Retrieval method and system for film and television extensive search, electronic equipment and storage medium

Abstract

The invention provides a retrieval method and system for film and television extensive search, electronic equipment and a storage medium, and relates to the technical field of natural language processing, and the method comprises the following steps: determining a target field involved by a user intention by using a user interaction text; under the condition that the target field is the movie and television search field, extracting a semantic slot position of the user interaction text; under the condition that the semantic slot position is not empty and the semantic slot position is inconsistent with a retrieval slot position corresponding to a knowledge text of each movie and television resource in the information source movie and television knowledge base, or the semantic slot position is empty, inquiring movie and television resource vectors in a movie and television vector knowledge base by utilizing vector representation of a user interaction text, and carrying out semantic splicing on the movie and television resource vectors to obtain a movie and television resource vector library; and obtaining the film and television resource semantic information, and sending the film and television resource semantic information to the user terminal. According to the method, film and television extensive search can be realized under the condition that the intention of the user is indefinite, the accuracy of semantic information of film and television resources is improved, more accurate target film and television resources can be conveniently recommended to the user, and the user experience is improved.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing technology, and in particular to a retrieval method, system, electronic device and storage medium for film and television general search. Background Art

[0002] With the rapid development of artificial intelligence technology, the application research of embedding film and television recommendations into smart devices has become an important part of current Internet and Internet of Things research. Film and television retrieval in the process of film and television recommendation has also received great attention and extensive research from academia and industry.

[0003] In the prior art, when searching for movies and TV shows, the corresponding fields of the injected movie and TV resources are determined according to the agreed protocol format, and the fields are cleaned and standardized before searching using an inverted index search engine. Common search engine technologies are as follows:

[0004] 1) Inverted index search engines, including Apache Lucene and Solr. Apache Lucene is a high-performance, scalable information retrieval library, and inverted index is its core function. Lucene itself is not a search system, but a toolkit for developers to build search engines. Solr is an open source full-text search platform based on Lucene, which supports complex queries and high scalability and is often used to process searches of large amounts of data.

[0005] 2) Column storage search engines, which can include HBase and Cassandra. HBase is an open source, non-relational, distributed database based on Hadoop and HFile technology, suitable for structured and semi-structured data storage and real-time read / write access. Cassandra is a distributed storage system developed by Facebook, characterized by high availability and scalability, suitable for processing large amounts of data.

[0006] 3) Full-text search engines, which can include Sphinx and Xapian. Sphinx is an independent full-text search engine that supports multiple languages ​​and character sets, is applicable to various database formats, and can efficiently perform full-text searches. Xapian is an open source search engine library that supports efficient text searches and can be embedded in other applications to provide advanced search capabilities.

[0007] 4) Meta search engines, which can include DuckDuckGo and BoardReader. DuckDuckGo is a privacy-preserving search engine that aggregates results from other search engines and emphasizes not tracking user information. BoardReader is a search engine specifically for forums and message boards, helping users find discussion content under specific topics.

[0008] 5) Special and vertical search engines, which can include SlideShare and CC Search. SlideShare is a search engine focused on PPT and document sharing, where users can search and download various professional presentations and documents. CC Search is a search engine that specifically searches for Creative Commons (CC) licensed content, helping users find images, music and other resources that can be legally shared and used.

[0009] Existing search engine technology can only search from fixed dimensions such as film and television names, types, actors, directors, etc. It cannot achieve accurate search when the user's intention is unclear, and thus cannot recommend film and television resources that meet the user's requirements. Summary of the invention

[0010] The present invention provides a retrieval method, system, electronic device and storage medium for film and television extensive search, which are used to solve the defects existing in the related technology.

[0011] The present invention provides a retrieval method for film and television extensive search, comprising:

[0012] Determine user interaction text;

[0013] Based on the user interaction text, determining a target domain involved in the user's intention;

[0014] If the target domain is a film and television search domain, extracting the semantic slots of the user interaction text;

[0015] If the semantic slot is not empty and the semantic slot is inconsistent with the retrieval slot corresponding to the knowledge text of each film and television resource in the source film and television knowledge base, or the semantic slot is empty, then based on the vector representation of the user interaction text, the film and television resource vector is queried in the film and television vector knowledge base, and the film and television resource vector is semantically spliced ​​to obtain film and television resource semantic information, and the film and television resource semantic information is sent to the user terminal; the user terminal is used to search for the target film and television resource based on the film and television resource semantic information;

[0016] Among them, the source film and television knowledge base is a structured film and television knowledge base, and the film and television vector knowledge base is an unstructured film and television knowledge base, and is constructed based on the source film and television knowledge base.

[0017] According to a retrieval method for film and television extensive search provided by the present invention, the film and television vector knowledge base is constructed based on the following steps:

[0018] Determine candidate film and television resources among the film and television resources based on the popularity and score of each film and television resource in the source film and television knowledge base;

[0019] Based on the names of the candidate film and television resources, a full-text search strategy is applied to perform a rough recall in a multi-source knowledge base to obtain a rough recall result;

[0020] Based on the brief introduction corresponding to the coarse recall result, a dynamic programming algorithm is applied to perform an initial fine recall to obtain an initial fine recall result;

[0021] Based on the initial fine recall result, applying the maximum entropy model and / or the MobileNet model to perform deep fine recall to obtain a target recall result;

[0022] Knowledge points are created for the target recall results, and the created knowledge points are vectorized to construct the film and television vector knowledge base.

[0023] According to a retrieval method for film and television extensive search provided by the present invention, the step of creating knowledge points for the target recall results and vectorizing the created knowledge points includes:

[0024] For any film and television resource in the target recall result, connect the knowledge fields of any film and television resource to obtain the initial knowledge points of any film and television resource;

[0025] Slice the initial knowledge points, each of which contains identification information, film and television name, and film and television type of any film and television resource in the source film and television knowledge base;

[0026] Each piece of knowledge point is input into the word vector representation model to obtain the vectorized representation of each piece of knowledge point output by the word vector representation model.

[0027] According to a retrieval method for film and television extensive search provided by the present invention, the film and television resource vector is queried in a film and television vector knowledge base based on the vector representation of the user interaction text, including:

[0028] Based on the vector representation, query the film and television vector knowledge base for a preset number of vector texts that have high similarity with the vector representation;

[0029] The vector representation and the preset number of vector texts are input into a reordering model to obtain the film and television resource vector output by the reordering model.

[0030] According to a retrieval method for film and television extensive search provided by the present invention, the target field involved in the user's intention is determined based on the user interaction text, and then includes:

[0031] If the target domain is the film and television chat domain, searching is performed in the film and television vector knowledge base based on the vector representation to obtain a search result;

[0032] If the search result is empty, inputting the vector representation into a large language model, and interacting with the user based on the large language model;

[0033] If the search result is not empty, the search result and the vector representation are input into the large language model, and based on the large language model, the prompt information is applied to polish the search result, and the polished result is sent to the user terminal; the user terminal is also used to search for the target film and television resources based on the polished result.

[0034] According to a retrieval method for film and television extensive search provided by the present invention, determining the target field involved in the user's intention based on the user interaction text includes:

[0035] Inputting the user interaction text into a neural network model to obtain the target domain output by the neural network model;

[0036] The target field includes one of a film and television search field, a film and television consultant field, and a skill field, and the film and television consultant field includes a film and television chat field.

[0037] According to a retrieval method for film and television extensive search provided by the present invention, the determining of the user interaction text includes:

[0038] Get user interaction voice;

[0039] The user interaction voice is recognized to obtain the user interaction text.

[0040] The present invention also provides a retrieval method for film and television extensive search, comprising:

[0041] Acquire user interaction voice and send the user interaction voice to a cloud server; the cloud server is used to determine the semantic information of film and television resources based on the user interaction voice and the retrieval method of film and television extensive search;

[0042] The film and television resource semantic information is received, and based on the film and television resource semantic information, a target film and television resource is retrieved in a film and television resource library.

[0043] The present invention also provides a retrieval system for film and television pan-search, comprising: an application layer, a cloud semantic analysis layer, and a knowledge base;

[0044] The application layer is used to execute the above-mentioned retrieval method for film and television extensive search with the user terminal as the execution subject;

[0045] The cloud semantic analysis layer is used to execute the above-mentioned retrieval method for film and television pan-search with the cloud server as the execution body;

[0046] The knowledge base includes a source film and television knowledge base and a film and television vector knowledge base.

[0047] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a retrieval method for extensive film and television search as described above is implemented.

[0048] The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the retrieval method for film and television extensive search as described in any one of the above is implemented.

[0049] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements any of the above-mentioned film and television extensive search retrieval methods.

[0050] The retrieval method, system, electronic device and storage medium for film and television pan-search provided by the present invention first determine the user interaction text; then determine the target field involved in the user intention based on the user interaction text; thereafter, when the target field is the film and television search field, extract the semantic slot of the user interaction text; when the semantic slot is not empty and the semantic slot is inconsistent with the retrieval slot corresponding to the knowledge text of each film and television resource in the source film and television knowledge base, or when the semantic slot is empty, use the vector representation of the user interaction text to query the film and television resource vector in the film and television vector knowledge base, and semantically splice the film and television resource vector to obtain the film and television resource semantic information, and send the film and television resource semantic information to the user terminal. The user terminal searches for the target film and television resource through the film and television resource semantic information. This method can realize film and television pan-search when the user intention is unclear, improve the accuracy of the film and television resource semantic information, facilitate the recommendation of more accurate target film and television resources to users, and improve user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the present invention or related technologies, the drawings required for use in the embodiments or related technical descriptions are briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0052] Figure 1 This is one of the flow charts of the retrieval method for film and television extensive search provided by the present invention.

[0053] Figure 2 This is the second flow chart of the retrieval method for film and television extensive search provided by the present invention.

[0054] Figure 3This is the third flow chart of the retrieval method for film and television general search provided by the present invention.

[0055] Figure 4 This is the fourth flow chart of the retrieval method for film and television extensive search provided by the present invention.

[0056] Figure 5 It is a schematic diagram of the process of film and television knowledge matching in the film and television pan-search retrieval method provided by the present invention.

[0057] Figure 6 It is a schematic diagram of the process of querying film and television resource vectors from a film and television vector knowledge base in the film and television pan-search retrieval method provided by the present invention.

[0058] Figure 7 It is a schematic diagram of the rough sorting process in the retrieval method for film and television general search provided by the present invention.

[0059] Figure 8 It is a schematic diagram of the detailed sorting process in the retrieval method for film and television general search provided by the present invention.

[0060] Fig. 9 It is a flow chart of the case where the target field in the retrieval method for film and television general search provided by the present invention is the field of film and television chat.

[0061] Fig.10 This is the second flow chart of the retrieval method for film and television extensive search provided by the present invention.

[0062] Fig.11 It is a schematic diagram of the structure of the retrieval system for film and television extensive search provided by the present invention.

[0063] Fig.12 This is one of the structural schematic diagrams of the retrieval device for film and television extensive search provided by the present invention.

[0064] Fig.13 This is the second structural schematic diagram of the retrieval device for film and television extensive search provided by the present invention.

[0065] Fig.14 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] Figure 1FIG. 1 is a flow chart of a retrieval method for a film and television extensive search provided in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0068] S11, determining user interaction text;

[0069] S12, determining a target domain involved in the user's intention based on the user interaction text;

[0070] S13, if the target domain is a film and television search domain, extracting semantic slots of the user interaction text;

[0071] S14, if the semantic slot is not empty and the semantic slot is inconsistent with the retrieval slot corresponding to the knowledge text of each film and television resource in the source film and television knowledge base, or the semantic slot is empty, then based on the vector representation of the user interaction text, query the film and television resource vector in the film and television vector knowledge base, and perform semantic splicing on the film and television resource vector to obtain film and television resource semantic information, and send the film and television resource semantic information to the user terminal; the user terminal is used to search for the target film and television resource based on the film and television resource semantic information;

[0072] Among them, the source film and television knowledge base is a structured film and television knowledge base, and the film and television vector knowledge base is an unstructured film and television knowledge base, and is constructed based on the source film and television knowledge base.

[0073] Specifically, the retrieval method for film and television pan-search provided in the embodiment of the present invention is executed by a cloud server. The cloud server can call different large language models (LLM) and different pre-trained language representation models (Bidirectional Encoder Representations from Transformers, BERT) and other neural network models. For large language models with different functions and pre-trained language representation models with different functions, the scale of their model parameters is different.

[0074] Here, it can be understood that film and television general search refers to the search for film and television resources when the user's intention is not clear. The dimensions of film and television general search can include the plot, comments, lines, introduction, etc. of film and television resources.

[0075] First, step S11 is executed to determine the user interaction text. The user interaction text can be obtained through user interaction information. The user interaction information can be text input by the user or text obtained by converting the voice input by the user, which is not specifically limited here.

[0076] Then, step S12 is executed to semantically understand the user interaction text and classify the domain to determine the target domain involved in the user's intention. Here, this process can be implemented by a neural network model. By inputting the user interaction text into the neural network model, the target domain can be obtained. The neural network model can be a model such as LLM, BERT, etc.

[0077] The target domain may include one of the following: film and television search domain, film and television consultant domain, and skill domain. The film and television search domain indicates that the user wants to find film and television resources, and the film and television consultant domain indicates that the user wants to learn about the relevant information of film and television resources. The film and television consultant domain can be subdivided into film and television chat domain, similar recommendation domain, other recommendation domain, etc. The skill domain can be subdivided into weather domain, stock domain, map domain, bottom chat domain, schedule domain, live broadcast domain, etc. Figure 2 shown.

[0078] Then, step S13 is executed to extract the user interaction text into a slot when the target domain is the film and television search domain. The semantic slot can be empty, that is, there is a slot extraction result, or it can be not empty, that is, there is no slot extraction result.

[0079] If the semantic slot is not empty, it can include slots such as film type, film name, actor, etc. Film type can include TV series, movies, variety shows, etc.

[0080] like Figure 2 As shown, if the semantic slot is not empty, the semantic slot can be further verified, and after the verification is successful, it can be standardized through dictionary mapping. After that, elasticsearch can be applied to search the source film and television knowledge base.

[0081] Here, the source film and television knowledge base is a structured film and television knowledge base, which stores structured knowledge texts of various film and television resources, and each knowledge text of a film and television resource corresponds to a search slot. Elasticsearch is used to search the source film and television knowledge base, that is, to determine whether the knowledge text of each film and television resource in the source film and television knowledge base has a search slot that is consistent with the semantic slot.

[0082] If there is a retrieval slot that is consistent with the semantic slot, it means that the user's semantic intention is clear, and the knowledge text of the film and television resources corresponding to the retrieval slot and the entity information of the semantic slot can be directly sent to the user terminal as the semantic information of the film and television resources. The user terminal can then use the received knowledge text and entity information to search for the corresponding target film and television resources. Here, the user terminal can be an application installation package (APK) installed in the terminal device. The terminal device can be a smart phone, a set-top box, a TV, etc. The terminal device can be configured with a large screen to support interaction with the user.

[0083] If there is no search slot that is consistent with the semantic slot, it means that the user's semantic intention is unclear, and step S14 is executed, that is, when the semantic slot is not empty and the semantic slot is inconsistent with the search slot corresponding to the knowledge text of each film and television resource in the source film and television knowledge base, or the semantic slot is empty, the vector representation of the user interaction text is used to query the film and television resource vector in the film and television vector knowledge base. The vector representation of the user interaction text can be obtained by vectorizing the user interaction text.

[0084] Among them, the film and television vector knowledge base is an unstructured film and television knowledge base, which contains unstructured film and television resource vectors of various film and television resources. The film and television resource vector of each film and television resource can include the vectorized knowledge points of the film and television resource. The knowledge points can include the plot, comments, lines, introduction, etc. of the film and television resource. The film and television vector knowledge base can be constructed by combining the knowledge texts of each film and television resource in the source film and television knowledge base with the knowledge texts of each film and television resource in the multi-source knowledge base.

[0085] After that, the film and television resource vectors can be semantically spliced ​​to obtain film and television resource semantic information. The film and television resource semantic information is the semantic representation of each knowledge point of the film and television resource vector after splicing, and the film and television resource semantic information can be sent to the user terminal. The user terminal can use the film and television resource semantic information and apply a search engine to search in the film and television resource library to obtain the target film and television resource corresponding to the film and television resource semantic information. Among them, the search engine can be built in the user terminal, and the film and television resource library contains various film and television resources that can be provided to the user terminal.

[0086] like Figure 3 As shown, the user can control the user terminal through the remote control and interact with the user terminal. After receiving the user interaction information, the user terminal can send the interaction information to the cloud server, determine the user interaction text through the cloud server, and perform semantic understanding of the user interaction text, and perform field classification to determine the target field involved in the user's intention, and when the target field is the film and television search field, query and determine the film and television resource semantic information from the source film and television knowledge base or from the film and television vector knowledge base.

[0087] The semantic information of film and television resources is returned to the user terminal, and the user terminal searches for target film and television resources in the film and television resource library by calling a search engine and recommends them to the user.

[0088] When the target domain is the film and television search domain, such as Figure 4As shown, the semantic slot of the user interaction text is extracted. If the semantic slot is not empty, the semantic slot can be first entity-checked. After the verification is successful, the semantic slot is searched in the source film and television knowledge base to determine whether the semantic slot is consistent with the retrieval slot corresponding to the knowledge text of each film and television resource in the source film and television knowledge base. If they are consistent, the semantic information of the film and television resource is obtained by semantically splicing the retrieval slot. If the entity verification fails or the semantic slot is inconsistent with the retrieval slot, the film and television resource vector is queried in the film and television vector knowledge base based on the vector representation of the user interaction text, and the film and television resource vector is semantically spliced ​​to obtain the semantic information of the film and television resource.

[0089] If the semantic slot is empty, the film and television resource vector is queried in the film and television vector knowledge base based on the vector representation of the user interaction text, and the film and television resource vector is semantically spliced ​​to obtain the film and television resource semantic information.

[0090] The retrieval method for film and television general search provided in the embodiment of the present invention first determines the user interaction text; then, based on the user interaction text, determines the target field involved in the user intention; thereafter, when the target field is the film and television search field, extracts the semantic slot of the user interaction text; when the semantic slot is not empty and the semantic slot is inconsistent with the retrieval slot corresponding to the knowledge text of each film and television resource in the source film and television knowledge base, or when the semantic slot is empty, uses the vector representation of the user interaction text to query the film and television resource vector in the film and television vector knowledge base, and semantically splices the film and television resource vector to obtain the film and television resource semantic information, and sends the film and television resource semantic information to the user terminal. The user terminal searches for the target film and television resource through the film and television resource semantic information. This method can realize film and television general search when the user intention is unclear, improve the accuracy of the film and television resource semantic information, facilitate the recommendation of more accurate target film and television resources to users, and improve user experience.

[0091] Based on the above embodiment, after obtaining the semantic information of film and television resources, the cloud server can also input the semantic information of film and television resources into a large language model, polish the semantic information of film and television resources through the large language model, and send the polishing result to the user terminal, so that the user terminal can search for the target film and television resources in the film and television resource library through the polishing result, thereby improving the accuracy of the target film and television resources.

[0092] Based on the above embodiment, the step of determining the user interaction text includes:

[0093] Get user interaction voice;

[0094] The user interaction voice is recognized to obtain the user interaction text.

[0095] Specifically, the user can input user interaction voice through the voice remote control. The user terminal receives the user interaction voice and can send the user interaction voice to the cloud server. The cloud server obtains the user interaction text by recognizing the user interaction voice.

[0096] In the embodiment of the present invention, the convenience of user operation can be improved through voice interaction between the user and the user terminal.

[0097] Based on the above embodiment, the film and television vector knowledge base is constructed based on the following steps:

[0098] Determine candidate film and television resources among the film and television resources based on the popularity and score of each film and television resource in the source film and television knowledge base;

[0099] Based on the names of the candidate film and television resources, a full-text search strategy is applied to perform a rough recall in a multi-source knowledge base to obtain a rough recall result;

[0100] Based on the brief introduction corresponding to the coarse recall result, a dynamic programming algorithm is applied to perform an initial fine recall to obtain an initial fine recall result;

[0101] Based on the initial fine recall result, applying the maximum entropy model and / or the MobileNet model to perform deep fine recall to obtain a target recall result;

[0102] Knowledge points are created for the target recall results, and the created knowledge points are vectorized to construct the film and television vector knowledge base.

[0103] Specifically, when constructing a film and television vector knowledge base, film and television knowledge screening can be performed first, that is, the popularity and score of each film and television resource in the source film and television knowledge base are used to determine the candidate film and television resources among the film and television resources. For example, in the source film and television knowledge base, film and television resources with a popularity greater than a first threshold and a score greater than a second threshold can be selected as candidate film and television resources. Among them, the first threshold and the second threshold can be set as needed, and are not specifically limited here. For example, both the first threshold and the second threshold can be set to 80. It should be noted that the name of the candidate film and television resource is in Chinese and there is a Chinese introduction in the source film and television knowledge base.

[0104] After that, the candidate film and television resources can be used to construct a film and television resource entity library. The film and television resource entity library may include entities such as the names of the candidate film and television resources. In addition, the knowledge fields required for each candidate film and television resource to construct a film and television vector knowledge base can be screened out, such as the name, type, identification information in the source film and television knowledge base, lines, comments, tags, etc. The tags may include martial arts, comedy, ethics, police and gangsters, etc.

[0105] After that, film and television knowledge matching is performed, that is, Figure 5 As shown in Figure 5 , first, using the name of the alternative video resource, a rough recall is performed in the multi-source knowledge base by applying the full-text retrieval strategy, and a rough recall result can be obtained. Here, the full-text retrieval strategy adopted can be the retrieval strategy of the search server (Elasticsearch, ES) based on Lucene. ES can implement the rough recall in the dual-recall strategy, simply referred to as rough recall. ES provides a full-text search engine with distributed multi-user capabilities based on the RESTful web interface. ES is developed in the Java language and released as open source under the Apache license terms. It is a popular enterprise-level search engine.

[0106] The data in the video search field has certain characteristics. The name of the video resource can roughly determine whether it is the same movie. However, just because the names are the same does not mean it is the same video resource, and different names do not necessarily mean the two are different video resources. Therefore, only video resources with similar names can be roughly recalled by name. To increase the possibility of recall, the aliases of the video resources are added for joint recall.

[0107] Here, introducing the ES rough recall can improve the calculation speed. In the implementation scheme of the ES rough recall, the tokenizer used for query is the standard tokenizer that tokenizes each character, and the similarity algorithm is the Term Frequency-Inverse Document Frequency (TF-IDF) algorithm.

[0108] The TF-IDF algorithm formula is as follows:

[0109] score = tf * idf * fieldNorm;

[0110] tf = sqrt(termFreq);

[0111] idf = loge((docCount + 1) / (docFreq + 1)) + 1;

[0112] fieldNorm = 1 / sqrt(length);

[0113] Among them, termFreq is the frequency of occurrence of the name or alias of the alternative video resource in a certain field of the knowledge text in the multi-source knowledge base. For example, "金" appears once in "金刚川", and the value of termFreq is 1. docCount is the total number of knowledge texts in the multi-source knowledge base, docFreq is the number of knowledge texts hit, and length is the tokenized length of the knowledge text. The tokenized length of one Chinese character is 1.

[0114] Compared with the BM25 algorithm, the TF-IDF algorithm has better performance in terms of recall quantity after comparative experiments. The number of rough recalls is currently set to 30, which means that the rough recall results of each candidate film and television resource can contain 30 pieces of knowledge text of film and television resources. The number of rough recalls is closely related to the quality of fusion. The lower the repetition, the theoretically lower the number of rough recalls can be.

[0115] After that, the dynamic programming algorithm is applied to the introduction corresponding to the rough recall result to perform initial fine recall and obtain the initial fine recall result. Here, the initial fine recall is achieved by calculating the introduction of the film and television resources and the alternative film and television resources in the rough recall result. Introduction calculation belongs to the initial fine recall in the double recall strategy. In the field of film and television search, if the introductions of two film and television resources are the same, the two film and television resources are considered to be the same. The film and television resources recalled by introduction calculation are accurate recall.

[0116] The scores of the maximum common sequence of the knowledge text of the film and television resources in the rough recall results and the knowledge text of the candidate film and television resources are used to determine whether the brief descriptions of the film and television resources in the rough recall results are consistent with those of the candidate film and television resources. The maximum common sequence can be solved using a dynamic programming algorithm.

[0117] Afterwards, the initial fine recall result can be used to apply the maximum entropy (ME) model and / or the MobileNet model to perform deep fine recall to obtain a target recall result. The target recall result can be the knowledge text of a film and television resource with the highest similarity to each candidate film and television resource in the initial fine recall result.

[0118] The maximum entropy model is widely used in the field of natural language processing (NLP). The maximum entropy principle is a general principle of statistical learning. Applying it to classification results in the maximum entropy model. Assume that the classification model is a conditional probability distribution P(Y|X), where X represents the input and Y represents the output. The classification model represents that for a given input X, the output Y is given with the conditional probability P(Y|X). Given a training data set T, the training goal is to select the best classification model using the maximum entropy principle. The classification model finally selected is the maximum entropy model.

[0119] According to the maximum entropy principle, it is necessary to ensure that the model satisfies all known constraints. Several features are extracted from the training data T, and then the expectation of these features on the empirical distribution on T is required to be equal to their mathematical expectation on p(x,y) in the model. In this way, one feature corresponds to one constraint. Core fields such as director, actor, film name, region, and introduction are selected as the features of the maximum entropy model for model training.

[0120] The MobileNet model belongs to the field of computer vision (CV). It is mainly used here to extract poster features of film and television resources, use poster features to calculate similarity, and improve the accuracy of film and television resource recall.

[0121] The ME model and MobileNet model can be used separately or in series. The ME model is used for judgment first. The film and television resources whose output of the ME model in the coarse recall result is lower than the specified threshold are used as the input of the MobileNet model, and the poster features are used for secondary confirmation.

[0122] After that, the sensitive word service can be called to verify the knowledge text of each film and television resource in the target recall result through the sensitive word service, temporarily store the risky knowledge text, and discard or persist it through the management system.

[0123] Finally, knowledge points are created for the target recall results, and the created knowledge points are vectorized to build a film and television vector knowledge base.

[0124] In an embodiment of the present invention, by making full use of the knowledge texts of various film and television resources in a multi-source knowledge base and relying on machine learning models in the NLP and / or CV fields, the recall rate of the knowledge texts of various film and television resources in the multi-source knowledge base can be improved, thereby improving the accuracy of the constructed film and television vector knowledge base.

[0125] Based on the above embodiment, the step of creating knowledge points for the target recall result and vectorizing the created knowledge points includes:

[0126] For any film and television resource in the target recall result, connect the knowledge fields of any film and television resource to obtain the initial knowledge points of any film and television resource;

[0127] Slice the initial knowledge points, each of which contains identification information, film name, film type, and film label of any film resource in the source film knowledge base;

[0128] Each piece of knowledge point is input into the word vector representation model to obtain the vectorized representation of each piece of knowledge point output by the word vector representation model.

[0129] Specifically, when constructing knowledge points, for any film and television resource in the target recall result, its knowledge field can be sequentially integrated and fragmented, and each piece of knowledge point can be vectorized to finally obtain a vectorized representation of each piece of knowledge point. The vectorized representation of all knowledge points of each film and television resource can construct a film and television vector knowledge base.

[0130] Among them, knowledge point integration is to unify structured data and unstructured data, filter out the knowledge fields of each film and television resource, and connect different knowledge fields with specific separators (such as "|") to convert structured data into plain text format.

[0131] Knowledge point segmentation is due to the fact that knowledge questions and answers have a limit on the text length. The excess part will be directly discarded. Therefore, the original knowledge points are cut according to a fixed length, mainly for longer knowledge fields (such as film and television introductions). The knowledge points after cutting retain the identification information, film and television type, film and television name, and film and television labels as fixed text, so that each piece of knowledge point can be located at the original film and television resources.

[0132] Knowledge point vectorization is to vectorize the segmented knowledge points to obtain the final vectorized representation. Here, vectorization can be implemented through the BERT model. The principle is as follows:

[0133] Vectorization is the process of gradually changing from the previous word embedding to the contextual word embedding. The BERT model enhances generalization ability by capturing the semantic relationship between characters, words and sentences, and perceives multi-granular semantic relationships. The BERT model implements the contextual word vector representation by encoding (encoder) through the Transformer encoder. The BERT model can use information from both the front and back directions at the same time. The Transformer encoder is the main part of the BERT model, through which encoding is performed. The specific steps of the BERT model are as follows:

[0134] Step 1: Segment the text sequence of each knowledge point into sentences and input them.

[0135] Step 2: Convert the input text sequence into word vectors, text vectors, and position vectors.

[0136] Step 3: Use the Transformer encoder to enhance the semantics of its word vector.

[0137] Step 4: Add the semantically enhanced word vector, text vector and position vector to obtain the context word vector sequence CE = (ce1, ce2, ..., ce i ,…,ce n ).

[0138] Thereafter, the vectorized representation of the obtained knowledge points may be stored in a film and television vector knowledge base, which may be Milvus, FAISS, etc., and Milvus with better scalability may be preferred here.

[0139] On the basis of the above embodiment, the step of searching for a film and television resource vector in a film and television vector knowledge base based on the vector representation of the user interaction text includes:

[0140] Based on the vector representation, query the film and television vector knowledge base for a preset number of vector texts that have high similarity with the vector representation;

[0141] The vector representation and the preset number of vector texts are input into a reordering model to obtain the film and television resource vector output by the reordering model.

[0142] Specifically, when querying the film and television resource vector from the film and television vector knowledge base using the vector representation of the user interaction text, the following can be done: Figure 6 As shown, the vector representation of the user interaction text is used to perform rough sorting and fine sorting in sequence. The rough sorting is to query the film and television vector knowledge base for a preset number of vector texts with high similarity to the vector representation of the user interaction text, and the fine sorting is to perform secondary sorting on the preset number of vector texts obtained by the rough sorting through the re-ranking model. Among them, the preset number of vector texts with high similarity can be the top N vector texts with high similarity, and the similarity here can be cosine similarity.

[0143] like Figure 7 As shown, the vector representation of the user interaction text and the film and television resource vectors in the film and television vector knowledge base can be passed through the BERT model and the pooling layer to obtain the corresponding vector features u and v, and then the top N vector texts can be found by calculating the cosine similarity between the vector features u and v.

[0144] like Figure 8 As shown, the vector representation of the user interaction text and a preset number of vector texts are input into the reordering model to obtain the film and television resource vector output by the reordering model. The reordering model may be a BERT model, which can achieve reordering by calculating the similarity between the vector representation of the user interaction text and a preset number of vector texts, so that the obtained film and television resource vector has a higher correlation with the user interaction text.

[0145] In the embodiment of the present invention, a rough sorting followed by a fine sorting is used, which can improve the search speed while ensuring the search effect. If all fine sorting algorithms are used, it will be very time-consuming and will not achieve the desired efficiency.

[0146] On the basis of the above embodiment, the step of determining the target domain involved in the user intention based on the user interaction text may further include:

[0147] If the target domain is the film and television chat domain, searching is performed in the film and television vector knowledge base based on the vector representation to obtain a search result;

[0148] If the search result is empty, inputting the vector representation into a large language model, and interacting with the user based on the large language model;

[0149] If the search result is not empty, the search result and the vector representation are input into the large language model, and based on the large language model, the prompt information is applied to polish the search result, and the polished result is sent to the user terminal; the user terminal is also used to search for the target film and television resources based on the polished result.

[0150] Specifically, after determining the target domain involved in the user's intention, if the target domain is the film and television chat domain, Fig. 9 As shown, the vector representation of the user interaction text can be used to search in the film and television vector knowledge base to obtain the search results.

[0151] If the search result is empty, the vector representation is input into the LLM, and the LLM is used to interact with the user to generate a reply message, and the reply message is sent to the user terminal. After receiving the reply message, the user terminal can use the reply message to search for the target film and television resources that meet the requirements in the film and television resource library.

[0152] If the search result is not empty, the search result and the vector representation of the user interaction text are input into the LLM, and the LLM is used to apply prompt information (prompt) to polish the search result, and the polished result is sent to the user terminal. Among them, the prompt information can be set in advance as needed to indicate the requirements for polishing. After receiving the polishing result, the user terminal can use the polishing result to search for target film and television resources that meet the requirements in the film and television resource library.

[0153] like Fig.10 As shown, based on the above embodiment, the embodiment of the present invention further provides a retrieval method for film and television extensive search, including:

[0154] S21, obtaining user interaction voice, and sending the user interaction voice to a cloud server; the cloud server is used to determine the semantic information of film and television resources based on the user interaction voice by applying the retrieval method of film and television extensive search provided in the above embodiments;

[0155] S22, receiving the film and television resource semantic information, and based on the film and television resource semantic information, searching for target film and television resources in a film and television resource library.

[0156] Specifically, the retrieval method for film and television pan-search provided in the embodiment of the present invention is executed by a user terminal. The user terminal can be an application installation package (APK) installed in a terminal device. The terminal device can be a smart phone, a set-top box, a television, etc. The terminal device can be configured with a large screen to support interaction with the user. The user terminal can have a built-in search engine.

[0157] First, step S21 is performed, where the user terminal obtains the user interaction voice through the terminal device and sends the user interaction voice to the cloud server. The cloud server receives the user interaction voice and applies the film and television extensive search retrieval method provided in the above embodiments to determine the film and television resource semantic information.

[0158] Table 1 Search fields

[0159]

[0160] Then, step S22 is executed, the user terminal receives the semantic information of the film and television resources through the network, and calls the search engine to retrieve the target film and television resources in the film and television resource library using the semantic information of the film and television resources. Here, the protocol adaptation can be performed first through the adapter component, and then the engine component is used for retrieval. When the engine component is used for retrieval, the fields used are shown in Table 1 above.

[0161] In the embodiment of the present invention, by determining the film and television resource semantic information used to search for the target film and television resource through the cloud server, the hardware cost of the film and television resource search can be greatly reduced.

[0162] like Fig.11 As shown, based on the above embodiment, the embodiment of the present invention further provides a retrieval system for film and television extensive search, including: an application layer, a cloud semantic analysis layer and a knowledge base;

[0163] The application layer is used to execute the retrieval method for film and television extensive search provided in the above-mentioned embodiments with the user terminal as the execution subject.

[0164] The cloud semantic parsing layer is used to execute the retrieval method for film and television general search provided in each embodiment with the cloud server as the execution body, including recognizing the user interaction voice received by the user terminal to obtain the user interaction text, semantically understanding the user interaction text, and performing field classification to obtain the target field. The target field can be one of the film and television search field, the film and television consultant field, and the skill field. When the target field is the film and television search field or the film and television chat field in the film and television consultant field, the film and television resource semantic information can be queried from the knowledge base and polished by a large language model, and returned to the user terminal, so that the user terminal uses the search engine to search for the target film and television resources in the film and television resource library.

[0165] like Fig.12 As shown, based on the above embodiment, an embodiment of the present invention provides a retrieval device for film and television extensive search, including:

[0166] A text determination module 121, used to determine a user interaction text;

[0167] A domain classification module 122, configured to determine a target domain involved in the user's intention based on the user interaction text;

[0168] A slot extraction module 123, configured to extract the semantic slots of the user interaction text if the target domain is a film and television search domain;

[0169] The film and television resource semantic information determination module 124 is used to query the film and television resource vector in the film and television vector knowledge base based on the vector representation of the user interaction text if the semantic slot is not empty and the semantic slot is inconsistent with the retrieval slot corresponding to the knowledge text of each film and television resource in the source film and television knowledge base, or the semantic slot is empty, and semantically splice the film and television resource vector to obtain the film and television resource semantic information, and send the film and television resource semantic information to the user terminal; the user terminal is used to search for the target film and television resource based on the film and television resource semantic information;

[0170] Among them, the source film and television knowledge base is a structured film and television knowledge base, and the film and television vector knowledge base is an unstructured film and television knowledge base, and is constructed based on the source film and television knowledge base.

[0171] Based on the above embodiment, a film and television vector knowledge base construction module is further included, which is used to:

[0172] Determine candidate film and television resources among the film and television resources based on the popularity and score of each film and television resource in the source film and television knowledge base;

[0173] Based on the names of the candidate film and television resources, a full-text search strategy is applied to perform a rough recall in a multi-source knowledge base to obtain a rough recall result;

[0174] Based on the brief introduction corresponding to the coarse recall result, a dynamic programming algorithm is applied to perform an initial fine recall to obtain an initial fine recall result;

[0175] Based on the initial fine recall result, applying the maximum entropy model and / or the MobileNet model to perform deep fine recall to obtain a target recall result;

[0176] Knowledge points are created for the target recall results, and the created knowledge points are vectorized to construct the film and television vector knowledge base.

[0177] Based on the above embodiment, the film and television vector knowledge base construction module is also used for:

[0178] For any film and television resource in the target recall result, connect the knowledge fields of any film and television resource to obtain the initial knowledge points of any film and television resource;

[0179] Slice the initial knowledge points, each of which contains identification information, film and television name, and film and television type of any film and television resource in the source film and television knowledge base;

[0180] Each piece of knowledge point is input into the word vector representation model to obtain the vectorized representation of each piece of knowledge point output by the word vector representation model.

[0181] Based on the above embodiment, the film and television resource semantic information determination module is further used to:

[0182] Based on the vector representation, query the film and television vector knowledge base for a preset number of vector texts that have high similarity with the vector representation;

[0183] The vector representation and the preset number of vector texts are input into a reordering model to obtain the film and television resource vector output by the reordering model.

[0184] Based on the above embodiment, the film and television resource semantic information determination module is further used to:

[0185] If the target domain is the film and television chat domain, searching is performed in the film and television vector knowledge base based on the vector representation to obtain a search result;

[0186] If the search result is empty, inputting the vector representation into a large language model, and interacting with the user based on the large language model;

[0187] If the search result is not empty, the search result and the vector representation are input into the large language model, and based on the large language model, the prompt information is applied to polish the search result, and the polished result is sent to the user terminal; the user terminal is also used to search for the target film and television resources based on the polished result.

[0188] Based on the above embodiment, the field classification module is specifically used for:

[0189] Inputting the user interaction text into a neural network model to obtain the target domain output by the neural network model;

[0190] The target field includes one of a film and television search field, a film and television consultant field, and a skill field, and the film and television consultant field includes a film and television chat field.

[0191] Based on the above embodiment, the text determination module is specifically used for:

[0192] Get user interaction voice;

[0193] The user interaction voice is recognized to obtain the user interaction text.

[0194] Specifically, the functions of each module in the retrieval of the film and television search provided in the embodiment of the present invention correspond one-to-one to the operation procedures of each step in the above-mentioned method class embodiments with the cloud server as the execution body, and the effects achieved are also consistent. Please refer to the above-mentioned embodiments for details, and no further description will be given in the embodiments of the present invention.

[0195] like Fig.13 As shown, based on the above embodiment, an embodiment of the present invention provides a retrieval device for film and television extensive search, including:

[0196] The voice sending module 131 is used to obtain the user interaction voice and send the user interaction voice to the cloud server; the cloud server is used to determine the semantic information of the film and television resources by applying the retrieval method of the film and television extensive search provided in each embodiment with the user terminal as the execution subject based on the user interaction voice;

[0197] The resource search module 132 is used to receive the film and television resource semantic information and search for target film and television resources in the film and television resource library based on the film and television resource semantic information.

[0198] Fig.14 An example of a physical structure diagram of an electronic device is shown in FIG. Fig.14 As shown, the electronic device may include: a processor 1410, a communication interface 1420, a memory 1430 and a communication bus 1440, wherein the processor 1410, the communication interface 1420 and the memory 1430 communicate with each other via the communication bus 1440. The processor 1410 may call the logic instructions in the memory 1430 to execute the retrieval method for film and television extensive search provided in the above embodiments.

[0199] In addition, the logic instructions in the above-mentioned memory 1430 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the relevant technology or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0200] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the retrieval method for film and television extensive search provided in the above-mentioned embodiments.

[0201] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the retrieval method for film and television extensive search provided in the above-mentioned embodiments.

[0202] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0203] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, by hardware. Based on this understanding, the above technical solution can be essentially or in other words, the part that contributes to the relevant technology can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiment.

[0204] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A retrieval method for film and television extensive search, characterized in that: include: Determine user interaction text; Based on the user interaction text, determining a target domain involved in the user's intention; If the target domain is a film and television search domain, extracting the semantic slots of the user interaction text; If the semantic slot is not empty and the semantic slot is inconsistent with the retrieval slot corresponding to the knowledge text of each film and television resource in the source film and television knowledge base, or the semantic slot is empty, then based on the vector representation of the user interaction text, the film and television resource vector is queried in the film and television vector knowledge base, and the film and television resource vector is semantically spliced ​​to obtain film and television resource semantic information, and the film and television resource semantic information is sent to the user terminal; the user terminal is used to search for the target film and television resource based on the film and television resource semantic information; Among them, the source film and television knowledge base is a structured film and television knowledge base, and the film and television vector knowledge base is an unstructured film and television knowledge base, and is constructed based on the source film and television knowledge base.

2. The retrieval method for film and television extensive search according to claim 1, characterized in that: The video vector knowledge base is constructed based on the following steps: Determine candidate film and television resources among the film and television resources based on the popularity and score of each film and television resource in the source film and television knowledge base; Based on the names of the candidate film and television resources, a full-text search strategy is applied to perform a rough recall in a multi-source knowledge base to obtain a rough recall result; Based on the brief introduction corresponding to the coarse recall result, a dynamic programming algorithm is applied to perform an initial fine recall to obtain an initial fine recall result; Based on the initial fine recall result, applying the maximum entropy model and / or the MobileNet model to perform deep fine recall to obtain a target recall result; Knowledge points are created for the target recall results, and the created knowledge points are vectorized to construct the film and television vector knowledge base.

3. The retrieval method for film and television extensive search according to claim 2 is characterized in that: The step of creating knowledge points for the target recall result and vectorizing the created knowledge points includes: For any film and television resource in the target recall result, connect the knowledge fields of any film and television resource to obtain the initial knowledge points of any film and television resource; Slice the initial knowledge points, each of which contains identification information, film and television name, and film and television type of any film and television resource in the source film and television knowledge base; Each piece of knowledge point is input into the word vector representation model to obtain the vectorized representation of each piece of knowledge point output by the word vector representation model.

4. The retrieval method for film and television extensive search according to claim 1, characterized in that: The step of searching a film and television resource vector in a film and television vector knowledge base based on the vector representation of the user interaction text includes: Based on the vector representation, query the film and television vector knowledge base for a preset number of vector texts that have high similarity with the vector representation; The vector representation and the preset number of vector texts are input into a reordering model to obtain the film and television resource vector output by the reordering model.

5. The retrieval method for film and television extensive search according to claim 1, characterized in that: The step of determining the target domain involved in the user intention based on the user interaction text may then include: If the target domain is the film and television chat domain, searching is performed in the film and television vector knowledge base based on the vector representation to obtain a search result; If the search result is empty, inputting the vector representation into a large language model, and interacting with the user based on the large language model; If the search result is not empty, the search result and the vector representation are input into the large language model, and based on the large language model, the prompt information is applied to polish the search result, and the polished result is sent to the user terminal; the user terminal is also used to search for the target film and television resources based on the polished result.

6. The retrieval method for film and television extensive search according to any one of claims 1 to 5, characterized in that: The determining, based on the user interaction text, a target domain involved in the user intention includes: Inputting the user interaction text into a neural network model to obtain the target domain output by the neural network model; The target field includes one of a film and television search field, a film and television consultant field, and a skill field, and the film and television consultant field includes a film and television chat field.

7. The retrieval method for film and television extensive search according to any one of claims 1 to 5, characterized in that: The determining of the user interaction text includes: Get user interaction voice; The user interaction voice is recognized to obtain the user interaction text.

8. A retrieval method for film and television general search, characterized in that: include: Acquire user interaction voice and send the user interaction voice to the cloud server; The cloud server is used to determine the semantic information of film and television resources based on the user interactive voice by applying the retrieval method of film and television extensive search according to any one of claims 1 to 7; The film and television resource semantic information is received, and based on the film and television resource semantic information, a target film and television resource is retrieved in a film and television resource library.

9. A retrieval system for film and television general search, characterized in that: include: Application layer, cloud semantic parsing layer, and knowledge base; The application layer is used to execute the retrieval method for film and television extensive search as claimed in claim 8; The cloud semantic parsing layer is used to execute the retrieval method for film and television extensive search as described in any one of claims 1 to 7; The knowledge base includes a source film and television knowledge base and a film and television vector knowledge base.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, it implements the retrieval method for film and television extensive search as described in any one of claims 1-8.

11. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the retrieval method for film and television extensive search as described in any one of claims 1 to 8 is implemented.

Images