WO2020166125A1

WO2020166125A1 - Translation data generating system

Info

Publication number: WO2020166125A1
Application number: PCT/JP2019/039337
Authority: WO
Inventors: 聡一朗村上
Original assignee: 株式会社Ｎｔｔドコモ
Priority date: 2019-02-12
Filing date: 2019-10-04
Publication date: 2020-08-20
Also published as: JP7194759B2; JPWO2020166125A1

Abstract

A translation data generating system 1 is provided with a noise imparting unit 12 that imparts noise to source language text to yield noise-imparted source language text, and a corpus constructing unit 13 that constructs a pseudo-parallel corpus where the noise-imparted source language text, and target language text corresponding to the source language text of the noise-imparted source language text before imparting noise, are correlated.

Description

Data generation system for translation

One aspect of the present invention relates to a translation data generation system.

In the machine translation system, the translation accuracy is reduced due to the fact that the user's natural utterance input includes stagnation, rewording, or fillers (hereinafter these may be collectively referred to as "noise"). There is a case.

With respect to such a problem, as disclosed in, for example, Patent Document 1 and Patent Document 2, there is known a technique of identifying a reworded portion in an utterance and correcting the utterance content of the user.

JP, 2010-079647, A JP, 2007-057844, A

However, it is not easy to identify and correct noise points, and it is difficult to ensure sufficient translation accuracy even with the techniques described above.

One aspect of the present invention has been made in view of the above situation, and an object thereof is to perform highly accurate translation even for natural speech including noise.

A translation data generation system according to an aspect of the present invention includes a noise adding unit that adds noise to a source language text to obtain a noise added source language text, a noise added source language text, and noise of the noise added source language text. And a corpus constructing unit for constructing a pseudo bilingual corpus in which a target language text corresponding to a source language text before being assigned is associated.

In the translation data generation system according to one aspect of the present invention, noise is added to the source language text, and the pseudo-translation corpus in which the noise-added source language text and the target language text corresponding to the source language text before the addition of noise are associated with each other. Is built. In this way, by constructing the bilingual corpus in which the source language text with noise is associated with the target language text corresponding to the source language text before adding noise, using such a bilingual corpus, for example, natural speech Even when noise such as filler is included in the input, it is possible to appropriately derive the target language text corresponding to the source language text before noise addition. That is, according to the translation data generation system of one aspect of the present invention, it is possible to construct a robust corpus (pseudo bilingual corpus) for natural utterances including noise, and for natural utterances including noise. However, it is possible to translate with high accuracy.

The translation data generation system may further include a translation model learning unit that learns a translation model using a pseudo bilingual corpus. By learning the translation model based on the constructed corpus, it is possible to translate the natural utterance containing noise with higher accuracy.

The translation data generation system further includes a noise model learning unit that uses a training data that is a source language text group containing noise to learn a noise model related to the addition of noise to the source language text, and the noise addition unit is , A noise model may be used to add noise to the source language text. A noise model is learned based on a source language text group that includes noise in advance, and noise is added based on the noise model, so that noise that is likely to be actually included is easily added. The accuracy can be further improved.

In the translation data generation system, the noise adding unit adds a noise label indicating the type of noise to each word of the source language text, and replaces the noise label with the word corresponding to the noise label Noise may be added to the text. Since the noise label corresponding to each word of the source language text is added and then the word (noise) corresponding to the noise label is derived, it is possible to ensure the easiness and validity of the noise addition.

In the above translation data generation system, the noise adding unit may derive a plurality of patterns of words to be replaced with respect to one noise label, and obtain a plurality of patterns of noise adding source language text from one source language text. As a result, it is possible to efficiently enhance the pseudo bilingual corpus from one source language text and further improve the translation accuracy.

In the above translation data generation system, the noise adding unit may derive a plurality of patterns of noise labels corresponding to each word and obtain a plurality of patterns of noise-added source language text from one source language text. As a result, it is possible to efficiently enhance the pseudo bilingual corpus from one source language text and further improve the translation accuracy.

In the above translation data generation system, the noise adding unit may add a noise label according to the characteristics of each word in the source language text. This makes it possible to appropriately give each word a noise label relating to noise that is likely to be included in association with each word.

In the above translation data generation system, the noise adding unit may add a noise label according to at least one of the morpheme, the part of speech, and the reading of the word, which are the features of each word of the source language text. This makes it possible to appropriately give each word a noise label relating to noise that is likely to be included in association with each word.

In the above translation data generation system, the noise adding unit samples the noise label according to the probability distribution of each noise level based on the score of each noise label output from the noise model with the feature of each word of the source language text as an input, You may decide the noise label given to a source language text. As a result, for example, it is possible to give a noise label having a high score output from the noise model, and it is possible to appropriately give each word a noise label relating to noise that is likely to be included in association with each word.

In the above translation data generation system, the noise model may be constructed by a method using a conditional random field or neural network. Thereby, the noise model can be appropriately configured by machine learning.

According to one aspect of the present invention, it is possible to perform highly accurate translation even for natural speech including noise.

It is a figure which shows typically the processing image of the data production system for translation which concerns on this embodiment. It is a figure which shows the function structure of the data generation system for translations which concerns on this embodiment. It is a figure explaining the outline of a noise model. It is a figure explaining a noise label. It is a figure which shows the construction image of a pseudo bilingual corpus. It is a flow chart which shows the processing which a translation data generation system performs. It is a table which shows the translation example of this embodiment and a comparative example. It is a figure which shows the hardware constitutions of a translation data generation apparatus.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the description of the drawings, the same or equivalent elements will be denoted by the same reference symbols, without redundant description.

First, a processing image of the translation data generation system 1 according to the present embodiment will be described with reference to FIG. FIG. 1 is a diagram schematically showing a processing image of the translation data generation system 1 according to the present embodiment. The translation data generation system 1 adds noise to a text (source language text) on the source language side of an existing bilingual corpus (generally used bilingual corpus), and at the same time, adds noise to the source language text ( A source language text (noise-added) and a target language side text (target language text) corresponding to the source language text before noise are associated to construct a pseudo-translation corpus, and a machine translation model is created using the pseudo-translation corpus. This is a system for learning (constructing) (for example, NMT (Neural Machine Translation) model). The noise here is stagnation, rewording, filler, or the like that may be included in the user's natural speech input.

In the example shown in Fig. 1, multiple patterns of noise are added to the source language text "I want to drive on a tourist route rather than on a major highway" in an existing bilingual corpus according to a predetermined rule ( (Details will be given later), "Well, I would like to drive on a sightseeing route than a major highway." "Oh, I would like to drive on a tourist route more than a major freeway." "Well than a major highway. I want to drive on a sightseeing route." Then, the target language text corresponding to the source language text before noise is added, "I would rather take a scenic route than a main highway." A corpus is constructed, and a machine translation model is learned (constructed) using the pseudo bilingual corpus. In this way, by constructing the pseudo bilingual corpus in which the noise-added source language text is associated with the target language text corresponding to the source language text before the addition of noise, using such a pseudo bilingual corpus, for example, Even when the natural utterance input includes noise such as filler, it is possible to appropriately derive the target language text corresponding to the source language text before noise addition. The details of the function of the translation data generation system 1 will be described below.

FIG. 2 is a diagram showing a functional configuration of the translation data generation system 1 according to the present embodiment. As shown in FIG. 2, the translation data generation system 1 includes a translation data generation device 10, a parallel translation corpus DB 20, a training information DB 30, a noise model learning device 40 (noise model learning unit), and a translation model learning. The apparatus 50 (translation model learning unit). It should be noted that the translation data generation system 1 does not necessarily have to have the above-described components, and may be configured with only the translation data generation device 10, for example, or the translation data generation device 10 and the noise model. It may be configured by only the learning device 40, or may be configured by only the translation data generation device 10, the noise model learning device 40, and the translation model learning device 50.

The parallel translation corpus DB 20 is a database that stores the parallel translation corpus. A bilingual corpus is a structured combination of source language text and target language text. The bilingual corpus stored in the bilingual corpus DB 20 may be a commonly used bilingual corpus, and is, for example, a Japanese/English bilingual corpus such as KFTT (Kyoto Free Translation Task) or BTEC. In the present embodiment, the translation data generation device 10 adds noise to the source language text of the bilingual corpus stored in the bilingual corpus DB 20 to generate a pseudo bilingual corpus (details will be described later).

The training information DB 30 is a database that stores training information (training data) for learning a noise model (details will be described later). The training information is a source language text group (transcription corpus of natural utterances, utterance data for learning) in which noise is annotated in advance. Such training information is constructed, for example, by annotating a source language text included in a normal corpus with noise.

The noise model learning device 40 uses the training information (training data that is a source language text group including noise) stored in the training information DB 30 to learn a noise model related to adding noise to the source language text. As the learning data (training data) of the noise model, for example, a transcription corpus of a Japanese utterance corpus (CSJ) or a spontaneous speech corpus such as Switch Board Corpus may be used. The noise model outputs noise label information related to the source language text when the source language text is input. The noise label is information indicating the type of noise. FIG. 4 is a diagram illustrating a noise label. As shown in FIG. 4, in this embodiment, there are three types of noise labels, <F>, <D>, and 0. <F> is a noise label indicating a filler. <D> is a noise label indicating stagnation or rewording. 0 is a noise label indicating that there is no noise. The noise label information is information in which the type of noise label (<F>, <D>, 0 described above) and a word (specifically, morpheme) associated with each noise label are associated with each other. Noise label sequence.

FIG. 3 is a diagram for explaining the outline of the noise model. As shown in FIG. 3, the noise model is constructed using, for example, a bi-directional recursive neural network (BiRNN) that is widely used in part-of-speech tagging and proper expression extraction tasks. ing. The noise model may be constructed by a method using another neural network such as RNN or a method using a conditional random field such as CRF (Conditional random field). The noise model has been trained to predict an appropriate noise label for each input element (word-specific morpheme) of the input source language text when noise is next to the element. In noise addition using a noise model, a noise label sequence l=(l ₀ , l ₁ ,..., L _n ) is predicted from a morpheme sequence w=(w ₀ , w ₁ ,..., W _n ) of a source language text. Think of it as a serial labeling problem.

Now, I will explain the method of learning the noise model by taking the learning utterance data as "<F-E> Then start <D>> <F-A>". Here, “<F>” included in the learning utterance data indicates that “E” corresponds to the filler <F>. In this case, first, the morpheme sequence w=(<BOS>, then conference, start, <EOS>) is extracted from the learning utterance data. As shown in FIG. 3, in the morpheme sequence, each morpheme (including <BOS> and <EOS>) is associated with the time step from t=0 to 6. Next, a noise label sequence l=(<F>,0,0,<D>,<F>,0,0) is generated based on the same learning utterance data in which noise is annotated. It Finally, BiRNN is learned as a sequence labeling problem that predicts the noise label sequence 1 from the morpheme sequence w. In BiRNN, parameter learning is performed using the prediction error of the output sequence with respect to the input sequence.

Returning to FIG. 2, the translation model learning device 50 learns a translation model using the pseudo-parallel translation corpus constructed in the translation data generation device 10. As the translation model, a Transformer or RNN-based Sequence-to-Sequence model or the like may be used.

The translation data generation device 10 includes, as its functions, an analysis unit 11, a noise addition unit 12, a corpus construction unit 13, and a storage unit 14.

The analysis unit 11 acquires a source language text from the bilingual corpus DB 20 and performs a morphological analysis on the acquired source language text. That is, for example, when the analysis unit 11 acquires a source language text "I want to drive on a tourist route rather than a major expressway.", the morpheme sequence w=(main, major, high speed, Road, more than sightseeing, route, one, run, want).

The noise adding unit 12 adds noise to the source language text (specifically, the morpheme sequence extracted by the analyzing unit 11) to obtain the noise adding source language text. The noise imparting unit 12 imparts noise to the source language text using the noise model learned by the noise model learning device 40. The noise imparting unit 12 imparts a noise label to each morpheme according to the characteristic (specifically, morpheme) of each word of the source language text, and assigns the noise label to the word corresponding to the noise label (the word as noise. ) To add noise to the source language text. The noise adding unit 12 predicts a noise label sequence corresponding to the morpheme sequence of the input source language text by using the noise model, and inserts a noise label next to the corresponding morpheme sequence. Then, the noise adding unit 12 replaces the inserted noise label with a word representing noise, and obtains a noise-added source language text that is the final output and is a source language text with noise added. The noise imparting unit 12 has been described as imparting a noise label according to the morpheme of the source language text, but the present invention is not limited to this, and the noise label is assigned according to the part of speech or reading (pronunciation) of each word in the source language text. May be given. Further, the noise adding unit 12 may add a noise label according to two or more pieces of information such as a morpheme of a word, a part of speech, and a reading.

Specifically, the noise adding unit 12 first inputs the morpheme sequence of the source language text into the noise model, and acquires the output vector h _t of the noise model at each time step (each morpheme sequence). In the present embodiment, regarding the noise label at each time step, the one that has the maximum posterior probability of the noise label is not simply used as the estimation result, but a value exp(h _t /τ that is an exponent of the output vector h _t ) Is determined by sampling based on the multinomial distribution defined in (1). That is, the noise label l _{t at} each time step is estimated based on the following equation (1).
l _t to exp(h _t /τ) (1)
In the above formula (1), l _t is the estimation result of the noise label, h _t is the output vector of the noise model, and τ is the temperature parameter. The output vector h _t is represented by a three-dimensional vector for three label types (<F>, <D>, 0). The temperature parameter τ is a parameter for operating the strength of variation of the noise label. When the value of the temperature parameter τ is increased (τ→∞), the noise label probability distribution approaches a uniform distribution, and when the temperature parameter τ is decreased (τ→0), the noise label with the highest probability is selected.

For example, the determination of the noise label when the temperature parameter τ is relatively small will be described. Now, the output vector h _t of the noise model = (-0.1 (weight score of 0), 0.3 (weight score of <F>), -0.3 (weight score of <D>)), It is assumed that the temperature parameter τ=0.15. In this case, h _t / _τ =(−0.6666..., 2,−2). Taking an index so that the weight score of each noise label is 0 or more, exp(h _t /τ)=(0.51, 7.39, 0.13). When the weighted score is treated as a probability value and normalized so that the range is [0, 1] and all values are added, the probability distribution becomes (0.06 (probability that 0 is selected as a noise label), 0 .92 (the probability that <F> is selected as the noise label) and 0.02 (the probability that <D> is selected as the noise label)). Sampling (trial) the noise label only once based on such a probability distribution (multinomial distribution) corresponds to sampling from the categorical distribution. In this case, the establishment of the noise label <F> is extremely high at 92%, and the possibility of being selected as the sampling result is extremely high.

For example, the determination of the noise label when the temperature parameter τ is relatively large will be described. Now, the output vector h _t of the noise model = (-0.1 (weight score of 0), 0.3 (weight score of <F>), -0.3 (weight score of <D>)), It is assumed that the temperature parameter τ=1.0. In this case, h _t /τ=(−0.1, 0.3, −0.3). When the weight score of each noise label take exponential order to zero or more, the exp _(h t /τ)=(0.90,1.35,0.74). When the weighted score is treated as a probability value and normalized so that the range is [0, 1] and all values are added to become 1, the probability distribution is (0.30 (0 is the probability that 0 is selected as a noise label), 0 .45 (the probability that <F> is selected as a noise label) and 0.25 (the probability that <D> is selected as a noise label)). Thus, it can be seen that when the temperature parameter τ is increased, the noise label 0 and the noise label <D> are more easily selected as compared with the case where the temperature parameter τ=0.15 described above. As the temperature parameter τ approaches ∞, the probability of establishment of each noise label approaches 33.333...%, and the probability distribution approaches a uniform distribution.

As described above, the noise adding unit 12 samples the noise label according to the probability distribution based on the score of each noise label output from the noise model, with the feature (morpheme sequence) of each word of the source language text as an input, and the source language text. The noise label to be given to is determined. In the above description, the probability distribution defined based on the output value of the noise model is described as representing a polynomial distribution, but the present invention is not limited to this, and the probability distribution represents a Poisson distribution, a normal distribution, or the like. It may be.

The noise adding unit 12 then replaces the noise label sequence predicted using the noise model with a word representing noise. The noise adding unit 12 performs sampling, for example, from the vocabulary set V _type corresponding to each noise label based on the unigram probability. For example, when replacing the noise label <F> of the filler with the word representing the filler, the word representing the filler is determined based on the following equation (2).
w _t ′～V _<F>・・・(2)
In the above formula (2), V _<F> is a vocabulary set of noise labels <F>, and w _t ′ is a word representing a filler (noise) inserted at time step t. From the above, from the morpheme sequence w=(w ₀ , w ₁ ,..., W _n ) of the source language text, a sequence w′=(w ₀ , w ₁ , w ₁ ′, w ₂ , w ₂ ) including a word expressing noise ,', wn) is obtained.

The noise imparting unit 12 obtains a plurality of patterns of noise imparting source language text from one source language text. The noise adding unit 12 may derive, for example, a plurality of patterns of words (words representing noise) to be replaced with respect to one noise label, and obtain a plurality of patterns of noise added source language text from one source language text. Further, the noise adding unit 12 may derive a plurality of patterns of noise labels corresponding to each morpheme, and obtain a plurality of patterns of noise-added source language text from one source language text.

The corpus construction unit 13 constructs a pseudo bilingual corpus in which the noise-added source language text and the target language text corresponding to the noise-added source language text before the noise addition are associated with each other. FIG. 5 is a diagram showing an image of constructing a pseudo bilingual corpus. In the example shown in FIG. 5, the noise-added source language text “I want to drive the tourist route better than the main expressway”, which is the source language text before adding noise (“Early tourism than main expressways I want to run on the route”, etc.), and the target language text corresponding to the source language text before noise addition, “I would rather take a scenic route than a main highway.” A pseudo-translation corpus (corresponding to a translation pair) associated with is constructed.

The storage unit 14 is a DB that stores the pseudo bilingual corpus constructed by the corpus construction unit 13. The translation model learning device 50 learns a translation model using the pseudo-parallel translation corpus stored in the storage unit 14.

Next, the processing executed by the translation data generation system 1 will be described with reference to FIG. FIG. 6 is a flowchart showing the processing executed by the translation data generation system 1. It is assumed that the noise model is constructed (learned) by the noise model learning device 40 as a premise for executing the processing shown in FIG.

As shown in FIG. 6, in the translation data generation system 1, first, the analysis unit 11 of the translation data generation device 10 acquires the source language text from the parallel translation corpus DB 20 (step S1). Subsequently, the analysis unit 11 executes morphological analysis on the acquired source language text (step S2).

Next, the noise adding unit 12 of the translation data generating device 10 adds noise to the morpheme sequence extracted by the analyzing unit 11 to obtain a noise-added source language text (step S3). Specifically, the noise adding unit 12 predicts a noise label sequence corresponding to the morpheme sequence of the input source language text by using the noise model, and inserts a noise label next to the corresponding morpheme sequence. Then, the noise adding unit 12 replaces the inserted noise label with a word representing noise, and obtains a noise-added source language text that is the final output and is a source language text with noise added.

Subsequently, the corpus construction unit 13 of the translation data generation device 10 associates the noise-added source language text with the target language text corresponding to the noise-added source language text before the noise addition. A corpus is constructed (step S4).

Finally, the translation model learning device 50 learns a translation model using the pseudo bilingual corpus constructed by the corpus construction unit 13 (step S5). The above is an example of the processing executed by the translation data generation system 1.

Next, the function and effect of this embodiment will be described.

The translation data generation system 1 according to the present embodiment includes a noise adding unit 12 that adds noise to a source language text to obtain a noise added source language text, a noise added source language text, and noise of the noise added source language text. And a corpus constructing unit 13 for constructing a pseudo bilingual corpus in which the target language text corresponding to the source language text before being assigned is associated.

In the translation data generation system 1 according to the present embodiment, noise is added to the source language text, and the pseudo-translation corpus that associates the noise-added source language text with the target language text corresponding to the source language text before the noise addition is created. Be built. In this way, by constructing the bilingual corpus in which the source language text with noise is associated with the target language text corresponding to the source language text before adding noise, using such a bilingual corpus, for example, natural speech Even when noise such as a filler is included in the input, it is possible to appropriately derive the target language text corresponding to the source language text before noise addition. That is, according to the translation data generation system 1 according to the present embodiment, a robust corpus (pseudo-translation corpus) can be constructed with respect to natural utterances containing noise, and non-fluent natural utterances containing noise. Can be translated with high accuracy. When the information generated by the translation data generating system 1 is used for translation, it is not necessary to correct the user's utterance content and input it into the translation model, and the user's utterance content is not changed. You can enter the translation model. Further, for example, in the systems described in Japanese Patent Laid-Open Nos. 2010-079647 and 2007-057844, a speech recognition device is used to sequentially receive a user's utterance and make a rephrasing determination. The translation data generation system 1 according to the embodiment does not require a voice recognition device, and only needs to use the text information of the recognition result. As described above, in the translation data generation system 1 according to the present embodiment, it is possible to suppress the correction processing of the utterance content and the rewording determination processing, so that the processing load on the processing unit such as the CPU is reduced. It also plays a dynamic effect.

FIG. 7 is a table showing a translation example of this embodiment and a comparative example. As shown in the upper part of FIG. 7, there is a lack of translation in the comparative example with respect to the natural utterance input including noise. Further, as shown in the lower part of FIG. 7, a natural utterance input including noise is translated in a state including noise in the comparative example, and desired translation cannot be performed. As shown in the comparative example, conventionally, it has been difficult to accurately translate a natural utterance containing noise. In this regard, as shown in the upper and lower parts of FIG. 7, when the translation is performed in consideration of the pseudo bilingual corpus constructed by the translation data generation system 1 of the present embodiment, noise is naturally included. Translation errors are unlikely to occur even when uttered, and translation can be performed with high accuracy.

The translation data generation system 1 includes a translation model learning device 50 that learns a translation model using a pseudo bilingual corpus. By learning the translation model based on the constructed corpus, it is possible to translate the natural utterance containing noise with higher accuracy.

The translation data generation system 1 includes a noise model learning device 40 that learns a noise model related to noise addition to a source language text by using training data that is a source language text group including noise, and the noise addition unit 12 Adds noise to the source language text using a noise model. A noise model is learned based on a source language text group that includes noise in advance, and noise is added based on the noise model, so that noise that is likely to be actually included is easily added. The accuracy can be further improved.

In the translation data generation system 1, the noise imparting unit 12 imparts a noise label indicating the type of noise to each word of the source language text, and replaces the noise label with the word corresponding to the noise label. Add noise to linguistic text. Since the noise label corresponding to each word of the source language text is added and then the word (noise) corresponding to the noise label is derived, it is possible to ensure the easiness and validity of the noise addition.

In the translation data generation system 1, the noise adding unit 12 derives a plurality of patterns of words to be replaced with respect to one noise label, and obtains a plurality of patterns of noise adding source language text from one source language text. As a result, it is possible to efficiently enhance the pseudo bilingual corpus from one source language text and further improve the translation accuracy.

In the translation data generation system 1, the noise adding unit 12 derives a plurality of patterns of noise labels corresponding to each word, and obtains a plurality of patterns of noise-added source language text from one source language text. As a result, it is possible to efficiently enhance the pseudo bilingual corpus from one source language text and further improve the translation accuracy.

In the translation data generation system 1, the noise adding unit 12 adds a noise label according to the characteristics of each word in the source language text. This makes it possible to appropriately give each word a noise label relating to noise that is likely to be included in association with each word.

In the translation data generation system 1, the noise adding unit 12 adds a noise label according to at least one of a morpheme, a part of speech, and a word reading, which are characteristics of each word in the source language text. This makes it possible to appropriately give each word a noise label relating to noise that is likely to be included in association with each word.

In the translation data generation system 1, the noise adding unit 12 samples the noise label according to the probability distribution of each noise level based on the score of each noise label output from the noise model with the feature of each word of the source language text as an input. , Determine the noise label added to the source language text. As a result, for example, it is possible to give a noise label having a high score output from the noise model, and it is possible to appropriately give each word a noise label relating to noise that is likely to be included in association with each word.

Finally, the hardware configuration of the translation data generation device 10 will be described with reference to FIG. The translation data generation device 10 described above may be physically configured as a computer device including a processor 1001, a memory 1002, a storage 1003, a communication device 1004, an input device 1005, an output device 1006, a bus 1007, and the like.

Note that in the following description, the word "device" can be read as a circuit, device, unit, or the like. The hardware configuration of the translation data generating device 10 may be configured to include one or a plurality of each device illustrated in the figure, or may be configured not to include some devices.

Each function in the translation data generation device 10 causes a predetermined software (program) to be loaded on hardware such as the processor 1001 and the memory 1002, so that the processor 1001 performs an operation to perform communication by the communication device 1004 and the memory 1002. It is realized by controlling the reading and/or writing of data in the storage 1003.

The processor 1001 operates an operating system to control the entire computer, for example. The processor 1001 may be configured by a central processing unit (CPU) including an interface with peripheral devices, a control device, a calculation device, a register, and the like. For example, the control function of the noise adding unit 12 and the like of the translation data generating device 10 may be realized by the processor 1001.

Further, the processor 1001 reads a program (program code), a software module and data from the storage 1003 and/or the communication device 1004 into the memory 1002, and executes various processes according to these. As the program, a program that causes a computer to execute at least part of the operations described in the above-described embodiments is used. For example, the control function of the noise adding unit 12 and the like of the translation data generating device 10 may be realized by a control program stored in the memory 1002 and operated by the processor 1001, and other function blocks are similarly realized. Good. Although it has been described that the various processes described above are executed by one processor 1001, they may be executed simultaneously or sequentially by two or more processors 1001. The processor 1001 may be implemented by one or more chips. The program may be transmitted from the network via an electric communication line.

The memory 1002 is a computer-readable recording medium, and is composed of, for example, at least one of ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (ElectricallyErasable Programmable ROM), RAM (Random Access Memory), and the like. May be done. The memory 1002 may be called a register, a cache, a main memory (main storage device), or the like. The memory 1002 can store an executable program (program code), a software module, etc. for implementing the wireless communication method according to the embodiment of the present invention.

The storage 1003 is a computer-readable recording medium, for example, an optical disc such as a CD-ROM (Compact Disc ROM), a hard disc drive, a flexible disc, a magneto-optical disc (for example, a compact disc, a digital versatile disc, a Blu-ray disc). (Registered trademark) disk), smart card, flash memory (eg, card, stick, key drive), floppy (registered trademark) disk, magnetic strip, and the like. The storage 1003 may be called an auxiliary storage device. The storage medium described above may be, for example, a database including the memory 1002 and/or the storage 1003, a server, or another appropriate medium.

The communication device 1004 is hardware (transmission/reception device) for performing communication between computers via a wired and/or wireless network, and is also called, for example, a network device, a network controller, a network card, a communication module, or the like.

The input device 1005 is an input device (for example, a keyboard, a mouse, a microphone, a switch, a button, a sensor, etc.) that receives an input from the outside. The output device 1006 is an output device (for example, a display, a speaker, an LED lamp, etc.) that performs output to the outside. The input device 1005 and the output device 1006 may be integrated (for example, a touch panel).

Each device such as the processor 1001 and the memory 1002 is connected by a bus 1007 for communicating information. The bus 1007 may be configured with a single bus or different buses among devices.

The translation data generation device 10 includes hardware such as a microprocessor, a digital signal processor (DSP), an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), and an FPGA (Field Programmable Gate Array). May be included, and a part or all of each functional block may be realized by the hardware. For example, processor 1001 may be implemented with at least one of these hardware.

Although the present embodiment has been described above in detail, it is obvious to those skilled in the art that the present embodiment is not limited to the embodiment described in this specification. The present embodiment can be implemented as modified and changed modes without departing from the spirit and scope of the present invention defined by the description of the claims. Therefore, the description of the present specification is for the purpose of exemplifying explanation, and does not have any restrictive meaning to the present embodiment. For example, the translation data generation system according to one aspect of the present invention may add a noise word (filler, stagnation, rewording, etc.) defined in advance to a random position of the source language text. The word (noise) to be added to the random position may be randomly selected from noise word candidates, for example. In such a configuration, even if there is no noise model learning data (labeled data), a noise model that randomly adds noise can be constructed as long as the noise word can be defined.

Each aspect/embodiment described in this specification is LTE (Long Term Evolution), LTE-A (LTE-Advanced), SUPER 3G, IMT-Advanced, 4G, 5G, FRA (Future Radio Access), W-CDMA. (Registered trademark), GSM (registered trademark), CDMA2000, UMB (Ultra Mobile Broad-band), IEEE 802.11 (Wi-Fi), IEEE 802.16 (WiMAX), IEEE 802.20, UWB (Ultra-Wide) Band), Bluetooth (registered trademark), a system using other appropriate systems, and/or a next-generation system extended based on the above.

As long as there is no contradiction, the order of the processing procedures, flowcharts, etc. of each aspect/embodiment described in this specification may be changed. For example, the methods described herein present elements of the various steps in a sample order, and are not limited to the specific order presented.

Information that has been input and output may be stored in a specific location (for example, memory), or may be managed in a management table. Information that is input/output may be overwritten, updated, or added. The output information and the like may be deleted. The input information and the like may be transmitted to another device.

The determination may be performed by a value represented by 1 bit (whether 0 or 1), may be performed by a Boolean value (Boolean: true or false), and may be performed by comparing numerical values (for example, a predetermined value). Value comparison).

Each aspect/embodiment described in the present specification may be used alone, may be used in combination, or may be switched according to execution. Further, the notification of the predetermined information (for example, the notification of “being X”) is not limited to the explicit notification, and is performed implicitly (for example, the notification of the predetermined information is not performed). Good.

Software, whether called software, firmware, middleware, microcode, hardware description language, or any other name, instructions, instruction sets, code, code segments, program code, programs, subprograms, software modules. , Application, software application, software package, routine, subroutine, object, executable, thread of execution, procedure, function, etc. should be construed broadly.

Also, software, instructions, etc. may be transmitted and received via a transmission medium. For example, the software may use wired technologies such as coaxial cable, fiber optic cable, twisted pair and digital subscriber line (DSL) and/or wireless technologies such as infrared, wireless and microwave to websites, servers, or other When transmitted from a remote source, these wireline and/or wireless technologies are included within the definition of transmission medium.

The information, signals, etc. described herein may be represented using any of a variety of different technologies. For example, data, instructions, commands, information, signals, bits, symbols, chips, etc. that may be referred to throughout the above description include voltage, current, electromagnetic waves, magnetic fields or magnetic particles, optical fields or photons, or any of these. May be represented by a combination of

The terms described in this specification and/or the terms necessary for understanding this specification may be replaced with terms having the same or similar meanings.

Further, the information, parameters, etc. described in this specification may be represented by absolute values, may be represented by relative values from predetermined values, or may be represented by other corresponding information. ..

User terminals are defined by those skilled in the art as mobile communication terminals, subscriber stations, mobile units, subscriber units, wireless units, remote units, mobile devices, wireless devices, wireless communication devices, remote devices, mobile subscriber stations, access terminals, It may also be referred to as a mobile terminal, wireless terminal, remote terminal, handset, user agent, mobile client, client, or some other suitable term.

As used herein, the terms "determining" and "determining" may encompass a wide variety of actions. "Judgment" and "decision" are, for example, calculating, computing, processing, deriving, investigating, looking up (e.g., table, database or another). (Search in data structure), ascertaining (ascertaining) can be regarded as "judgment" and "decision". In addition, “decision” and “decision” include receiving (eg, receiving information), transmitting (eg, transmitting information), input (input), output (output), access (accessing) (for example, accessing data in a memory) may be regarded as “judging” and “deciding”. In addition, "judgment" and "decision" are considered to be "judgment" and "decision" when things such as resolving, selecting, choosing, selecting, establishing, and comparing are done. May be included. That is, the “judgment” and “decision” may include considering some action as “judgment” and “decision”.

As used herein, the phrase “based on” does not mean “based only on,” unless expressly specified otherwise. In other words, the phrase "based on" means both "based only on" and "based at least on."

Where the designations "first," "second," etc. are used herein, any reference to that element does not generally limit the amount or order of those elements. These designations may be used herein as a convenient way of distinguishing between two or more elements. Thus, references to the first and second elements do not imply that only two elements may be employed therein, or that the first element must precede the second element in any way.

As long as the terms “include”, “including” and variations thereof are used in the present specification or claims, these terms are the same as the term “comprising”. It is intended to be comprehensive. Furthermore, the term "or" as used in the specification or claims is not intended to be an exclusive OR.

In the present specification, a plurality of devices are also included, unless it is a device in which only one is clearly present in terms of context or technology.

The whole of this disclosure includes a plurality unless the context clearly shows the singular.

1... Translation data generation system, 12... Noise adding unit, 13... Corpus building unit, 40... Noise model learning device (noise model learning unit), 50... Translation model learning device (Translation model learning unit).

Claims

A noise adding unit that adds noise to the source language text to obtain noise
Translation data generation including a corpus construction unit that constructs a pseudo bilingual corpus in which the noise-added source language text and the target language text corresponding to the noise-added source language text before noise addition are associated with each other. system.
The translation data generation system according to claim 1, further comprising a translation model learning unit that learns a translation model using the pseudo bilingual corpus.
Using training data that is a source language text group containing noise, a noise model learning unit that learns a noise model related to the addition of noise to the source language text is further provided.
The translation data generation system according to claim 1, wherein the noise adding unit adds noise to the source language text using the noise model.
The noise adding unit adds a noise label indicating a type of noise to each word of the source language text, and replaces the noise label with a word corresponding to the noise label to add noise to the source language text, The data generation system for translation according to claim 3.
The translation data generation system according to claim 4, wherein the noise adding unit derives a plurality of patterns of words to be replaced with respect to one noise label, and obtains a plurality of patterns of the noise added source language text from one source language text. ..
The translation data generation system according to claim 4, wherein the noise adding unit derives a plurality of patterns of the noise label corresponding to each word, and obtains a plurality of patterns of the noise-added source language text from one source language text. ..
The translation data generation system according to any one of claims 4 to 6, wherein the noise adding unit adds the noise label according to the characteristics of each word of the source language text.
The translation data generation system according to claim 7, wherein the noise adding unit adds the noise label according to at least one of a morpheme, a part of speech, and a reading of the word, which is a feature of each word of the source language text. ..
The noise adding unit samples the noise label according to the probability distribution of each noise level based on the score of each noise label output from the noise model with the feature of each word of the source language text as an input, and adds the noise label to the source language text. 9. The translation data generation system according to claim 7, which determines a noise label.
The translation data generation system according to any one of claims 3 to 9, wherein the noise model is constructed by a method using a conditional random field or a neural network.