Scientific research topic selection information generation method and device

Abstract

The invention relates to the technical field of data processing, and discloses a scientific research topic selection information generation method and device. The method comprises the following steps: inputting demand information to be identified into a text embedding model, and determining a feature vector of the demand information; determining a target vector matched with the feature vector in a vector data set according to the feature vector; determining target literature data associated with the target vector in a preset literature database according to the target vector; and inputting the target literature data into a preset large model, and determining scientific research topic selection information corresponding to the demand information. According to the demand information input by the user, the target vector is extracted from the vector data set established according to the comprehensive literature data, the target literature data most related to the user demand is obtained from the preset literature database according to the target vector, and the diversified scientific research topic selection information is intelligently generated through the preset large model; therefore, the processes of literature retrieval, scientific research novelty search and the like are integrated with scientific research topic selection, and the scientific research efficiency is improved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method and device for generating scientific research topic information. Background Art

[0002] Research topic selection is an important part of scientific research activities. During the topic selection process, researchers may need to go beyond their own professional fields and query and read knowledge and methods from other disciplines on their own. However, manual literature query is very limited and inefficient.

[0003] At present, in addition to relying on researchers' mastery of scientific research knowledge and cutting-edge research in this field, research topic selection can also be carried out through information retrieval methods such as literature retrieval and scientific research query. In related technologies, information retrieval mostly only covers part of the database or literature resources, and due to the limitations of the data source, it is impossible to fully retrieve relevant information. Moreover, in related technologies, information retrieval and research topic selection are usually two independent processes. In most cases, researchers need to manually judge the required research topic based on the results of the retrieval, which makes it impossible to effectively integrate information retrieval and research topic selection, and it is difficult to efficiently integrate relevant search results to carry out innovative topic selection. Summary of the invention

[0004] In view of this, the present application provides a method and device for generating scientific research topic information. According to the demand information input by the user, a target vector is extracted from a vector data set established based on comprehensive literature data, and the target literature data most relevant to the user's needs is obtained in a preset literature database based on the target vector. Through a preset large model, diverse scientific research topic information is intelligently generated according to the target literature data, thereby realizing the integration of literature retrieval, scientific research novelty search and other processes with scientific research topic generation, thereby improving scientific research efficiency.

[0005] According to one aspect of the present application, a method for generating scientific research topic information is provided, including: inputting demand information to be identified into a text embedding model to determine a feature vector of the demand information, wherein the text embedding model is trained based on document data in a preset document database; determining a target vector matching the feature vector in a vector data set based on the feature vector, wherein the vector data set is obtained by vectorizing the document data in the preset document database based on the text embedding model; determining target document data associated with the target vector in the preset document database based on the target vector; and inputting the target document data into a preset large model to determine the scientific research topic information corresponding to the demand information.

[0006] According to another aspect of the present application, a device for generating scientific research topic information is provided, including: a determination module, used to input demand information to be identified into a text embedding model, and determine a feature vector of the demand information, wherein the text embedding model is trained based on document data in a preset document database; and, based on the feature vector, determine a target vector that matches the feature vector in a vector data set, wherein the vector data set is obtained by vectorizing the document data in the preset document database based on the text embedding model; and, based on the target vector, determine target document data associated with the target vector in the preset document database; and, input the target document data into a preset large model to determine the target scientific research topic information corresponding to the demand information.

[0007] By means of the above technical scheme, the present application provides a method and device for generating scientific research topic information, which uses a pre-trained text embedding model that conforms to the scientific research scenario to vectorize the comprehensive literature data and convert it into a vector data set for efficient data storage and retrieval. Furthermore, by retrieving the target vector that matches the feature vector of the demand information input by the user in the vector data set, it is possible to quickly extract literature data that highly matches the user's needs from a large amount of literature data, thereby improving the accuracy and comprehensiveness of the retrieval. And according to the target vector, the target literature data that is most relevant to the user's needs is obtained in the preset literature database, reducing redundant information and providing better input data for the preset large model. Thus, the preset large model is used to intelligently generate diverse scientific research topic information, so as to integrate the processes of literature retrieval, scientific research novelty search and scientific research topic generation, and improve the efficiency and effectiveness of scientific research.

[0008] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0010] Figure 1 A schematic diagram of the process of generating scientific research topic information provided by an embodiment of the present application is shown;

[0011] Figure 2 The structure block diagram of the scientific research topic information generation device provided in the embodiment of the present application is shown;

[0012] Figure 3A schematic diagram of a vector data set provided in an embodiment of the present application is shown;

[0013] Figure 4 A schematic diagram showing the requirement information input by a user provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0014] The present application will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

[0015] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be interpreted as limiting the present application.

[0016] Those skilled in the art will appreciate that, unless expressly stated, the singular forms "a", "an", "said" and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present application refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "connected" to another element, it may be directly connected or connected to the other element, or there may be intermediate elements. In addition, the "connection" or "connection" used herein may include wireless connection or wireless fusion. The term "and / or" used herein includes all or any unit and all combinations of one or more associated listed items.

[0017] Now, exemplary embodiments according to the present application will be described in more detail with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in a variety of different forms and should not be interpreted as being limited to the embodiments set forth herein. It should be understood that these embodiments are provided to make the disclosure of the present application thorough and complete, and to fully convey the concepts of these exemplary embodiments to those of ordinary skill in the art.

[0018] The method for generating scientific research topic information provided in the embodiment of the present application can be applied to a terminal, can be applied to a server, and can also be software running in a terminal or a server. In some embodiments, the terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, etc.; the server side can be configured as an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud data sets, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application that implements the method for generating scientific research topic information, etc., but is not limited to the above forms.

[0019] The selection of research topics is an important part of scientific research activities and the primary issue of scientific and technological innovation. The scope of research topic selection is relatively wide. The basis for topic selection is not only related to the latest scientific and technological development, but also to economic development and social needs. Researchers have a basic understanding of the fields they are studying, but more data support is needed for specific research directions. In the process of research topic selection, there are many difficulties: research topics need to be cutting-edge and innovative, but how to accurately judge the cutting-edge dynamics of a field and propose innovative research questions on this basis requires researchers to master deep professional knowledge, keen insight and broad academic vision. Modern scientific research is increasingly tending towards interdisciplinary integration. Therefore, when selecting topics, researchers may need to transcend their own professional fields to understand and master the knowledge and methods of other disciplines. It is inefficient to search and read the knowledge and methods of other disciplines on their own.

[0020] At present, the selection of research topics not only depends on the researchers' mastery of scientific research knowledge and cutting-edge research in this field, but also can be done through information retrieval methods such as literature retrieval and scientific research query. Taking the proposal of graduate thesis as an example, the tutor only provides a general research scope, and a more detailed research direction needs to be determined by the graduate students themselves through investigation of existing research situations. However, on the one hand, due to the limitations of the retrieval data source, it is impossible to fully retrieve relevant information. On the other hand, information retrieval and research topic selection are two independent processes, and it is difficult to efficiently integrate relevant search results to carry out innovative and intelligent topic selection.

[0021] In this embodiment, a method for generating scientific research topic information is provided, which integrates the processes of literature retrieval, scientific research novelty search, etc. with the generation of scientific research topic, thereby improving the efficiency and effectiveness of scientific research. Figure 1 As shown, the method includes:

[0022] Step 101: input the demand information to be identified into a text embedding model to determine a feature vector of the demand information.

[0023] Among them, the text embedding model is trained based on the document data in the preset document database.

[0024] In this embodiment, the text embedding model is used to convert the user input demand information into a vector representation, so that in the subsequent steps, rapid retrieval and intelligent generation of scientific research topics can be performed based on the feature vector.

[0025] Exemplarily, the requirement information input by the user is received, wherein the requirement information can be a keyword, a combination of keywords or a text fragment related to the research content, for example, "generate the application of artificial intelligence technology in the medical field". Further, a pre-trained text embedding model that conforms to the scientific research scenario is used to accurately capture the semantic features of the requirement information, and the requirement information is represented as a feature vector for vector retrieval based on semantic content in subsequent steps.

[0026] In one embodiment, the method for generating scientific research topic information further includes: fine-tuning a preset cross-language text embedding model according to document data in a preset document database to determine a text embedding model.

[0027] In this embodiment, fine-tuning is performed based on the open source cross-language text embedding model to make it more suitable for the text characteristics in the scientific research field.

[0028] Exemplarily, an open source cross-language text embedding model is selected as a preset cross-language text embedding model, and data for fine-tuning is determined based on the literature data in a preset literature database. For example, similar text pairs, Chinese-English title pairs, Chinese-English abstract pairs, and data pairs consisting of subheadings at all levels and paragraphs corresponding to subheadings at all levels in the literature are determined in the preset literature database, and the open source cross-language text embedding model is fine-tuned to improve the open source cross-language text embedding model's ability to understand and represent scientific research literature, and the text embedding model is determined based on the fine-tuned open source cross-language text embedding model, so that the vector generated by the text embedding model can more accurately reflect the semantics and theme of the literature.

[0029] Among them, similar text pairs are texts with similar semantics, such as "Gene mutations may cause diseases." and "Diseases may be caused by gene mutations, so that the text embedding model can more effectively handle complex sentences and professional terms in scientific research documents. The Chinese-English title pair is the Chinese title and English title of the same document. Similarly, the Chinese-English abstract is the Chinese abstract and English abstract of the same document, so as to enhance the versatility and adaptability of the text embedding model in a multilingual environment and provide researchers with more comprehensive document information. The data pairs consisting of subheadings at all levels in the document and the corresponding paragraphs at all levels of the subheadings help the text embedding model understand the structural hierarchy of scientific research documents. The subheadings represent different parts or themes of the document, corresponding to the paragraph content, which allows the text embedding model to learn the organizational structure and logical relationship of the document, and improve the processing ability and overall understanding level of long scientific research documents.

[0030] It should be noted that this embodiment collects a large amount of Chinese and English academic literature data in advance, sorts out the collected literature data to distinguish the various components in the literature, and marks the various components in the literature, processes the literature data with marks into various fields, and thus constructs a preset literature database. Exemplarily, the literature data is marked according to the publication time of the literature, the classification of the literature, the title, abstract, keywords, subheadings at all levels and the paragraphs corresponding to the subheadings at all levels. Among them, the classification of the literature can be divided into core journals and non-core journals.

[0031] It is worth mentioning that the preset literature database is equipped with an interface that can access data from other databases or systems related to scientific research literature, providing researchers with a wider range of literature resources, covering more subject areas, research topics and time ranges, and meeting different scientific research needs. In addition, the preset literature NC library is equipped with an open source large model for identification, which identifies the newly accessed data and ensures the validity of the preset literature database.

[0032] In one embodiment, the method for generating scientific research topic information also includes: combining the literature data according to preset identifiers in a preset literature database according to preset combination rules to determine first target data and second target data; determining a first vector of the first target data and a second vector of the second target data based on a text embedding model; and determining a vector data set based on the first vector and the second vector.

[0033] The preset identifiers include at least the title of the document, keywords in the document, the abstract of the document, subheadings of various levels in the document, and paragraphs corresponding to subheadings of various levels in the document.

[0034] In this embodiment, the document data are combined in multiple ways from different dimensions according to preset identifiers, and the combined data are converted into a vector data set using a text embedding model, so that the vector data set can more comprehensively represent the semantic features, structural information, contextual relationships, etc. of the document data, so that subsequent retrieval in the vector data set based on the feature vector of the demand information can provide users with more accurate results.

[0035] Exemplarily, the document data is combined according to the preset identifier and the preset combination rule to obtain the first target data and the second target data of different structures. Then, the first target data and the second target data are represented as the first vector and the second vector respectively using the text embedding model, so as to construct a vector data set according to the first vector and the second vector.

[0036] It is worth mentioning that in the vector dataset obtained through the text embedding model, similar vectors have similar semantic features, which provides a basis for subsequent fast and accurate vector retrieval.

[0037] Further, as a refinement and extension of the specific implementation methods of the above-mentioned embodiment, in order to fully illustrate the specific implementation process of this embodiment, according to the preset identifier of the document data in the preset document database, the document data is combined according to the preset combination rule to determine the first target data and the second target data, including: according to the preset identifier, the document data is combined according to the first preset combination rule to determine the first target data; according to the preset identifier, the document data is combined according to the second preset combination rule to determine the sub-data, and the second target data is determined according to the sub-data that meets the same second preset combination rule.

[0038] The preset combination rule includes a first preset combination rule and a second preset combination rule.

[0039] In this embodiment, the combination method of the first preset combination rule is to perform a single combination of document data that conforms to the same first preset combination rule to determine the first target data, thereby using the first vector after the first target data is vectorized to semantically express the document from a holistic perspective.

[0040] Different from the first preset combination rule, the second preset combination rule is to combine a certain data with other multiple different data in the document data that conforms to the same second preset combination rule, determine multiple sub-data, and determine the second target data based on the multiple sub-data. In this way, the second vector after the second target data is vectorized can reflect the semantic features of the document from a finer granularity, provide richer and more specific document information representation, and adapt to different retrieval needs and scenarios.

[0041] Exemplarily, the first preset combination rule includes at least: combining the title of the same document with keywords, combining the title of the same document with abstracts, combining the title of the same document with all subheadings in the document, and taking subheadings of all levels in the document as the first target data. The hierarchy of subheadings in the document is specifically distinguished according to different documents. For example, it can be divided into parent subheadings and last subheadings. The parent subheading is the subheading in the title hierarchy structure that is second only to the document title, and the last subheading is the bottom-most title in the title hierarchy structure and no longer contains any subheadings.

[0042] For example, if the first preset combination rule is to combine the title and keywords of the same document, then the document data of the title and keywords of the same document are called in the preset document database according to the preset identifier, and they are combined into the first target data in the form of a string. Similarly, if the first preset combination rule is to combine the title and abstract of the same document, then the document data of the title and abstract of the same document are called in the preset document database according to the preset identifier, and they are combined into the first target data in the form of a string. Taking the example that the subheadings in the document include the parent subheading and the last subheading, if the first preset combination rule is to combine the title of the same document with all the subheadings in the document, then the document data of the title, parent subheading and last subheading of the same document can be called in the preset document database according to the preset identifier, and they are combined into the first target data in the form of a string. If the first preset combination rule is to use the subheadings of each level in the document as the first target data, the document data with the preset identifier as the last subheading can be called in the preset document database and used as the first target data in the form of a string, or the document data with the preset identifier as the parent subheading can be called in the preset document database and used as the first target data in the form of a string. Thus, the first target data can be used as the input of the text embedding model to obtain the first vector of the first target data, and the first vector can be stored separately in a field as a single-value vector field.

[0043] Furthermore, the second preset combination rule at least includes: combining the title of the same document with each sentence in the abstract, and combining each level of subheadings in the same document with each paragraph corresponding to each level of subheadings.

[0044] For example, if the second preset combination rule is to combine the title of the same document with each sentence in the abstract, taking the abstract including four sentences as an example, the title of the same document and the four sentences in the abstract are combined in the form of a character string according to the preset identifier to determine four sub-data, and then the second target data corresponding to the second preset combination rule is determined according to the four sub-data. If the second preset combination rule is to combine the sub-headings of each level in the same document with each paragraph corresponding to the sub-headings of each level, taking the sub-heading as the last sub-heading as an example, if there are three paragraphs corresponding to the last sub-heading in the body of the document, then the last sub-heading is combined with the three paragraphs in the form of a character string according to the preset identifier to determine three sub-data, and then the second target data corresponding to the second preset combination rule is determined according to the three sub-data.

[0045] In this embodiment, unlike the related art that simply matches based on keywords, the vector data set constructed by the first preset combination rule and the second preset combination rule can cover different parts and complex semantics of the document when performing vector retrieval in the vector data set later, and more comprehensively determine the relevance of the document data to the user's needs. For example, this embodiment can not only query the semantic features of simple content such as document titles, abstracts, keywords, etc., but also query the semantic features of different parts of the content in the document body, thereby obtaining document data that is more relevant to the user's needs.

[0046] In one embodiment, a first vector subset may be determined based on a first vector vectorized from the first target data. Similarly, a second vector subset may be determined based on a second vector vectorized from the second target data. Thus, a vector data set is determined based on the first vector subset and the second vector subset.

[0047] For example, Figure 3 As shown, taking the first preset combination rule and the second preset combination rule in the above-mentioned example as an example, four first vector subsets and two second vector subsets can be obtained, so that all the document data in the preset document database can be represented from the document granularity and the full-text fragment granularity according to these four first vector subsets and the two second vector subsets, thereby determining the vector data set.

[0048] Furthermore, as a refinement and extension of the specific implementation methods of the above-mentioned embodiment, in order to fully illustrate the specific implementation process of this embodiment, the step of determining the second vector of the second target data based on the text embedding model specifically includes: inputting the sub-data in the second target data into the text embedding model, determining the sub-vector corresponding to the sub-data, and determining the second vector of the second target data based on the sub-vector.

[0049] In this embodiment, based on the different structures of the second target data, the text embedding model is used to respectively vectorize each sub-data in the second target data to obtain the sub-vectors corresponding to the sub-data, and the sub-vectors are stored together in the same field as a multi-value vector field to reflect the structure and contextual relationship of the document, so as to better capture the correlation between different parts of the document, thereby determining the second vector of the second target data according to the sub-vectors.

[0050] Exemplarily, the second target data is obtained according to the second preset combination rule of combining subheadings of different levels in the same document with paragraphs corresponding to the subheadings of different levels, and the subheading is taken as the last subheading. If there are three paragraphs corresponding to the last subheading, the last subheading is combined with the three paragraphs to determine three sub-data and form the second target data. Then, each sub-data in the second target data is input into the text embedding model to obtain a sub-vector of the sub-data, and the three sub-vectors are stored in the same field to obtain a multi-value vector field, and the second vector of the second target data is determined according to the sub-vector.

[0051] Step 102: determining a target vector that matches the feature vector in the vector data set according to the feature vector.

[0052] Among them, the vector data set is obtained by vectorizing the document data in the preset document database according to the text embedding model.

[0053] In this embodiment, a retrieval based on semantic features is performed in a vector data set constructed according to a preset document database, and the target vector obtained through the vector retrieval is used to quickly screen out document data that is highly relevant to user needs from a large amount of document data, thereby improving the accuracy of the retrieval.

[0054] Further, as a refinement and extension of the specific implementation methods of the above-mentioned embodiment, in order to fully illustrate the specific implementation process of this embodiment, step 102, that is, the step of determining the target vector matching the feature vector in the vector data set based on the feature vector, specifically includes: clustering each vector in the vector data set to determine a first preset number of clusters; calculating a first distance between the feature vector and the first cluster center in the clusters, and sorting the clusters according to the first distance to determine the order of the clusters; determining the vectors in the clusters that are located before the first preset position in the order of the clusters as candidate vectors; calculating the second distance between the feature vector and the candidate vector, and sorting the candidate vectors according to the second distance to determine the vector order; determining the candidate vector that is located before the second preset position in the vector order as the target vector.

[0055] In this example, similarity retrieval is performed by calculating the distance between the feature vector and each vector in the vector data set, and clustering processing is used to reduce the number of vectors that need to calculate the distance, thereby ensuring the retrieval effect while improving the retrieval efficiency.

[0056] It should be noted that, since similar vectors in the vector data set have similar semantic features, the closer the distance between the vectors, the higher the similarity between the vectors. In this embodiment, the similarity between the vectors is represented by the distance between the vectors.

[0057] Exemplarily, this embodiment uses the IVFPQ (Inverted File and Product Quantization) algorithm to index vector data. First, the KMeans Clustering algorithm is used to cluster each vector in the vector data set, and similar vectors in the vector data set are clustered together to obtain a first preset number of clusters. Next, the first distance between the feature vector and the first cluster center in the cluster is calculated to reduce the number of vectors for which the distance needs to be calculated by several orders of magnitude. Then, the clusters are sorted from small to large according to the first distance to determine the order of the clusters. The vectors in the clusters that are located before the first preset position in the order of the clusters are determined as candidate vectors, so as to quickly find several vectors that are close to the feature vector in the vector data set, and realize preliminary screening of each vector in the vector data set.

[0058] Among them, the first distance can be quickly determined by calculating the residual between the vectors, thereby improving the speed of retrieval and ensuring the effect of retrieval.

[0059] Exemplarily, the second distance between the feature vector and the candidate vector is calculated, and the candidate vectors are sorted from small to large according to the second distance to determine the vector order. The candidate vector located before the second preset position in the vector order is determined as the target vector, so as to determine the target vector with the most similar semantic features to the feature vector in the vector data set, so as to obtain the target document data most relevant to the user's needs according to the target vector in the subsequent steps.

[0060] In one embodiment, if the vector data set includes multiple vector subsets, such as the first vector subset and the second vector subset, as described in the above embodiment, the above clustering process and the calculation of the second distance after clustering can be used to search in each vector subset to obtain the target number of candidate target vectors that match the feature vector in each vector subset. Then, all candidate target vectors are sorted from small to large according to the distance between the candidate target vector and the feature vector to obtain a candidate vector sequence, so that the candidate target vector located before the second preset position in the candidate vector sequence is determined as the target vector.

[0061] Further, as a refinement and extension of the specific implementation of the above embodiment, in order to fully illustrate the specific implementation process of this embodiment, the step of calculating the second distance between the feature vector and the candidate vector specifically includes: dividing the candidate vector into a second preset number of first sub-segments, and determining the position sequence of the first sub-segments according to the order in which the first sub-segments are generated; clustering the first sub-segments with the same position sequence to determine a third preset number of second cluster centers; calculating the third distance between the first sub-segment and the second cluster center corresponding to the first sub-segment, and sorting the second cluster centers corresponding to the first sub-segment according to the third distance to determine the center order; The second cluster center that is at the first position in the center order updates the first sub-segment, and updates the candidate vector according to the updated first sub-segment; divides the feature vector into a second preset number of second sub-segments, and determines the position sequence of the second sub-segments according to the order in which the second sub-segments are generated; calculates the fourth distance between the second sub-segment and the second cluster center corresponding to the first sub-segment with the same position sequence as the second sub-segment, and determines the distance list according to the fourth distance; determines the target distance between the second sub-segment and the first sub-segment with the same position sequence in the feature vector and the updated candidate vector according to the distance list, and determines the second distance between the feature vector and the candidate vector according to the target distance.

[0062] In this embodiment, the high-dimensional candidate vector is decomposed into multiple low-dimensional first sub-segments, and each first sub-segment is quantized and encoded, so as to map the high-dimensional vector into a low-dimensional discrete space, so as to quickly calculate the distance and perform the search.

[0063] Exemplarily, the candidate vector is divided into a second preset number of first sub-segments, and the bit sequence of the first sub-segments is determined according to the order in which the first sub-segments are generated. Then, the first sub-segments of the same bit sequence are clustered to determine the second cluster centers of the third preset number. For example, if the second preset number is 4, each candidate vector is divided into four first sub-segments, and the first first sub-segment in each candidate vector is clustered to obtain the second cluster center corresponding to the first first sub-segment, and so on, to respectively determine the second cluster center corresponding to the second first sub-segment, the second cluster center corresponding to the third first sub-segment, and the second cluster center corresponding to the fourth first sub-segment.

[0064] Next, for any first subsegment in the candidate vector, the first subsegment is updated according to the identification number of the second cluster center corresponding to the first subsegment that is closest to the first subsegment, so that the candidate vector becomes a vector composed of the identification numbers of a second preset number of second cluster centers.

[0065] Exemplarily, the third distance between the first sub-segment and the second cluster center corresponding to the first sub-segment is calculated, and the second cluster centers corresponding to the first sub-segment are sorted from small to large according to the third distance to determine the center order. The first sub-segment is updated according to the second cluster center that is first in the center order, and the candidate vector is updated according to the updated first sub-segment. For example, taking the second preset number as 4 as an example, the candidate vector may be quantized to [124, 56, 132, 222].

[0066] Furthermore, when the demand information input by the user is represented as a feature vector, the feature vector is also divided into a second preset number of second sub-segments, and then the distance between each second sub-segment in the feature vector and its corresponding previously pre-trained second cluster center is calculated to obtain a distance list. Among them, the second cluster center corresponding to the second sub-segment is the second cluster center corresponding to the first sub-segment with the same position order as the second sub-segment.

[0067] When calculating the second distance between the feature vector and the candidate vector, the candidate vector has been quantized into the identification number of the second cluster center, and the distance between the second sub-segment and the second cluster center in the feature vector has been sorted into a distance list. Therefore, the second distance can be converted into the sum of the distances between the first sub-segment and the second sub-segment of the same order, and the distance between the first sub-segment and the second sub-segment of the same order can be obtained by querying the distance list, thereby reducing the amount of calculation, ensuring the retrieval effect, and improving the retrieval efficiency.

[0068] Exemplarily, a fourth distance between the second subsegment and the second cluster center corresponding to the first subsegment having the same position sequence as the second subsegment is calculated, and a distance list is determined according to the fourth distance. According to the distance list, a target distance between the second subsegment and the first subsegment having the same position sequence in the feature vector and the updated candidate vector is determined, and a second distance between the feature vector and the candidate vector is determined according to the target distance.

[0069] For example, continuing to take the second preset number as 4, if a candidate vector a is quantized to [124, 56, 132, 222], when calculating the second distance between the feature vector and the candidate vector a, first query the distance list to obtain the distance d1 between the first second sub-segment in the feature vector and the second cluster center with the identification number 124, and then query the distance list to obtain the distance d2 between the second second sub-segment in the feature vector and the second cluster center with the identification number 56. By analogy, four distances d1, d2, d3, and d4 can be obtained, and the second distance d between the feature vector and the candidate vector a is converted to d=d1+d2+d3+d4.

[0070] Among them, the third distance and the fourth distance can be determined by Euclidean distance, Hamming distance, cosine distance, etc., so as to improve the retrieval speed.

[0071] It should be noted that the first preset number, the second preset number and the third preset number are specifically set according to the search accuracy required in the actual application scenario, and the larger the values ​​of the first preset number, the second preset number and the third preset number are, the higher the search accuracy is. The first preset position sequence, the second preset position sequence and the target number are specifically set according to the number of search results required in the actual application scenario.

[0072] In one embodiment, the method for generating scientific research topic information also includes: if the candidate vector includes multiple sub-vectors, dividing the sub-vectors in the candidate vector into a second preset number of first sub-segments, and calculating the candidate second distances between the sub-vectors in the candidate vector and the feature vector; sorting the candidate second distances to determine the distance order; and determining the candidate second distance that is ranked first in the distance order as the second distance between the feature vector and the candidate vector.

[0073] In this example, multi-value vector indexing is supported. If there are multiple sub-vectors in the candidate vector, each sub-vector in the candidate vector will participate in the vector retrieval, integrating the various features included in the candidate vector, and using the sub-vector in the candidate vector that is most similar to the feature vector to represent the entire candidate vector, thereby improving the effectiveness of the retrieval results and being applicable to complex data structures.

[0074] Exemplarily, if the candidate vector includes multiple sub-vectors, for example, if the candidate vector is obtained according to the second target data corresponding to the second preset combination rule in the above embodiment, it is used to more carefully represent the content of each part of the document. Then, according to the method for calculating the second distance in the above embodiment, each sub-vector in the candidate vector is divided into the first sub-segments of the second preset number, so as to calculate the candidate second distances between each sub-vector in the candidate vector and the feature vector respectively, so as to reflect the correlation between each feature included in the candidate vector and the feature vector. Further, the candidate second distances are sorted from small to large to determine the distance order, and the candidate second distance that is at the first place in the distance order is determined as the second distance between the feature vector and the candidate vector, so as to use the feature in the candidate vector that is most similar to the feature vector to represent the entire candidate vector.

[0075] Step 103: determining target document data associated with the target vector in a preset document database according to the target vector.

[0076] In this embodiment, the results after vector retrieval are optimized, so that target document data most relevant to user needs can be found in a large amount of document data according to the target vector, providing better input for the large model.

[0077] Further, as a refinement and extension of the specific implementation of the above embodiment, in order to fully illustrate the specific implementation process of this embodiment, step 103, that is, the step of determining the target document data associated with the target vector in the preset document database according to the target vector, specifically includes: determining the candidate document titles corresponding to the target vector in the preset document database according to the target vector; removing the same titles in the candidate document titles to update the candidate document titles; sorting the candidate document titles according to the target vector to determine the document order; determining the target document data according to the document data associated with the candidate document titles located before the third preset position in the document order in the preset document database.

[0078] In this example, duplicate documents are removed from the documents associated with the target vector, and the documents associated with the target vector are sorted to optimize the search results.

[0079] Exemplarily, the titles of each document associated with the target vector are determined in a preset document database according to the target vector, i.e., the candidate document titles. For example, it is determined by the preset identifier of the document data in the preset document database. Then, the same titles in the candidate document titles are removed to remove duplicate documents, thereby updating the candidate document titles and reducing redundant information. Further, according to the parameters used when determining the target vector in the vector data set, the candidate document titles corresponding to the target vector are sorted to obtain the document sequence. For example, if the target vector is determined according to the distance or similarity between each vector and the feature vector in the vector data set, the candidate document titles associated with the target vector are sorted from small to large according to the distance corresponding to the target vector, or the candidate document titles associated with the target vector are sorted from large to small according to the similarity corresponding to the target vector. Then, the document data associated with the candidate document titles in the preset document database that are located before the third preset position in the document sequence are determined as the target document data, so that the document data most relevant to the user's needs can be accurately searched in a large amount of document data.

[0080] The third preset order is determined according to the number of documents required in the actual application scenario. For example, the third preset order is set to 6 to obtain data of 5 documents that are most relevant to the user's needs.

[0081] For example, the target document data may be the title and abstract of the document, or the document title, subheadings at various levels in the document, and paragraphs in the text corresponding to the subheadings at various levels.

[0082] Step 104, input the target document data into the preset large model to determine the scientific research topic information corresponding to the demand information.

[0083] In this embodiment, the target document data is input into a preset large model as prior knowledge, and the powerful language understanding and generation capabilities of the preset large model are utilized to generate diversified scientific research topic information based on the target document data, thereby providing researchers with innovative research ideas and directions.

[0084] It should be noted that the preset big model is an open source generative big model, which can automatically perform related operations and generate diverse output content based on the knowledge fragments and instructions input by the user.

[0085] For example, the target document data can be combined with preset prompt words and input into the preset large model to generate the scientific research topic information required by the user. For example, the target document data combined with the preset prompt words is: You are currently a well-known expert in the field of XXX. Based on the information input below and all journal documents in the past X years, summarize X research topics. The input information is as follows: XXXXXX (target document data).

[0086] In one embodiment, after constructing a vector data set based on a preset literature database, the IVFPQ algorithm is used to index the vector data, wherein the dimension of each vector in the vector data set is 1024 dimensions, the number of clustering centers involved in the clustering process, i.e., the first preset number and the third preset number are both set to 262144, and the number of segments into which the vector is divided is 256, i.e., the second preset number is set to 256.

[0087] In this embodiment, if Figure 4 As shown, the user inputs the demand information, for example, the user inputs "artificial intelligence field, generate artificial intelligence applications in the medical field". Then, the input "generate artificial intelligence applications in the medical field" is converted into a feature vector using a text embedding model suitable for the scientific research field. Further, a vector search is performed in the vector data set based on the feature vector to obtain 300 target vectors. The literature data associated with the target vector is sorted and deduplicated based on the target vector to obtain the target literature data corresponding to 10 documents, and the target literature data is combined with the preset prompt words to generate the following format information:

[0088] "You are currently a well-known expert in the field of artificial intelligence. Based on the information entered below and all journal articles in the past five years, summarize five research topics. The information entered is as follows:

[0089] 1. Exploration of the application of artificial intelligence technology in the medical field

[0090] Artificial intelligence is an important branch of computer science that involves the research, design and application of intelligent machines. Especially in the field of medical and health care, artificial intelligence has a broader application prospect and higher practical value. The medical expert system created by artificial intelligence technology is applied to various aspects of the medical industry: a universal medical insurance system based on fingerprint recognition technology; a medical examination interpretation expert system; a diagnosis and treatment expert system; a patient monitoring system; a multimedia medical guide system; a nursing plan expert system; and a doctor's order auxiliary processing system. A computer model of the medical expert system is constructed on the basis of production, and scientific steps are taken to establish a medical expert system of a certain scale that is becoming more and more perfect. On the basis of better serving patients, it brings good social response and economic benefits to the hospital.

[0091] …

[0092] 10. Application of artificial intelligence technology in the medical field; 2 Application of artificial intelligence in the medical field; 2.2 Drug development

[0093] Through artificial intelligence technology, computer deep learning and data mining capabilities are used to simulate the drug synthesis process and predict the effects and side effects of drug trials, thereby shortening the time used for drug development and reducing the number of trials, helping to improve research and development efficiency and reduce drug development costs and prices. Currently, artificial intelligence is mainly used in the field of drug development for cardiovascular disease control drugs, tumor control drugs and preventive drugs for certain infectious diseases. "

[0094] The above information is input into the preset large model, and the preset large model generates 5 research topics according to the instruction requirements. The example is as follows:

[0095] 1. Application of generative artificial intelligence in drug development;

[0096] 2. Application of generative artificial intelligence in disease diagnosis and treatment;

[0097] 3. Application of generative artificial intelligence in gene editing and cell therapy;

[0098] 4. Application of generative artificial intelligence in clinical trial design and data analysis;

[0099] 5. The application of generative artificial intelligence in medical image analysis and pathology analysis.

[0100] The present application provides a method and device for generating scientific research topic information, which has a high degree of intelligence and fully understands the user's scientific research topic intention. The preset cross-language text embedding model is fine-tuned according to the document data in the preset document database, so that the text embedding model can learn the semantic relationship between words in the scientific research field. The document data in the preset document database is further combined, and the combined first target data and the second target data are represented as vectors through the text embedding model to obtain single-value vector fields and multi-value vector fields. These vector fields can more comprehensively and accurately represent the content of various aspects of the document to obtain a vector data set. Therefore, vector indexing is performed in the vector data set according to the feature vector of the demand information, and the content most relevant to the user's needs can be found more accurately from a large-scale database in a short time, greatly improving the retrieval speed. Finally, a preset large model is used to automatically generate a variety of scientific research topic information, thereby integrating the processes of document retrieval, scientific research search, etc. with the generation of scientific research topics, opening up user thinking, and improving scientific research efficiency and effectiveness.

[0101] It should be noted that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0102] Furthermore, if Figure 2 As shown, as a specific implementation of the above-mentioned scientific research topic information generation method, the embodiment of the present application provides a scientific research topic information generation device 200, and the scientific research topic information generation device 200 includes: a determination module 201.

[0103] Among them, the determination module 201 is used to input the demand information to be identified into the text embedding model to determine the feature vector of the demand information, wherein the text embedding model is trained based on the document data in the preset document database; and, according to the feature vector, determine the target vector matching the feature vector in the vector data set, wherein the vector data set is obtained by vectorizing the document data in the preset document database according to the text embedding model; and, according to the target vector, determine the target document data associated with the target vector in the preset document database; and, input the target document data into the preset large model to determine the target scientific research topic information corresponding to the demand information.

[0104] In one embodiment, the scientific research topic information generating device 200 further includes:

[0105] A combination module is used to combine the document data according to preset identifiers of the document data in a preset document database according to preset combination rules to determine the first target data and the second target data; wherein the preset identifiers include at least the title of the document, keywords in the document, abstracts in the document, subheadings at all levels in the document, and paragraphs corresponding to the subheadings at all levels in the document; based on a text embedding model, determine a first vector of the first target data and a second vector of the second target data; and determine a vector data set according to the first vector and the second vector.

[0106] In one embodiment, the combination module is specifically used to combine the document data according to a preset identifier and a first preset combination rule to determine the first target data; combine the document data according to a preset identifier and a second preset combination rule to determine the sub-data, and determine the second target data based on the sub-data that meets the same second preset combination rule.

[0107] In one embodiment, the combination module is specifically used to input sub-data in the second target data into the text embedding model, determine the sub-vector corresponding to the sub-data, and determine the second vector of the second target data according to the sub-vector.

[0108] In one embodiment, the determination module 201 is specifically used to perform clustering processing on each vector in the vector data set to determine a first preset number of clusters; calculate a first distance between a feature vector and a first cluster center in the cluster, and sort the clusters according to the first distance to determine an order of the clusters; determine the vectors in the clusters that are located before the first preset position in the order of the clusters as candidate vectors; calculate a second distance between the feature vector and the candidate vector, and sort the candidate vectors according to the second distance to determine the vector order; determine the candidate vector that is located before the second preset position in the vector order as the target vector.

[0109] In one embodiment, the determination module 201 is specifically used to divide the candidate vector into a second preset number of first sub-segments, and determine the position sequence of the first sub-segments according to the order in which the first sub-segments are generated; cluster the first sub-segments of the same position sequence to determine a third preset number of second cluster centers; calculate the third distance between the first sub-segment and the second cluster center corresponding to the first sub-segment, and sort the second cluster centers corresponding to the first sub-segment according to the third distance to determine the center order; update the first sub-segment according to the second cluster center that is first in the center order, and update the candidate vector according to the updated first sub-segment; divide the feature vector into a second preset number of second sub-segments, and determine the position sequence of the second sub-segments according to the order in which the second sub-segments are generated; calculate the fourth distance between the second sub-segment and the second cluster center corresponding to the first sub-segment of the same position sequence as the second sub-segment, and determine the distance list according to the fourth distance; determine the target distance between the feature vector and the second sub-segment and the first sub-segment of the same position sequence in the updated candidate vector according to the distance list, and determine the second distance between the feature vector and the candidate vector according to the target distance.

[0110] In one embodiment, the scientific research topic information generating device 200 further includes:

[0111] A judgment model is used for respectively calculating the candidate second distances between each sub-vector in the candidate vector and the feature vector if the candidate vector includes multiple sub-vectors; sorting the candidate second distances to determine the distance order; and determining the candidate second distance that is at the first position in the distance order as the second distance between the feature vector and the candidate vector.

[0112] In one embodiment, the determination module 201 is specifically used to determine the candidate document title corresponding to the target vector in a preset document database according to the target vector; remove the same titles in the candidate document titles to update the candidate document titles; sort the candidate document titles according to the target vector to determine the document order; determine the target document data according to the document data associated with the candidate document titles located before the third preset position in the document order in the preset document database.

[0113] In one embodiment, the determination module 201 is specifically configured to fine-tune a preset cross-language text embedding model according to document data in a preset document database to determine a text embedding model.

[0114] For the specific limitations of the scientific research topic information generation device, please refer to the limitations of the scientific research topic information generation method above, which will not be repeated here. Each module in the above-mentioned scientific research topic information generation device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0115] Through the description of the above implementation methods, those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform, or the embodiments of the present application can be implemented by hardware.

[0116] Those skilled in the art will appreciate that the accompanying drawings are only schematic diagrams of a preferred implementation scenario, and the modules or processes in the accompanying drawings are not necessarily necessary for implementing the present application. Those skilled in the art will appreciate that the modules in the devices in the implementation scenario can be distributed in the devices of the implementation scenario according to the description of the implementation scenario, or can be changed accordingly and located in one or more devices different from the present implementation scenario. The modules of the above-mentioned implementation scenario can be combined into one module, or can be further split into multiple submodules.

[0117] The above serial numbers of this application are only for description and do not represent the advantages and disadvantages of the implementation scenarios. The above disclosure is only a few specific implementation scenarios of this application, but this application is not limited to them, and any changes that can be thought of by technicians in this field should fall within the scope of protection of this application.

Claims

1. A method for generating scientific research topic information, characterized in that: The method comprises: Inputting the demand information to be identified into a text embedding model to determine a feature vector of the demand information, wherein the text embedding model is trained based on document data in a preset document database; Determining a target vector matched by the feature vector in a vector data set according to the feature vector, wherein the vector data set is obtained by vectorizing the document data in the preset document database according to the text embedding model; Determining target document data associated with the target vector in the preset document database according to the target vector; The target document data is input into a preset macro model to determine the scientific research topic information corresponding to the demand information.

2. The method for generating scientific research topic information according to claim 1, characterized in that: The method further comprises: According to the preset identifiers of the document data in the preset document database, the document data are combined according to the preset combination rules to determine the first target data and the second target data; wherein the preset identifiers at least include the title of the document, the keywords in the document, the abstract of the document, the subheadings at various levels in the document, and the paragraphs corresponding to the subheadings at various levels in the document; Based on the text embedding model, determining a first vector of the first target data and a second vector of the second target data; The vector data set is determined according to the first vector and the second vector.

3. The method for generating scientific research topic information according to claim 2, characterized in that: The preset combination rule includes a first preset combination rule and a second preset combination rule, and combining the document data according to the preset identifier of the document data in the preset document database according to the preset combination rule to determine the first target data and the second target data includes: According to the preset identifier, the document data is combined according to the first preset combination rule to determine the first target data; According to the preset identifier, the document data is combined according to the second preset combination rule to determine the sub-data, and the second target data is determined according to the sub-data that conforms to the same second preset combination rule.

4. The method for generating scientific research topic information according to claim 3, characterized in that: The determining, based on the text embedding model, a second vector of the second target data comprises: The sub-data in the second target data is input into the text embedding model, a sub-vector corresponding to the sub-data is determined, and a second vector of the second target data is determined according to the sub-vector.

5. The method for generating scientific research topic information according to claim 1, characterized in that: The step of determining, in a vector data set according to the feature vector, a target vector that matches the feature vector comprises: Performing clustering processing on each vector in the vector data set to determine a first preset number of clusters; Calculating a first distance between the feature vector and a first cluster center in the clusters, and sorting the clusters according to the first distance to determine an order of the clusters; Determine the vectors in the clusters that are located before the first preset position sequence in the cluster sequence as candidate vectors; Calculating a second distance between the feature vector and the candidate vector, and sorting the candidate vectors according to the second distance to determine a vector order; The candidate vector located before a second preset position in the vector sequence is determined as the target vector.

6. The method for generating scientific research topic information according to claim 5, characterized in that: The calculating a second distance between the feature vector and the candidate vector comprises: Dividing the candidate vector into a second preset number of first sub-segments, and determining a bit sequence of the first sub-segments according to the order in which the first sub-segments are generated; Performing clustering processing on the first sub-segments of the same bit sequence to determine a third preset number of second cluster centers; calculating a third distance between the first subsegment and the second cluster centers corresponding to the first subsegment, and sorting the second cluster centers corresponding to the first subsegment according to the third distance to determine a center order; updating the first subsegment according to the second cluster center that is at the first position in the center sequence, and updating the candidate vector according to the updated first subsegment; Dividing the feature vector into the second preset number of second sub-segments, and determining the bit sequence of the second sub-segments according to the order in which the second sub-segments are generated; calculating a fourth distance between the second subsegment and the second cluster center corresponding to the first subsegment having the same bit sequence as the second subsegment, and determining a distance list according to the fourth distance; According to the distance list, a target distance between the second subsegment and the first subsegment of the same order in the feature vector and the updated candidate vector is determined, and a second distance between the feature vector and the candidate vector is determined according to the target distance.

7. The method for generating scientific research topic information according to claim 6, characterized in that: The method further comprises: If the candidate vector includes a plurality of sub-vectors, dividing the sub-vectors in the candidate vector into the second preset number of first sub-segments, and calculating a candidate second distance between the sub-vectors in the candidate vector and the feature vector; Sorting the candidate second distances to determine a distance order; The candidate second distance that is at the first position in the distance order is determined as the second distance between the feature vector and the candidate vector.

8. The method for generating scientific research topic information according to claim 1, characterized in that: Determining the target document data associated with the target vector in the preset document database according to the target vector includes: Determining, in the preset document database, a candidate document title corresponding to the target vector according to the target vector; Removing identical titles from the candidate document titles to update the candidate document titles; Sort the candidate document titles according to the target vector to determine the document order; The target document data is determined according to the document data associated with the candidate document title located before the third preset position in the document sequence in the preset document database.

9. The method for generating scientific research topic information according to claim 1, characterized in that: The method further comprises: The preset cross-language text embedding model is fine-tuned according to the document data in the preset document database to determine the text embedding model.

10. A device for generating scientific research topic information, characterized in that: The device comprises: A determination module, used for inputting the demand information to be identified into a text embedding model to determine a feature vector of the demand information, wherein the text embedding model is trained based on document data in a preset document database; and Determining a target vector matched by the feature vector in a vector data set according to the feature vector, wherein the vector data set is obtained by vectorizing the document data in the preset document database according to the text embedding model; and Determining target document data associated with the target vector in the preset document database according to the target vector; and The target document data is input into a preset macro model to determine the target scientific research topic information corresponding to the demand information.

Images