corpus associated with each feature tag in the first example. In this example, the source language is Japanese. In this example, <NHK-S> and <MT-S> are used as source language side feature tags. The tag <NHK-S> corresponds to a corpus of Japanese news sentences used in NHK news. The tag <MT-S> corresponds to a corpus of sentences obtained by translating English news sentences used in NHK news into Japanese using machine translation processing.

FIG. 6 is a table showing the relationship between examples of target language feature tags and the types of target language corpora associated with each feature tag in the first example. In this example, the target language is English. In this example, <NHK-T>, <MT-T>, and <AMA-T> are used as the target language side feature tags. The tag <NHK-T> corresponds to a corpus of English news sentences used in NHK news. The tag <MT-T> corresponds to a corpus of sentences obtained by translating Japanese news sentences used in NHK news into English using machine translation processing. The tag <AMA-T> corresponds to a corpus of sentences translated into English by amateur translators from Japanese news sentences used in NHK news.

FIG. 7 is a schematic diagram showing an example of the combination use of source language side feature tags and target language side feature tags in the first example. In the same figure,
The combination of tags <NHK-S> and <NHK-T> in (1) means that the sentence pair of the Japanese news sentence used in NHK news programs and the corresponding English news sentence is Indicates that it is used during learning processing.
The combination of tags <NHK-S> and <AMA-T> in (2) is a combination of the Japanese news text used in NHK news programs and the Japanese news text translated into English by an amateur translator. Indicates that the sentence pair with the English sentence obtained as a result of translation is used during machine learning processing.
The tag combination <NHK-S> and <MT-T> in (3) is the Japanese news text used in NHK news programs and the Japanese news text translated into English by machine translation. Indicates that the resulting sentence pair with an English sentence will be used during machine learning processing.
The combination of tags <MT-S> and <NHK-T> in (4) is the combination of the English news sentences (target language side sentences) used in NHK's English news programs and the English news sentences. This indicates that a sentence pair with a Japanese sentence (source language side sentence) obtained as a result of translation into Japanese by machine translation is used during machine learning processing.

In this first example, the feature of combination (1) is that the quality of the sentences is high in both the source language (Japanese) and the target language (English), but there is an excess or deficiency in content between the sentences included in the sentence pair. may occur. The characteristic of combination (2) is that the quality of the sentence in the source language (Japanese) is high, but the quality of the sentence in the target language (English) is poor as a translation of the sentence in the source language (translation quality). It is possible that it is not that high. The characteristics of the combinations (3) and (4) are that, because machine translation processing is used, a large number of sentence pairs can be generated at low cost, but the quality of the sentences generated by machine translation is not necessarily high. be.

As described above, in the first example, the tags indicate whether the sentence is a news sentence created for broadcasting, whether the sentence is translated by an amateur, and whether the sentence is generated by machine translation. Represents information such as whether or not. By performing machine learning using learning data with such tags added, the translation model 60 can acquire knowledge in a form associated with tag information.

[Example of tag to be added (second example)]
FIG. 8 is a table showing the relationship between examples of source language feature tags and the type of source language corpus associated with each feature tag in the second example. In this example as well, the source language is Japanese. In this example, <PRO-JA>, <MTA-JA>, and <MTB-JA> are used as source language side feature tags. The tag <PRO-JA> corresponds to a corpus of Japanese sentences created by professional sentence creators. The tag <MTA-JA> corresponds to a corpus of Japanese sentences generated by machine translation processing by translation device A based on English sentences. The tag <MTB-JA> corresponds to a corpus of Japanese sentences generated by machine translation processing by translation device B based on English sentences. Note that translation device A and translation device B are different translation devices. In other words, the Japanese sentences generated by translation device A and the Japanese sentences generated by translation device B have different characteristics.

FIG. 9 is a table showing the relationship between examples of target language feature tags and the types of target language corpora associated with each feature tag in the second example. In this example as well, the target language is English. In this example, <PRO-EN> and <AMA-EN> are used as the target language side feature tags. The tag <PRO-EN> corresponds to a corpus of English sentences created by a professional English sentence writer. The tag <AMA-EN> corresponds to a corpus of sentences resulting from translation into English by an amateur translator based on Japanese sentences.

FIG. 10 is a schematic diagram showing an example of the combination use of source language side feature tags and target language side feature tags in the second example. In the same figure, the combination of tags <PRO-JA> and <PRO-EN> in (1) is a sentence created by a professional in the source language (Japanese) and a sentence created by a professional in the target language (English). This indicates that machine learning is performed using a sentence pair containing the following sentences. The combination of tags <PRO-JA> and <AMA-EN> in (2) is a sentence in the source language (Japanese) created by a professional, and the Japanese sentence translated into English by an amateur translator. Machine learning is performed using sentence pairs that include the resulting English sentence. The combination of tags <MTA-JA> and <PRO-EN> in (3) is a sentence (news sentence) in the target language (English) created by a professional and a machine translation of the English sentence using translation device A. This indicates that machine learning is performed using sentence pairs that include the resulting sentences in the source language (Japanese). The combination of tags <MTB-JA> and <PRO-EN> in (4) is a sentence (news sentence) in the target language (English) created by a professional and a machine translation of the English sentence using translation device B. This indicates that machine learning is performed using sentence pairs that include the resulting sentences in the source language (Japanese).

FIG. 11 is a schematic diagram showing an example of learning data when using the second example of feature tags. As shown in the figure, the pair of translated sentences before being tagged by the learning data tagging unit 20 is: ``Takahashi is scheduled to participate in the Summer Jump Grand Prix tournament that will begin in France on the 11th of this month, and on the morning of the 6th. "Saya Takahashi, will be participating in the Ski Jumping Summer Grand Prix due to start in France on August 11. She "responded to an interview at Haneda Airport on the morning of August 6, just before departure." is the sentence in the target language (English). The learning data tagging unit 20 adds the tag combination <PRO-JA><PRO-EN> to the beginning of this sentence pair. The learning processing unit 30 provides this tagged bilingual sentence pair to the translation model 60 when performing machine learning.

As mentioned above, in the second example as well, the tags are whether the sentence was created by a professional, whether the sentence was translated by an amateur translator, and whether the sentence was obtained as a result of machine translation processing. (and what translation device was used at that time). By performing machine learning using learning data with such tags added, the translation model 60 can acquire knowledge in a form associated with tag information.

[Example of tag to be added (third example)]
In the first and second examples, a total of two feature tags, one feature tag for the source language and one feature tag for the target language, were assigned. In the third example, a feature tag related to noise is further added.

Learning data generally includes noise. Noise contained in training data leads to a decline in the quality of translation models that are subjected to machine learning using the training data. The noise-related feature tag is a tag that indicates, for each type of noise, whether or not the corpus includes that type of noise. Here, back-translated noise is noise generated by back-translation processing from the target language side to the source language side. Furthermore, non-equivalent noise is noise that occurs when sentences in the source language and sentences in the target language do not have equivalent content. Non-equivalence noise may occur, for example, when translation is performed without special consideration of equivalence. As an example, when a person writes a news sentence in Japanese and uses some method of translation to create an English news sentence based on the Japanese news sentence, generally speaking, the original Japanese sentence is Therefore, some content may be added or deleted. When the learning data includes noise as described above, a tag indicating that the learning data includes the noise is added to each type of noise. Additionally, a corpus can be generated so that the learning data does not include the above-mentioned noise. If the learning data does not contain noise, a tag indicating that it does not contain noise is added to each type of noise.

FIG. 12 is a schematic diagram showing the configuration of learning data when a third example of tags is added in this embodiment. Similar to the learning data shown in FIG. 3, the learning data shown in FIG. 12 includes data on the input side and data on the output side to the translation model 60. Furthermore, in the learning data shown in FIG. 12, similarly to the learning data shown in FIG. 3, the data on the input side includes a source language side feature tag, a target language side feature tag, and a source language side sentence. The data on the input side of the learning data shown in FIG. 12 is further configured to include feature tags related to noise (also referred to as "noise-related feature tags"). In the example shown in FIG. 12, the data on the input side is arranged in the following order: source language feature tag, target language feature tag, noise feature tag, and source language sentence. However, the arrangement within the data on the input side is arbitrary. The noise-related feature tag is information indicating the presence or absence of noise as described above. Note that the feature tags related to noise shown in FIG. 12 may include a plurality of tags for each type of noise. Note that the learning data tag adding unit 20 adds a source language side tag, a target language side tag, and a feature tag related to noise to the learning data. Which noise-related feature tag the learning data tagging unit 20 adds depends on instruction information given from outside the learning data tagging unit 20 .

By performing machine learning using such learning data shown in FIG. 12, the translation model 60 is able to differentiate between the source language side sentence and the target language side sentence in a manner that depends on the presence or absence of noise for each noise type. Learn relationships. That is, by performing such machine learning, the trained translation model 60 represents the relationship between the source language side sentence and the target language side sentence depending on the presence or absence of noise.

The data shown in FIG. 12 corresponds to a pair of bilingual sentences for machine learning. When actually performing machine learning on the translation model 60, a large amount of data having the configuration shown in FIG. 12 is prepared.

FIG. 13 is a schematic diagram showing the configuration of input data for translation processing for the translation model 60 to perform translation processing when adding the third example tag in this embodiment. Similar to the data shown in FIG. 4, the input data for translation processing shown in FIG. 13 includes a source language feature tag, a target language feature tag, and a source language sentence. The input data for translation processing shown in FIG. 13 is further configured to include feature tags related to noise. As explained in FIG. 12, the feature tag regarding noise is a tag representing information regarding the presence or absence of noise. The noise-related feature tags shown in FIG. 13 may include a plurality of tags for each type of noise. The translation model 60 that has been subjected to machine learning using the training data shown in FIG. Then, translation processing that also depends on noise-related feature tags is performed. That is, the translation model 60 that has been subjected to machine learning using the learning data in FIG. 12 performs processing to output a target language side sentence that also depends on the noise-related feature tag based on the source language side sentence. Note that the translation target sentence tag adding unit 50 adds source language side feature tags, target language side feature tags, and noise-related feature tags to the translation target sentence. Which noise-related feature tag the translation target sentence tag assigning unit 50 assigns depends on instruction information given from outside the translation target sentence tag assigning unit 50 .

FIG. 14 is a table showing the relationship between examples of source language feature tags and the type of source language corpus associated with each feature tag in the third example. In this example, the source language is not limited to any particular language. In this example, <NS-S> and <CE-S> are used as source language side feature tags. The tag <NS-S> corresponds to a corpus of news sentences written in the source language (for example, news sentences used in broadcasting). The tag <CE-S> corresponds to a corpus of sentences (sentences written in the source language) that have been back-translated by machine translation processing based on <NS-T> described in FIG. 15 below.

FIG. 15 is a table showing the relationship between examples of target language feature tags and the types of target language corpora associated with each feature tag in the third example. In this example, the target language is also not limited to a specific language. In this example, <CE-T> and <NS-T> are used as target language side feature tags. The tag <CE-T> corresponds to a corpus of sentences in the target language that have the same content as the news sentences in the source language (the <NS-S> sentences shown in FIG. 14). Note that "CE" means "content equivalent". The tag <NS-T> corresponds to a corpus of news sentences written in the target language (for example, news sentences used in broadcasting in the target language).

FIG. 16 is a table showing the relationship between examples of noise-related feature tags and the type of corpus associated with each feature tag in the third example. In this example, <NO-BT>, <BT>, <NO-AN>, and <AN> are used as feature tags related to noise, for example. The tag <NO-BT> corresponds to a corpus that does not have back translation noise. The tag <BT> corresponds to a corpus with back translation noise. The tag <NO-AN> corresponds to a corpus that does not have non-equivalent noise. The tag <AN> corresponds to a corpus with non-equivalent noise.

FIG. 17 is a schematic diagram showing an example of the combination use of source language side feature tags and target language side feature tags in the third example. In the same figure, the combination of tags <NS-S>, <CE-T>, <NO-BT>, and <NO-AN> in (1) is used to translate news sentences written in the source language. Indicates that a sentence in the target language with content equivalent to the news sentence in the source language is used. Furthermore, in the case of (1), it means that there is no back translation noise or non-equivalence noise between the sentence pair. The tag combination <NS-S>, <NS-T>, <NO-BT>, and <AN> in (2) is used to translate a news sentence written in the source language into a target language. Indicates that a news text is used. Furthermore, in the case of (2), it means that there is no back translation noise between the sentence pair, but there may be non-equivalence noise. The combination of tags <CE-S>, <NS-T>, <BT>, and <NO-AN> in (3) was back-translated by machine translation processing based on the news text written in the target language. This indicates that a sentence in the source language is used as a sentence to be translated in the learning data, and a news sentence written in the target language, which is the basis for the above-mentioned back translation process, is used on the target language side. However, the sentences in the source language are created to have the same content as the sentences in the target language. Furthermore, in the case of (3), it means that the sentence pair contains back translation noise and does not have non-equivalent noise. The combination of tags <NS-S>, <NS-T>, <BT>, and <AN> in (4) uses a news sentence written in the source language and a news sentence written in the target language. represents something. Furthermore, in the case of (4), it means that the sentence pair may contain back translation noise and may also contain non-equivalent noise.

When machine learning of the translation model 60 is performed using the feature tags of this third example, tags related to noise are selected and assigned in the translation process so as to obtain a desired result. In order to obtain high-quality translation results in translation processing, it is usually considered to use noise-related feature tags to perform noise-free translation.

As described above, in the machine translation device 1 according to the present embodiment, machine learning of the translation model 60 can be performed using a plurality of corpora having a plurality of characteristics in machine translation processing. Furthermore, during learning using a plurality of corpora having a plurality of features, the machine translation device 1 adds a tag representing the feature of each corpus to the learning data. Thereby, the translation model 60 can acquire knowledge in a form associated with the tag. In other words, the tags assigned during learning have the effect of classifying knowledge according to the characteristics of the corpus in the information held by the translation model 60. Thereby, the machine translation device 1 stores knowledge for translation in a classified and organized manner without mixing knowledge obtained from corpora having a plurality of characteristics.

By using the machine translation device 1 of this embodiment that produces the above-mentioned effects, it is possible not only to utilize a high-quality corpus at a high cost, but also to utilize a large quantity at a low cost, although the quality is not so high. It is also possible to utilize corpora that provide data on In other words, the machine translation device 1 can acquire a large amount of knowledge at a low cost overall.

Note that the machine translation device 1 can distinguish between features of the corpus on the source language side and features of the corpus on the target language side.

Furthermore, the machine translation device 1 according to the present embodiment can also use feature tags related to noise, like the tags shown in the third example. The noise-related feature tag can represent noise features that also depend on the relationship between the source language corpus and the target language corpus. The machine translation device 1 can also associate noise-related features, acquire knowledge, and utilize that knowledge in machine translation processing.

Note that if the number and types of tags to be added are too large, knowledge will be overclassified and learning efficiency will deteriorate. The number of tags to be added is desirably between about 2 and 4, as exemplified in this embodiment. Also, the number of tag types should not be too large.

Note that at least some of the functions of the machine translation device in the embodiments described above can be realized by a computer. In that case, a program for realizing this function may be recorded on a computer-readable recording medium, and the program recorded on the recording medium may be read into a computer system and executed. Note that the "computer system" herein includes hardware such as an OS and peripheral devices. Furthermore, "computer-readable recording media" refers to portable media such as flexible disks, magneto-optical disks, ROM, CD-ROM, DVD-ROM, and USB memory, and storage devices such as hard disks built into computer systems. Say something. Furthermore, a "computer-readable recording medium" refers to a medium that temporarily and dynamically stores a program, such as a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In that case, it may also include something that retains a program for a certain period of time, such as a volatile memory inside a computer system that is a server or client. Further, the program may be one for realizing a part of the above-mentioned functions, or may be one that can realize the above-mentioned functions in combination with a program already recorded in the computer system.

Note that the machine translation device 1 may be implemented as the following modification example.
The machine translation device 1 may not include the translation target sentence input unit 40, the translation target sentence tagging unit 50, and the translation result sentence output unit 70. In this case as well, the machine translation device 1 can perform machine learning processing on the translation model 60. That is, the machine translation device 1 can function as a learning device.
The machine translation device 1 may not include the corpus acquisition section 10, the learning data tagging section 20, and the learning processing section 30. However, in this case, the translation model 60 is set to have undergone machine learning. Thereby, although the machine translation device 1 does not have a function of learning using learning data, it can perform translation processing on a sentence to be translated.
The machine translation device 1 may or may not handle noise-related feature tags (see the third example of tags). In either case, the machine translation device 1 performs processing for handling source language feature tags and target language feature tags.

Although the embodiments of the present invention have been described above in detail with reference to the drawings, the specific configuration is not limited to these embodiments, and includes designs within the scope of the gist of the present invention.

[Evaluation experiment]
An experiment was conducted to evaluate the method of this embodiment. The BLEU score was used as the evaluation scale. The BLEU score is a value calculated based on the number of n-gram matches between the result of machine translation and a reference translation, and is often used to evaluate machine translation.
As a first comparison target, the BLEU score when neural machine translation was performed without using tags was 20.36.
As a second comparison target, the BLEU score when using the single tag method (Kobus et al., 2017, Non-Patent Document 2 mentioned above) was 22.41.
As a third comparison target, the BLEU score when using a method using two types of tags (Kobus et al., 2017) was 24.25.
On the other hand, when using the method of this embodiment, the BLEU score was 24.56, which is higher than the conventional technology.

The present invention can be used, for example, in any industry that uses machine translation processing of sentences written in natural language. However, the scope of use of the present invention is not limited to what is exemplified here.

1 Machine translation device 10 Corpus acquisition unit 20 Learning data tagging unit 30 Learning processing unit 40 Translation target sentence input unit 50 Translation target sentence tagging unit 60 Translation model 70 Translation result sentence output unit

Claims (7)

A translation target sentence written in the source language, a source language feature tag representing the feature of the sentence in the source language, and a target language feature tag representing the feature of the sentence in the target language are input, and the translation target is The target sentence, the source language feature tag, and the target are configured to output a translation result sentence written in the target language in correspondence with the sentence, the source language feature tag, and the target language feature tag. a translation model that retains information on the relationship between language-side feature tags and the translation result sentence;
A source language feature tag corresponding to the type of corpus to which the source language sentence belongs is attached to a sentence pair consisting of a source language sentence written in the source language and a target language sentence written in the target language. and a training data tagging unit that adds a target language side feature tag according to the type of corpus to which the target language side sentence belongs;
a learning processing unit that supplies the set of sentence pairs to which the learning data tagging unit has added the source language feature tag and the target language feature tag to the translation model as learning data for use in machine learning processing; and,
Equipped with
The translation model further receives as input a noise-related feature tag related to noise between a sentence in the source language and a sentence in the target language, and outputs the translated result sentence corresponding to the noise-related feature tag. and retains information on the relationship between the translation target sentence, the source language side feature tag, the noise related feature tag, the target language side feature tag, and the translation result sentence,
The learning data tag adding unit further adds the predetermined noise-related feature tag to the sentence pair,
The learning processing unit further supplies the set of sentence pairs to which the noise-related feature tag has been added to the translation model as the learning data.
Machine translation device. The translation model is configured using a neural network.
The machine translation device according to claim 1. A source language feature tag representing a feature of the sentence in the source language and a target language feature tag representing a desired feature of the sentence in the target language are added to the sentence to be translated written in the source language, and a translation target sentence tagging unit that supplies the translation target sentence to which a source language side feature tag and a target language side feature tag have been added to the translation model for translation processing;
The machine translation device according to claim 1 or 2, further comprising: A source language feature tag representing a feature of the sentence in the source language, a target language feature tag representing a desired feature of the sentence in the target language, and the predetermined noise are added to the sentence to be translated written in the source language. a related feature tag, and the source language side feature tag, the target language side feature tag, and the noise related feature tag are attached, and the translation target sentence is supplied to the translation model for translation processing. Sentence tagging section,
The machine translation device according to any one of claims 1 to 3, comprising: A translation target sentence written in the source language, a source language feature tag representing the feature of the sentence in the source language, and a target language feature tag representing the feature of the sentence in the target language are input, and the translation target is The target sentence, the source language feature tag, and the target are configured to output a translation result sentence written in the target language in correspondence with the sentence, the source language feature tag, and the target language feature tag. a translation model that retains information on the relationship between language-side feature tags and the translation result sentence;
A source language feature tag representing a feature of the sentence in the source language and a target language feature tag representing a desired feature of the sentence in the target language are added to the sentence to be translated written in the source language, and a translation target sentence tagging unit that supplies the translation target sentence to which a source language side feature tag and a target language side feature tag have been added to the translation model for translation processing;
Equipped with
The translation model has been machine learned in advance using learning data,
The translation model further receives as input a noise-related feature tag related to noise between a sentence in the source language and a sentence in the target language, and outputs the translated result sentence corresponding to the noise-related feature tag. and retains information on the relationship between the translation target sentence, the source language side feature tag, the noise related feature tag, the target language side feature tag, and the translation result sentence,
The translation target sentence tag adding unit further adds the noise-related feature tag to the translation target sentence, and applies the translation target sentence to which the noise-related feature tag has also been added to the translation model for translation processing. supply to,
Machine translation device. The translation model is configured using a neural network.
The machine translation device according to claim 5 . computer,
The machine translation device according to any one of claims 1 to 6 ,
A program to function as

Images