Academic paper retrieval method and device based on large language model driving and medium

Abstract

The invention discloses an academic paper retrieval method and device based on large language model driving, and a medium, and the method comprises the steps: obtaining a preliminary academic paper retrieval result through retrieval of a plurality of academic paper databases according to an input retrieval request; combining the preliminary academic paper retrieval results, and performing weighted sorting processing according to correlation, popularity and time; inputting the sorted preliminary academic paper retrieval results into a pre-trained first large language model to obtain article abstracts of academic papers and comparative analysis results of similar academic papers; wherein the training process of the first large language model comprises the steps of constructing diversified task samples including abstract generation, question and answer, keyword extraction, logical reasoning, information extraction and literature review generation, and constructing a total loss function so as to train the first large language model; the total loss function is the weighted sum of abstract generation, question and answer, keyword extraction, logical reasoning, information extraction and literature review loss functions.

Description

Technical Field

[0001] The present invention belongs to the technical field of information retrieval and natural language processing, and in particular relates to an academic paper retrieval method, device, and medium driven by a large language model. Background Art

[0002] At present, researchers and practitioners in the academic community are increasingly relying on online platforms to obtain academic papers and other scientific research resources. However, existing academic search platforms have the following shortcomings in terms of search accuracy, recommendation quality, and search result analysis:

[0003] (1) Users use independent search tools on different platforms, which often makes it impossible to effectively integrate search results from different sources;

[0004] (2) The number of search results is usually huge, which requires users to spend a lot of time and energy to screen academic papers that meet their needs. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides an academic paper retrieval method, device, and medium driven by a large language model.

[0006] In a first aspect, an embodiment of the present invention provides an academic paper retrieval method driven by a large language model, the method comprising:

[0007] According to the input search request, preliminary academic paper search results are obtained by searching several academic paper databases;

[0008] Merge the preliminary academic paper search results and sort them by weight according to relevance, popularity and time;

[0009] Input the sorted preliminary academic paper search results into the pre-trained first language model to obtain the article abstracts of the academic papers and the comparative analysis results of similar academic papers;

[0010] Among them, the training process of the first language model includes: constructing diversified task samples including summary generation, question and answering, keyword extraction, logical reasoning, information extraction and literature review generation, constructing a total loss function, and training the first language model through the total loss function; the total loss function is the weighted sum of the summary generation task loss function, the question and answer task loss function, the keyword extraction task loss function, the logical reasoning task loss function, the information extraction task loss function, and the literature review task loss function.

[0011] In a second aspect, an embodiment of the present invention provides an electronic device, comprising a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-mentioned academic paper retrieval method driven by a large language model.

[0012] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned academic paper retrieval method driven by a large language model.

[0013] In a fourth aspect, an embodiment of the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the above-mentioned academic paper retrieval method driven by a large language model.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

[0015] The present invention provides an academic paper retrieval method driven by a large language model, which obtains preliminary academic paper retrieval results by searching several academic paper databases, thereby integrating the results of multiple data sources. The present invention inputs the academic paper retrieval results into a pre-trained first large language model to obtain the article abstracts of the academic paper and the comparative analysis results of similar academic papers. The present invention uses the large language model to efficiently analyze and recommend the retrieval results, which can significantly improve the user experience and retrieval efficiency. At the same time, during the training process of the first large language model, the present invention constructs a variety of task samples including abstract generation, question answering, keyword extraction, logical reasoning, information extraction and literature review generation, so that the large language model can better understand the complex information in the literature. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0017] Figure 1 An academic paper retrieval method based on a large language model driven by an embodiment of the present invention;

[0018] Figure 2 A schematic diagram of sorting preliminary academic paper search results provided by an embodiment of the present invention;

[0019] Figure 3 A schematic diagram of a first language model training process provided by an embodiment of the present invention;

[0020] Figure 4 A schematic diagram of scoring the comparative analysis result output by the first language model provided by an embodiment of the present invention;

[0021] Figure 5 A schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] It should be noted that, in the absence of conflict, the features in the following embodiments and implementations may be combined with each other.

[0024] like Figure 1 As shown, an embodiment of the present invention provides an academic paper retrieval method based on a large language model drive, the method comprising:

[0025] Step S1, based on the input search request, obtain preliminary academic paper search results by searching several academic paper databases.

[0026] Furthermore, the search request includes: target academic paper database, document type, target time range, target subject, and target keywords.

[0027] It should be noted that in this example, preliminary paper results from multiple search sources are obtained through the service APIs of multiple academic paper databases or paper platforms, thus avoiding the problem of incomplete or biased information caused by a single source.

[0028] Step S2, merge the preliminary academic paper retrieval results from different data sources, and perform weighted sorting processing based on relevance, popularity, and time.

[0029] Furthermore, the preliminary academic paper search results include a series of academic papers, including title, author information, conference journal, paper abstract, publication year, keywords, etc.

[0030] Specifically, Figure 2 As shown, the sorting rules of step S2 include but are not limited to:

[0031] According to the keywords and / or themes in the search request, the matching degree with the keywords and / or themes of each academic paper in the preliminary academic paper search results is calculated, so as to set the relevance ranking factor;

[0032] The popularity ranking factor is set according to the number of citations, downloads, and sharing frequency of each academic paper in the preliminary academic paper search results; thereby helping users quickly discover key research results in the target field.

[0033] A time sorting factor is set according to the publication time of each academic paper in the preliminary academic paper search results; thereby ensuring that users can obtain the latest research results first.

[0034] Define sorting rules, perform weighted processing on relevance sorting factor, heat sorting factor, and time sorting factor, so as to complete the sorting of preliminary academic paper retrieval results and improve retrieval efficiency.

[0035] It should be noted that in this example, a custom hybrid sorting rule is used to weight different sorting factors to obtain the order of paper recommendations that best meets its needs; for example, for academic papers that tend to focus on technical applications, a higher weight needs to be given to papers that contain open source code libraries.

[0036] Step S3, performing copyright marking and / or paper type marking on the sorted preliminary academic paper search results; wherein the paper type marking includes: journal paper, conference paper, technical report, preprint.

[0037] It should be noted that in this example, the marking will provide users with prominent prompts (such as through color marking or adding tag identifiers) in the subsequent paper recommendation process. When users download or cite literature, copyright and usage restrictions will be prompted to ensure that users comply with usage requirements. For papers with stricter restrictions (such as paid or non-citeable preprints), alternative literature recommendations can be provided to help users find suitable open access versions.

[0038] Step S4, inputting the sorted preliminary academic paper retrieval results into the pre-trained first language model to obtain the article abstracts of the academic papers and the comparative analysis results of similar academic papers.

[0039] In this example, Alibaba Cloud's Tongyi Qianwen large model qwen2 (Meta's LLAMA3 can also be selected) is used as the base model. By building professional corpus to train the base model, we can obtain the paper analysis large language model used for scientific literature interpretation and analysis, that is, the first large language model that has been trained, denoted as SemanticLLM.

[0040] Among them, Figure 3 As shown, the training process of the first language model includes:

[0041] Step S100, select academic papers in several scientific fields (such as chemistry, biology, physics, engineering, etc.) to ensure that the model has a broad understanding of interdisciplinary scientific terms and concepts. Constructing a variety of task samples including summary generation, question answering, keyword extraction, logical reasoning, information extraction, and literature review generation helps the model adapt to different literature analysis needs and enables the large language model to better understand the complex information in the literature.

[0042] Specifically, the process of constructing diverse task samples includes:

[0043] Constructing a summary generation task sample for the summary generation task; the summary generation task sample includes an academic paper as input and a summary as output;

[0044] It should be noted that the content of the summary generation task sample should include the main arguments, research background, key methods and conclusions of the document, so that the model can identify and extract core information when processing the summary generation task. The training data can contain a variety of document types, such as journal articles, conference papers, technical reports, etc., to ensure that the model can handle different formats and writing styles.

[0045] For example, a sample summary generation task is:

[0046] Enter the full text of the document:

[0047] “1 Introduction

[0048] According to the global satellite mission statistics of CEOS (Committee on Earth Observation Satellites), since the successful launch of the first US meteorological satellite in 1960, there have been 558 earth observation satellites in Earth orbit as of the end of September 2020. Humanity has accumulated more than half a century of global-scale historical remote sensing data. Remote sensing technology has developed rapidly, and technical indicators such as spatial resolution, temporal resolution, and spectral resolution have been continuously improved. The acquisition of remote sensing data has also tended to be multi-platform, multi-sensor, and multi-angle. The continuous improvement and increase of remote sensing platforms and sensors have led to a rapid increase in the amount of various remote sensing data. At the data level, it has reflected the "5V" characteristics of large volume (Volume), variety (Variety), fast change (Velocity), accuracy (Veracity) and high value (Value). We have entered an unprecedented era of remote sensing big data. Compared with traditional remote sensing data with low spatial and temporal resolution or sparse and discontinuous spatial and temporal sampling, remote sensing big data has many advantages: (1) it pursues high-resolution, spatial and temporal continuous global data rather than random or sparse sampling, (2) it is more conducive to insight into macro-level trends rather than local accuracy, and (3) it focuses more on multi-factor correlation analysis to solve problems rather than on constructing more difficult causal relationships. Remote sensing big data can provide more detailed attribute information of land objects, which helps to study the Earth system as a whole…”

[0049] Output summary:

[0050] "Humanity has accumulated global historical remote sensing data for more than half a century. The remote sensing big data formed by these emerging massive remote sensing data provides rich data support for earth science research; rapid processing, analysis and mining of remote sensing big data is a new challenge. The emergence of remote sensing cloud computing platforms has provided unprecedented opportunities for remote sensing big data mining, and has completely changed the mode of traditional remote sensing data processing and analysis, making it possible to quickly analyze and apply long-term series on a global scale. This article systematically sorts out the development status of remote sensing cloud computing platforms at home and abroad, and summarizes the main directions of application of remote sensing cloud computing platforms in the field of earth science so far. On this basis, the limitations of the current remote sensing cloud computing platform are discussed, and the key technologies and core issues that need to be solved in the future are prospected. Suggestions for the development of China's remote sensing cloud computing platform are put forward: As human beings' demand for understanding the earth increases, remote sensing cloud computing platforms will play a greater role in earth science research, serving the deepening of earth science knowledge and the sustainable development of human society."

[0051] Constructing question-answering task samples for question-answering tasks; the question-answering task samples include questions and academic papers as inputs, and answers as outputs;

[0052] It should be noted that by designing question and answer pairs based on the content of the literature, the model learns how to extract and generate answers from the literature. Such task samples can include specific questions such as "What is the research goal of this literature?", "What do the experimental results show?", etc., so that the model can understand and locate the key information in the literature to support the implementation of the automated literature question answering system.

[0053] For example, the sample question answering task is:

[0054] Input question:

[0055] “What is the research goal of this document?”

[0056] Full text of the article:

[0057] “The main goal of this study is to develop an efficient environmental impact assessment model that can quantify the impact of industrial emissions on air quality and provide a scientific basis for policy making,…”

[0058] Output answer:

[0059] “The research objective of this paper is to develop an assessment model to quantify the impact of industrial emissions on air quality.”

[0060] Constructing a keyword extraction task sample for a keyword extraction task; the keyword extraction task sample includes an academic paper as input and a keyword as output;

[0061] It should be noted that the keyword extraction task samples are set to guide the model to identify the core keywords in the document. To improve the accuracy and representativeness of the keywords, manually annotated keywords or keywords preliminarily screened by statistical methods can be used in the training set to let the model learn to screen and extract the words that best reflect the theme of the document during the training process. This task plays an important role in document classification, retrieval and recommendation.

[0062] For example, a sample of a keyword extraction task is:

[0063] Enter the document snippet:

[0064] "This paper proposes a financial market prediction model based on a large language model that can predict future stock price trends using historical data."

[0065] Output keywords:

[0066] "Financial markets, machine learning, large language models, stock prices"

[0067] Constructing a logical reasoning task sample for a logical reasoning task; the logical reasoning task sample includes an academic paper as input and a reasoning result as output;

[0068] It should be noted that the logical reasoning task samples enable the model to handle complex logic such as causal relationships and hypothetical reasoning in the literature. Logical reasoning tasks can include text inference, premise hypothesis verification, etc. By marking the causal chain, logical relationship or reasoning conclusion in the literature, the model can generate logical explanations and conclusions in literature analysis.

[0069] For example, a sample of logical reasoning tasks is:

[0070] Enter the document snippet:

[0071] "The experimental results showed that the growth rate of plants treated with organic fertilizers was significantly higher than that of the chemical fertilizer group, which may be because organic fertilizers contain more trace elements."

[0072] Output the inference results:

[0073] “Organic fertilizers contain more trace elements, which may be the reason why they increase plant growth rate.”

[0074] Constructing information extraction task samples for information extraction tasks; the information extraction task samples include academic papers as input and target information as output; the target information includes experimental methods, experimental data and / or research conclusions;

[0075] It should be noted that the information extraction task samples focus on extracting target types of information from the literature, such as experimental methods, data, research conclusions, etc. The implementation of such tasks can help the model learn to identify and extract these key information by annotating the corresponding data and conclusion parts in the literature, thereby providing users with rapid information summary and analysis.

[0076] For example, the information extraction task sample examples are:

[0077] Enter the document snippet:

[0078] "The experimental results show that the autonomous driving system using the new sensor has significantly improved traffic accident prevention, reducing the accident rate by 30%."

[0079] Output:

[0080] “Experimental data: accident rate reduced by 30%. Main conclusion: new sensors improve the safety of autonomous driving systems.”

[0081] Constructing a literature review task sample for the literature review task; the literature review task sample includes a number of academic papers as input and a review as output;

[0082] It should be noted that the literature review task sample enables the model to conduct a comprehensive analysis of multiple documents and extract common research themes, methodological progress and trends. This task requires the model to be able to find related information in multiple documents and form a well-structured review to help researchers understand the overall picture of a research field.

[0083] For example, a sample literature review task example is:

[0084] Enter a collection of document snippets:

[0085] Reference 1: Full text discussing the application of IoT in smart manufacturing.

[0086] Document 2: Full text of the article discussing how artificial intelligence can optimize enterprise production processes.

[0087] Reference 3: Full text of the study on smart factory architecture supported by 5G networks.

[0088] Output summary:

[0089] "In recent years, research on intelligent manufacturing technology has focused on the application of the Internet of Things, artificial intelligence and 5G networks. The Internet of Things technology makes information exchange between devices more efficient, artificial intelligence plays an important role in process optimization, and 5G networks provide support for the high bandwidth and low latency of smart factories."

[0090] Step S200, constructing a total loss function, wherein the total loss function is a weighted sum of the summary generation task loss function, the question answering task loss function, the keyword extraction task loss function, the logical reasoning task loss function, the information extraction task loss function, and the literature review task loss function.

[0091] Among them, the loss function of the summary generation task, the question answering task, the information extraction task, and the literature review task are constructed based on the cross-entropy loss, the keyword extraction task loss function is constructed based on the binary cross-entropy loss, and the logical reasoning task loss function is constructed based on the similarity function.

[0092] Among them, the weight corresponding to each loss function is adjusted according to the corresponding loss change in the previous training cycle.

[0093] Specifically, the loss function of the summary generation task is expressed as follows:

[0094]

[0095] In the formula, T is the total number of words in the abstract, is the tth word generated, For input literature.

[0096] The expression of the question-answering task loss function is as follows:

[0097]

[0098] In the formula, T is the total number of words in the answer, is the tth word generated, For input literature and questions.

[0099] The keyword extraction task can be viewed as a multi-label classification problem, using the binary cross-entropy loss to optimize the probability prediction of whether each word is a keyword. By optimizing this loss, the model can gradually learn the characteristics of keywords. The loss function of the keyword extraction task is expressed as follows:

[0100]

[0101] In the formula, N is the total number of words in the document, is the label of whether the i-th word is a keyword (1 means yes, 0 means no), is the probability that the model predicts that the word is a keyword.

[0102] The loss function of the logical reasoning task uses logical consistency loss to evaluate whether the explanations and conclusions generated by the model are semantically consistent. To this end, the similarity score between sentences is introduced to ensure that there is no logical conflict between the conclusion and the explanation. The expression of the loss function of the logical reasoning task is as follows:

[0103] L_Logical reasoning = 1 - cosine_sim(E(large model output), E(training sample output))

[0104] In the formula, E represents the embedding of the sentence, which is obtained using the sentence-bert model, and cosine_sim represents the cosine similarity. The smaller the loss of the logical reasoning task, the higher the logical consistency of the explanation and conclusion. By minimizing the loss function of the logical reasoning task, the logical reasoning ability of the generated content can be improved.

[0105] The information extraction task involves the location and extraction of specific information (such as experimental data and conclusions). The cross-entropy loss for sequence labeling can be used to label each word as an information category (such as data, conclusion, or irrelevant). The expression of the information extraction task loss function is as follows:

[0106]

[0107] Where N is the total number of words in the document, C is the number of categories (such as data, conclusion, irrelevant), is the label of the i-th word belonging to category c, It is the probability that the model predicts that the word belongs to category c. The information extraction task loss function can help the model gradually learn to identify the target information type.

[0108] The expression of the literature review task loss function is as follows:

[0109]

[0110] In the formula, T is the total number of words in the literature review, is the tth word generated, The content of multiple documents is input.

[0111] The expression of the total loss function is as follows:

[0112]

[0113] In the formula, is the weight of each task, which is used to balance the training of different tasks. These weights can be adjusted dynamically according to the training effect to ensure that the performance of each task meets the expectations. The dynamic adjustment of weights is set by using the loss change rate, which is expressed as follows:

[0114]

[0115] In the formula, is the change in the loss of task i in the last training cycle. Tasks with larger changes indicate slower convergence or higher difficulty, so they are given higher weights, while tasks with smaller changes have lower weights. This approach allows the model to pay more attention to tasks that have not yet converged well.

[0116] Step S300, using supervised fine-tuning (SFT), trains the first large language model through diversified task samples and total loss function, and records the trained first large language model as the paper analysis large language model SemanticLLM.

[0117] Specifically, the process of inputting the sorted preliminary academic paper search results into the trained first language model (i.e., the paper analysis language model, SemanticLLM) to obtain the article abstracts of the academic papers and the comparative analysis results of similar academic papers includes:

[0118] For example, the first prompt word template is set as follows: "Now you are a professional with extensive knowledge. Please write a summary of the given article, highlighting the key points of the document, including innovation, advantages and possible shortcomings, and a structured output is required." The academic paper document and the first prompt word template are input into SemanticLLM, so that each paper can be quickly extracted and a high-quality brief summary can be generated to help users quickly understand the main content of the paper.

[0119] For example, the second prompt word template is set as follows: "Please mix all the above literature and help me write a professional literature review of about 1,000 words, covering the similarities and differences between papers, comparing different research methods and conclusions, and making structured outputs." Multiple academic papers and the second prompt word template are input into SemanticLLM to analyze the similarities and differences between similar papers, generate corresponding comparative analysis reports, and point out the similarities and differences between different papers in research methods, conclusions or experimental designs.

[0120] For example, for opinion-type academic papers, the third prompt word template is set: "For papers that expound on opinions and conduct corresponding analysis, statistics of opinions are conducted on the analyzed and compared literature, including but not limited to the proportion statistics of positive and negative opinions. Examples of positive and negative opinions are respectively the positive opinion: 'General artificial intelligence can be achieved through large language models', and the corresponding negative opinion is 'General artificial intelligence cannot be achieved through large language models'". The opinion-type academic papers and the third prompt word template are input into SemanticLLM, so as to perform statistical analysis and output of the paper opinions, helping users to quickly grasp the relevant research trends.

[0121] Furthermore, if Figure 4 As shown, the method also includes:

[0122] A prompt word template is set, and the comparative analysis results of similar academic papers output by the first largest language model and the prompt word template are input into the second largest language model, so as to score the comparative analysis results of similar academic papers output by the first largest language model; when the score output by the second largest language model is lower than the threshold, the prompt words corresponding to the comparative analysis of similar academic papers are adjusted, and the comparative analysis results of similar academic papers are regenerated until the score output by the second largest language model is greater than the threshold.

[0123] Among them, the second largest language model can use Open AI's Gpt-4, which needs to be different from the base model of SemanticLLM. The scoring dimensions include text fluency, contextual relevance, logical rigor, and data richness.

[0124] Among them, the scoring prompt word template can be set as:

[0125] Role:

[0126] You are a professional scientific research expert, proficient in the use of various large language models (such as ChatGPT, Tongyi Qianwen, etc.), have rich practical experience, and are good at comprehensively summarizing and summarizing the usage scenarios and functions of models in the field of scientific research.

[0127] Skill:

[0128] Skill 1: Overall Rating

[0129] 1. Conduct a comprehensive evaluation of the generated analysis report and give an overall score of 1-10.

[0130] 2. The scoring will take into account the fluency of the text, the logical rigor of the coherence between the previous and the next paragraphs, and whether it contains comparisons of key data.

[0131] Skill 2: Scoring dimensions individually

[0132] 1. For the text fluency dimension, give a score of 1-10 based on the natural coherence of the text content and the logical consistency of the context.

[0133] 2. For contextual relevance, a score of 1-10 is given based on the model's degree of relevance to the contextual information of the text and the level of support.

[0134] 3. Regarding the comparison of key data, carefully review the proportion of key data, the support of key data for the argument, and the consistency of data scale, and give a score of 1-10.

[0135] limit:

[0136] The operations only revolve around the scientific research field and do not involve other irrelevant content.

[0137] All scoring must be carried out strictly in accordance with the given standards and scoring ranges.

[0138] The output content must be organized according to the given format and must not deviate from it.

[0139] Furthermore, the method further comprises:

[0140] Extract features from the user's historical behaviors and explicit preferences to form a user profile; the historical behaviors include browsing, downloading, collecting, and liking; the explicit preferences include academic fields, target topics, target keywords, and language preferences;

[0141] Set prompt words to guide SemanticLLM to make personalized academic paper recommendations based on user portraits.

[0142] For example, the prompt word can be set as: "Based on the user's preference for <Topic 1, such as deep learning> and <Topic 2, such as reinforcement learning>, recommend some papers that have made breakthrough progress in these two fields in recent years. The recommended content should include the paper title, key research questions, innovation points and potential applications."

[0143] Furthermore, compared with the strong reliance of traditional academic paper recommendations on relevance retrieval and popular papers, using a large language model to set temperature parameter control can discover more long-tail papers and help find innovative research directions. For scenarios that require more hypothesis analysis, a higher temperature parameter is given to SemanticLLM, such as 0.7; for objective factual paper scenarios such as theoretical analysis and reviews, a lower temperature parameter is given, such as 0.1.

[0144] At the same time, in order to overcome the deviation of single paper recommendation, prompts with different angles are designed to generate multiple sets of candidate paper recommendation content. Then, post-processing methods (such as topic clustering and text similarity calculation) are used to filter redundant or irrelevant content and retain the best results. The angles are divided into research topic angles (mainstream, niche or cross-domain research) and research method angles (innovation of algorithms and models, analysis of applications and cases, experimental design and verification). In addition, the content of the prompt can be adjusted according to user feedback (such as whether to download or collect recommended papers), and the recommendation strategy of the model can be optimized through A / B testing. For example, if the user is not very interested in the recommended paper, the keywords or topics of the prompt can be dynamically adjusted to gradually adjust to a recommendation strategy that is more in line with the user's interests.

[0145] In summary, the present invention provides an academic paper retrieval method driven by a large language model, which obtains preliminary academic paper retrieval results by searching several academic paper databases, thereby integrating the results of multiple data sources. The present invention inputs the academic paper retrieval results into a pre-trained first large language model to obtain the article abstracts of the academic paper and the comparative analysis results of similar academic papers. The present invention uses a large language model to efficiently analyze and recommend the retrieval results, which can significantly improve user experience and retrieval efficiency. At the same time, during the training process of the first large language model, the present invention constructs a variety of task samples including abstract generation, question and answer, keyword extraction, logical reasoning, information extraction, and literature review generation, so that the large language model can better understand the complex information in the literature.

[0146] Accordingly, the present application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned academic paper retrieval method based on large language model drive. Figure 5As shown, a hardware structure diagram of any device with data processing capability for the academic paper retrieval method driven by a large language model provided by an embodiment of the present invention is shown, except Figure 5 In addition to the processor, memory and network interface shown, any device with data processing capability in which the apparatus in the embodiment is located may also include other hardware, generally based on the actual functions of the device with data processing capability, which will not be described in detail.

[0147] Correspondingly, the present application also provides a computer-readable storage medium on which computer instructions are stored, and when the instructions are executed by the processor, the above-mentioned academic paper retrieval method based on large language model drive is implemented. The computer-readable storage medium can be an internal storage unit of any device with data processing capability described in any of the aforementioned embodiments, such as a hard disk or a memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), an SD card, a flash card (Flash Card), etc. equipped on the device. Furthermore, the computer-readable storage medium can also include both an internal storage unit and an external storage device of any device with data processing capability. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capability, and can also be used to temporarily store data that has been output or is to be output.

[0148] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A method for academic paper retrieval based on a large language model, characterized in that: The method comprises: According to the input search request, preliminary academic paper search results are obtained by searching several academic paper databases; Merge the preliminary academic paper search results and sort them by relevance, popularity, and time; Input the sorted preliminary academic paper search results into the pre-trained first language model to obtain the article abstracts of the academic papers and the comparative analysis results of similar academic papers; Among them, the training process of the first language model includes: constructing diversified task samples including summary generation, question and answering, keyword extraction, logical reasoning, information extraction and literature review generation, constructing a total loss function, and training the first language model through the total loss function; the total loss function is the weighted sum of the summary generation task loss function, the question and answer task loss function, the keyword extraction task loss function, the logical reasoning task loss function, the information extraction task loss function, and the literature review task loss function.

2. According to claim 1, the academic paper retrieval method based on large language model driving is characterized in that: The search request includes: target academic paper database, document type, target time range, target subject, and target keywords.

3. The academic paper retrieval method based on a large language model drive according to claim 1 is characterized in that: The process of weighted sorting of preliminary academic paper search results by relevance, popularity, and time includes: According to the keywords and / or themes in the search request, the matching degree with the keywords and / or themes of each academic paper in the preliminary academic paper search results is calculated, so as to set the relevance ranking factor; Set the popularity ranking factor based on the number of citations, downloads, and sharing frequency of each academic paper in the preliminary academic paper search results; Set the time ranking factor according to the publication time of each academic paper in the preliminary academic paper search results; Define sorting rules, perform weighted processing on relevance sorting factor, heat sorting factor, and time sorting factor, so as to complete the sorting of preliminary academic paper retrieval results.

4. The method for academic paper retrieval based on a large language model according to claim 1, characterized in that: After ranking the preliminary academic paper search results, and before processing the ranked preliminary academic paper search results using the pre-trained first language model, the method further includes: Apply copyright mark and / or paper type mark to the sorted preliminary academic paper search results; Among them, the paper type tags include: journal papers, conference papers, technical reports, and preprints.

5. The method for academic paper retrieval based on a large language model drive according to claim 1, characterized in that: The process of training the first language model includes: Construct multi-task samples; including: Constructing a summary generation task sample for the summary generation task; the summary generation task sample includes an academic paper as input and a summary as output; Constructing question-answering task samples for question-answering tasks; the question-answering task samples include questions and academic papers as inputs, and answers as outputs; Constructing a keyword extraction task sample for a keyword extraction task; the keyword extraction task sample includes an academic paper as input and a keyword as output; Constructing a logical reasoning task sample for a logical reasoning task; the logical reasoning task sample includes an academic paper as input and a reasoning result as output; Constructing information extraction task samples for information extraction tasks; the information extraction task samples include academic papers as input and target information as output; the target information includes experimental methods, experimental data and / or research conclusions; Constructing a literature review task sample for the literature review task; the literature review task sample includes a number of academic papers as input and a review as output; Constructing a total loss function, and training the first language model through the total loss function; the total loss function is a weighted sum of the summary generation task loss function, the question answering task loss function, the keyword extraction task loss function, the logical reasoning task loss function, the information extraction task loss function, and the literature review task loss function; Among them, the loss function of the summary generation task, the question answering task, the information extraction task, and the literature review task are constructed based on the cross entropy loss, the keyword extraction task loss function is constructed based on the binary cross entropy loss, and the logical reasoning task loss function is constructed based on the similarity function; Among them, the weight corresponding to each loss function is adjusted according to the corresponding loss change in the previous training cycle.

6. The method for academic paper retrieval based on a large language model according to claim 1, characterized in that: The method further comprises: Setting a prompt word template, inputting the comparative analysis results of similar academic papers output by the first language model and the prompt word template into the second language model, thereby scoring the comparative analysis results of similar academic papers output by the first language model; Among them, the scoring dimensions include text fluency, contextual relevance, logical rigor, and data richness; When the score output by the second largest language model is lower than the threshold, the prompt words corresponding to the comparative analysis of similar academic papers are adjusted, and the comparative analysis results of similar academic papers are regenerated until the score output by the second largest language model is greater than the threshold.

7. The method for academic paper retrieval based on a large language model according to claim 1, characterized in that: The method further comprises: Extract features from the user's historical behaviors and explicit preferences to form a user profile; the historical behaviors include browsing, downloading, collecting, and liking; the explicit preferences include academic fields, target topics, target keywords, and language preferences; Personalized academic paper recommendations based on user portraits.

8. An electronic device, comprising a memory and a processor, characterized in that: The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the academic paper retrieval method based on large language model drive as described in any one of claims 1-7 above.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the academic paper retrieval method based on large language model drive as described in any one of claims 1 to 7 is implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the method for academic paper retrieval based on large language model drive described in any one of claims 1-7 is implemented.

Images