Intelligent analysis and service combination method and device for science and technology service demand

Abstract

The application provides a demand intelligent analysis and service combination method for science and technology services, and relates to the technical fields of big data and artificial intelligence, wherein the method comprises the following steps: obtaining demand documents corresponding to historical orders of science and technology services completed through a platform and services contained in the demand documents as training samples; performing document processing on the demand documents in the training samples to generate new training samples; training a constructed contrast learning model using the new training samples and a global loss function, constantly optimizing model parameters until the model converges, and obtaining a trained model; obtaining a demand document to be tested, performing document processing on the demand document to generate atomic demand documents, inputting the atomic demand documents into the trained model, outputting an optimal atomic service and an optimal scheme composed of atomic service combinations, and presenting the optimal scheme to a user in the form of a visual service graph. The application adopting the above scheme can make full use of scientific research resources and reasonably allocate scientific research personnel, and significantly improve the service efficiency of modern science and technology service industry.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data and artificial intelligence, and in particular to a demand intelligent analysis and service combination method and device for technology services. BACKGROUND

[0002] At present, the technology service industry in China is still in its initial stage, and the conversion rate of numerous scientific research achievements of enterprises and research institutions is not high. The technology service platforms cannot exchange, and the resources cannot be converged, which makes the development of the technology service industry in China seriously out of line, hindering the innovation-driven development of technology, economy and society. Therefore, the research and innovation of cross-platform technology resource and service collaboration technology are of great significance to improve the technology innovation support capability and level of modern service industry. However, the current research status of domestic and foreign technology service collaboration platforms is that the technology service providers on the platform cannot automatically obtain the accurate demands of technology service demanders according to demand documents, especially for large-granularity technology service demands, how to scientifically and reasonably decompose them is a great challenge. SUMMARY

[0003] The present application aims to at least solve one of the technical problems in the related art to some extent.

[0004] To this end, the first object of the present application is to provide a demand intelligent analysis and service combination method for technology services, which solves the technical problem that the existing method cannot automatically obtain the accurate demands of technology service demanders according to demand documents, especially for large-granularity technology service demands, how to scientifically and reasonably decompose them is a great challenge, and realizes the full use of scientific research resources and the reasonable allocation of scientific research personnel, significantly improving the service efficiency of modern technology service industry.

[0005] The second object of the present application is to provide a demand intelligent analysis and service combination device for technology services.

[0006] To achieve the above object, the first aspect of the present application provides a demand intelligent analysis and service combination method for technology services, comprising:

[0007] Obtaining the technology service demand documents corresponding to the technology service historical orders completed through the platform and the services contained in the technology service historical orders as training samples;

[0008] Performing document processing on the demand documents in the training samples to generate new training samples, wherein the new training samples include atomic demand graphs and corresponding services;

[0009] Building a contrast learning model, training the contrast learning model using the new training samples and a global loss function, constantly optimizing the model parameters until the model converges, completing the training, and obtaining the trained model;

[0010] Obtaining a requirement document to be tested, performing document processing on the requirement document to be tested, inputting an atomic requirement document generated after processing into a trained model, outputting an optimal scheme composed of an optimal atomic service and an atomic service combination, and presenting the optimal scheme to a user in the form of a visual service graph.

[0011] Optionally, in an embodiment of the present application, the document processing includes:

[0012] Preprocessing and labeling the requirement document;

[0013] Performing event extraction on the preprocessed and labeled requirement document to obtain a large-granularity complex scientific and technological service requirement;

[0014] Decomposing the large-granularity complex scientific and technological service requirement to obtain a requirement triple;

[0015] Performing requirement entity extraction on the requirement triple and the requirement document to generate an atomic requirement entity;

[0016] Establishing an atomic requirement graph based on the extracted atomic requirement entity and a logical relationship between requirement entities.

[0017] Optionally, in an embodiment of the present application, preprocessing and labeling the requirement document includes recognizing and performing semantic analysis on pictures and text information in the document.

[0018] Optionally, in an embodiment of the present application, performing event extraction on the preprocessed and labeled requirement document to obtain a large-granularity complex scientific and technological service requirement includes:

[0019] The BERT model processing includes: performing dropout processing on neurons with weights less than a preset threshold in a hidden layer of the BERT model, and then assigning new weight values to the remaining neurons through a sigmoid function; performing pruning on the hidden layer of the BERT model using saliency, and then retraining the BERT model;

[0020] Using a skip-gram model to construct word embedding vectors for the preprocessed and labeled requirement document, and generating three kinds of embeddings constituted by the requirement document, including word vectors, position vectors, and segmentation vectors;

[0021] Inputting the three kinds of embeddings constituted by the requirement document into the processed BERT model, and outputting a feature sequence;

[0022] Using a log-linear model to represent the joint probability of the entire feature sequence through conditional random field (CRF), and predicting labels in sequence labeling;

[0023] The trigger word extraction and detection is regarded as a word-level labeling task, a BMES sequence labeling rule is adopted, three types of embeddings formed by the demand document are labeled according to the labels in the predicted sequence labeling, and a labeling result is output, the labeling result is a large-granularity complex scientific service demand.

[0024] Optionally, in an embodiment of the present application, the large-granularity complex scientific service demand is decomposed, including:

[0025] All potential relationships of the sentences in the demand document are predicted, so as to complete relationship limitation;

[0026] The overlapping problem between the subject and the object is processed by the sequence labeling component of the relationship limitation;

[0027] A global correspondence component is designed, the subject, the object and the relationship are aligned into a low-complexity triple, and a demand triple is obtained.

[0028] Optionally, in an embodiment of the present application, the demand triple and the demand document are subjected to demand entity extraction, including:

[0029] The demand document is subjected to secondary manual labeling, and the labeling is in the form of triple, the triple includes a demand name, a demand content and a demand category label;

[0030] A word segmenter is created, the demand triple is subjected to feature extraction according to the demand document after secondary labeling, and a content string of the demand triple is converted into an integer index list;

[0031] The demand triple is subjected to one-hot coding according to the integer index list, a feature information sequence matrix is constructed, and the association between the demand triples is represented by the coded feature information sequence matrix;

[0032] According to the association between the demand triples, an adjacency matrix is created to represent the relationship between entities;

[0033] The constructed feature information sequence matrix and the adjacency matrix are input into a weight attention network, and an atomic demand entity is generated.

[0034] Optionally, in an embodiment of the present application, the contrast learning model includes a multi-knowledge graph construction module, a graph contrast learning module and a classifier learning module, and the global loss function is:

[0035] L total =L cls +γL contract

[0036] Wherein, L cls represents the loss function of the classifier, L contract represents the contrast loss function, and γ represents a weight parameter.

[0037]

[0038]

[0039] wherein, W ij represents the output category of the graph node embedding of the knowledge graph after inputting the classifier, represents the true category of the graph node embedding of the knowledge graph, T is a transpose symbol, 1<=j<=dimension d of the feature vector, 1<=i<=number X of known categories, A P represents the output corresponding to the input graph contrast learning module of the knowledge graph G P , i and j represent the i-th and j-th nodes of the knowledge graph respectively, m and n represent the number of nodes in the corresponding processed knowledge graph, and sigma is a sigmoid nonlinear activation function.

[0040] Optionally, in an embodiment of the present application, the generated atomic demand document after processing is input into the trained model, and the optimal atomic service and the optimal scheme composed of atomic service combination are output, including:

[0041] Using the graph construction module, the knowledge graph is constructed from the perspective of scientific and technological services according to the input atomic demand graph;

[0042] Using the contrast learning module, the graph node embedding of the knowledge graph is generated by processing the constructed knowledge graph, wherein the processing of the knowledge graph includes: by means of the constraint of the contrast learning loss function, the consistent representation of the corresponding nodes in different graphs is generated, and the feature representation of different nodes in the same graph is distinctive;

[0043] The classification learning module is used for fusing the graph node embedding of the knowledge graph, and the fused knowledge graph is input into the classifier for classification to obtain the optimal atomic service and the optimal scheme composed of atomic service combination.

[0044] To achieve the above purpose, the second aspect embodiment of the present application proposes a demand intelligent analysis and service combination device for scientific and technological services, comprising an acquisition module, a processing module, a training module and a result generation module, wherein,

[0045] The acquisition module is used for acquiring the scientific and technological service demand documents corresponding to the historical orders of scientific and technological services completed through the platform and the services contained in the historical orders of scientific and technological services as training samples;

[0046] The processing module is used for document processing on the demand documents in the training samples to generate new training samples, wherein the new training samples include atomic demand graphs and corresponding services;

[0047] The training module is configured to build a contrast learning model, train the contrast learning model using new training samples and a global loss function, and constantly optimize model parameters until the model converges, complete the training, and obtain the trained model.

[0048] The result generation module is configured to obtain a to-be-tested requirement document, perform document processing on the to-be-tested requirement document, input an atomic requirement document generated after the processing into the trained model, output an optimal scheme composed of an optimal atomic service and an atomic service combination, and present the optimal scheme to a user in the form of a visual service graph.

[0049] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0050] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the following description, including the accompanying drawings, wherein:

[0051] Figure 1 A flowchart of a requirement intelligent analysis and service combination method for a scientific and technological service provided by an embodiment of the present application;

[0052] Figure 2 A schematic diagram of decomposing a complex large-granularity scientific and technological service requirement in a requirement document by a model constructed by an embodiment of the present application;

[0053] Figure 3 A contrast learning model based on multiple knowledge graphs according to an embodiment of the present application;

[0054] Figure 4 A structural schematic diagram of a requirement intelligent analysis and service combination device for a scientific and technological service provided by an embodiment of the present application. DETAILED DESCRIPTION

[0055] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, in which the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below are exemplary and are intended to explain the present application, and cannot be understood as limiting the present application.

[0056] The requirement intelligent analysis and service combination method and device for a scientific and technological service according to the embodiments of the present application are described below with reference to the accompanying drawings.

[0057] Figure 1 A flowchart of a requirement intelligent analysis and service combination method for a scientific and technological service provided by an embodiment of the present application;

[0058] AsFigure 1 As shown, the demand intelligent analysis and service combination method for technology services includes the following steps:

[0059] Step 101, obtaining the technology service demand document corresponding to the historical technology service order completed through the platform and the service contained in the technology service historical order as a training sample;

[0060] Step 102, document processing is performed on the demand document in the training sample to generate a new training sample, wherein the new training sample includes an atomic demand graph and corresponding services;

[0061] Step 103, constructing a contrast learning model, training the contrast learning model using the new training sample and a global loss function, constantly optimizing the model parameters until the model converges, completing the training, and obtaining the trained model;

[0062] Step 104, obtaining a demand document to be tested, performing document processing on the demand document to be tested, inputting the atomic demand document generated after processing into the trained model, outputting an optimal atomic service and an optimal scheme composed of atomic service combination, and presenting the scheme to the user in the form of a visual service graph.

[0063] The demand intelligent analysis and service combination method for technology services of the embodiments of the present application can automatically and reasonably split large-granularity technology service demands through intelligent analysis of demand documents, and combine and arrange corresponding technology services, and finally generate an optimal service combination scheme. The method proposed in the present application is applied and deployed in a technology service collaboration platform, which can make full use of scientific research resources and reasonably allocate scientific research personnel, and significantly improve the service efficiency of the modern technology service industry.

[0064] Further, in the embodiments of the present application, the document processing includes:

[0065] Pretreatment and labeling of the demand document;

[0066] Event extraction is performed on the pretreated and labeled demand document to obtain a large-granularity complex technology service demand;

[0067] The large-granularity complex technology service demand is decomposed to obtain a demand triple;

[0068] Demand entity extraction is performed on the demand triple and the demand document to generate an atomic demand entity;

[0069] An atomic demand graph is established based on the extracted atomic demand entity and the logical relationship between the demand entities.

[0070] Further, in the embodiments of the present application, all atomic service information published by multiple technology service platforms and historical technology service orders completed through the platforms, including technology service demand documents corresponding to the orders, services contained in the orders and service completion evaluation data, are preprocessed and labeled by using computer vision and natural language processing technology, including identifying and performing semantic analysis on pictures and text information in the documents.

[0071] Specifically, in the embodiments of the present application, the preprocessed and labeled text is input into a BERT-CRF model for event extraction to obtain large-granularity complex technology service demands, including:

[0072] 1) A BERT model pre-trained for Chinese corpus published by Google is used to crop neurons in a hidden layer of the model, dropout processing is performed on neurons with a weight less than a threshold of 0.5 in the hidden layer, and then new weight values are assigned to the remaining neurons by using a sigmoid function; then, the hidden layer of the model is cropped (to 128 layers) by using saliency, and then retrained.

[0073] 2) A user demand document is used to construct a word embedding vector by using a skip-gram model. The input representation of BERT is three Embeddings constituted by the demand document, which are word vectors, position vectors and segmentation vectors.

[0074] 3) A conditional random field CRF is used to predict labels in sequence labeling by using a log-linear model to represent joint probability of the entire feature sequence. Assuming that a sentence length is t, a sentence sequence is N=(n1,n2,…,n n ), a corresponding predicted label sequence is M=(m1,m2,…,m n ), and a final total score of the predicted sequence is:

[0075]

[0076] wherein λ represents a set weighting parameter, is 0.5; D represents a transition score between labels; F i,mi represents a score of each word to a corresponding m i label, and b represents an adjustment parameter, is 5. Since there are multiple possibilities for the predicted sequence, all possible sequences are globally normalized to generate a probability of the original sequence to the predicted sequence.

[0077] 4) Trigger word extraction and detection are regarded as a word-level labeling task, and BMES sequence labeling rules are adopted. Three types of embeddings formed by the demand document are labeled according to the predicted label in the sequence labeling, and the labeling result is output. The labeling result is a complex technology service demand with large granularity. Among them, B represents the beginning character of the word, M represents the middle character, E represents the end character, and S represents the single word.

[0078] Further, in the embodiment of the present application, the present application constructs a model to decompose the complex large-granularity technology service demand in the demand document from the perspective of technology service, as shown in Figure 2 . Specifically, the entity-relation joint extraction task is decomposed into three sub-tasks of relation judgment, entity extraction and subject-object alignment, and the proposed triple joint extraction framework effectively improves the problems of relationship prediction redundancy, poor generalization ability based on span extraction and low efficiency in the existing method.

[0079] 1) First, the model predicts all potential relationships of the sentence in the demand document. It limits the subsequent entity extraction to the predicted relationship subset rather than all relationships. Given a high-dimensional embedding vector For a sentence with n tokens, each element of the component is as follows:

[0080] A max = Maxpool (A)

[0081] F r = σ (W r A max + λ r )

[0082] Where Maxpool represents the maximum pooling operation, A represents the high-dimensional embedding vector output by the encoder, F r is the probability of the predicted potential relationship, Wr represents the training weight, σ represents the Sigmoid function, and λ r represents the adjustment factor.

[0083] 2) Then, the overlapping problem between the subject and the object is handled by the relationship-limited sequence labeling component. The relationship-limited formula is as follows:

[0084] X i,j s = Soft max (W s (α i + β j )+ ε s )

[0085] X i,j o = Soft max (W o (αi +β j )+ε o )

[0086] where X i,j s denotes the existence probability of the subject after the i,j relationship restriction calculation, X i,j o denotes the existence probability of the object after the i,j relationship restriction calculation, β j is the encoding representation of the jth relationship restriction in the trainable embedding matrix B, α i denotes the encoding representation of the ith relationship restriction, W s , W o denotes a trainable weight, ε s , ε o denotes an adjustment factor.

[0087] 3) Finally, a global correspondence component is designed to align the subject, object and relationship into a low complexity triple. Each element of the matrix is about the starting position of the paired subject and object, indicating the confidence level of the subject-object pair. The higher the value, the higher the confidence level of the triple.

[0088]

[0089] where, denotes the confidence of the triple, A s i and A o j denotes the encoding representation of the ith token and the jth token in the input sentence, forming a potential subject and object pair, W k denotes a trainable weight, τ is a Sigmoid function.

[0090] Further, in the embodiments of the present application, the above-mentioned demand triple and the above-mentioned demand document are subjected to demand entity extraction, including:

[0091] 1) The demand document is manually labeled again as a (demand name, demand content, demand category label) triple.

[0092] 2) A common word tokenizer with 1024 dimensions is created, the demand triple is extracted according to the twice-labeled demand document, the content string of the demand triple is converted into an integer index list, the demand triple is one-hot encoded according to the index, a feature information sequence matrix is constructed, the association between the demand triples is represented by the encoded feature information sequence matrix, and then an adjacency matrix is created to represent the relationship between entities according to the association between the extracted atomic demand triples.

[0093] 3) According to the feature sequence matrix and the adjacency matrix of the constructed demand triplets, input them into a WAB-GCN (Weight Attention Based-GCN) weight attention network, the input dimension is 512, the hidden layer dimension is set to 32, the number of neurons in the last layer is set to 4 (the number of service categories), and the activation function selected is a Mish function, thereby generating atomic demand entities.

[0094] Based on the extracted atomic demand entities and the logical relationships between them, a demand graph is established, and a visual atomic demand graph for science and technology services is obtained using VUE.js technology.

[0095] Further, in the embodiments of the present application, for the zero-shot classification problem in the science and technology service collaboration platform, the present application proposes a novel contrast learning model based on multiple knowledge graphs (research and development, technology transfer, intellectual property, testing and detection), as shown in the structure of the contrast learning model, Figure 3 including a multi-knowledge graph construction module, a graph contrast learning module, and a classifier learning module, the proposed method can simultaneously utilize multiple knowledge relationships between categories to train a classifier that is robust and has high discriminative ability for unknown category services.

[0096] As shown in Figure 3 , four knowledge graphs corresponding to science and technology services (G a , G b , G c , and G d ) are constructed as inputs, the model aims to learn more internal relationship representation information between different categories, and learn an effective classifier for unseen categories (i.e. zero-shot learning), the proposed model includes three main modules, namely multi-knowledge graph construction, graph contrast learning, and classifier learning.

[0097] • Multi-knowledge graph construction, four knowledge graphs are constructed from the perspective of science and technology services to capture more internal relationship information between different categories, and each category of science and technology services corresponds to a knowledge graph.

[0098] • Graph contrast learning, the purpose of the graph contrast learning module is to force the corresponding node representations in different graphs to be consistent, and the feature representations of different nodes in the same graph to be distinctive.

[0099] • Classifier learning, we obtain the final classifier by fusing the graph node embeddings from the four different knowledge graphs, and further learn the classifier in a semi-supervised manner, specifically:

[0100] First, by fusing the node embeddings of the four knowledge graphs, we obtain a classifier for n categories

[0101]

[0102] where, contains x known classes and y unknown classes, i.e.,

[0103]

[0104] The loss function of the classifier is defined as:

[0105]

[0106] In addition, if there is an unknown class sample in the test sample, first use the pre-trained CNN network to extract the feature f t ∈R d Then use the learned classifier to obtain its classification label: i.e.

[0107] In the graph contrastive learning module, G a ,G b ,G c and G d represent the input knowledge graph of the GCN network, A a ,A b ,A c and A d represent the output of the four GCN network branches, and their target loss function is defined as follows:

[0108]

[0109]

[0110]

[0111]

[0112] where σ is the sigmoid nonlinear activation function. In addition,

[0113] The definition of is similar to the above loss function.

[0114] Finally, the overall loss function of the proposed graph contrastive learning model is defined as:

[0115]

[0116] where W ij represents the output class of the graph node embedding of the knowledge graph after inputting the classifier, True class of graph node embedding representing a knowledge graph, T is a transpose symbol, 1 <= j <= dimension d of a feature vector, 1 <= i <= number X of known classes, A P A knowledge graph G P Output corresponding to the input graph contrast learning module, i and j represent the i and j nodes of the knowledge graph respectively, m and n represent the number of nodes in the corresponding processed knowledge graph, and sigma is a sigmoid nonlinear activation function.

[0117] The classifier loss function and the loss function of the graph contrast learning model are added to obtain a global loss function, defined as:

[0118] L total = L cls + γL contract

[0119] Wherein, gamma is a weight parameter.

[0120] Further, in the embodiment of the present application, the generated atomic demand document after processing is input into the trained contrast learning model, and the optimal atomic service and the optimal scheme composed of atomic service combination are output, including:

[0121] Using the graph construction module, a knowledge graph is constructed from the perspective of scientific and technological services according to the input atomic demand graph;

[0122] Using the contrast learning module, the graph node embedding of the constructed knowledge graph is generated by processing the constructed knowledge graph, wherein the knowledge graph is processed, including: through the constraint of the contrast learning loss function, the corresponding node representation in different graphs generated is consistent, and the feature representation of different nodes in the same graph is distinctive;

[0123] Using the classification learning module, the graph node embedding of the knowledge graph is fused, and the fused knowledge graph is input into the classifier for classification to obtain the optimal atomic service and atomic service combination,

[0124] Finally, through the VUE.js technology, the results output by the model form a service optimal combination scheme, and are presented to the user in the form of a visual service graph.

[0125] In order to realize the above-mentioned embodiment, the present application further provides a demand intelligent analysis and service combination device for scientific and technological services,

[0126] Figure 4 A structural schematic diagram of a demand intelligent analysis and service combination device for scientific and technological services provided in the embodiment of the present application.

[0127] As Figure 4As shown, the demand intelligent analysis and service combination device for technology service-oriented services includes an acquisition module, a processing module, a training module, and a result generation module, wherein,

[0128] The acquisition module is configured to acquire technology service demand documents corresponding to historical technology service orders completed through a platform and services contained in the historical technology service orders as training samples.

[0129] The processing module is configured to perform document processing on the demand documents in the training samples to generate new training samples, wherein the new training samples include atomic demand graphs and corresponding services.

[0130] The training module is configured to construct a contrast learning model, train the contrast learning model using the new training samples and a global loss function, constantly optimize model parameters until the model converges, complete the training, and obtain the trained model.

[0131] The result generation module is configured to acquire a demand document to be tested, perform document processing on the demand document to be tested, input the atomic demand documents generated after the processing into the trained model, output an optimal atomic service and an optimal scheme composed of atomic service combinations, and present the optimal scheme to a user in the form of a visual service graph.

[0132] It should be noted that the foregoing description of the demand intelligent analysis and service combination method for technology service-oriented services is also applicable to the demand intelligent analysis and service combination device for technology service-oriented services of this embodiment, which will not be described here.

[0133] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples without contradiction.

[0134] In addition, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise specifically limited.

[0135] Any processes or methods described in the flowcharts or otherwise described herein can be understood as representing modules, segments, or portions of code that include one or more executable instructions for implementing specific logic functions (or steps) and / or can be implemented entirely in hardware. The various embodiments of the application can be further implemented in a computer program product tangibly embodied in a machine-readable storage medium (e.g., memory) including instructions

[0136] Logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be embodied in computer-readable instructions, modules, segments, or portions of codes, which are executable by a processing system, apparatus, or device, such as a computer-based system, processor- based system, or other system that can fetch instructions from a machine-readable medium and execute the instructions. For the purposes of this specification, a "computer-readable medium" can be any apparatus that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. The computer-readable medium can specifically include a hardware apparatus (e.g., a wired or wireless communication device) that is programmed to operate as described herein, as well as a hardware apparatus (e.g., a wired or wireless communication device) that is structurally configured to operate as described herein. The computer-readable medium can be non-transitory, in that it does not include a transitory, propagating signal per se (e.g., a

[0137] It should be understood that aspects of the application can be implemented in hardware, software, firmware, or combinations thereof. In the above embodiments, various steps or methods can be implemented in software or firmware that is stored in memory and executed by a suitable instruction execution system. As well, if desired, the functions can be implemented in hardware, as opposed to software, or combinations thereof. If desired, the functions can be implemented using any of or a combination of the following technologies, which are all well known in the art: a discrete logic circuit(s) having logic gates for implementing logic functions upon an application of data signals; an application specific integrated circuit(s) having appropriate combinational logic gates; a programmable gate array(s) (PGA), a field programmable gate array(s) (FPGA), etc.

[0138] Those skilled in the art of the present technology can understand that all or part of the steps carried out by the above-mentioned embodiment method can be completed by programs instructing related hardware, and the programs can be stored in a computer readable storage medium. When the program is executed, it includes one of the steps of the method embodiment or a combination thereof.

[0139] In addition, each functional unit in each embodiment of the present application can be integrated into one processing module, or each unit can exist physically alone, or two or more units can be integrated into one module. The integrated module can be realized in the form of hardware or in the form of a software functional module. When the integrated module is realized in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer readable storage medium.

[0140] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc. Although the embodiments of the present application have been shown and described above, it should be understood that the above-mentioned embodiments are exemplary and cannot be understood as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above-mentioned embodiments within the scope of the present application.

Claims

1. A method for intelligent demand analysis and service combination for scientific and technological services, characterized in that: The following steps are involved: Obtain the technology service demand documents corresponding to the historical technology service orders completed through the platform and the services included in the historical technology service orders as training samples; Performing document processing on the requirement documents in the training samples to generate new training samples, wherein the new training samples include atomic requirement graphs and corresponding services; Constructing a contrastive learning model, training the contrastive learning model using the new training samples and the global loss function, continuously optimizing model parameters until the model converges, completing the training, and obtaining a trained model; Obtaining a requirement document to be tested, performing the document processing on the requirement document to be tested, inputting the atomic requirement document generated after the processing into the trained model, outputting the optimal atomic service and the optimal solution composed of the atomic service combination, and presenting it to the user in the form of a visual service map; The document processing includes: Preprocess and annotate requirement documents; Perform event extraction on pre-processed and annotated requirement documents to obtain large-scale, complex technology service requirements; Decomposing the large-scale complex technology service requirements to obtain requirement triples; Extracting requirement entities from the requirement triples and the requirement document to generate atomic requirement entities; Establish an atomic requirement graph based on the extracted atomic requirement entities and the logical relationships between the requirement entities; The event extraction is performed on the pre-processed and annotated demand documents to obtain large-scale complex technology service requirements, including: Processing the BERT model includes: dropping out neurons in the hidden layer of the BERT model whose weights are less than a preset threshold, and then assigning new weights to the remaining neurons using the sigmoid function; pruning the hidden layer of the BERT model using significance, and then retraining the BERT model; Use the skip-gram model to construct word embedding vectors for the preprocessed and annotated requirements documents, generating three embeddings of the requirements documents, including word vectors, position vectors, and segmentation vectors. The three embeddings of the requirement document are input into the processed BERT model, and a feature sequence is output; Through the conditional random field CRF, the log-linear model is used to represent the joint probability of the entire feature sequence and predict the label in the sequence annotation, including: Assume that the sentence length is t and the sentence sequence is N=(n1,n2,…,n n ), the corresponding predicted label sequence is M=(m1,m2,…,m n ), the final total score of the predicted sequence is: Among them, λ represents the weighted parameter set, D represents the transfer score between labels, Indicates that each word corresponds to m i The score of the label, b represents the adjustment parameter, and global normalization is performed on all possible sequences to generate the probability of the original sequence to the predicted sequence; The extraction and detection of trigger words are regarded as word-level annotation tasks. The BMES sequence annotation rules are adopted to annotate the three embeddings of the requirement document according to the labels in the predicted sequence annotation, and the annotation results are output. The annotation results are the large-grained complex scientific and technological service requirements.

2. The method according to claim 1, wherein The pre-processing and annotation of the demand document includes: identifying and semantically analyzing the images and text information in the document.

3. The method according to claim 2, wherein Decomposing the large-scale complex technology service requirements includes: Predicting all potential relationships between sentences in the requirement document, thereby completing relationship definition; The overlap between subject and object is handled through a relation-qualified sequence tagging component; Design a global correspondence component to align the subject, object, and relation into a low-complexity triple to obtain the requirement triple.

4. The method according to claim 3, wherein The extracting requirement entities from the requirement triples and the requirement document includes: Perform secondary manual annotation on the requirement document, where the annotation is in the form of triples, which include the requirement name, requirement content, and requirement category label; Create a word segmenter, extract features from the requirement triples according to the secondary annotated requirement document, and convert the content string of the requirement triples into an integer index list; One-hot encode the demand triples according to the integer index list, construct a feature information sequence matrix, and represent the association between the demand triplets through the encoded feature information sequence matrix; According to the association between the requirement triples, an adjacency matrix is ​​created to represent the relationship between entities; The constructed feature information sequence matrix and adjacency matrix are input into a weighted attention network to generate atomic demand entities.

5. The method according to claim 1, wherein The contrastive learning model includes a multi-knowledge graph construction module, a graph contrastive learning module, and a classifier learning module. The global loss function is: THE total =L cls +γL contract Among them, L cls represents the loss function of the classifier, L contract represents the contrast loss function, γ represents the weight parameter, Among them, W ij Represents the output category of the graph node of the knowledge graph after embedding the input classifier, Represents the true category of the graph node embedding of the knowledge graph, T is the transposition symbol, 1<=j<=the dimension d of the feature vector, 1<=i<=the number of known categories X, A P Represents the knowledge graph G P The corresponding output after the input graph comparison learning module, i, j respectively represent the i-th and j-th nodes of the knowledge graph, m, n represent the number of nodes in the corresponding processed knowledge graph, and σ is the sigmoid nonlinear activation function.

6. The method according to claim 5, wherein The step of inputting the processed atomic requirement document into the trained model and outputting the optimal solution consisting of the optimal atomic service and the combination of the atomic services includes: Using the graph construction module, a knowledge graph is constructed based on the input atomic demand graph from the perspective of scientific and technological services; Using the contrastive learning module, processing the constructed knowledge graph to generate graph node embeddings for the knowledge graph, wherein the processing of the knowledge graph includes: constraining the knowledge graph using the contrastive learning loss function so that corresponding node representations in different generated graphs are consistent and feature representations of different nodes in the same graph are distinctive; The classification learning module is used to fuse the graph node embeddings of the knowledge graph, and the fused knowledge graph is input into the classifier for classification to obtain the optimal atomic service and atomic service combination.

7. A device for intelligent analysis of demand and service combination for scientific and technological services, characterized in that: It includes acquisition module, processing module, training module and result generation module, among which, The acquisition module is used to acquire the technology service demand documents corresponding to the historical technology service orders completed through the platform and the services included in the historical technology service orders as training samples; The processing module is configured to perform document processing on the requirement documents in the training samples to generate new training samples, wherein the new training samples include atomic requirement graphs and corresponding services; The training module is used to build a contrastive learning model, train the contrastive learning model using the new training samples and the global loss function, continuously optimize the model parameters until the model converges, complete the training, and obtain a trained model; The result generation module is used to obtain the requirement document to be tested, perform the document processing on the requirement document to be tested, input the atomic requirement document generated after the processing into the trained model, output the optimal atomic service and the optimal solution composed of the atomic service combination, and present it to the user in the form of a visual service map; The document processing includes: Preprocess and annotate requirement documents; Perform event extraction on pre-processed and annotated requirement documents to obtain large-scale, complex technology service requirements; Decomposing the large-scale complex technology service requirements to obtain requirement triples; Extracting requirement entities from the requirement triples and the requirement document to generate atomic requirement entities; Establish an atomic requirement graph based on the extracted atomic requirement entities and the logical relationships between the requirement entities; The event extraction is performed on the pre-processed and annotated demand documents to obtain large-scale complex technology service requirements, including: Processing the BERT model includes: dropping out neurons in the hidden layer of the BERT model whose weights are less than a preset threshold, and then assigning new weights to the remaining neurons using the sigmoid function; pruning the hidden layer of the BERT model using significance, and then retraining the BERT model; Use the skip-gram model to construct word embedding vectors for the preprocessed and annotated requirements documents, generating three embeddings of the requirements documents, including word vectors, position vectors, and segmentation vectors. The three embeddings of the requirement document are input into the processed BERT model, and a feature sequence is output; Through the conditional random field CRF, the log-linear model is used to represent the joint probability of the entire feature sequence and predict the label in the sequence annotation, including: Assume that the sentence length is t and the sentence sequence is N=(n1,n2,…,n n ), the corresponding predicted label sequence is M=(m1,m2,…,m n ), the final total score of the predicted sequence is: Among them, λ represents the weighted parameter set, D represents the transfer score between labels, Indicates that each word corresponds to m i The score of the label, b represents the adjustment parameter, and global normalization is performed on all possible sequences to generate the probability of the original sequence to the predicted sequence; The extraction and detection of trigger words are regarded as word-level annotation tasks. The BMES sequence annotation rules are adopted to annotate the three embeddings of the requirement document according to the labels in the predicted sequence annotation, and the annotation results are output. The annotation results are the large-grained complex scientific and technological service requirements.

Images