Medical machine translation self-learning method, device and electronic equipment

Abstract

The present application relates to the field of machine translation, and provides a self-learning method, device and electronic device for medical machine translation, which first constructs a medical machine translation model, and then collects an initial parallel corpus consisting of a first language text and its corresponding second language text. Then, using these corpora, the medical machine translation model translates the first language text into a translated text in the second language. Subsequently, the translated text is compared with the original second language text to calculate the translation quality score. Finally, based on the translation quality score, it is decided whether to use these initial parallel corpora for training to perform self-learning optimization on the medical machine translation model. In this way, it is conducive to optimizing the medical machine translation model, thereby improving the accuracy of medical translation.

Description

Technical Field

[0001] The present application relates to the field of machine translation, and more specifically, to a medical machine translation self-learning method, device and electronic device. Background Art

[0002] At present, the global pharmaceutical and medical service markets are more closely connected than ever before. The improvement of medical consumption and clinical trial standards has increased the demand for medical translation. Medical translation must not only ensure the accuracy and fluency of the translation, but also ensure the accuracy of medical terminology and reading habits suitable for clinical research to accurately convey the original information. Any minor error may lead to serious consequences. However, manual medical translation relies on the collaboration of experts in multiple fields, but due to the large differences in specialties, it is difficult to unify standards, which can easily lead to inaccurate translation.

[0003] In this regard, patent CN116629277A proposes a medical machine translation method based on reinforcement learning, which improves the translation accuracy and efficiency of traditional manual medical professional content by building a parallel corpus, establishing and pre-training a machine translation model, calculating a translation quality score, building a reinforcement learning training set, and fine-tuning the machine translation model.

[0004] In the above patent, the first language text is translated into the second language through the Transformer architecture model. Specifically, the Transformer architecture uses a multi-head attention mechanism and a convolutional neural network to capture the contextual information of text sentences. Although the self-attention mechanism and convolutional model of the Transformer model can capture local contextual information, the ability to capture long-distance global dependencies is limited. Especially when dealing with medical texts that contain a large number of professional terms and complex sentence structures, it may still be impossible to fully capture all important and critical contextual information due to the limitations of the attention mechanism and the fixed convolution kernel in the convolutional network. This may lead to the omission or misinterpretation of key information in the translation results, which in turn affects clinical decisions.

[0005] Therefore, an optimized self-learning scheme for medical machine translation is desired. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present application provides a medical machine translation self-learning method, device and electronic device.

[0007] According to one aspect of the present application, a medical machine translation self-learning method is provided, which includes: step 1: establishing a medical machine translation model; step 2: obtaining a first initial parallel corpus, wherein the first initial parallel corpus consists of a first language, a first language text, a second language, and a second language text; step 3: based on the first language and the second language of the first initial parallel corpus, using the medical machine translation model to translate the first language text to generate a second language translation text; step 4: calculating a translation quality score between the second language translation text and the second language text; step 5: based on the translation quality score, determining whether to include the first initial parallel corpus as a training corpus to perform self-learning optimization on the medical machine translation model to adjust the The model parameters of the medical machine translation model; wherein, the step 3 includes: respectively embedding the first language and the second language to obtain the first language embedding features and the second language embedding features; performing word granularity semantic analysis on the first language text to obtain a sequence of word granularity semantic features of the first language text; performing field effect sequence semantic feature enhancement on the sequence of word granularity semantic features of the first language text to obtain the global word granularity semantic enhancement features of the first language text; performing semantic response fusion on the global word granularity semantic enhancement features of the first language text, the first language embedding features and the second language embedding features to obtain the second language translation text semantic representation, and obtaining the second language translation text based on the second language translation text semantic representation.

[0008] In the above-mentioned medical machine translation self-learning method, the first language and the second language are respectively embedded encoded to obtain first language embedding features and second language embedding features, including: embedding encoding the first language through a first language embedding matrix to obtain a first language embedding coding vector as the first language embedding feature, and embedding encoding the second language through a second language embedding matrix to obtain a second language embedding coding vector as the second language embedding feature.

[0009] In the above-mentioned medical machine translation self-learning method, word-granularity semantic analysis is performed on the first language text to obtain a sequence of word-granularity semantic features of the first language text, including: performing word segmentation processing on the first language text and then passing it through a first language text semantic encoder based on the BERT model to obtain a sequence of first language text word-granularity semantic feature vectors as a sequence of word-granularity semantic features of the first language text.

[0010] In the above-mentioned medical machine translation self-learning method, the sequence of word-granularity semantic features of the first language text is subjected to field-effect sequence semantic feature enhancement to obtain global word-granularity semantic enhancement features of the first language text, including: using a semantic reconstructor to perform feature reconstruction on each first language text word-granularity semantic feature vector in the sequence of word-granularity semantic feature vectors of the first language text to obtain a sequence of reconstructed first language text word-granularity semantic feature vectors; based on the sequence of reconstructed first language text word-granularity semantic feature vectors, performing deep implicit semantic global enhancement on the sequence of word-granularity semantic features of the first language text to obtain a first language text global word-granularity semantic enhancement feature vector as the first language text global word-granularity semantic enhancement feature.

[0011] In the above-mentioned medical machine translation self-learning method, based on the sequence of reconstructed first language text word granularity semantic feature vectors, the sequence of the first language text word granularity semantic features is deeply implicitly semantically globally enhanced to obtain the first language text global word granularity semantic enhanced feature vector, including: calculating the semantic deep implicit representation value of each reconstructed first language text word granularity semantic feature vector in the sequence of reconstructed first language text word granularity semantic feature vectors to obtain a sequence of first language text word granularity semantic deep implicit representation values; based on the sorting of the sequence of the first language text word granularity semantic deep implicit representation values, the sequence of the first language text word granularity semantic feature vectors is reordered to obtain an updated sequence of the first language text word granularity semantic feature vectors; and the updated sequence of the first language text word granularity semantic feature vectors is respectively subjected to bidirectional global semantic encoding fusion to obtain the first language text global word granularity semantic enhanced feature vector.

[0012] In the above-mentioned medical machine translation self-learning method, the update sequence of the word granularity semantic feature vector of the first language text is subjected to bidirectional global semantic encoding fusion to obtain the global word granularity semantic enhanced feature vector of the first language text, including: inputting the update sequence of the word granularity semantic feature vector of the first language text into a first semantic encoder based on a forward LSTM model to obtain a forward global first language text word granularity semantic encoding feature vector; inputting the update sequence of the word granularity semantic feature vector of the first language text into a first semantic encoder based on a backward LSTM model to obtain a backward global first language text word granularity semantic encoding feature vector; fusing the forward global first language text word granularity semantic encoding feature vector and the backward global first language text word granularity semantic encoding feature vector to obtain the global word granularity semantic enhanced feature vector of the first language text.

[0013] In the above-mentioned medical machine translation self-learning method, the global word granularity semantic enhancement feature of the first language text, the first language embedding feature and the second language embedding feature are semantically fused to obtain a semantic representation of a second language translation text, and based on the semantic representation of the second language translation text, the second language translation text is obtained, including: passing the global word granularity semantic enhancement feature vector of the first language text, the first language embedding coding vector and the second language embedding coding vector through a semantic translation response fuser to obtain a second language translation text semantic representation vector as the second language translation text semantic representation; passing the second language translation text semantic representation vector through a translation text generator to obtain the second language translation text.

[0014] In the above-mentioned medical machine translation self-learning method, the translation text generator is an AIGC-based translation text generator.

[0015] According to another aspect of the present application, a medical machine translation self-learning device is provided, which is used to execute the medical machine translation self-learning method as described above, wherein the medical machine translation self-learning device includes: a language embedding encoding module, which is used to embed the first language and the second language respectively to obtain first language embedding features and second language embedding features; a language text word granularity semantic analysis module, which is used to perform word granularity semantic analysis on the first language text to obtain a sequence of word granularity semantic features of the first language text; a semantic feature enhancement module, which is used to perform field effect sequence semantic feature enhancement on the sequence of word granularity semantic features of the first language text to obtain global word granularity semantic enhancement features of the first language text; a second language translation text generation module, which is used to perform semantic response fusion on the global word granularity semantic enhancement features of the first language text, the first language embedding features and the second language embedding features to obtain a second language translation text semantic representation, and obtain the second language translation text based on the second language translation text semantic representation.

[0016] According to another aspect of the present application, an electronic device is provided, comprising a memory and a processor, wherein the processor is configured to execute the medical machine translation self-learning method as described in any one of the above items based on instructions stored in the memory.

[0017] The present application has significant technical effects due to the adoption of the above technical solutions: the medical machine translation self-learning method, device and electronic device provided by the present application adopt language processing and translation technology based on artificial intelligence and machine translation models to embed the first language and the second language, segment and semantically encode the first language text, and then perform field-effect global semantic enhancement on the word granularity semantic features of each first language text, so as to automatically generate the second language translation text according to the response fusion representation between the global word granularity semantic enhancement features of the first language text, the first language embedding features and the second language embedding features. In this way, contextual information can be captured in a larger range, especially when dealing with long sentences and complex structures, the logical relationship and long-distance dependency between sentences can be better understood, and through field-effect global semantic enhancement, the omission of key information can be reduced, ensuring the integrity and accuracy of the translation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0019] Figure 1 Flow chart of the medical machine translation self-learning method according to an embodiment of the present application.

[0020] Figure 2 This is a flowchart of step 3 in the medical machine translation self-learning method according to an embodiment of the present application.

[0021] Figure 3 Flow chart of step 33 in the medical machine translation self-learning method according to an embodiment of the present application.

[0022] Figure 4 Flow chart of step 34 in the medical machine translation self-learning method according to an embodiment of the present application.

[0023] Figure 5 It is a block diagram of a medical machine translation self-learning device according to an embodiment of the present application.

[0024] Figure 6 is a block diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0026] Based on the technical problems raised in the above background technology, this application proposes a medical machine translation self-learning method. Figure 1 Flowchart of the self-learning method for medical machine translation according to an embodiment of the present application. Figure 1 As shown, according to the embodiment of the present application, the medical machine translation self-learning method includes: step 1, establishing a medical machine translation model; step 2, obtaining a first initial parallel corpus, wherein the first initial parallel corpus consists of a first language, a first language text, a second language and a second language text; step 3, based on the first language and the second language of the first initial parallel corpus, using the medical machine translation model to translate the first language text to generate a second language translation text; step 4, calculating the translation quality score between the second language translation text and the second language text; step 5, based on the translation quality score, determining whether to include the first initial parallel corpus as a training corpus to perform self-learning optimization on the medical machine translation model to adjust the model parameters of the medical machine translation model.

[0027] In step 1, a medical machine translation model is established. It should be understood that the medical machine translation model established in this application is a translation model that combines the BERT model, the LSTM model, the AIGC model and other models. The BERT model can better process the contextual information of the input data through bidirectional training, which helps to understand the terms and concepts that are highly dependent on the context in medical translation; the LSTM model has a memory unit that can remember long-term dependency information, which is conducive to processing complex structures and long-distance dependencies in medical texts; and the AIGC model refers to an artificial intelligence model that can generate content. It performs well in generating fluent and natural language, which can improve the fluency and naturalness of translation. Different model structures have their own advantages, and they can complement each other by being used in combination as a medical machine translation model, thereby improving the accuracy, fluency and robustness of medical machine translation.

[0028] Establishing a medical machine translation model is a complex and multi-step process involving model selection, data preparation, model training and optimization. The following is a detailed description of this process: The first step in establishing a medical machine translation model is to select a suitable model architecture. Medical texts are highly professional and complex, so it is necessary to select a model that can effectively handle long dependencies and contextual information. In this application, a combination of BERT model, LSTM model and AIGC model is selected as the medical machine translation model.

[0029] High-quality parallel corpus data is the basis of model training, so a large amount of parallel corpus needs to be collected and prepared. These corpora should cover a wide range of medical topics, including but not limited to diagnostic reports, treatment plans, clinical trial results, medical papers, etc. Specific data preparation includes the following steps: a. Collect parallel corpus: Collect parallel corpus through various channels, such as from medical journals, clinical reports, academic papers, medical websites and other sources. In addition, high-quality parallel corpus can also be obtained through channels such as international cooperation projects, medical databases and professional translation agencies. b. Preprocessing: The collected parallel corpus needs to be preprocessed to ensure its quality and consistency. Specific operations include but are not limited to cleaning, segmentation, annotation and alignment. The preprocessed parallel corpus can provide reliable data for model training.

[0030] Model training is the core step in building a medical machine translation model. At this stage, the model will learn the mapping relationship between the source language text and the target language text, and then achieve translation. The specific process of model training includes: a. Initialize the model parameters according to the selected model architecture, which includes setting hyperparameters such as learning rate, batch size, and number of iterations. b. Load the prepared parallel corpus data into the training environment and divide it into training set, validation set, and test set. The training set is used for model training, the validation set is used to adjust the model parameters, and the test set is used to evaluate the model performance. c. Use the training set data to train the model. During the training process, the model will continuously adjust the parameters to minimize the loss function and improve the translation quality. Optimization strategies such as gradient clipping and learning rate adjustment may also be required during the training process to avoid problems such as overfitting and gradient disappearance. d. During the training process, the validation set data is regularly used to evaluate the model performance.

[0031] After the model training is completed, the model needs to be fully evaluated to ensure its translation quality and performance. For example, use an independent test set to fully evaluate the model. The test set is data that has never been involved in training and verification, and is used to evaluate the performance of the model on new data. Through the test set evaluation, various evaluation indicators such as BLEU score, ROUGE score, accuracy, etc. can be calculated to comprehensively evaluate the translation quality of the model. Analyze the model's errors on the test set to find out common problems, such as improper vocabulary selection, grammatical errors, semantic understanding deviations, etc. Error analysis can help understand the weaknesses of the model and provide direction for further optimization. For example, if it is found that the model performs poorly when processing certain specific medical terms, the training data of these terms can be increased to improve the model's recognition ability. Based on the results of the error analysis, the model is further optimized.

[0032] After model training and optimization are completed, the model needs to be deployed in actual application environments, such as hospital information systems, medical research platforms, etc. During the deployment process, factors such as the model's operating environment, resource requirements, and security need to be considered. For example, the model can be deployed on a cloud platform to provide efficient services using cloud computing resources. In addition, it is necessary to design reasonable interfaces and user interfaces to facilitate users to use the model for translation.

[0033] In summary, by selecting a suitable model architecture, preparing high-quality parallel corpus data, conducting effective model training and evaluation, and continuously optimizing model performance, an efficient and accurate medical machine translation model can be established.

[0034] In step 2, the first initial parallel corpus is obtained, and the first initial parallel corpus consists of a first language, a first language text, a second language, and a second language text. It should be understood that the first language is the source language, that is, the language to be translated. It can be any natural language, such as English, Chinese, etc., depending on the actual application scenario. The first language text refers to the original text content written in the first language, which is the object to be translated. The second language refers to the target language, that is, the translated language. Like the first language, it can also be any natural language, but is usually different from the first language in order to achieve a cross-language translation function. The second language text refers to a translation text written in the second language corresponding to the first language text. It is the product of the translation process and is used to achieve language conversion and information transmission. The first initial parallel corpus refers to a corresponding text set between two languages, in which each pair of texts is the same in content, but is written in different languages. It provides a translation sample from the first language to the second language, and a qualified initial parallel corpus can be used as training data for a medical machine translation model.

[0035] In step 3, based on the first language and the second language of the first initial parallel corpus, the medical machine translation model is used to translate the first language text to generate a second language translation text.

[0036] Accordingly, in the first language and the second language based on the first initial parallel corpus, the medical machine translation model is used to translate the first language text to generate the second language translation text, the technical concept of the present application is to use language processing and translation technology based on artificial intelligence and machine translation models to embed the first language and the second language, segment and semantically encode the first language text, and then perform field-effect global semantic enhancement on the word granularity semantic features of each first language text, so as to automatically generate the second language translation text according to the response fusion representation between the global word granularity semantic enhancement features of the first language text, the first language embedding features and the second language embedding features. In this way, contextual information can be captured in a larger range, especially when dealing with long sentences and complex structures, the logical relationship and long-distance dependency between sentences can be better understood, and through field-effect global semantic enhancement, the omission of key information can be reduced to ensure the completeness and accuracy of the translation results.

[0037] Specifically, Figure 2 FIG. 3 is a flowchart of step 3 in the medical machine translation self-learning method according to an embodiment of the present application. Figure 2 As shown, the step 3 includes: step 31, respectively embedding the first language and the second language to obtain the first language embedding feature and the second language embedding feature; step 32, word granularity semantic analysis of the first language text to obtain a sequence of word granularity semantic features of the first language text; step 33, field effect sequence semantic feature enhancement of the sequence of word granularity semantic features of the first language text to obtain the global word granularity semantic enhancement feature of the first language text; step 34, semantic response fusion of the global word granularity semantic enhancement feature of the first language text, the first language embedding feature and the second language embedding feature to obtain the semantic representation of the second language translation text, and obtain the second language translation text based on the semantic representation of the second language translation text.

[0038] In step 31, embedding coding is performed on the first language and the second language respectively to obtain first language embedding features and second language embedding features. Specifically, in an embodiment of the present application, the step 31 includes: embedding coding the first language through the first language embedding matrix to obtain the first language embedding coding vector as the first language embedding feature, and embedding coding the second language through the second language embedding matrix to obtain the second language embedding coding vector as the second language embedding feature. It should be understood that the first language and the second language contain semantic information about different languages, such as vocabulary and grammatical structure. Based on this, in order to effectively represent the semantic information of words to help the model better understand and generate translated texts, in the technical solution of the present application, the first language is embedded coded through the first language embedding matrix to obtain the first language embedding coding vector, and the second language is embedded coded through the second language embedding matrix to obtain the second language embedding coding vector. In this way, the semantic information of words in the language can be captured, and the role of words in the context can be reflected, providing more reliable support for subsequent text generation.

[0039] In step 32, the first language text is subjected to word-granular semantic analysis to obtain a sequence of word-granular semantic features of the first language text. Specifically, in an embodiment of the present application, the step 32 includes: performing word segmentation processing on the first language text and passing it through a first language text semantic encoder based on a BERT model to obtain a sequence of word-granular semantic feature vectors of the first language text as a sequence of word-granular semantic features of the first language text. It should be understood that the first language text contains a plurality of vocabulary information, such as anterior chamber puncture, ALPI and argon laser pupilloplasty, corneal indentation, etc. in Chinese vocabulary, and each word in these words or phrases contains a semantic association relationship between contexts. Therefore, in order to more finely capture and represent the semantic information in the text, in the technical solution of the present application, the first language text is subjected to word segmentation processing and passed through a first language text semantic encoder based on a BERT model to obtain a sequence of word-granular semantic feature vectors of the first language text as a sequence of word-granular semantic features of the first language text. Those skilled in the art should be aware that the BERT model is unsupervised pre-trained on a large-scale text corpus, which gives it a great advantage in understanding the semantics of the input first-language text, and by performing bidirectional processing on the input first-language text, it can learn rich contextual representations, thereby more accurately capturing the semantic information of the first-language text.

[0040] In step 33, field effect sequence semantic feature enhancement is performed on the sequence of word granularity semantic features of the first language text to obtain global word granularity semantic enhancement features of the first language text. Figure 3FIG. 3 is a flowchart of step 33 in the medical machine translation self-learning method according to an embodiment of the present application. Figure 3 As shown, the step 33 includes: step 331, using a semantic reconstructor to perform feature reconstruction on each first language text word granularity semantic feature vector in the sequence of the first language text word granularity semantic feature vectors to obtain a sequence of reconstructed first language text word granularity semantic feature vectors; step 332, based on the sequence of reconstructed first language text word granularity semantic feature vectors, performing deep implicit semantic global enhancement on the sequence of the first language text word granularity semantic features to obtain a first language text global word granularity semantic enhancement feature vector as the first language text global word granularity semantic enhancement feature.

[0041] It should be understandable that the semantic sparsity and feature imbalance of the word-granularity semantic features of the first language text will lead to understanding bias in the contextual semantic encoding process. Semantic sparsity means that the word embedding vector may not be able to fully capture the rich semantic information of the word, while ignoring the global relationship between words, while feature imbalance may cause the semantic features of certain words to be over- or under-represented in the vector space. This phenomenon may cause misunderstanding of certain words or phrases during semantic encoding, thereby affecting the overall quality of semantic understanding. Based on this, in the technical solution of the present application, the sequence of word-granularity semantic features of the first language text is subjected to field effect sequence semantic feature enhancement to obtain the global word-granularity semantic enhancement features of the first language text.

[0042] Specifically, first, the semantic reconstructor used performs feature reconstruction on each first language text word granularity semantic feature vector in the sequence of the first language text word granularity semantic feature vectors to obtain a sequence of reconstructed first language text word granularity semantic feature vectors. The above process can be expressed as: ;in, is a sequence of word-granular semantic feature vectors of the first language text, and are the first, second, and third words in the sequence of the word-granularity semantic feature vectors of the first language text. and First language text word-granular semantic feature vectors, and are the first weight matrix and the second weight matrix respectively, yes The corresponding first language text word granularity semantic feature vector is reconstructed. That is, the semantic reconstructor enables the original word feature vectors in the sequence of the original word feature vectors to achieve dimensional unification, and can capture the semantic information of the words in more detail, especially for polysemous words and context-sensitive words, so that the feature vector is more in line with the meaning in the specific context.

[0043] Specifically, in an embodiment of the present application, the step 332 includes: calculating the semantic depth implicit representation value of each reconstructed first language text word granularity semantic feature vector in the sequence of reconstructed first language text word granularity semantic feature vectors to obtain a sequence of first language text word granularity semantic depth implicit representation values; based on the sorting of the sequence of the first language text word granularity semantic depth implicit representation values, reordering the sequence of the first language text word granularity semantic feature vectors to obtain an updated sequence of the first language text word granularity semantic feature vectors; performing bidirectional global semantic encoding fusion on the updated sequence of the first language text word granularity semantic feature vectors to obtain the first language text global word granularity semantic enhanced feature vector.

[0044] Next, the semantic depth implicit representation value of each reconstructed first language text word granularity semantic feature vector in the sequence of reconstructed first language text word granularity semantic feature vectors is calculated to obtain a sequence of first language text word granularity semantic depth implicit representation values. The above process can be expressed as: ;in, yes The corresponding reconstructed first language text word granularity semantic feature vector, For calculation The square of the norm, represents the logarithmic function value with base 2, For calculation The semantic depth implicit representation value of for The corresponding first language text word granularity semantic depth implicit representation value. In this way, by calculating the semantic depth implicit representation value of these reconstructed first language text word granularity semantic feature vectors, the role of the word in the entire sentence or document can be inferred. In other words, the semantic depth implicit representation value can be used to measure the criticality and influence of each first language text word in its first language text. Based on these quantitative values, the first language text word feature vectors can be reordered to ensure that the most important information can be identified and processed first, so as to ensure that key words are given proper attention in the process of semantic understanding and information extraction.

[0045] Then, based on the order of the sequence of the first language text word granularity semantic depth implicit representation values, the sequence of the first language text word granularity semantic feature vectors is reordered to obtain an updated sequence of the first language text word granularity semantic feature vectors. The above process can be expressed as: ;in, , and are the first, second and third words in the update sequence of the first language text word-granular semantic feature vector. Update the word-granular semantic feature vector of the first language text, is an update sequence of the semantic feature vector of the word granularity of the first language text. In this way, by adjusting the structure of the sentence according to the importance of the first language text vocabulary, the overall semantic encoding can be optimized to be more precise and unified, which helps to ensure that key information is properly focused and processed during text processing and understanding, thereby improving the accuracy and reliability of semantic analysis.

[0046] More specifically, in an embodiment of the present application, the update sequence of the word granularity semantic feature vector of the first language text is respectively subjected to bidirectional global semantic encoding fusion to obtain the global word granularity semantic enhancement feature vector of the first language text, including: inputting the update sequence of the word granularity semantic feature vector of the first language text into the first semantic encoder based on the forward LSTM model to obtain the forward global first language text word granularity semantic encoding feature vector; inputting the update sequence of the word granularity semantic feature vector of the first language text into the first semantic encoder based on the backward LSTM model to obtain the backward global first language text word granularity semantic encoding feature vector; fusing the forward global first language text word granularity semantic encoding feature vector and the backward global first language text word granularity semantic encoding feature vector to obtain the global word granularity semantic enhancement feature vector of the first language text. The above process can be expressed by the formula: .

[0047] in, is an update sequence of the first language text word-granular semantic feature vector, For Perform forward LSTM semantic encoding, is the forward global first language text word granularity semantic encoding feature vector, For Perform backward LSTM semantic encoding, is the backward global first language text word granularity semantic encoding feature vector, and are the forward coding weight matrix and the backward coding weight matrix respectively, is the bias vector, It is the global word granularity semantically enhanced feature vector of the first language text.

[0048] That is to say, the updated first language text word granularity semantic feature vector sequence is doubly encoded using a bidirectional long short-term memory model: on the one hand, the forward LSTM tracks the flow of information from the beginning to the end of the sentence, and on the other hand, the backward LSTM identifies the dependency of information from the end to the beginning of the sentence. This bidirectional encoding strategy can fully grasp the dependencies between contexts in the text, thereby achieving a deep understanding of the text semantics. Through this mechanism, the model can capture the subtle differences in language from two directions, enhancing the ability to parse sentence structure and meaning. And finally, the forward global first language text word granularity semantic encoding feature vector and the backward global first language text word granularity semantic encoding feature vector are fused to combine the global semantic encoding information in the two directions and integrate the bidirectional semantic data to generate a comprehensive and semantically rich first language text global word granularity semantic enhancement feature vector.

[0049] In step 34, the semantic response fusion is performed on the global word granularity semantic enhancement feature of the first language text, the first language embedding feature and the second language embedding feature to obtain a semantic representation of the second language translation text, and the second language translation text is obtained based on the semantic representation of the second language translation text. Specifically, Figure 4 FIG. 3 is a flowchart of step 34 in the medical machine translation self-learning method according to an embodiment of the present application. Figure 4 As shown, the step 34 includes: step 341, passing the first language text global word granularity semantic enhanced feature vector, the first language embedding coding vector and the second language embedding coding vector through a semantic translation response amalgamator to obtain a second language translation text semantic representation vector as the second language translation text semantic representation; step 342, passing the second language translation text semantic representation vector through a translation text generator to obtain the second language translation text.

[0050] In step 341, the global word granularity semantic enhancement feature vector of the first language text, the first language embedding coding vector and the second language embedding coding vector are passed through a semantic translation response fusion device to obtain a second language translation text semantic representation vector as the second language translation text semantic representation. Accordingly, taking into account the global word granularity semantic enhancement features of the first language text, the first language embedding features and the second language embedding features, and in order to integrate a variety of feature information, more comprehensively capture and integrate semantic information from different sources to generate high-quality translation results, in the technical solution of the present application, the global word granularity semantic enhancement feature vector of the first language text, the first language embedding coding vector and the second language embedding coding vector are passed through a semantic translation response fusion device to obtain a second language translation text semantic representation vector. The above process can be expressed as: .

[0051] in, embedding encoding vector for the first language, embedding the encoding vector for the second language, A global word granularity semantically enhanced feature vector for the first language text, represents the logarithmic function value with base 2, represents the exponential function value with the natural constant e as the base, yes function, is the weight vector, is the semantic representation vector of the second language translation text.

[0052] In step 342, the second language translation text semantic representation vector is passed through the translation text generator to obtain the second language translation text. It should be understood that in the medical machine translation process, converting the source language text into the target language text usually involves two key steps: semantic understanding and text generation. In the semantic understanding stage, the medical machine translation model analyzes the content, context and meaning of the source language text and converts it into an internal representation, namely the second language translation text semantic representation vector. This vector captures the core information and intention of the second language text. In order to convert this internal representation into an actual target language text, a translation text generator is required. Specifically, in an embodiment of the present application, the translation text generator is a translation text generator based on AIGC. The translation text generator generally includes important components such as an encoder, a decoder and an attention mechanism. During processing, the second language translation text semantic representation vector is first received as input. Then, it uses an encoder to convert these vectors into an internal representation. In the decoding stage, the generator uses a decoder and an attention mechanism to gradually generate an output text. During this process, the generator selects appropriate words or characters for generation based on the semantic and grammatical information of the input vector, as well as the internal state and attention weight of the decoder, and ultimately generates a second language translation text.

[0053] In particular, when the first language embedding coding vector and the second language embedding coding vector respectively represent the embedding coding features of the first language and the second language of the first initial parallel corpus, and the sequence of the global word granularity semantic enhancement feature vectors of the first language text represents the field effect sequence enhanced semantic features of the first language text, it is expected to improve the semantically consistent aggregation expression effect of the semantic representation vector of the second language translation text obtained by the semantic translation response aggregator.

[0054] Preferably, the step of obtaining the second language translation text by passing the second language translation text semantic representation vector through an AIGC-based translation text generator comprises: performing feature clustering on a feature set of the second language translation text semantic representation vector to obtain a second language translation text semantic class feature set and a second language translation text semantic class feature set, that is: ;in, For each feature value in the feature set within the semantic class of the second language translation text, Each feature value in the feature set outside the semantic class of the second language translation text is translated.

[0055] The ratio of the number of feature values ​​in the feature set within the semantic class of the second language translation text to the number of feature values ​​in the feature set of the semantic representation vector of the second language translation text is calculated, that is: ;in, The number of feature values ​​in the feature set within the semantic class of the second language translation text, The number of feature values ​​of the feature set of the semantic representation vector of the second language translation text, for and The ratio between .

[0056] Calculate the sum of the absolute values ​​of all feature values ​​in the feature set within the semantic class of the second language translation text The power of the sum of the absolute values ​​of all feature values ​​in the feature set of the semantic representation vector of the second language translation text The ratio of the powers is used to obtain the semantic modulation weight of the second language translation text, namely: ;in, For each feature value in the feature set within the semantic class of the second language translation text, The number of feature values ​​in the feature set within the semantic class of the second language translation text, for and The ratio between Represents the calculated value Power, For each feature value in the feature set of the semantic representation vector of the second language translation text, The number of feature values ​​of the feature set of the semantic representation vector of the second language translation text, The weights are modulated for the semantics of the second language translation text.

[0057] Calculate the sum of the squares of all feature values ​​in the feature set within the semantic class of the second language translation text The power of the sum of the squares of all feature values ​​in the feature set of the semantic representation vector of the second language translation text The ratio of the powers is used to obtain the semantic reconciliation weight of the second language translation text, namely: ;in, For each feature value in the feature set within the semantic class of the second language translation text, The number of feature values ​​in the feature set within the semantic class of the second language translation text, for and The ratio between Represents the calculated value Power, For each feature value in the feature set of the semantic representation vector of the second language translation text, The number of feature values ​​of the feature set of the semantic representation vector of the second language translation text, The semantics of the second language translation text is adjusted and weighted.

[0058] For each feature value in the feature set within the semantic class of the second language translation text, the product of the feature value and the semantic reconciliation weight of the second language translation text is calculated, and then the semantic reconciliation weight of the second language translation text is added to obtain the optimized feature value, that is: ;in, The semantics of the second language translation text is adjusted and weighted, For each feature value in the feature set within the semantic class of the second language translation text, modulate the weights for the semantics of the second language translation text, The feature value after optimization is obtained for each feature value in the feature set within the semantic class of the second language translation text.

[0059] For each feature value in the semantic out-of-class feature set of the second language translation text, the product of the feature value and the semantic modulation weight of the second language translation text is calculated to obtain the optimized feature value, that is: ;in, modulate the weights for the semantics of the second language translation text, For each feature value in the semantic out-of-class feature set of the second language translation text, The feature value after optimization is obtained for each feature value in the semantic out-of-class feature set of the second language translation text.

[0060] The optimized second language translation text semantic representation vector based on the optimized feature values ​​of the second language translation text semantic in-class feature set and the second language translation text semantic out-class feature set is passed through the AIGC-based translation text generator to obtain the second language translation text.

[0061] That is, while performing feature clustering on the semantic representation vector of the second language translation text, the interaction description of the key feature information of the semantic representation vector of the second language translation text in the clustering process is carried out, and the geometric equivariant topology of the features is constructed by low-rank harmonic modulation of the features on the basis of equivariance of the clustering features and the features as a whole of the semantic representation vector of the second language translation text, so as to obtain the schematic distribution translation and rotation symmetry of the clustering features of the semantic representation vector of the second language translation text relative to the features as a whole, thereby introducing geometric message passing in the feature expression of the semantic representation vector of the second language translation text, and realizing the cluster mapping symmetry of the semantic representation vector of the second language translation text through the manipulation of irreducible low-rank coefficients, thereby improving the consistency of the clustering-based feature representation of the semantic representation vector of the second language translation text, and improving the text quality of the second language translation text obtained by the semantic representation vector of the second language translation text through the AIGC-based translation text generator.

[0062] Therefore, step 3 is explained clearly, which uses language processing and translation technology based on artificial intelligence and machine translation models to embed the first language and the second language, segment and semantically encode the first language text, and then perform field-effect global semantic enhancement on the word-granularity semantic features of each first language text, so as to automatically generate the second language translation text based on the response fusion representation between the global word-granularity semantic enhancement features of the first language text, the first language embedding features and the second language embedding features. In this way, contextual information can be captured in a larger range, especially when dealing with long sentences and complex structures, the logical relationship and long-distance dependency between sentences can be better understood, and through field-effect global semantic enhancement, the omission of key information can be reduced, ensuring the integrity and accuracy of the translation results.

[0063] In step 4, the translation quality score between the second language translation text and the second language text is calculated. It should be understood that the translation quality score is an important indicator for measuring the accuracy of machine translation results. By comparing the second language translation text and the second language text, the quality of the translation result can be quantitatively evaluated to determine whether the machine translation model accurately conveys the meaning of the original text. More importantly, it can serve as a guiding basis for model optimization. According to the scoring results, the problems existing in the model during the translation process, such as improper vocabulary selection, grammatical errors, semantic understanding deviations, etc., can be analyzed, and the model can be optimized and adjusted in a targeted manner.

[0064] In the process of medical machine translation, calculating the translation quality score is an important step in evaluating model performance and optimizing the model. It can quantitatively evaluate the accuracy, fluency and consistency of the translation results. The following is the specific implementation process of calculating the translation quality score: First, clarifying the goals and standards of the evaluation is the primary task. Translation quality evaluation aims to measure the extent to which the translated text accurately and fluently conveys the information and meaning of the original text, while maintaining consistency with the original text in terms of vocabulary, grammar, sentence structure, etc. Therefore, before the evaluation begins, a set of clear and specific evaluation criteria needs to be established, which may cover multiple aspects such as translation accuracy, semantic integrity, language fluency, consistency of format and structure, etc. These standards will serve as the benchmark and basis for subsequent evaluation work.

[0065] Next, collecting and preparing relevant data is the basis for the evaluation work. This includes obtaining the second language translation text (i.e. the translation results to be evaluated) and the second language text. Ensuring the accuracy and completeness of this data is crucial to the accuracy of the subsequent evaluation. At the same time, if possible, multiple reference translations can also be collected for use in automated evaluation tools to improve the accuracy and reliability of the evaluation.

[0066] In the selection of evaluation methods, usually a combination of automated evaluation tools and manual evaluation is used. Automated evaluation tools such as BLEU score calculate the score by comparing the n-gram matching degree between the translated text and the reference translation, so as to quickly and objectively evaluate the translation quality. The higher the BLEU score, the higher the similarity between the translated text and the reference translation, and the higher the translation quality. However, automated evaluation tools also have their limitations, and they may not be able to fully capture the nuances and contextual information in the language. Therefore, manual evaluation is still indispensable. Manual evaluation is carried out by professional translators or language experts, who carefully proofread and evaluate the accuracy, fluency, consistency and other aspects of the translated text according to the evaluation criteria. During the manual evaluation process, experts will carefully review every word and sentence in the translated text to ensure that they not only accurately convey the information of the original text, but also conform to the expression habits and grammatical rules of the target language. At the same time, manual evaluation can also find errors and deficiencies that may be missed by automated evaluation tools, such as semantic misunderstanding errors and improper handling of cultural differences. In the process of implementing the evaluation, attention should be paid to the fairness and objectivity of the evaluation. Whether it is automated or manual evaluation, strict evaluation standards and processes must be followed to ensure the accuracy and reliability of the evaluation results. After the evaluation is completed, the automated evaluation score and the manual evaluation score need to be comprehensively considered to obtain the final translation quality score.

[0067] In this way, by clarifying appropriate evaluation indicators, implementing automatic and manual evaluations, and comprehensively evaluating the results, the translation quality score can be effectively calculated to provide a reference for subsequent model optimization.

[0068] In step 5, based on the translation quality score, it is determined whether to include the first initial parallel corpus as training corpus to perform self-learning optimization on the medical machine translation model to adjust the model parameters of the medical machine translation model. It should be understood that based on the translation quality score, parallel corpora with low translation quality can be screened out as basic corpora, and the basic corpora are expanded to construct a learning training set with the expanded corpora. These low-quality fine-tuned corpora can be used as training data for further optimization of the model, helping the model to adjust parameters and then learn how to improve these specific translation errors so that the model can better adapt to the needs of medical translation tasks.

[0069] Based on the translation quality score, it is determined whether the first initial parallel corpus should be included in the training corpus, and then the medical machine translation model is self-learning optimized to adjust the model parameters. This process is crucial to improving model performance and translation accuracy. The following is a description of the implementation of this process: After the translation quality score is obtained, the processing method of the first initial parallel corpus will be determined according to the pre-set standards. A threshold is usually set. For example, if the translation quality score is lower than 80 points (this threshold can be flexibly adjusted according to actual conditions), the parallel corpus is regarded as low-quality corpus and will be considered for inclusion in the optimization process; if the score is higher than this threshold, it may be considered that the translation effect of the corpus under the current model is better, and it will not be taken as the focus of optimization for the time being.

[0070] For the first initial parallel corpus that is judged to be of low quality, it will be expanded next. The purpose of expansion is to enrich the content of the corpus and increase the diversity of model learning. There are many ways to expand, such as using medical field knowledge to interpret terms and replace synonyms. In medical texts, many terms have multiple expressions. If the translation of a term in the translated text is not accurate or comprehensive enough, you can refer to professional medical dictionaries or authoritative medical literature to find other synonymous expressions of the term and integrate them into the parallel corpus. At the same time, for some complex grammatical structures, appropriate adjustments and transformations can also be made to generate new corpus instances. For example, for sentences containing clauses, the position of the clauses can be changed or the guide words of the clauses can be adjusted to change the sentence structure, thereby increasing the model's understanding and processing capabilities of different grammatical structures.

[0071] After completing the expansion of the low-quality parallel corpus, we began to build a learning training set. The expanded corpus was sorted and organized to meet the format requirements of model training. This training set will be used specifically for targeted training of medical machine translation models, with a focus on helping the model learn how to improve specific translation errors that have occurred previously. For example, if in previous translations, the model often mistranslates vocabulary collocations in certain specific medical scenarios, then the expanded training set will contain more correct translation examples of such vocabulary collocations, allowing the model to correct incorrect translation patterns by learning from these examples.

[0072] Subsequently, the constructed learning training set is used to carry out self-learning optimization of the medical machine translation model. In this process, the model adjusts its own model parameters based on the corpus in the training set. The algorithm inside the model calculates the difference between the predicted result and the actual correct translation in the training set, and continuously updates the parameters such as weights and biases in the model based on this difference through the back propagation algorithm. The weights determine the importance of different features in the model's decision-making process, and the biases are used to adjust the model's output. As the training progresses, the model gradually learns more appropriate parameter values, enabling it to make more accurate predictions when faced with similar medical translation tasks.

[0073] During the self-learning optimization of the model, continuous evaluation of the model's performance is an indispensable step. This can be achieved by recalculating the translation quality score, comparing the translation results at different stages, and analyzing the model's performance when processing specific types of medical texts. For example, after a round of training, reselect some samples for translation and calculate the score to see if the score has improved; or compare the translation effect of the model on complex medical sentences before and after training to see if there is improvement in structure and semantic understanding. If it is found that there are still problems with the model or there is room for further improvement, it is necessary to repeat the above steps of expanding the corpus, building a training set, training the model, and evaluating the performance, and continuously adjust and improve the model until the model achieves a relatively satisfactory translation accuracy and effect, can accurately and fluently handle various medical translation tasks, and meet the strict requirements of the medical field for translation accuracy, terminology accuracy, and adaptability to clinical research reading habits. Through this cyclical optimization process, the medical machine translation model can continue to evolve and gradually improve its capabilities and performance in the field of medical translation.

[0074] In summary, the self-learning method of medical machine translation based on the embodiment of the present application is explained, which first constructs a medical machine translation model, and then collects an initial parallel corpus consisting of a first language text and its corresponding second language text. Then, using these corpora, the medical machine translation model translates the first language text into a translated text of a second language. Subsequently, the translated text is compared with the original second language text to calculate the translation quality score. Finally, based on the translation quality score, it is decided whether to use these initial parallel corpora for training to perform self-learning optimization on the medical machine translation model. In this way, it is conducive to optimizing the medical machine translation model, thereby improving the accuracy of medical translation.

[0075] In an embodiment of the present application, a medical machine translation self-learning device is also provided for executing the medical machine translation self-learning method as described above. Figure 5 FIG. 1 is a block diagram of a medical machine translation self-learning device according to an embodiment of the present application. Figure 5 As shown, according to the embodiment of the present application, the medical machine translation self-learning device 100 includes: a language embedding encoding module 110, which is used to embed the first language and the second language respectively to obtain the first language embedding feature and the second language embedding feature; a language text word granularity semantic analysis module 120, which is used to perform word granularity semantic analysis on the first language text to obtain a sequence of word granularity semantic features of the first language text; a semantic feature enhancement module 130, which is used to perform field effect sequence semantic feature enhancement on the sequence of word granularity semantic features of the first language text to obtain the global word granularity semantic enhancement feature of the first language text; a second language translation text generation module 140, which is used to perform semantic response fusion on the global word granularity semantic enhancement feature of the first language text, the first language embedding feature and the second language embedding feature to obtain the second language translation text semantic representation, and obtain the second language translation text based on the second language translation text semantic representation.

[0076] Here, those skilled in the art will appreciate that the specific functions and operations of the various units and modules in the above-mentioned medical machine translation self-learning device 100 have been described in detail above. Figures 1 to 4 The medical machine translation self-learning method has been introduced in detail, and therefore, its repeated description will be omitted.

[0077] In summary, the medical machine translation self-learning device 100 based on the embodiment of the present application is explained, which first constructs a medical machine translation model, and then collects an initial parallel corpus consisting of a first language text and its corresponding second language text. Then, using these corpora, the medical machine translation model translates the first language text into a translated text in the second language. Subsequently, the translated text is compared with the original second language text to calculate the translation quality score. Finally, based on the translation quality score, it is decided whether to use these initial parallel corpora for training to perform self-learning optimization on the medical machine translation model. In this way, it is conducive to optimizing the medical machine translation model, thereby improving the accuracy of medical translation.

[0078] Below, reference Figure 6 To describe an electronic device according to an embodiment of the present application. Figure 6 FIG. 1 is a block diagram of an electronic device according to an embodiment of the present application. Figure 6 As shown, the electronic device may include: a processor 10, a communication interface 12, a memory 13 and a communication bus 14, wherein the processor 11, the communication interface 12 and the memory 13 communicate with each other through the communication bus 14. The processor 11 may call the logic instructions in the memory 13 to execute the above-mentioned medical machine translation self-learning method.

[0079] In addition, the logic instructions in the above-mentioned memory 13 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

Claims

1. A medical machine translation self-learning method, characterized in that: The method comprises: step 1: establishing a medical machine translation model; step 2: obtaining a first initial parallel corpus, wherein the first initial parallel corpus consists of a first language, a first language text, a second language and a second language text; step 3: based on the first language and the second language of the first initial parallel corpus, using the medical machine translation model to translate the first language text to generate a second language translation text; step 4: calculating a translation quality score between the second language translation text and the second language text; step 5: based on the translation quality score, determining whether to include the first initial parallel corpus as a training corpus to perform self-learning optimization on the medical machine translation model to adjust the model parameters of the medical machine translation model ;Wherein, the step 3 includes: respectively performing embedding coding on the first language and the second language to obtain first language embedding features and second language embedding features; performing word granularity semantic analysis on the first language text to obtain a sequence of word granularity semantic features of the first language text; performing field effect sequence semantic feature enhancement on the sequence of word granularity semantic features of the first language text to obtain global word granularity semantic enhancement features of the first language text; performing semantic response fusion on the global word granularity semantic enhancement features of the first language text, the first language embedding features and the second language embedding features to obtain a second language translation text semantic representation, and obtaining the second language translation text based on the second language translation text semantic representation; The method of performing field effect sequence semantic feature enhancement on the sequence of word granularity semantic features of the first language text to obtain global word granularity semantic enhancement features of the first language text comprises: using a semantic reconstructor to perform feature reconstruction on each first language text word granularity semantic feature vector in the sequence of first language text word granularity semantic feature vectors to obtain a sequence of reconstructed first language text word granularity semantic feature vectors; based on the sequence of reconstructed first language text word granularity semantic feature vectors, performing deep implicit semantic global enhancement on the sequence of first language text word granularity semantic features to obtain a first language text global word granularity semantic enhancement feature vector as the first language text global word granularity semantic enhancement feature; Among them, based on the sequence of reconstructed first language text word granularity semantic feature vectors, the sequence of the first language text word granularity semantic features is globally enhanced with deep implicit semantics to obtain the first language text global word granularity semantic enhanced feature vector, including: calculating the semantic deep implicit representation value of each reconstructed first language text word granularity semantic feature vector in the sequence of reconstructed first language text word granularity semantic feature vectors to obtain a sequence of first language text word granularity semantic deep implicit representation values; based on the sorting of the sequence of the first language text word granularity semantic deep implicit representation values, the sequence of the first language text word granularity semantic feature vectors is reordered to obtain an updated sequence of the first language text word granularity semantic feature vectors; and the updated sequence of the first language text word granularity semantic feature vectors is respectively subjected to bidirectional global semantic encoding fusion to obtain the first language text global word granularity semantic enhanced feature vector.

2. The medical machine translation self-learning method according to claim 1, characterized in that: The first language and the second language are respectively embedded and encoded to obtain a first language embedding feature and a second language embedding feature, including: embedding the first language through a first language embedding matrix to obtain a first language embedding coding vector as the first language embedding feature, and embedding the second language through a second language embedding matrix to obtain a second language embedding coding vector as the second language embedding feature.

3. The medical machine translation self-learning method according to claim 2, characterized in that: The first language text is subjected to word-granularity semantic analysis to obtain a sequence of word-granularity semantic features of the first language text, including: performing word segmentation processing on the first language text and then passing the first language text semantic encoder based on the BERT model to obtain a sequence of first language text word-granularity semantic feature vectors as the sequence of word-granularity semantic features of the first language text.

4. The medical machine translation self-learning method according to claim 3, characterized in that: The update sequence of the word granularity semantic feature vector of the first language text is subjected to bidirectional global semantic encoding fusion to obtain the global word granularity semantic enhancement feature vector of the first language text, including: inputting the update sequence of the word granularity semantic feature vector of the first language text into a first semantic encoder based on a forward LSTM model to obtain a forward global first language text word granularity semantic encoding feature vector; inputting the update sequence of the word granularity semantic feature vector of the first language text into a first semantic encoder based on a backward LSTM model to obtain a backward global first language text word granularity semantic encoding feature vector; and fusing the forward global first language text word granularity semantic encoding feature vector and the backward global first language text word granularity semantic encoding feature vector to obtain the global word granularity semantic enhancement feature vector of the first language text.

5. The medical machine translation self-learning method according to claim 4, characterized in that: The global word granularity semantic enhancement feature of the first language text, the first language embedding feature and the second language embedding feature are semantically fused to obtain a semantic representation of a second language translation text, and the second language translation text is obtained based on the semantic representation of the second language translation text, including: passing the global word granularity semantic enhancement feature vector of the first language text, the first language embedding coding vector and the second language embedding coding vector through a semantic translation response fuser to obtain a second language translation text semantic representation vector as the second language translation text semantic representation; passing the second language translation text semantic representation vector through a translation text generator to obtain the second language translation text.

6. The medical machine translation self-learning method according to claim 5, characterized in that: The translation text generator is an AIGC-based translation text generator.

7. A medical machine translation self-learning device, used to execute the medical machine translation self-learning method according to claim 1, characterized in that: The medical machine translation self-learning device includes: a language embedding encoding module, which is used to embed the first language and the second language respectively to obtain the first language embedding feature and the second language embedding feature; a language text word granularity semantic analysis module, which is used to perform word granularity semantic analysis on the first language text to obtain a sequence of word granularity semantic features of the first language text; a semantic feature enhancement module, which is used to perform field effect sequence semantic feature enhancement on the sequence of word granularity semantic features of the first language text to obtain the first language text global word granularity semantic enhancement feature; a second language translation text generation module, which is used to perform semantic response fusion on the first language text global word granularity semantic enhancement feature, the first language embedding feature and the second language embedding feature to obtain the second language translation text semantic representation, and obtain the second language translation text based on the second language translation text semantic representation.

8. An electronic device, comprising a memory and a processor, characterized in that: The processor is configured to execute the medical machine translation self-learning method according to any one of claims 1 to 6 based on the instructions stored in the memory.

Images