Information retrieval method

Abstract

The invention relates to the field of natural language processing and the technical field of deep learning, and discloses an information retrieval method which comprises the following steps: S1, information data conversion and cleaning: converting data in different formats into data in a unified text format, and cleaning the converted text; s2, information extraction: extracting information of the text; and S3, data storage and index construction: storing the information of the intelligence data into a database, and establishing an index for the information. According to the method, the problems of low retrieval comprehensiveness and accuracy and the like in the prior art are solved.

Description

Technical Field

[0001] The present invention relates to the fields of natural language processing and deep learning technology, and specifically to an intelligence retrieval method. Background Art

[0002] In the era of big data, open source websites have generated massive amounts of multi-source scientific and technological intelligence data, which have become core resources in the field of science and technology. Scientific and technological intelligence data is diverse and has inconsistent structures, making it difficult to store and retrieve directly.

[0003] The existing technology still has some shortcomings when dealing with the retrieval of multi-source scientific and technological intelligence data. First, the text structure of multi-source scientific and technological intelligence data is diverse. After converting the data into a unified structure, the existing technology fails to effectively evaluate the conversion quality and clean up low-quality data. Secondly, information extraction includes the extraction of abstracts and keywords, which can effectively summarize the content of scientific and technological intelligence. For abstract extraction and keyword extraction, most existing studies use different algorithms; in addition, text content in different languages ​​also needs to be processed using different algorithms. Finally, in terms of data storage and retrieval, MongoDB is a commonly used text data storage tool that helps users find the required data through full-text retrieval technology. However, its core component, the word segmenter, is often difficult to accurately segment when processing intelligence text in the field of science and technology, resulting in low query accuracy. In order to improve the accuracy of scientific and technological intelligence retrieval, it is necessary to further analyze and solve the shortcomings of existing methods. Summary of the invention

[0004] In order to overcome the deficiencies of the prior art, the present invention provides an information retrieval method to solve the problems of the prior art such as low comprehensiveness and accuracy of retrieval.

[0005] The technical solution adopted by the present invention to solve the above problems is:

[0006] An information retrieval method comprises the following steps:

[0007] S1, intelligence data conversion and cleaning: convert data of different formats into unified text format data, and clean the converted text;

[0008] S2, information extraction: extracting information from text;

[0009] S3, data storage and index construction: save the intelligence data information into the database and index the information.

[0010] As a preferred technical solution, in step S1, when cleaning the converted text, first use Python to write rules to automatically count the data length, and delete the blank text with a length of 0; then use a box plot to identify and clean the garbled English text; and use the Jieba algorithm to segment the garbled Chinese text, remove stop words, and use regular expressions to remove punctuation marks, and clean up the abnormal text whose average word length is lower than the threshold.

[0011] As a preferred technical solution, in step S2, the summary and keyword information of the text are extracted through ChatGLM.

[0012] As a preferred technical solution, in step S2, based on an input text x=[x1,…x n ]Get multiple text spans {s1,…s m}, each span is replaced with a token, forming a mangled text x corrupt , randomly arrange multiple text spans of the input text to obtain z, and the set of all possible arrangements is Z m , the training formula of ChatGLM is formula (1):

[0013]

[0014] Among them, x1,…x n Represent the 1st to nth input characters respectively, i represents the text span number, s1,…s m Respectively represent the 1st to mth text spans, represents the maximization objective function, θ represents the model parameters, and p θ (·) represents the conditional probability associated with θ, Represents the expectation of all possible span arrangements z, Z m represents a set of span permutations, s z＜i express The span of previous predictions, Indicates that at a given x corrupt and S z＜i Predict the i-th span under the condition The logarithmic conditional probability of .

[0015] As a preferred technical solution, in step S3, the information of the intelligence data is saved in a database, and ascending indexes are established for the title, keyword, and abstract respectively.

[0016] As a preferred technical solution, in step S3, a domain dictionary is constructed through three steps: ChatGLM and Qwen reasoning word segmentation, TF-IDF evaluation of word importance, and expansion of the vocabulary.

[0017] As a preferred technical solution, when using ChatGLM and Qwen to infer word segmentation, the Jaccard similarity is used to calculate the consistency of the word segmentation results of the two models to obtain the initial domain dictionary: the word segmentation results with a Jaccard threshold equal to 1 are directly retained; the word segmentation results with a Jaccard threshold less than 1 and greater than or equal to 0.8 are manually processed and one is selected as the word segmentation result; the word segmentation results with a Jaccard threshold less than 0.8 are manually processed; the calculation formula of Jaccard similarity is formula (2):

[0018]

[0019] Among them, for a corpus, A is the word set obtained by ChatGLM reasoning segmentation, B is the word set obtained by Qwen reasoning segmentation, Jaccard(A,B) represents the Jaccard similarity between set A and set B, |A| represents the number of elements in set A, |B| represents the number of elements in set B, |A∩B| represents the number of elements in the intersection of set A and set B, and |A∪B| represents the number of elements in the union of set A and set B.

[0020] As a preferred technical solution, the TF-IDF algorithm is used to process the initial domain dictionary, and key words with TF-IDF values ​​higher than the set threshold are retained. The calculation formula of the TF-IDF algorithm is as follows:

[0021]

[0022] Among them, i represents the number of the entry, j represents the number of the document, and k represents the number of the entry, belonging to document d j A kind of word in Chinese, n i,j Indicates the term t i In the document d j The number of occurrences in n k,j Indicates the term t k In the document d j The number of occurrences in k n k,j Represents document d j The total number of words, TF ij Indicates the term t i In the document d j The frequency of occurrence in

[0023]

[0024] Among them, |D| represents the number of all documents, |{j:t i ∈d j}| indicates that the term t is included i the number of

[0025] TF-IDF=TF*IDF (5)

[0026] Among them, TF-IDF represents the importance of a word.

[0027] As a preferred technical solution, in step S3, the English text is directly saved in the database, the Chinese text is segmented using a domain dictionary and saved in the database, and a text index is established for the text.

[0028] As a preferred technical solution, after establishing the ascending index and the text index, search logic symbols and operators are introduced to search for various index combinations.

[0029] As a preferred technical solution,

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] (1) The present invention ensures the consistency and quality of data by converting and cleaning multi-source scientific and technological intelligence data;

[0032] (2) The present invention utilizes ChatGLM to extract keywords and abstracts from data, and on this basis saves the text information into a database and further establishes multiple indexes. After establishing multiple indexes, they are combined with search logic symbols and operators to achieve comprehensive and accurate retrieval of scientific and technological intelligence data. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Establish a flow chart for scientific and technological information retrieval methods;

[0034] Figure 2 It is a data conversion flow chart;

[0035] Figure 3 A flowchart for data cleaning;

[0036] Figure 4 It is a schematic diagram of the box plot structure;

[0037] Figure 5 Build a flow chart for data preservation and indexing;

[0038] Figure 6 Constructing a flow chart for a science and technology domain dictionary. DETAILED DESCRIPTION

[0039] The present invention will be further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0040] Example 1

[0041] like Figures 1 to 6As shown, the present invention provides an efficient scientific and technological intelligence retrieval method to convert, clean and extract information from data, save it in a database with an appropriate data structure, and design efficient indexing and retrieval strategies to ensure that users can quickly and accurately retrieve the required information, thereby providing strong support for technical decision-making and strategic planning.

[0042] The purpose of the present invention is to propose a scientific and technological intelligence retrieval method based on a large language model, aiming to improve the comprehensiveness and accuracy of multi-source scientific and technological intelligence data retrieval.

[0043] The present invention adopts a unified method to process text content and information extraction tasks, thereby significantly improving the efficiency of information extraction.

[0044] To achieve the above purpose, the present invention first converts and cleans multi-source scientific and technological intelligence data, converts data with diverse text formats into data with unified formats, and proposes a text cleaning method based on statistical rules to clean low-quality data, which is conducive to further analysis and processing. Then, the converted and cleaned data is extracted using a large language model so as to establish an index later, thereby realizing a comprehensive retrieval of scientific and technological intelligence data. Finally, after the scientific and technological intelligence data is converted, cleaned, and information extracted, the results are stored in a MongoDB database, and ascending indexes are established for titles, keywords, and abstracts; a dictionary is constructed based on a large language model to segment the text and establish a text index for the text. Through the constructed multi-category index, it is used in conjunction with search logic symbols and operators to achieve comprehensive and accurate retrieval of scientific and technological intelligence materials. The present invention extracts information from scientific and technological intelligence data based on a large language model, so that users can retrieve the required information more comprehensively, and constructs a science and technology field dictionary based on a large language model and a TF-IDF algorithm to establish a text index for the text, so that users can retrieve full-text information more accurately.

[0045] First, convert multi-source scientific and technological intelligence data into unified txt format data and clean up low-quality data. Then, use the open source large language model ChatGLM to extract keywords and abstracts from the text data to facilitate subsequent indexing and achieve comprehensive retrieval of the data. Finally, save the title, keywords, and abstract to the database, and establish ascending indexes for each. Build a science and technology dictionary based on the large language model, segment the text and save it to the database, and then establish a text index to achieve full-text retrieval. Users can combine the above-mentioned multiple indexes containing rich information to achieve comprehensive and accurate retrieval of scientific and technological intelligence. The specific process is as follows: Figure 1 shown.

[0046] Data conversion and cleaning. Multi-source scientific and technological intelligence data has various formats, including pdf, docx, html, caj, etc. It needs to be converted into a unified text (txt) format to eliminate the problem of information being difficult to use effectively due to inconsistent data formats. After data conversion, it is also necessary to evaluate the data quality and clean up low-quality data, such as blank data and garbled data that failed to be converted, so as to ensure the quality of the converted scientific and technological intelligence data set.

[0047] Information extraction part. After data conversion and cleaning, information is extracted from the data so that a variety of indexes containing rich document information can be established later to achieve comprehensive retrieval of scientific and technological intelligence data. Information extraction includes keyword extraction and abstract extraction. It uses an open source large language model to extract keywords and abstracts from the data at the same time through model dialogue reasoning.

[0048] Data preservation and index building part. Save the title and extracted keywords and abstracts to MongoDB, and create ascending indexes for each. In addition, it is necessary to save the body text after data conversion to MongoDB and create a text index. Inverted index is a typical index structure commonly used in text indexing, which can greatly improve the speed of full-text retrieval. By constructing multiple types of indexes, combined with search logic symbols and operators, the required scientific and technological intelligence information can be fully and accurately retrieved.

[0049] For multi-source scientific and technological intelligence data conversion and cleaning. When processing multi-source scientific and technological intelligence data, the present invention solves the challenge of inconsistent multi-source data formats by converting data in multiple formats such as pdf, docx, html, caj, etc., ensuring that data from different sources can be analyzed and effectively used under a unified framework. In addition, the converted data is cleaned, using box plots to clean English data, and using jieba word segmentation and statistical methods to clean Chinese data. Data cleaning can identify and correct incomplete data and outliers, improve data quality, and thus ensure the efficiency and credibility of subsequent analysis.

[0050] For information extraction, we use the open source large language model ChatGLM to extract the abstracts and keywords of scientific and technological intelligence data. The information extraction method is extractive rather than generative, which ensures the authenticity of the extracted information. The extracted abstracts and keywords provide rich text information for the subsequent establishment of multiple indexes, providing strong support for more comprehensive and multi-angle data retrieval.

[0051] For data preservation and index construction. The title, keywords, abstract, text, source file address and other information of the intelligence data are saved in MongoDB, and ascending indexes are established for the title, keywords and abstract respectively, and a text index is established for the text. The establishment of multiple indexes greatly improves the breadth and depth of data retrieval. A text index is directly constructed for the English text, and a text index is constructed after the Chinese text is segmented using a science and technology dictionary. The present invention constructs a science and technology dictionary through three steps of ChatGLM and Qwen reasoning segmentation, TF-IDF evaluation of word importance, and expansion of the vocabulary, thereby ensuring the accuracy of the text index.

[0052] A scientific and technological intelligence retrieval method based on a large language model includes three aspects: multi-source scientific and technological intelligence data conversion and cleaning, information extraction, data preservation and index construction.

[0053] The so-called multi-source scientific and technological intelligence data conversion and cleaning refers to the conversion of data in various formats such as pdf, docx, html, caj, etc. into a unified text format, and further using box plots and statistical methods to clean the converted text to ensure data consistency and quality.

[0054] Information extraction refers to extracting text summaries and keyword information through ChatGLM to support the subsequent construction of a comprehensive index.

[0055] The so-called data preservation and index construction refers to saving the title, keywords, abstract, text, source file address and other information of scientific and technological intelligence data into the database, and establishing ascending indexes for the title, keywords and abstract respectively, and text indexes for the text. The establishment of multiple indexes has greatly improved the comprehensiveness and accuracy of data retrieval.

[0056] The following steps are involved:

[0057] S1. Transformation and cleaning of multi-source scientific and technological intelligence data;

[0058] S2, information extraction;

[0059] S3, data storage and index construction.

[0060] Preferably, data in various formats such as pdf, docx, html, caj, etc. are converted into a unified text format, and the converted text is further cleaned using a box plot, jieba algorithm and statistical methods, specifically including the following steps:

[0061] S11. Multi-source scientific and technological intelligence data collection, the collected data types include pdf, docx, html, caj and other file formats.

[0062] S12. For the data set obtained in step S11, a program is written using Python to convert the multi-source data into unified text data.

[0063] S13. For the data set formed in step S12, a text cleaning method based on statistical rules is proposed. First, Python is used to write rules to automatically count the length of data, and blank text with a length of 0 is deleted. Then, for garbled English text, a box plot is used to identify anomalies and clean them. Finally, for garbled Chinese text, the Jieba algorithm is used to segment the text, remove stop words, and use regular expressions to remove punctuation marks, and to clean abnormal texts whose average word length is lower than the threshold.

[0064] Preferably, extracting abstracts and keywords from scientific and technological information specifically includes the following steps:

[0065] S21, for the data set cleaned in step S13, the text summary and keywords are extracted simultaneously using ChatGLM reasoning, and the extraction method is extractive.

[0066] Preferably, the title, keywords, abstract, text, source file address and other information of the intelligence data are saved in the database, and appropriate indexes are constructed respectively, combined with search logic symbols and operators, so as to achieve comprehensive and accurate retrieval of scientific and technological intelligence data.

[0067] S31. Save the title, keywords, summary, source file address and other information of the intelligence data into the database, and build an ascending index for the title, keywords and summary.

[0068] S32. Construct a special dictionary in the field of science and technology. Construct a dictionary in the field of science and technology through three steps: ChatGLM and Qwen reasoning word segmentation, TF-IDF evaluation of word importance, and expansion of the vocabulary.

[0069] S33. The English text is directly saved in the database, and the Chinese text is segmented into words using a dictionary in the field of science and technology and saved in the database, and a text index is further constructed for the text.

[0070] S34. Combining search logic symbols and operators together greatly improves the breadth and depth of scientific and technological intelligence data retrieval.

[0071] The invention discloses a scientific and technological intelligence retrieval method based on a large language model. The system first transforms and cleans multi-source scientific and technological intelligence data, then uses ChatGLM to extract abstracts and keyword information, and finally saves the multi-dimensional information of the text into a database and constructs an ascending index. When saving the Chinese text of the text, ChatGLM, Qwen and TF-IDF are used to construct a domain dictionary to segment the text before saving it, and further construct a text index for the Chinese and English texts. Using multiple ascending indexes and text indexes in combination with retrieval logic symbols and operators greatly improves the comprehensiveness and accuracy of scientific and technological intelligence data retrieval.

[0072] Example 2

[0073] like Figures 1 to 6 As shown, as a further optimization of Example 1, based on Example 1, this embodiment also includes the following technical features:

[0074] Conversion and cleaning of multi-source scientific and technological intelligence data. Scientific and technological intelligence data contains multiple formats such as pdf, docx, html, caj, etc., which need to be converted into unified text (txt) format data. Use Python to write a program, use Python libraries such as pdfplumber, win32com, BeautifulSoup and fitz, etc., to convert multi-source data into unified text data. The specific process is as follows Figure 2 shown.

[0075] After data conversion, it is necessary to evaluate the quality of the converted text data and clean up low-quality text. The generated low-quality text includes blank text where the conversion failed, and text with a large amount of garbled data. The present invention proposes a text cleaning method based on statistical rules. The specific process is as follows: Figure 3 As shown. For blank text, we use Python to write rules to automatically count the length of the text and delete blank text with a length of 0. For text with a large amount of garbled data, we adopt different solutions based on the different nature of English text and Chinese text. For a very small amount of text in other languages, we do not make quality judgments due to cost considerations.

[0076] English text cleaning. For English text, since words are composed of multiple letters, only part of the letters in the successful word will be converted during the conversion, resulting in word errors and shortened lengths, such as "mohd" and "dhe". Therefore, 10,000 pieces of information were randomly selected from scientific and technological intelligence materials, and the sum of the lengths of the first 500 words in each piece of information was counted. Similarly, the sum of the lengths of the first 500 words in each converted text was counted, and a box plot was used to determine whether it was garbled, thereby cleaning up abnormal data. The box plot details are as follows: Figure 4As shown in the figure. The box plot is a commonly used tool for data cleaning, which can quickly identify abnormal data and contains 7 important parameters, namely: the lower quartile Q1, the median Q2, the upper quartile Q3, the lower limit, the upper limit, the interquartile range IQR, and outliers. Among them, Q1 is the 25% data after sorting all the values of the sample from smallest to largest; Q2 is the 50% data after sorting all the values of the sample from smallest to largest; Q3 is the 75% data after sorting all the values of the sample from smallest to largest; IQR is equal to Q3 - Q1; the lower limit is Q1 - 1.5IQR; the upper limit is Q3 + 1.5IQR; data beyond the upper and lower limits are abnormal data.

[0077] Chinese text cleaning. For Chinese text, words will be converted into character combinations without semantic meaning, such as combinations like "Zi Dai" and "Kuang Lao". Therefore, the common word segmentation algorithm jieba is used to segment the text, remove stop words, and remove punctuation marks using regular expressions. For normal text, the average length of the segmented words is greater than or equal to 2; for abnormal text, since these character combinations cannot be recognized as words, they will be segmented into single characters during word segmentation, and their average length is much less than 2. Segment the first 1000 characters of each text, calculate the average length of the words, set the length threshold to 1.8, and if it is lower than this threshold, clean up this text.

[0078] Information extraction. Large language models have strong language understanding and generation capabilities, can process large-scale natural language data, and have significantly improved natural language processing technologies, including text generation, machine translation, multi-modal integration and other technical problems are easily solved. The present invention uses the large language model ChatGLM to extract information from scientific and technological intelligence data, including abstracts and keywords, to facilitate the establishment of retrieval indexes from multiple perspectives. ChatGLM is an open-source Chinese-English bilingual dialogue model based on the GLM architecture, trained based on the autoregressive blank filling task, introduces two-dimensional position encoding to process variable-length filling tasks, and realizes multi-task learning by changing the number and length of masked segments. It aligns with human intentions through technologies such as supervised fine-tuning, feedback self-help, and RLHF. In addition, the quantization technology can greatly reduce the user deployment threshold. The training process is shown in Equation (1). Given an input text x = [x1,...x n , multiple text spans {s1,...s m} can be sampled, and each span is replaced with a [MASK] token to form a corrupted text x corrupt . In order to fully capture the interdependencies between text spans, their order can be randomly permuted to obtain all possible permutation sets Z m , where: The present invention uses the ChatGLM reasoning method to simultaneously extract keywords and summaries of intelligence data, and limits the large language model information extraction method to an extractive method rather than a generative method, thereby ensuring the authenticity of the information extraction.

[0079]

[0080] Data preservation and index building. The intelligence data is saved in MongoDB. The saved information includes title, keywords, abstract, text, source file address, etc., and further builds a query index. Create ascending indexes for title, keywords, and abstract respectively, and then create a text index for the text. The specific process is as follows: Figure 5 As shown. MongoDB has good support for building text indexes for English data. The English text content is directly saved as a field in the database. The database will automatically segment English words according to spaces to build an index and establish a text index. However, MongoDB lacks support for Chinese data, so it is necessary to pre-segment the Chinese text, use spaces to splice it into a new string and save it in the database. Then the database can build a text index in the same way as it processes English data. When users search, after retrieving data that meets the requirements, they can splice the Chinese text and return the original text data. When segmenting Chinese words, based on the large language model to infer massive scientific and technological intelligence data, combined with the TF-IDF algorithm, a dictionary in the field of science and technology is built to achieve accurate segmentation of texts in this field.

[0081] The process of constructing a dictionary in the field of science and technology is as follows: Figure 6 As shown in the figure, first, 100,000 Chinese texts are randomly selected from the dataset. Two commonly used open source large language models, ChatGLM and Qwen, are used to segment the texts through dialogue reasoning and remove stop words. The Jaccard similarity is used to calculate the consistency of the segmentation of each text. The calculation of the Jaccard similarity is shown in formula (2). The segmentation results with a Jaccard threshold equal to 1 are directly retained; the segmentation results with a Jaccard threshold less than 1 and greater than or equal to 0.8 are manually processed and the one with better effect is selected as the segmentation result; the segmentation results with a Jaccard threshold less than 0.8 are manually processed to obtain the initial domain dictionary. The TF-IDF algorithm is used to process the initial domain dictionary and retain key words with higher TF-IDF values. The idea of ​​the TF-IDF algorithm is that if a word appears more frequently in an article and rarely appears in other articles, the word is more important. The calculation of the TF-IDF algorithm is shown in formulas (3)-(5). The initial domain dictionary is further expanded by adding relevant technology dictionaries of Baidu Input Method and Sogou Input Method, as well as the technology terminology library, to obtain the final technology domain dictionary.

[0082] The calculation of Jaccard similarity is:

[0083]

[0084] Among them, A and B are two sets, |A∩B| represents the number of elements in the intersection of sets A and B, and |A∪B| represents the number of elements in the union of sets A and B. The value of Jaccard similarity is between 0 and 1. The larger the value, the higher the similarity of the two sets. The value is 1 when they are exactly the same, and the value is 0 when there are no common elements.

[0085] The calculation of TF-IDF is:

[0086]

[0087] Among them, i represents the number of the term, j represents the number of the document, k represents the number of the term, and the term belongs to document d j A kind of word in Chinese, n i,j Indicates the term t i In the document d j The number of occurrences in n k,j Indicates the term t k In the document d j The number of occurrences in k n k,j Represents document d j The total number of words, TF ij Indicates the term t i In the document d j The frequency of occurrence in .

[0088]

[0089] Among them, |D| represents the number of all documents, |{j:t i ∈d j}| indicates that the term t is included i The number of

[0090] TF-IDF=TF*IDF (5)

[0091] Among them, the value of TF-IDF is equal to the product of TF and IDF.

[0092] After building the ascending index of title, keyword, abstract and the text index of the main text, we introduce the search logic AND, OR and NOT to search for each index combination. In addition, for a single search item, we support +, - and *. If you need to adjust the order of operations, you can add "()". By building multiple ascending indexes and text indexes and using them with search logic symbols and operators, the comprehensiveness and accuracy of the search are greatly improved.

[0093] As described above, the present invention can be preferably implemented.

[0094] All features disclosed in all embodiments in this specification, or steps in all methods or processes implicitly disclosed, except for mutually exclusive features and / or steps, can be combined and / or expanded or replaced in any manner.

[0095] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. According to the technical essence of the present invention, within the spirit and principles of the present invention, any simple modification, equivalent replacement and improvement made to the above embodiment still falls within the protection scope of the technical solution of the present invention.

Claims

1. An information retrieval method, characterized in that: The following steps are involved: S1, intelligence data conversion and cleaning: convert data of different formats into unified text format data, and clean the converted text; S2, information extraction: extracting information from text; S3, data storage and index construction: save the intelligence data information into the database and index the information.

2. An information retrieval method according to claim 1, characterized in that: In step S1, when cleaning the converted text, first use Python to write rules to automatically count the data length and delete blank text with a length of 0; then use a box plot to identify and clean the garbled English text; and use the Jieba algorithm to segment the garbled Chinese text, remove stop words, and use regular expressions to remove punctuation marks, and clean up abnormal texts whose average word length is lower than the threshold.

3. The information retrieval method according to claim 1, characterized in that: In step S2, the summary and keyword information of the text are extracted through ChatGLM.

4. The information retrieval method according to claim 3, characterized in that: In step S2, based on an input text x=[x1,…x n ]Get multiple text spans {s1,…s m }, each span is replaced with a token, forming a mangled text x corrupt , randomly arrange multiple text spans of the input text to obtain z, and the set of all possible arrangements is Z m , the training formula of ChatGLM is formula (1): Among them, x1,…x n Represent the 1st to nth input characters respectively, i represents the text span number, s1,…s m Respectively represent the 1st to mth text spans, represents the maximization objective function, θ represents the model parameters, and p θ (·) represents the conditional probability associated with θ, represents the expectation of all possible span arrangements z, Z m represents a collection of span permutations, represents the span selected according to the order selected by the permutation z, s z＜i express The span of previous predictions, Indicates that at a given x corrupt and S z＜i Predict the i-th span under the condition The logarithmic conditional probability of .

5. The information retrieval method according to claim 1, characterized in that: In step S3, the information of the intelligence data is saved in the database, and ascending indexes are established for the title, keyword, and abstract respectively.

6. An information retrieval method according to claim 5, characterized in that: In step S3, a domain dictionary is constructed through three steps: ChatGLM and Qwen reasoning word segmentation, TF-IDF evaluation of word importance, and expansion of the vocabulary.

7. An information retrieval method according to claim 6, characterized in that: When using ChatGLM and Qwen to infer word segmentation, the Jaccard similarity is used to calculate the consistency of the word segmentation results of the two models to obtain the initial domain dictionary: the word segmentation results with a Jaccard threshold equal to 1 are directly retained; the word segmentation results with a Jaccard threshold less than 1 and greater than or equal to 0.8 are manually processed and one is selected as the word segmentation result; the word segmentation results with a Jaccard threshold less than 0.8 are manually processed; the calculation formula of Jaccard similarity is formula (2): Among them, for a corpus, A is the word set obtained by ChatGLM reasoning segmentation, B is the word set obtained by Qwen reasoning segmentation, Jaccard(A,B) represents the Jaccard similarity between set A and set B, |A| represents the number of elements in set A, |B| represents the number of elements in set B, |A∩B| represents the number of elements in the intersection of set A and set B, and |a∪B| represents the number of elements in the union of set A and set B.

8. An information retrieval method according to claim 7, characterized in that: The TF-IDF algorithm is used to process the initial domain dictionary and retain key words with TF-IDF values ​​higher than the set threshold. The calculation formula of the TF-IDF algorithm is as follows: Among them, i represents the number of the entry, j represents the number of the document, and k represents the number of the entry, belonging to document d j A kind of word in Chinese, n i,j Indicates the term t i In the document d j The number of occurrences in n k,j Indicates the term t k In the document d j The number of times it appears in k n k,j Represents document d j The total number of words, TF ij Indicates the term t i In the document j The frequency of occurrence in Among them, |D| represents the number of all documents, |{j:t i ∈d j }| means it contains the term d i the number of TF-IDF=TF*IDF (5) Among them, TF-IDF represents the importance of a word.

9. An information retrieval method according to any one of claims 5 to 8, characterized in that: In step S3, the English text is directly saved in the database, the Chinese text is segmented using the domain dictionary and saved in the database, and a text index is established for the text.

10. An information retrieval method according to claim 9, characterized in that: After establishing the ascending index and text index, introduce search logic symbols and operators to search for each index combination.

Images