CN114911944A - A method for building knowledge graph of multi-source media resources based on TextRank algorithm combined with convolutional neural network model - Google Patents

Abstract

The invention relates to a method for constructing a multi-source media asset knowledge graph based on a TextRank algorithm and a convolutional neural network model. The method mainly comprises the steps of information extraction, information fusion and knowledge graph construction. Due to the multiple sources of the media asset data, when a knowledge graph facing the multi-source media asset data is constructed, a TextRank keyword extraction algorithm, entity attribute secondary classification, a convolutional neural network and an attention mechanism are introduced into entity alignment, so that semantic association between entities and attributes can be better understood, and the accuracy of media asset entity alignment is improved.

Description

A method for building knowledge graph of multi-source media resources based on TextRank algorithm combined with convolutional neural network model

technical field

The invention relates to the field of artificial intelligence, and mainly relates to the technology of knowledge graph of media resources.

Background technique

In 2012, Google proposed a knowledge graph to optimize its search engine. The search method of knowledge graph is different from that of traditional search engines. The traditional search engine is based on keyword search, while knowledge graph is based on the relationship between entities, which optimizes the search method and improves the search quality.

With the advent of the era of big data, the TV business has entered a stage of stock development, which requires professional and refined operation capabilities. Unified media assets is the function of aggregating, integrating, processing and outputting media asset information. Through user behavior data and media asset information data, users' viewing behaviors can be deeply mined, which can improve service usage perception and increase user activity and stickiness. However, in the existing media asset management system, media asset information is seriously missing or wrong, and manual calibration and matching can no longer meet the urgent needs of precision marketing.

The patent "a method and device for merging media assets" (CN202010128799.8) uses name similarity and theme similarity to merge media assets, but its matching effect is low, and it does not have automatic learning ability. The analysis mainly relies on manual, automatic To a lesser degree, the results are presented in a traditional way.

The patent "a method and system for integrating media asset data" (CN201610777461.9) adopts the method of field-by-field comparison to match the pending media asset with the standard media asset, lacking the association relationship between media asset entities and attributes, and the media asset data is presented in the same way. Weak association, the constructed media library does not have cognitive ability.

Therefore, there is a need for a method that can efficiently and accurately align media asset entities to construct a media asset knowledge graph.

SUMMARY OF THE INVENTION

This Summary is provided to introduce a simplified form of some concepts that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used as an aid in determining the scope of the claimed subject matter.

In the present invention, in the media asset information fusion, instead of the traditional entity alignment method, a method combining manual intervention and machine learning algorithm is adopted. Firstly, the TextRank keyword extraction algorithm is introduced to obtain keywords that are more relevant to the entity content as media asset tags. On the one hand, it supplements the semantics of entity alignment, and on the other hand, it also supplements the extracted attribute values. Then, according to the entity attributes, detailed secondary classification of entities is carried out, the scope of entity alignment is narrowed, the matching results are intervened in advance, and the complexity of the alignment algorithm is reduced. Finally, the neural network technology is used to learn the deep semantic relationship between entities and attributes, and it has the ability to learn and automatically learn the characteristics of different types of entities and attributes; the attention mechanism is introduced to weight the local feature vectors to realize the integration of input and output. Alignment can also utilize more contextual information of the original data, which can continuously improve the entity alignment accuracy of different data sources.

According to an embodiment of the present invention, a method for constructing a knowledge graph for multi-source media resource data is disclosed, including: extracting media resource information from the multi-source data to form an entity-centric knowledge graph III. Tuple, wherein extracting media asset information includes entity extraction, relationship extraction and entity attribute extraction, the multi-source data comes from web pages including semi-structured data and/or web pages including unstructured data, and the triples have { entity, relationship, attribute}; information fusion is performed on the media asset information to form an updated triple, wherein the information fusion includes entity alignment and attribute unification, wherein the entity alignment adopts attention-based An entity matching model of a mechanism of convolutional neural networks; the knowledge graph is constructed based on the updated triples.

According to another embodiment of the present invention, a system for building a knowledge graph for multi-source media resource data is disclosed, comprising: an information extraction module, the information extraction module is configured to: extract media from multi-source data resource information to form entity-centric knowledge graph triples, wherein the extraction of media resource information includes entity extraction, relation extraction and entity attribute extraction, and the multi-source data comes from web pages including semi-structured data and/or including non-structured data. A web page of structured data, the triplet has the form of {entity, relationship, attribute}; an information fusion module, the information fusion module is configured to: perform information fusion on the media asset information to form an updated triplet, wherein the information fusion includes entity alignment and attribute unification, wherein the entity alignment includes an entity matching model using an attention mechanism-based convolutional neural network; a knowledge graph building module, the knowledge graph building module is is configured to build the knowledge graph based on the updated triples.

According to yet another embodiment of the present invention, a computing device for building a knowledge graph for multi-source media resource data is disclosed, including: a processor; The processor can execute the method as described above when executed.

These and other features and advantages will become apparent upon reading the following detailed description with reference to the associated drawings. It is to be understood that both the foregoing general description and the following detailed description are illustrative only and not restrictive of the claimed aspects.

Description of drawings

In order that the manner in which the above-described features of the present invention can be understood in detail, what has been briefly summarized above may be described in more detail with reference to various embodiments, some aspects of which are illustrated in the accompanying drawings. It is to be noted, however, that the appended drawings illustrate only certain typical aspects of the invention and are therefore not to be considered limiting of its scope, for the description may admit to other equally effective aspects.

1 shows a block diagram of a system 100 for building a knowledge graph for multi-source media resource data according to an embodiment of the present invention;

2 shows a schematic diagram of an entity matching model 200 of an attention mechanism-based convolutional neural network according to an embodiment of the present invention;

Figure 3 shows a flowchart of a method 300 for building a knowledge graph for multi-source property data according to one embodiment of the present invention; and

FIG. 4 shows a block diagram of an exemplary computing device 400 according to one embodiment of the present invention.

Detailed ways

The present invention will be described in detail below in conjunction with the accompanying drawings, and the features of the present invention will be further revealed in the following detailed description.

The following are explanations of terms used in the present invention, which include the ordinary meanings well known to those skilled in the art:

Knowledge graph: a graph-based data structure, each node represents an "entity" (which can be a specific person, thing, object, or an abstract concept), and each edge represents a "relationship" between entities and entities . Both "entities" and "relationships" can have their own "properties". Therefore, "entities", "relationships" and "attributes" constitute the three components of the knowledge graph.

TextRank algorithm: a graph-based ranking algorithm for keyword extraction and document summarization. It uses the co-occurrence information (semantics) between words in a document to extract keywords. It can extract keywords from a given text. Extract the keywords and keyword groups of the text from the text, and extract the key sentences of the text using the extractive automatic summarization method.

Neo4j: A high-performance NOSQL graph database that stores structured data on the network rather than in tables. The bottom layer of Neo4j will store the user-defined nodes and relationships in a graph. In this way, the efficient implementation starts from a certain node and finds the relationship between two nodes through the relationship between nodes. Neo4j queries graph data using Cypher, a descriptive graph query language.

D3 technology: The full name of D3 is Data-Driven Documents, that is, data-driven documents, which are used for data visualization.

At present, it generally takes 5 working days to manually complete the matching of 700,000 program media assets. In order to realize the automatic association of media asset contents, improve the accuracy of search and recommendation, and provide media asset data empowerment for media asset information management in the TV industry, the present invention utilizes the network Media asset information data, through knowledge graph technology to build a knowledge graph for multi-source media asset data.

Due to the multi-source of media asset data, there are differences and redundancy in entity and attribute information. The entity alignment in the current knowledge graph adopts some general entity alignment methods, which cannot well understand the relationship between media asset entities and attributes. Semantic association, the alignment accuracy of media asset entities from multiple data sources is low. When constructing a knowledge graph oriented to multi-source media resource data, the present invention introduces TextRank keyword extraction algorithm, entity attribute secondary classification, convolutional neural network and attention mechanism into entity alignment, which can better understand the relationship between entity and attribute. Semantic association between media assets to improve the alignment accuracy of media asset entities.

Figure 1 shows a block diagram of a system 100 for multi-source property data-oriented knowledge graph construction according to one embodiment of the present invention. As shown in Figure 1, the system 100 is divided into modules, each of which communicates and exchanges data in a manner known in the art. In the present invention, each module can be implemented by software or hardware or a combination thereof. As shown in Figure 1, the system 100 includes an information extraction module 101, an information fusion module 102 and a knowledge graph construction module 103.

Generally speaking, the construction of a media resource knowledge graph mainly includes: selecting appropriate multi-source data, extracting media resource information from the multi-source data, performing data processing on the extracted media resource information, and constructing a knowledge graph based on the processed media resource information. , and publish it in practical applications (eg, search, presentation, etc.).

Specifically, referring to Fig. 1 , the information extraction module 101 extracts media information from different data sources, mainly including entity extraction, relationship extraction, and attribute extraction. The information fusion module 102 fuses the media asset information extracted by the information extraction module 101. The knowledge graph building module 103 builds a knowledge graph according to the fused media asset information.

According to one embodiment of the present invention, the information extraction module 101 is configured to extract media resource information from the multi-source data 104, such as entity extraction, relation extraction, attribute extraction, to form entity-centric knowledge graph triples.

As known to those skilled in the art, media resource information mainly includes information of media resources, such as TV series, movies, variety shows, animation, children, documentaries, music, education, mothers and babies, sports, actors, and stars. Comprehensively considering the scale, quality and difficulty of data acquisition of media resource information data, the multi-source data 104 of the present invention comes from two different types of data sources, namely data sources mainly including unstructured data and data sources mainly including semi-structured data The data source of the data. For example, the data source of unstructured data can be Baidu Encyclopedia, which is an open and free network platform that covers almost all industry knowledge fields. Due to manual editing, the consistency of information structure is poor, especially for film and television media information, the quality of which is also quite different. For this reason, in order to improve the information quality of media resources such as film and television, a data source of semi-structured data, such as Douban Movie, is selected as a supplement to the information of film and television media resources. Douban Movie is a film and television review website with the latest film and television information. Its data structure is better. Of course, the above websites are only given for exemplary purposes, and those skilled in the art can completely obtain media asset information by using other data sources. Of course, it is also possible to use only one type of data source to obtain media asset information.

According to an embodiment of the present invention, extracting the media resource information from the multi-source data 104 mainly includes (1) formulating a crawler program using python language to obtain the knowledge graph data source of the media resource data; (2) obtaining the media resource through web page parsing labels and structured information; and (3) extract triples from the acquired structured information to construct a knowledge graph.

According to an embodiment of the present invention, continuing the above examples of Baidu Encyclopedia and Douban Movie, (1) using python language to formulate a crawler program to obtain media resource data The knowledge graph data source includes: 1.1) directional crawling web page entries according to basic conditions (For example, Baidu Encyclopedia entries). Specifically, the webpage content can be searched according to category feature words, wherein the category feature words refer to words that can distinguish various types of media asset data (for example, ten categories). For example, for TV drama media data, the category feature words select "episode", "diversity", and "TV drama"; for education media data, the category feature words select "education", "preschool", "primary school", " 1.2) According to the category of film and television, including movies, TV series, variety shows, animation, documentaries, crawl Douban movie webpage data.

According to an embodiment of the present invention, (2) obtaining media asset tags and structured information through web page parsing may further include: 2.1) extracting keywords from a knowledge graph data source of media asset data to construct a media asset tag library; 2.2) Extract structured information from media asset data knowledge graph data sources.

According to an embodiment of the present invention, 2.1) extracting keywords from the media resource data knowledge graph data source to construct a media resource tag library may further include using the TextRank algorithm to extract keywords from the text of the web page (such as the introduction of Baidu Encyclopedia). The keyword extraction may include the following steps:

① First, the jieba word segmentation tool is used to segment the text extracted from the web page, and at the same time, part of speech is marked, and stop words are filtered out, and only words with nouns, verbs and adjectives are retained.

②Construct the candidate keyword graph G=(V,E) from the segmented words, where V is composed of the candidate keywords generated by ①, and then use the co-occurrence relationship to construct the edge between any two points. An edge exists between two nodes only if their corresponding words co-occur in a window of length K, where K represents the window size, that is, at most K words co-occur.

③ Iteratively propagate the weight of each node until convergence.

④ Sort the node weights in reverse order to get the most important T words as candidate keywords.

⑤ Obtain the most important T words from ④ and mark them in the original text. If adjacent phrases are formed, they are combined into multi-word keywords as media asset labels.

According to an embodiment of the present invention, 2.2) extracting structured information from the knowledge graph data source of media asset data may further include grabbing the entry name of the web page, that is, the entity name, according to font characteristics. For example, the entity name can be fetched according to the title data format of the fetched web page url (the title is in the h1 sub-tag under the class lemmaWgt-lemmaTitle-title).

Generally speaking, in the field of media assets, entities are mainly divided into audiovisual entities (eg, video, audio, etc.), non-audiovisual entities (eg, books, exhibitions, etc.) and character entities. The entity name is the name of the entity, such as the title of a TV series, the title of a movie, the title of a book, and so on. The attributes of an entity mainly include information describing the entity. For example, the attributes of audio-visual entities mainly include Chinese name, alias, genre, production company, first broadcast time, director, starring and so on. Non-audio-visual attributes mainly include Chinese name, alias, venue, time, organizer, etc. Character entity attributes mainly include Chinese name, alias, nationality, ethnicity, blood type, height, constellation, representative works, main achievements, occupation, etc.

For example, Baidu Encyclopedia entries are mainly composed of videos, pictures, texts and diagrams, and the present invention mainly extracts table data in web pages. The corresponding relationship between the field name and field value in the table just corresponds to the entity, relationship and attribute in the knowledge graph. The relationship between entities mainly includes the "participating/directing/writing/hosting/participating" relationship between character entities and audiovisual/non-audiovisual entities. The parsing of the Douban Movie web page is similar to that described above. For example, the source code for storing entity (eg, movie) information can be found in the source code of the webpage, and the webpage entry name and corresponding field name and field value can be parsed according to the tag data format of the title and entity (eg, movie) information. . For example, the web page entry name and the corresponding field name and field value of the web page can be parsed by identifying "$$" in the web page source code.

According to an embodiment of the present invention, (3) extracting the triples for constructing the knowledge graph from the acquired structured information further includes constructing an entity-centric {entity , relation, attribute } triplet.

According to an embodiment of the present invention, the information fusion module 102 is configured to fuse the media asset information based on the triplet of the knowledge graph formed by the information extraction module 101.

According to an embodiment of the present invention, the fusion of media resource information further includes (1) aligning media resource entities; (2) unifying attribute vocabulary in different and heterogeneous data sources and (3) combining entity keywords (ie, the media asset tag) incorporates the attribute value as a complement.

According to an embodiment of the present invention, for the same entity extracted from different data sources may have different names, it is necessary to compare the similarity between the entity and the corresponding attribute. (1) Entity alignment of media assets may include entity classification and entity matching.

According to an embodiment of the present invention, in the entity classification, the entity is classified into two levels according to the entity attributes extracted by the information extraction module 101, thereby effectively narrowing the entity matching range. In the above-mentioned example of the present invention, the secondary classification is carried out as follows, and the classification result categories are as follows:

Of course, the above primary and secondary classifications are only exemplary, and those skilled in the art can completely define different multi-level classifications according to different media resources.

According to an embodiment of the present invention, in entity matching, based on entities under each category, entities, entity attributes and entity keywords (ie, information extraction The keywords of the media assets acquired by the module 101) are used for semantic matching. FIG. 2 shows a schematic diagram of an entity matching model 200 of an attention mechanism-based convolutional neural network according to an embodiment of the present invention. As can be seen with reference to Figure 2, entity matching may include the following steps:

1) Build a corpus. The media asset entity matching training set is constructed by using the film and television entry names, attribute values and the extracted media asset labels. First, the attribute values of different source entities to be matched are aligned according to the attribute names, and the word2Vec model is used to match the entity names and entities in the training set. Attribute values and entity keywords are vectorized to obtain word vector representations of entities, attributes and keywords;

2) The entity matching model of the convolutional neural network based on the attention mechanism performs entity matching. The entity matching model consists of an input layer, a convolutional layer, an attention layer, a pooling layer, and an output layer.

· Input layer. Preprocess entities, attribute values and keywords to obtain the sequence word vector representation x=(x ₁ ,x ₂ ,...,x _i ,...,x _n ), X _i ∈R ^d represents the i-th word vector in the sequence vector , the input word vectors are arranged from top to bottom according to the order of entity names, attributes, and keywords. The dimension of each word vector is k, then the dimension of the generated matrix is n*k, and the unregistered words are filled with 0;

· Convolutional layers. The local features and position features of the input matrix are obtained through the convolution operation. The convolution formula is as follows:

Among them, w∈R ^d*k is the filter, b∈R is the bias, f is the nonlinear function ReLU, k is the size of the convolution convolution kernel, and represents the number of words in the vertical convolution operation. 4 and 5 three convolution kernel lengths, the calculated local semantic feature vector h _i of the window, and finally output three sequence local semantic feature vector matrices.

· Attention layer. The attention mechanism is introduced to calculate the local semantic feature vector matrix of the sequence obtained through three convolution kernels. Take the mean of the local semantic feature vectors of the sequence as auxiliary information:

Using attention mechanism:

m _j =v ^T tanh(W _s h ^j +W _h h'+b)

c=∑ _l a _i h ^j

A series of feature information codes c _i are obtained through the above attention mechanism, where a _i represents the attention weight of the local semantic feature vector, W _s and W _h are the weight matrix, b is the bias, h _j is the local semantic feature vector, m _j represents the matching score between s _j and h'.

· Pooling layer. The max-pooling (maximum pooling) method is used for the feature information code c output by the attention layer, and some features that are weakly related to the topic are discarded, and the maximum value of each sequence feature information code is obtained, which can effectively avoid two input words. For the problem of inconsistency in numbers, the final pooling layer outputs the maximum value of each feature information code, and after connection, a final semantic representation vector of a one-dimensional vector is obtained;

· Output layer. The output layer mainly realizes cosine similarity matching. The final semantic representation vector of the sequence to be matched is obtained through the pooling layer. The cosine similarity is calculated according to the following formula to obtain the matching score l. The threshold is set to 0.85. If the calculated matching score l is greater than 0.85, it is considered that If the match is successful, it means that it is the same entity. If it is less than 0.85, it is considered that the match fails, indicating that there are two types of entities.

Of course, the above thresholds are only exemplary, and other thresholds can be used to perform probability matching as needed.

According to an embodiment of the present invention, in (2) unifying attribute vocabulary in different and heterogeneous data sources, if there is a successfully matched media resource entity in the target data source, update its attributes; if the matching fails , it means that the media asset entity does not exist in the target data source, and the entity needs to be added to the target data source. When the attribute value is aligned and updated after the entity matching is successful, there is an attribute value ambiguity, and the newly updated attribute value is retained. Specifically, in attribute unification, the "attribute" value in the triple is updated.

For example, if a certain video A exists in both Baidu Baike and Douban Movie, after performing entity alignment as described above, the attributes for video A from Baidu Baike and Douban Movie can be unified, and Update attribute values in triples for video A. That is, if the triplet is constructed based on the entry from Baidu Baike, the Baidu entry attribute is updated with the data in Douban Movie.

According to an embodiment of the present invention, after entity alignment and attribute unification, entity keywords (media asset tags) may be included in attribute values as supplements to further update triples for entities. For example, attribute values in triples will be further updated by including media asset tags.

According to an embodiment of the present invention, the knowledge graph building module 103 is configured to build a knowledge graph according to the fused media asset information. Specifically, the knowledge graph building module 103 is configured to store and build a media resource knowledge graph with the graph database Neo4j according to the finally obtained entity-centric triples. Those skilled in the art can fully understand how to use the graph database Neo4j to construct a knowledge graph, and this construction method is not within the protection scope of the present invention.

In practice, after constructing the knowledge graph of media assets, D3 technology can be used to realize the visualization of media asset data knowledge graph web pages. Specifically, the Cypher statement query can be performed according to the keywords input from the web page (for example, from the user), the result of the Neo4j query can be encapsulated into JSON, and passed to D3 to complete the drawing and display on the web page.

FIG. 3 shows a flowchart of a method 300 for building a knowledge graph for multi-source media resource data according to an embodiment of the present invention. In the context of the present invention, multi-source property data may, for example, come from semi-structured web pages and/or unstructured web pages. Of course, multi-source media asset data can also come from other types of sources, such as specific media libraries, and the like.

In step 301, media asset information is extracted from multi-source data to form entity-centric knowledge graph triples. According to an embodiment of the present invention, extracting media asset information includes entity extraction, relation extraction and attribute extraction. According to an embodiment of the present invention, extracting the media asset information further includes parsing the web page to extract keywords for establishing a media asset tag library and extracting structured information.

Specifically, the keywords of the webpage are extracted as the media asset tag library through the TextRank algorithm; the entry name of the webpage, that is, the entity name, is grabbed according to the font characteristics. According to an embodiment of the present invention, according to the characteristics of media asset information in the television industry, the media asset entities are divided into audiovisual entities, non-audiovisual entities and character entities, and the relationship between the entities is determined through the association relationship between these three types of entities , and get properties corresponding to entities and relationships. Next, the triples for constructing the knowledge graph are extracted from the acquired structured information to form entity-centered triples {entity, relationship, attribute}.

In step 302, information fusion is performed on the media asset information extracted in step 301 to form an updated triplet. According to one embodiment of the present invention, step 302 further includes entity alignment and attribute unification. In entity alignment, according to entity attributes, the entity is classified into two levels, and the scope of entity matching is narrowed. And, through an entity matching model based on an attention mechanism convolutional neural network, semantic matching is performed on entities, entity attributes, and entity keywords to align entities. After entity alignment, attribute vocabularies from different data sources are unified, and media asset tags are supplemented as attribute vocabularies to update entity-centric triples.

At step 303, a media asset knowledge graph is constructed based on the updated triples. According to an embodiment of the present invention, a graph database Neo4j is used to store and build a knowledge graph of media assets.

To sum up, compared with the prior art, the present invention has the following main advantages:

1. Efficient: Compared with the manual matching media asset management system, the use of knowledge graph algorithm can realize the automatic association of media asset content and improve the accuracy of search and recommendation;

2. Versatility: Media asset data comes from general web pages, such as Baidu Baike and Douban Movies, with complete and real-time data, suitable for media asset business platforms of various application scenarios;

3. Possess learning ability: Using TextRank algorithm and neural network technology in entity alignment, it can be automatically tuned through training, and the accuracy of final entity matching is continuously improved;

4. It has the ability to correct by artificial experience: Use entity attributes in entity alignment to perform secondary classification of entities, narrow the scope of entity matching, and effectively reduce the matching error rate of unsupervised machine learning algorithms based on human experience.

4 illustrates a block diagram of an exemplary computing device 400, which is one example of a hardware device that may be applied to aspects of the present invention, according to one embodiment of the present invention. Computing device 400 can be any machine that can be configured to perform processing and/or computing, and can be, but is not limited to, workstations, servers, desktops, laptops, tablets, personal digital processors, smartphones , an on-board computer, or any combination of these. Computing device 400 may include components that may connect or communicate with bus 402 via one or more interfaces. For example, computing device 400 may include a bus 402 , one or more processors 404 , one or more input devices 406 , and one or more output devices 408 . The one or more processors 404 may be any type of processor and may include, but are not limited to, one or more general-purpose processors and/or one or more special-purpose processors (eg, specialized processing chips). Input device 406 may be any type of device capable of inputting information to a computing device and may include, but is not limited to, a mouse, keyboard, touch screen, microphone, and/or remote controller. Output device 408 may be any type of device capable of presenting information and may include, but is not limited to, displays, speakers, video/audio output terminals, vibrators, and/or printers. Computing device 400 may also include or be connected to non-transitory storage device 410, which may be any storage device that is non-transitory and capable of data storage, and that is non-transitory. Transient storage devices may include, but are not limited to, magnetic disk drives, optical storage devices, solid state memory, floppy disks, floppy disks, hard disks, magnetic tape or any other magnetic medium, optical disks or any other optical medium, ROM (read only memory), RAM (random memory) fetch memory), cache memory, and/or any memory chip or tape cartridge, and/or any other medium from which a computer may read data, instructions, and/or code. Non-transitory storage device 410 may be detached from the interface. The non-transitory storage device 410 may have data/instructions/code for implementing the methods and steps described above. Computing device 400 may also include communication device 412. Communication device 412 may be any type of device or system capable of communicating with internal devices and/or with a network and may include, but is not limited to, modems, network cards, infrared communication devices, wireless communication devices, and/or chipsets, such as Bluetooth devices , IEEE 1302.11 devices, WiFi devices, WiMax devices, cellular communication devices and/or similar devices.

The bus 402 may include, but is not limited to, an Industry Standard Architecture (ISA) bus, a Microchannel Architecture (MCA) bus, an Enhanced ISA (EISA) bus, a Video Electronics Standards Association (VESA) local bus, and a Peripheral Component Interconnect (PCI) bus.

Computing device 400 may also include working memory 414, which may be any type of working memory capable of storing instructions and/or data that facilitates the operation of processor 404 and may include, but is not limited to, random access memory and/or read-only storage device.

Software components may be located in working memory 414, including, but not limited to, operating system 416, one or more application programs 418, drivers, and/or other data and code. Instructions for implementing the above-described methods and steps of the present invention may be included in the one or more application programs 418, and the present invention may be implemented by the processor 404 reading and executing the instructions of the one or more application programs 418 The above method 300.

It should also be recognized that variations may be made according to specific needs. For example, custom hardware may also be used, and/or certain components may be implemented in hardware, software, firmware, middleware, microcode, hardware description voice, or any combination thereof. Additionally, connections to other computing devices, such as network input/output devices, etc., may be employed. For example, programmable logic circuits (eg, programmable logic circuits including Field Programmable Gate Arrays (FPGAs) and/or Programmable Logic Arrays (PLAs) may be programmed with assembly language or hardware programming languages (eg, VERILOG, VHDL, C++). ) utilize logic and algorithms in accordance with the present invention to implement some or all of the disclosed methods and apparatus.

Although aspects of the present invention have so far been described with reference to the accompanying drawings, the above-described method and apparatus are merely examples, and the scope of the present invention is not limited to these aspects, but only by the appended claims and their equivalents. Various components may be omitted or may be replaced by equivalent components. Additionally, the steps may also be performed in an order different from that described in this disclosure. Furthermore, the various components can be combined in various ways. It is also important that, as technology develops, many of the components described may be replaced by equivalent components that appear later.

Claims (10)

1. A method for building a knowledge graph for multi-source media resource data, comprising: Extract media resource information from multi-source data to form entity-centric knowledge graph triples, wherein extracting media resource information includes entity extraction, relation extraction and entity attribute extraction, the multi-source data comes from including semi-structured data web pages and/or web pages including unstructured data, the triples have the form {entity, relationship, attribute}; Information fusion is performed on the media asset information to form an updated triple, wherein the information fusion includes entity alignment and attribute unification, wherein the entity alignment employs entity matching based on an attention mechanism convolutional neural network Model; The knowledge graph is constructed based on the updated triples. 2. The method of claim 1, wherein the entities are classified into audiovisual entities, non-audiovisual entities, and character entities. 3 . The method according to claim 1 , wherein extracting the media asset information further comprises: using the TextRank algorithm to perform entity keyword extraction on the text of the web page to construct a media asset tag library. 4 . 4. The method of claim 3, wherein the entity alignment further comprises performing a secondary classification of the entities. 5. The method according to claim 4, wherein the entity alignment further comprises: based on the entities under each classification, through the entity matching model of the convolutional neural network based on the attention mechanism, the entity, the entity attribute and the entity keyword are analyzed. Do semantic matching. 6. The method of claim 3, wherein the attribute unification further comprises unifying attribute vocabularies from different data sources to update entity-centric triples. 7. The method of claim 6, wherein the information fusion further comprises supplementing the entity keyword as an attribute vocabulary to update entity-centric triples. 8. A system for building a knowledge graph for multi-source media resource data, comprising: an information extraction module, the information extraction module is configured to: extract media resource information from multi-source data to form an entity-centric knowledge graph triple, wherein the extraction of media resource information includes entity extraction, relationship extraction and entity attribute Extracting, the multi-source data comes from web pages comprising semi-structured data and/or web pages comprising unstructured data, and the triplet has the form {entity, relationship, attribute}; an information fusion module configured to perform information fusion on the media asset information to form an updated triple, wherein the information fusion includes entity alignment and attribute unification, wherein the entity alignment Including an entity matching model using an attention-based convolutional neural network; A knowledge graph building module configured to build the knowledge graph based on the updated triples. 9 . The system of claim 8 , wherein extracting the media asset information further comprises: extracting entity keywords from the text of the web page by using the TextRank algorithm to construct a media asset tag library. 10 . 10. A computing device for building a knowledge graph for multi-source media resource data, comprising: processor; a memory storing instructions which, when executed by the processor, are capable of performing the method of any of claims 1-7.

Images