Language translation processing method and device, equipment, medium and product

Abstract

The invention provides a language translation processing method and device, equipment, a medium and a product. The method comprises the steps of obtaining to-be-translated source language data; translating the source language data into target language data through a language translation model to obtain a target language translation result; wherein the language translation model is obtained through training of training samples of a source language and similar languages thereof, the similar languages are languages meeting corresponding similarity requirements with the source language under each language feature dimension, and each language feature dimension comprises at least one of a sub-word dimension, a phoneme dimension and a semantic dimension. The translation quality of the low-resource language can be improved.

Description

Technical Field

[0001] The present application relates to the field of machine translation, and in particular to a language translation processing method, apparatus, device, medium and product. Background Art

[0002] Machine translation refers to the use of computers to automatically translate one language into another, thereby effectively promoting information exchange and communication between different languages.

[0003] Neural Machine Translation (NMT) is currently the most widely used translation method. Although this method offers excellent translation quality, its performance depends heavily on the availability of corpus resources. This means that NMT performs well only on languages with abundant resources, but often suffers from low translation accuracy in low-resource languages. Summary of the Invention

[0004] Based on the above-mentioned technical status, the present application provides a language translation processing method, device, equipment, medium and product, which can improve the accuracy of language translation of low-resource languages.

[0005] In order to achieve the above technical objectives, this application specifically proposes the following technical solutions: According to a first aspect of an embodiment of the present application, a language translation processing method is provided, comprising: obtaining source language data to be translated; translating the source language data into target language data using a language translation model to obtain a target language translation result; wherein the language translation model is trained using training samples of the source language and similar languages thereof, wherein the similar languages are languages that meet corresponding similarity requirements with the source language in various language feature dimensions, wherein the various language feature dimensions include at least one of a subword dimension, a phoneme dimension, and a semantic dimension.

[0006] In some implementations, the language translation model is trained using the following steps: determining similar languages to the source language; generating a training sample set based on training samples corresponding to the source language and training samples corresponding to the similar languages; pre-training a preset neural network architecture based on the training sample set to obtain a pre-trained translation model; and performing transfer learning training on the pre-trained translation model based on the training samples corresponding to the source language to obtain the language translation model.

[0007] In some implementations, determining similar languages to the source language includes: determining various related languages of the source language; determining, for each of the various language feature dimensions, the similarity between the source language and the various related languages in the language feature dimension; determining related languages whose similarity with the source language in the language feature dimension is greater than a preset similarity threshold as similar languages to the source language in the language feature dimension; and determining the union of the similar languages of the source language in each language feature dimension as the similar languages of the source language.

[0008] In some implementations, when the language feature dimension includes the subword dimension, determining the similarity between the source language and each of the related languages in the language feature dimension includes: determining the overlap of subwords in a subword dictionary of the source language and subword dictionaries corresponding to each of the related languages, to obtain the similarity between the source language and each of the related languages in the subword dimension; and / or, when the language feature dimension includes the phoneme dimension, determining the similarity between the source language and each of the related languages in the language feature dimension includes: determining the similarity of phonemes in a phoneme dictionary of the source language and phoneme dictionaries corresponding to each of the related languages, to obtain the similarity between the source language and each of the related languages in the phoneme dimension; and / or, when the language feature dimension includes the semantic dimension, determining the similarity between the source language and each of the related languages in the language feature dimension includes: determining the similarity between a training sample of the source language and a source language translation result corresponding to each of the related languages, to obtain the similarity between the source language and each of the related languages in the semantic dimension.

[0009] In some implementations, the method further includes: inputting the training samples of the similar language into the language translation model to obtain source language translation data; determining the source language translation data that meets preset scoring requirements as optimized training samples; and performing optimization training on the language translation model based on the optimized training samples to obtain an optimized language translation model.

[0010] In some implementations, determining source language translation data that meets preset scoring requirements as optimized training samples includes: determining a score for the source language translation data, the score including at least one of a language fluency score, a back-translation quality score, and a match score with a preset source language terminology database; determining a comprehensive score for the source language translation data based on at least one of the language fluency score, the back-translation quality score, and the match score; and if the comprehensive score is greater than or equal to a preset scoring threshold, determining that the source language translation data meets the preset scoring requirements, and determining the source language translation data as the optimized training sample.

[0011] In some implementations, when the score includes the language fluency score, determining the language fluency score of the source language translation data includes: dividing the source language translation data into n-grams; determining the conditional probability of each n-gram using a language model; and determining the fluency score of the source language translation data based on a cumulative multiplication result of the conditional probabilities of each n-gram and a weight coefficient corresponding to the language fluency score.

[0012] In some implementations, when the score includes the back-translation quality score, determining the back-translation quality score of the source language translation data includes: translating the source language translation data into translation data in respective similar languages using the language translation model, and back-translating the translation data in respective similar languages into the source language to obtain respective source language back-translation data; and obtaining the back-translation score of the source language translation data based on a weighted sum of similarities between each of the respective source language back-translation data and the source language translation data and a weight coefficient corresponding to the back-translation quality score.

[0013] According to a second aspect of an embodiment of the present application, a language translation processing device is provided, comprising: an acquisition unit, configured to acquire source language data to be translated; and a translation processing unit, configured to translate the source language data into target language data using a language translation model to obtain a target language translation result; wherein the language translation model is trained using training samples of the source language and similar languages thereof, wherein the similar languages are languages that meet corresponding similarity requirements with the source language in various language feature dimensions, wherein the various language feature dimensions include at least one of a subword dimension, a phoneme dimension, and a semantic dimension.

[0014] According to a third aspect of an embodiment of the present application, an electronic device is provided, comprising a memory and a processor; the memory is connected to the processor and is used to store a program; the processor is used to implement the language translation processing method as described in the first aspect by running the program in the memory.

[0015] According to a fourth aspect of the embodiments of the present application, a computer program product is provided, comprising computer program instructions, which, when executed by a processor, cause the processor to execute: the language translation processing method as described in the first aspect.

[0016] Embodiments of the present application provide a language translation processing method, apparatus, device, medium, and product. The method comprises obtaining source language data to be translated and translating the source language data into target language data using a language translation model to obtain a target language translation result. The language translation model is trained using training samples of the source language and similar languages, where the similar languages are languages that meet corresponding similarity requirements with the source language in various language feature dimensions, including at least one of a subword dimension, a phoneme dimension, and a semantic dimension. Because the language translation model is trained not only based on training samples of the source language but also includes training samples of "similar languages" that are similar to the source language in multiple language feature dimensions, similar languages often share some basic language building blocks (e.g., common subwords or phonemes) across different languages. This allows the model to identify and utilize these commonalities across different languages to enhance understanding of the source language, thereby better capturing the linguistic characteristics of the source language, enhancing the model's ability to understand and generate the source language, promoting positive transfer, and improving the translation quality of the translation processing model in low-resource translation scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without any creative work.

[0018] Figure 1 A flowchart of a language translation processing method provided in an embodiment of the present application.

[0019] Figure 2 A flowchart of the training process of a language translation model provided in an embodiment of the present application.

[0020] Figure 3 A schematic diagram of the principle of determining similar languages to a source language provided in an embodiment of the present application.

[0021] Figure 4 A flowchart of a model optimization training provided in an embodiment of the present application.

[0022] Figure 5 A flowchart for determining optimized training samples provided in an embodiment of the present application.

[0023] Figure 6 Schematic diagram of the principle of screening and optimizing training samples provided in an embodiment of the present application.

[0024] Figure 7 A schematic diagram of the structure of a language translation processing device provided in an embodiment of the present application.

[0025] Figure 8 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0026] The technical solution proposed in the embodiments of this application is applicable to various scenarios such as cross-language real-time communication systems, automatic generation of bilingual subtitles for multimedia content, localization processing of product information on cross-border e-commerce platforms, and large-scale online translation services.

[0027] The technical solutions provided in the embodiments of the present application can be exemplarily applied to hardware devices such as processors, electronic devices, and servers (including cloud servers), or packaged into software programs and run. When the hardware devices execute the processing of the technical solutions in the embodiments of the present application, or the above-mentioned software programs are run, the automatic splitting of target tasks and the automatic calling of the application program interfaces required for the tasks can be achieved, thereby completing the purpose of the target tasks. The embodiments of the present application only provide an illustrative introduction to the specific processing of the technical solutions in the present application, and do not limit the specific implementation form of the technical solutions in the present application. Any technical implementation form that can execute the processing of the technical solutions in the present application can be adopted by the embodiments of the present application.

[0028] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0029] Before introducing this application solution, we first introduce the relevant technologies: Neural machine translation systems demonstrate significant performance differences between resource-rich languages like Chinese, English, and Korean, and low-resource languages like Uyghur and Tibetan. Because high-resource languages have vast parallel corpora and rich linguistic resources, their translation quality is gradually approaching that of professional translators. In contrast, low-resource languages like Uyghur and Tibetan often face higher error rates, weaker syntactic and semantic understanding, and poor cross-domain generalization due to scarce training data, a lack of pre-trained models, and difficult-to-capture linguistic characteristics.

[0030] To improve translation quality for low-resource languages, improvements are currently being made primarily at the data and model building levels. At the data level, parallel corpora are expanded through methods such as lexical substitution, synonymous syntactic transformation, back-translation, and its iterative variants. At the model building level, universal translation models based on a shared multilingual representation space, such as the Transformer-based multilingual encoder-decoder, are utilized to adapt to low-resource languages through cross-lingual knowledge transfer and efficient parameter fine-tuning. Zero-shot / few-shot translation techniques, combined with temperature-controlled decoding strategies, further enhance the model's generalization capabilities in resource-poor scenarios.

[0031] While the aforementioned methods can improve translation quality for low-resource languages to a certain extent, there is still a significant gap between their translation quality and that of mainstream languages. For example, at the data level, the target language text generated by back-translation may retain the syntactic features of the source language, resulting in insufficient domain adaptability. Furthermore, because back-translation relies on the effectiveness of existing translation models, if the model itself is ineffective, the quality of the back-translated data may decline and contain more noise.

[0032] At the model level, the efficiency of cross-lingual knowledge transfer is affected by the distance between the language lines. For languages with complex morphological structures or widely varying linguistic features, the transfer effect is often limited. Therefore, when performing translation tasks based on a universal multilingual translation model, since multiple languages share the same parameter space, negative transfer may occur. This means that high-resource languages tend to dominate the model learning process, thereby weakening the representational capabilities of low-resource languages.

[0033] In view of this, the embodiments of the present application are dedicated to providing a language translation processing method, apparatus, device, medium, and product. During the model training process, languages that meet corresponding similarity requirements with the source language in various language feature dimensions are obtained as similar languages, wherein each language feature dimension includes at least one of a subword dimension, a phoneme dimension, and a semantic dimension. The source language and its similar languages are then used for model training. In this way, when the trained model is applied to a translation task, the translation quality can be improved. Detailed descriptions are given one by one in the following embodiments.

[0034] Exemplary Methods Figure 1 This is a flow chart of a language translation processing method provided in an embodiment of the present application. Figure 1 As shown, the language translation processing method provided in this embodiment includes steps S101-S102: S101: Acquire source language data to be translated.

[0035] This step aims to collect the target content that the user needs to translate, typically including source language text data. Source language text can be obtained in a variety of ways, such as by having the user directly enter the text to be translated into an input box within a translation application, translation app, or web interface; or by using voice recognition technology to automatically transcribe the user's voice input into the corresponding text content. Additionally, source language data can be obtained from methods such as document uploads and optical character recognition (OCR) in images.

[0036] S102: Translate the source language data into the target language data using a language translation model to obtain a target language translation result.

[0037] Among them, the language translation model is obtained by training samples of the source language and its similar languages. The similar languages are languages that meet the corresponding similarity requirements with the source language in various language feature dimensions. Each language feature dimension includes at least one of the subword dimension, phoneme dimension and semantic dimension.

[0038] Among them, the subword dimension is used to evaluate the similarity of vocabulary structure between different languages, such as whether they share similar word formation rules, prefixes, suffixes or word segmentation methods.

[0039] The phoneme dimension is used to evaluate the similarity between languages from the perspective of the phonetic system, such as the overlap of vowel and consonant systems, the similarity of syllable structures, etc.

[0040] The semantic dimension is evaluated based on the similarity of lexical semantics and syntactic structure. Multilingual embedding space or pre-trained language models can be used to measure the semantic closeness of sentences or words between different languages.

[0041] This embodiment integrates training samples from the source language and similar languages. That is, the language translation model is trained not only based on training samples from the source language, but also incorporates training samples from "similar languages" that share similarities with the source language across multiple linguistic feature dimensions. Since similar languages often share some basic language building blocks (e.g., common subwords or phonemes), the model can identify and utilize these commonalities across different languages to enhance its understanding of the source language. For example, in low-resource languages, by incorporating data from similar languages, the model can learn a broader set of common linguistic rules, such as vocabulary representation, syntactic patterns, and semantic associations, from more examples. This allows the model to learn additional linguistic information and contextual understanding capabilities for the language translation model, even if the two languages appear to be quite different on the surface, due to significant commonalities in subword structure, phoneme system, or semantic expression. This allows the model to better capture the linguistic characteristics of the source language, enhance the model's understanding and generation capabilities for the source language, promote positive transfer, and improve the translation quality of the translation processing model in low-resource translation scenarios.

[0042] Figure 2 The following is a flow chart of the training process of a language translation model provided in an embodiment of the present application. Figure 2 As shown, the training process of the language translation model includes the following steps S201-S204: S201: Determine languages similar to the source language.

[0043] Similar languages are those that share at least one linguistic feature dimension with the source language and meet a preset similarity threshold. For example, these languages share similar grammar and a high degree of phrase overlap with the source language. Source languages can include low-resource languages such as Uyghur and Tibetan.

[0044] In some embodiments, the specific implementation of step S201 includes the following steps a1-a4: Step a1: Determine various related languages of the source language.

[0045] Figure 3 This is a flow chart for determining similar languages to a source language provided in an embodiment of the present application. Figure 3 As shown, languages that are close to the source language in terms of language family or region can be preliminarily screened out as related languages from the dimensions of language family and / or region.

[0046] For example, taking Uyghur and Tibetan as examples, the following related languages can be screened out from the perspective of language families: Then, the filtering rules are used to select languages that meet the similarity requirements as similar languages. The specific filtering process can be found in the detailed description of steps a2-a4 below: Step a2: for each of the language feature dimensions, determine the similarity between the source language and each of the related languages in the language feature dimension.

[0047] Each language feature dimension includes at least one of a subword dimension, a phoneme dimension, and a semantic dimension. For different language feature dimensions, step a2 has different similarity calculation methods, as follows: Continue reading Figure 3 In some examples, when the language feature dimension includes a subword dimension, determining the similarity between the source language and each related language in the language feature dimension includes: determining the overlap of subwords in the subword dictionary of the source language and the subword dictionaries corresponding to each related language, and obtaining the similarity between the source language and each related language in the subword dimension.

[0048] In this example, we first need to use the sentence piece modeling method based on the existing text data of the source language and each related language to build a source language seed word dictionary and a subword dictionary corresponding to each related language.

[0049] Specifically, we first need to determine the dictionary size required to fully cover the language characters, and then add 1,000 characters to this size as the final dictionary capacity; then, using this setting, we train a subword dictionary of the source language.

[0050] Similarly, a similar approach can be used for other languages related to the source language. That is, based on the existing text data, sentence piece modeling is used, and the dictionary size is set to the size required for full character coverage plus 1000, thereby training a sub-dictionary for the relevant language.

[0051] Next, we can calculate the overlap between each pair of subwords in the subword dictionary of the source language and the subword dictionary of each related language. The specific calculation formula is shown in the following formula (1): ; (1) Where V 源语种 V represents the dictionary set of subwords in the source language; 相关语种 Represents a dictionary set of subwords in related languages; ∩ means taking the intersection, ∪ means taking the union, and |·| means counting the number of elements in the set.

[0052] The subword overlap calculated by the above formula (1) can be used as the similarity between the source language and the related language in the subword dimension.

[0053] In other examples, when the language feature dimension includes a phoneme dimension, determining the similarity between the source language and each related language in the language feature dimension includes: determining the similarity between the phonemes in the phoneme dictionary of the source language and the phoneme dictionaries corresponding to each related language, and obtaining the similarity between the source language and each related language in the phoneme dimension.

[0054] Continue reading Figure 3 In this example, the existing text data in the source language and related languages must first be preprocessed to build a phoneme dictionary for the source language and a phoneme dictionary for each related language. This preprocessing includes cleaning punctuation marks and processing Arabic numerals in the text data.

[0055] Next, we can calculate the similarity between each pair of phonemes in the phoneme dictionary of the source language and the phoneme dictionary of each related language. The specific calculation formula is shown in the following formula (2): ; (1) Where U 源语种 Represents the dictionary set of phonemes in the source language; U 相关语种Represents the dictionary set of phonemes in each relevant language; ∩ means taking the intersection, ∪ means taking the union, and |·| means calculating the number of elements in the set.

[0056] The similarity between the source language and the related language in the phoneme dimension can be calculated using the above formula (2).

[0057] In some other examples, when the language feature dimension includes a semantic dimension, determining the similarity between the source language and each related language in the language feature dimension includes: determining the similarity between the training sample of the source language and the source language translation results corresponding to each related language, and obtaining the similarity between the source language and each related language in the semantic dimension.

[0058] Continue reading Figure 3 In this example, we first need to obtain parallel translation corpora between the source language and various related languages. Parallel translation corpora refer to sentence pairs translated between different languages, such as a Uyghur sentence and its corresponding Chinese translation.

[0059] The parallel translation corpus was then processed using the Word2vec model to generate word vectors for each word in different languages. A word vector is a numerical vector that maps words into a high-dimensional space and can reflect the semantic relationship between words.

[0060] Next, we calculate the cosine similarity between the word vectors of each word in the source language and the word vectors of the corresponding words in the related languages. This serves as the semantic similarity between the source language and each related language. Cosine similarity measures the degree of directional similarity between two vectors, ranging from -1 to 1. Values closer to 1 indicate closer semantic similarity between the two words.

[0061] Step a3: Determine the related languages whose similarity with the source language in the language feature dimension is greater than a preset similarity threshold as similar languages to the source language in the language feature dimension.

[0062] Continue reading Figure 3 In the case where the language feature dimension includes a subword dimension, step a3 includes: selecting related languages whose subword overlap is greater than or equal to a preset overlap threshold as similar languages to the source language in the subword dimension, that is, the similar language set 1 shown in the figure.

[0063] When the language feature dimension includes the phoneme dimension, step a3 includes: selecting related languages whose phoneme similarity is greater than or equal to a preset similarity threshold as similar languages of the source language in the phoneme dimension, i.e., the similar language set 2 shown in the figure.

[0064] When the language feature dimension includes a semantic dimension, step a3 includes: selecting related languages whose cosine similarity is greater than or equal to a preset similarity threshold as similar languages of the source language in the semantic dimension, i.e., similar language set 3 shown in the figure.

[0065] Step a4: determine the union of similar languages of the source language in each language feature dimension as the similar languages of the source language.

[0066] Continue reading Figure 3 Since languages with high subword overlap, high phoneme similarity, or high semantic relevance with the source language are helpful for the subsequent training of the neural translation model, and in order to obtain more training data, the "similar language set 1", "similar language set 2", and "similar language set 3" selected according to the above three language feature dimensions can be unioned and deduplicated to finally obtain languages similar to the source language.

[0067] Continue reading Figure 2 After step S201, the following step S202 may also be included.

[0068] S202: Generate a training sample set based on the training samples corresponding to the source language and the training samples corresponding to similar languages.

[0069] In step S202, in order to obtain more training data and improve the model training effect, the text data corresponding to the standard language of the source language, the text data corresponding to the dialect of the source language and the text data corresponding to each similar language can be integrated to construct a training sample set.

[0070] S203: Pre-training a preset neural network architecture according to the training sample set to obtain a pre-trained translation model.

[0071] In this step, by inputting the constructed training sample set into the preset neural network architecture for multilingual pre-training, a many-to-many translation model covering multiple languages and dialects, namely a pre-trained translation model, can be obtained.

[0072] The architecture can be a multilingual neural machine translation model based on the Transformer structure. This model has an encoder-decoder structure and supports joint modeling and parameter sharing for multiple languages.

[0073] Furthermore, by performing multi-task learning and cross-lingual representation learning on a large-scale, multilingual training set, the model can automatically extract common linguistic features across different languages and establish cross-lingual semantic mappings. During training, the model not only learns to translate from the source language to the target language, but also simultaneously masters the ability to convert between multiple languages, thereby forming a unified multilingual semantic space.

[0074] After sufficient training, the obtained pre-trained translation model has the ability to handle translation tasks between multiple languages, that is, "many-to-many" translation capability. It can not only achieve high-quality translation between high-resource languages, but also support translation reasoning of low-resource languages, dialects or unseen languages, thereby improving the generalization performance and adaptability of the model.

[0075] S204: Perform transfer learning training on the pre-trained translation model based on training samples corresponding to the source language to obtain a language translation model.

[0076] After obtaining a pre-trained translation model with multi-language understanding and translation capabilities, in order to further improve its translation performance in specific source languages, especially low-resource languages or dialect scenarios, this embodiment can also use transfer learning to refine the model training.

[0077] Specifically, text data related to the source language can be used as training samples, including the standard language text of the language and its dialect variant texts, to further fine-tune the pre-trained model so that the model can learn the language structure, vocabulary usage, syntactic features and expression habits of the source language more deeply, thereby significantly enhancing its language understanding and generation capabilities in that language.

[0078] During transfer learning, the model gradually adjusts its parameter distribution to better adapt to the linguistic characteristics of the source language, while retaining the cross-lingual commonality learned during the initial pre-training phase. This "general first, specialized later" training strategy not only improves the model's adaptability to low-resource languages but also mitigates overfitting issues caused by limited data.

[0079] Ultimately, the language translation model obtained after transfer learning training can provide more accurate, natural and contextual translation results in translation tasks targeting the source language (especially its dialects or resource-scarce variants).

[0080] This embodiment introduces linguistic knowledge to screen the related languages of the source language from multiple language feature dimensions, including subword overlap, phoneme similarity and semantic relevance. Through this multi-dimensional similarity screening, similar languages that have commonalities with the source language at different language levels can be identified. On this basis, the text data of these similar languages are combined with the dialect data of the source language and used together in the pre-training process of the neural translation model. Due to the integration of multi-level language resources, the model can learn a wider range of language feature representations and cross-language semantic mapping relationships. Subsequently, the model pre-trained on the above-mentioned multilingual and dialect data is migrated to a specific source language translation task to promote forward migration, enhance the model's understanding and generation capabilities of low-resource languages, and improve its translation quality and model generalization capabilities in low-resource languages.

[0081] To further improve the model's translation performance for low-resource languages, the language translation model trained above can be used to filter out high-quality source language translation data. This high-quality source language translation data can then be used to retrain the model to optimize its translation performance for low-resource languages. The following diagram details the model optimization and training process: Figure 4 This is a flow chart of a model optimization training provided in the embodiment of the present application. Figure 4 As shown, the optimization training process of the model includes the following steps S401-S403: S401: Input training samples of similar languages into a language translation model to obtain source language translation data.

[0082] The training samples of similar languages include text data. The text data of similar languages is translated into source language data through a language translation model to obtain source language translation data.

[0083] S402: Determine the source language translation data that meets the preset scoring requirements as optimized training samples.

[0084] After obtaining the source language translation data, we can filter out high-quality source language translation data as training samples for model optimization, so as to optimize the language translation model and further improve the model's translation performance in low-resource languages. The following details the specific implementation of this screening process: In some embodiments, as Figure 5 As shown, the specific implementation of step S402 includes the following steps S501-S503: S501: Determine a score for the source language translation data, where the score includes at least one of a language fluency score, a back-translation quality score, and a match score with a preset source language terminology database.

[0085] This step aims to determine the score of the source language translation data under each screening dimension. Among them, each screening dimension includes at least one of: n-grams screening dimension, back-translation quality screening dimension and term library matching dimension. Specifically, the language fluency score can be obtained through the n-grams screening dimension; the back-translation quality score can be obtained through the back-translation quality screening dimension; and the matching degree score with the preset source language term library can be obtained through the matching dimension with the preset source language term library. Among them, the scores under each screening dimension can be obtained in different implementation methods, as follows: In some implementations, when the score includes a language fluency score, determining the language fluency score of the source language translation data includes: dividing the source language translation data into individual n-grams; determining the conditional probability of each n-gram using a language model; and obtaining the fluency score of the source language translation data based on a cumulative multiplication result of the conditional probabilities of each n-gram and a weight coefficient corresponding to the language fluency score.

[0086] like Figure 6 As shown, this example aims to filter out high-quality source language translation data based on the n-grams screening dimension. Specifically, the n-gram Bi-Gram model can be used to calculate the logarithmic probability of each sentence in the source language translation data, thereby obtaining a fluency score of 1 for the source language translation data under the n-grams screening dimension. The process specifically includes: dividing the source language translation text into multiple consecutive n-gram phrases, and calculating the conditional probability of each n-gram phrase through a pre-trained Bi-Gram model; then, by accumulating the conditional probabilities of all n-gram phrases and multiplying them by the weight coefficient w1 corresponding to the language fluency dimension, the fluency score of the source language translation data is obtained. Here, n can take a value of 2.

[0087] In other implementations, when the scoring includes a back-translation quality score, determining the back-translation quality score of the source language translation data includes: translating the source language translation data into translation data in respective similar languages using a language translation model, and back-translating the translation data in respective similar languages into the source language to obtain respective source language back-translation data; and obtaining the back-translation score of the source language translation data based on a weighted sum of similarities between each respective source language back-translation data and the source language translation data and a weight coefficient corresponding to the back-translation quality score thereof.

[0088] In the dimension of back-translation quality screening, assuming that each similar language includes three similar languages, which are respectively recorded as similar language 1, similar language 2, and similar language 3, the implementation method includes: First, the source language translation data is translated into similar language 1, similar language 2, and similar language 3 through a language translation model to obtain translation data corresponding to similar language 1, similar language 2, and similar language 3.

[0089] Afterwards, the translation data in similar language 1, similar language 2, and similar language 3 are translated back to the source language through the language translation model to obtain source language back-translation data 1, source language back-translation data 2, and source language back-translation data 3.

[0090] Finally, a similarity algorithm is used to determine the similarity between the source language translation data and source language back-translation data 1, source language back-translation data 2, and source language back-translation data 3. These three similarities are weighted and summed, and multiplied by the corresponding weight coefficient w2 to obtain the back-translation score 2 of the source language translation data. Similarity algorithms include edit distance and cosine similarity.

[0091] In yet other implementations, when the score includes a match score with a preset source language terminology library, determining the match score between the source language translation data and the preset source language terminology library includes: determining the number of matches between the source language translation data and terms in the source language terminology library; and determining the match score between the source language translation data and the preset source language terminology library based on the number of matches.

[0092] Continue reading Figure 6 First, it is necessary to build a source language terminology library, which specifically includes: extracting source language terms from existing high-quality source language text data to obtain a source language terminology library. Afterwards, the source language translation data output by the language translation model is matched with the words in the terminology library to obtain the number of matching terms, which is recorded as m. The matching score 3 between the source language translation data and the preset source language terminology library can be expressed as the following formula (3): ; (3) Where, is a preset parameter in the interval (1, 1.5); m is the number of completely matched terms, which is an integer greater than or equal to 0; w3 is the weight coefficient corresponding to the matching score.

[0093] S502: Determine a comprehensive score of the source language translation data based on at least one of the language fluency score, the back-translation quality score, and the matching score.

[0094] Continue reading Figure 6 Step S502 can be implemented in a variety of ways, as follows: In some implementations, any one of the language fluency score, the back-translation quality score, and the matching score can be used as a comprehensive score for the source language translation data.

[0095] In other implementations, the comprehensive score of the source language translation data may be determined as the cumulative sum of any two of the language fluency score, the back-translation quality score, and the matching score.

[0096] In some other implementations, the language fluency score, the back-translation quality score, and the matching score may be accumulated to determine the comprehensive score of the source language translation data.

[0097] Continue reading Figure 5 After step S502, the method may further include step S503: S503: If the comprehensive score is greater than or equal to the preset score threshold, it is determined that the source language translation data meets the preset screening rule, and the source language translation data is determined as an optimized training sample.

[0098] After determining the comprehensive score, high-quality source language translation data can be screened out as optimized training samples based on the comparison results between the comprehensive score and the preset score threshold.

[0099] Continue reading Figure 6 By judging whether the comprehensive score is greater than or equal to the preset score threshold, it is determined whether the source language translation data is high-quality translation data. If so, the source language translation data is determined as an optimized training sample; otherwise, it is discarded.

[0100] Continue reading Figure 4 , after step S402, step S403 may also be included.

[0101] S403: Optimize and train the language translation model based on the optimized training samples to obtain an optimized language translation model.

[0102] After obtaining the optimal training samples, the optimized training samples can be input into the language translation model for iterative training to continuously optimize the model performance and continuously improve its translation quality, especially for low-resource languages.

[0103] In summary, the embodiments of the present application introduce linguistic knowledge to screen out languages similar to low-resource languages such as Uyghur and Tibetan from multiple dimensions such as subword overlap, phoneme similarity, and semantic relevance, while incorporating their dialect data into training. Furthermore, by combining low-resource languages such as Uyghur and Tibetan and their dialect data, as well as data from similar languages, a neural translation model is pre-trained and transferred to translation tasks for low-resource languages such as Uyghur and Tibetan, thereby effectively promoting positive transfer and improving the performance of the translation model.

[0104] Exemplary devices Corresponding to the above-mentioned language translation processing method, an embodiment of the present application also provides a language translation processing device. Figure 7 This is a structural diagram of a language translation processing device provided in an embodiment of the present application. Figure 7As shown, the language translation processing device provided in the embodiment of the present application includes: an acquisition unit 701 and a translation processing unit 702; wherein the acquisition unit 701 is used to acquire source language data to be translated; the translation processing unit 702 is used to translate the source language data into target language data through a language translation model to obtain a target language translation result; wherein the language translation model is trained by training samples of the source language and its similar languages, and the similar languages are languages that meet corresponding similarity requirements with the source language in various language feature dimensions, and the various language feature dimensions include at least one of a subword dimension, a phoneme dimension, and a semantic dimension.

[0105] In some embodiments, the language translation model is trained using a training unit 703, and the training unit 703 is used to perform the following steps: determining similar languages to the source language; generating a training sample set based on training samples corresponding to the source language and training samples corresponding to the similar languages; pre-training a preset neural network architecture based on the training sample set to obtain a pre-trained translation model; and performing transfer learning training on the pre-trained translation model based on the training samples corresponding to the source language to obtain the language translation model.

[0106] In some embodiments, the training unit 703 determines similar languages of the source language, including: determining various related languages of the source language; determining, for each of the various language feature dimensions, the similarity between the source language and the various related languages in the language feature dimension; determining the related languages whose similarity with the source language in the language feature dimension is greater than a preset similarity threshold as similar languages of the source language in the language feature dimension; and determining the union of the similar languages of the source language in each language feature dimension as the similar languages of the source language.

[0107] In some embodiments, when the language feature dimension includes the subword dimension, the training unit 703 determines the similarity between the source language and the related languages in the language feature dimension, including: determining the overlap of subwords in the subword dictionary of the source language and the subword dictionaries corresponding to the related languages, to obtain the similarity between the source language and the related languages in the subword dimension; and / or, when the language feature dimension includes the phoneme dimension, the training unit 703 determines the similarity between the source language and the related languages in the language feature dimension, including : determining the similarity between the phonemes in the phoneme dictionary of the source language and the phoneme dictionaries corresponding to the respective related languages, and obtaining the similarity between the source language and the respective related languages in the phoneme dimension; and / or, when the language feature dimension includes the semantic dimension, the training unit 703 determines the similarity between the source language and the respective related languages in the language feature dimension, including: determining the similarity between the training samples of the source language and the source language translation results corresponding to the respective related languages, and obtaining the similarity between the source language and the respective related languages in the semantic dimension.

[0108] In some embodiments, the training unit 703 is further configured to perform the following steps: inputting the training samples of the similar language into the language translation model to obtain source language translation data; determining the source language translation data that meets preset scoring requirements as optimized training samples; and optimizing the language translation model based on the optimized training samples to obtain an optimized language translation model.

[0109] In some embodiments, the training unit 703 determines source language translation data that meets preset scoring requirements as optimized training samples, including: determining a score for the source language translation data, the score including at least one of a language fluency score, a back-translation quality score, and a match score with a preset source language terminology database; determining a comprehensive score for the source language translation data based on at least one of the language fluency score, the back-translation quality score, and the match score; and if the comprehensive score is greater than or equal to a preset scoring threshold, determining that the source language translation data meets the preset scoring requirement, and determining the source language translation data as an optimized training sample.

[0110] In some embodiments, when the score includes the language fluency score, the training unit 703 determines the language fluency score of the source language translation data, including: dividing the source language translation data into n-grams; determining the conditional probability of each n-gram through a language model; and determining the fluency score of the source language translation data based on a cumulative multiplication result of the conditional probabilities of each n-gram and a weight coefficient corresponding to the language fluency score.

[0111] In some embodiments, when the score includes the back-translation quality score, the training unit 703 determines the back-translation quality score of the source language translation data, including: translating the source language translation data into translation data in respective similar languages using the language translation model, and back-translating the translation data in respective similar languages into the source language to obtain respective source language back-translation data; and obtaining the back-translation score of the source language translation data based on a weighted sum of similarities between each respective source language back-translation data and the source language translation data and a weight coefficient corresponding to the back-translation quality score.

[0112] The language translation processing device provided in this embodiment is based on the same concept as the language translation processing method provided in the above-mentioned embodiments of this application. It can execute the language translation processing method provided in any of the above-mentioned embodiments of this application and has the corresponding functional modules and beneficial effects of executing the language translation processing method. For technical details not fully described in this embodiment, please refer to the specific processing content of the language translation processing method provided in the above-mentioned embodiments of this application and will not be repeated here.

[0113] The functions implemented by the above acquisition unit 701, translation processing unit 702 and training unit 703 can be implemented by the same or different processors respectively, and this embodiment of the application does not limit this.

[0114] It should be understood that the units in the above devices can be implemented in the form of a processor calling software. For example, the device includes a processor, the processor is connected to a memory, and the memory stores instructions. The processor calls the instructions stored in the memory to implement any of the above methods or realize the functions of each unit of the device. The processor can be a general-purpose processor, such as a CPU or a microprocessor, and the memory can be a memory within the device or a memory outside the device. Alternatively, the units in the device can be implemented in the form of hardware circuits. The functions of some or all units can be realized by designing the hardware circuits. The hardware circuit can be understood as one or more processors. For example, in one implementation, the hardware circuit is an ASIC, and the functions of some or all of the above units can be realized by designing the logical relationships between the components within the circuit. For another example, in another implementation, the hardware circuit can be implemented by a PLD. For example, an FPGA can include a large number of logic gate circuits. The connection relationships between the logic gate circuits are configured through a configuration file to realize the functions of some or all of the above units. All units of the above devices can be implemented entirely in the form of a processor calling software, or entirely in the form of hardware circuits, or partially in the form of a processor calling software, with the remaining parts implemented in the form of hardware circuits.

[0115] In an embodiment of the present application, a processor is a circuit with the ability to process signals. In one implementation, the processor may be a circuit with the ability to read and execute instructions, such as a CPU, a microprocessor, a GPU, or a DSP. In another implementation, the processor may implement certain functions through the logical relationship of a hardware circuit, and the logical relationship of the hardware circuit may be fixed or reconfigurable, such as a hardware circuit implemented by an ASIC or PLD, such as an FPGA. In a reconfigurable hardware circuit, the process of the processor loading a configuration document to implement the configuration of the hardware circuit can be understood as the process of the processor loading instructions to implement the functions of some or all of the above units. In addition, it can also be a hardware circuit designed for artificial intelligence, which can be understood as an ASIC, such as an NPU, TPU, DPU, etc.

[0116] It can be seen that each unit in the above device can be one or more processors (or processing circuits) configured to implement the above method, such as: CPU, GPU, NPU, TPU, DPU, microprocessor, DSP, ASIC, FPGA, or a combination of at least two of these processor forms.

[0117] In addition, the various units in the above apparatus may be fully or partially integrated together, or may be implemented independently. In one implementation, these units are integrated together and implemented in the form of a system-on-chip (SOC). The SOC may include at least one processor for implementing any of the above methods or implementing the functions of the various units of the apparatus. The at least one processor may be of different types, such as a CPU and an FPGA, a CPU and an artificial intelligence processor, a CPU and a GPU, etc.

[0118] Exemplary electronic devices The present application embodiment provides an electronic device, see Figure 8 As shown, the electronic device includes: Memory 200 and processor 210; The memory 200 is connected to the processor 210 and is used to store programs; The processor 210 is configured to implement the language translation processing method disclosed in any of the above embodiments by running the program stored in the memory 200 .

[0119] Specifically, the electronic device may further include: a bus, a communication interface 220 , an input device 230 and an output device 240 .

[0120] The processor 210, the memory 200, the communication interface 220, the input device 230 and the output device 240 are interconnected via a bus. A bus may include a pathway that transfers information between components of a computer system.

[0121] Processor 210 can be a general-purpose processor, such as a general-purpose central processing unit (CPU), a microprocessor, or the like. It can also be an application-specific integrated circuit (ASIC), or one or more integrated circuits used to control the execution of the program of the present invention. It can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic device, discrete gate or transistor logic device, or discrete hardware components.

[0122] The processor 210 may include a main processor, and may also include a baseband chip, a modem, and the like.

[0123] Memory 200 stores programs that implement the technical solutions of the present invention and may also store an operating system and other key services. Specifically, the programs may include program code, which includes computer operating instructions. More specifically, memory 200 may include read-only memory (ROM), other types of static storage devices capable of storing static information and instructions, random access memory (RAM), other types of dynamic storage devices capable of storing information and instructions, disk storage, flash memory, and the like.

[0124] The input device 230 may include a device for receiving data and information input by a user, such as a keyboard, a mouse, a camera, a scanner, a light pen, a voice input device, a touch screen, a pedometer, or a gravity sensor.

[0125] Output device 240 may include devices that allow information to be output to a user, such as a display screen, printer, speakers, etc.

[0126] The communication interface 220 may include any device such as a transceiver to communicate with other devices or communication networks, such as Ethernet, a radio access network (RAN), a wireless local area network (WLAN), etc.

[0127] The processor 210 executes the program stored in the memory 200 and calls other devices, and can be used to implement each step of any language translation processing method provided in the above embodiments of the present application.

[0128] An embodiment of the present application also proposes a chip, which includes a processor and a data interface. The processor reads and runs a program stored in a memory through the data interface to execute the language translation processing method introduced in any of the above embodiments. The specific processing process and its beneficial effects can be found in the embodiment introduction of the above-mentioned language translation processing method.

[0129] Exemplary computer program products and storage media In addition to the above-mentioned methods and devices, an embodiment of the present application may also be a computer program product, which includes computer program instructions. When the computer program instructions are executed by a processor, the processor executes the steps of the language translation processing method according to various embodiments of the present application described in any of the above embodiments of this specification.

[0130] The computer program product may be written in any combination of one or more programming languages to implement the program code for performing the operations of the embodiments of the present application, including object-oriented programming languages such as Java, C++, and conventional procedural programming languages such as "C" or similar programming languages. The program code may be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0131] In addition, an embodiment of the present application may also be a storage medium having a computer program stored thereon, and the computer program is used by a processor to execute the steps of the language translation processing method according to various embodiments of the present application described in any of the above embodiments of this specification, specifically the following steps: Step S101: Obtain facial video data of a target user.

[0132] Step S102: Input the facial video data into a language translation processing model to obtain a target cognitive state corresponding to the target user, wherein the language translation processing model is obtained by optimizing and training a pre-trained baseline language translation processing model based on manually annotated video samples, wherein the accuracy of the cognitive state prediction result of the baseline language translation processing model for the video sample is lower than a set accuracy threshold.

[0133] For the sake of simplicity, the aforementioned method embodiments are described as a series of action combinations. However, those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily required by this application.

[0134] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similarities between the various embodiments can be referred to in conjunction with each other. For device embodiments, since they are generally similar to method embodiments, their description is relatively simple, and for relevant details, reference can be made to the description of the method embodiments.

[0135] The steps in the methods of each embodiment of the present application can be adjusted in sequence, merged, and deleted according to actual needs, and the technical features recorded in each embodiment can be replaced or combined.

[0136] The modules and sub-modules in the devices and terminals of the various embodiments of the present application can be merged, divided, and deleted according to actual needs.

[0137] In the several embodiments provided in this application, it should be understood that the disclosed terminals, devices, and methods can be implemented in other ways. For example, the terminal embodiments described above are merely illustrative. For example, the division of modules or submodules is merely a logical function division. In actual implementation, there may be other division methods, such as multiple submodules or modules can be combined or integrated into another module, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, device or module, which can be electrical, mechanical or other forms.

[0138] The modules or submodules described as separate components may or may not be physically separate, and the components of the modules or submodules may or may not be physical modules or submodules, that is, they may be located in one place or distributed across multiple network modules or submodules. Some or all of the modules or submodules may be selected to achieve the purpose of this embodiment according to actual needs.

[0139] In addition, each functional module or submodule in each embodiment of the present application may be integrated into a processing module, or each module or submodule may exist physically separately, or two or more modules or submodules may be integrated into a single module. The above-mentioned integrated modules or submodules may be implemented in the form of hardware or software functional modules or submodules.

[0140] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0141] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, software executed by a processor, or a combination of the two. The software may be stored in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0142] Finally, it should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0143] The above description of the disclosed embodiments will enable those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is to be construed in the widest manner consistent with the principles and novel features disclosed herein.

Claims

1. A language translation processing method, characterized in that: include: Obtain source language data to be translated; Translating the source language data into target language data using a language translation model to obtain a target language translation result; The language translation model is obtained by training with training samples of the source language and its similar languages, wherein the similar languages are languages that meet corresponding similarity requirements with the source language in various language feature dimensions, and the various language feature dimensions include at least one of a subword dimension, a phoneme dimension, and a semantic dimension.

2. The method according to claim 1, characterized in that The language translation model is trained using the following steps: determining languages similar to the source language; generating a training sample set based on the training samples corresponding to the source language and the training samples corresponding to the similar language; Pre-training a preset neural network architecture according to the training sample set to obtain a pre-trained translation model; The pre-trained translation model is subjected to transfer learning training based on training samples corresponding to the source language to obtain the language translation model.

3. The method according to claim 2, characterized in that Determining a language similar to the source language includes: determining various related languages of the source language; For each of the language feature dimensions, determining the similarity between the source language and each of the related languages in the language feature dimension; Determine the related languages whose similarity with the source language in the language feature dimension is greater than a preset similarity threshold as similar languages of the source language in the language feature dimension; The union of similar languages of the source language in each language feature dimension is determined as the similar language of the source language.

4. The method according to claim 3, characterized in that In a case where the language feature dimension includes the subword dimension, determining the similarity between the source language and each of the related languages in the language feature dimension includes: Determining the overlap of subwords in the subword dictionary of the source language and the subword dictionaries corresponding to the respective related languages, and obtaining similarities between the source language and the respective related languages in the subword dimension; and / or, In a case where the language feature dimension includes the phoneme dimension, determining the similarity between the source language and each of the related languages in the language feature dimension includes: Determining the similarity between the phoneme dictionary of the source language and the phoneme dictionaries corresponding to the respective related languages, to obtain the similarity between the source language and the respective related languages in the phoneme dimension; and / or, In a case where the language feature dimension includes the semantic dimension, determining the similarity between the source language and each of the related languages in the language feature dimension includes: Determine the similarity between the training sample of the source language and the source language translation results corresponding to each of the related languages, and obtain the similarity between the source language and each of the related languages in the semantic dimension.

5. The method according to claim 2, characterized in that The method further comprises: Inputting the training samples of the similar language into the language translation model to obtain source language translation data; The source language translation data that meets the preset scoring requirements are determined as optimized training samples; The language translation model is optimized and trained based on the optimized training samples to obtain an optimized language translation model.

6. The method according to claim 5, characterized in that The source language translation data that meets the preset scoring requirements is determined as the optimized training samples, including: Determining a score for the source language translation data, the score comprising at least one of a language fluency score, a back-translation quality score, and a score for matching with a preset source language terminology database; determining a comprehensive score of the source language translation data based on at least one of the language fluency score, the back-translation quality score, and the matching score; If the comprehensive score is greater than or equal to a preset score threshold, it is determined that the source language translation data meets the preset score requirement, and the source language translation data is determined as an optimized training sample.

7. The method according to claim 6, characterized in that In a case where the score includes the language fluency score, determining the language fluency score of the source language translation data includes: Segmenting the source language translation data into n-grams; Determining the conditional probability of each n-gram using a language model; The fluency score of the source language translation data is determined based on the cumulative multiplication result of the conditional probabilities of the respective n-grams and the weight coefficient corresponding to the language fluency score.

8. The method according to claim 6, characterized in that In a case where the score includes the back-translation quality score, determining the back-translation quality score of the source language translation data comprises: Translating the source language translation data into translation data of each similar language using the language translation model, and back-translating the translation data of each similar language into the source language to obtain back-translated data of each source language; The back-translation score of the source language translation data is obtained based on the weighted sum of the similarities between each source language back-translation data and the source language translation data, and the weight coefficient corresponding to the back-translation quality score.

9. A language translation processing device, characterized in that: include: An acquisition unit, used to acquire source language data to be translated; A translation processing unit, configured to translate the source language data into target language data using a language translation model to obtain a target language translation result; The language translation model is obtained by training with training samples of the source language and its similar languages, wherein the similar languages are languages that meet corresponding similarity requirements with the source language in various language feature dimensions, and the various language feature dimensions include at least one of a subword dimension, a phoneme dimension, and a semantic dimension.

10. An electronic device, characterized in that: including memory and processor; The memory is connected to the processor and is used to store programs; The processor is configured to implement the method according to any one of claims 1 to 8 by running the program in the memory.

11. A computer program product, characterized in that The method comprises computer program instructions, which, when executed by a processor, cause the processor to implement the method according to any one of claims 1 to 8.

Images