Zero-resource minority language data generation method and device, minority language resource library construction method and device and medium

Abstract

The invention belongs to the technical field of artificial intelligence, and provides a zero-resource minority language data generation method and device, a minority language resource library construction method and device and a medium. The method comprises the following steps: S1, acquiring zero-resource minority language data based on an existing channel, wherein the zero-resource minority language data comprises text data and voice data; s2, performing data expansion on the obtained zero-resource small language data to obtain original corpus data; s3, performing data cleaning and labeling on the original corpus data to obtain processed corpus data; and S4, classifying and organizing the processed corpus data, and constructing a diversified small language resource library. According to the method, the original corpus data is generated through corpus data collection and data expansion, the processed corpus data is classified and organized, a diversified small language resource library is constructed, and the problems that zero-resource small language corpus data is short, cleaning of the corpus data is complex, and labeling is difficult are solved.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and specifically, relates to a method, device and medium for generating language data of a zero-resource minority language and building a minority language resource library. Background Art

[0002] With the continuous development of global informatization, the demand for communication in minority languages ​​is increasing. Especially with the rapid development of artificial intelligence technology in recent years, communication and exchange between minority languages ​​have gained unprecedented convenience and opportunities. The widespread application of these technologies, such as machine translation, speech recognition and natural language processing, are highly dependent on rich language data resources. The vast majority of minority language data are still in a zero-resource state, that is, lack of corresponding corpus support.

[0003] Corpus is an important cornerstone in the field of artificial intelligence. It contains a large number of language examples for training and optimizing AI models. However, for minority languages, the relevant corpus data is extremely scarce due to the relatively small number of users. What's worse, due to the lack of automatic corpus generation and processing tools, the construction of existing minority language resource libraries faces many difficulties. This not only limits the application of minority languages ​​in the field of AI, but also hinders the communication and understanding of global multilingual cultures. Summary of the invention

[0004] The purpose of the present invention is to provide a method for generating language data of a zero-resource minority language and building a minority language resource library to solve the technical problems existing in the prior art.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is as follows: A method for generating language data of a zero-resource minority language and building a minority language resource library comprises the following steps: Step S1: Acquire language data of zero-resource minority languages ​​based on existing channels, including text data and voice data; Step S2: perform data expansion on the acquired zero-resource minority language data to obtain original corpus data; Step S3: Clean and annotate the original corpus data to obtain processed corpus data; Step S4: Classify and organize the processed corpus data to build a diversified minority language resource library.

[0006] In a preferred embodiment, the text data in step S1 includes monolingual data and bilingual data, and the acquisition method includes text analysis, image analysis, and manual editing. The acquired corpus includes a dictionary, sentence-level corpus, and document-level corpus.

[0007] In a preferred embodiment, in step S1, for voice data, when the audio source is audio data, the audio data is obtained by an automatic audio extraction method; when the audio source is video data, the audio data in the video data is extracted by a subtitle extraction method or an OCR method, and then the audio data is converted into text data.

[0008] In a preferred embodiment, the data expansion in step S2 includes expansion for text data and expansion for voice data.

[0009] In a preferred embodiment, the method for expanding the text data includes any one or more combinations of the following: Dictionary-based expansion method: Based on text data, using existing dictionary resources, a dictionary of synonyms of the original text and the translation is constructed, and keywords in the text are identified and replaced with synonyms through word frequency statistics and part-of-speech analysis methods; Phrase-based expansion method: Through syntactic analysis and statistical analysis, common phrases and fixed collocations in text data are identified, high-frequency phrases are extracted, and high-frequency phrases are replaced and combined to generate new sentence structures; Expansion method based on back translation: Based on text data, through machine translation method, find a high-resource language, train a machine translation model from the high-resource language to the low-resource language, and use the machine translation model to translate the high-resource language sentences into the low-resource language sentences to generate new corpus.

[0010] In a preferred embodiment, the method for expanding the voice data includes any one or more combinations of the following: Dynamic time warping method: The dynamic time warping algorithm is used to distort the time axis of the speech signal to generate new speech data; Frequency masking method: masking signals of certain frequency bands in the spectrum graph, using the auditory masking effect of the human ear to generate new speech data by selectively reducing or completely masking signals within a specific frequency range; Hybrid method: Mix two or more speech signals together to generate new synthetic speech data; Enhancement method: denoising, filtering and enhancing the original speech signal.

[0011] In a preferred embodiment, in step S3, data cleaning for text data includes deduplication, special characters, number inconsistency, and language inconsistency; cleaning for voice data includes denoising, noise reduction, text checking, and speech speed adjustment; data annotation includes either or both of manual annotation and model annotation.

[0012] In a preferred embodiment, in step S4, the classification method is as follows: a multi-level classification system is adopted to design categories of different levels for the corpus data.

[0013] In order to achieve the above objectives, the present invention also provides a computer-readable storage medium on which a computer program is stored. The computer program is executed by a processor to implement the method for generating zero-resource minority language data and building a minority language resource library as described above.

[0014] In order to achieve the above-mentioned purpose, the present invention also provides a device for generating language data of a zero-resource minority language and building a minority language resource library, comprising: a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory, and is used to execute the computer program stored in the memory, so that the device for generating language data of a zero-resource minority language and building a minority language resource library executes the method for generating language data of a zero-resource minority language and building a minority language resource library as described above.

[0015] Compared with the prior art, the present invention has the following beneficial effects: (1) According to the present invention, the problem of lack of corpus data for zero-resource minority languages ​​is solved by generating original corpus data through corpus data collection and data expansion.

[0016] (2) According to the present invention, the automatic processing of corpus data can be realized, solving the problems of complex corpus data cleaning and difficult labeling.

[0017] (3) According to the present invention, through the construction of a resource library, the language data of zero-resource minority languages ​​is managed, providing convenient services for users to use the language data of minority languages. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a schematic diagram of the process of Example 1 of the present invention. DETAILED DESCRIPTION

[0019] In order to enable those skilled in the art to have a clearer understanding and knowledge of the present invention, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described below are only used to explain the present invention, and are convenient for understanding. The technical solutions provided by the present invention are not limited to the technical solutions provided by the following embodiments, and the technical solutions provided by the embodiments should not limit the protection scope of the present invention.

[0020] It should be noted that the illustrations provided in the following embodiments are only used to illustrate the basic concept of the present invention in a schematic manner. Therefore, the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the form, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0021] Example 1 like Figure 1 As shown, this embodiment provides a method for generating zero-resource minority language data and building a minority language resource library, which can be used to solve the problems of scarcity of minority language corpus data and lack of automated tools and resource library management.

[0022] In this embodiment, the method for generating zero-resource minority language data and building a minority language resource library mainly includes the following steps: 1. Step S1: Obtaining language data of zero-resource minority languages ​​based on existing channels Minority language data include text data and voice data, among which text data include monolingual data and bilingual data, and its acquisition methods include text analysis, image analysis, and manual organization. The acquired corpus includes dictionaries, sentence-level corpora, document-level corpora, etc.; for voice data, when the audio source is audio data, the audio data is obtained through an automatic audio extraction method. When the audio source is video data, the audio data in the video data is extracted through a subtitle extraction method or an OCR method. The purpose is to retain the correspondence between voice data and text data, which can be used to train voice recognition models, voice synthesis models, direct voice translation and other voice-related models, and then convert the audio data into text data.

[0023] Step S2: Expand the acquired zero-resource minority language data to obtain the original corpus data In this step, data expansion includes expansion for text data and expansion for voice data. The expansion method for text data is as follows: Dictionary-based expansion: Use existing dictionary resources to build a dictionary of synonyms of the original text and the translated text. Identify keywords in the text and replace them with synonyms through word frequency statistics, part-of-speech analysis, etc., to increase the diversity of the text and ensure that the replaced words are still reasonable in the context; apply word form change rules (such as verb tense, noun plural form, etc.) to generate different forms of vocabulary, thereby expanding the expression of the text.

[0024] Phrase-based expansion: Through syntactic analysis and statistical analysis, common phrases and fixed collocations in the text are identified, and high-frequency phrases are extracted. The identified phrases are replaced and combined to generate new sentence structures.

[0025] Expansion based on back translation: Through machine translation methods, find a high-resource language (such as English, Japanese, etc., the high-resource language is not specifically limited), train a machine translation model from the high-resource language to a low-resource language, and use the machine translation model to translate a large number of high-resource language sentences into low-resource language sentences to generate new corpus.

[0026] Combining the above three expansion methods, a multi-level, comprehensive text expansion strategy is formed to improve the quality and quantity of data generation. A feedback mechanism is established to continuously optimize the expansion algorithm by evaluating the use effect of the generated data.

[0027] The method for expanding voice data is as follows: Through voice data augmentation technology, the spectrogram of audio data is expanded. Dynamic Time Warping (DTW): Through the dynamic time warping algorithm, the time axis of the voice signal is distorted to adapt to different speaking speeds and rhythms. This method can effectively increase the variability of speech and generate new deformed speech data.

[0028] Frequency masking technology: masks signals in certain frequency bands in the spectrum graph, and uses the auditory masking effect of the human ear to generate new speech data by selectively reducing or completely masking signals in a specific frequency range. This method can simulate speech heard in a noisy environment, making the model more robust; by masking the pronunciation of certain time periods, randomly selecting and masking a continuous time frame, new speech data with part of the signal missing is generated. This method can help the model learn to perform effective recognition when some information is missing.

[0029] Mixing and enhancement: Mix multiple speech signals together to generate new synthetic speech data. This can be done by mixing voice clips from different speakers or combining speech with background noise to improve the model's adaptability to real-world scenarios; denoising, filtering, and enhancing the original speech signal to improve speech quality and clarity. Through these processes, the generated new speech data can be made closer to natural human speech, thereby improving the model's training effect.

[0030] Through the above expansion, we can get the original corpus data Step S3: Clean and label the original corpus data to obtain processed corpus data Use advanced artificial intelligence technology, machine learning algorithms, etc. to clean and annotate the collected language data of small languages. In the data cleaning stage: build a data cleaning framework and configure dynamically adjustable cleaning rules; data cleaning for text data includes deduplication, special characters, number inconsistencies, and language inconsistencies; cleaning for voice data includes denoising, noise reduction, text inspection, and speech speed adjustment; data annotation includes either or both manual annotation and model annotation.

[0031] Data labeling: Based on the labeling platform, it provides basic functions such as text highlighting, selection, and modification, and supports filtering data by category, time, label, and other conditions. According to the requirements of downstream tasks, the labeling format is designed, such as entity name, attribute, etc., and the functions of adding, deleting, and modifying labels are provided. Data labeling is performed through manual labeling, model labeling, or a combination of the two. Professional training is provided for manual labeling personnel to ensure the quality of labeling, and labeling data is regularly checked to ensure the accuracy of labeling. For model labeling, the appropriate pre-trained model is selected according to the task requirements, and the model labeling results are manually proofread and adjusted. Among them, voice labeling needs to support common audio formats and support proofreading and correction of automatic transcription results.

[0032] Step S4: Classify and organize the processed corpus data to build a diverse minority language resource library The processed language data is classified and organized to build a diverse library of small language resources, provide language data support for relevant users, and support automated model training.

[0033] In this step, the classification method is as follows: adopt a multi-level classification system to design different levels of categories for the corpus data. Through multi-level classification, it helps users find specific category data more quickly, such as language, field, time period, etc., and classifies and stores language data according to different categories.

[0034] Through the above steps, the generation of language data for zero-resource minority languages ​​and the construction of a minority language resource library are achieved.

[0035] Based on the construction of the above-mentioned small language resource library, we can subsequently develop a language data calling interface for users to support users in querying dictionaries, phrases, attributes, sentences and other language data; develop a model automation training data calling interface, and users can directly call relevant data from the resource library for model training and iteration by setting model training parameters.

[0036] Example 2 This embodiment provides a computer-readable storage medium on which a computer program is stored, and the computer program is executed by a processor to implement the method for generating language data of zero-resource minority languages ​​and building a minority language resource library provided in Embodiment 1. A person of ordinary skill in the art can understand that all or part of the steps of implementing the method provided in Embodiment 1 can be completed by hardware related to the computer program, and the above-mentioned computer program can be stored in a computer-readable storage medium, and when the program is executed, the steps of the method provided in Embodiment 1 are executed; and the above-mentioned storage medium includes: ROM, RAM, magnetic disk or optical disk and other media that can store program codes.

[0037] Example 3 This embodiment provides a device for generating language data of a zero-resource minority language and building a minority language resource library, comprising: a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory, and is used to execute the computer program stored in the memory, so that the device for generating language data of a zero-resource minority language and building a minority language resource library executes the method for generating language data of a zero-resource minority language and building a minority language resource library provided in Example 1.

[0038] Specifically, the memory includes: ROM, RAM, disk, USB flash drive, memory card or CD and other media that can store program codes.

[0039] Preferably, the processor can be a general-purpose processor, including a central processing unit, a network processor, etc.; it can also be a digital signal processor, an application-specific integrated circuit, a field programmable gate array or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component.

[0040] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by a person of ordinary skill in the art without departing from the spirit and technical concept disclosed by the present invention shall still be covered by the claims of the present invention.

Claims

1. A method for generating language data of a zero-resource minority language and building a minority language resource library, characterized in that: The following steps are involved: Step S1: Acquire language data of zero-resource minority languages ​​based on existing channels, including text data and voice data; Step S2: perform data expansion on the acquired zero-resource minority language data to obtain original corpus data; Step S3: Clean and annotate the original corpus data to obtain processed corpus data; Step S4: Classify and organize the processed corpus data to build a diversified minority language resource library.

2. The method for generating zero-resource minority language data and building a minority language resource library according to claim 1, characterized in that: The text data in step S1 includes monolingual data and bilingual data, and the acquisition method includes text analysis, image analysis, and manual sorting. The acquired corpus includes a dictionary, sentence-level corpus, and document-level corpus.

3. The method for generating zero-resource minority language data and building a minority language resource library according to claim 2, characterized in that: In the step S1, for voice data, when the audio source is audio data, the audio data is obtained by an automatic audio extraction method; when the audio source is video data, the audio data in the video data is extracted by a subtitle extraction method or an OCR method, and then the audio data is converted into text data.

4. The method for generating zero-resource minority language data and building a minority language resource library according to claim 3, characterized in that: The data expansion in step S2 includes expansion for text data and expansion for voice data.

5. The method for generating zero-resource minority language data and building a minority language resource library according to claim 4, characterized in that: The method for expanding the text data includes any one or more combinations of the following: Dictionary-based expansion method: Based on text data, using existing dictionary resources, a dictionary of synonyms of the original text and the translation is constructed, and keywords in the text are identified and replaced with synonyms through word frequency statistics and part-of-speech analysis methods; Phrase-based expansion method: Through syntactic analysis and statistical analysis, common phrases and fixed collocations in text data are identified, high-frequency phrases are extracted, and high-frequency phrases are replaced and combined to generate new sentence structures; Expansion method based on back translation: Based on text data, through machine translation method, find a high-resource language, train a machine translation model from the high-resource language to the low-resource language, and use the machine translation model to translate the high-resource language sentences into the low-resource language sentences to generate new corpus.

6. The method for generating zero-resource minority language data and building a minority language resource library according to claim 5, characterized in that: The method for expanding the voice data includes any one or more combinations of the following: Dynamic time warping method: The dynamic time warping algorithm is used to distort the time axis of the speech signal to generate new speech data; Frequency masking method: masking signals of certain frequency bands in the spectrum graph, using the auditory masking effect of the human ear to generate new speech data by selectively reducing or completely masking signals within a specific frequency range; Hybrid method: Mix two or more speech signals together to generate new synthetic speech data; Enhancement method: denoising, filtering and enhancing the original speech signal.

7. The method for generating zero-resource minority language data and building a minority language resource library according to claim 6, characterized in that: In step S3, data cleaning for text data includes deduplication, special characters, number inconsistency, and language inconsistency; cleaning for voice data includes denoising, noise reduction, text checking, and speech speed adjustment; data annotation includes any one or both of manual annotation and model annotation.

8. The method for generating zero-resource minority language data and building a minority language resource library according to claim 7, characterized in that: In step S4, the classification method is as follows: a multi-level classification system is adopted to design categories of different levels for the corpus data.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: The computer program is executed by a processor to implement the method for generating zero-resource minority language data and building a minority language resource library as described in any one of claims 1 to 8.

10. A device for generating language data of a zero-resource minority language and building a minority language resource library, characterized in that: include: Processor and memory; The memory is used to store computer programs; The processor is connected to the memory and is used to execute the computer program stored in the memory, so that the zero-resource minority language data generation and minority language resource library construction device executes the zero-resource minority language data generation and minority language resource library construction method as described in any one of claims 1 to 8.

Images