Data processing method and device, computer equipment and computer readable storage medium

Abstract

The application discloses a data processing method and device, computer equipment and a computer readable storage medium, and applies to the technical field of computers. The method comprises the following steps: acquiring a to-be-processed data sequence; calling a data conversion model to process the to-be-processed data sequence, and obtaining target data corresponding to each to-be-processed data in the to-be-processed data sequence, wherein the training sample corresponding to the data conversion model comprises a source word sequence and a corresponding target word sequence, the data conversion model is obtained by training based on the training sample and an alignment bias of each target word in the target word sequence, and the alignment bias is offset data of each target word relative to a corresponding source word in the source word sequence; and outputting the target data corresponding to each to-be-processed data. The application can be applied to various data conversion scenes such as synchronous machine translation, and the accuracy of data conversion can be improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to data processing methods, data processing apparatus, computer equipment, and computer-readable storage media. Background Technology

[0002] Simultaneous Machine Translation (SiMT) is a subtask of machine translation. It starts by reading an incomplete source and gradually produces a translation. Its aim is to translate a continuously input stream of text into another language with the lowest latency and the highest quality.

[0003] Currently, the samples used to train and evaluate simultaneous translation models come from readily available offline translation corpora. However, these samples suffer from word order misalignment between the source and translation texts. Using these samples for training can impair model performance to some extent, thereby reducing the accuracy of simultaneous machine translation. Summary of the Invention

[0004] This application provides data processing methods, apparatus, computer equipment, and computer-readable storage media, which can improve the accuracy of data conversion in scenarios such as synchronous machine translation.

[0005] On one hand, embodiments of this application provide a data processing method, the method comprising:

[0006] Obtain the sequence of data to be processed;

[0007] The data transformation model is called to process the data sequence to be processed, and the target data corresponding to each data to be processed in the data sequence is obtained. The training samples of the data transformation model include the source word sequence and the corresponding target word sequence. The data transformation model is trained based on the alignment bias of each target word in the training samples and the target word sequence. The alignment bias is the offset data of each target word relative to the corresponding source word in the source word sequence.

[0008] Output the target data corresponding to each piece of data to be processed.

[0009] On one hand, embodiments of this application provide a data processing apparatus, the apparatus comprising:

[0010] The acquisition unit is used to acquire the data sequence to be processed;

[0011] The processing unit is used to call the data transformation model to process the data sequence to be processed, and obtain the target data corresponding to each data to be processed in the data sequence to be processed. The training samples corresponding to the data transformation model include the source word sequence and the corresponding target word sequence. The data transformation model is trained based on the alignment bias of each target word in the training samples and the target word sequence. The alignment bias is the offset data of each target word relative to the corresponding source word in the source word sequence.

[0012] The output unit is used to output the target data corresponding to each piece of data to be processed.

[0013] On one hand, embodiments of this application provide a computer device, which includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the data processing method described above.

[0014] On one hand, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when read and executed by a processor of a computer device, causes the computer device to perform the aforementioned data processing method.

[0015] On one hand, embodiments of this application provide a computer program product, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the aforementioned data processing method.

[0016] In this embodiment, the data sequence to be processed is first obtained; then, a data conversion model is called to process the data sequence to obtain the target data corresponding to each data to be processed in the data sequence. The training samples corresponding to the data conversion model include the source word sequence and the corresponding target word sequence. The data conversion model is trained based on the alignment bias of each target word in the training samples and the target word sequence. Here, the alignment bias refers to the offset data of each target word relative to the corresponding source word in the source word sequence. Finally, the target data corresponding to each data to be processed is output. By using the alignment bias of each target word in the target word sequence, the learning difficulty of each target word is evaluated. Then, incremental training can be performed step by step using the corresponding training samples in a learning manner from easy to difficult to obtain the data conversion model. This can effectively improve the performance of the trained data conversion model and improve the accuracy of data conversion in scenarios such as synchronous machine translation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the structure of a data processing system provided in an embodiment of this application;

[0019] Figure 2 This is a flowchart illustrating a data processing method provided in an embodiment of this application;

[0020] Figure 3 This is a flowchart illustrating another data processing method provided in an embodiment of this application;

[0021] Figure 4 This is a schematic diagram of the structure of a data conversion model provided in an embodiment of this application;

[0022] Figure 5 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application;

[0023] Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0025] It should be noted that the terms "first," "second," etc., used in the embodiments of this application are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature specified with "first" or "second" may explicitly or implicitly include at least one of those features.

[0026] This application relates to Artificial Intelligence (AI) technology. AI refers to the theories, methods, technologies, and application systems that utilize digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce new intelligent machines capable of reacting in a manner similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess perception, reasoning, and decision-making capabilities. Specifically, AI technology covers a wide range of fields, including both hardware and software technologies. At the hardware level, AI technology generally includes technologies such as sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. At the software level, AI technology mainly includes computer vision, speech processing, natural language processing, machine learning / deep learning, autonomous driving, and intelligent transportation.

[0027] Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning / deep learning typically includes techniques such as artificial neural networks, supervised learning, machine translation (MT), and curriculum learning (CL). Among these, curriculum learning techniques have been widely used in Natural Language Processing (NLP) tasks and can be divided into model-based methods and data-based methods. For model-based methods, the main approach is to introduce curriculum learning into non-autoregressive translation (NAT). Compared with autoregressive translation models, NAT is more difficult to learn and has lower translation accuracy. Therefore, a curriculum was designed during the fine-tuning process to gradually transform the training from a simple autoregressive translation model to a more difficult NAT model. For data-based methods, the main approach is to introduce curriculum learning into the question-answering system, gradually learning from simple samples to more difficult samples.

[0028] Machine translation, also known as automatic translation, is the process of using computers to convert one natural language (source language) into another natural language (target language). It is a branch of computational linguistics and one of the ultimate goals of artificial intelligence, possessing significant scientific research value. The development of machine translation technology has been closely intertwined with advancements in computer technology, information theory, and linguistics. Simultaneous Machine Translation (SiMT) is a subtask within machine translation, starting with an incomplete source text and progressively generating a translation. Its aim is to translate continuously input streams of text into another language with minimal latency and high quality.

[0029] Based on the aforementioned machine learning / deep learning techniques and synchronous machine translation techniques, this application proposes a data processing method to improve the accuracy of synchronous machine translation. Specifically, the general principle of this data processing method is as follows: First, a data sequence to be processed is obtained; then, a data transformation model is called to process the data sequence to obtain the target data corresponding to each piece of data in the data sequence. The training samples for this data transformation model include a source word sequence and corresponding target word sequences. The data transformation model is trained based on the alignment bias of each target word in the training samples and the target word sequence. Here, the alignment bias refers to the offset data of each target word relative to the corresponding source word in the source word sequence; finally, the target data corresponding to each piece of data to be processed is output.

[0030] In practical implementation, the data processing methods mentioned above can be executed by computer devices, which can be terminal devices or servers. Terminal devices can be, for example, smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, smart in-vehicle systems, etc., but are not limited to these. Servers can be, for example, independent physical servers, server clusters or distributed systems composed of multiple physical servers, or cloud servers providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0031] Alternatively, the data processing methods mentioned above can be executed jointly by the terminal device and the server. For example, see [link to relevant documentation]. Figure 1As shown: Terminal device 101 can first obtain the data sequence to be processed and send it to server 102. Correspondingly, after receiving the data sequence, server 102 can call a data transformation model to process the data sequence, obtaining the target data corresponding to each piece of data to be processed, and outputting the target data corresponding to each piece of data to be processed. Alternatively, server 102 can also send the target data corresponding to each piece of data to be processed to terminal device 101, so that terminal device 101 can process the target data corresponding to each piece of data to be processed.

[0032] In this embodiment, by utilizing the alignment bias of each target word in the target word sequence, the learning difficulty of each target word is evaluated. Then, incremental training can be performed step by step using the corresponding training samples in a learning manner from easy to difficult to obtain a data conversion model. This can effectively improve the performance of the trained data conversion model and enhance the accuracy of data conversion in scenarios such as synchronous machine translation.

[0033] It is understood that the system architecture diagrams described in the embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided in the embodiments of this application. As those skilled in the art will know, with the evolution of system architecture and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0034] Based on the above explanation, the following will be combined with Figure 2 The flowchart shown further illustrates the data processing method proposed in the embodiments of this application. In this embodiment, the data processing method is mainly described using the aforementioned computer device as an example. Please refer to... Figure 2 The data processing method may specifically include steps S201 to S203:

[0035] S201. Obtain the data sequence to be processed.

[0036] In this embodiment, the data sequence to be processed can be either a text sequence or an audio sequence, without limitation. If the data sequence to be processed is an audio sequence, further speech recognition processing is required to obtain the corresponding text sequence. It should be noted that the data to be processed in the data sequence has a sequential order.

[0037] S202. Call the data transformation model to process the data sequence to be processed, and obtain the target data corresponding to each data to be processed in the data sequence.

[0038] In the embodiments of the present application, a computer device implements processing of a data sequence to be processed by using a data conversion model. The training samples corresponding to the data conversion model include a source word sequence and a corresponding target word sequence. The data conversion model is trained based on the training samples and an Alignment Offset (AO) of each target word in the target word sequence, where the alignment offset is offset data of each target word relative to a corresponding source word in the source word sequence. It should be noted that one training sample includes a source word sequence and a corresponding target word sequence, and the alignment offset can be used to indicate the learning difficulty of each target word. A larger alignment offset of a target word indicates that it is more difficult to learn the target word. The alignment offset may be position offset data of each target word relative to a corresponding source word in the source word sequence, or may be offset data of each target word relative to a corresponding source word in the source word sequence based on a predefined offset, which is not limited herein.

[0039] For example, the source word sequence included in one training sample is "Jia Yi Bing Ding", and the corresponding target word sequence is "ABCD", where the target word "A" corresponds to the source word "Ding", and the target word "B" corresponds to the source word "Bing". The alignment offset of the target word "A" is 3, and the alignment offset of the target word "B" is 1, so it is more difficult to learn the target word "A" than to learn the target word "B".

[0040] S203: Output target data corresponding to each piece of data to be processed.

[0041] In the embodiments of the present application, the computer device may directly display the target data corresponding to each piece of data to be processed on a display screen, or may further convert the target data corresponding to each piece of data to be processed into audio data for output, which is not limited herein.

[0042] In summary, in the embodiments of the present application, a to-be-processed data sequence is obtained first; then a data conversion model is invoked to process the to-be-processed data sequence, to obtain target data corresponding to each piece of to-be-processed data in the to-be-processed data sequence, where training samples corresponding to the data conversion model include a source word sequence and a corresponding target word sequence, the data conversion model is trained based on the training samples and an alignment offset of each target word in the target word sequence, and the alignment offset herein refers to offset data of each target word relative to a corresponding source word in the source word sequence; finally, target data corresponding to each piece of data to be processed is output. By using the alignment offset of each target word in the target word sequence, the learning difficulty of each target word is evaluated, and then corresponding training samples can be gradually used for incremental training according to a learning manner from easy to difficult to obtain the data conversion model, which can effectively improve the performance of the trained data conversion model and improve the accuracy of data conversion in scenarios such as simultaneous machine translation.

[0043] The following is combined Figure 3 The flowchart shown further illustrates the data processing method proposed in the embodiments of this application. In this embodiment, the data processing method is mainly described using the aforementioned computer device as an example. Please refer to... Figure 3 The data processing method may specifically include steps S301 to S307. Wherein:

[0044] S301. Obtain multiple training samples, each of which includes a source word sequence and a corresponding target word sequence.

[0045] In this embodiment, the computer device trains an initial neural network model using multiple training samples. Each training sample includes a source word sequence and a target word sequence corresponding to that source word sequence. As shown in Table 1, a training sample includes the source word sequence: "A is ready to oppose B on steel materials," and the target word sequence corresponding to that source word sequence is: "A ready to oppose B on steel materials."

[0046] Table 1

[0047] English <![CDATA[A ready to oppose B on steel materials]]>

[0048] S302. Determine the alignment bias of each target word in the target word sequence included in each training sample.

[0049] In this embodiment, the computer device uses external tools (such as alignment software) to determine the alignment information of each target word in the target word sequence included in each training sample, and calculates the alignment bias of each target word using the alignment information. In this way, the learning difficulty of each target word is evaluated by using the alignment bias of each target word, which facilitates the incremental training of the initial neural network model in a progressive manner from easy to difficult.

[0050] In one possible implementation, the computer device determines the alignment bias of each target word in the target word sequence included in each training sample by: obtaining the corresponding source words in the source word sequence for each target word in the target word sequence included in each training sample; obtaining the positions of each source word in the source word sequence and the position of each target word in the target word sequence; and determining the alignment bias of each target word based on predefined offsets, the positions of each source word in the source word sequence, and the positions of each target word in the target word sequence. It should be noted that the alignment bias of each target word can be considered as the positional offset data of each target word relative to its corresponding source word in the source word sequence. Therefore, by obtaining the positions of each source word in the source word sequence and the position of each target word in the target word sequence, the alignment bias of each target word can be determined.

[0051] For example, please refer to the foregoing Table 1. Assuming that the target word is "B" and the corresponding source word is "Yi", obtain the position of the source word in the source word sequence and the position of the target word in the target word sequence, that is, the source word is at the 7th position in the source word sequence and the target word is at the 5th position in the target word sequence. According to the positional relationship between the two, the alignment bias of the target word is determined to be 2.

[0052] Optionally, the specific implementation mode in which the computer device determines the alignment bias of each target word based on the predefined offset, the position of each source word in the source word sequence, and the position of each target word in the target word sequence is: determining the alignment bias component of each target word relative to each source word respectively based on the predefined offset, the position of each source word in the source word sequence, and the position of each target word in the target word sequence; obtaining the average alignment bias of the alignment bias components of each target word relative to each source word respectively; and taking the average alignment bias as the alignment bias of each target word.

[0053] It should be noted that the alignment bias component of each target word relative to each source word can be specifically calculated by using formula (1), which is shown as follows:

[0054]

[0055] wherein, (x i ,y j ) is an alignment pair, which consists of a source word and a corresponding target word, representing that the i-th source word x i is aligned to the j-th target word y j , i is the position of the source word in the source word sequence, j is the position of the target word in the target word sequence; k represents a predefined offset, that is, k words are read in before starting to translate the first word, which is beneficial to improving the accuracy of prediction results; PAO represents the alignment bias component of the target word relative to each source word.

[0056] Since each target word may correspond to multiple source words, it is necessary to calculate the average alignment bias of the alignment bias components of the target word relative to each source word, that is, perform summation and average calculation on the alignment bias components of the target word relative to each source word, which can be specifically calculated by using formula (2). Formula (2) is shown as follows:

[0057]

[0058] wherein, (c,y j ) is an alignment pair, which consists of a source word and a corresponding target word, y j represents the j-th target word, j is the target word y jIn the target word sequence, c represents the position of the target word y. j Corresponding source word; C j Indicates target word y j The set of all source words, PAO represents the alignment bias component of the target word relative to each source word, and AO represents the alignment bias of the target word.

[0059] For example, please refer to Table 1 above. Assume that the predefined offset k is 1, the target word is "oppose", and the source word corresponding to the target word is "adopt opposition". That is, the target word corresponds to two source words, namely "adopt" and "oppose". Among them, the source word "adopt" is at the 8th position in the source word sequence, the source word "oppose" is at the 9th position in the source word sequence, and the target word "oppose" is at the 4th position in the target word sequence. First, the alignment offset components of the target word relative to the two source words are calculated using formula (1): the alignment offset component of the target word "oppose" relative to the source word "adopt" is 4, and the alignment offset component of the target word "oppose" relative to the source word "oppose" is 5. Then, the average alignment offset of the target word relative to the two source words is calculated using formula (2), which is 4.5. Finally, this average alignment offset is used as the alignment offset of the target word "oppose", that is, the alignment offset of the target word "oppose" is 4.5.

[0060] For example, please refer to Table 1 above. Assume that the predefined offset k is 2, the target word is "oppose", and the source word corresponding to the target word is "adopt against". That is, the target word corresponds to two source words, namely "adopt" and "oppose". Among them, the source word "adopt" is at the 8th position in the source word sequence, the source word "oppose" is at the 9th position in the source word sequence, and the target word "oppose" is at the 4th position in the target word sequence. First, the alignment offset components of the target word relative to the two source words are calculated using formula (1): the alignment offset component of the target word "oppose" relative to the source word "adopt" is 3, and the alignment offset component of the target word "oppose" relative to the source word "oppose" is 4; then, the average alignment offset of the target word relative to the two source words is calculated using formula (2), which is 3.5; finally, this average alignment offset is used as the alignment offset of the target word "oppose", that is, the alignment offset of the target word "oppose" is 3.5.

[0061] S303. Based on the alignment bias of each target word in each training sample and the target word sequence in each training sample, determine the target word subsequence corresponding to each training sample.

[0062] In this embodiment, the computer device employs a course-based learning approach, incrementally training based on the alignment bias of each target word in each training sample. This involves first learning simple words and then gradually adding more difficult words to the training process, which helps avoid catastrophic forgetting in the model. Therefore, the target word learned for each training step can be a subsequence of target words determined from the target word sequence in each training sample. This subsequence is then continuously updated in subsequent training steps until the target word sequence in the training sample has been learned. The number of training steps can be preset; the number of target words learned for each training step can be preset or dynamically changed, and is not limited here.

[0063] For example, if the target word sequence in a training sample is "ABCD", the target word subsequence taken in the first training step can be "AB", "ABC", etc., depending on the alignment bias of each target word.

[0064] In one possible implementation, the training of the initial neural network model includes multiple course stages. The computer device determines the target word subsequence corresponding to each training sample based on the alignment bias of each target word in each training sample and the target word sequence in each training sample. This includes: obtaining the target training sample corresponding to the current course stage, and obtaining the alignment bias threshold corresponding to the target training step number of the current course stage. The target training sample includes at least one of multiple training samples. Based on the alignment bias threshold and the alignment bias of each target word in the target training sample, at least one target word is determined from the target word sequence in the target training sample. The target word subsequence corresponding to the target training sample is determined based on the at least one target word. It should be noted that the alignment bias threshold corresponding to each target training step number can be preset or dynamically changed, and is not limited here. For the target training sample corresponding to the current course stage, when the alignment bias of the target words in the target training sample is less than or equal to the alignment bias threshold, these target words are selected from the target word sequence to determine the target word subsequence corresponding to the target training sample.

[0065] In one possible implementation, if the alignment bias threshold corresponding to each target training step is preset, then as the target training step increases, the corresponding alignment bias threshold also increases, so that target words with greater learning difficulty can be added to the target word subsequence corresponding to the target training sample.

[0066] For example, for the target training sample in the first course stage, assuming the alignment bias threshold corresponding to the target training step number in the current course stage is 3, and the target word sequence of the target training sample is "abc", where the alignment bias of target word "a" is 1, the alignment bias of target word "b" is 2, and the alignment bias of target word "c" is 4. The alignment biases of target word "a" and target word "b" are less than the alignment bias threshold corresponding to the target training step number in the current course stage. Therefore, the target word subsequence corresponding to the target training sample, namely "ab", is determined based on target word "a" and target word "b".

[0067] Optionally, the specific implementation method for the computer device to obtain the alignment bias threshold corresponding to the target training steps of the current course stage is as follows: obtain the target training steps of the current course stage, the number of the multiple course stages, and the total training steps corresponding to the multiple course stages; determine the alignment bias threshold corresponding to the target training steps based on the target training steps of the current course stage, the number of the multiple course stages, and the total training steps corresponding to the multiple course stages.

[0068] It should be noted that the number of multiple course stages and the total number of training steps corresponding to multiple course stages can be preset. The alignment bias threshold corresponding to the target number of training steps can be dynamically changed according to the target number of training steps in the current course stage. The alignment bias threshold corresponding to the target number of training steps can be calculated by formula (3), as shown below:

[0069]

[0070] Where, N C N represents the number of course stages. S Let represent the total number of training steps across multiple course phases, where i represents the target number of training steps for the current course phase, and f(i) represents the alignment bias threshold corresponding to the i-th training step. This indicates rounding down to the nearest integer.

[0071] For example, suppose the number of course stages is 10, the total number of training steps corresponding to the multiple course stages is 100, and the number of training steps corresponding to each course stage is 10. In the third course stage, the 25th training step is the target training step for the current course stage, and the alignment bias threshold corresponding to the target training step is calculated to be 2 using formula (3).

[0072] S304. Using the source word sequence included in each training sample and the target word subsequence corresponding to each training sample, train the initial neural network model to obtain the data conversion model.

[0073] In this embodiment of the application, the computer device trains the initial neural network model in each training step using the source word sequence included in each training sample and the target word subsequence corresponding to each training sample. Then, in subsequent training steps, the target word subsequence is continuously updated to achieve incremental training of the initial neural network model, thereby obtaining the data conversion model.

[0074] In one possible implementation, the computer device trains an initial neural network model using the source word sequence included in each training sample and the target word subsequence corresponding to each training sample. The specific implementation of the data conversion model is as follows: the encoder of the initial neural network model encodes the source words in the source word sequence included in the target training sample to obtain a first feature representation of the source words in the target training sample; the decoder of the initial neural network model decodes the target word subsequence corresponding to the target training sample and the first feature representation to obtain a second feature representation of the target words in the target training sample; the initial neural network model processes the second feature representation to obtain the prediction result of each source word in the target training sample; based on the prediction result of each source word in the target training sample and the corresponding target word, a loss value is determined, and the model parameters of the initial neural network model are adjusted according to the loss value to obtain the data conversion model.

[0075] Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of a data conversion model provided in an embodiment of this application. For example... Figure 4 As shown, the source words in the source word sequence included in the target training sample are used as the input to the encoder of the initial neural network model to determine the word embeddings of the source words in the target training sample. Then, fixed sine and cosine functions are used to obtain the positional encodings of the source words in the target training sample. A multi-head attention mechanism and a fully connected feedforward network are used to process the word embeddings and positional encodings of the source words in the target training sample. After normalization, the first feature representation of the source words in the target training sample is obtained. The multi-head attention mechanism used is an N-layer multi-head attention mechanism, where each sub-layer's multi-head attention mechanism is a unidirectional encoding, meaning each source word can only see itself and the words preceding it.

[0076] Furthermore, the target word subsequence corresponding to the target training sample is used as input to the decoder of the initial neural network model to determine the word embedding encoding of the target word subsequence, and the position encoding of the target word subsequence is obtained using fixed sine and cosine functions. A masked multi-head attention mechanism is then used to process the word embedding encoding and position encoding of the target word subsequence, and after regularization, the third feature representation of the target word in the target training sample is obtained. Next, a residual connection mechanism is used to add the first and third feature representations. The first and third feature representations are then processed through a multi-head attention mechanism and a fully connected feedforward network, and after regularization, the second feature representation of the target word in the target training sample is obtained. Here, an N-layer multi-head attention mechanism is used, where each sub-layer's multi-head attention mechanism is a unidirectional encoding, meaning each source word can only see itself and the words preceding it. Furthermore, the second feature representation is processed through a linear transformation layer and a normalized exponential function Softmax layer to obtain the prediction results for each source word in the target training sample.

[0077] Finally, based on the prediction results of each source word in the target training sample and the matching degree between the corresponding target words, the loss value is determined. The model parameters of the initial neural network model are then adjusted in the direction of reducing this loss value to obtain the data transformation model. The matching degree can be calculated using cosine similarity or other similarity algorithms; no specific method is specified here.

[0078] S305. Obtain the data sequence to be processed.

[0079] S306. Call the data transformation model to process the data sequence to be processed, and obtain the target data corresponding to each data to be processed in the data sequence.

[0080] S307. Output the target data corresponding to each data to be processed.

[0081] The specific implementation methods of steps S305 to S307 can refer to the specific implementation methods of steps S201 to S203 above, and are not limited here.

[0082] The following specific example illustrates the data processing method proposed in the embodiments of this application:

[0083] Firstly, for the data conversion model, multiple training samples are mainly used to train the initial neural network model. Each training sample includes a source word sequence and a corresponding target word sequence. Please refer to Table 1 above. The source word sequence of a training sample (target training sample) is: "A is ready to oppose B on steel materials". The target word sequence corresponding to this source word sequence is: "A ready to oppose B on steel materials". Based on the predefined offset, the position of each source word in the source word sequence and the position of each target word in the target word sequence, the alignment bias of each target word is determined using formulas (1) and (2). Assuming that the target training step number for the current course stage is 25, the number of multiple course stages is 10, and the total number of training steps corresponding to the multiple course stages is 100, the alignment bias threshold corresponding to the target training step number is determined to be 2 according to formula (3). If the alignment bias of the target word is less than or equal to the alignment bias threshold, these target words are selected from the target word sequence to determine the target word subsequence corresponding to the training sample. The alignment biases of target word "A", target word "ready", target word "to", and target word "B" are all less than or equal to the alignment bias threshold. Therefore, the target word subsequence corresponding to this training sample is "A ready to B". Using the source word sequence and the corresponding target word subsequence of this training sample, the initial neural network model is trained. Through continuous incremental training, a data conversion model is obtained. In applying this data conversion model, the computer first acquires the data sequence to be processed, then directly calls the data conversion model to process the data sequence, obtaining the target data corresponding to each piece of data in the data sequence. Finally, the target data corresponding to each piece of data is output, thus realizing the processing of the data sequence and completing the synchronous machine translation task.

[0084] In summary, the data processing method provided in this application can effectively improve the accuracy of synchronous machine translation, and the trained data conversion model can be used in machine translation systems. As shown in Table 2, this table compares the translation quality and latency of the proposed method with a baseline system. Wait-k and MultipathWait-k models were selected as baseline systems, and experiments and validations were conducted on the dataset. Table 2 lists the Bilingual Evaluation Understudy (BLEU) and Average Lagging (AL) metrics, both standard methods for evaluating synchronous machine translation tasks. A higher BLEU value indicates better translation performance, while a lower AL value indicates lower latency. As shown in Table 2, this method improves the BLEU value by 0.47 compared to Wait-k, and by 0.35 compared to MultipathWait-k; the AL value decreases by 0.02 compared to Wait-k, and by 0.08 compared to MultipathWait-k. This demonstrates that the proposed method not only improves model performance and increases the accuracy of synchronous machine translation, but can also be applied to different synchronous machine translation model structures.

[0085] Table 2

[0086]

[0087]

[0088] In summary, in this embodiment, the training of the initial neural network model first involves acquiring multiple training samples, each including a source word sequence and a corresponding target word sequence. Then, the alignment bias of each target word in the target word sequence of each training sample is determined. Based on the alignment bias and the target word sequence of each training sample, a target word subsequence corresponding to each training sample is determined. Finally, the initial neural network model is trained using the source word sequence and the corresponding target word subsequence of each training sample to obtain a data conversion model. The training of this initial neural network model includes multiple course stages, each using different training samples. In each course stage, the learning difficulty of each target word is evaluated using the alignment bias of each target word in the target word sequence. The initial neural network model is incrementally trained from easy to difficult, resulting in a data conversion model. This approach first teaches simple words and then gradually adds more difficult words to the training process, which helps avoid catastrophic forgetting in the model. Since the initial neural network model is first trained using simple words, it has already been optimized. Then, by further training the optimized initial neural network model using difficult words, the performance of the data conversion model can be effectively improved, thereby increasing the accuracy of data conversion in scenarios such as synchronous machine translation.

[0089] Based on the above data processing method, this application provides a data processing apparatus. Please refer to... Figure 5 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application. The data processing device 500 can operate the following units:

[0090] Acquisition unit 501 is used to acquire the data sequence to be processed;

[0091] The processing unit 502 is used to call the data conversion model to process the data sequence to be processed, and obtain the target data corresponding to each data to be processed in the data sequence to be processed. The training samples corresponding to the data conversion model include the source word sequence and the corresponding target word sequence. The data conversion model is trained based on the alignment bias of each target word in the training samples and the target word sequence. The alignment bias is the offset data of each target word relative to the corresponding source word in the source word sequence.

[0092] Output unit 503 is used to output the target data corresponding to each data to be processed.

[0093] In one embodiment, the device further includes a training unit, which is configured to: acquire a plurality of training samples, each of the plurality of training samples including a source word sequence and a corresponding target word sequence; determine the alignment bias of each target word in the target word sequence included in each training sample; determine a target word subsequence corresponding to each training sample based on the alignment bias of each target word in each training sample and the target word sequence in each training sample; and train an initial neural network model using the source word sequence included in each training sample and the target word subsequence corresponding to each training sample to obtain a data conversion model.

[0094] In another implementation, when determining the alignment bias of each target word in the target word sequence included in each training sample, the training unit may specifically be used to: obtain each source word in the source word sequence corresponding to each target word in the target word sequence included in each training sample; obtain the position of each source word in the source word sequence and the position of each target word in the target word sequence; and determine the alignment bias of each target word based on the predefined offset, the position of each source word in the source word sequence and the position of each target word in the target word sequence.

[0095] In another implementation, when the training unit determines the alignment bias of each target word based on predefined offsets, the positions of each source word in the source word sequence, and the positions of each target word in the target word sequence, it can specifically be used to: determine the alignment bias component of each target word relative to each of the respective source words based on predefined offsets, the positions of each source word in the source word sequence, and the positions of each target word in the target word sequence; obtain the average alignment bias of each target word relative to each of the respective source words; and use the average alignment bias as the alignment bias of each target word.

[0096] In another implementation, the training of the initial neural network model includes multiple course stages. When the training unit determines the target word subsequence corresponding to each training sample based on the alignment bias of each target word in each training sample and the target word sequence in each training sample, it can be specifically used to: obtain the target training sample corresponding to the current course stage, and obtain the alignment bias threshold corresponding to the target training step number of the current course stage, wherein the target training sample includes at least one of the multiple training samples; determine at least one target word from the target word sequence in the target training sample based on the alignment bias threshold and the alignment bias of each target word in the target training sample; and determine the target word subsequence corresponding to the target training sample based on the at least one target word.

[0097] In another implementation, when the training unit obtains the alignment bias threshold corresponding to the target training steps of the current course stage, it may specifically be used to: obtain the target training steps of the current course stage, the number of the multiple course stages, and the total training steps corresponding to the multiple course stages; and determine the alignment bias threshold corresponding to the target training steps based on the target training steps of the current course stage, the number of the multiple course stages, and the total training steps corresponding to the multiple course stages.

[0098] In another implementation, when the training unit trains the initial neural network model using the source word sequence included in each training sample and the target word subsequence corresponding to each training sample to obtain the data conversion model, it can specifically be used to: call the encoder of the initial neural network model to encode the source words in the source word sequence included in the target training sample to obtain the first feature representation of the source words in the target training sample; call the decoder of the initial neural network model to decode the target word subsequence corresponding to the target training sample and the first feature representation to obtain the second feature representation of the target words in the target training sample; call the initial neural network model to process the second feature representation to obtain the prediction result of each source word in the target training sample; determine the loss value based on the prediction result of each source word in the target training sample and the corresponding target word, and adjust the model parameters of the initial neural network model according to the loss value to obtain the data conversion model.

[0099] According to another embodiment of this application, Figure 5 The data processing apparatus shown can be constructed by combining each unit individually or entirely into one or more other units, or one or more of the units can be further divided into multiple functionally smaller units. This achieves the same operation without affecting the technical effects of the embodiments of this application. The above-mentioned units are based on logical function division. In practical applications, the function of one unit can be implemented by multiple units, or the function of multiple units can be implemented by one unit. In other embodiments of this application, the data processing apparatus may also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented collaboratively by multiple units.

[0100] According to another embodiment of this application, the following can be achieved by running on a general-purpose computing device, such as a computer, which includes processing elements and storage elements such as a central processing unit (CPU), random access memory (RAM), and read-only memory (ROM), a device capable of performing operations such as... Figure 2 or Figure 3 The computer program (including program code) for each step involved in the corresponding method shown, to construct such... Figure 5The data processing apparatus shown herein, and the data processing method for implementing the embodiments of this application, are described. The computer program may be recorded on, for example, a computer-readable recording medium, loaded onto the aforementioned computing device via the computer-readable recording medium, and run therein.

[0101] In this embodiment, the data sequence to be processed is first obtained; then, a data conversion model is called to process the data sequence to obtain the target data corresponding to each data to be processed in the data sequence. The training samples corresponding to the data conversion model include the source word sequence and the corresponding target word sequence. The data conversion model is trained based on the alignment bias of each target word in the training samples and the target word sequence. Here, the alignment bias refers to the offset data of each target word relative to the corresponding source word in the source word sequence. Finally, the target data corresponding to each data to be processed is output. By using the alignment bias of each target word in the target word sequence, the learning difficulty of each target word is evaluated. Then, incremental training can be performed step by step using the corresponding training samples in a learning manner from easy to difficult to obtain the data conversion model. This can effectively improve the performance of the trained data conversion model and improve the accuracy of data conversion in scenarios such as synchronous machine translation.

[0102] Based on the description of the above method and apparatus embodiments, this application also provides a computer device. Please refer to... Figure 6 The computer device 600 includes at least a processor 601, a communication interface 602, and a computer storage medium 603. The processor 601, communication interface 602, and computer storage medium 603 can be connected via a bus or other means. The computer storage medium 603 can be stored in the memory 604 of the computer device 600. The computer storage medium 603 is used to store computer programs, which include program instructions. The processor 601 is used to execute the program instructions stored in the computer storage medium 603. The processor 601 (or CPU (Central Processing Unit)) is the computing and control core of the computer device, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve corresponding method flows or corresponding functions.

[0103] In one embodiment, the processor 601 described in this application embodiment can be used to perform a series of data processing operations, specifically including: acquiring a data sequence to be processed; calling a data conversion model to process the data sequence to be processed, obtaining target data corresponding to each data to be processed in the data sequence to be processed, wherein the training samples corresponding to the data conversion model include a source word sequence and a corresponding target word sequence, and the data conversion model is trained based on the alignment bias of each target word in the training samples and the target word sequence, wherein the alignment bias is the offset data of each target word relative to the corresponding source word in the source word sequence; outputting the target data corresponding to each data to be processed, etc.

[0104] This application embodiment also provides a computer storage medium (memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer storage medium provides storage space that stores the operating system of the computer device. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by the processor 601. These instructions can be one or more computer programs (including program code). It should be noted that the computer storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device; optionally, it can also be at least one computer storage medium located remotely from the aforementioned processor.

[0105] In one embodiment, a processor may load and execute one or more instructions stored in a computer storage medium to achieve the aforementioned... Figure 2 or Figure 3 The corresponding steps of the method in the data processing method embodiment shown; in specific implementation, one or more instructions in the computer storage medium are loaded by the processor 601 and executed as follows:

[0106] Obtain the sequence of data to be processed;

[0107] The data transformation model is called to process the data sequence to be processed, and the target data corresponding to each data to be processed in the data sequence is obtained. The training samples of the data transformation model include the source word sequence and the corresponding target word sequence. The data transformation model is trained based on the alignment bias of each target word in the training samples and the target word sequence. The alignment bias is the offset data of each target word relative to the corresponding source word in the source word sequence.

[0108] Output the target data corresponding to each piece of data to be processed.

[0109] In one implementation, the one or more instructions may be loaded by a processor and specifically executed as follows: acquiring multiple training samples, each training sample including a source word sequence and a corresponding target word sequence; determining the alignment bias of each target word in the target word sequence included in each training sample; determining the target word subsequence corresponding to each training sample based on the alignment bias of each target word in each training sample and the target word sequence in each training sample; and training an initial neural network model using the source word sequence included in each training sample and the target word subsequence corresponding to each training sample to obtain a data conversion model.

[0110] In another implementation, when determining the alignment offset of each target word in the target word sequence included in each training sample, the one or more instructions can be loaded and specifically executed by the processor: obtaining each source word in the source word sequence corresponding to each target word in the target word sequence included in each training sample; obtaining the position of each source word in the source word sequence and the position of each target word in the target word sequence; determining the alignment offset of each target word based on the predefined offset, the position of each source word in the source word sequence and the position of each target word in the target word sequence.

[0111] In another implementation, when determining the alignment offset of each target word based on predefined offsets, the positions of each source word in the source word sequence, and the positions of each target word in the target word sequence, the one or more instructions can be loaded and specifically executed by the processor: determining the alignment offset component of each target word relative to each source word based on predefined offsets, the positions of each source word in the source word sequence, and the positions of each target word in the target word sequence; obtaining the average alignment offset of each target word relative to each source word; and using the average alignment offset as the alignment offset of each target word.

[0112] In another implementation, the training of the initial neural network model includes multiple course stages. When determining the target word subsequence corresponding to each training sample based on the alignment bias of each target word in each training sample and the target word sequence in each training sample, one or more instructions can be loaded and specifically executed by the processor: obtaining the target training sample corresponding to the current course stage, and obtaining the alignment bias threshold corresponding to the target training step number of the current course stage, wherein the target training sample includes at least one of the multiple training samples; determining at least one target word from the target word sequence in the target training sample based on the alignment bias threshold and the alignment bias of each target word in the target training sample; and determining the target word subsequence corresponding to the target training sample based on the at least one target word.

[0113] In another implementation, when obtaining the alignment bias threshold corresponding to the target training steps of the current course stage, the one or more instructions can be loaded and specifically executed by the processor: obtaining the target training steps of the current course stage, the number of the multiple course stages, and the total training steps corresponding to the multiple course stages; and determining the alignment bias threshold corresponding to the target training steps based on the target training steps of the current course stage, the number of the multiple course stages, and the total training steps corresponding to the multiple course stages.

[0114] In another implementation, when training an initial neural network model using the source word sequence included in each training sample and the target word subsequence corresponding to each training sample to obtain the data conversion model, one or more instructions can be loaded and executed by the processor: calling the encoder of the initial neural network model to encode the source words in the source word sequence included in the target training sample to obtain the first feature representation of the source words in the target training sample; calling the decoder of the initial neural network model to decode the target word subsequence corresponding to the target training sample and the first feature representation to obtain the second feature representation of the target words in the target training sample; calling the initial neural network model to process the second feature representation to obtain the prediction result of each source word in the target training sample; determining the loss value based on the prediction result of each source word in the target training sample and the corresponding target word, and adjusting the model parameters of the initial neural network model according to the loss value to obtain the data conversion model.

[0115] In this embodiment, the data sequence to be processed is first obtained; then, a data conversion model is called to process the data sequence to obtain the target data corresponding to each data to be processed in the data sequence. The training samples corresponding to the data conversion model include the source word sequence and the corresponding target word sequence. The data conversion model is trained based on the alignment bias of each target word in the training samples and the target word sequence. Here, the alignment bias refers to the offset data of each target word relative to the corresponding source word in the source word sequence. Finally, the target data corresponding to each data to be processed is output. By using the alignment bias of each target word in the target word sequence, the learning difficulty of each target word is evaluated. Then, incremental training can be performed step by step using the corresponding training samples in a learning manner from easy to difficult to obtain the data conversion model. This can effectively improve the performance of the trained data conversion model and improve the accuracy of data conversion in scenarios such as synchronous machine translation.

[0116] It should be noted that, according to one aspect of this application, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the aforementioned... Figure 2 or Figure 3 The data processing method embodiments shown are provided in various optional ways. Furthermore, it should be understood that the above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, equivalent variations made in accordance with the claims of this application are still within the scope of this application.

Claims

1. A data processing method, characterized in that, include: Obtain the sequence of data to be processed; The data transformation model is invoked to process the data sequence to obtain the target data corresponding to each data to be processed in the data sequence; Output the target data corresponding to each of the data to be processed; The methods for obtaining the data transformation model include: Multiple training samples are obtained, and each training sample includes a source word sequence and a corresponding target word sequence; Determine the alignment bias of each target word in the target word sequence included in each training sample, wherein the alignment bias is the offset data of each target word relative to the corresponding source word in the source word sequence; Based on the alignment bias of each target word in each training sample and the target word sequence in each training sample, the target word subsequence corresponding to each training sample is determined; The initial neural network model is trained using the source word sequence included in each training sample and the target word subsequence corresponding to each training sample to obtain the data conversion model.

2. The method according to claim 1, characterized in that, Determining the alignment bias of each target word in the target word sequence included in each training sample includes: Obtain the corresponding source words in the source word sequence for each target word in the target word sequence included in each training sample; Obtain the position of each source word in the source word sequence and the position of each target word in the target word sequence; The alignment offset of each target word is determined based on a predefined offset, the position of each source word in the source word sequence, and the position of each target word in the target word sequence.

3. The method according to claim 2, characterized in that, The process of determining the alignment offset of each target word based on a predefined offset, the position of each source word in the source word sequence, and the position of each target word in the target word sequence includes: Based on the predefined offset, the position of each source word in the source word sequence, and the position of each target word in the target word sequence, the alignment offset component of each target word relative to each source word is determined; Obtain the average alignment offset of each target word relative to the alignment offset components of each source word; The average alignment bias is used as the alignment bias for each target word.

4. The method according to any one of claims 1 to 3, characterized in that, The training of the initial neural network model includes multiple course stages. The determination of the target word subsequence corresponding to each training sample, based on the alignment bias of each target word in each training sample and the target word sequence in each training sample, includes: Obtain the target training sample corresponding to the current course stage, and obtain the alignment bias threshold corresponding to the target training step number of the current course stage, wherein the target training sample includes at least one of the plurality of training samples; Based on the alignment bias threshold and the alignment bias of each target word in the target training sample, at least one target word is determined from the target word sequence in the target training sample; Determine the target word subsequence corresponding to the target training sample based on the at least one target word.

5. The method according to claim 4, characterized in that, The step of obtaining the alignment bias threshold corresponding to the target training steps of the current course stage includes: Obtain the target training steps for the current course stage, the number of the multiple course stages, and the total training steps corresponding to the multiple course stages; Based on the target training steps for the current course stage, the number of multiple course stages, and the total training steps corresponding to the multiple course stages, the alignment bias threshold corresponding to the target training steps is determined.

6. The method according to claim 4, characterized in that, The step of training an initial neural network model using the source word sequence included in each training sample and the target word subsequence corresponding to each training sample to obtain the data conversion model includes: The encoder of the initial neural network model is invoked to encode the source words in the source word sequence included in the target training sample, thereby obtaining the first feature representation of the source words in the target training sample; The decoder of the initial neural network model is invoked to decode the target word subsequence and the first feature representation corresponding to the target training sample, thereby obtaining the second feature representation of the target word in the target training sample; The initial neural network model is invoked to process the second feature representation to obtain the prediction results of each source word in the target training sample; Based on the prediction results of each source word in the target training sample and the corresponding target word, a loss value is determined, and the model parameters of the initial neural network model are adjusted according to the loss value to obtain the data conversion model.

7. A data processing apparatus, characterized in that, The device includes: The acquisition unit is used to acquire the sequence of data to be processed. The processing unit is used to call the data conversion model to process the data sequence to be processed, and obtain the target data corresponding to each data to be processed in the data sequence to be processed. The training samples corresponding to the data conversion model include the source word sequence and the corresponding target word sequence. The data conversion model is trained based on the training samples and the alignment bias of each target word in the target word sequence. The alignment bias is the offset data of each target word relative to the corresponding source word in the source word sequence. The output unit is used to output the target data corresponding to each of the data to be processed; A training unit is configured to acquire multiple training samples, each of which includes a source word sequence and a corresponding target word sequence; determine the alignment bias of each target word in the target word sequence of each training sample, wherein the alignment bias is the offset data of each target word relative to the corresponding source word in the source word sequence; determine the target word subsequence corresponding to each training sample based on the alignment bias of each target word in each training sample and the target word sequence of each training sample; and train an initial neural network model using the source word sequence and the target word subsequence corresponding to each training sample to obtain the data conversion model.

8. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the data processing method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more computer programs, which are adapted to be loaded by a processor and executed by the data processing method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes computer instructions, which, when executed by a processor, implement the data processing method according to any one of claims 1 to 6.

Images