Intelligent search recommendation method and system

Abstract

The application provides a kind of intelligent retrieval recommendation method and system, wherein the method comprises: determining the semantic label corresponding to each document in corpus, and the entity and inter-entity relationship corresponding to each document; based on the entity and inter-entity relationship corresponding to each document in the corpus, and the transE model trained, construct fragmentation knowledge net; based on the search term input by user, determine the search result matched with the search term; according to the user portrait of the user, determine the recommendation order of each target document in the target document set.The method provided by the application enables the user to obtain the required related document according to the search term, while obtaining other fragmentation knowledge associated with the semantic label of the document, in addition, the priority of all documents obtained by retrieval can be recommended according to the user portrait, to provide the retrieval result to the user more flexibly, and to improve the effect of retrieval more effectively.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mass data resource analysis processing, and particularly relates to an intelligent retrieval recommendation method and system. BACKGROUND

[0002] With the continuous development of Web2.0 technology, the amount of information on the Internet is growing exponentially. The mass of information resources far exceeds the information needs of individuals, so that users cannot obtain effective information in time according to the demand and use, which causes the sharp contradiction between information producers and content consumers - information overload.

[0003] Under the background of big data, the development of deep learning and neural network improves the user's processing capacity of information, but does not alleviate the impact of information overload on users. The search engine is one of the most effective ways to solve information overload, which can display relevant document information to users according to a certain strategy. This search is a coarse-grained retrieval, that is, returning a document list meeting the user's demand from the document set. In this era of information overload, document-level retrieval will bring users the burden of reading, and the user needs to read multiple documents to obtain a relatively complete knowledge structure through a number of scattered knowledge fragments. Thus, another coarse-grained information overload problem is caused.

[0004] The search engine provides a very important technical means to solve the traditional cognitive information overload problem, which matches information in the background according to the query keywords input by the user. However, if the user cannot provide keywords accurately describing his own demand, the search engine cannot serve the user well. Therefore, how to mine user demand and recommend appropriate document content to the user according to the user demand has gradually become a research hotspot in the information field. SUMMARY

[0005] In view of the problems in the prior art, the present application provides an intelligent retrieval recommendation method, device, equipment and storage medium, to solve the technical problems of information overload and inability to intelligently recommend related information resources according to user demand in the prior art.

[0006] In the first aspect, the present application provides an intelligent retrieval recommendation method, comprising:

[0007] determining semantic tags corresponding to each document in a corpus, and entities and inter-entity relationships corresponding to each document;

[0008] constructing a fragmented knowledge network based on the entities and inter-entity relationships corresponding to each document in the corpus and a trained translation distance model transE;

[0009] determining a search result matched with the search term based on the search term input by the user; the search result comprising a target document set matched with the search term, and a fragmented knowledge sub-network matched with a target semantic label;

[0010] determining a recommended order of each target document in the target document set according to a user portrait of the user;

[0011] wherein the fragmented knowledge sub-network is determined based on a matching relationship between the target semantic label and the entity in the fragmented knowledge network, and comprises a target entity matched with the target semantic label, and a sub-network composed of a triple having an inter-entity relationship with the target entity; the target semantic label is a semantic label corresponding to any target document in the target document set.

[0012] Optionally, the determining of the semantic label corresponding to each document in the corpus comprises:

[0013] determining all topic words corresponding to all documents in the corpus as a standard topic word library;

[0014] determining a first candidate word set corresponding to a target document based on a term frequency-inverse document frequency (TF-IDF) algorithm and the standard topic word library; the target document being any document in the corpus;

[0015] determining whether a first target word belongs to a candidate semantic label set corresponding to the target document based on a first judgment condition; the first target word being any word in the first candidate word set;

[0016] the first judgment condition comprising whether the first target word belongs to the standard topic word library and a similarity between the first target word and a target topic word;

[0017] determining the top N1 candidate semantic labels with the highest correlation degrees as the semantic labels corresponding to the target document based on a correlation degree between any candidate semantic label in the candidate semantic label set and the target document; wherein N1 is a positive integer greater than or equal to 1.

[0018] Optionally, before the determining of all topic words corresponding to all documents in the corpus, the method comprises:

[0019] extracting topics corresponding to all documents in the corpus and determining the distribution of documents and topics and the distribution of topics and words based on a partitioned latent Dirichlet allocation model.

[0020] Optionally, the determining of all topic words corresponding to all documents in the corpus comprises:

[0021] determine a second candidate word set belonging to the target theme based on the theme and the distribution of the words, the target theme being any one of themes corresponding to all documents in the corpus;

[0022] determine the first N2 target words with the maximum first chi-square values as theme words belonging to the target theme, and add the theme words to the standard theme word library, the first chi-square value being a chi-square value of the second target word and the target theme, the second target word being any word in the second candidate word set, and N2 being a positive integer greater than or equal to 1.

[0023] Optionally, the determining whether the first target word belongs to the candidate semantic label set corresponding to the target document based on the first judgment condition comprises:

[0024] determining whether the first target word belongs to the standard theme word library;

[0025] if yes, determining that the first target word belongs to the candidate semantic label set corresponding to the target document;

[0026] if no, determining, based on a similarity between the first target word and any theme word in the standard theme word library, to add the first target word to the candidate semantic label set corresponding to the target document by converting the first target word into the theme word with the highest similarity to the first target word.

[0027] Optionally, the determining the entity and the inter-entity relationship corresponding to each document comprises:

[0028] marking each word corresponding to a target sentence as a target sequence, the target sentence being any sentence in a target document, and the target document being any document in the corpus;

[0029] determining the entity and the inter-entity relationship corresponding to the target sequence based on a joint model LSTM-LSTM-Bias;

[0030] the entity at least includes all theme words corresponding to all documents in the corpus.

[0031] Optionally, the training method of the transE model comprises:

[0032] determining a positive sample based on a triple formed by the entity and the inter-entity relationship corresponding to each document in the corpus, and a non-similarity measure d(S) corresponding to the positive sample;

[0033] replacing a head entity or a tail entity in the positive sample so as to satisfy that the positive sample is not a triple in the positive sample, determining a negative sample and a non-similarity measure d(S') corresponding to the negative sample.

[0034] determining the boundary function when the boundary function

[0035] the boundary function is expressed as:

[0036]

[0037] and S' (h，l，t) ={(h',l,t)|h'∈E}∪{(h,l,t')|t'∈E};

[0038] S={(h,l,t)|h∈E,t∈E}

[0039] wherein [γ+d(S)-d(S')] + represents that γ+d(S)-d(S') takes a positive value, γ is a boundary hyperparameter, S represents a positive sample, S' (h，l，t) represents a negative sample, (h',l,t) and (h,l,t') both represent a triple belonging to a negative sample, h' and h both represent a head entity, t and t' both represent a tail entity, l represents a relationship between entities, E represents a set of all entities, and d() represents a dissimilarity measure.

[0040] Optionally, before determining the recommendation order of the target document set according to the user portrait of the user, the method comprises:

[0041] determining, based on a search log of the user, commonly used search words of the target user and corresponding search frequencies, and selecting the first N3 search words with the most search frequencies as a search history portrait of the target user;

[0042] determining, based on basic information of the user, commonly used search words of other users belonging to the same department as the target user and corresponding search frequencies, and selecting the first N3 search words with the most search frequencies as a department portrait of the target user;

[0043] determining, based on an operation log of the user, semantic tags corresponding to all documents operated by the target user, and selecting the first N3 search words with the highest appearance frequencies as a document operation portrait of the target user;

[0044] determining, based on a first topic set, topic words and auxiliary words corresponding to each topic in the first topic set as a research field portrait of the target user, the first topic set being composed of topics corresponding to all documents operated by the target user;

[0045] determine a customized profile of the target user based on user preferences or needs;

[0046] determine a search history profile of the target user, a department profile of the target user, a document operation profile of the target user, a research field profile of the target user, and the customized profile of the target user as the user profile of the target user;

[0047] wherein N3 is a positive integer greater than or equal to 1.

[0048] Optionally, the determining of the recommendation order of each of the target documents in the target document set based on the user profile of the user comprises:

[0049] determining a matching value of the target semantic label and the user profile of the target user; the target semantic label is any semantic label corresponding to any target document in the target document set;

[0050] determining a comprehensive matching value based on the matching value of the target document and the search term, and the matching value of the target semantic label and the user profile of the target user;

[0051] determining the recommendation order of each of the target documents in the target document set in descending order of the comprehensive matching value.

[0052] In a second aspect, the present application further provides an intelligent search recommendation system, comprising:

[0053] a determining module configured to determine semantic labels corresponding to each document in a corpus, and entities and inter-entity relationships corresponding to each of the documents;

[0054] a constructing module configured to construct a fragmented knowledge network based on the entities and inter-entity relationships corresponding to each of the documents in the corpus, and a trained transE model;

[0055] a search module configured to determine search results matching a search term input by a user based on the search term; the search results comprising a target document set matching the search term, and a fragmented knowledge subnetwork matching a target semantic label;

[0056] a sorting module configured to determine a recommendation order of each of the target documents in the target document set based on a user profile of the user;

[0057] wherein the fragmented knowledge subnetwork is determined based on a matching relationship between the target semantic label and the entities in the fragmented knowledge network, and comprises a subnetwork composed of a target entity matching the target semantic label, and triplets having an inter-entity relationship with the target entity; the target semantic label is a semantic label corresponding to any target document in the target document set.

[0058] In a third aspect, the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the intelligent search recommendation method according to the first aspect when executing the program.

[0059] In a fourth aspect, the present application provides a non-transitory computer readable storage medium, having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the intelligent search recommendation method according to the first aspect.

[0060] In a fifth aspect, the present application provides a computer program product, comprising a computer program, wherein the computer program, when executed by a processor, implements the intelligent search recommendation method according to any one of the above aspects.

[0061] The intelligent search recommendation method and system provided by the present application determines the corresponding semantic labels, entities and relationships between entities by analyzing each document in the corpus, constructs a fragmented knowledge network based on the trained transE model, and then determines the matching target document set and the fragmented knowledge sub-network having a matching relationship with the semantic label of the target document according to the user search term, and can perform recommendation according to different priorities according to the user portrait. The user can obtain the required related documents according to the search term, and at the same time obtain other fragmented knowledge associated with the semantic label of the document. In addition, the user portrait can be used to sort the recommendation priority of all documents obtained by searching, to provide more flexible search results to the user and more effectively improve the search effect. BRIEF DESCRIPTION OF DRAWINGS

[0062] In order to more clearly illustrate the technical solutions in the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0063] Figure 1 is a flowchart of the intelligent search recommendation method provided by the embodiment of the present application;

[0064] Figure 2 is a schematic diagram of the LDAP model provided by the embodiment of the present application;

[0065] Figure 3 is a flowchart of determining the subject words by using the chi-square value provided by the embodiment of the present application;

[0066] Figure 4 is a flowchart of determining the semantic labels of the documents provided by the embodiment of the present application;

[0067] Figure 5 is a structural schematic diagram of the intelligent retrieval recommendation system provided by an embodiment of the present application;

[0068] Figure 6 is a flow schematic diagram of the intelligent retrieval recommendation system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0069] To make the objectives, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described below in detail with reference to the drawings in the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0070] Figure 1 is a flow schematic diagram of the intelligent retrieval recommendation method provided by an embodiment of the present application, as shown in Figure 1 the method comprises:

[0071] Step 101, determining the semantic tags corresponding to each document in the corpus, and the entities and the relationships between the entities corresponding to each document;

[0072] Specifically, there are various existing document types, such as patent files, academic papers, journal articles, conference articles, design documents, etc. In order to obtain a corpus with a wider coverage and richer content, more types of documents can be collected as the corpus. This is more conducive to users to obtain more relevant knowledge when retrieving.

[0073] On this basis, the semantic tags corresponding to each document are determined, which can be simply understood as the related words that can summarize the main content of the document, such as the form of key words in the abstract in academic papers or journal articles, or the subject words corresponding to design documents, etc.

[0074] Further, the extraction of entities and entity relationships of the above-mentioned documents can be understood as information extraction (IR), which is a text processing technology that extracts specified types of entity, relationship, event, etc. from natural language text and forms structured data output. It includes named entity recognition and entity relationship extraction. The task of named entity recognition (NER) is to find each named entity mentioned in the text and mark its type, that is, to determine the entity corresponding to the document. Entity relationship extraction refers to extracting predefined entity relationships from unstructured text on the basis of entity recognition.

[0075] The current learning methods of named entity and entity relation extraction are mainly divided into two categories: pipeline learning method and joint learning method. The pipeline learning method means that the relation extraction between entities is directly carried out on the basis of the completion of entity recognition; the joint learning method is mainly an end-to-end model based on neural network, which simultaneously completes the identification of entities and the extraction of relations between entities. Although the pipeline model has achieved good results in extraction, the pipeline model still has many problems: the errors of the entity recognition module will affect the performance of the relation classification; the relationship between the two sub-tasks is ignored; the entity pair without relation will cause information redundancy, thereby increasing the error rate. Compared with the pipeline method, the joint model can utilize the close interaction information between entities and relations, simultaneously extract entities and classify the relations of entity pairs, and well solve the problems existing in the pipeline method.

[0076] Step 102, based on the entities and the relations between entities corresponding to each document in the corpus and the transE model trained, a fragmented knowledge network is constructed;

[0077] Specifically, after extracting the entities and the relations between entities corresponding to each document in the corpus, the document information can be converted into numerous triples, which are specifically represented as: (h, l, t), h represents the head entity, t represents the tail entity, and l represents the relation between entities, h, t ∈ E represents any two entities h, t in the entity set E; l ∈ L represents a relation l in the relation set L.

[0078] However, the expression ability of the triple is limited. If the head entity and the tail entity in the triple are connected through a path in the graph, more rich information can be obtained. First, the triple itself will be retained in the path; second, the path contains long chain information between multiple triples, which can express more complex relations.

[0079] Further, according to the transE model trained, the triples contained in the document set under each topic are mapped into a multi-relation directed graph, so as to generate a fragmented knowledge network. The fragmented knowledge network is mainly established in units of topics. A same topic corresponds to multiple topic words, and each topic word is associated with multiple documents, so that the fragmented knowledge network associated with the topic words and the documents is constructed.

[0080] The translation distance model transE is relatively simple and has good performance, and mainly realizes learning of low-dimensional embedding vectors of entities and relations in the knowledge base.

[0081] Step 103, based on the search term input by the user, determining a search result matched with the search term; the search result includes a target document set matched with the search term, and a fragmented knowledge sub-network matched with the target semantic label;

[0082] The fragmented knowledge sub-network is determined based on a matching relationship of the target semantic label and the entity in the fragmented knowledge network, and includes a sub-network composed of a target entity matched with the target semantic label and a triple having an inter-entity relationship with the target entity. The target semantic label is a semantic label corresponding to any target document in the target document set.

[0083] Specifically, after the fragmented knowledge network is constructed, the search engine retrieves one or more target documents matched with the search term input by the user to form a target document set, and each target document has a corresponding semantic label. Each semantic label has a matching relationship with an entity in the fragmented knowledge network, that is, according to any semantic label, a target entity can be found in the fragmented knowledge network, and the target entity can have other entities having an inter-entity relationship with the target entity in the fragmented knowledge network. Thus, the target entity, other entities associated with the target entity, and the inter-entity relationship therebetween, that is, the triple corresponding to the target entity, form a fragmented knowledge sub-network. Of course, the other entities having an inter-entity relationship with the target entity are mainly one-layer inter-entity relationship, or direct inter-entity relationship, rather than a connection between two entities through multi-layer inter-entity relationship. Each target document has a semantic label, and thus a corresponding fragmented knowledge sub-network also exists.

[0084] The above is for determining the fragmented knowledge sub-network corresponding to the semantic label of a single target document. Similarly, each target document can obtain a corresponding fragmented knowledge sub-network in the same way.

[0085] Step 104: determining a recommended order of each target document in the target document set according to the user portrait of the user.

[0086] Specifically, after the search engine matches the target document set and the fragmented knowledge sub-network matched with the target semantic label according to the search term input by the user, the intelligent search recommendation method provided by the application can further prioritize the search results matched with the search term according to the user portrait of the user.

[0087] The user portrait mainly reflects the personal habits, preferences, fields of interest, and related information of other users associated with the user, and comprehensively reflects the characteristics of each user from multiple aspects. The search results can be filtered and sorted according to the characteristics of the user to obtain the final search recommendation results.

[0088] The intelligent retrieval recommendation method provided by the application determines corresponding semantic tags, entities and relationships between entities by analyzing each document in the corpus, constructs a fragmented knowledge network based on a trained transE model, and then determines a matched target document set and a fragmented knowledge subnetwork having a matching relationship with the semantic tags of the target document according to a user search term, and can perform recommendation according to different priorities according to a user portrait. The user obtains the required related documents according to the search term, and simultaneously obtains other fragmented knowledge associated with the semantic tags of the document. In addition, the user portrait can be used to sort the recommendation priorities of all the documents obtained by searching, so that the user is provided with more flexible search results and the search effect is improved more effectively.

[0089] Optionally, the determining of the semantic tags corresponding to each document in the corpus comprises:

[0090] determining all the subject words corresponding to all the documents in the corpus as a standard subject word library;

[0091] determining a first candidate word set corresponding to a target document based on a TF-IDF algorithm and the standard subject word library; the target document is any document in the corpus;

[0092] determining whether a first target word belongs to a candidate semantic tag set corresponding to the target document based on a first judgment condition; the first target word is any word in the first candidate word set;

[0093] The first judgment condition comprises whether the first target word belongs to the standard subject word library and the similarity between the first target word and a target subject word.

[0094] determining the first N1 candidate semantic tags with the highest correlation degrees as the semantic tags corresponding to the target document based on the correlation degrees between any candidate semantic tag in the candidate semantic tag set and the target document.

[0095] Specifically, in the process of determining the semantic tags corresponding to each document in the corpus, all the subject words corresponding to all the documents in the corpus need to be determined in advance. The subject words can be obtained in various ways, such as a Term Frequency-Inverse Document Frequency (TF-IDF) algorithm, a TextRank algorithm, and the distribution of topics and words in the document. The standard subject word library is obtained by determining all the subject words corresponding to all the documents in the corpus through the above methods.

[0096] The automatic standard word indexing of the document is a basic work of the automatic text processing, and the standard theme word library is used to determine the theme word contained in each document, so that the precision of the retrieval and recommendation can be effectively improved. However, the general semantic labeling method has the following problems: the general concept is labeled, the effective labeling cannot be performed according to the characteristics of different field knowledge; the manual and semi-automatic labeling needs manual intervention, and is not suitable for large-scale application; and the accuracy of the automatic semantic labeling still needs to be improved. For different fields, there are some common themes. Therefore, the automatic semantic label labeling method of the document used in the application is based on the standard theme word library, has strong field, and has no strict format requirement for the labeled document and high algorithm performance.

[0097] The theme word table is added to the word library of the word segmenter, each document is segmented, the TF-IDF value of each segmented word is determined according to the TF-IDF algorithm, the first M1 words with the maximum TF-IDF value are selected as the first candidate word set of the document.

[0098] Then, one word in the first candidate word set is selected as a first target word, whether the first target word belongs to the standard theme word library is determined, and the similarity of the first target word and the target theme word is determined, the target theme word is any theme word in the standard theme word library. It is equivalent to determining the same target theme word or similar target theme word as the first target word, and the same or similar target theme words are classified into the semantic label candidate set of the document.

[0099] Further, the association degree of any semantic label in the semantic label candidate set and the document is determined, and the first N1 semantic labels with the highest association degree are selected as the corresponding semantic labels of the target document. Wherein, M1 and N1 are positive integers, and M1≥N1≥1.

[0100] The calculation of the association degree can consider the position and frequency of the candidate semantic label in the corresponding document, the corresponding weight is set based on the position and frequency of the semantic label, and the association degree of the semantic label and the document is obtained.

[0101] Optionally, before determining all theme words corresponding to all documents in the corpus, the method comprises:

[0102] Based on the partition-based latent Dirichlet allocation model, the themes corresponding to all documents in the corpus are extracted, and the distribution of the document and the theme and the distribution of the theme and the word are determined.

[0103] In particular, in order to effectively construct a standard thesaurus, it is necessary to first extract the implied theme from various types of documents. Various types of documents include scientific literature and design documents. A typical theme model is Latent Dirichlet Allocation (LDA), which is a three-layer Bayesian probability generation model that can be used to estimate the properties of multiple observations through unsupervised learning. That is, the theme is discovered in the document set, and then any single document in the set is automatically classified according to the "relevance" of the document to each discovered theme. However, the traditional LDA ignores some semantic features hidden in the semantic structure of the long text document. In addition, the diversified theme distribution in the long text will limit the quality of the theme vector generated by the LDA.

[0104] To solve the above problems, the improved LDA theme modeling method based on partition is adopted, and the LDA based on partition (LDAP) is used for resource theme extraction. The schematic diagram of the LDAP model is shown in Figure 2 As LDA, LDAP is also a probability model describing the process of generating a document set. The difference is that LDAP introduces the concept of semantic topic unit and model at the semantic topic unit level instead of the document level. By dividing the longer text into topic units, the hidden semantic information in the text structure can be fully utilized to achieve better modeling effect. In summary, LDAP not only retains the advantages of the original LDA, but also refines the modeling granularity from the document level to the semantic topic level, and is particularly suitable for theme modeling of medium and long text.

[0105] The LDAP algorithm is used to divide the paragraphs of the document into topic units, and each paragraph of each document in the corpus will form a sub-document, and the sub-document will reconstitute a large corpus. Then, the LDAP is used to model the newly generated sub-document corpus to obtain the sub-document-theme distribution and the theme-word distribution. Finally, in order to obtain the theme distribution corresponding to the original document, the obtained sub-document-theme distribution needs to be combined. The method used here is to assign a weight value to each sub-document, which represents the importance of the sub-document in the original document. Considering the weight value and the theme distribution of the sub-document, the theme distribution of the original document is obtained. The whole process of the LDAP algorithm can be described as follows: (1) divide the document into multiple semantic topic units, i.e. sub-documents; (2) model the sub-document corpus to obtain the theme distribution of the sub-document and the theme-word distribution; (3) weighted summation to obtain the distribution of the original document and the theme.

[0106] Define a transition matrix R from the document theme distribution θ m to the sub-document theme distribution R can be obtained by the weight vector r of the sub-document r=(r1, r2, …, rp ) T The calculation is as follows:

[0107] R = r + (1)

[0108]

[0109] where p i is the p i th sub-document, m is the document; length() is the length of the text, r + is the generalized inverse matrix of matrix r. Here, r refers to the semantic importance of each topic unit (paragraph) in the entire document. Considering the complexity of the calculation, LDAP assumes that the number of valid terms (words or phrases) contained in a paragraph can to some extent reflect the semantic richness of the paragraph.

[0110] Finally, using the LDAP model, the topics corresponding to all documents in the corpus are determined, and the distribution of documents and topics, as well as the distribution of topics and words, can be represented as:

[0111] 1) For each topic k ∈ [1, K] in the corpus D:

[0112] 1.1 Generate the multinomial distribution of words β ~ φ k ;

[0113] 2) For each document d in the corpus D:

[0114] 2.1 Generate the document topic distribution θ m ;

[0115] 2.2 For each sub-document p in the document d:

[0116] 2.2.1 Generate the sub-document topic distribution

[0117] 2.2.2 For each sub-document p:

[0118] According to generate the topic z p,n to which the word n belongs;

[0119] According to φ k generate the word w p,n .

[0120] where D represents the corpus (also represents the total number of documents in the corpus); K is the total number of topics in the corpus; α, β are the Dirichlet prior distribution hyperparameters of the multinomial distribution; φ k represents the multinomial distribution of words belonging to the kth topic; θ m represents the document topic distribution of document d, θp represents the multinomial distribution of topics for sub-document p, where θ m is the weighted sum of θ p ; R is the transition matrix from document topic distribution θ m to sub-document topic distribution θ p ; w p,n is a word n in sub-document p; z p,n is the topic that word n in sub-document p belongs to.

[0121] Each document in the corpus is divided into sub-documents, and a new corpus is formed by sub-documents. The topics are modeled on the new corpus. For any sub-document in the corpus, given parameters α, β, the joint distribution of words z p , w p , θ p , φ in the sub-document is:

[0122]

[0123] The joint distribution of documents z m , w m , θ m , φ in the corpus is:

[0124]

[0125] where r p is the weight of sub-document p in the original document, M is the total number of documents in the corpus; P m represents the total number of sub-documents belonging to the same document m; N p represents the total number of words in sub-document p.

[0126] The main difference between LDAP and LDA is that LDAP uses LDA to process the paragraphs of a document, and then weights each paragraph to obtain the topic distribution of the document.

[0127] The initialization of the parameters of the LDAP model uses the Gibbs sampling method to estimate the model parameters, the process is as follows:

[0128]

[0129]

[0130]

[0131] where V represents the total number of different words in the corpus; represents the number of the t-th word in the document belonging to topic k; β t is the β hyperparameter corresponding to the t-th word; represents the number of the kth word in the sub-document p; a k represents the a hyper-parameter corresponding to the topic k; a k，t represents the topic-word distribution, a p,k represents the sub-document-topic distribution, a m,k represents the document-topic distribution.

[0132] Optionally, the determining of all topic words corresponding to all documents in the corpus comprises:

[0133] determining, in units of target topics, a second candidate word set belonging to the target topic based on the topic and word distribution; the target topic is any one of the topics corresponding to all documents in the corpus;

[0134] determining the first N2 target words with the largest first Chi-square value as the topic words belonging to the target topic, and adding the topic words to the standard topic word library; the first Chi-square value is the Chi-square value of the target word and the target topic; the target word is any word in the second candidate word set.

[0135] Specifically, after determining the topics of all documents included in the corpus, the document and topic distribution, and the topic and word distribution based on the corpus, it is necessary to determine all topic words included in each topic.

[0136] Each topic is represented by a group of documents belonging to the topic, a document set belonging to the same topic, and the topics corresponding to all documents have been obtained through the LDAP model. For a topic and the document set belonging to the topic, it is necessary to calculate a group of topic words representing the topic. Mutual information (Mutual Information) and Chi-square test (Chi-square Feature Selection) are two commonly used feature selection methods. The present application adopts Chi-square test for feature selection, and screens out the first Top N2 words that can best represent the topic as a group of topic words of the topic by calculating the Chi-square value of each word and the topic, wherein N2 is a positive integer greater than or equal to 1. The process of determining the topic words by using the Chi-square value is shown in Figure 3 .

[0137] All topics and word distributions in the corpus D are obtained through the LDAP model. It is assumed that based on the topic and word distribution, the candidate set of topic words of all documents under any topic c is W={w1,w2,…,wn}, i.e. the second candidate word set of the target topic c. n

[0138] Further, the Chi-square value of each candidate word in the second candidate word set and the target topic c is determined, and the specific calculation method comprises:

[0139] ​In the present application, whether a document contains a certain topic word and whether a document belongs to a certain topic are two independent events. Based on this, when topic word screening is performed, define random variables e t and e c : when a document contains topic word t, e t = 1; when a document does not contain topic word t, e t = 0; when a document belongs to topic c, e c = 1. When a document does not belong to topic c, e c = 0.

[0140] A word t in a topic c, calculate the chi-square value of word t and topic c:

[0141]

[0142] Wherein, represents the observed frequency of e t and e c in corpus D; is the corresponding expected frequency, for example, E 11 is the expected frequency of the simultaneous appearance of word t and topic c in a document; an equivalent method of calculating chi-square value X 2 is as follows:

[0143]

[0144] Wherein N 10 represents the number of documents containing topic word t but not in topic c; N 11 represents the number of documents containing topic word t and also in topic c; N 01 represents the number of documents not containing topic word t but in topic c; N 00 represents the number of documents not containing topic word t and not in topic c.

[0145] According to formula 8 or formula 9, determine the chi-square value of each word (i.e. second target word) w i and the target topic c in the second candidate word set W in turn, and sort them in descending order of chi-square value, and finally take the second target words corresponding to the first N2 chi-square values as the topic words of the topic. N2 is a positive integer and is greater than or equal to 1.

[0146] After calculating the topic words, manual intervention can be used to review the topic words, and the topic words that pass the review are included in the standard topic word library. At the same time, the synonyms of each topic word in the standard topic word library can be calculated using the HowNet synonym dictionary to enrich the description of each topic.

[0147] Optionally, the determining whether the first target word belongs to the candidate semantic label set corresponding to the target document based on the first judgment condition comprises:

[0148] determining whether the first target word belongs to the standard subject heading library;

[0149] if yes, determining that the first target word belongs to the candidate semantic label set corresponding to the target document;

[0150] if no, based on the similarity between the first target word and any subject heading in the standard subject heading library, determining to convert the first target word into a subject heading with the highest similarity to the first target word and add the subject heading to the candidate semantic label set corresponding to the target document.

[0151] Specifically, after adding the standard subject heading library W to the word library of the word segmenter, all documents in the corpus are segmented, and the TF-IDF value of each segmented word is calculated by using TF-IDF, taking a single document as a unit (i.e., a target document), and the first M1 words with the largest TF-IDF value are selected as the first candidate word set of the target document. and i≤M1. represents any word in the first candidate word set, i.e., a first target word.

[0152] if if the first target word is added to the candidate semantic label set i.e.

[0153] if if the similarity between the first target word and any subject heading w in the standard subject heading library W is calculated by using the word vector model Word2Vec, the first target word is converted into a subject heading with the highest similarity to the first target word, and the converted subject heading is added to i The conversion method can use Word2Vec. The calculation formula of the similarity is as follows:

[0154]

[0155] If the similarity between the first target word and all subject headings is equal, the first target word is not converted and is directly added to If the similarity between the first target word and several subject headings is equal, the subject headings with the equal similarity to the first target word are added together to

[0156] According to the above steps, the semantic label candidate set of the target document is obtained ​​For Computing the first target word The weight value of the target document. The first N1 target words with the largest weight value are taken as the semantic label of the target document, thereby obtaining Wherein, M1 and N1 are positive integers, and M1≥N1.

[0157] Considering the position and frequency of the semantic label of the target document in the target document, the weight of the word is calculated by using the following formula:

[0158]

[0159] Wherein, w t , w a , w fls , w c Respectively, the semantic label (the first target word) In the title, abstract, beginning and end of the paragraph and other parts of the text; f t , f a , f fls , f c Respectively, the semantic label (the first target word) In the title, abstract, beginning and end of the paragraph and other parts of the text, w len The length of the semantic label (the first target word) ; w tf / idf The TF-IDF value of the semantic label (the first target word) According to the TF-IDF algorithm. The process of determining the semantic label of the document is shown in Figure 4 .

[0160] Optionally, the determining the entity and the inter-entity relationship corresponding to each of the documents comprises:

[0161] Taking the target sentence as a unit, marking each word corresponding to the target sentence as a target sequence; the target sentence is any sentence in the target document; the target document is any document in the corpus;

[0162] Determining the entity and the inter-entity relationship corresponding to the target sequence based on the joint model LSTM-LSTM-Bias;

[0163] The entity at least includes all the topic words corresponding to all the documents in the corpus.

[0164] Specifically, this invention uses a joint model based on sequence labeling, LSTM-LSTM-Bias, which is built upon Long Short Term Memory (LSTM) networks. LSTM is a special type of Recurrent Neural Network (RNN). This joint model designs a special label type, transforming entity recognition and relation classification into a sequence labeling problem. Then, an end-to-end model is proposed to jointly extract entities and relations. This method is an end-to-end neural network model, mainly consisting of three steps: First, the encoding layer uses Bi-LSTM to encode the input sentence; then, the decoding layer uses LSTM-d (an improved version of LSTM) for decoding; finally, the model-labeled entity-relation triples are output.

[0165] (1) Coding layer

[0166] Using sentences as units in the document, represent all the words in a sentence as a sequence S = {s1, ..., s2}. t s t+1 , ..., s n} represents a sequence of words, where s t ∈R d It is the d-dimensional word vector corresponding to the t-th word in the sentence, where n is the length of the given sentence; R represents the real number field, R d This represents the d-dimensional real number field. For example, the coordinates of the two-dimensional vector (2, 3), where 2 and 3 are both real numbers.

[0167] In sequence labeling problems, Bi-LSTM encoding layers have been proven to effectively capture the semantic information of each word. It consists of forward LSTM layers, backward LSTM layers, and connection layers. The word embedding layer converts words with one-hot representations into embedding vectors. Following the word embedding layer are two parallel LSTM layers: a forward LSTM layer and a backward LSTM layer. The LSTM architecture consists of a set of circularly connected subnets called memory blocks. Each time step is an LSTM memory block. The LSTM memory blocks in the Bi-LSTM encoding layer are used to calculate the semantic information based on the previous hidden vector h. t-1 The previous unit vector c t-1 and embedding the current input word w t Calculate the current hidden vector h t .

[0168] For each word s t The forward LSTM layer passes through words s1 to s2. t Contextual information for s t Encode and mark as Similarly, the backward LSTM layer encodes s n according to the context information from s t to s t , denoted as Finally, the and are concatenated to represent the encoding information of the word t, denoted as

[0169] (2) Decoding layer

[0170] The model uses an LSTM structure to generate a label sequence. When detecting the label of the word s t , the input of the decoding layer is: h t obtained from the Bi-LSTM encoding layer, the previously predicted label embedding T t-1 , the previous cell value The hidden vector layer at the previous decoding Finally, the softmax layer calculates the normalized entity label probability according to the label prediction vector T t .

[0171] (3) Bias objective function

[0172] The model training uses the maximum log-likelihood of the data, and the optimization method uses RMSRrop.

[0173] According to the above joint model LSTM-LSTM-Bias, all entities and entity relationships in the corpus are extracted, the entity is mainly represented by a word, and at least includes all the keywords corresponding to all the documents in the corpus. The entity relationship is represented as the semantic connection between any two entities, such as apple is a fruit with enough water, and the entity relationship between “apple” and “fruit” is “belongs to”, and nuclear energy is energy, and the entity relationship between “nuclear energy” and “energy” is “belongs to”. The entity relationship can be one-to-one, one-to-many, or many-to-many.

[0174] Optionally, the training method of the transE model comprises:

[0175] Based on the triples formed by the entities and the entity relationships corresponding to each of the documents in the corpus, determine the positive samples and the non-similarity measure d(S) corresponding to the positive samples;

[0176] Replace the head entity or the tail entity in the positive sample so that it does not satisfy the triple in the positive sample, determine the negative sample and the non-similarity measure d(S') corresponding to the negative sample;

[0177] Determine the boundary function when the positive sample and the negative sample are used as training samples.At the minimum time, the related parameters of the trained transE model are obtained;

[0178] The boundary function is expressed as:

[0179]

[0180] And S' (h，l，t) ={(h', l, t) | h' E} U {(h, l, t') | t' E};

[0181] S={(h, l, t) | h E, t E}

[0182] Wherein, [gamma+d(S)-d(S')] + Indicates that gamma+d(S)-d(S') takes positive value, gamma is a boundary hyperparameter, S represents a positive sample, S' (h，l，t) Indicates a negative sample, (h', l, t) and (h, l, t') both represent a triple belonging to a negative sample, h' and h represent head entities, t and t' represent tail entities, l represents an inter-entity relationship, E represents a set of all entities, and d() represents a dissimilarity measure.

[0183] Specifically, in order to establish a connection between the semantic label corresponding to the document and the theme and the subject word of the document, the application constructs a fragmented knowledge network of entities and inter-entity relationships corresponding to each document in the corpus through the trained transE model.

[0184] All entities corresponding to the documents in the corpus at least include the subject words corresponding to the documents, so that the fragmented knowledge network constructed by the entities and the inter-entity relationships can be understood as a fragmented knowledge network constructed between the subject words of each document, and there is also a corresponding relationship or matching relationship between the subject word and the semantic label. For example, the subject word can be the same as the semantic label, or the relevant description of the subject word includes the semantic label. Then the matching part of the network can be found in the above-mentioned fragmented knowledge network through the semantic label, which includes the entity matched with the semantic label, other entities having an inter-entity relationship with the entity, and a part of the fragmented knowledge network formed by the inter-entity relationship.

[0185] And the training method of the transE model specifically includes:

[0186] 1. Determine the training sample set

[0187] The entity set E and the relation set L extracted from all the documents in the corpus, h E, t E represent the head entity h and the tail entity t in the entity set E respectively; l L represents a relation l in the relation set L; the triple (h, l, t) represents a positive sample in the training sample set, and the triple (h', l, t) or (h, l, t') represents a negative sample, also known as a corrupted triple, in the training sample set. The negative sample is obtained by replacing the head entity or the tail entity in the positive sample with an optional entity in the entity set E, and the replaced triple does not belong to the positive sample. The training sample set includes all the positive samples and the negative samples.

[0188] 2. Determining the boundary function

[0189] The vector embeddings of the entities and the relations between the entities are learned by the transE model, and the embeddings are in R k (k is a model hyperparameter); when the triple (h, l, t) is correct, h + l ≈ t, that is, t should be close to h + l, and h + l should be far away from other t; following the energy-based framework, the energy of the triple is approximately equal to the dissimilarity measure d(h + l, t);

[0190] The boundary function of the transE model is:

[0191]

[0192] and S' = {(h', l, t) | h' E} U {(h, l, t') | t' E}; (h，l，t)

[0193] S = {(h, l, t) | h E, t E}

[0194] where [γ + d(S) - d(S')] + represents that γ + d(S) - d(S') is positive, γ is a boundary hyperparameter, S represents the positive sample, S' (h，l，t) represents the negative sample, (h', l, t) and (h, l, t') both represent a triple belonging to the negative sample, h' and h represent the head entity, t and t' represent the tail entity, l represents the relation between the entities, E represents the set of all entities, and d() represents the dissimilarity measure.

[0195] The transE model is trained by using the above training sample set, and when the corresponding boundary function is minimized, the related model parameters of the transE model are obtained.

[0196] The boundary function is minimized, and the distance between the positive sample and the negative sample is maximized. The ideal situation that the minimization of the boundary function hopes to achieve is that the d(h + l, t) of the positive sample is as small as possible, and the d(h​′ + l, t) or d(h, l, t ′ ) is as large as possible, so that the margin function is minimized and the overall loss tends to 0. The role of the hyperparameter in the margin function is to correct the gap between positive and negative samples. The larger the margin hyperparameter, the larger the gap between the two samples is corrected, and the more stringent the correction of the word vector is. Maximizing the distance between positive and negative samples means that the margin hyperparameter value is large, and minimizing the margin function means that the value of the margin function is minimum and tends to 0.

[0197] The detailed optimization process is as follows: all embeddings of entities and inter-entity relationships are initialized through a random process. In each main iteration of the algorithm, first, the embedding vectors of the entities are normalized. Then a small part of triples is extracted from the training set as a small batch of training triples. For each such triple, a corrupted triple is sampled. Finally, the parameters are updated by a gradient step with a constant learning rate. The algorithm stops based on its performance on the validation set.

[0198] Optionally, before determining the recommendation order of the target document set according to the user portrait of the user, the method comprises:

[0199] Based on the retrieval log of the user, determine the target user's commonly used retrieval words and the corresponding retrieval times, and select the top N3 retrieval words with the most retrieval times as the retrieval history portrait of the target user;

[0200] Based on the basic information of the user, determine the commonly used retrieval words of other users in the same department as the target user and the corresponding retrieval times, and select the top N3 retrieval words with the most retrieval times as the department portrait of the target user;

[0201] Based on the operation log of the user, determine the semantic tags corresponding to all documents operated by the target user, and select the top N3 retrieval words with the highest frequency of occurrence as the document operation portrait of the target user;

[0202] Based on the first topic set, determine the topic words and auxiliary words corresponding to each topic in the first topic set as the research field portrait of the target user in units of a single topic; the first topic set is composed of topics corresponding to all documents operated by the target user;

[0203] Based on the user's preferences or needs, determine the target user's custom portrait;

[0204] Determine the target user's retrieval history portrait, the target user's department portrait, the target user's document operation portrait, the target user's research field portrait, and the target user's custom portrait as the user portrait of the target user;

[0205] where N3 is a positive integer greater than or equal to 1.

[0206] Specifically, the needs of users are biased, but in most cases, users cannot accurately express their needs, which makes it difficult to conduct personalized retrieval and recommendation based on user needs. User profile can formally describe user needs, thereby solving this problem. User profile, i.e. user information tagging, describes the characteristic attributes of users by collecting user attributes, historical behavior, preference features and other dimensional data, and analyzes and mines the potential value information of these features, thereby abstracting the overall information of a user. User profile is the foundation of applying big data, and is also a prerequisite for personalized retrieval and recommendation.

[0207] Before the user's user profile is based on the user, the target documents in the user's retrieval results are recommended and prioritized, the user profile needs to be determined.

[0208] User profile can be analyzed based on many factors: user search history; document content viewed, previewed and downloaded by the user; the number of times the user views, previews and downloads the document; the user's basic information, such as the organization and department to which the user belongs; the field studied by the user, etc. At the same time, the user profile also needs to consider the user-defined profile to achieve personalized feature analysis based on the user. In order to improve the accuracy and standardization of the user profile, the features of the research field profile should come from a standard thesaurus. On the other hand, auxiliary words can be added as a supplement to the research field profile, thereby improving the breadth of the user profile.

[0209] The user profile can be represented as M u = {S u , O u , D u , F u , C u}, where S u represents the search history profile, O u represents the document operation profile, D u represents the department profile, F u represents the research field profile, C u represents the user-defined profile; each sub-profile is a collection of a certain attribute of the user, s i , o i , d i , f i , c i represent the search terms with higher search frequency of the user, the semantic tags corresponding to the documents frequently previewed and downloaded by the user, the user profile words of other users in the department of the user, the subject headings related to the research field and the auxiliary words, and the user-defined profile words, respectively.

[0210] For a user u, the user portrait construction specifically includes:

[0211] (1) According to the user's search log, the user's search words and corresponding search times are counted, and the top N3 search words in search times are selected as the user's search history portrait S u ={s1, s2, …, s N}; Wherein, N3 is a positive integer greater than or equal to 1;

[0212] (2) Through the user's basic information, other users in the user's department are obtained, the search history of other users is counted, and the top N3 search words of other users in the department are selected as the department portrait D u ={d1, d2, …, d N} of the user;

[0213] (3) According to the user's operation log, the document set DS viewed, previewed and downloaded by the user is obtained;

[0214] The semantic tags of the document set DS are counted, and the top N3 semantic tags with the highest frequency are selected as the user's document operation portrait O u ={o1, o2, …, o N};

[0215] (4) According to the topic extraction, the topic set FD of the document set DS is obtained; for each topic F i in the topic set FD, the steps of determining the research field portrait include:

[0216] 4.1 Find the document subset i belonging to the topic F

[0217] 4.2 For each document in the document subset , the theme word is calculated by the theme word, and is added to the theme word set i of the topic F

[0218] 4.3 For each theme word RW i in , add its synonyms in the theme word table to the auxiliary word set SRW i ;

[0219] 4.4 Add the element {RW i , SRW i : {sw1, sw2, …, sw n}} to the topic F i .

[0220] Finally, the user u's research field profile F is determined u , F u is expressed as:

[0221] F u ={…F i : {…{RW i , SRW i : {sw1, sw2, …, sw n}}…}…} (i = 1, 2, …, n) (12)

[0222] wherein F i (i = 1, 2, …, n) is any field topic of interest to the user u, and is a set composed of multiple elements, the ith element of F i corresponds to a topic word set {RW i , SRW i : {sw1, sw2, …, sw n}}, RW i is a topic word used to describe the topic F i , SRW i is a set of auxiliary words of the topic word RW i , and is a synonym of the topic word RW i in the standard topic word library, used to supplement the description of the topic F i .

[0223] (5) According to the user's preferences or needs, add custom profile C u ={c1, c2, …, c N} by yourself.

[0224] Optionally, the determining of the recommended order of each target document in the target document set according to the user profile of the user comprises:

[0225] determining a target semantic label based on the semantic label corresponding to any target document in the target document set;

[0226] determining a matching value of the target semantic label and the user profile of the target user;

[0227] reordering the target document set in descending order of the matching value.

[0228] Specifically, data intelligent push can make users change from retrieving data to automatically receiving the data they need, greatly improving the reuse effect and use efficiency of data. Collaborative filtering recommendation algorithm recommends items through user rating data, but sparse rating data can lead to inaccurate similarity calculation, thereby affecting the recommendation quality. On the other hand, the recommendation method based on collaborative filtering does not consider user background, user preference, behavior purpose and other information, and cannot well meet the personalized needs of different users. Like existing retrieval methods, the existing recommendation method returns coarse-grained document information, which is difficult to accurately locate the knowledge of interest to the user.

[0229] The present application uses a user portrait-based and improved fragmented knowledge network recommendation algorithm. First, a quantifiable user portrait is constructed, and a combination of user portrait use features is used to fuse the improved fragmented knowledge network recommendation algorithm, to obtain a recommendation algorithm combined with the user portrait. Then, aiming at the problem of sparse rating data of the traditional collaborative filtering recommendation algorithm, the user similarity calculation method is improved, and a multi-relation directed graph composed of entities of interest to the user is returned, to obtain an improved fragmented knowledge network recommendation algorithm. Finally, the two algorithms are combined in a weighted hybrid manner to obtain the final user portrait-based and improved fragmented knowledge network hybrid recommendation algorithm.

[0230] The user portrait-based recommendation algorithm fully utilizes the function of the user portrait and provides good results for personalized recommendation for users. When the algorithm recommends to the target user, the specific steps include:

[0231] According to the search term, after the search engine retrieves the target document set matching the search term, each target document corresponds to a first matching value;

[0232] For these target documents, according to the user portrait of the target user, determine the semantic label corresponding to each target document and the second matching value corresponding to the user portrait;

[0233] Based on the first matching value and the second matching value of each target document, and the preset relevant weight, obtain the comprehensive matching value of each target document;

[0234] According to the high and low of the comprehensive matching value, determine the recommendation priority order of the target document set.

[0235] The intelligent retrieval recommendation method provided by the application determines corresponding semantic tags, entities and entity relationships of each document in a corpus, constructs a fragmented knowledge network based on a trained transE model, and then determines a matched target document set and a fragmented knowledge subnetwork having a matching relationship with the semantic tags of the target document according to a user search term, and can perform recommendation according to a user portrait. The user obtains the required related documents according to the search term, and simultaneously obtains other fragmented knowledge associated with the semantic tags of the document. In addition, the user portrait can be used to sort the recommendation priorities of all the documents obtained through retrieval, so that the user is provided with more flexible retrieval results and the retrieval effect is improved more effectively.

[0236] Figure 5 is a structural schematic diagram of the intelligent retrieval recommendation system provided by the embodiment of the application, as shown in the figure, the system comprises: Figure 5

[0237] A determination module 110 is configured to determine the corresponding semantic tags of each document in a corpus, and the corresponding entities and entity relationships of each document.

[0238] A construction module 120 is configured to construct a fragmented knowledge network based on the corresponding entities and entity relationships of each document in the corpus and a trained transE model.

[0239] A retrieval module 130 is configured to determine a retrieval result matched with a search term input by a user based on the search term, wherein the retrieval result comprises a matched target document set and a fragmented knowledge subnetwork matched with a target semantic tag.

[0240] A sorting module 140 is configured to determine a recommendation order of each target document in the target document set according to a user portrait of the user.

[0241] The fragmented knowledge subnetwork is determined based on the matching relationship between the target semantic tag and the entities in the fragmented knowledge network, and comprises a target entity matched with the target semantic tag and a triple composed of other entities having an entity relationship with the target entity. The target semantic tag is a semantic tag corresponding to any target document in the target document set.

[0242] It should be noted that the above system provided by the embodiment of the application can realize all the method steps realized by the method embodiment and achieve the same technical effects, and thus the same parts and beneficial effects of the method embodiment will not be described in detail.

[0243] Optionally, the determination module 110 further comprises:

[0244] ​The subject word sub-module 111 is configured to determine all subject words corresponding to all documents in the corpus as a standard subject word library.

[0245] The semantic label preliminary screening sub-module 112 is configured to determine a first candidate word set corresponding to a target document based on a TF-IDF algorithm and the standard subject word library; the target document is any document in the corpus.

[0246] The first judgment sub-module 113 is configured to determine whether a first target word belongs to a candidate semantic label set corresponding to the target document based on a first judgment condition; the first target word is any word in the first candidate word set; the first judgment condition includes whether the first target word belongs to the standard subject word library and a similarity between the first target word and a target subject word.

[0247] The determined semantic label sub-module 114 is configured to determine a first N1 candidate semantic label set with a highest association degree as semantic labels corresponding to the target document based on an association degree between any candidate semantic label in the candidate semantic label set and the target document.

[0248] Optionally, the system further comprises a subject extraction module 150 configured to extract subjects corresponding to all documents in the corpus based on a partitioned latent Dirichlet allocation model, and determine a distribution of documents and subjects and a distribution of subjects and words.

[0249] Optionally, the determining module 110 further comprises:

[0250] The candidate subject word sub-module 115 is configured to determine a second candidate word set belonging to a target subject based on the distribution of subjects and words in units of target subjects; the target subject is any one of the subjects corresponding to all documents in the corpus.

[0251] The subject word determination module 116 is configured to determine a first N2 target word corresponding to a second candidate word set with a maximum first chi-square value as a subject word belonging to the target subject, and add the subject word to the standard subject word library; the first chi-square value is a chi-square value of the second target word and the target subject; the second target word is any word in the second candidate word set.

[0252] Optionally, the first judgment sub-module 113 is further configured to:

[0253] determine whether the first target word belongs to the standard subject word library;

[0254] if yes, determine that the first target word belongs to the candidate semantic label set corresponding to the target document;

[0255] If not, based on the similarity between the first target word and any of the standard subject words, determine to convert the first target word into the subject word with the highest similarity to the first target word, and add the subject word to the candidate semantic label set corresponding to the target document.

[0256] Optionally, the determining module 110 further comprises:

[0257] A marking module 117 is configured to mark each word corresponding to a target sentence as a target sequence in units of the target sentence; the target sentence is any sentence in a target document; the target document is any document in the corpus;

[0258] A triple determining module 118 is configured to determine entities and inter-entity relationships corresponding to the target sequence based on a joint model LSTM-LSTM-Bias.

[0259] The entities at least include all subject words corresponding to all documents in the corpus.

[0260] Optionally, the system further comprises a user portrait module 160 configured to determine a user portrait of a target user.

[0261] Optionally, the user portrait module 160 further comprises:

[0262] A search history portrait module 161 is configured to determine commonly used search words and corresponding search frequencies of a target user based on a search log of the user, and select the top N3 search words with the highest search frequencies as a search history portrait of the target user.

[0263] A department portrait 162 is configured to determine commonly used search words and corresponding search frequencies of other users belonging to the same department as the target user based on basic information of the user, and select the top N3 search words with the highest search frequencies as a department portrait of the target user.

[0264] A document operation portrait 163 is configured to determine semantic labels corresponding to all documents operated by the target user based on an operation log of the user, and select the top N3 search words with the highest frequencies as a document operation portrait of the target user.

[0265] A research field portrait 164 is configured to determine subject words and auxiliary words corresponding to each subject in a first subject set in units of a single subject as a research field portrait of the target user, wherein the first subject set is composed of subjects corresponding to all documents operated by the target user.

[0266] A custom portrait 165 is configured to determine a custom portrait of the target user based on user preferences or needs.

[0267] determining a search history profile of the target user, a department profile of the target user, a document operation profile of the target user, a research field profile of the target user, and a custom profile of the target user as the user profile of the target user.

[0268] Optionally, the sorting module 140 further comprises:

[0269] a first matching module 141, configured to determine a matching value of the target semantic label and the user profile of the target user; the target semantic label is any semantic label corresponding to any target document in the target document set;

[0270] a comprehensive matching module 142, configured to determine a comprehensive matching value based on the matching value of the target document and the search term, and the matching value of the target semantic label and the user profile of the target user;

[0271] a recommendation module 143, configured to determine a recommendation order of each of the documents in the target document set in descending order of the comprehensive matching value.

[0272] The intelligent search recommendation system provided by the embodiment of the present application is based on Java Web, mainly adopts SpringMVC mode, Spring and Hibernate Java EE technologies to establish SSM architecture, wherein SpringMVC is responsible for implementing distribution of RestFul requests; Spring is responsible for management and research of all business services; Hibernate is a persistence framework, used for database access and operation. The search bottom layer uses ElasticSearch distributed multi-user full-text search engine, and performs ordered organization and efficient search of data based on Lucene inverted index. ElasticSearch is a distributed, high-extensible and high-real-time search and data analysis engine, which has good support for multi-modal data contained in the module. The system uses Hadoop platform, and establishes an index through a MapReduce distributed computing method. Figure 6 Figure 6

[0273] The intelligent search recommendation system provided by the embodiment of the present application is based on Java Web, mainly adopts SpringMVC mode, Spring and Hibernate Java EE technologies to establish SSM architecture, wherein SpringMVC is responsible for implementing distribution of RestFul requests; Spring is responsible for management and research of all business services; Hibernate is a persistence framework, used for database access and operation. The search bottom layer uses ElasticSearch distributed multi-user full-text search engine, and performs ordered organization and efficient search of data based on Lucene inverted index. ElasticSearch is a distributed, high-extensible and high-real-time search and data analysis engine, which has good support for multi-modal data contained in the module. The system uses Hadoop platform, and establishes an index through a MapReduce distributed computing method.

[0274] (1) User inputs a search term

[0275] 1.1 Analyzing and calculating hot search terms. The search terms and corresponding search frequencies are obtained by counting the search histories of all users, and the top N search terms with the most search frequencies are sorted and recommended as search hot words for user inputless search terms.​​

[0276] (2) Prompt related keywords

[0277] Search hotword intelligent prompt. The way of intelligent prompt is used to regulate the user input search condition. Intelligent prompt is based on user search log and keyword table to complete, when the user inputs part of the search word (i.e. the user is still inputting the search word, and has not determined the search operation), such as "smart", first find the first N search words starting with "smart" in the log, and check whether these words belong to the keywords in the keyword table. If yes, display these words to the user for selection; if not, find the keywords starting with "smart" in the keyword table, and display them to the user for selection.

[0278] (3) Keyword query expansion, submit search

[0279] Keyword table-based query expansion. When the user submits a search request, the system uses IK segmenter to segment the search sentence. The correlation between the search segmentation and the keywords is calculated by Jensen-Shannon divergence measurement method, and the keywords with high correlation are taken for query expansion to construct a new search vector.

[0280] (4) Switch index, open corresponding search engine

[0281] Database metadata represents a row of data in a relational database table, which represents the information of each document in the system. The present application creates an index by MapReduce, and stores the document information and the user's search history in the index, i.e. realizes the recommendation of search hotwords corresponding to search words, and the search result obtained after inputting search words to perform search, respectively through different indexes, switch different search engines. After receiving the user's search request, match the index where the document data is located and open the corresponding search engine to perform search.

[0282] (5) Lucene search and sorting

[0283] Through the position of the keywords in the matched document and the correlation with the document, the search result can also be sorted in multiple dimensions by combining the user's user portrait.

[0284] (6) Determine the corresponding fragmented knowledge subnetwork according to the semantic label corresponding to the matched document.

[0285] (7) Intelligent recommendation result display.

[0286] In another aspect, the present application also provides a computer program product comprising a computer program, the computer program being stored in a non-transitory computer-readable storage medium, and the computer program being executable by a processor to cause a computer to perform the method for quantifying rotor response uncertainty based on path-dependent coordinate system according to any of the above embodiments.

[0287] In yet another aspect, the present application also provides a non-transitory computer-readable storage medium having stored thereon a computer program, the computer program being executable by a processor to cause a computer to perform the method for quantifying rotor response uncertainty based on path-dependent coordinate system according to any of the above embodiments.

[0288] The system embodiments described above are merely illustrative, wherein the modules described as separate components can or can not be physically separated, and the components shown as modules can or can not be physical units, i.e., can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purposes of the embodiments according to actual needs, and those of ordinary skill in the art can understand and implement without creative labor.

[0289] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be implemented by means of software plus necessary general hardware platforms, and of course can also be implemented by hardware. Based on such understanding, the above technical solutions, essentially or in terms of contribution to the prior art, can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions to cause a computer device (which can be a personal computer, a server, or a network device, etc.) to perform the methods described in the various embodiments or some parts of the embodiments.

[0290] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the above-mentioned embodiments of the present application have been described in detail, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the above-mentioned embodiments, or make equivalent replacements to some technical features; and these modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An intelligent retrieval and recommendation method, characterized in that, include: Determine the semantic tags corresponding to each document in the corpus, as well as the entities corresponding to each document and the relationships between entities; Based on the entities and relationships between entities corresponding to each document in the corpus, and the trained translation distance model transE, a fragmented knowledge network is constructed. Based on the user-input search terms, determine the search results that match the search terms; the search results include a set of target documents that match the search terms and a fragmented knowledge subnet that matches the target semantic tags; Based on the user profile of the user, determine the recommendation order of each target document in the target document set; The fragmented knowledge subnet is determined based on the matching relationship between the target semantic tag and the entities in the fragmented knowledge network. It includes a subnet consisting of target entities that match the target semantic tag and triples of other entities that have inter-entity relationships with the target entity. The target semantic tag is the semantic tag corresponding to any target document in the target document set.

2. The intelligent retrieval and recommendation method according to claim 1, characterized in that, The determination of semantic tags corresponding to each document in the corpus includes: Determine all the subject terms corresponding to all documents in the corpus and use them as a standard subject terminology library; Based on the TF-IDF algorithm and the standard topic lexicon, a first candidate word set corresponding to the target document is determined; the target document is any document in the corpus. Based on the first judgment condition, it is determined whether the first target word belongs to the candidate semantic tag set corresponding to the target document; the first target word is any word in the first candidate word set. The first judgment condition includes: whether the first target word belongs to the standard topic thesaurus, and the similarity between the first target word and the target topic word; Based on the correlation between any candidate semantic tag in the candidate semantic tag set and the target document, the top N1 candidate semantic tags with the highest correlation are determined as the semantic tags corresponding to the target document; where N1 is a positive integer greater than or equal to 1.

3. The intelligent retrieval and recommendation method according to claim 2, characterized in that, Before determining all the topic terms corresponding to all documents in the corpus, the following steps are included: Based on the partitioned latent Dirichlet allocation model, the topics corresponding to all documents in the corpus are extracted, and the distribution of documents and topics, as well as the distribution of topics and words, are determined.

4. The intelligent retrieval and recommendation method according to claim 2 or 3, characterized in that, The process of determining all topic terms corresponding to all documents in the corpus includes: Based on the distribution of the target topic and words, a second set of candidate words belonging to the target topic is determined; the target topic is any one of the topics corresponding to all documents in the corpus. The top N2 words with the largest first chi-square values ​​are identified as the second target words and added to the standard topic thesaurus. The first chi-square value is the chi-square value of the second target word and the target topic. The second target word is any word in the second candidate word set. N2 is a positive integer greater than or equal to 1.

5. The intelligent retrieval and recommendation method according to claim 2, characterized in that, The step of determining whether the first target word belongs to the candidate semantic tag set corresponding to the target document based on the first judgment condition includes: Determine whether the first target word belongs to the standard thesaurus; If it belongs to the target document, then the first target word is determined to belong to the candidate semantic tag set corresponding to the target document; If it does not belong to the category, then based on the similarity between the first target word and any topic word in the standard topic thesaurus, the first target word is determined to be converted into the topic word with the highest similarity to the first target word, and added to the candidate semantic tag set corresponding to the target document.

6. The intelligent retrieval and recommendation method according to claim 1, characterized in that, Determining the entities corresponding to each document and the relationships between entities includes: Each word corresponding to a target sentence is labeled as a target sequence, with the target sentence being the unit of measurement; the target sentence is any sentence in the target document; the target document is any document in the corpus. Based on the joint model LSTM-LSTM-Bias, the entities and relationships between entities corresponding to the target sequence are determined. The entity includes at least all the topic terms corresponding to all documents in the corpus.

7. The intelligent retrieval and recommendation method according to claim 1, characterized in that, The training method for the transE model includes: Based on the triples formed by the entities and relationships between entities corresponding to each document in the corpus, positive samples and the dissimilarity measure corresponding to the positive samples are determined. ; The head or tail entity in the positive sample is replaced to satisfy a triplet that is not in the positive sample, thus identifying the negative sample and its corresponding dissimilarity metric. ; Using the positive and negative samples as training samples, determine the boundary function. When the value is minimized, the relevant parameters of the trained transE model are obtained; The boundary function Represented as: ； and ; ； in, express The value is positive. These are boundary hyperparameters. Indicates a positive sample. Indicates a negative sample. and All three indicate that they belong to a negative sample triple. and Both represent head entities. and Both represent tail entities. Indicates the relationship between entities. Represents the set of all entities. This represents a measure of dissimilarity.

8. The intelligent retrieval and recommendation method according to claim 1, characterized in that, Before determining the recommendation order of the target document set based on the user's user profile, the process includes: Based on the user's search logs, determine the target user's frequently used search terms and their corresponding search counts, and select the top N3 search terms with the most search counts as the target user's search history profile. Based on the user's basic information, determine the commonly used search terms and corresponding search counts of other users in the same department as the target user, and select the top N3 search terms with the most search counts as the department profile of the target user; Based on the user's operation logs, determine the semantic tags corresponding to all documents operated by the target user, and select the top N3 search terms with the highest frequency to serve as the document operation profile of the target user. Based on the first topic set, the topic words and auxiliary words corresponding to each topic in the first topic set are determined on a single topic basis, which serve as the research field profile of the target user; the first topic set consists of the topics corresponding to all documents operated by the target user; Based on user preferences or needs, a custom profile of the target user is determined; The target user's search history profile, department profile, document operation profile, research field profile, and custom profile are determined as the target user's user profile. Where N3 is a positive integer greater than or equal to 1.

9. The intelligent retrieval and recommendation method according to claim 8, characterized in that, Determining the recommendation order of each target document in the target document set based on the user's user profile includes: Determine the matching value between the target semantic tag and the user profile of the target user; the target semantic tag is any semantic tag corresponding to any target document in the target document set; Based on the matching values ​​between the target document and the search terms, as well as the matching values ​​between the target semantic tags and the user profile of the target user, a comprehensive matching value is determined; The recommendation order of each document in the target document set is determined according to the comprehensive matching value from high to low.

10. A system for intelligent retrieval and recommendation, characterized in that, include: The determination module is used to determine the semantic tags corresponding to each document in the corpus, as well as the entities and relationships between entities corresponding to each document; The construction module is used to construct a fragmented knowledge network based on the entities and relationships between entities corresponding to each document in the corpus and the trained transE model; The retrieval module is used to determine retrieval results that match the user-inputted search terms; the retrieval results include a set of matched target documents and a fragmented knowledge subnet that matches the target semantic tags; The sorting module is used to determine the recommendation order of each target document in the target document set based on the user's user profile. The fragmented knowledge subnet is determined based on the matching relationship between the target semantic tag and the entities in the fragmented knowledge network. It includes a subnet consisting of target entities that match the target semantic tag and triples of other entities that have inter-entity relationships with the target entity. The target semantic tag is the semantic tag corresponding to any target document in the target document set.

Images