Language data processing method, device, storage medium and electronic device

Abstract

The present application discloses a method, device, storage medium and electronic device for processing language data. The method comprises: translating source language data into first target language data; performing code conversion processing on the source language data to obtain first code conversion data, and performing code conversion processing on the first target language data to obtain second code conversion data; training multiple models using the source language data, the first target language data, the first code conversion data and the second code conversion data; inputting the second target language data into the multiple models and outputting multiple first prediction results; training a target model using the multiple first prediction results; inputting the third target language data into the target model and outputting a second prediction result. The present application solves the technical problem in the prior art that the processing efficiency of sentiment analysis tasks in multiple target languages ​​is low due to the need for fine-grained annotation processing for aspect-level sentiment analysis.

Description

Technical Field

[0001] The present application relates to the field of computer data processing, and more specifically, to a method, device, storage medium, and electronic device for processing language data. Background Art

[0002] Currently, aspect-based sentiment analysis is the most widely used sentiment analysis task in scenarios such as mining sentiment trends in user product reviews. In real-world business environments, sentiment analysis often involves multiple target languages. For example, product review mining algorithms for cross-border e-commerce platforms like AE and Lazada require support for dozens of target languages.

[0003] To extract user sentiment from product reviews, existing techniques typically label data for different target languages ​​to serve as model training data. However, aspect-level sentiment analysis requires fine-grained labeling, so labeling for each target language individually consumes significant time, money, and effort, hindering the rapid deployment of analysis algorithms.

[0004] To address the above-mentioned problems, no effective solutions have been proposed so far. Summary of the Invention

[0005] The embodiments of the present application provide a method, device, storage medium and electronic device for processing language data, so as to at least solve the technical problem in the prior art that aspect-level sentiment analysis requires fine-grained annotation processing, resulting in low processing efficiency of sentiment analysis tasks in multiple target languages, which is not conducive to mining user emotional tendencies.

[0006] According to one aspect of an embodiment of the present application, a method for processing language data is provided, comprising: translating source language data into first target language data, wherein the source language data is pre-labeled data; performing code conversion on the source language data to obtain first code-conversion data, and performing code conversion on the first target language data to obtain second code-conversion data; training multiple models using the source language data, the first target language data, the first code-conversion data, and the second code-conversion data; inputting second target language data into the multiple models and outputting multiple first prediction results, wherein the second target language data is pre-collected unlabeled data; training a target model using the multiple first prediction results; and inputting third target language data into the target model and outputting a second prediction result, wherein the third target language data is unlabeled data to be analyzed.

[0007] According to another aspect of an embodiment of the present application, a language data processing device is provided, comprising: a translation module for translating source language data into first target language data, wherein the source language data is pre-labeled data; a conversion processing module for performing code conversion processing on the source language data to obtain first code-conversion data, and performing code conversion processing on the first target language data to obtain second code-conversion data; a first training module for training multiple models using the source language data, the first target language data, the first code-conversion data, and the second code-conversion data; a first input module for inputting second target language data into the multiple models and outputting multiple first prediction results, wherein the second target language data is pre-collected unlabeled data; a second training module for training a target model using the multiple first prediction results; and a second input module for inputting third target language data into the target model and outputting a second prediction result, wherein the third target language data is unlabeled data to be analyzed.

[0008] According to another aspect of an embodiment of the present application, a computer-readable storage medium is further provided, wherein the computer-readable storage medium includes a stored program, wherein when the program is run, the device where the computer-readable storage medium is located is controlled to execute any one of the above-mentioned language data processing methods.

[0009] According to another aspect of an embodiment of the present application, an electronic device is provided, comprising: a processor; and a memory connected to the processor and configured to provide the processor with instructions for processing the following processing steps: translating source language data into first target language data, wherein the source language data is pre-labeled data; performing code conversion on the source language data to obtain first code-conversion data, and performing code conversion on the first target language data to obtain second code-conversion data; training multiple models using the source language data, the first target language data, the first code-conversion data, and the second code-conversion data; inputting second target language data into the multiple models and outputting multiple first prediction results, wherein the second target language data is pre-collected unlabeled data; training a target model using the multiple first prediction results; and inputting third target language data into the target model and outputting a second prediction result, wherein the third target language data is unlabeled data to be analyzed.

[0010] In an embodiment of the present application, a cross-language migration method that does not require target language annotated data is adopted, wherein source language data is translated into first target language data, wherein the above-mentioned source language data is pre-annotated data; code conversion processing is performed on the above-mentioned source language data to obtain first code-conversion data, and code conversion processing is performed on the above-mentioned first target language data to obtain second code-conversion data; multiple models are trained using the above-mentioned source language data, the above-mentioned first target language data, the above-mentioned first code-conversion data, and the above-mentioned second code-conversion data; second target language data is input into the above-mentioned multiple models, and multiple first prediction results are output, wherein the above-mentioned second target language data is pre-collected unlabeled data; a target model is trained using the above-mentioned multiple first prediction results; third target language data is input into the above-mentioned target model, and a second prediction result is output, wherein the above-mentioned third target language data is unlabeled data to be analyzed.

[0011] It is easy to notice that the embodiment of the present application uses a translation engine to translate the given annotated source language data to obtain the target language data; by annotating the aspect words with special symbols in the source language data, the aspect words and their corresponding emotional tendencies in the translation data are obtained according to the correspondence between the special symbols in the translation data, thereby mining the aspect-level opinions in the product reviews written by users in the target language. Therefore, the embodiment of the present application achieves the purpose of obtaining high-quality target language annotated data by a label mapping method without alignment, and based on the code conversion of the aspect words and the unlabeled target language data, the cross-language word vectors are further aligned, thereby achieving the technical effect of cross-language aspect-level sentiment analysis, thereby solving the technical problem in the prior art that the aspect-level sentiment analysis requires fine-grained annotation processing, resulting in low processing efficiency of sentiment analysis tasks in multiple target languages, which is not conducive to mining user emotional tendencies. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0013] Figure 1 This is a hardware structure block diagram of a computer terminal (or mobile device) for implementing a method for processing language data according to an embodiment of the present application;

[0014] Figure 2 is a schematic diagram of a method for processing language data according to an embodiment of the present application;

[0015] Figure 3 This is a schematic diagram of an application scenario of a method for processing language data according to an embodiment of the present application;

[0016] Figure 4 This is a schematic diagram of an application scenario of a method for processing language data according to an embodiment of the present application;

[0017] Figure 5 is a structural diagram of a language data processing device according to an embodiment of the present application;

[0018] Figure 6 It is a structural block diagram of another computer terminal according to an embodiment of the present application. DETAILED DESCRIPTION

[0019] In order to enable those skilled in the art to better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0020] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in a sequence other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0021] First, some nouns or terms that appear in the description of the embodiments of the present application are subject to the following interpretations:

[0022] Aspect: used to represent a characteristic of an object.

[0023] Aspect-based sentiment analysis: refers to analyzing the emotional tendencies towards specific aspects.

[0024] Code-switching: Code-switching is the process of switching words from one language to another.

[0025] Word embedding: refers to the vectorized representation of words.

[0026] To address the aforementioned annotation processing issues for cross-lingual sentiment analysis tasks, a common traditional approach is to use translation tools to translate source language data into target language data, and then map the labels of the source language data to the translated data. However, this mapping mostly relies on word-level or phrase-level mapping tools. Due to the accuracy of the mapping tools themselves, the quality of the labels obtained for the target language data is often poor, which limits the performance of the model. Another common approach is to use cross-lingual word vectors: first, cross-lingual word vectors are trained on large-scale bilingual parallel corpora, and then the model is trained using the annotated source language data. When implementing the model on the target language, it is only necessary to switch the input word vectors from the source language to the target language without changing the model architecture. However, due to the different characteristics of different languages, this approach often ignores the task-related knowledge in each language, and it is still difficult to obtain high-quality target language annotated data.

[0027] Example 1

[0028] According to an embodiment of the present application, an embodiment of a method for processing language data is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0029] The method embodiment provided in Example 1 of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Figure 1 FIG1 shows a hardware structure block diagram of a computer terminal (or mobile device) for implementing a method for processing language data. Figure 1 As shown, the computer terminal 10 (or mobile device 10) may include one or more (illustrated as 102a, 102b, ..., 102n) processors 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, and a transmission module 106 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the BUS bus), a network interface, a power supply and / or a camera. It will be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 1 More or fewer components than shown, or with Figure 1 Different configurations shown.

[0030] It should be noted that the one or more processors 102 and / or other data processing circuits described above may generally be referred to herein as "data processing circuitry". The data processing circuitry may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuitry may be a single independent processing module, or may be incorporated in whole or in part into any of the other components of the computer terminal 10 (or mobile device). As described in the embodiments of the present application, the data processing circuitry serves as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0031] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the language data processing method in the embodiment of the present application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, implementing the above-mentioned language data processing method. The memory 104 may include a high-speed random access memory and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory remotely located relative to the processor 102, and these remote memories may be connected to the computer terminal 10 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0032] The transmission device 106 is configured to receive or transmit data via a network. A specific example of the aforementioned network may include a wireless network provided by the communications provider of the computer terminal 10. In one embodiment, the transmission device 106 includes a network interface controller (NIC), which can be connected to other network devices via a base station to enable communication with the Internet. In another embodiment, the transmission device 106 may be a radio frequency (RF) module, which is configured to communicate with the Internet wirelessly.

[0033] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the computer terminal 10 (or mobile device).

[0034] It should be noted that, in some optional embodiments, the above Figure 1 The computer device (or mobile device) shown may include hardware elements (including circuits), software elements (including computer code stored on a computer-readable medium), or a combination of hardware elements and software elements. Figure 1This is merely one example of a particular embodiment and is intended to illustrate the types of components that may be present in the aforementioned computer device (or mobile device).

[0035] Under the above operating environment, this application provides Figure 2 A method for processing language data is shown. Figure 2 is a flow chart of a method for processing language data according to an embodiment of the present application. Figure 2 As shown, the above method can be implemented by the following steps:

[0036] Step S202: translating the source language data into first target language data, wherein the source language data is pre-annotated data;

[0037] Step S204: performing code conversion on the source language data to obtain first code-converted data, and performing code conversion on the first target language data to obtain second code-converted data;

[0038] Step S206, training multiple models using the source language data, the first target language data, the first code-conversion data, and the second code-conversion data;

[0039] Step S208: inputting the second target language data into the plurality of models, and outputting a plurality of first prediction results, wherein the second target language data is pre-collected unlabeled data;

[0040] Step S210, training a target model using the multiple first prediction results;

[0041] Step S212: input the third target language data into the target model and output a second prediction result, wherein the third target language data is unlabeled data to be analyzed.

[0042] In an optional embodiment, the above-mentioned language data processing method is applied to one of the following scenarios: multilingual product review mining scenario, multilingual named entity recognition scenario, and multilingual relationship extraction scenario.

[0043] Taking the application scenario of the above-mentioned language data processing method as an example, the multilingual product review mining scenario, since e-commerce websites have become the preferred way for users to purchase products, there are massive user product reviews on e-commerce websites. In order to mine users' views and opinions on specific aspects from product reviews, the aspect-based sentiment analysis problem intends to mine the aspect words and their corresponding sentiment tendencies in a given review.

[0044] This application scheme addresses the cross-language data processing problem in aspect-level sentiment analysis existing in the prior art, and proposes a language data processing method that does not require target language labeled data. This method uses aspect word code conversion and unlabeled target language data to strengthen cross-language word vector alignment, so that even for product reviews written in the unlabeled target language, the mentioned aspect words and their corresponding emotional tendencies can be automatically mined. In addition, in addition to being widely used in multilingual product review mining scenarios, the cross-language data processing method proposed in this scheme can also be applied to other similar multilingual extraction task scenarios, such as the above-mentioned multilingual named entity recognition scenario, multilingual relationship extraction scenario, etc.

[0045] The cross-language transfer method proposed in the embodiment of this application can mine the aspect-level opinions in the product reviews written by users in the target language by transferring the knowledge in the annotated English data to the unannotated target language data. Figure 3 As shown, a translation engine is used to obtain target language data by giving annotated source language data, that is, source language data with special marks; by marking aspect words with special symbols in the source language, the aspect words in the translation data and their corresponding emotional tendencies are obtained according to the corresponding relationship of the special symbols in the translation data.

[0046] As an optional embodiment, the aspect words in the source sentence and the translated sentence are code-converted, that is, the aspect words in each sentence are replaced with another language, thereby constructing two new bilingual sentences. The source sentence and the translated sentence and the sentence after code conversion can be combined together as training data to train a cross-language aspect-level sentiment analysis model. Specifically, the embodiment of the present application utilizes a pre-trained multilingual model, including Multilingual BERT and XLM-Roberta as the main part of the model, and then adds a layer of word-level linear classification network to predict the label of each word. For the collected unlabeled target language data, the embodiment of the present application can use knowledge distillation to obtain the target language-related knowledge: the teacher model makes predictions on the unlabeled data and obtains the predicted probability distribution of each word (called soft label); the student model attempts to learn the probability distribution of each word.

[0047] It should be noted that the language data processing method provided in the embodiments of the present application is essentially to adopt a label mapping method that does not require alignment to obtain high-quality target language annotation data, and based on the code conversion of aspect words and unlabeled target language data, cross-language word vectors are further aligned, thereby realizing cross-language aspect-level sentiment analysis.

[0048] Taking English (i.e., language 1, language 2, language 3, ..., language 10) as the source language and French (language 2, language 2, language 3, ..., language 2, language 10) as the target language as an example, the embodiment of the present application can also use special marks (e.g., special symbols) to highlight aspect words. Figure 3 The source language sentence x in S (language 1, [language 2], language 3…{language 8}…language 10), which is translated into the target language sentence x by the translation system. T (Language word 1, [Language word 2], Language word 3...{Language word 8}...Language word 10). Since the source language sentence is specially marked with sentiment polarity in advance, the target language sentence will also have special marks, so that the aspect words can be recovered from it and linked with the aspect words of the source language to obtain the sentiment polarity (e.g., positive polarity and negative polarity). The corresponding relationship between aspect words and sentiment polarity is as follows: Figure 3 In the implementation of this application, as shown. Figure 3 As shown in , the ACS model can be used to further transform the source language sentence x S (language word 1, [language word 2], language word 3...{language word 8}...language word 10) performs word code conversion to obtain the conversion processing result x St (language word 1, [language word 2], language word 3...{language word 8}...language word 10); Similarly, you can also translate sentence x T (Language word 1, [Language word 2], Language word 3... {Language word 8}... Language word 10) performs word code conversion to obtain another conversion processing result x Ts (Language word 2 1, [Language word 1 2], Language word 2 3…{Language word 1 8}…Language word 2 10).

[0049] The method of the embodiment of the application can correctly match "Language 1 word 2" in English with "Language 2 word 2" in French, and match "Language 1 word 8" in English with "Language 2 word 8" in French, thereby providing correct labeled data in the target language. The high-quality target language labeled data establishes a competitive benchmark for cross-language ABSA tasks. To solve cross-language ABSA tasks, in order to obtain high-quality labeled data in the target language, the embodiment of the application designs an alignment-free label projection method, which establishes a strong translation-based baseline, proposes an aspect word code conversion strategy to enhance cross-language alignment, and extracts the method proposed in the embodiment of the application on unlabeled target data to utilize rich language-specific knowledge.

[0050] In an optional embodiment, performing code conversion on the source language data to obtain the first code conversion data, and performing code conversion on the first target language data to obtain the second code conversion data include:

[0051] Step S302: Using preset symbols to mark aspect words in the source language data, and restoring target language words corresponding to the aspect words in the first target language data, wherein the aspect words and the target language words are used to represent target characteristics of the target object;

[0052] Step S304 , converting the aspect words marked in the source language data into the target language words to obtain the first code-converted data, and converting the target language words restored in the first target language data into the aspect words to obtain the second code-converted data.

[0053] As an optional embodiment, the embodiment of the present application can first annotate each aspect word with a special symbol and then input it into the translation system. For example, if there are multiple aspect words in a sentence, the embodiment of the present application will use a predefined list of special symbols to annotate them in order. After translation, the embodiment of the present application extracts the target language words with special symbols and matches them with the corresponding aspect words in the source sentence to project the emotional label. By doing so, the embodiment of the present application can reconstruct a label sequence representing the sequence boundaries and their corresponding emotional polarity for the translated data.

[0054] As an optional embodiment, since explicitly mixing data from different languages ​​has been proven to be an effective strategy to enhance cross-language consistency with multilingual pre-trained models. Based on this observation, the embodiment of the present application performs fine-grained aspect code conversion (ACS) on the source language data and the translated target language data, that is, code conversion processing is performed on the above-mentioned source language data to obtain first code-converted data, and code conversion processing is performed on the above-mentioned first target language data to obtain second code-converted data; to enhance cross-language consistency with task-specific knowledge.

[0055] Still Figure 3 As shown, in an optional embodiment, by giving a source sentence x S and its translation sentence x T , the embodiment of the present application switches the aspect words in the two sentences to construct two bilingual sentences. The first is the source sentence with aspect words in the target language, represented by x St , the second one is the translated sentence with aspect words in the source language, denoted as x TsAfter code switching, the embodiment of the present application uses the updated lengths of the two sentences after switching to refine the corresponding label sequence y St and y Ts .

[0056] The present application examples show that there are some interesting relationships between these data. For example, Ts and x T have the same sentence context, but the aspect words are expressed in different languages; while x T and x St With different sentence contexts but the same aspect words in the target language, by training on a combination of original and code-swapped data, we can better align the embedding spaces of different languages ​​with task-specific knowledge by using different aspects as alignment anchors.

[0057] In addition, assuming that multiple translation systems are available, the embodiment of the present application can be expanded to a multi-language setting method. For the source sentence x S , the embodiment of the present application can obtain multiple translation sentences x T1 ;x T2 ;…;x Tn n target languages ​​are used. Similarly, their aspect words can be switched with the source sentences, resulting in code-switched sentences x T1s ;x T2s ;…;x Tns and x St1 ;x St2 ; …;Models trained on this data combination can then benefit from task-specific knowledge across multiple languages.

[0058] In an optional embodiment, training the multiple models using the source language data, the first target language data, the first code-conversion data, and the second code-conversion data includes:

[0059] Step S402: combining the source language data and the first target language data into first category training data;

[0060] Step S404, combining the first code-conversion data and the first target language data into second category training data;

[0061] Step S406, combining the second code-converted data and the first target language data into third category training data;

[0062] Step S408: Use the first category training data, the second category training data, and the third category training data to train and obtain a first model, a second model, and a third model, respectively.

[0063] like Figure 4 As shown, the embodiment of the present application designs a variety of knowledge distillation methods, namely single teacher distillation method, multi-teacher distillation method, and multi-language distillation method. Figure 4 In the left shadow box (D s 、D T 、D Ts 、D st etc.) means using specific data (D U 、D U1 、D U2 、D Un ) trained models, such as the ACS model and the MACS model. The rightmost box (stu) represents the student model whose parameters are to be updated. KD Represents the loss function, and the student model uses this loss function to update the parameters from the soft labels predicted by the teacher model on the left (such as the teacher model composed of w1, w2, and w3 in the multi-teacher distillation method). The same is true for other distillation methods.

[0064] The embodiment of the present application uses the source sentence, the translated sentence and the sentence after code conversion as the model trained out of training data as teacher model, then makes prediction on the target language data without labels, and trains a student model. Because a data set comprises the translation data DT of the target language with specific language knowledge, another data set shares the same sentence semantics (D1), the same context sentence (D2) and the same aspect word (D3) respectively. In order to make full use of the different features in these data sets, the embodiment of the present application utilizes different data combinations to carry out multi-teacher knowledge distillation, and the two types of data obtained by code conversion of the source language data are combined with the translated sentence respectively to obtain the training data of three categories, thereby training three teacher models respectively, i.e. the first model, the second model and the third model. These three teacher models are trained to obtain probability distribution on the target language without labels respectively, and then the embodiment of the present application combines them to train a student model.

[0065] It should be noted that although the annotation cost of labeled ABSA data is high, it is much easier to collect unlabeled data in the target language. This embodiment of the present application uses unlabeled target data to extract the model proposed in this embodiment to train a student model for prediction. As an optional embodiment, knowledge extraction has been applied to various NLP tasks. When multiple target language data are available, these data can be combined to train a multilingual teacher model, thereby obtaining a multilingual student model.

[0066] In an optional embodiment, inputting the second target language data into the multiple models and outputting the multiple first prediction results includes:

[0067] Step S502: input the second target language data into the first model, and output a first predicted probability distribution of each word in the second target language data by the first model;

[0068] Step S504: input the second target language data into the second model, and output a second predicted probability distribution of each word in the second target language data by the second model;

[0069] Step S506 : Input the second target language data into the third model, and output a third predicted probability distribution of each word in the second target language data by the third model.

[0070] In the above optional embodiment, the above first model, the above second model and the above third model all include: a linear classification network at the word level, and the above linear classification network is used to obtain the predicted probability distribution of each word.

[0071] In the above optional embodiment, the distribution inputs the above second target language data into the above first model, the second model and the second model, and then respectively outputs the first predicted probability distribution of the above first model for each word in the above second target language data; the second predicted probability distribution of the above second model for each word in the above second target language data; and the third predicted probability distribution of the above third model for each word in the above second target language data.

[0072] In an optional embodiment, using the plurality of first prediction results to train the target model includes:

[0073] Step S602: combining the first predicted probability distribution, the second predicted probability distribution, and the third predicted probability distribution into fourth category training data;

[0074] Step S604: Use the fourth category training data to train and obtain the target model.

[0075] Optionally, the target model, or student model, can take any form, as the present embodiment is intended only to make its predictions similar to soft labels. For simplicity, the present embodiment uses the same neural architecture as the teacher model to train the student model. The trained student model can be used to make predictions on unseen data in the target language.

[0076] The main results of the cross-lingual ABSA task with English as the source language are shown in this example. This example reports the average F1 score of five runs using different random seeds. In this cross-lingual setting, this example uses English as the source language and other languages ​​as targets. For unlabeled target data, following previous research, this example discards the labels of the training set for each target language and uses the original sentences as unlabeled data.

[0077] As an optional embodiment, the language-specific knowledge in the unlabeled data is appropriately combined with the task-specific knowledge in the teacher model. Since these teacher models capture different features, the soft labels predicted by them can better "teach" the student model, so multiple teacher models are better than a single teacher model. In addition to the cross-lingual results (i.e., from one source language to one target language), the embodiment of the present application also trains the target data for multilingual translation (i.e., MTL-AF) through a label projection method, establishing a very strong baseline that may have surpassed MTL-WS, which is the latest model of the prior art with a parameter warm-up mechanism. Similar to observations in cross-lingual transmission, the aspect code conversion method and the refinement of unlabeled data proposed in the embodiment of the present application further improve the adaptive performance.

[0078] In an optional embodiment, inputting the third target language data into the target model and outputting the second prediction result includes:

[0079] Step S702: When the third target language data is product review data described in the target language, the product review data is input into the target model, and the target aspect words in the product review data and the sentiment tendency of the target aspect words are output.

[0080] In the above optional embodiment, taking the above third target language data as product review data described in the target language as an example, the above product review data is input into the student model, and the student model outputs the target aspect words in the above product review data and the emotional tendency of the above target aspect words.

[0081] Different from the existing methods based on word or phrase alignment, the present application solution can obtain high-quality target language annotated data through a label mapping method that does not require alignment. In addition, the code conversion of aspect words and the distillation on unlabeled target language data can effectively capture the task-related and language-related knowledge therein, thereby improving cross-language prediction capabilities. In bilingual and multilingual cases, compared with the prediction scores of 55.69 and 61.89 achieved by the previous baseline model, the prediction scores of the present application solution using multi-teacher knowledge distillation can reach up to 63.73 and 65.33 respectively.

[0082] It should be noted that for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but 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 the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by this application.

[0083] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus the necessary general hardware platform, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a computer-readable storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the above-mentioned methods of each embodiment of the present application.

[0084] Example 2

[0085] According to an embodiment of the present application, a device for implementing the above-mentioned language data processing method is also provided. Figure 5 is a structural diagram of a language data processing device according to an embodiment of the present application. Figure 5 As shown, the apparatus includes: a translation module 500, a conversion processing module 502, a first training module 504, a first input module 506, a second training module 508, and a second input module 510, wherein:

[0086] The translation module 500 is configured to translate source language data into first target language data, wherein the source language data is pre-labeled data. The conversion processing module 502 is configured to perform code conversion on the source language data to obtain first code-conversion data, and to perform code conversion on the first target language data to obtain second code-conversion data. The first training module 504 is configured to train multiple models using the source language data, the first target language data, the first code-conversion data, and the second code-conversion data. The first input module 506 is configured to input second target language data into the multiple models and output multiple first prediction results, wherein the second target language data is pre-collected unlabeled data. The second training module 508 is configured to train a target model using the multiple first prediction results. The second input module 510 is configured to input third target language data into the target model and output a second prediction result, wherein the third target language data is unlabeled data to be analyzed.

[0087] In an embodiment of the present application, a cross-language migration method that does not require target language annotated data is adopted, wherein source language data is translated into first target language data, wherein the above-mentioned source language data is pre-annotated data; code conversion processing is performed on the above-mentioned source language data to obtain first code-conversion data, and code conversion processing is performed on the above-mentioned first target language data to obtain second code-conversion data; multiple models are trained using the above-mentioned source language data, the above-mentioned first target language data, the above-mentioned first code-conversion data, and the above-mentioned second code-conversion data; second target language data is input into the above-mentioned multiple models, and multiple first prediction results are output, wherein the above-mentioned second target language data is pre-collected unlabeled data; a target model is trained using the above-mentioned multiple first prediction results; third target language data is input into the above-mentioned target model, and a second prediction result is output, wherein the above-mentioned third target language data is unlabeled data to be analyzed.

[0088] It is easy to notice that the embodiment of the present application uses a translation engine to translate the given annotated source language data to obtain the target language data; by annotating the aspect words with special symbols in the source language data, the aspect words and their corresponding emotional tendencies in the translation data are obtained according to the correspondence between the special symbols in the translation data, thereby mining the aspect-level opinions in the product reviews written by users in the target language. Therefore, the embodiment of the present application achieves the purpose of obtaining high-quality target language annotated data by a label mapping method without alignment, and based on the code conversion of the aspect words and the unlabeled target language data, the cross-language word vectors are further aligned, thereby achieving the technical effect of cross-language aspect-level sentiment analysis, thereby solving the technical problem in the prior art that the aspect-level sentiment analysis requires fine-grained annotation processing, resulting in low processing efficiency of sentiment analysis tasks in multiple target languages, which is not conducive to mining user emotional tendencies.

[0089] It should be noted that the translation module 500, conversion processing module 502, first training module 504, first input module 506, second training module 508, and second input module 510 described above correspond to steps S202 to S212 in Example 1. The examples and application scenarios implemented by these modules and the corresponding steps are the same, but are not limited to those disclosed in Example 1. It should be noted that these modules, as part of the apparatus, can be run in the computer terminal 10 provided in Example 1.

[0090] It should be noted that the preferred implementation of this embodiment can be found in the relevant description in Method Example 1, which will not be repeated here.

[0091] Example 3

[0092] According to an embodiment of the present application, an embodiment of an electronic device is also provided, which can be any computing device in a computing device group. The electronic device includes: a processor and a memory, wherein:

[0093] A processor; and a memory, connected to the processor, for providing the processor with instructions for processing the following processing steps: translating source language data into first target language data, wherein the source language data is pre-labeled data; performing code conversion on the source language data to obtain first code-conversion data, and performing code conversion on the first target language data to obtain second code-conversion data; training multiple models using the source language data, the first target language data, the first code-conversion data, and the second code-conversion data; inputting second target language data into the multiple models and outputting multiple first prediction results, wherein the second target language data is pre-collected unlabeled data; training a target model using the multiple first prediction results; inputting third target language data into the target model and outputting a second prediction result, wherein the third target language data is unlabeled data to be analyzed.

[0094] In an embodiment of the present application, a cross-language migration method that does not require target language annotated data is adopted, wherein source language data is translated into first target language data, wherein the above-mentioned source language data is pre-annotated data; code conversion processing is performed on the above-mentioned source language data to obtain first code-conversion data, and code conversion processing is performed on the above-mentioned first target language data to obtain second code-conversion data; multiple models are trained using the above-mentioned source language data, the above-mentioned first target language data, the above-mentioned first code-conversion data, and the above-mentioned second code-conversion data; second target language data is input into the above-mentioned multiple models, and multiple first prediction results are output, wherein the above-mentioned second target language data is pre-collected unlabeled data; a target model is trained using the above-mentioned multiple first prediction results; third target language data is input into the above-mentioned target model, and a second prediction result is output, wherein the above-mentioned third target language data is unlabeled data to be analyzed.

[0095] It is easy to notice that the embodiment of the present application uses a translation engine to translate the given annotated source language data to obtain the target language data; by annotating the aspect words with special symbols in the source language data, the aspect words and their corresponding emotional tendencies in the translation data are obtained according to the correspondence between the special symbols in the translation data, thereby mining the aspect-level opinions in the product reviews written by users in the target language. Therefore, the embodiment of the present application achieves the purpose of obtaining high-quality target language annotated data by a label mapping method without alignment, and based on the code conversion of the aspect words and the unlabeled target language data, the cross-language word vectors are further aligned, thereby achieving the technical effect of cross-language aspect-level sentiment analysis, thereby solving the technical problem in the prior art that the aspect-level sentiment analysis requires fine-grained annotation processing, resulting in low processing efficiency of sentiment analysis tasks in multiple target languages, which is not conducive to mining user emotional tendencies.

[0096] It should be noted that the preferred implementation of this embodiment can be found in the relevant description in Method Example 1, which will not be repeated here.

[0097] Example 4

[0098] According to an embodiment of the present application, a computer terminal embodiment is also provided, which can be any computer terminal device in a computer terminal group. Optionally, in this embodiment, the computer terminal can also be replaced by a terminal device such as a mobile terminal.

[0099] Optionally, in this embodiment, the computer terminal may be located in at least one network device among a plurality of network devices of a computer network.

[0100] In this embodiment, the computer terminal can execute the program code of the following steps in the language data processing method: translating source language data into first target language data, wherein the source language data is pre-labeled data; performing code conversion on the source language data to obtain first code-conversion data, and performing code conversion on the first target language data to obtain second code-conversion data; training multiple models using the source language data, the first target language data, the first code-conversion data, and the second code-conversion data; inputting second target language data into the multiple models and outputting multiple first prediction results, wherein the second target language data is pre-collected unlabeled data; training a target model using the multiple first prediction results; inputting third target language data into the target model and outputting a second prediction result, wherein the third target language data is unlabeled data to be analyzed.

[0101] Optionally, Figure 6 is a structural block diagram of another computer terminal according to an embodiment of the present application, such as Figure 6 As shown, the computer terminal may include: one or more (only one is shown in the figure) processors 602 , a memory 604 , and a peripheral interface 606 .

[0102] Among them, the memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the language data processing method and device in the embodiments of the present application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, that is, realizing the above-mentioned language data processing method. The memory may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include a memory remotely arranged relative to the processor, and these remote memories may be connected to the computer terminal via a network. Examples of the above-mentioned network include but are not limited to the Internet, corporate intranet, local area network, mobile communication network and combinations thereof.

[0103] The processor can call the information and application stored in the memory through the transmission device to perform the following steps: translating source language data into first target language data, wherein the above-mentioned source language data is pre-labeled data; performing code conversion processing on the above-mentioned source language data to obtain first code-conversion data, and performing code conversion processing on the above-mentioned first target language data to obtain second code-conversion data; training multiple models using the above-mentioned source language data, the above-mentioned first target language data, the above-mentioned first code-conversion data and the above-mentioned second code-conversion data; inputting the second target language data into the above-mentioned multiple models and outputting multiple first prediction results, wherein the above-mentioned second target language data is pre-collected unlabeled data; training a target model using the above-mentioned multiple first prediction results; inputting third target language data into the above-mentioned target model and outputting a second prediction result, wherein the above-mentioned third target language data is unlabeled data to be analyzed.

[0104] Optionally, the processor may also execute the program code of the following steps: using preset symbols to mark aspect words in the source language data, and restoring target language words corresponding to the aspect words in the first target language data, wherein the aspect words and the target language words are used to represent the target characteristics of the target object; converting the aspect words marked in the source language data into the target language words to obtain the first code conversion data, and converting the target language words restored in the first target language data into the aspect words to obtain the second code conversion data.

[0105] Optionally, the processor may further execute program code for the following steps: combining the source language data and the first target language data into first category training data; combining the first code conversion data and the first target language data into second category training data; combining the second code conversion data and the first target language data into third category training data; and using the first category training data, the second category training data and the third category training data to respectively train to obtain a first model, a second model and a third model.

[0106] Optionally, the processor may further execute program code for the following steps: inputting the second target language data into the first model, and outputting a first predicted probability distribution of each word in the second target language data by the first model; inputting the second target language data into the second model, and outputting a second predicted probability distribution of each word in the second target language data by the second model; inputting the second target language data into the third model, and outputting a third predicted probability distribution of each word in the second target language data by the third model; wherein the first model, the second model, and the third model all include: a linear classification network at the word level, and the linear classification network is used to obtain a predicted probability distribution for each word.

[0107] Optionally, the processor may also execute the program code of the following steps: combining the first predicted probability distribution, the second predicted probability distribution and the third predicted probability distribution into a fourth category of training data; and training the target model using the fourth category of training data.

[0108] Optionally, the processor may also execute the program code of the following steps: when the third target language data is product review data described in the target language, the product review data is input into the target model, and the target aspect words in the product review data and the emotional tendency of the target aspect words are output.

[0109] An embodiment of the present application provides a language data processing solution. The solution comprises translating source language data into first target language data, wherein the source language data is pre-labeled data; performing code conversion on the source language data to obtain first code-conversion data; and performing code conversion on the first target language data to obtain second code-conversion data; training multiple models using the source language data, the first target language data, the first code-conversion data, and the second code-conversion data; inputting second target language data into the multiple models, and outputting multiple first prediction results, wherein the second target language data is pre-collected unlabeled data; training a target model using the multiple first prediction results; and inputting third target language data into the target model, and outputting a second prediction result, wherein the third target language data is unlabeled data to be analyzed.

[0110] It is easy to notice that the embodiment of the present application uses a translation engine to translate the given annotated source language data to obtain the target language data; by annotating the aspect words with special symbols in the source language data, the aspect words and their corresponding emotional tendencies in the translation data are obtained according to the correspondence between the special symbols in the translation data, thereby mining the aspect-level opinions in the product reviews written by users in the target language. Therefore, the embodiment of the present application achieves the purpose of obtaining high-quality target language annotated data by a label mapping method without alignment, and based on the code conversion of the aspect words and the unlabeled target language data, the cross-language word vectors are further aligned, thereby achieving the technical effect of cross-language aspect-level sentiment analysis, thereby solving the technical problem in the prior art that the aspect-level sentiment analysis requires fine-grained annotation processing, resulting in low processing efficiency of sentiment analysis tasks in multiple target languages, which is not conducive to mining user emotional tendencies.

[0111] It can be understood by those skilled in the art that Figure 6 The structure shown is for illustration only, and the computer terminal may also be a smart phone (such as an Android phone, an iOS phone, etc.), a tablet computer, a handheld computer, a mobile Internet device (MID), a PAD, or other terminal devices. Figure 6 It does not limit the structure of the above electronic device. For example, the computer terminal may also include Figure 6 More or fewer components (such as network interfaces, display devices, etc.) shown in, or with Figure 6 Different configurations shown.

[0112] A person skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the hardware related to the terminal device through a program, and the program can be stored in a computer-readable computer-readable storage medium. The computer-readable storage medium may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0113] Example 5

[0114] According to an embodiment of the present application, an embodiment of a computer-readable storage medium is further provided. Optionally, in this embodiment, the computer-readable storage medium can be used to store program codes executed by the language data processing method provided in the above embodiment 1.

[0115] Optionally, in this embodiment, the computer-readable storage medium may be located in any computer terminal in a computer terminal group in a computer network, or in any mobile terminal in a mobile terminal group.

[0116] Optionally, in this embodiment, the computer-readable storage medium is configured to store program code for executing the following steps: translating source language data into first target language data, wherein the source language data is pre-labeled data; performing code conversion on the source language data to obtain first code-conversion data, and performing code conversion on the first target language data to obtain second code-conversion data; training multiple models using the source language data, the first target language data, the first code-conversion data, and the second code-conversion data; inputting the second target language data into the multiple models and outputting multiple first prediction results, wherein the second target language data is pre-collected unlabeled data; training a target model using the multiple first prediction results; inputting third target language data into the target model and outputting a second prediction result, wherein the third target language data is unlabeled data to be analyzed.

[0117] Optionally, in this embodiment, the computer-readable storage medium is configured to store program codes for executing the following steps: marking aspect words in the source language data using preset symbols, and restoring target language words corresponding to the aspect words in the first target language data, wherein the aspect words and the target language words are used to represent target characteristics of the target object; converting the aspect words marked in the source language data into the target language words to obtain the first code conversion data, and converting the target language words restored in the first target language data into the aspect words to obtain the second code conversion data.

[0118] Optionally, in this embodiment, the computer-readable storage medium is configured to store program codes for executing the following steps: combining the source language data and the first target language data into first category training data; combining the first code conversion data and the first target language data into second category training data; combining the second code conversion data and the first target language data into third category training data; and using the first category training data, the second category training data and the third category training data to respectively train to obtain a first model, a second model and a third model.

[0119] Optionally, in this embodiment, the computer-readable storage medium is configured to store program code for executing the following steps: inputting the second target language data into the first model, and outputting a first predicted probability distribution of the first model for each word in the second target language data; inputting the second target language data into the second model, and outputting a second predicted probability distribution of the second model for each word in the second target language data; inputting the second target language data into the third model, and outputting a third predicted probability distribution of the third model for each word in the second target language data; wherein the first model, the second model, and the third model all include: a word-level linear classification network, and the linear classification network is used to obtain a predicted probability distribution for each word.

[0120] Optionally, in this embodiment, the computer-readable storage medium is configured to store program code for performing the following steps: combining the above-mentioned first predicted probability distribution, the above-mentioned second predicted probability distribution and the above-mentioned third predicted probability distribution into a fourth category of training data; and using the above-mentioned fourth category of training data to train and obtain the above-mentioned target model.

[0121] Optionally, in this embodiment, the computer-readable storage medium is configured to store program code for executing the following steps: when the above-mentioned third target language data is product review data described in the target language, the above-mentioned product review data is input into the above-mentioned target model, and the target aspect words in the above-mentioned product review data and the emotional tendency of the above-mentioned target aspect words are output.

[0122] The serial numbers of the above embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0123] In the above embodiments of the present application, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0124] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0125] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0126] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0127] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a computer-readable storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned computer-readable storage medium includes: various media that can store program codes, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0128] The above is only a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for processing language data, characterized in that: include: Translating source language data into first target language data, wherein the source language data is pre-annotated data; performing code conversion on the source language data to obtain first code conversion data, and performing code conversion on the first target language data to obtain second code conversion data; training a plurality of models using the source language data, the first target language data, the first code-conversion data, and the second code-conversion data; Inputting second target language data into the multiple models and outputting multiple first prediction results, wherein the second target language data is pre-collected unlabeled data; Using the multiple first prediction results to train a target model; Input third target language data into the target model and output a second prediction result, wherein the third target language data is unlabeled data to be analyzed.

2. The method for processing language data according to claim 1, wherein: Performing code conversion on the source language data to obtain the first code conversion data, and performing code conversion on the first target language data to obtain the second code conversion data includes: Using preset symbols to mark aspect words in the source language data, and restoring target language words corresponding to the aspect words in the first target language data, wherein the aspect words and the target language words are used to represent target characteristics of a target object; The aspect words marked in the source language data are converted into the target language words to obtain the first code conversion data, and the target language words restored in the first target language data are converted into the aspect words to obtain the second code conversion data.

3. The method for processing language data according to claim 1, wherein: The multiple models are trained using the source language data, the first target language data, the first code-conversion data, and the second code-conversion data, including: combining the source language data and the first target language data into first category training data; combining the first code-conversion data and the first target language data into second category training data; combining the second code-conversion data and the first target language data into third category training data; The first category training data, the second category training data and the third category training data are used to train a first model, a second model and a third model respectively.

4. The method for processing language data according to claim 3, wherein: Inputting the second target language data into the multiple models and outputting the multiple first prediction results comprises: Inputting the second target language data into the first model, and outputting a first predicted probability distribution of each word in the second target language data by the first model; Inputting the second target language data into the second model, and outputting a second predicted probability distribution of each word in the second target language data by the second model; inputting the second target language data into the third model, and outputting a third predicted probability distribution of each word in the second target language data by the third model; The first model, the second model and the third model all include: a linear classification network at the word level, and the linear classification network is used to obtain the predicted probability distribution of each word.

5. The method for processing language data according to claim 4, characterized in that: The step of training the target model using the plurality of first prediction results includes: combining the first predicted probability distribution, the second predicted probability distribution, and the third predicted probability distribution into fourth category training data; The target model is obtained by training using the fourth category training data.

6. The method for processing language data according to claim 1, wherein: Inputting the third target language data into the target model, and outputting the second prediction result includes: When the third target language data is product review data described in the target language, the product review data is input into the target model, and the target aspect words in the product review data and the sentiment tendency of the target aspect words are output.

7. The method for processing language data according to claim 1, wherein: The language data processing method is applied to one of the following scenarios: Multilingual product review mining scenarios, multilingual named entity recognition scenarios, and multilingual relationship extraction scenarios.

8. A language data processing device, characterized in that: include: a translation module, configured to translate source language data into first target language data, wherein the source language data is pre-annotated data; a conversion processing module, configured to perform code conversion processing on the source language data to obtain first code-converted data, and to perform code conversion processing on the first target language data to obtain second code-converted data; a first training module, configured to train a plurality of models using the source language data, the first target language data, the first code-conversion data, and the second code-conversion data; a first input module, configured to input second target language data into the multiple models and output multiple first prediction results, wherein the second target language data is pre-collected unlabeled data; A second training module is configured to train a target model using the plurality of first prediction results; The second input module is configured to input third target language data into the target model and output a second prediction result, wherein the third target language data is unlabeled data to be analyzed.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute the language data processing method according to any one of claims 1 to 7.

10. An electronic device, characterized in that: include: processor; as well as A memory, connected to the processor, configured to provide the processor with instructions for processing the following processing steps: Translating source language data into first target language data, wherein the source language data is pre-annotated data; performing code conversion on the source language data to obtain first code conversion data, and performing code conversion on the first target language data to obtain second code conversion data; training a plurality of models using the source language data, the first target language data, the first code-conversion data, and the second code-conversion data; Inputting second target language data into the multiple models and outputting multiple first prediction results, wherein the second target language data is pre-collected unlabeled data; Using the multiple first prediction results to train a target model; Input third target language data into the target model and output a second prediction result, wherein the third target language data is unlabeled data to be analyzed.

Images