Acronym recognition method and device, computer equipment and storage medium
Through the similarity retrieval of word embedding model and vector database, the coverage and efficiency problems of abbreviation recognition in the prior art are solved, and more accurate and faster understanding and recognition of domain-specific terms are achieved.
Patent Information
- Application Number
- CN202510506839.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-22
- Publication Date
- 2025-08-01
AI Technical Summary
The prior art has limited coverage, untimely updates, lack of semantic correlation, low retrieval efficiency and low matching accuracy in the recognition of abbreviations of proper nouns in the processing field, especially in multilingual support and ambiguity processing.
By obtaining the target abbreviation, inputting the word embedding model for reasoning, obtaining the word embedding vector, and similarity search is performed in the target vector database, combining the search enhancement generation process to obtain the corresponding target term information.
It improves the accuracy and efficiency of abbreviation recognition, enhances the semantic understanding of domain-specific terms, covers more professional abbreviations and abbreviations, reduces error matching, and meets real-time processing needs.
Smart Images
Figure CN120409480A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of data processing, and particularly to an abbreviation recognition method and apparatus thereof, a computer device, and a storage medium. Background Art
[0002] In multiple professional fields such as computer science, medicine, and automotive engineering, abbreviations and contractions are widely used as efficient communication tools. However, these abbreviations and contractions often pose an understanding barrier for non-professionals or newbies in the field. Currently, although there are some dictionary-like tools on the market, such as embedding models like Word2Vec and GloVe, which can be used to query the meanings of common abbreviations and contractions, they perform poorly when dealing with domain-specific terms.
[0003] The existing retrieval methods have the following technical defects: First, the coverage is limited, usually only being able to query general and common abbreviations and contractions, while the professional abbreviations and contractions in specific fields are not fully included; Second, the update is not timely. With the development of technology, new abbreviations and contractions emerge continuously, but the traditional dictionaries have a long update cycle and cannot reflect these changes in a timely manner; Third, there is a lack of semantic relevance, only providing simple correspondence query and unable to give more accurate and relevant full names or phrases according to the context semantics of the input word; At the same time, the existing embedding models have limited understanding ability for domain-specific terms; In addition, the traditional database retrieval method has poor performance and low retrieval efficiency when dealing with large-scale vector data; Finally, the existing string matching-based methods cannot handle the diversity of abbreviations and contractions, and the matching accuracy is low. Summary of the Invention
[0004] An object of an embodiment of the present invention is to provide an abbreviation recognition method and apparatus thereof, a computer device, and a storage medium for solving the technical problems that the traditional retrieval method is inefficient in the face of large-scale data, cannot meet the real-time processing requirements, and has low retrieval accuracy.
[0005] In a first aspect, an embodiment of the present invention provides an abbreviation recognition method, and the method includes:
[0006] Obtain a target abbreviation;
[0007] Input the target abbreviation into a word embedding model for inference to obtain a target word embedding vector output by the word embedding model;
[0008] In a target vector database, perform a similarity retrieval on the target word embedding vector to obtain a term vector corresponding to the target word embedding vector;
[0009] Perform retrieval enhancement generation processing on the term vector corresponding to the target word embedding vector to obtain target term information corresponding to the target abbreviation.
[0010] In combination with the first aspect, in a possible implementation, in the target vector database, performing a similarity retrieval on the target word embedding vector to obtain the term vector corresponding to the target word embedding vector includes: performing a preset process on the target word embedding vector and all block vectors in the target vector database to obtain a plurality of similarity scores; screening out at least one target similarity score greater than a preset similarity score threshold from the plurality of similarity scores; obtaining a first term vector according to the at least one target similarity score; and performing a calibration process on the first term vector to determine the first term vector that passes the calibration as the term vector corresponding to the target word embedding vector.
[0011] It can be seen that in this embodiment, it is ensured that the target word embedding vector can be accurately matched with the term vector in the database, providing a reliable basis for subsequent natural language processing tasks.
[0012] In combination with the first aspect, in a possible implementation, the performing a preset process on the target word embedding vector and all block vectors in the target vector database to obtain a plurality of similarity scores includes: calculating the target word embedding vector and each block vector in the target vector database respectively to obtain the similarity score corresponding to the target word embedding vector and each block vector.
[0013] It can be seen that in this embodiment, the similarity degree between the target word embedding vector and each block vector in the database can be quantified as a similarity score, providing a basis for subsequent similarity ranking and screening.
[0014] In combination with the first aspect, in a possible implementation, the performing a retrieval enhancement generation process on the term vector corresponding to the target word embedding vector to obtain the target term information corresponding to the target abbreviation includes: obtaining first term information according to the term vector corresponding to the target word embedding vector; obtaining first input data according to the first term information and the target abbreviation; performing a format conversion process on the first input data to obtain target input data; and inputting the target input data into a preset large language model for a retrieval enhancement generation process to obtain the target term information corresponding to the target abbreviation output by the preset large language model.
[0015] It can be seen that in this embodiment, through the first term information and the target abbreviation, combined with retrieval enhancement generation, it is possible to give the most appropriate full name or phrase according to the context of the input data, enhancing the understanding and expansion ability of domain terms and improving the accuracy and practicality of recognition.
[0016] In combination with the first aspect, in a possible implementation, obtaining the first input data according to the first term information and the target abbreviation includes: obtaining a preset text format; arranging the first term information and the target abbreviation according to the preset text format to obtain the first input data.
[0017] It can be seen that in this embodiment, through the preset text format, the term information and the abbreviation are ensured to be displayed and processed in a clear and structured manner.
[0018] In combination with the first aspect, in a possible implementation, before performing a similarity search on the target word embedding vector in the target vector database, the method further includes: obtaining preset text data; performing text preprocessing on the preset text data to obtain a training phrase corresponding to the preset text data; inputting the training phrase into the word embedding model for inference to obtain a word embedding vector output by the word embedding model; storing the word embedding vector in a preset vector database to obtain a target vector database.
[0019] It can be seen that in this embodiment, the preset text data is effectively converted into structured word embedding vectors and stored in the target vector database, ensuring the quality of the data and the effectiveness of the model.
[0020] In combination with the first aspect, in a possible implementation, performing text preprocessing on the preset text data to obtain a target phrase corresponding to the preset text data includes: performing word segmentation on the preset text data to obtain a plurality of first word segmentation phrases; performing denoising processing on the plurality of first word segmentation phrases to obtain second word segmentation phrases after denoising; performing part-of-speech processing on the second word segmentation phrases to obtain a training phrase corresponding to the preset text data.
[0021] It can be seen that in this embodiment, through word segmentation, denoising, and part-of-speech processing, the preset text data is systematically converted into structured training phrases, providing high-quality input for subsequent model training, further ensuring the cleanliness and semantic accuracy of the data, and supporting more effective machine learning and natural language processing tasks.
[0022] In a second aspect, an embodiment of the present invention provides an abbreviation recognition device, and the device includes:
[0023] An acquisition unit, configured to acquire a target abbreviation;
[0024] A training unit, configured to input the target abbreviation into a word embedding model for inference to obtain a target word embedding vector output by the word embedding model;
[0025] A retrieval unit, configured to perform a similarity retrieval on the target word embedding vector in a target vector database to obtain a term vector corresponding to the target word embedding vector;
[0026] A processing unit, configured to perform retrieval enhancement generation processing on the term vector corresponding to the target word embedding vector to obtain target term information corresponding to the target abbreviation.
[0027] In a third aspect, a computer device is provided. The computer device includes a memory and a processor. The memory is connected to the processor. The processor is configured to execute one or more computer programs stored in the memory. When the processor executes the one or more computer programs, the computer device implements the abbreviation recognition method as described in the first aspect.
[0028] In a fourth aspect, a computer-readable storage medium is provided. The computer-readable storage medium stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, the processor executes the abbreviation recognition method as described in the first aspect.
[0029] In the embodiments implemented by the above abbreviation recognition method, its device, computer device, and storage medium, first, a target abbreviation is obtained. Secondly, the target abbreviation is input into a word embedding model for inference to obtain a target word embedding vector output by the word embedding model. Then, in a target vector database, a similarity retrieval is performed on the target word embedding vector to obtain a term vector corresponding to the target word embedding vector. Finally, retrieval enhancement generation processing is performed on the term vector corresponding to the target word embedding vector to obtain target term information corresponding to the target abbreviation. In this embodiment, by inputting the target abbreviation into the word embedding model for inference, the corresponding word embedding vector can be obtained more accurately, thereby improving the accuracy of abbreviation recognition. Further, by using the target word embedding vector output by the word embedding model for similarity retrieval, the semantic association between the abbreviation and the full-term term can be better captured, enhancing the semantic understanding ability of domain-specific terms. Performing similarity retrieval in the target vector database can find the corresponding term vector more quickly compared to the traditional database retrieval method, improving the retrieval efficiency. Through efficient similarity retrieval and retrieval enhancement generation processing, the target term information corresponding to the target abbreviation can be output in real time to meet the requirements of real-time processing. Further, by performing retrieval enhancement generation processing on the term vector corresponding to the target word embedding vector, more professional abbreviations and acronyms in specific fields can be covered, improving the matching accuracy and reducing the situation of incorrect matching. Description of the Drawings
[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.
[0031] Figure 1 It is a flowchart of a method for identifying abbreviations in an embodiment of the present invention;
[0032] Figure 2 It is a structural diagram of a device for identifying abbreviations in an embodiment of the present invention;
[0033] Figure 3 It is a structural diagram of a computer device in an embodiment of the present invention. Detailed implementation manners
[0034] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.
[0035] It should be noted that if there is no conflict, the various features in the embodiments of the present invention can be combined with each other, and all are within the scope of protection of the present invention. In addition, although the functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order from the module division in the device or the flowchart. Furthermore, the terms "first", "second", "third", etc. used in the present invention do not limit the data and the execution order, but only distinguish the same items or similar items with basically the same functions and effects.
[0036] In addition to the problems existing in the related technologies pointed out in the background art, other related problems are also found: (1) Existing tools may not support multiple languages comprehensively enough, resulting in poor performance when processing non-English abbreviations and terms; (2) Some abbreviations may have multiple meanings, and existing tools may not be able to effectively distinguish and process these polysemous meanings, resulting in incorrect or ambiguous expansion results.
[0037] To solve these problems, this method proposes an acronym recognition method. In this embodiment, by inputting the target acronym into the word embedding model for inference, the corresponding word embedding vector can be obtained more accurately, thereby improving the accuracy of acronym recognition. Further, using the target word embedding vector output by the word embedding model for similarity retrieval can better capture the semantic association between the acronym and the full-term term, enhancing the semantic understanding ability of domain-specific terms. Conducting similarity retrieval in the target vector database can find the corresponding term vector more quickly compared to traditional database retrieval methods, improving the retrieval efficiency. Through efficient similarity retrieval and retrieval-enhanced generation processing, the target term information corresponding to the target acronym can be output in real time to meet the requirements of real-time processing. Further, through retrieval-enhanced generation processing on the term vector corresponding to the target word embedding vector, more professional abbreviations and contractions in specific fields can be covered, improving the matching accuracy and reducing the situation of incorrect matching.
[0038] See Figure 1 , Figure 1 is a schematic flowchart of the acronym recognition method provided by the embodiment of the present invention. The method includes the following steps:
[0039] S10. Obtain the target acronym.
[0040] Among them, the target acronym refers to the acronym or abbreviated word to be recognized, queried, or expanded, and this acronym is not limited to a specific language, such as Chinese, English, etc.
[0041] Among them, the obtaining process can be input by the user himself or obtained by text extraction, and there is no unique limitation here.
[0042] Specifically, if the obtaining process is obtained by text extraction, then use a preset rule, algorithm, or machine learning model to identify possible acronyms from the text. For example, match according to common acronym formats (such as initial letter abbreviations, numerical abbreviations, etc.); or query a predefined acronym vocabulary table to see if the input contains known acronyms; or use a trained machine learning model, such as a sequence labeling model, to identify acronyms; or extract the identified acronyms from the text to form a separate list or data structure for subsequent inference and retrieval processing.
[0043] Optionally, preprocess the input target acronym, including removing irrelevant information, formatting, word segmentation, etc., to better recognize the acronym.
[0044] S20. Input the target acronym into the word embedding model for inference to obtain the target word embedding vector output by the word embedding model.
[0045] Among them, the above word embedding model is a trained machine learning model used to convert text data into embedding vectors, such as models like BERT and GPT in natural language processing.
[0046] Among them, the inference process refers to the process of using the trained model to make predictions or generate outputs for new data.
[0047] Among them, the target word embedding vector is a multi-dimensional vector output by the target acronym in the word embedding model, representing the semantic features of the target acronym in the vector space.
[0048] Optionally, before inputting the target acronym into the word embedding model for inference, the method further includes: performing data preprocessing on the target acronym to obtain the input text, such as removing special characters, converting to lowercase, etc., to meet the input requirements of the word embedding model. Further, convert the input text into a format acceptable to the word embedding model, such as a Tensor or numpy array.
[0049] It can be seen that in this embodiment, through the word embedding model, the target acronym can be converted into a word embedding vector with semantic information, providing a basis for subsequent term information generation and retrieval.
[0050] S30. In the target vector database, perform a similarity search on the target word embedding vector to obtain the term vector corresponding to the target word embedding vector.
[0051] Among them, the target vector database is a database that stores a large number of word embedding vectors, used to support similarity search and semantic analysis. The block vectors in this target vector database are usually generated by a pre-trained deep learning model and can capture the semantic information of the terms.
[0052] Among them, similarity search refers to finding the closest vector in the vector space by calculating the similarity between the target word embedding vector and other vectors in the target vector database.
[0053] Among them, the term vector refers to the embedding vector stored in the target vector database corresponding to a specific term. The term vector represents the position and semantic information of the term in the vector space.
[0054] For example, assume that the input target abbreviation is "CPU", and its corresponding target word embedding vector is [0.2, -0.5, 0.8, …]. In the target vector database, there are the following terms and their block vectors: "Central Processing Unit" -> [0.18, -0.49, 0.82, …], "Computer Hardware" -> [0.1, -0.3, 0.7, …], "Processor Architecture" -> [0.15, -0.4, 0.78, …]; through similarity retrieval, calculate the similarity between the embedding vector of "CPU" and these term vectors, and it is found that the similarity score of "Central Processing Unit" is the highest. Therefore, it is determined that the embedding vector of "Central Processing Unit" is the corresponding term vector of "CPU".
[0055] It can be seen that in this embodiment, the target word embedding vector is effectively matched with the term vectors in the database, providing a basis for subsequent semantic analysis and information processing.
[0056] S40. Perform retrieval enhancement generation processing on the term vector corresponding to the target word embedding vector to obtain the target term information corresponding to the target abbreviation.
[0057] Among them, the retrieval enhancement generation processing is a processing method for optimizing and enhancing the retrieval results, aiming to improve the accuracy and relevance of the retrieval.
[0058] Among them, the target term information refers to information such as the full name, definition, relevant context, etc. corresponding to the target abbreviation.
[0059] In a specific implementation, input the target word embedding vector into the target vector database to find the term vector most similar to the target word embedding vector; return the most similar term vector as the retrieval result; perform context analysis on the retrieved term vector, considering the context information of the target abbreviation; machine learning or deep learning models can be applied to optimize the term vector to enhance its representation ability. For example, the attention mechanism model can be used to readjust the weights of the term vector to make more important features more prominent; according to the enhanced term vector, generate the target term information corresponding to the target abbreviation, which can include but is not limited to the full name of the term, definition, relevant context, usage examples, etc., and use natural language generation technology to convert the vector information into a human-readable text form; finally, perform post-processing on the generated target term information, such as format adjustment, duplicate removal, error correction, etc.
[0060] For example, assume that the target acronym is "CPU", and the retrieved corresponding term vector is the embedding vector of "central processing unit". In the enhancement stage, it is further considered that "CPU" often appears in the context related to computer hardware. Therefore, the vector of "central processing unit" is optimized to make its features related to hardware more prominent. In the generation stage, information such as the full name, definition, and role in the computer system of "central processing unit" can be output as the target term information of "CPU".
[0061] It can be seen that in this embodiment, the retrieval enhancement generation processing can be effectively performed on the term vector corresponding to the target word embedding vector, so as to provide accurate and relevant target term information to help users better understand and use the acronyms.
[0062] In this embodiment, by inputting the target acronym into the word embedding model for inference, the corresponding word embedding vector can be obtained more accurately, thereby improving the accuracy of acronym recognition; further, using the target word embedding vector output by the word embedding model for similarity retrieval can better capture the semantic association between the acronym and the full-term term, enhancing the semantic understanding ability of domain-specific terms; performing similarity retrieval in the target vector database can find the corresponding term vector more quickly compared with the traditional database retrieval method, improving the retrieval efficiency; through efficient similarity retrieval and retrieval enhancement generation processing, the target term information corresponding to the target acronym can be output in real time to meet the requirements of real-time processing; further, through the retrieval enhancement generation processing of the term vector corresponding to the target word embedding vector, more professional abbreviations and acronyms in specific fields can be covered, improving the matching accuracy and reducing the situation of incorrect matching.
[0063] In one embodiment, in the target vector database, performing similarity retrieval on the target word embedding vector to obtain the term vector corresponding to the target word embedding vector includes: performing preset processing on the target word embedding vector and all block vectors in the target vector database to obtain a plurality of similarity scores; screening out at least one target similarity score greater than the preset similarity score threshold from the plurality of similarity scores; obtaining a first term vector according to the at least one target similarity score; and performing calibration processing on the first term vector to determine the first term vector that passes the calibration as the term vector corresponding to the target word embedding vector.
[0064] Among them, the target vector database is a database that stores a large number of term vectors and their corresponding full names or phrases.
[0065] Among them, the block vector is a single vector in the target vector database, which can be an entry or a subset in the database.
[0066] Among them, the preset processing is to calculate the similarity between two vectors using predefined similarity methods (such as cosine similarity, Euclidean distance, etc.). Among them, the cosine similarity is to calculate the cosine value of the angle between two vectors to determine their similarity; the Euclidean distance is to measure the distance between two vector points, and the smaller the distance, the more similar.
[0067] Among them, the similarity score refers to the numerical value representing the similarity degree between two vectors obtained through the preset processing.
[0068] Among them, the preset similarity score threshold refers to a preset numerical value used to screen out sufficiently similar vectors. This preset similarity score threshold can be an empirical value or set artificially, and there is no unique limitation here.
[0069] Among them, the target similarity score is the similarity score greater than the preset similarity score threshold.
[0070] Among them, the first term vector refers to the term vector with the highest similarity score to the target word embedding vector.
[0071] Among them, the calibration processing is a process of further verifying and adjusting the selected first term vector to ensure its accuracy and relevance. For example, checking whether the corresponding term conforms to the context semantics of the target abbreviation and whether there is polysemy, etc. The calibration processing can also use deep learning models, such as recurrent neural networks (RNNs), Transformers, etc., to verify the selected candidate terms to ensure that the mapped terms are reasonable in the context semantics of the user input.
[0072] In a specific implementation, calculate the similarity between the target word embedding vector and each block vector in the database, screen the calculated similarity scores, retain those scores that exceed the preset similarity score threshold, and extract the corresponding term vectors according to the screened similarity scores as candidate matches for the target word embedding vector; calibrate the extracted term vectors; finally determine the term vector that passes the calibration as the term vector corresponding to the target word embedding vector to ensure its accuracy in the semantic space.
[0073] For example, assume the target acronym is "AI". Input "AI" into a pre-trained embedding model (such as Word2Vec), and the model outputs a high-dimensional vector representing the position of "AI" in the vector space. For example, the vector might be [0.1, -0.3, 0.5, …]; Multiple terms and their corresponding block vectors are stored in the database. For example: "Artificial Intelligence" -> [0.09, -0.29, 0.52, …], "Art History" -> [0.02, 0.15, -0.38, …], "Aviation Industry" -> [-0.20, 0.35, 0.40, …]…; Further, using cosine similarity as the calculation method, calculate the similarity scores between the embedding vector of "AI" and the block vectors of each term in the database, obtaining multiple similarity scores: The similarity score between "AI" and "Artificial Intelligence": 0.98, the similarity score between "AI" and "Art History": 0.20, the similarity score between "AI" and "Aviation Industry": 0.45…; Set a preset similarity score threshold of 0.90. From the above scores, only the score of "Artificial Intelligence" (0.98) is greater than the threshold; According to the selected scores, the block vector of "Artificial Intelligence" [0.09, -0.29, 0.52, …] can be obtained as the first term vector; Further check whether "Artificial Intelligence" is semantically consistent with the context of "AI". For example, if the context is about computer science, then "Artificial Intelligence" is appropriate, and it can also be checked whether "Artificial Intelligence" has multiple meanings. In this example, it has no obvious multiple meanings. Therefore, after calibration processing, it is determined that the block vector of "Artificial Intelligence" [0.09, -0.29, 0.52, …] is the corresponding term vector of "AI"; Output "Artificial Intelligence" as the full-term of "AI".
[0074] It can be seen that in this embodiment, it is ensured that the target word embedding vector can be accurately matched with the term vectors in the database, providing a reliable basis for subsequent natural language processing tasks.
[0075] In one embodiment, the preset processing of the target word embedding vector and all block vectors in the target vector database to obtain multiple similarity scores includes: calculating the target word embedding vector and each block vector in the target vector database respectively to obtain the similarity scores corresponding to the target word embedding vector and each block vector.
[0076] Among them, the calculation process is to calculate the similarity between the target word embedding vector and each term vector in the database using a preset similarity measurement method (such as cosine similarity, Euclidean distance, etc.). Commonly used measurement methods include cosine similarity, Euclidean distance, Manhattan distance, etc. For example, the dot product of two vectors is calculated using cosine similarity, and then divided by the product of their magnitudes. It can be seen that in this embodiment, the similarity degree between the target word embedding vector and each block vector in the database can be quantified as a similarity score, providing a basis for subsequent similarity ranking and screening.
[0077] In one embodiment, the retrieval enhancement generation process for the term vector corresponding to the target word embedding vector to obtain the target term information corresponding to the target abbreviation includes: obtaining first term information according to the term vector corresponding to the target word embedding vector; obtaining first input data according to the first term information and the target abbreviation; performing format conversion processing on the first input data to obtain target input data; inputting the target input data into a preset large language model for retrieval enhancement generation processing to obtain the target term information corresponding to the target abbreviation output by the preset large language model.
[0078] Among them, the first term information refers to the term identified as being most relevant to the target word embedding vector. It can be the text form of the term, ID, database record, etc.
[0079] Among them, in the process of obtaining the first term information according to the term vector corresponding to the target word embedding vector, the term vector corresponding to the target word embedding vector queries the corresponding term information and extracts the corresponding term information to obtain the first term information. That is, each term vector contains the corresponding term information.
[0080] Among them, the first input data refers to the data structure formed by combining the first term information and the target abbreviation, which is used as the input of the large language model. The first input data usually contains sufficient information to indicate the task that the large language model needs to perform, that is, to generate the corresponding term according to the target abbreviation.
[0081] Among them, in the specific implementation of obtaining the first input data according to the first term information and the target abbreviation, the first term information and the target abbreviation are combined into a data structure, such as a string, a dictionary, or a JSON object. This data structure needs to contain sufficient information to indicate the task performed by the preset large language model, that is, to generate the corresponding term according to the abbreviation.
[0082] Among them, format conversion processing refers to the process of converting data from one format to another. Format conversion processing may include, but is not limited to, string operations, data type conversions, JSON serialization / deserialization, and other operations. In this embodiment, the format conversion processing is to convert the first input data into a specific input format acceptable to the large language model, that is, to obtain the target input data.
[0083] Among them, the target input data is the data that meets the input requirements of the large language model after format conversion processing.
[0084] Among them, the preset large language model is a pre-trained large language model with powerful language understanding and generation capabilities, such as GPT-3, GPT-4, etc., and is not limited to this here.
[0085] In a specific implementation, the preset large language model combines the retrieved information and the target input data to generate target term information corresponding to the target acronym.
[0086] It can be seen that in this embodiment, through the first term information and the target acronym, and combined with retrieval-augmented generation, it is possible to give the most appropriate full name or phrase according to the context of the input data, enhancing the understanding and expansion ability of domain terms, and improving the accuracy and practicality of recognition.
[0087] In one embodiment, obtaining the first input data according to the first term information and the target acronym includes: obtaining a preset text format; arranging the first term information and the target acronym according to the preset text format to obtain the first input data.
[0088] Among them, the preset text format is a predefined text structure or template for organizing and arranging information. The preset text format may include different text layouts and styles such as headings, paragraphs, lists, tables, etc. For example, CSV (Comma-Separated Values) format, JSON (JavaScript Object Notation) format, XML (eXtensible Markup Language) format, etc. are all common preset text formats.
[0089] Among them, the obtaining process may refer to obtaining the definition or template of the preset text format from a preset source or storage location. The obtaining process may involve operations such as reading files, accessing databases, and calling APIs.
[0090] Specifically, in the process of arranging the first term information and the target acronym, according to the preset text format, the first term information and the target acronym are placed in specific positions or structures. For example, the first term information is placed in the body paragraph, and the target acronym is used as a heading or label.
[0091] It can be seen that in this embodiment, by presetting the text format, the term information and abbreviations are ensured to be displayed and processed in a clear and structured manner.
[0092] In one embodiment, before performing similarity retrieval on the target word embedding vectors in the target vector database, the method further includes: obtaining preset text data; performing text preprocessing on the preset text data to obtain training phrase groups corresponding to the preset text data; inputting the training phrase groups into the word embedding model for inference to obtain word embedding vectors output by the word embedding model; and storing the word embedding vectors in a preset vector database to obtain a target vector database.
[0093] Among them, the preset text data refers to a large amount of text data collected and prepared in advance, usually from a specific field or multiple fields, and is used to train and generate word embedding vectors. The preset text data can collect relevant text data from a variety of reliable sources (such as publicly available scientific research papers, patent documents, industry standard documents, relevant domain term databases, and open source data sets) in the field, and there is no unique limitation here.
[0094] Among them, text preprocessing refers to a series of cleaning and formatting operations performed on the original text data to eliminate noise and unify the format for subsequent processing.
[0095] Among them, the training phrase groups are a set of words or phrases obtained after text preprocessing and are ready to be input into the model for training.
[0096] Among them, the word embedding model can be a pre-trained pre-training embedding model, such as BERT (Bidirectional Encoder Representations from Transformers), Word2Vec, FastText, etc. BERT has advantages in word understanding and context capture due to its bidirectional encoding ability.
[0097] Among them, the word embedding vectors are vectors generated by the inference of the word embedding model, representing the positions of words or phrases in the vector space, and capturing the semantic information of the words.
[0098] Among them, the preset vector database is a database for storing word embedding vectors, which can be a relational database, a NoSQL database, etc.; the target vector database is a database containing all the word embedding vectors corresponding to the preset text data after the storage operation, and is used for subsequent retrieval and matching tasks.
[0099] Specifically, the process of vector storage is to store the generated word embedding vectors in a database, including the storage of indexes and metadata, for subsequent retrieval and analysis.
[0100] For example, assume there is a preset text dataset on computer science; collect text data from papers, books, websites, etc. in the field of computer science; clean the collected text (remove HTML tags, special symbols, etc.), tokenize it (split the text into independent words), remove stop words (such as "de", "shi", etc.), and lemmatize it (unify "computer" into "computer" instead of variants like "компьютér"); input the preprocessed training phrases into a pre-trained Word2Vec model, and the model generates a word embedding vector for each phrase by learning the relationships between phrases; store the generated word embedding vectors and their corresponding phrase information in a preset vector database, such as using the Faiss vector database to store vector values.
[0101] It can be seen that in this embodiment, the preset text data is effectively converted into structured word embedding vectors and stored in the target vector database, ensuring the quality of the data and the effectiveness of the model.
[0102] In one embodiment, the preprocessing of the preset text data to obtain the target phrases corresponding to the preset text data includes: performing a tokenization process on the preset text data to obtain a plurality of first tokenized phrases; performing a denoising process on the plurality of first tokenized phrases to obtain the second tokenized phrases after denoising; performing a part-of-speech process on the second tokenized phrases to obtain the training phrases corresponding to the preset text data.
[0103] Among them, the tokenization process refers to the process of splitting continuous text into independent words or phrases according to certain rules. In specific implementations, natural language processing (NLP) tokenization algorithms (such as Jieba tokenization, word_tokenize in NLTK, SpaCy, etc.) can be used to split the text into individual words or phrases, and split long text into independent words or phrases for subsequent analysis and processing.
[0104] Among them, the first tokenized phrases are a set of words or phrases obtained through preliminary tokenization, which may contain noise or irrelevant content.
[0105] Among them, the denoising process refers to the process of removing noise, irrelevant or redundant information in the tokenization results to improve data quality.
[0106] Specifically, define noise recognition rules, such as removing punctuation marks, numbers, special characters, etc.; traverse the first tokenized phrases and apply the noise recognition rules; remove the identified noise content to obtain a clean tokenization result, that is, the second tokenized phrases after denoising.
[0107] Among them, the second segmented phrase refers to a set of clean words or phrases after denoising processing.
[0108] Among them, part-of-speech processing refers to the process of performing part-of-speech tagging on the segmented results, that is, identifying the grammatical attributes of each word (such as noun, verb, adjective, etc.).
[0109] Specifically, select a part-of-speech tagging tool or algorithm, such as Stanford CoreNLP, spaCy, etc.; input the denoised second segmented phrase into the part-of-speech tagging tool; the part-of-speech tagging tool assigns corresponding part-of-speech tags to each word according to grammatical rules and models.
[0110] Among them, the phrase to be trained refers to the final set of words or phrases after word segmentation, denoising, and part-of-speech processing, which is ready for model training or inference.
[0111] For example, assume there is a preset text data: "The central processing unit (CPU) is the core component of a computer."; use the jieba word segmentation tool to split the text into the first segmented phrases such as "central", "processing unit", "CPU", "is", "computer", "of", "core", "component"; identify and remove punctuation marks such as "(", ")", etc., to obtain the denoised second segmented phrases: "central", "processing unit", "CPU", "is", "computer", "core", "component"; use Stanford CoreNLP for part-of-speech tagging to obtain the following phrases to be trained: "central" (adjective), "processing unit" (noun), "CPU" (noun), "is" (verb), "computer" (noun), "core" (noun), "component" (noun).
[0112] Optionally, the phrases to be trained corresponding to the preset text data can be identified and audited as follows: use the named entity recognition (NER) algorithm to calibrate important entity information such as proper nouns, locations, and times in this field. At the same time, manually review and correct and supplement professional nouns to ensure high accuracy and comprehensiveness of term recognition.
[0113] It can be seen that in this embodiment, through word segmentation, denoising, and part-of-speech processing, the preset text data is systematically transformed into structured phrases to be trained, providing high-quality input for subsequent model training, further ensuring the data cleanliness and semantic accuracy, and supporting more effective machine learning and natural language processing tasks.
[0114] It should be noted that in the above various embodiments, there is not necessarily a certain order between the above steps. Those of ordinary skill in the art can understand from the description of the embodiments of the present application that in different embodiments, the above steps can have different execution orders, that is, they can be executed in parallel, or exchanged, etc.
[0115] As another aspect of the embodiments of the present application, the embodiments of the present application provide an abbreviation recognition device. Among them, the abbreviation recognition device can be a software module, and the software module includes several instructions, which are stored in a memory, and a processor can access the memory and call the instructions for execution to complete the abbreviation recognition method described in the above various embodiments.
[0116] See Figure 2 , Figure 2 is a structural schematic diagram of an abbreviation recognition device provided by an embodiment of the present application. As Figure 2 shown, the abbreviation recognition device 200 includes:
[0117] An acquisition unit 201, configured to acquire a target abbreviation;
[0118] A training unit 202, configured to input the target abbreviation into a word embedding model for inference to obtain a target word embedding vector output by the word embedding model;
[0119] A retrieval unit 203, configured to perform a similarity retrieval on the target word embedding vector in a target vector database to obtain a term vector corresponding to the target word embedding vector;
[0120] A processing unit 204, configured to perform a retrieval enhancement generation process on the term vector corresponding to the target word embedding vector to obtain target term information corresponding to the target abbreviation.
[0121] In this embodiment, by inputting the target abbreviation into the word embedding model for inference, the corresponding word embedding vector can be obtained more accurately, thereby improving the accuracy of abbreviation recognition; further, by using the target word embedding vector output by the word embedding model for similarity retrieval, the semantic association between the abbreviation and the full-term term can be better captured, enhancing the semantic understanding ability of domain-specific terms; performing similarity retrieval in the target vector database can find the corresponding term vector more quickly compared with the traditional database retrieval method, improving the retrieval efficiency; through efficient similarity retrieval and retrieval enhancement generation processing, the target term information corresponding to the target abbreviation can be output in real time to meet the requirements of real-time processing; further, by performing a retrieval enhancement generation process on the term vector corresponding to the target word embedding vector, more professional abbreviations and contractions in specific fields can be covered, improving the matching accuracy and reducing the situation of incorrect matching.
[0122] In one embodiment, in the target vector database, when performing a similarity search on the target word embedding vector to obtain the term vector corresponding to the target word embedding vector, the retrieval unit 203 is further configured to: perform a preset process on the target word embedding vector and all block vectors in the target vector database to obtain a plurality of similarity scores; screen out at least one target similarity score greater than a preset similarity score threshold from the plurality of similarity scores; obtain a first term vector according to the at least one target similarity score; perform a calibration process on the first term vector, and determine that the first term vector that passes the calibration is the term vector corresponding to the target word embedding vector.
[0123] In one embodiment, in the step of performing a preset process on the target word embedding vector and all block vectors in the target vector database to obtain a plurality of similarity scores, the retrieval unit 203 is further configured to: calculate the target word embedding vector and each block vector in the target vector database respectively to obtain the similarity score corresponding to the target word embedding vector and each block vector.
[0124] In one embodiment, in the step of performing a retrieval enhancement generation process on the term vector corresponding to the target word embedding vector to obtain the target term information corresponding to the target abbreviation, the processing unit 204 is further configured to: obtain first term information according to the term vector corresponding to the target word embedding vector; obtain first input data according to the first term information and the target abbreviation; perform a format conversion process on the first input data to obtain target input data; input the target input data into a preset large language model for a retrieval enhancement generation process to obtain the target term information corresponding to the target abbreviation output by the preset large language model.
[0125] In one embodiment, in the step of obtaining first input data according to the first term information and the target abbreviation, the processing unit 204 is further configured to: obtain a preset text format; arrange the first term information and the target abbreviation according to the preset text format to obtain first input data.
[0126] In one embodiment, before performing a similarity search on the target word embedding vector in the target vector database, the processing unit 204 is further configured to: obtain preset text data; perform text preprocessing on the preset text data to obtain a training phrase group corresponding to the preset text data; input the training phrase group into the word embedding model for inference to obtain the word embedding vector output by the word embedding model; store the word embedding vector in a preset vector database to obtain a target vector database.
[0127] In one embodiment, in the process of performing text preprocessing on the preset text data to obtain the target phrase corresponding to the preset text data, the processing unit 204 is further configured to: perform word segmentation on the preset text data to obtain a plurality of first word segmentation phrases; perform noise reduction processing on the plurality of first word segmentation phrases to obtain second word segmentation phrases after noise reduction; and perform part-of-speech processing on the second word segmentation phrases to obtain the phrase to be trained corresponding to the preset text data.
[0128] It should be noted that the above-mentioned abbreviation recognition device can execute the abbreviation recognition method provided by the embodiments of the present application, and has corresponding functional modules and beneficial effects for executing the method. For technical details not described in detail in the embodiments of the abbreviation recognition device, reference can be made to the abbreviation recognition method provided by the embodiments of the present application.
[0129] See Figure 3 , Figure 3 is a schematic structural diagram of a computer device provided by an embodiment of the present application. As Figure 3 shown, the computer device 300 includes a processor 301 and a memory 302. The processor 301 is communicatively connected to the memory 302.
[0130] The processor 301 is configured to support the computer device to execute the corresponding functions in the abbreviation recognition method in the above-mentioned method embodiments. The processor 301 may be a central processing unit (CPU), a network processor (NP), a hardware chip, or any combination thereof. The above-mentioned hardware chip may be an application specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The above-mentioned PLD may be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL), or any combination thereof.
[0131] Specifically, the processor 301 may include a sending card, a receiving card, and a driving chip.
[0132] The memory 302 is used to store program codes and the like. The memory 302 may include volatile memory (VM), such as random access memory (RAM); the memory 302 may also include non-volatile memory (NVM), such as read-only memory (ROM), flash memory, hard disk drive (HDD) or solid-state drive (SSD); the memory 302 may further include a combination of the above types of memories.
[0133] An embodiment of the present application further provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer is caused to execute the abbreviation recognition method as described in the foregoing embodiments.
[0134] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it may include the processes of the embodiments of the above methods. Among them, the storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.
[0135] The foregoing disclosure is only for the preferred embodiments of the present application, and of course cannot be used to limit the scope of rights of the present application. Therefore, equivalent changes made according to the claims of the present application still fall within the scope covered by the present application.
Claims
1. A method for identifying abbreviations, characterized in that, Including: Obtain a target abbreviation; Input the target abbreviation into a word embedding model for inference to obtain a target word embedding vector output by the word embedding model; In a target vector database, perform a similarity search on the target word embedding vector to obtain a term vector corresponding to the target word embedding vector; Perform retrieval enhancement generation processing on the term vector corresponding to the target word embedding vector to obtain target term information corresponding to the target abbreviation.
2. The method according to claim 1, wherein The step of performing a similarity search on the target word embedding vector in the target vector database to obtain a term vector corresponding to the target word embedding vector includes: Perform a preset process on the target word embedding vector and all block vectors in the target vector database to obtain a plurality of similarity scores; Among the plurality of similarity scores, filter out at least one target similarity score greater than a preset similarity score threshold; According to the at least one target similarity score, obtain a first term vector; Perform a calibration process on the first term vector, and determine that the first term vector that passes the calibration is the term vector corresponding to the target word embedding vector.
3. The method according to claim 2, wherein The step of performing a preset process on the target word embedding vector and all block vectors in the target vector database to obtain a plurality of similarity scores includes: Calculate the target word embedding vector and each block vector in the target vector database respectively to obtain a similarity score corresponding to the target word embedding vector and each block vector.
4. The method according to claim 1, wherein The step of performing retrieval enhancement generation processing on the term vector corresponding to the target word embedding vector to obtain target term information corresponding to the target abbreviation includes: According to the term vector corresponding to the target word embedding vector, obtain first term information; According to the first term information and the target abbreviation, obtain first input data; Perform format conversion processing on the first input data to obtain target input data; Input the target input data into a preset large language model for retrieval enhancement generation processing to obtain the target term information corresponding to the target abbreviation output by the preset large language model.
5. The method according to claim 4, wherein The step of obtaining first input data according to the first term information and the target abbreviation includes: Obtain a preset text format; Arrange the first term information and the target abbreviation according to the preset text format to obtain first input data.
6. The method according to claim 1, wherein Before performing a similarity search on the target word embedding vector in the target vector database, the method further includes: Obtain preset text data; Perform text preprocessing on the preset text data to obtain a training phrase group corresponding to the preset text data; Input the training phrase group into the word embedding model for inference to obtain a word embedding vector output by the word embedding model; Store the word embedding vector in a preset vector database to obtain a target vector database.
7. The method according to claim 6, wherein The step of performing text preprocessing on the preset text data to obtain a target phrase group corresponding to the preset text data includes: Perform word segmentation processing on the preset text data to obtain a plurality of first word segmentation phrases; Perform denoising processing on the plurality of first word segmentation phrases to obtain second word segmentation phrases after denoising; Perform part-of-speech processing on the second participle phrase to obtain the training phrase corresponding to the preset text data.
8. An acronym recognition device, the device comprising: An acquisition unit for acquiring a target acronym; A training unit for inputting the target acronym into a word embedding model for inference to obtain a target word embedding vector output by the word embedding model; A retrieval unit for performing similarity retrieval on the target word embedding vector in a target vector database to obtain a term vector corresponding to the target word embedding vector; A processing unit for performing retrieval enhancement generation processing on the term vector corresponding to the target word embedding vector to obtain target term information corresponding to the target acronym.
9. A computer device, characterized in that, Comprising a memory and a processor, the memory is connected to the processor, and the processor is configured to execute one or more computer programs stored in the memory. When the processor executes the one or more computer programs, the computer device implements the acronym recognition method according to any one of claims 1-7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program includes program instructions, and when the program instructions are executed by a processor, the processor executes the acronym recognition method according to any one of claims 1-7.