16 results about "Back translation" patented technology

The utility model discloses a stop ruler swing angle, front-back translation and quick locking mechanism of a sliding table saw, which comprises a material placing frame, a stop ruler arranged on the material placing frame, a front-back translation device, a swing angle device, a left-right sliding device and a quick locking device, the front-back translation device comprises two translation guide rails fixed to the discharging frame, a translation sliding block installed on the two translation guide rails in a sliding mode and a translation plate fixedly installed on the translation sliding block. Two of the four quick locking devices are located at the two ends of the front end of the discharging frame, and the other two quick locking devices are located at the two ends of the rear end of the discharging frame and used for locking the two fixing shafts located at the two ends of the bottom of the blocking ruler; the blocking ruler can translate front and back, slide left and right and rotate, can be quickly locked, is convenient to use, and can be suitable for processing plates with different specifications.

Taking a translation task from Naxi Dongba to Chinese as an example, the invention provides a weak semantic low-resource character machine translation method based on semantic enhancement, which comprises the following steps: S1, designing a Naxi Dongba encoding system, and establishing a Naxi Dongba electronic dictionary; s2, a sufficient number of Naxi Dongba text-Chinese parallel sentence pairs are collected and marked, and a Naxi Dongba text-Chinese parallel corpus is constructed; s3, dividing the data set into a fine adjustment data set and a test data set, and further dividing the fine adjustment data set into a training set and a verification set; s4, constructing a semantic enhancement model based on fine tuning and custom word list embedding; s5, providing an iterative reverse translation method combined with word replacement, and constructing an extended data set; s6, constructing a weak semantic low-resource text machine translation model based on semantic enhancement, adopting an increment updating mechanism, taking the high-quality pseudo-parallel corpus generated in the step S5 as increment, inputting the increment into the semantic enhancement model in the step S3, and adjusting and optimizing the weight of the model through parameters; and S7, inputting the Naxi Dongba coded sentences to be translated into the updated model for translation, and outputting a result. According to the method, translation research from the Naxi Dongba text to Chinese is carried out based on traditional expert experience, automatic translation of the Naxi Dongba text can be achieved, meanwhile, the method has the capacity of continuous learning and adapting to new data, the machine translation effect of weak-semantic low-resource characters is improved, and technical support is provided for research in related fields.

The invention relates to a data enhancement method based on a two-way translation adversarial network, and belongs to the field of natural language processing. In low-resource language machine translation, the problem that a machine translation result deviates in a specific field due to the lack of large-scale diversified training corpora is solved. In order to relieve the influence caused by deviation in a specific field, the invention provides the method, and the method combines adversarial training and a bidirectional back translation technology, and generates a high-quality adversarial sample and diversified training data by using a pre-trained shielded language model so as to enhance the robustness of the model. Firstly, adversarial training is used for carrying out data enhancement on a data set, then a bidirectional translation method is used for generating texts with similar semantics but different expressions, then a pre-training shielding language model is used for generating text variants with reasonable semantics, finally, bidirectional translation and MLM are fused to generate more reasonable and diversified enhanced data, and the data are used for retraining the model. According to the method, the robustness and the performance of the NMT model are remarkably enhanced on the data level.

The invention relates to a training method of an unsupervised language translation model and a language translation method and device. The implementation scheme is as follows: performing grammatical analysis and coding processing on unsupervised training corpora to obtain a grammatical feature vector; respectively inputting the training corpora into the corresponding monolingual word embedding layer and the shared word embedding layer to obtain monolingual word embedding and shared word embedding, and generating semantic vectors according to the monolingual word embedding and the shared word embedding; the shared word embedding layer is constructed based on a mixed corpus containing a source language and a target language; performing fusion processing on the grammar feature vector and the semantic vector to obtain a fusion vector of the training corpus; training an unsupervised pre-training model by using the fusion vector; and initializing an encoder of the language translation model by adopting a weight parameter of the trained unsupervised pre-training model, and performing unsupervised training on the language translation model by alternately executing a de-noising automatic encoder and a reverse translation task. According to the invention, the accuracy, fluency and robustness of translation can be improved.

The application discloses a machine translation quality estimation method based on multi-semantic space, which is divided into machine translation quality estimation model training and machine translation quality estimation. The model training steps are as follows: generating pseudo reference translation and back translation by using a dialogue large language model; extracting machine translation quality features in the source language semantic space; extracting machine translation quality features in the target language semantic space; extracting machine translation quality features in the cross-language semantic space; extracting a machine translation quality feature vector in the multi-semantic space; predicting a machine translation quality score based on the multi-semantic space; and training a machine translation quality estimation model based on the multi-semantic space. The machine translation quality estimation method steps are as follows: normalizing a source language sentence and machine translation to be quality estimated; inputting the normalized source language sentence and machine translation into the trained machine translation quality estimation model based on the multi-semantic space, and predicting a machine translation quality score.

A translation device (1) is configured so as to comprise: an original text acquisition unit (11) that acquires an original text to be translated; a first translation unit (12) that causes the original text acquired by the original text acquisition unit (11) to be translated into a first language; a back translation unit (13) that causes the translated text by the first translation unit (12) to be translated back; a display processing unit (16) that causes a display device (3) to display the translated-back text by the back translation unit (13); and a correction reception unit (14) that receives correction of the original text after the translated-back text is displayed by the display processing unit (16). The translation device (1) also comprises a second translation unit (15) that, when the correction reception unit (14) receives the correction of the original text, causes the corrected original text received by the correction reception unit (14) to be translated into a second language.

The application provides a multilingual large language model content security defense method, and belongs to the technical field of artificial intelligence security, which can at least partially solve the problems of lack of multilingual evaluation benchmark, easy general ability decline caused by full fine-tuning, and lack of pertinence in defense in the prior art. The application comprises: constructing a multilingual jailbreak evaluation data set with semantic retention, screening the adversarial samples through double checking of back translation similarity and cross-lingual consistency; positioning the security layer by layer pruning of the target large language model; calculating the security score by decoding the hidden state of each layer to the vocabulary space to identify the toxicity layer; freezing the parameters of the non-security layer and the non-toxicity layer, and only updating the weight of the security layer and the auxiliary layer to suppress the toxic response by taking the hidden state of the toxicity layer as the constraint. The application reduces the success rate of multilingual jailbreak attacks while avoiding the general ability decline caused by full fine-tuning.

The application provides a question generation model training method and device, wherein the question generation model training method comprises the following steps: obtaining a triple in a target knowledge base; creating an initial question template according to the triple, and performing back translation processing on the initial question template to obtain an extended question template; determining a mapping relationship between the triple and the initial question template and the extended question template based on a relationship contained in the triple; constructing a sample set based on the mapping relationship, and training a question generation model through the sample set until a target question generation model meeting a training stop condition is obtained.

The invention discloses a translation model training method, a text translation method and a related device, and relates to the technical field of translation, and the method comprises the following steps: carrying out professional term marking on each source language text in a source language text set to obtain a term marking text corresponding to each source language text, and carrying out back translation on the term marking text by utilizing an initial translation model to obtain a translation model; the method comprises the steps of obtaining a back-translated text marked with back-translated terms, determining term back-translating accuracy of the back-translated terms in the back-translated text relative to professional terms, determining a first source language text enabling the term back-translating accuracy to be smaller than an accuracy threshold value from a source language text set, obtaining a standard translated text of the first source language text, and obtaining the standard translated text of the first source language text. And forming a training set with the first source language text to perform fine tuning training on the initial translation model to obtain a target translation model. According to the method, the technical terms with non-ideal translation effects are accurately positioned based on the term back translation accuracy, and then the technical terms are subjected to targeted training, so that the translation accuracy of the model is improved.

The application discloses a kind of Chinese-tibetan bidirectional machine translation method, device, equipment and medium for low-resource scene, it is related to machine translation technical field.The method comprises: adding multilingual embedding layer on the basis of multilingual pre-training model, constructs Chinese-tibetan bidirectional translation fine-tuning starting point;Through dynamic adjustment decoder position coding, intermediate pre-training of mask sequence reconstruction is executed on monolingual corpus;Single-way and bidirectional translation model are constructed, pseudo-parallel data is generated by back translation, and the model is iteratively trained by mixed with real parallel corpus;In inference stage, use hidden state to construct inverted index to retrieve similar context Token, interpolation fusion is carried out between retrieval probability distribution and decoder prediction probability distribution, and the final translation result is output.The application effectively overcomes the semantic drift problem in cross-language training, maximizes the corpus utilization efficiency under low-resource condition, and significantly improves the translation accuracy of rare words and complex long sentences.

The present disclosure relates to a data generation method, device and readable storage medium, the method comprising: performing noise adding processing on initial first language text to obtain noise-added first language text; processing the noise-added first language text according to a pre-trained language model to obtain target first language text; performing back translation processing on the target first language text to obtain second language text; and obtaining training data for training a translation model based on the target first language text and the second language text. The method of the present disclosure can improve the diversity of the training data for training the translation model and solve the problem of training data shortage.

The design is a voice input image, and as shown in the reference diagram showing the usage state, pressing the voice input button allows for voice input of the native language into the native language input section. Below the input section, multiple language names corresponding to the language to be translated are displayed, and switching between languages ​​is possible by selecting the desired language. Using a "back translation function," which translates content translated into another language back into the native language, the back translation of the selected language is displayed in the back translation display section. After comparing the content of the back translation display section with the content of the native language input section, if the user wishes to correct the content in the native language input section, for example, the user can press the "discard button" to reset the display of the native language input section and then press the voice input button again to perform voice input. On the other hand, if the user does not need to correct the native language input section and wishes to send the content in the native language to another medium, the user can press the "send button" to send the content.

A computer-implemented method is disclosed, comprising: receiving, by one or more large language models, LLMs, one or more natural language system prompts which, when processed, causes the one or more LLMs to perform a system method. The system method comprises: detecting the source language of a source text, wherein the source text is provided in natural language; determining, based on a user prompt, a target language; generating, a first translation by translating the source text from the source language to the target language; generating, a back translation by translating the first translation from the target language to the source language; comparing, the back translation with the source text to determine one or more inconsistencies; determining whether the one or more inconsistencies exceed an error threshold; based on the one or more inconsistencies exceeding the error threshold, generating, based on the one or more inconsistencies, a second translation by translating the source text from the source language to the target language; and based on the one or more inconsistencies not exceeding the error threshold, outputting, to a user, the first translation; processing, by the one or more LLMs, the one or more system prompts; receiving, by the one or more LLMs, via a user device, a user prompt comprising the source text; and providing, by the one or more LLMs, using the system method, the first translation or the second translation.

This invention discloses a sliding table saw, including a feeding rack, a stopper mounted on the feeding rack, and further including a front-to-back translation device, a swing angle device, a left-to-right sliding device, and a quick-locking device. The front-to-back translation device includes two translation guide rails fixed on the feeding rack, a translation slider slidably mounted on the two translation guide rails, and a translation plate fixedly mounted on the translation slider. There are four quick-locking devices, two of which are located at the front ends of the feeding rack and the other two are located at the rear ends of the feeding rack, used to lock two fixed shafts located at the bottom ends of the stopper. The stopper of this invention can be translated front-to-back, slid left-to-right, and rotated, and can also be quickly locked. It is convenient to use and suitable for processing boards of different specifications.

Provided are a multi-modal model training method, apparatus and device, and a storage medium. The method includes the following steps: obtaining a training sample set, and training a multi-modal model for a plurality of rounds by successively using each of training sample pair in the training sample set: during use of any one of the training sample pairs for training, obtaining an image feature of a target visual sample firstly, and then determining whether back translation needs to be performed on a target original text; when back translation needs to be performed on the target original text, performing corresponding back translation to obtain a target back-translated text, and obtaining a text feature of the target back-translated text; and training the multi-modal model based on the image feature and the text feature.

Embodiments of the present application disclose a parallel corpus data pair construction method and device and a storage medium. The method comprises: inputting first corpus data with first style characteristics into a first style conversion model to obtain second corpus data with second style characteristics; the first style conversion model is obtained based on back translation training; inputting the first corpus data into a second style conversion model to obtain third corpus data with the second style characteristics; the second style conversion model is obtained based on adversarial training; calculating a first score corresponding to the second style characteristics of the second corpus data; calculating a second score corresponding to the second style characteristics of the third corpus data; if the first score is greater than the second score, constructing a first parallel corpus data pair using the second corpus data and the first corpus data; if the first score is not greater than the second score, constructing the first parallel corpus data pair using the third corpus data and the first corpus data, thereby solving the problem of scarcity of parallel corpus data pairs.