CN112686060B - Text translation method, device, electronic equipment and storage medium - Google Patents
Text translation method, device, electronic equipment and storage medium Download PDFInfo
- Publication number
- CN112686060B CN112686060B CN202011593186.8A CN202011593186A CN112686060B CN 112686060 B CN112686060 B CN 112686060B CN 202011593186 A CN202011593186 A CN 202011593186A CN 112686060 B CN112686060 B CN 112686060B
- Authority
- CN
- China
- Prior art keywords
- text
- source language
- layer
- feature
- domain
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000013519 translation Methods 0.000 title claims abstract description 344
- 238000000034 method Methods 0.000 title claims abstract description 80
- 238000012549 training Methods 0.000 claims abstract description 28
- 230000004927 fusion Effects 0.000 claims description 165
- 238000004364 calculation method Methods 0.000 claims description 32
- 230000003993 interaction Effects 0.000 claims description 22
- 238000004590 computer program Methods 0.000 claims description 10
- 230000015654 memory Effects 0.000 claims description 10
- 230000000694 effects Effects 0.000 abstract description 27
- 230000008569 process Effects 0.000 description 11
- 238000005457 optimization Methods 0.000 description 7
- 238000013528 artificial neural network Methods 0.000 description 6
- 238000004891 communication Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 6
- 230000007246 mechanism Effects 0.000 description 5
- 230000008859 change Effects 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 238000012015 optical character recognition Methods 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 238000013145 classification model Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000013518 transcription Methods 0.000 description 1
- 230000035897 transcription Effects 0.000 description 1
- 239000013598 vector Substances 0.000 description 1
Landscapes
- Machine Translation (AREA)
Abstract
The invention provides a text translation method, a text translation device, electronic equipment and a storage medium, wherein the method comprises the following steps: determining a source language text and the field to which the source language text belongs; inputting the source language text into a domain machine translation model corresponding to the domain to obtain a target language text output by the domain machine translation model; the domain machine translation model is obtained based on training of sample source language texts and sample target language texts under the corresponding domain; the domain machine translation model is used for translating the text based on the text characteristics of the source language text in the belonging domain and the general scene. According to the method, the device, the electronic equipment and the storage medium, machine translation is carried out by combining text features in the specific field and the general scene, the translation effect of the text in the specific field is improved, and meanwhile, the translation effect of the text in the general scene is ensured not to be reduced, so that the problem that the text translation effect is greatly reduced due to field classification errors is avoided.
Description
Technical Field
The present invention relates to the field of natural language processing technologies, and in particular, to a text translation method, a device, an electronic apparatus, and a storage medium.
Background
Machine translation is a process of converting one natural language (source language) into another natural language (target language) by using a computer, and is currently focused on machine translation of text in a source language in combination with a user's field of use, that is, an application field in which the content of speech of a user is considered in machine translation.
The machine translation in the specific field is generally based on a general machine translation system, and model iterative optimization is performed by combining training data in the specific field, so that a machine translation model in the specific field is obtained, and the machine translation effect in the specific field is greatly improved. However, the machine translation model in a specific field is difficult to guarantee the translation effect of the text in other fields, and if the field classification error is encountered, the text in a certain field is input into the machine translation model in other fields, which directly leads to that the translation effect of the text cannot meet the expectations.
Disclosure of Invention
The invention provides a text translation method, a text translation device, electronic equipment and a storage medium, which are used for solving the defect of low reliability of machine translation in a specific field in the prior art.
The invention provides a text translation method, which comprises the following steps:
determining a source language text and the field to which the source language text belongs;
Inputting the source language text into a domain machine translation model corresponding to the domain to obtain a target language text output by the domain machine translation model;
The domain machine translation model is obtained based on training of sample source language texts and sample target language texts under corresponding domains; the domain machine translation model is used for translating the text based on the text characteristics of the source language text in the domain and the general scene.
According to the present invention, a text translation method is provided, in which the source language text is input into a domain machine translation model corresponding to a domain to which the source language text belongs, to obtain a target language text output by the domain machine translation model, including:
Inputting the source language text to a domain coding layer of the domain machine translation model to obtain domain text characteristics output by the domain coding layer;
inputting the source language text to a universal coding layer of the domain machine translation model to obtain universal text characteristics output by the universal coding layer;
Inputting the field text features and the general text features into a source language fusion layer of the field machine translation model to obtain source language text features output by the source language fusion layer;
And inputting the source language text characteristics to a characteristic decoding layer of the domain machine translation model to obtain the target language text output by the characteristic decoding layer.
According to the present invention, the text translation method is provided, the inputting the text features in the domain and the general text features into a source language fusion layer of a machine translation model in the domain, to obtain the text features in the source language output by the source language fusion layer, and the method includes:
inputting the field text features and the general text features to a source language weight calculation layer of the source language fusion layer to obtain source language weights output by the source language weight calculation layer;
And inputting the field text features, the general text features and the source language weights to a source language weighted fusion layer of the source language fusion layer to obtain the source language text features output by the source language weighted fusion layer.
According to the text translation method provided by the invention, the parameters of the universal coding layer are consistent with the parameters of the coding part in the universal machine translation model;
The parameters of the domain coding layer are obtained based on the training of the sample source language text and the sample target language text under the corresponding domain on the basis of the parameters of the coding part in the universal machine translation model.
According to the invention, the text translation method is provided, the source language text feature is input to a feature decoding layer of the domain machine translation model, and the target language text output by the feature decoding layer is obtained, and the method comprises the following steps:
inputting the source language text characteristics and the translation text pair characteristics into the characteristic decoding layer to obtain the target language text output by the characteristic decoding layer;
The translation text pair characteristics are obtained by text encoding of a text translation pair matched with the source language text.
According to the present invention, the text translation method is provided, in which the source language text feature and the translated text pair feature are input to the feature decoding layer to obtain the target language text output by the feature decoding layer, and the method includes:
inputting the source language text features, the translation text pair features and the decoding state of the last decoding moment into a feature fusion layer of the feature decoding layer, and fusing the source language text features and the translation text pair features by the feature fusion layer based on the decoding state of the last decoding moment to obtain fusion features of the current decoding moment output by the feature fusion layer;
Inputting the fusion characteristic of the current decoding moment and the decoding result of the last decoding moment into a decoding layer of the characteristic decoding layer to obtain the decoding state and decoding result of the current decoding moment output by the decoding layer;
the target language text is the decoding result of the final decoding moment.
According to the invention, the method for translating text includes inputting the source language text feature, the translated text pair feature and the decoding state of the last decoding time to a feature fusion layer of the feature decoding layer, and fusing the source language text feature and the translated text pair feature by the feature fusion layer based on the decoding state of the last decoding time to obtain the fused feature of the current decoding time output by the feature fusion layer, wherein the method comprises the following steps:
Inputting the source language text feature, the translation text pair feature and the decoding state of the last decoding moment to an attention interaction layer of the feature fusion layer, and performing attention interaction on the source language text feature, the decoding state of the last decoding moment and the decoding state of the translation text pair feature by the attention interaction layer to obtain the source language context feature and the translation pair context feature of the current decoding moment output by the attention interaction layer;
and inputting the source language context characteristics and the translation pair context characteristics at the current decoding time to a context fusion layer of the characteristic fusion layer to obtain fusion characteristics at the current decoding time output by the context fusion layer.
According to the present invention, the text translation method inputs the context feature of the source language and the translation pair context feature of the current decoding time to the context fusion layer of the feature fusion layer, and obtains the fusion feature of the current decoding time output by the context fusion layer, including:
Inputting the context characteristics of the source language at the current decoding time and the translation pair context characteristics into a fusion weight calculation layer of the context fusion layer to obtain fusion weights at the current decoding time output by the fusion weight calculation layer;
and inputting the source language context feature and the translation pair context feature at the current decoding time and the fusion weight at the current decoding time into a context weighted fusion layer of the context fusion layer to obtain the fusion feature at the current decoding time output by the context weighted fusion layer.
According to the invention, a text translation method is provided, and the translation text pair is determined based on the following steps:
Respectively carrying out similarity calculation on each candidate translation text pair belonging to the same field as the source language text and the source language text to obtain the similarity between the source language text and each candidate translation text pair;
and using the candidate translation text pair corresponding to the maximum similarity as the translation text pair matched with the source language text.
The invention also provides a text translation device, which comprises:
The text determining unit is used for determining a source language text and the field to which the source language text belongs;
The machine translation unit is used for inputting the source language text into a domain machine translation model corresponding to the domain to obtain a target language text output by the domain machine translation model;
The domain machine translation model is obtained based on training of sample source language texts and sample target language texts under corresponding domains; the domain machine translation model is used for translating the text based on the text characteristics of the source language text in the domain and the general scene.
The invention also provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the steps of any one of the text translation methods described above when executing the computer program.
The present invention also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of the text translation method as described in any of the above.
According to the text translation method, the device, the electronic equipment and the storage medium, the text features of the source language text in the belonging field and the general scene are embedded into the field machine translation model in the corresponding field, so that the field machine translation model can be combined with the text features in the specific field and the general scene to carry out machine translation, the translation effect of the text in the specific field is improved, the translation effect of the text in the general scene is not reduced, and the problem that the text translation effect is greatly reduced due to field classification errors is avoided.
Drawings
In order to more clearly illustrate the invention or the technical solutions of the prior art, the following description will briefly explain the drawings used in the embodiments or the description of the prior art, and it is obvious that the drawings in the following description are some embodiments of the invention, and other drawings can be obtained according to the drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flow chart of a text translation method provided by the invention;
FIG. 2 is a flow chart of an embodiment of step 120 in the text translation method provided by the present invention;
FIG. 3 is a flowchart illustrating an embodiment of step 123 in the text translation method according to the present invention;
FIG. 4 is a schematic diagram of a domain machine translation model provided by the present invention;
FIG. 5 is a flow chart of an embodiment of step 124 in the text translation method provided by the present invention;
FIG. 6 is a flowchart illustrating an embodiment of step 1241 in the text translation method provided by the present invention;
FIG. 7 is a flow chart of an embodiment of step 1241-1 of the text translation method provided by the present invention;
FIG. 8 is a schematic diagram of a text translation device according to the present invention;
fig. 9 is a schematic structural diagram of an electronic device provided by the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is apparent that the described embodiments are some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Currently, machine translation for a specific field generally needs to perform model iterative optimization based on a general machine translation model by combining training data of the specific field and a scene. The machine translation model in the specific field obtained by the method has difficult guarantee on the translation effect of texts in other fields. If a wrong domain classification is encountered, for example, the sentence "Michael Jordan WILL PLAY THE next game" in the sports domain "should be translated into" Michael Gerdan will participate in the next game. "and if misclassified into the" game "field, will be translated into" Michael Gerdan "and will play a game. ", resulting in a significant drop in the translation effect of the sentence.
In order to solve the problem, the embodiment of the invention provides a text translation method in a specific field. Fig. 1 is a schematic flow chart of a text translation method provided by the present invention, as shown in fig. 1, the method includes:
Step 110, determining the source language text and the domain to which the source language text belongs.
The source language text is the text to be translated, the language applied by the source language text is the source language, the language applied by the text is obtained after translation, namely the target language, and the corresponding obtained translation result is the target language text. The source language text may be directly input by a user, or may be obtained by performing voice transcription on the acquired audio, or may be obtained by acquiring an image through an image acquisition device such as a scanner, a mobile phone, a camera, etc., and performing OCR (Optical Character Recognition ) on the image, which is not particularly limited in the embodiment of the present invention.
The field to which the source language text belongs may be selected and input by the user while the source language text is input, or may be obtained by performing field classification on the source language text based on a field classification model obtained by training in advance, which is not particularly limited in the embodiment of the present invention. The fields referred to herein may be divided according to the actual application scenario, for example, the source language text "michael jordan will participate in the next game" belongs to the sports field, and "Zhou Jielun released a new album" belongs to the music field.
Step 120, inputting the source language text into a domain machine translation model corresponding to the domain to obtain a target language text output by the domain machine translation model;
The domain machine translation model is obtained based on training of sample source language texts and sample target language texts under the corresponding domain; the domain machine translation model is used for translating the text based on the text characteristics of the source language text in the belonging domain and the general scene.
Specifically, unlike the machine translation model in the general scene, the domain machine translation model is used for implementing machine translation of the source language text belonging to the specific domain, and when the domain machine translation model translates the source language text of the specific domain, the translation effect of the domain machine translation model is better than that of the machine translation model in the general scene.
Before executing step 120, domain machine translation models corresponding to the respective domains may be obtained in advance, where a correspondence between domains and domain machine translation models may be one-to-one, that is, one domain corresponds to one domain machine translation model, or may be many-to-one, that is, a plurality of domains share one domain machine translation model, which is not limited in this embodiment of the present invention. For example, a game field may correspond to a game field machine translation model, and a sports field may also correspond to a sports field machine translation model.
After determining the domain to which the source language text belongs, determining a domain machine translation model corresponding to the domain to which the source language text belongs, and inputting the source language text into the domain machine translation model of the corresponding domain for machine translation. In the machine translation process, the field machine translation model can respectively extract text characteristics of the source language text under the field and text characteristics of the source language text under the general scene, so that in the text translation process of the specific field, the text characteristics under the general scene are referred to, and the translation effect of the text of the specific field is improved while the translation effect of the text of the general scene is not reduced.
The domain machine translation model may also be obtained by training, before performing step 120, and specifically may be trained by the following steps: firstly, collecting a large number of sample source language texts and corresponding sample target language texts in the corresponding field; and then training the initial model based on the sample source language text and the sample target language text in the corresponding field, thereby obtaining a field machine translation model in the corresponding field. Preferably, the initial model herein may be a machine translation model in a general scenario.
According to the method provided by the embodiment of the invention, the text features of the source language text in the belonging field and the general scene are embedded into the field machine translation model in the corresponding field, so that the field machine translation model can be combined with the text features in the specific field and the general scene to carry out machine translation, the translation effect of the text in the specific field is improved, and meanwhile, the translation effect of the text in the general scene is ensured not to be reduced, so that the problem of greatly reduced text translation effect caused by field classification errors is avoided.
Based on the embodiment, the domain machine translation model can respectively extract the domain text characteristics of the source language text in the domain and the general text characteristics in the general scene, and the domain text characteristics and the general text characteristics are fused for text translation; fig. 2 is a schematic flow chart of an embodiment of step 120 in the text translation method provided in the present invention, as shown in fig. 2, step 120 includes:
Step 121, inputting the source language text to a domain coding layer of a domain machine translation model to obtain domain text characteristics output by the domain coding layer;
And step 122, inputting the source language text to a universal coding layer of the domain machine translation model to obtain universal text characteristics output by the universal coding layer.
Specifically, the domain coding layer is used for extracting the semantics of the source language text in the corresponding domain and outputting the corresponding text characteristics, namely the domain text characteristics. The universal coding layer is used for extracting the semantics of the source language text in a universal scene and outputting corresponding text features, namely universal text features. Here, since the domain coding layer and the universal coding layer code the source language text in different domain scenes, the semantics of the source language text represented in different domain scenes are different, and the corresponding obtained domain text features and universal text features are also different.
There are many ways to perform feature encoding on the text, for example, context feature extraction of word vectors of individual segmented words in the input text through a long short-term memory network, or text feature extraction through a combination of self-attention mechanisms and FNN (Feedforward Neural Network, feed-forward neural network). The domain coding layer and the general coding layer may adopt the same network structure and different parameters, or may adopt different network structures and different parameters, which is not particularly limited in the embodiment of the present invention.
In addition, the execution sequence of step 121 and step 122 is not limited in the embodiment of the present invention, and step 121 may be performed before or after step 122, or may be performed synchronously with step 122.
And step 123, inputting the field text features and the general text features into a source language fusion layer of the field machine translation model to obtain the source language text features output by the source language fusion layer.
And 124, inputting the source language text characteristics into a characteristic decoding layer of the domain machine translation model to obtain the target language text output by the characteristic decoding layer.
Specifically, the source language fusion layer is used for carrying out feature fusion on the field text features and the general text features, so that the source language text features containing the semantics of the source language text in the field and the general scene are output. The feature fusion mode can be direct splicing or weighted fusion, the weight during weighted fusion can be learned in the training process of the domain machine translation model, and can be dynamically changed according to the domain text features and the general text features input into the source language fusion layer, and the embodiment of the invention is not particularly limited.
The feature decoding layer is used for carrying out feature decoding on the text features of the source language, and fully references the text features in the corresponding field and the general scene in the decoding process, so that more accurate target language text is obtained. Further, the decoding portion may be implemented by a beam search method of a general machine translation model, which is not described herein.
According to the method provided by the embodiment of the invention, the machine translation is performed by fusing the text features in the field and the general text features, so that the translation effect of the text in the specific field is improved, the translation effect of the text in the general scene is not reduced, and the reliability and the accuracy of the text translation are comprehensively ensured.
Based on any of the above embodiments, fig. 3 is a schematic flow chart of an implementation of step 123 in the text translation method provided by the present invention, and as shown in fig. 3, step 123 includes:
Step 1231, inputting the text features of the domain and the general text features to a source language weight calculation layer of a source language fusion layer to obtain source language weights output by the source language weight calculation layer;
And step 1232, inputting the text features of the domain, the general text features and the source language weights into a source language weighted fusion layer of the source language fusion layer to obtain the text features of the source language output by the source language weighted fusion layer.
Specifically, the source language weight calculation layer is used for calculating the weight required by the two when the weighted fusion is carried out according to the field text characteristics and the general text characteristics, namely the source language weight. The source language weight calculation layer can be realized through a feedforward neural network, for example, can be realized by utilizing a forward full-connection layer, and can be specifically expressed as the following formula:
g＝σ(Wdomain*Cdomain+Wcom*Ccom)
where the source language weight g is a gating unit scalar having a value between 0 and 1. For a scale representing that the source language text feature originates from the domain text feature C domain, respectively 1-g may be used to represent a scale of the source language text feature originating from the generic text feature C com. σ is the activation function, and W domain and W com are the parameters obtained by training.
The source language weighting fusion layer can perform weighted summation on the field text feature and the general text feature based on the source language weight, so as to obtain a source language text feature C, which can be expressed as the following formula:
C＝g*Cdomain+(1-g)*Ccom
according to the method provided by the embodiment of the invention, the source language weight is calculated based on the field text feature and the general text feature, the flexible fusion of the field text feature and the general text feature is realized based on the weighted summation of the source language weight, so that the translation effect is more stable, and the problem of robustness to text translation in the field of the current hard allocation source language text is avoided.
Based on any of the above embodiments, in the domain machine translation model, parameters of the universal coding layer are consistent with parameters of the coding part in the universal machine translation model; the parameters of the domain coding layer are obtained based on the training of the sample source language text and the sample target language text under the corresponding domain on the basis of the parameters of the coding part in the universal machine translation model.
Specifically, the universal machine translation model is a machine translation model applied to a universal scene, and an encoding part of the universal machine translation model is used for encoding the input source language text in the universal scene. Thereby obtaining a source language text is a common text feature of (c). Due to the coding part of the general machine translation model and the tasks executed by the general coding layer in the domain machine translation model the scenes are all consistent and the scene is uniform, the coding part of the general machine translation model can be directly used parameters migrate to the generic coding layer of the domain machine translation model.
Compared with the coding part of the general machine translation model, the field coding layer has the same coding task and different field scenes, and can take the parameters of the coding part of the general machine translation model as the initialization parameters of the field coding layer, and based on the initialization parameters, the initialized parameters are iteratively updated based on the sample source language text and the sample target language text in the corresponding field, thereby obtaining the field coding layer.
Based on any one of the above embodiments, fig. 4 is a schematic structural diagram of a domain machine translation model provided by the present invention, and as shown in fig. 4, a text translation method based on the domain machine translation model includes the following steps:
The method comprises the steps of inputting source language texts into a coding layer of a domain machine translation model, coding the source language texts by the coding layer to obtain coded representations of the source language texts, inputting the coded representations of the source language texts into a domain coding layer and a general coding layer respectively, and extracting text features of the source language texts from corresponding domains and general scenes respectively to obtain domain text features C domain and general text features C com correspondingly. Here, the coding layer is provided with an N-layer coding structure, and the domain coding layer and the general coding layer are respectively provided with an M-layer coding structure, and each coding structure is in a form of combining a self-attention mechanism SelfAtt with the feedforward neural network FNN.
Thereupon, the domain text feature and the general text feature are input to the source language weight calculation layer, and the source language weight g is calculated by the source language weight calculation layer based on the formula g=σ (W domain*Cdomain+Wcom*Ccom). On the basis, the field text feature, the general text feature and the source language weight are input into a source language weighting fusion layer, and the source language weighting fusion layer calculates the source language text feature C based on a formula C=g×C domain+(1-g)*Ccom.
Then, the source language text features are input into a feature decoding layer, and feature decoding is carried out on the source language text features by the feature decoding layer, so that more accurate target language text is obtained. Here, the feature decoding layer arrangement is formed of L-layer decoding structures, each in the form of a combination of self-attention mechanism SelfAtt, cross-attention mechanism CrossAtt, and feed-forward neural network FNN.
Based on any of the above embodiments, model training of the domain machine translation model may be achieved by:
First, a sample source language text and a sample target language text in a generic scene may be determined based on a generic domain data set, and a generic machine translation model may be trained based on the sample source language text and the sample target language text in the generic scene. The general machine translation model herein may be a machine translation model of a conventional encoder-decoder structure applied to a general field, and specifically may include an encoding layer, a general encoding layer, and a feature decoding layer shown in fig. 4.
And copying the parameters of the universal coding layer in the universal machine translation model obtained through training to serve as initialization parameters of the field coding layer in the field machine translation model, and initializing the parameters of the source language weight calculation layer in the field machine learning model.
Based on the sample source language text and the sample target language text in the specific field, under the condition of fixing the parameters of the rest parts except the field coding layer and the source language weight calculation layer in the field machine translation model, the parameters of the field coding layer and the source language weight calculation layer are trained and updated, so that a final field machine translation model is obtained.
In the process, the parameters of the universal coding layer are used as the initialization parameters of the domain coding layer, so that the domain machine translation model can be converged more quickly. And the parameter training of the source language weight calculation layer can enable the adjustment of the source language weight to be more accurate.
Currently, domain machine translation generally needs model iterative optimization based on a general machine translation model by combining training data of a specific domain and a scene. Whereas iterative optimization of models generally requires a longer time, resulting in a time-limited improvement of the machine translation effect. In addition, training data in a specific field needs to be prepared in advance for iterative optimization of the model, the preparation time of the training data further prolongs the period of iterative optimization of the model, and the machine translation model cannot adapt to the change of an actual scene in time. In this regard, based on any of the embodiments described above, step 124 includes:
Inputting the source language text features and the translation text pair features into a feature decoding layer to obtain target language text output by the feature decoding layer;
the translation text pair characteristics are obtained by text encoding of a text translation pair matched with the source language text.
Specifically, the translation text pair is a set of text pairs which have completed translation, the translation text pair includes two texts, one of the languages of the text application is a source language, the other of the languages of the text application is a target language, and the translation text pair can be obtained by translating the source language into target voice or translating the target language into the source language.
The translation text pair matched with the source language text refers to a translation text pair similar to the source language text, and the translation text pair matched with the source language text can be one pair or multiple pairs. The similarity referred to herein may be similarity on a semantic level, similarity on a syntactic structure, similarity of an application word in a text, or the like, which is not particularly limited in the embodiment of the present invention. For example, the source language text is "I love work", and the translated text pair "I love you, I love you" that matches the source language text is retrieved.
When the feature decoding layer decodes the features of the source language text, the feature decoding layer can also refer to the thought of translating the text similar to the source language text in the translation text pair, decode the features of the source language text, and improve the accuracy of similar text translation by fully applying the existing translation text pair features, thereby obtaining more accurate target language text. For example, when the feature decoding layer decodes the source language text feature of "I love work", the text feature of "I love you" may be referred to by the translation text pair, so as to determine that the source language text feature of "I love work" is decoded as "I love.
According to the method provided by the embodiment of the invention, the feature of the text translation pair matched with the source language text is introduced in the feature decoding process, so that the feature decoding layer can take the information of the translation pair as a reference for translating and decoding the source language text, and the translation effect is optimized. According to the method provided by the embodiment of the invention, only the translation text pairs matched with the source language text are accumulated in advance, iterative optimization is not required to be carried out on the domain machine translation model again, and when the actual scene changes, the machine translation can be ensured to meet the scene change requirement in time by only accumulating the translation text pairs after the actual scene changes.
Based on any of the above embodiments, fig. 5 is a schematic flow chart of an embodiment of step 124 in the text translation method provided by the present invention, and as shown in fig. 5, step 124 includes:
step 1241, inputting the source language text feature, the translation text pair feature and the decoding state of the last decoding time to a feature fusion layer of a feature decoding layer, and fusing the source language text feature and the translation text pair feature by the feature fusion layer based on the decoding state of the last decoding time to obtain a fusion feature of the current decoding time output by the feature fusion layer;
Step 1242, inputting the fusion feature of the current decoding moment and the decoding result of the last decoding moment to the decoding layer of the feature decoding layer to obtain the decoding state and decoding result of the current decoding moment output by the decoding layer;
The target language text is the decoding result of the final decoding moment.
Specifically, the feature fusion layer is used for realizing feature fusion of the source language text features and the translation text pair features. And unlike feature fusion in the conventional sense, the feature fusion in the embodiment of the invention is dynamic, and the fusion mode is changed along with the change of the condition of feature decoding. The decoding status here reflects what is the case for feature decoding,
Further, the decoding status at the last decoding time contains history information generated in the decoding process before the decoding time. When feature fusion is carried out on the source language text features and the translation text features, the feature fusion layer can analyze and judge which information in the source language text features and which information in the translation text features should be focused on at the current decoding time based on the decoding state of the last decoding time, and can analyze and judge whether information in the source language text features or information in the translation text features should be focused more or not at the current decoding time, so that in the process of feature fusion, the information which needs focused attention is highlighted, the information which does not need focused attention is weakened, and the fusion features which are more suitable for the current decoding time are obtained.
After the fusion characteristic of the current decoding moment is obtained, the decoding layer can determine the decoding state of the current decoding moment based on the fusion characteristic of the current moment and the decoding state and decoding result of the last decoding moment. On the basis, the decoding layer can also decode based on the decoding state and the fusion characteristic of the current decoding moment and the decoding result of the last decoding moment, so as to obtain and output the decoding result of the current decoding moment. Here, the decoding result at any decoding time is a character sequence in which characters decoded at the decoding time are spliced with the decoding result corresponding to the previous decoding time.
According to the method provided by the embodiment of the invention, the source language text features and the translation text pair features are dynamically fused based on the decoding state of the last decoding moment, so that the fusion features themselves can highlight the information needing to be focused at the current decoding moment in the decoding process, and the accuracy of text translation is improved.
Based on any of the above embodiments, the feature fusion layer includes an attention interaction layer and a context fusion layer; fig. 6 is a flowchart of an embodiment of step 1241 in the text translation method provided by the present invention, where, as shown in fig. 6, step 1241 includes:
step 1241-1, inputting the source language text feature, the translation text pair feature, and the decoding state of the last decoding time to an attention interaction layer of the feature fusion layer, and performing attention interaction on the source language text feature, the decoding state of the last decoding time, and the translation text pair feature and the decoding state of the last decoding time by the attention interaction layer to obtain the source language context feature and the translation pair context feature of the current decoding time output by the attention interaction layer;
And step 1241-2, inputting the context characteristics of the source language at the current decoding time and the translation pair context characteristics into a context fusion layer of the characteristic fusion layer to obtain fusion characteristics of the current decoding time output by the context fusion layer.
Specifically, the attention interaction layer may analyze the importance of each feature included in the source language text feature and the importance of each feature included in the translated text to the feature based on the attention mechanism, so as to highlight the information needing to be focused in the source language text feature and the translated text feature, weaken the information not needing to be focused, and obtain the corresponding source language context feature and the corresponding translated text context feature.
Further, the attention interaction layer can perform attention interaction on the text features of the source language and the decoding state at the last decoding time, so that the attention weights of all the features in the text features of the source language are determined, all the features in the text features of the source language are weighted based on the attention weights of all the features, and the context features of the source language are obtained; similarly, the attention interaction layer may perform attention interaction on the translated text to the features and the decoding state at the last decoding time, thereby determining the attention weights of the translated text to each of the features, and weighting each of the features based on the attention weights of each of the features, thereby obtaining the translated text context feature.
The context fusion layer is used for carrying out feature fusion on the context features of the source language at the current decoding time and the translation pair context features, wherein the feature fusion can be carried out by carrying out weighted summation according to the fixed weight obtained by pre-learning, or carrying out the weighted summation on the context features according to the source language context features and the translation pair context features by carrying out dynamic calculation on the weight on the basis, or directly splicing the source language context features and the translation pair context features, and the embodiment of the invention is not limited in particular.
Based on any of the above embodiments, the context fusion layer includes a weight calculation layer and a weighted fusion layer; fig. 7 is a schematic flow chart of an embodiment of step 1241-1 in the text translation method provided by the present invention, where, as shown in fig. 7, step 1241-1 includes:
Step 1241-11, inputting the source language context feature and the translation pair context feature at the current decoding time to a fusion weight calculation layer of the context fusion layer to obtain the fusion weight at the current decoding time output by the fusion weight calculation layer;
and step 1241-12, inputting the source language context feature and the translation pair context feature at the current decoding time and the fusion weight at the current decoding time into a context weighted fusion layer of the context fusion layer to obtain the fusion feature at the current decoding time output by the context weighted fusion layer.
Specifically, the weight calculation layer is used for calculating the fusion weight required by the source language context feature and the translation pair context feature at the current decoding moment when the source language context feature and the translation pair context feature are subjected to weighted fusion. The weight calculation layer can be realized by a feedforward neural network, for example, can be realized by a forward full-connection layer, and can be specifically represented by the following formula:
g′＝σ(WS*CS+Wm*Cm)
Where the fusion weight g 'is a gating unit scalar having a value between 0 and 1, and is used to represent the proportion of the fusion feature derived from the source language context feature C S, and correspondingly 1-g' can be used to represent the proportion of the fusion feature derived from the translation to the context feature C m. σ is the activation function, and W S and W m are the parameters obtained by training.
The weighted fusion layer may perform weighted summation on the context feature and the translation of the source language context feature at the current decoding time based on the fusion weight at the current decoding time, so as to obtain a fusion feature C' at the current time, which may be specifically expressed as the following formula:
C′＝g′*CS+(1-g′)*Cm
According to the method provided by the embodiment of the invention, the fusion weight is calculated based on the source language context feature and the translation pair context feature at the current decoding moment, so that the dynamic fusion of the source language context feature and the translation pair context feature at the current decoding moment is realized, the fusion feature is more fit with the decoding requirement at the current decoding moment, the decoding precision is effectively improved, and the text translation effect is greatly improved.
Based on any of the above embodiments, the translation text pair that matches the source language text is determined based on the steps of:
Respectively carrying out similarity calculation on each candidate translation text pair belonging to the same field as the source language text and the source language text to obtain the similarity between the source language text and each candidate translation text pair;
and taking the candidate translation text pair corresponding to the maximum similarity as a translation text pair matched with the source language text.
Specifically, after determining the domain to which the source language text belongs, the source language text and each candidate translation text pair belonging to the same domain can be matched one by one, so that the matched translation text pair is selected. The candidate translation text pairs belonging to the same field are selected, the fact that the finally obtained translation text belongs to the same field for the source language text can be guaranteed, in the process of translating the text by referring to the translation text pairs, the similar translation ideas of the translation text pairs and the source language text on the semantic level, the syntactic structure or the containing word segmentation can be referred to, and information on the field level carried in the translation text pairs can be applied, so that the translation effect cannot be influenced by the field difference of the source language text, and the accuracy and the reliability of text translation are guaranteed.
The text translation device provided by the invention is described below, and the text translation device described below and the text translation method described above can be referred to correspondingly.
Based on any one of the above embodiments, fig. 8 is a schematic structural diagram of a text translation device provided by the present invention, and as shown in fig. 8, the text translation device includes a text determining unit 810 and a machine translation unit 820;
The text determining unit 810 is configured to determine a source language text and a domain to which the source language text belongs;
The machine translation unit 820 is configured to input the source language text into a domain machine translation model corresponding to the domain to obtain a target language text output by the domain machine translation model;
The domain machine translation model is obtained based on training of sample source language texts and sample target language texts under corresponding domains; the domain machine translation model is used for translating the text based on the text characteristics of the source language text in the domain and the general scene.
According to the method provided by the embodiment of the invention, the text features of the source language text in the belonging field and the general scene are embedded into the field machine translation model in the corresponding field, so that the field machine translation model can be combined with the text features in the specific field and the general scene to carry out machine translation, the translation effect of the text in the specific field is improved, and meanwhile, the translation effect of the text in the general scene is ensured not to be reduced, so that the problem of greatly reduced text translation effect caused by field classification errors is avoided.
Based on any of the above embodiments, the machine translation unit 820 includes:
a domain coding subunit, configured to input the source language text to a domain coding layer of the domain machine translation model, and obtain a domain text feature output by the domain coding layer;
The universal coding subunit is used for inputting the source language text to a universal coding layer of the domain machine translation model to obtain universal text characteristics output by the universal coding layer;
The source language fusion subunit is used for inputting the field text features and the general text features into a source language fusion layer of the field machine translation model to obtain source language text features output by the source language fusion layer;
And the feature decoding subunit is used for inputting the source language text features to a feature decoding layer of the domain machine translation model to obtain target language text output by the feature decoding layer.
Based on any of the above embodiments, the source language fusion subunit is configured to:
inputting the field text features and the general text features to a source language weight calculation layer of the source language fusion layer to obtain source language weights output by the source language weight calculation layer;
And inputting the field text features, the general text features and the source language weights to a source language weighted fusion layer of the source language fusion layer to obtain the source language text features output by the source language weighted fusion layer.
Based on any of the above embodiments, the parameters of the generic coding layer are consistent with the parameters of the coding part in the generic machine translation model;
The parameters of the domain coding layer are obtained based on the training of the sample source language text and the sample target language text under the corresponding domain on the basis of the parameters of the coding part in the universal machine translation model.
Based on any of the above embodiments, the signature decoding subunit is to:
inputting the source language text characteristics and the translation text pair characteristics into the characteristic decoding layer to obtain the target language text output by the characteristic decoding layer;
The translation text pair characteristics are obtained by text encoding of a text translation pair matched with the source language text.
Based on any of the above embodiments, the signature decoding subunit comprises:
The feature fusion module is used for inputting the source language text feature, the translation text pair feature and the decoding state of the last decoding moment into the feature fusion layer of the feature decoding layer, and the feature fusion layer fuses the source language text feature and the translation text pair feature based on the decoding state of the last decoding moment to obtain a fusion feature of the current decoding moment output by the feature fusion layer;
The decoding module is used for inputting the fusion characteristic of the current decoding moment and the decoding result of the last decoding moment into a decoding layer of the characteristic decoding layer to obtain the decoding state and the decoding result of the current decoding moment output by the decoding layer;
the target language text is the decoding result of the final decoding moment.
Based on any of the above embodiments, the feature fusion module is configured to:
Inputting the source language text feature, the translation text pair feature and the decoding state of the last decoding moment to an attention interaction layer of the feature fusion layer, and performing attention interaction on the source language text feature, the decoding state of the last decoding moment and the decoding state of the translation text pair feature by the attention interaction layer to obtain the source language context feature and the translation pair context feature of the current decoding moment output by the attention interaction layer;
and inputting the source language context characteristics and the translation pair context characteristics at the current decoding time to a context fusion layer of the characteristic fusion layer to obtain fusion characteristics at the current decoding time output by the context fusion layer.
Based on any of the above embodiments, the feature fusion module is configured to:
Inputting the context characteristics of the source language at the current decoding time and the translation pair context characteristics into a fusion weight calculation layer of the context fusion layer to obtain fusion weights at the current decoding time output by the fusion weight calculation layer;
and inputting the source language context feature and the translation pair context feature at the current decoding time and the fusion weight at the current decoding time into a context weighted fusion layer of the context fusion layer to obtain the fusion feature at the current decoding time output by the context weighted fusion layer.
Based on any of the above embodiments, the apparatus further includes a translation pair determining unit configured to:
Respectively carrying out similarity calculation on each candidate translation text pair belonging to the same field as the source language text and the source language text to obtain the similarity between the source language text and each candidate translation text pair;
and using the candidate translation text pair corresponding to the maximum similarity as the translation text pair matched with the source language text.
Fig. 9 illustrates a physical schematic diagram of an electronic device, as shown in fig. 9, which may include: processor (processor) 99, communication interface (Communications Interface) 920, memory (memory) 930, and communication bus 940, wherein processor 99, communication interface 920, memory 930 complete communication with each other through communication bus 940. Processor 99 may invoke logic instructions in memory 930 to perform a text translation method comprising: determining a source language text and the field to which the source language text belongs; inputting the source language text into a domain machine translation model corresponding to the domain to obtain a target language text output by the domain machine translation model; the domain machine translation model is obtained based on training of sample source language texts and sample target language texts under corresponding domains; the domain machine translation model is used for translating the text based on the text characteristics of the source language text in the domain and the general scene.
Further, the logic instructions in the memory 930 described above may be implemented in the form of software functional units and may be stored in a computer-readable storage medium when sold or used as a stand-alone product. Based on this understanding, the technical solution of the present invention may be embodied essentially or in a part contributing to the prior art or in a part of the technical solution, in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device (which may be a personal computer, a server, a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a usb disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.
In another aspect, the present invention also provides a computer program product comprising a computer program stored on a non-transitory computer readable storage medium, the computer program comprising program instructions which, when executed by a computer, enable the computer to perform a text translation method provided by the above methods, the method comprising: determining a source language text and the field to which the source language text belongs; inputting the source language text into a domain machine translation model corresponding to the domain to obtain a target language text output by the domain machine translation model; the domain machine translation model is obtained based on training of sample source language texts and sample target language texts under corresponding domains; the domain machine translation model is used for translating the text based on the text characteristics of the source language text in the domain and the general scene.
In yet another aspect, the present invention also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, is implemented to perform the text translation methods provided above, the method comprising: determining a source language text and the field to which the source language text belongs; inputting the source language text into a domain machine translation model corresponding to the domain to obtain a target language text output by the domain machine translation model; the domain machine translation model is obtained based on training of sample source language texts and sample target language texts under corresponding domains; the domain machine translation model is used for translating the text based on the text characteristics of the source language text in the domain and the general scene.
The apparatus embodiments described above are merely illustrative, wherein the elements illustrated as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art will understand and implement the present invention without undue burden.
From the above description of the embodiments, it will be apparent to those skilled in the art that the embodiments may be implemented by means of software plus necessary general hardware platforms, or of course may be implemented by means of hardware. Based on this understanding, the foregoing technical solution may be embodied essentially or in a part contributing to the prior art in the form of a software product, which may be stored in a computer readable storage medium, such as ROM/RAM, a magnetic disk, an optical disk, etc., including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the method described in the respective embodiments or some parts of the embodiments.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims (12)
1. A method of text translation, comprising:
determining a source language text and the field to which the source language text belongs;
Inputting the source language text into a domain machine translation model corresponding to the domain to obtain a target language text output by the domain machine translation model;
The domain machine translation model is obtained based on training of sample source language texts and sample target language texts under corresponding domains; the domain machine translation model is used for respectively extracting text features of the source language text in the domain and text features of the source language text in the general scene, and performing text translation based on the text features of the source language text in the domain and the general scene, wherein the text features in the domain represent semantics of the source language text in the domain, and the text features in the general scene represent semantics of the source language text in the general scene.
2. The text translation method according to claim 1, wherein the step of inputting the source language text into a domain machine translation model corresponding to the domain to obtain a target language text output by the domain machine translation model comprises the steps of:
Inputting the source language text to a domain coding layer of the domain machine translation model to obtain domain text characteristics output by the domain coding layer;
inputting the source language text to a universal coding layer of the domain machine translation model to obtain universal text characteristics output by the universal coding layer;
Inputting the field text features and the general text features into a source language fusion layer of the field machine translation model to obtain source language text features output by the source language fusion layer;
And inputting the source language text characteristics to a characteristic decoding layer of the domain machine translation model to obtain the target language text output by the characteristic decoding layer.
3. The text translation method according to claim 2, wherein the step of inputting the domain text feature and the general text feature to a source language fusion layer of the domain machine translation model to obtain the source language text feature output by the source language fusion layer includes:
inputting the field text features and the general text features to a source language weight calculation layer of the source language fusion layer to obtain source language weights output by the source language weight calculation layer;
And inputting the field text features, the general text features and the source language weights to a source language weighted fusion layer of the source language fusion layer to obtain the source language text features output by the source language weighted fusion layer.
4. The text translation method according to claim 2, wherein the parameters of the generic coding layer are consistent with the parameters of the coding part in the generic machine translation model;
The parameters of the domain coding layer are obtained based on the training of the sample source language text and the sample target language text under the corresponding domain on the basis of the parameters of the coding part in the universal machine translation model.
5. The text translation method according to claim 2, wherein the inputting the source language text feature to a feature decoding layer of the domain machine translation model to obtain the target language text output by the feature decoding layer comprises:
inputting the source language text characteristics and the translation text pair characteristics into the characteristic decoding layer to obtain the target language text output by the characteristic decoding layer;
The translation text pair characteristics are obtained by text encoding of a text translation pair matched with the source language text.
6. The text translation method according to claim 5, wherein said inputting the source language text feature and the translated text pair feature to the feature decoding layer to obtain the target language text output by the feature decoding layer comprises:
inputting the source language text features, the translation text pair features and the decoding state of the last decoding moment into a feature fusion layer of the feature decoding layer, and fusing the source language text features and the translation text pair features by the feature fusion layer based on the decoding state of the last decoding moment to obtain fusion features of the current decoding moment output by the feature fusion layer;
Inputting the fusion characteristic of the current decoding moment and the decoding result of the last decoding moment into a decoding layer of the characteristic decoding layer to obtain the decoding state and decoding result of the current decoding moment output by the decoding layer;
the target language text is the decoding result of the final decoding moment.
7. The text translation method according to claim 6, wherein the inputting the source language text feature and the translated text pair feature, and the decoding status of the last decoding time to the feature fusion layer of the feature decoding layer, the feature fusion layer fusing the source language text feature and the translated text pair feature based on the decoding status of the last decoding time to obtain the fused feature of the current decoding time output by the feature fusion layer includes:
Inputting the source language text feature, the translation text pair feature and the decoding state of the last decoding moment to an attention interaction layer of the feature fusion layer, and performing attention interaction on the source language text feature, the decoding state of the last decoding moment and the decoding state of the translation text pair feature by the attention interaction layer to obtain the source language context feature and the translation pair context feature of the current decoding moment output by the attention interaction layer;
and inputting the source language context characteristics and the translation pair context characteristics at the current decoding time to a context fusion layer of the characteristic fusion layer to obtain fusion characteristics at the current decoding time output by the context fusion layer.
8. The text translation method according to claim 7, wherein the inputting the source language context feature and the translation pair context feature at the current decoding time to the context fusion layer of the feature fusion layer, to obtain the fusion feature at the current decoding time output by the context fusion layer, includes:
Inputting the context characteristics of the source language at the current decoding time and the translation pair context characteristics into a fusion weight calculation layer of the context fusion layer to obtain fusion weights at the current decoding time output by the fusion weight calculation layer;
and inputting the source language context feature and the translation pair context feature at the current decoding time and the fusion weight at the current decoding time into a context weighted fusion layer of the context fusion layer to obtain the fusion feature at the current decoding time output by the context weighted fusion layer.
9. A method of text translation according to any of claims 5 to 8, wherein the translated text pairs are determined based on the steps of:
Respectively carrying out similarity calculation on each candidate translation text pair belonging to the same field as the source language text and the source language text to obtain the similarity between the source language text and each candidate translation text pair;
and using the candidate translation text pair corresponding to the maximum similarity as the translation text pair matched with the source language text.
10. A text translation device, comprising:
The text determining unit is used for determining a source language text and the field to which the source language text belongs;
The machine translation unit is used for inputting the source language text into a domain machine translation model corresponding to the domain to obtain a target language text output by the domain machine translation model;
The domain machine translation model is obtained based on training of sample source language texts and sample target language texts under corresponding domains; the domain machine translation model is used for respectively extracting text features of the source language text in the domain and text features of the source language text in the general scene, and performing text translation based on the text features of the source language text in the domain and the general scene, wherein the text features in the domain represent semantics of the source language text in the domain, and the text features in the general scene represent semantics of the source language text in the general scene.
11. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the steps of the text translation method according to any of claims 1 to 9 when the program is executed.
12. A non-transitory computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when executed by a processor, implements the steps of the text translation method according to any one of claims 1 to 9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011593186.8A CN112686060B (en) | 2020-12-29 | 2020-12-29 | Text translation method, device, electronic equipment and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011593186.8A CN112686060B (en) | 2020-12-29 | 2020-12-29 | Text translation method, device, electronic equipment and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112686060A CN112686060A (en) | 2021-04-20 |
| CN112686060B true CN112686060B (en) | 2024-04-30 |
Family
ID=75455034
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011593186.8A Active CN112686060B (en) | 2020-12-29 | 2020-12-29 | Text translation method, device, electronic equipment and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112686060B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2023243946A1 (en) * | 2022-06-14 | 2023-12-21 | Samsung Electronics Co., Ltd. | Machine translation method, devices, and storage media |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR101646926B1 (en) * | 2015-04-16 | 2016-08-09 | 서울대학교산학협력단 | Method and system of deep concept hioerarchy for reconstruction in multi-modality data |
| CN109190134A (en) * | 2018-11-21 | 2019-01-11 | 科大讯飞股份有限公司 | A kind of text interpretation method and device |
| CN109408834A (en) * | 2018-12-17 | 2019-03-01 | 北京百度网讯科技有限公司 | Auxiliary machinery interpretation method, device, equipment and storage medium |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11138392B2 (en) * | 2018-07-26 | 2021-10-05 | Google Llc | Machine translation using neural network models |
-
2020
- 2020-12-29 CN CN202011593186.8A patent/CN112686060B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR101646926B1 (en) * | 2015-04-16 | 2016-08-09 | 서울대학교산학협력단 | Method and system of deep concept hioerarchy for reconstruction in multi-modality data |
| CN109190134A (en) * | 2018-11-21 | 2019-01-11 | 科大讯飞股份有限公司 | A kind of text interpretation method and device |
| CN109408834A (en) * | 2018-12-17 | 2019-03-01 | 北京百度网讯科技有限公司 | Auxiliary machinery interpretation method, device, equipment and storage medium |
Non-Patent Citations (2)
| Title |
|---|
| Unsupervised Neural Machine Translation with Cross-Ligual Language Representation Agreement;Haipeng Sun et al;《IEEE/ACM Transaction on Audio，Speech，and Language Processing》(第28期);第1170-1182页 * |
| 一种面向统计机器翻译的协同权重训练方法;刘树杰 等;《软件学报》;第23卷(第12期);第3101-3114页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112686060A (en) | 2021-04-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11106714B2 (en) | Summary generating apparatus, summary generating method and computer program | |
| KR102449842B1 (en) | Method for training language model and apparatus therefor | |
| WO2019200923A1 (en) | Pinyin-based semantic recognition method and device and human-machine conversation system | |
| CN112711948B (en) | Named entity recognition method and device for Chinese sentences | |
| US11797772B2 (en) | Word lattice augmentation for automatic speech recognition | |
| KR102057184B1 (en) | Interest determination system, interest determination method, and storage medium | |
| JP2020505643A (en) | Voice recognition method, electronic device, and computer storage medium | |
| CN112464993A (en) | Multi-mode model training method, device, equipment and storage medium | |
| JP7111464B2 (en) | Translation method, translation device and translation system | |
| JP2023544336A (en) | System and method for multilingual speech recognition framework | |
| CN112750419A (en) | Voice synthesis method and device, electronic equipment and storage medium | |
| CN112016271A (en) | Language style conversion model training method, text processing method and device | |
| CN113591462A (en) | Bullet screen reply generation method and device and electronic equipment | |
| CN115455946A (en) | Voice recognition error correction method and device, electronic equipment and storage medium | |
| CN112686060B (en) | Text translation method, device, electronic equipment and storage medium | |
| CN112686059B (en) | Text translation method, device, electronic equipment and storage medium | |
| CN114783405B (en) | Speech synthesis method, device, electronic equipment and storage medium | |
| Wang et al. | Emily: Developing An Emotion-affective Open-Domain Chatbot with Knowledge Graph-based Persona | |
| Kheang et al. | Solving the phoneme conflict in grapheme-to-phoneme conversion using a two-stage neural network-based approach | |
| CN115048102A (en) | Code generation method, device, equipment and storage medium | |
| CN114299930A (en) | End-to-end speech recognition model processing method, speech recognition method and related device | |
| CN114519358A (en) | Translation quality evaluation method and device, electronic equipment and storage medium | |
| US20230140480A1 (en) | Utterance generation apparatus, utterance generation method, and program | |
| KR102383043B1 (en) | Learning method and cognition method for omission restoration and apparatus for executing the method | |
| KR102592623B1 (en) | Method for learning real-time simultaneous translation model based on alignment information, method and system for simutaneous translation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| TA01 | Transfer of patent application right | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20230512 Address after: 230026 No. 96, Jinzhai Road, Hefei, Anhui Applicant after: University of Science and Technology of China Applicant after: IFLYTEK Co.,Ltd. Address before: 230088 666 Wangjiang West Road, Hefei hi tech Development Zone, Anhui Applicant before: IFLYTEK Co.,Ltd. |
|
| GR01 | Patent grant | ||
| GR01 | Patent grant |