Multi-language neural machine translation model construction method based on hybrid expert network

Abstract

The invention relates to the technical field of neural machine translation model construction, and discloses a hybrid expert network-based multi-language neural machine translation model construction method, which comprises a data terminal, the output end of the data terminal is connected with a server through a signal, and the output end of the server is connected with a database through a signal. The output end of the database is in signal connection with an input layer module, the output end of the input layer module is in signal connection with an expert module, and the output end of the input layer module is in signal connection with a dynamic routing layer module. The problems that challenges confronted by current neural machine translation include sharing of cross-language knowledge, expandability of models and improvement of translation quality, and when multiple languages are processed in the prior art, an independent model needs to be constructed for each pair of languages, so that resource consumption is huge, and high-quality translation is difficult to achieve are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of neural machine translation model construction, and specifically to a method for constructing a multilingual neural machine translation model based on a hybrid expert network. Background Art

[0002] Neural machine translation is an important branch of machine translation. It uses deep learning technology, especially neural network models, to achieve automatic translation from one language to another. Unlike traditional statistical or rule-based machine translation methods, NMT can learn the complex mapping relationship between the source language and the target language, thereby generating more natural, fluent and accurate translation results. Encoder: The encoder encodes the source language sentence into one or a series of vector representations. This process is usually completed using models such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs) or gated recurrent units (GRUs). The self-attention mechanism is also widely used due to its parallel processing capabilities and efficiency. The decoder (Deco der): The decoder gradually generates the translation result of the target language based on the vector representation generated by the encoder. The decoder can also use RNN, LSTM, GRU or Transformer structure. During the generation process, the decoder may use the attention mechanism to focus on different parts of the source language sentence to generate a more accurate translation. The development of neural machine translation has greatly promoted the performance of machine translation, making the quality of automatic translation close to or even exceeding human translation in some scenarios. The training of NMT models usually requires a large amount of bilingual corpus and high computing resources. With the continuous advancement of technology, NMT still has a lot of research space in processing low-resource languages, improving translation speed and reducing computing costs. The current challenges facing neural machine translation include sharing cross-language knowledge, scalability of models, and improvement of translation quality. When dealing with multiple languages, existing technologies often need to build separate models for each pair of languages, which not only consumes a lot of resources, but also makes it difficult to achieve high-quality translation. Summary of the invention

[0003] The purpose of the present invention is to provide a method for constructing a multilingual neural machine translation model based on a hybrid expert network, so as to achieve the purpose of solving the problems raised in the above background technology.

[0004] To achieve the above-mentioned purpose, the present invention provides the following technical solution: a method for constructing a multilingual neural machine translation model based on a hybrid expert network, comprising a data terminal, the output end of the data terminal is connected to a server via a signal, the output end of the server is connected to a database via a signal, the output end of the database is connected to an input layer module via a signal, and the output end of the input layer module is connected to an expert module via a signal.

[0005] Preferably, the output end of the input layer module is connected to the dynamic routing layer module via a signal, the output end of the dynamic routing layer module is connected to the output layer module via a signal, and the output end of the output layer module is connected to the audio processing module via a signal.

[0006] Preferably, the output end of the output layer module is connected to the data processing module through a signal, the output end of the data processing module is connected to the model architecture module through a signal, the output end of the model architecture module is connected to the training module through a signal, and the output end of the training module is connected to the input end signal of the input layer module.

[0007] Preferably, the output end of the audio processing module is connected to multilingual audio through signals, the output end of the multilingual audio is connected to an ASR system module through signals, the output end of the ASR system module is connected to multilingual text through signals, the output end of the multilingual text is connected to a translation system through signals, and the output end of the translation system is connected to a target text through signals.

[0008] Preferably, the output end of the multilingual audio is connected to a TTS system module via a signal, and the output end of the TTS system module is connected to the input end of the multilingual text via a signal.

[0009] Preferably, the output end of the training module is connected to the pre-training module via a signal, and the output end of the pre-training module is connected to the distillation module via a signal.

[0010] Preferably, the output end of the distillation module is connected to a target model construction module via a signal, and the output end of the target model construction module is connected to a model training module via a signal.

[0011] Preferably, S1: Perform data cleaning on the source language corpus and the target language corpus through a preset cleaning scheme, and use the cleaned source language corpus and the target language corpus as parallel corpora; S2: According to the reverse multilingual neural machine translation model, a pseudo corpus corresponding to the target language corpus is obtained, and the pseudo corpus and the target language corpus are used as pseudo parallel corpus; S3: Use the preset cleaning scheme to clean the initial training samples; S4: Train the initial multilingual neural machine translation model through the target training samples to construct a trained target multilingual neural machine translation model.

[0012] The present invention provides a method for constructing a multilingual neural machine translation model based on a hybrid expert network. It has the following beneficial effects: (1) The present invention transmits the data to be translated into the input layer module, analyzes the data through the dynamic routing layer module, and then transports the data to the data processing module through the output layer module. At this time, the model is constructed through the model architecture module, which solves the challenges currently faced by neural machine translation, including cross-language knowledge sharing, model scalability, and translation quality improvement. When processing multiple languages, the existing technology often needs to build a separate model for each pair of languages, which not only consumes a huge amount of resources, but also makes it difficult to achieve high-quality translation.

[0013] (2) The present invention trains the model by using training data to obtain a multilingual speech translation model. First, multilingual audio features are extracted to extract useful information. Then, a first encoding pre-training is performed on a transcription system from multilingual audio to multilingual text constructed based on the audio features, so that the model has the ability to convert speech to text. A second decoding pre-training is performed on a text system from the constructed multilingual text to the target text, so that the model has a sense of language and a multilingual speech translation model is constructed.

[0014] (3) The present invention can reduce the noise in the target training samples and improve the quality of the target training samples by means of data cleaning. The noise in the initial training samples can be removed by data cleaning, and low-quality training samples can be filtered out to improve the quality of the target training samples. On the other hand, sufficient target training samples can be provided by means of data enhancement. The present application combines data cleaning and data enhancement to improve the translation effect of the target multilingual neural machine translation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a system block diagram of the present invention; Figure 2 is a system block diagram of the audio processing module of the present invention; Figure 3 It is a system module of the training module of the present invention.

[0016] In the figure: 1. Data terminal; 2. Server; 3. Database; 4. Audio processing module; 411. Multilingual audio; 412. TTS system module; 413. ASR system module; 414. Multilingual text; 415. Translation system; 416. Target text; 5. Training module; 502. Pre-training module; 503. Distillation module; 504. Target model building module; 505. Model training module; 6. Output layer module; 7. Data processing module; 8. Model architecture module; 9. Dynamic routing layer module; 10. Expert module. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0018] Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0019] In the description of the present invention, it is to be understood that the terms “center”, “longitudinal”, “lateral”, “length”, “width”, “thickness”, “up”, “down”, “front”, “back”, “left”, “right”, “vertical”, “horizontal”, “top”, “bottom”, “inside”, “outside”, “clockwise”, “counterclockwise”, “axial”, “radial”, “circumferential”, etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0020] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0021] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described with reference to the accompanying drawings.

[0022] Example 1 A preferred embodiment of the method for constructing a multilingual neural machine translation model based on a hybrid expert network provided by the present invention is as follows: Figures 1 to 3 As shown: A method for constructing a multilingual neural machine translation model based on a hybrid expert network includes a data terminal 1, an output end of the data terminal 1 is connected to a server 2 through a signal, an output end of the server 2 is connected to a database 3 through a signal, an output end of the database 3 is connected to an input layer module 11 through a signal, and an output end of the input layer module 11 is connected to an expert module 10 through a signal.

[0023] The output end of the input layer module 11 is connected to the dynamic routing layer module 9 via a signal, the output end of the dynamic routing layer module 9 is connected to the output layer module 6 via a signal, and the output end of the output layer module 6 is connected to the audio processing module 4 via a signal.

[0024] The output end of the output layer module 6 is connected to the data processing module 7 through a signal, the output end of the data processing module 7 is connected to the model architecture module 8 through a signal, the output end of the model architecture module 8 is connected to the training module 5 through a signal, and the output end of the training module 5 is connected to the input end signal of the input layer module 11.

[0025] In the specific implementation process, the data to be translated is transmitted to it through the input layer module 11, and then the data is analyzed through the dynamic routing layer module 9. Later, the data is transported to the data processing module 7 through the output layer module 6. At this time, the model is constructed through the model architecture module 8.

[0026] Example 2 On the basis of Example 1, a preferred embodiment of the method for constructing a multilingual neural machine translation model based on a hybrid expert network provided by the present invention is as follows: Figures 1 to 3 As shown: the output end of the audio processing module 4 is connected to the multilingual audio 411 through a signal, the output end of the multilingual audio 411 is connected to the ASR system module 413 through a signal, the output end of the ASR system module 413 is connected to the multilingual text 414 through a signal, the output end of the multilingual text 414 is connected to the translation system 415 through a signal, and the output end of the translation system 415 is connected to the target text 416 through a signal.

[0027] The output end of the multilingual audio 411 is connected to the TTS system module 412 via a signal, and the output end of the TTS system module 412 is connected to the input end of the multilingual text 414 via a signal.

[0028] The output end of the training module 5 is connected to the pre-training module 502 via a signal, and the output end of the pre-training module 502 is connected to the distillation module 503 via a signal.

[0029] The output end of the distillation module 503 is connected to the target model construction module 504 through a signal, and the output end of the target model construction module 504 is connected to the model training module 505 through a signal.

[0030] By using training data to train the model, a multilingual speech translation model is obtained. First, multilingual audio features are extracted to extract useful information. Then, a first encoding pre-training is performed on a transcription system from multilingual audio to multilingual text constructed based on audio features, so that the model has the ability to convert speech to text. By performing a second decoding pre-training on a text system from the constructed multilingual text to the target text, the model has a sense of language and a multilingual speech translation model is constructed.

[0031] S1: Perform data cleaning on the source language corpus and the target language corpus through a preset cleaning scheme, and use the cleaned source language corpus and the target language corpus as parallel corpora; S2: According to the reverse multilingual neural machine translation model, a pseudo corpus corresponding to the target language corpus is obtained, and the pseudo corpus and the target language corpus are used as pseudo parallel corpus; S3: Use the preset cleaning scheme to clean the initial training samples; S4: Train the initial multilingual neural machine translation model through the target training samples to construct a trained target multilingual neural machine translation model.

[0032] Data cleaning can be used to remove noise from the initial training samples, filter out low-quality training samples, and improve the quality of the target training samples. On the other hand, sufficient target training samples can be provided through data enhancement. This application combines data cleaning and data enhancement to improve the translation effect of the target multilingual neural machine translation model.

[0033] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

[0034] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for constructing a multilingual neural machine translation model based on a hybrid expert network, comprising a data terminal (1); characterized in that: The output end of the data terminal (1) is connected to a server (2) via a signal, the output end of the server (2) is connected to a database (3) via a signal, the output end of the database (3) is connected to an input layer module (11) via a signal, and the output end of the input layer module (11) is connected to an expert module (10) via a signal.

2. The method for constructing a multilingual neural machine translation model based on a hybrid expert network according to claim 1, characterized in that: The output end of the input layer module (11) is connected to the dynamic routing layer module (9) via a signal, the output end of the dynamic routing layer module (9) is connected to the output layer module (6) via a signal, and the output end of the output layer module (6) is connected to the audio processing module (4) via a signal.

3. The method for constructing a multilingual neural machine translation model based on a hybrid expert network according to claim 2, characterized in that: The output end of the output layer module (6) is connected to the data processing module (7) via a signal, the output end of the data processing module (7) is connected to the model architecture module (8) via a signal, the output end of the model architecture module (8) is connected to the training module (5) via a signal, and the output end of the training module (5) is connected to the input end of the input layer module (11) via a signal.

4. The method for constructing a multilingual neural machine translation model based on a hybrid expert network according to claim 3, characterized in that: The output end of the audio processing module (4) is connected to multilingual audio (411) via signals, the output end of the multilingual audio (411) is connected to an ASR system module (413) via signals, the output end of the ASR system module (413) is connected to multilingual text (414) via signals, the output end of the multilingual text (414) is connected to a translation system (415) via signals, and the output end of the translation system (415) is connected to a target text (416) via signals.

5. The method for constructing a multilingual neural machine translation model based on a hybrid expert network according to claim 4, characterized in that: The output end of the multilingual audio (411) is connected to a TTS system module (412) via a signal, and the output end of the TTS system module (412) is connected to an input end of the multilingual text (414) via a signal.

6. The method for constructing a multilingual neural machine translation model based on a hybrid expert network according to claim 5, characterized in that: The output end of the training module (5) is connected to the pre-training module (502) via a signal, and the output end of the pre-training module (502) is connected to the distillation module (503) via a signal.

7. The method for constructing a multilingual neural machine translation model based on a hybrid expert network according to claim 6, characterized in that: The output end of the distillation module (503) is connected to the target model construction module (504) through a signal, and the output end of the target model construction module (504) is connected to the model training module (505) through a signal.

8. The method for constructing a multilingual neural machine translation model based on a hybrid expert network according to claim 7, characterized in that: S1: Perform data cleaning on the source language corpus and the target language corpus through a preset cleaning scheme, and use the cleaned source language corpus and the target language corpus as parallel corpora; S2: According to the reverse multilingual neural machine translation model, a pseudo corpus corresponding to the target language corpus is obtained, and the pseudo corpus and the target language corpus are used as pseudo parallel corpus; S3: Use the preset cleaning scheme to clean the initial training samples; S4: Train the initial multilingual neural machine translation model through the target training samples to construct a trained target multilingual neural machine translation model.

Images