CN109635171A - A kind of fusion reasoning system and method for news program intelligent label - Google Patents
A kind of fusion reasoning system and method for news program intelligent label Download PDFInfo
- Publication number
- CN109635171A CN109635171A CN201811528577.4A CN201811528577A CN109635171A CN 109635171 A CN109635171 A CN 109635171A CN 201811528577 A CN201811528577 A CN 201811528577A CN 109635171 A CN109635171 A CN 109635171A
- Authority
- CN
- China
- Prior art keywords
- label
- library
- reasoning
- entity
- recognition
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 230000004927 fusion Effects 0.000 title claims abstract description 38
- 238000000034 method Methods 0.000 claims abstract description 30
- 239000000463 material Substances 0.000 claims abstract description 25
- 238000000605 extraction Methods 0.000 claims abstract description 15
- 238000004458 analytical method Methods 0.000 claims abstract description 14
- 230000001502 supplementation Effects 0.000 claims abstract description 4
- 230000001537 neural Effects 0.000 claims description 29
- 239000011159 matrix material Substances 0.000 claims description 8
- 238000005516 engineering process Methods 0.000 claims description 7
- 238000010586 diagram Methods 0.000 claims description 5
- 239000000284 extract Substances 0.000 claims description 4
- 238000010276 construction Methods 0.000 claims description 3
- 238000004064 recycling Methods 0.000 claims description 3
- 230000011218 segmentation Effects 0.000 claims description 3
- 230000000875 corresponding Effects 0.000 description 2
- 238000007418 data mining Methods 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 238000003058 natural language processing Methods 0.000 description 2
- 230000003321 amplification Effects 0.000 description 1
- 230000000295 complement Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000035800 maturation Effects 0.000 description 1
- 238000002844 melting Methods 0.000 description 1
- 238000003199 nucleic acid amplification method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING; COUNTING
- G06K—GRAPHICAL DATA READING; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
- G06K9/00—Methods or arrangements for recognising patterns
- G06K9/62—Methods or arrangements for pattern recognition using electronic means
- G06K9/6288—Fusion techniques, i.e. combining data from various sources, e.g. sensor fusion
Abstract
The invention discloses a kind of fusion reasoning system and method for news program intelligent label, it is related to news program label technique field, the present invention includes intelligent recognition actuator, history tab library, internal knowledge base, internal case library and analysis ratiocination device, intelligent recognition actuator executes the identification mission of all kinds of news program materials, carries out basic label extraction to video image, voice and text information；History tab library stores material, metadata and label；Internal knowledge base is for supplementing intelligent recognition as a result, providing more information for subsequent analysis reasoning；Internal case library is the case set established based on history tab library；Analysis ratiocination device is used for the fusion reasoning of intelligent label, reasoning device comprising rule-based reasoning device and based on deep learning, the present invention comprehensively utilizes intelligent identification Method, internal knowledge base and internal case library are established based on history tab library, the automatic fusion reasoning of news program label is completed, precise and high efficiency of classifying.
Description
Technical field
The present invention relates to news program label technique fields, more particularly to a kind of melting for news program intelligent label
Close inference system and method.
Background technique
With the development of the times with progress, the label of media content is ubiquitous, and many practical businesses require to mark
The participation of label, such as content retrieval and examination etc..Tradition is all that media content is identified and marked by artificial mode
The addition of label, but with the amplification of the increase of media data amount and business demand, conventional inefficiencies manually add tagged side
Formula can no longer meet business demand rapidly and efficiently.At the same time, intelligent identification technology continuous development and deep learning etc.
Technology gradually maturation make for Multi-media Material it is automatic, accurately generate content tab and become a reality gradually.
In one section of Press release, a variety of expression forms of video, audio, text are usually contained, these media contents are not
Same media event is described in same dimension, the general mutual of the information for including is overlapped or is complementary to one another, can be by artificial
It browses Press release and obtains the keywords, including time, place, personage, event, reason such as five "W's" of news etc..To media content
Intelligent identification Method it is gradually mature, for example video can be identified using recognition of face and OCR, to audio using voice knowledge
Not, and the information for how integrating different dimensions carrys out reasoning and concludes to form intelligent label, is the important topic under medium technique.
The mode that label in past media content is generally manually extracted, but manually extracted can not be adapted to the matchmaker of magnanimity
Hold in vivo, the inefficiency manually extracted and easy error；On the other hand, the multiclass recognition result of video and audio, text is had no
Reasonable comprehensive utilization, the difficulty utilized are also bigger.
Summary of the invention
It is an object of the invention to: it is extracted to solve the label in existing news program media content by artificial,
Inefficiency is easy the problem of error, and the present invention provides a kind of fusion reasoning system and method for news program intelligent label, comprehensive
It closes and utilizes intelligent identification Method, internal knowledge base and internal case library are established based on history tab library, complete news program label
Automatic fusion reasoning, classify precise and high efficiency.
The present invention specifically uses following technical scheme to achieve the goals above:
A kind of fusion reasoning system of news program intelligent label, including intelligent recognition actuator, history tab library, inside
Knowledge base, internal case library and analysis ratiocination device,
Intelligent recognition actuator: for executing the identification mission of all kinds of news program materials, to video image, voice and
Text information carries out basic label extraction；
History tab library: for storing material, metadata and label；
Internal knowledge base: the external knowledge map established based on history tab library, the entity packet of the external knowledge map
Personage, country, the political situation of the time, event etc. are included, internal knowledge base is for supplementing intelligent recognition as a result, providing more for subsequent analysis reasoning
Information；
Internal case library: the case set established based on history tab library, including it is representative, possess complete first number
According to the case with label, it to be used for deep learning training process；
Analysis ratiocination device: for the fusion reasoning of intelligent label, comprising rule-based reasoning device and it is based on deep learning
Reasoning device；Rule-based reasoning device comprehensive utilization intelligent recognition result and internal knowledge base make inferences；Based on depth
Using internal case library training deep learning model, the model for recycling training to obtain makes inferences the reasoning device of habit.
Further, the intelligent recognition actuator identifies all kinds of news program materials, and identification process uses people
Face identifies actuator, OCR actuator, speech recognition actuator and NLP actuator.
Further, material includes text, voice and video image in the history tab library；Metadata includes material
The information such as title, keyword, creation time, place, type；When label includes field, Sentiment orientation involved in material, occurs
Between, personage, scene etc..
A kind of fusion reasoning method of news program intelligent label, includes the following steps:
S1, building news knowledge base: history tab library, history material database, internet and other construction of knowledge base are utilized
News knowledge base；
S2, it constructs internal case library: extracting the text material with field label from history tab library and have field
The picture of scape label forms internal case library, and by the case textual data value in internal case library；
S3, training deep learning model: the case training deep neural network by numeralization, the depth nerve are utilized
The training process of network includes text classification training process and scene Recognition training process；
S4, it carries out fusion reasoning: according to the type of input program, using the corresponding intelligent recognition of intelligent recognition actuator
Strategy carries out content recognition, then carries out rule-based reasoning to recognition result using internal knowledge base to judge program class
Not, the first candidate newly-increased tally set is obtained, and carries out text classification and scene Recognition using trained deep neural network,
Obtain the second candidate newly-increased tally set, user to the first candidate newly-increased tally set and the second candidate newly-increased tally set carry out selection and
Correction, exports final program label.
Further, the entity of the news knowledge base in the S1 includes time, place, event, personage etc., the news
The building of knowledge base includes the following steps:
S1.1, global ontology library is constructed: according to the amusement of news, sport, finance and economics, the people's livelihood, the political situation of the time, tourism, military affairs
Equal classification carry out each field ontology library building respectively, and building range includes concept, concept hierarchy, attribute, attribute Value Types, attribute
It is worth codomain, relationship, contextual definition domain concept set and relation value codomain；
S1.2, it obtains entity: history tab library entity is obtained, if history tab library entity information is imperfect, in history
Entity information supplement is carried out in material database and internet；
S1.3, entity assessment: the confidence level of the entity of acquisition is quantified, the lower entity of confidence level is given up；
S1.4, knowledge fusion: by entity disambiguate and coreference resolution method, will after entity is assessed remaining chain of entities
It is connected in current knowledge library, and the related entities in third party's knowledge base is merged into current knowledge library；
S1.5, knowledge reasoning: the relationship, entity attributes in reasoning current knowledge library between entity and the layer between ontology
Knowledge reasoning result Jing Guo manual examination and verification is added in current knowledge library, obtains news knowledge base by secondary relationship.
Further, global ontology library is constructed in the S1.1 to include the following steps:
S1.1.1: label is divided into multiple fields according to news category from history tab library；
S1.1.2: generating according to the conclusion that history tab carries out each field ontology library, and artificial according to news rule
Additions and deletions adjust each field ontology library；
S1.1.3: each field ontology library and existing knowledge map sheet are merged using rules such as similitude detection and Conflict solvings
Body library obtains global ontology library.
Further, entity acquisition includes the following steps: in the S1.2
S1.2.1: the table name and field information of analysis of history tag library extract what history tab library interior label was characterized
Relationship between correspondent entity, attribute and entity；
S1.2.2: by the methods of speech recognition, subtitle recognition, images steganalysis by video/audio in history material
It is converted into text and then news manuscript is extracted by the methods of natural language processing and data mining and identification obtains text
In entity, relationship and attribute；
S1.2.3: entity information is acquired from internet, the keyword of retrieval is entity name, and then obtains entity attribute
And relation information, such as birthday, nationality, the companion of personage etc., the administrative area type in place, alias, weather etc..
Further, the text classification training process in the S3 based on deep neural network includes the following steps:
S3.1.1: word segmentation processing is carried out to text and is serialized, removal obtains word sequence to meaningless word of classifying；
S3.1.2: word numbered sequence is converted by word sequence；
S3.1.3: word order column number is separately converted to the term vector of n dimension；
S3.1.4: term vector is formed into text matrix by word order, the every a line of matrix is all the term vector of a word；
S3.1.5: it is used for text matrix to train deep neural network.
Further, the scene Recognition training process in the S3 based on deep neural network includes the following steps:
S3.2.1: expand scene Recognition image pattern by operations such as image cut, rotation, scalings；
S3.2.2: adjustment scene Recognition image size and carry out other pretreatment after, for training deep neural network.
Further, content recognition is carried out using intelligent recognition actuator in the S4, using following strategy:
Tactful A: when input program category is picture, picture input recognition of face device is obtained into people tag, and will figure
Piece inputs OCR identifier and recognition result input NLP actuator is carried out entity extraction, the mark such as acquisition time, place, event
Label；
Tactful B: when input program category is voice, which is inputted into speech recognition device and recognition result is inputted into NLP
Actuator carries out entity extraction, obtains the labels such as time, place, event；
Tactful C: when input program category is video, picture frame is handled using tactful A, using tactful B to video
Voice is handled；
Tactful D: when input program category is text, text input NLP actuator is subjected to entity extraction, when acquisition
Between, place, the labels such as event.
Further, rule-based reasoning includes the following steps: in the S4
S4.1.1: respectively using the initial associated metadata of the extracted basic label of intelligent recognition actuator and program as
Entity；
S4.1.2: position of each entity in internal knowledge base is obtained using the methods of entity disambiguation and coreference resolution；
S4.1.3: go out adjacent node and relationship as subgraph by center Node extraction of each entity；
S4.1.4: the sub-graph data of each entity is inputted into trained GCN and is calculated using distributed diagram technology, is pushed away
Reason obtains program field；
S4.1.5: the program field that basic label and reasoning are obtained is as the first candidate newly-increased tally set.
Further, text classification is carried out based on deep neural network in the S4 and scene Recognition includes the following steps:
S4.2.1: trained deep neural network will be inputted by pretreated text data, and will obtain field label；
S4.2.2: trained deep neural network will be inputted by pretreated scene Recognition image data, must shown up
Scape label；
S4.2.3: using obtained field label and scene tag as the second candidate newly-increased tally set.
Further, in the S4 after final program label, the deposit of program, final program label and metadata information is gone through
History tag library is updated history tab library, and updates internal knowledge base and internal case library simultaneously.
Beneficial effects of the present invention are as follows:
1, fusion reasoning system of the invention is based on history tab library and establishes internal case library and internal knowledge base, for fusion
Reasoning process provides authentic and valid foundation, and proposes rule-based reasoning device and the reasoning device based on deep learning, from knowing
Know map and the multiple dimensions of deep learning carry out tag fusion reasoning, updates history tab library, shape while exporting the reasoning results
At reasoning process closed loop, inference system accuracy will be continuously improved.
2, fusion reasoning method of the invention is based on intelligent recognition process, and mature intelligent identification technology is incorporated fusion and is pushed away
Reason process, reduces the data dimension of unstructured data, while also reducing subsequent reasoning difficulty；On the other hand it comprehensively utilizes
Distributed figure computing technique and figure neural network construct news program disaggregated model, disaggregated model precise and high efficiency.
Detailed description of the invention
Fig. 1 is the overall schematic of fusion reasoning system of the present invention.
Fig. 2 is the building flow diagram of news knowledge base of the present invention.
Fig. 3 is fusion reasoning flow diagram of the present invention.
Specific embodiment
In order to which those skilled in the art better understand the present invention, with reference to the accompanying drawing with following embodiment to the present invention
It is described in further detail.
Embodiment 1
As shown in Figure 1, the present embodiment provides a kind of fusion reasoning system of news program intelligent label, including intelligent recognition
Actuator, history tab library, internal knowledge base, internal case library and analysis ratiocination device,
Intelligent recognition actuator: for executing the identification mission of all kinds of news program materials, identification process is known using face
Other actuator, OCR actuator, speech recognition actuator and NLP actuator carry out video image, voice and text information
Basic label extracts；
History tab library: for storing material, metadata and label, in history tab library material include text, voice and
Video image；Metadata includes the information such as title, keyword, creation time, place, the type of material；Label includes material institute
Field, Sentiment orientation, time of origin, personage, scene for being related to etc., the field include amusement, the political situation of the time, people's livelihood etc., feelings
Sense tendency includes positive and negative；
Internal knowledge base: the external knowledge map established based on history tab library, the entity packet of the external knowledge map
Personage, country, the political situation of the time, event etc. are included, internal knowledge base is for supplementing intelligent recognition as a result, providing more for subsequent analysis reasoning
Information；
Internal case library: the case set established based on history tab library, including it is representative, possess complete first number
According to the case with label, it to be used for deep learning training process；
Analysis ratiocination device: for the fusion reasoning of intelligent label, comprising rule-based reasoning device and it is based on deep learning
Reasoning device；Rule-based reasoning device comprehensive utilization intelligent recognition result and internal knowledge base make inferences；Based on depth
Using internal case library training deep learning model, the model for recycling training to obtain makes inferences the reasoning device of habit.
Based on above-mentioned fusion reasoning system, the present embodiment also provides a kind of fusion reasoning side of news program intelligent label
Method includes the following steps:
S1, building news knowledge base: history tab library, history material database, internet and other construction of knowledge base are utilized
News knowledge base；The entity of news knowledge base includes time, place, event, personage etc., and the building of the news knowledge base includes
Following steps:
S1.1, global ontology library is constructed: according to the amusement of news, sport, finance and economics, the people's livelihood, the political situation of the time, tourism, military affairs
Equal classification carry out each field ontology library building respectively, and building range includes concept, concept hierarchy, attribute, attribute Value Types, attribute
It is worth codomain, relationship, contextual definition domain concept set and relation value codomain, specifically:
S1.1.1: label is divided into multiple fields according to news category from history tab library；
S1.1.2: generating according to the conclusion that history tab carries out each field ontology library, and artificial according to news rule
Additions and deletions adjust each field ontology library；
S1.1.3: each field ontology library and existing knowledge map sheet are merged using rules such as similitude detection and Conflict solvings
Body library obtains global ontology library；
S1.2, it obtains entity: obtaining history tab library entity, the entity refers to the vertex in knowledge mapping, if history
Tag library entity information is imperfect, then entity information supplement is carried out in history material database and internet, specifically:
S1.2.1: the table name and field information of analysis of history tag library extract what history tab library interior label was characterized
Relationship between correspondent entity, attribute and entity；
S1.2.2: by the methods of speech recognition, subtitle recognition, images steganalysis by video/audio in history material
It is converted into text and then news manuscript is extracted by the methods of natural language processing and data mining and identification obtains text
In entity, relationship and attribute；
S1.2.3: entity information is acquired from internet, the keyword of retrieval is entity name, and then obtains entity attribute
And relation information, such as birthday, nationality, the companion of personage etc., the administrative area type in place, alias, weather etc.；
S1.3, entity assessment: the confidence level of the entity of acquisition is quantified, the lower entity of confidence level is given up；
S1.4, knowledge fusion: by entity disambiguate and coreference resolution method, will after entity is assessed remaining chain of entities
It is connected in current knowledge library, and the related entities in third party's knowledge base is merged into current knowledge library；
S1.5, knowledge reasoning: the relationship, entity attributes in reasoning current knowledge library between entity and the layer between ontology
Knowledge reasoning result Jing Guo manual examination and verification is added in current knowledge library, obtains news knowledge base by secondary relationship；
S2, it constructs internal case library: extracting the text material with field label from history tab library and have field
The picture of scape label forms internal case library, and by the case textual data value in internal case library, the field label includes
Science and technology, sport, amusement, political situation of the time etc.；
S3, training deep learning model: the case training deep neural network by numeralization, the depth nerve are utilized
The training process of network includes text classification training process and scene Recognition training process；
The text classification training process based on deep neural network includes the following steps:
S3.1.1: word segmentation processing is carried out to text and is serialized, removal obtains word sequence to meaningless word of classifying；
S3.1.2: word numbered sequence is converted by word sequence；
S3.1.3: word numbered sequence is separately converted to the term vector of n dimension；
S3.1.4: term vector is formed into text matrix by word order, the every a line of matrix is all the term vector of a word；
S3.1.5: it is used for text matrix to train deep neural network.
The scene Recognition training process based on deep neural network includes the following steps:
S3.2.1: expand scene Recognition image pattern by operations such as image cut, rotation, scalings；
S3.2.2: adjustment scene Recognition image size and carry out other pretreatment after, for training deep neural network；
S4, it carries out fusion reasoning: according to the type of input program, using the corresponding intelligent recognition of intelligent recognition actuator
Strategy carries out content recognition, then carries out rule-based reasoning to recognition result using internal knowledge base to judge program class
Not, the first candidate newly-increased tally set is obtained, and carries out text classification and scene Recognition using trained deep neural network,
Obtain the second candidate newly-increased tally set, user to the first candidate newly-increased tally set and the second candidate newly-increased tally set carry out selection and
Correction, exports final program label；
The intelligent recognition strategy includes following strategy:
Tactful A: when input program category is picture, picture input recognition of face device is obtained into people tag, and will figure
Piece inputs OCR identifier and recognition result input NLP actuator is carried out entity extraction, the mark such as acquisition time, place, event
Label；
Tactful B: when input program category is voice, which is inputted into speech recognition device and recognition result is inputted into NLP
Actuator carries out entity extraction, obtains the labels such as time, place, event；
Tactful C: when input program category is video, picture frame is handled using tactful A, using tactful B to video
Voice is handled；
Tactful D: when input program category is text, text input NLP actuator is subjected to entity extraction, when acquisition
Between, place, the labels such as event；
The rule-based reasoning includes the following steps:
S4.1.1: respectively using the initial associated metadata of the extracted basic label of intelligent recognition actuator and program as
Entity；
S4.1.2: position of each entity in internal knowledge base is obtained using the methods of entity disambiguation and coreference resolution；
S4.1.3: go out adjacent node and relationship as subgraph by center Node extraction of each entity；
S4.1.4: the sub-graph data of each entity is inputted into trained GCN and is calculated using distributed diagram technology, is pushed away
Reason obtains program field；
S4.1.5: the program field that basic label and reasoning are obtained is as the first candidate newly-increased tally set；
It is described to be included the following steps: based on deep neural network progress text classification and scene Recognition
S4.2.1: trained deep neural network will be inputted by pretreated text data, and will obtain field label；
S4.2.2: trained deep neural network will be inputted by pretreated scene Recognition image data, must shown up
Scape label；
S4.2.3: using obtained field label and scene tag as the second candidate newly-increased tally set；
In the S4 after final program label, program, final program label and metadata information are stored in history tab library,
History tab library is updated, and updates internal knowledge base and internal case library simultaneously.
The above, only presently preferred embodiments of the present invention, are not intended to limit the invention, patent protection model of the invention
It encloses and is subject to claims, it is all to change with equivalent structure made by specification and accompanying drawing content of the invention, similarly
It should be included within the scope of the present invention.
Claims (10)
1. a kind of fusion reasoning system of news program intelligent label, it is characterised in that: including intelligent recognition actuator, history mark
Library, internal knowledge base, internal case library and analysis ratiocination device are signed,
Intelligent recognition actuator: for executing the identification mission of all kinds of news program materials, to video image, voice and text
Information carries out basic label extraction；
History tab library: for storing material, metadata and label, in the history tab library material include text, voice and
Video image；Metadata includes the information such as title, keyword, creation time, place, the type of material；Label includes material institute
Field, Sentiment orientation, time of origin, personage, scene for being related to etc.；
Internal knowledge base: the external knowledge map established based on history tab library, the entity of the external knowledge map includes people
Object, country, the political situation of the time, event etc., internal knowledge base is for supplementing intelligent recognition as a result, providing more letters for subsequent analysis reasoning
Breath；
Internal case library: based on history tab library establish case set, including it is representative, possess complete metadata and
The case of label is used for deep learning training process；
Analysis ratiocination device: for the fusion reasoning of intelligent label, include rule-based reasoning device and pushing away based on deep learning
Manage device；Rule-based reasoning device comprehensive utilization intelligent recognition result and internal knowledge base make inferences；Based on deep learning
Using internal case library training deep learning model, the model for recycling training to obtain makes inferences reasoning device.
2. a kind of fusion reasoning system of news program intelligent label according to claim 1, it is characterised in that: the intelligence
Can identify that actuator identifies all kinds of news program materials, identification process using recognition of face actuator, OCR actuator,
Speech recognition actuator and NLP actuator.
3. a kind of fusion reasoning method of news program intelligent label, which comprises the steps of:
S1, building news knowledge base: history tab library, history material database, internet and other construction of knowledge base news are utilized
Knowledge base；
S2, it constructs internal case library: extracting the text material with field label from history tab library and have scene mark
The picture of label forms internal case library, and by the case textual data value in internal case library；
S3, training deep learning model: the case training deep neural network by numeralization, the deep neural network are utilized
Training process include text classification training process and scene Recognition training process；
S4, it carries out fusion reasoning: according to the type of input program, carrying out content recognition using intelligent recognition actuator, extract base
Then plinth label carries out rule-based reasoning to basic label to judge program category using internal knowledge base, obtains first
The newly-increased tally set of candidate, and text classification or scene Recognition are carried out using trained deep neural network, obtain the second time
Newly-increased tally set is selected, user selects and corrects to the first candidate newly-increased tally set and the second candidate newly-increased tally set, exports
Final program label.
4. a kind of fusion reasoning method of news program intelligent label according to claim 3, which is characterized in that the S1
In the entity of news knowledge base include time, place, event, personage etc., the building of the news knowledge base includes following step
It is rapid:
S1.1, global ontology library is constructed: according to the amusement of news, sport, finance and economics, the people's livelihood, the political situation of the time, tourism, military equal part
Class carries out each field ontology library building respectively, and building range includes concept, concept hierarchy, attribute, attribute Value Types, attribute value value
Domain, relationship, contextual definition domain concept set and relation value codomain；
S1.2, it obtains entity: history tab library entity is obtained, if history tab library entity information is imperfect, in history material
Entity information supplement is carried out in library and internet；
S1.3, entity assessment: the confidence level of the entity of acquisition is quantified, the lower entity of confidence level is given up；
S1.4, knowledge fusion: being disambiguated by entity and coreference resolution method, will be arrived remaining entity link after entity is assessed
In current knowledge library, and the related entities in third party's knowledge base are merged into current knowledge library；
S1.5, knowledge reasoning: relationship, entity attributes in reasoning current knowledge library between entity and the level between ontology close
Knowledge reasoning result Jing Guo manual examination and verification is added in current knowledge library, obtains news knowledge base by system.
5. a kind of fusion reasoning method of news program intelligent label according to claim 3, which is characterized in that the S3
In the text classification training process based on deep neural network include the following steps:
S3.1.1: word segmentation processing is carried out to text and is serialized, removal obtains word sequence to meaningless word of classifying；
S3.1.2: word numbered sequence is converted by word sequence；
S3.1.3: obtained word order column number is separately converted to the term vector of n dimension；
S3.1.4: term vector is formed into text matrix by word order；
S3.1.5: it is used for text matrix to train deep neural network.
6. a kind of fusion reasoning method of news program intelligent label according to claim 3, which is characterized in that the S3
In the scene Recognition training process based on deep neural network include the following steps:
S3.2.1: expand scene Recognition image pattern by operations such as image cut, rotation, scalings；
S3.2.2: adjustment scene Recognition image size and carry out other pretreatment after, for training deep neural network.
7. a kind of fusion reasoning method of news program intelligent label according to claim 3, which is characterized in that the S4
It is middle to carry out content recognition using intelligent recognition actuator, using following strategy:
Tactful A: when input program category is picture, picture input recognition of face device is obtained into people tag, and picture is defeated
Enter OCR identifier and recognition result input NLP actuator is subjected to entity extraction, the labels such as acquisition time, place, event；
Tactful B: when input program category is voice, which is inputted into speech recognition device and executes recognition result input NLP
Device carries out entity extraction, obtains the labels such as time, place, event；
Tactful C: when input program category is video, picture frame is handled using tactful A, using tactful B to video speech
It is handled；
Tactful D: input program category be text when, by the text input NLP actuator carry out entity extraction, obtain the time,
The labels such as point, event.
8. a kind of fusion reasoning method of news program intelligent label according to claim 3, which is characterized in that the S4
Middle rule-based reasoning includes the following steps:
S4.1.1: respectively using the initial associated metadata of the extracted basic label of intelligent recognition actuator and program as real
Body；
S4.1.2: position of each entity in internal knowledge base is obtained using the methods of entity disambiguation and coreference resolution；
S4.1.3: go out adjacent node and relationship as subgraph by center Node extraction of each entity；
S4.1.4: the sub-graph data of each entity is inputted into trained GCN and is calculated using distributed diagram technology, reasoning obtains
To program field；
S4.1.5: the program field that basic label and reasoning are obtained is as the first candidate newly-increased tally set.
9. a kind of fusion reasoning method of news program intelligent label according to claim 3, which is characterized in that the S4
In text classification and scene Recognition carried out based on deep neural network include the following steps:
S4.2.1: trained deep neural network will be inputted by pretreated text data, and will obtain field label；
S4.2.2: trained deep neural network will be inputted by pretreated scene Recognition image data, and will obtain scene mark
Label；
S4.2.3: using obtained field label and scene tag as the second candidate newly-increased tally set.
10. a kind of fusion reasoning method of news program intelligent label according to claim 3, which is characterized in that described
After obtaining final program label in S4, program, final program label and metadata information are stored in history tab library, to history mark
Label library is updated, and updates internal knowledge base and internal case library simultaneously.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811528577.4A CN109635171A (en) | 2018-12-13 | 2018-12-13 | A kind of fusion reasoning system and method for news program intelligent label |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811528577.4A CN109635171A (en) | 2018-12-13 | 2018-12-13 | A kind of fusion reasoning system and method for news program intelligent label |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109635171A true CN109635171A (en) | 2019-04-16 |
Family
ID=66073756
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811528577.4A Pending CN109635171A (en) | 2018-12-13 | 2018-12-13 | A kind of fusion reasoning system and method for news program intelligent label |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109635171A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110059201A (en) * | 2019-04-19 | 2019-07-26 | 杭州联汇科技股份有限公司 | A kind of across media program feature extracting method based on deep learning |
| CN110784759A (en) * | 2019-08-12 | 2020-02-11 | 腾讯科技（深圳）有限公司 | Barrage information processing method and device, electronic equipment and storage medium |
| CN110827351A (en) * | 2020-01-09 | 2020-02-21 | 西南交通大学 | Automatic generation method of voice tag of new target for robot audio-visual collaborative learning |
| CN111598239A (en) * | 2020-07-27 | 2020-08-28 | 江苏联著实业股份有限公司 | Method and device for extracting process system of article based on graph neural network |
| CN111857551A (en) * | 2019-04-29 | 2020-10-30 | 杭州海康威视数字技术股份有限公司 | Video data aging method and device |
| WO2021120181A1 (en) * | 2019-12-20 | 2021-06-24 | 京东方科技集团股份有限公司 | Inference computing apparatus, model training apparatus, and inference computing system |
| CN110245259B (en) * | 2019-05-21 | 2021-09-21 | 北京百度网讯科技有限公司 | Video labeling method and device based on knowledge graph and computer readable medium |
| CN113473182A (en) * | 2021-09-06 | 2021-10-01 | 腾讯科技（深圳）有限公司 | Video generation method and device, computer equipment and storage medium |
-
2018
- 2018-12-13 CN CN201811528577.4A patent/CN109635171A/en active Pending
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110059201A (en) * | 2019-04-19 | 2019-07-26 | 杭州联汇科技股份有限公司 | A kind of across media program feature extracting method based on deep learning |
| CN111857551A (en) * | 2019-04-29 | 2020-10-30 | 杭州海康威视数字技术股份有限公司 | Video data aging method and device |
| CN110245259B (en) * | 2019-05-21 | 2021-09-21 | 北京百度网讯科技有限公司 | Video labeling method and device based on knowledge graph and computer readable medium |
| CN110784759A (en) * | 2019-08-12 | 2020-02-11 | 腾讯科技（深圳）有限公司 | Barrage information processing method and device, electronic equipment and storage medium |
| WO2021120181A1 (en) * | 2019-12-20 | 2021-06-24 | 京东方科技集团股份有限公司 | Inference computing apparatus, model training apparatus, and inference computing system |
| CN110827351A (en) * | 2020-01-09 | 2020-02-21 | 西南交通大学 | Automatic generation method of voice tag of new target for robot audio-visual collaborative learning |
| CN110827351B (en) * | 2020-01-09 | 2020-04-14 | 西南交通大学 | Automatic generation method of voice tag of new target for robot audio-visual collaborative learning |
| CN111598239A (en) * | 2020-07-27 | 2020-08-28 | 江苏联著实业股份有限公司 | Method and device for extracting process system of article based on graph neural network |
| CN113473182A (en) * | 2021-09-06 | 2021-10-01 | 腾讯科技（深圳）有限公司 | Video generation method and device, computer equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109635171A (en) | A kind of fusion reasoning system and method for news program intelligent label | |
| CN108595708A (en) | A kind of exception information file classification method of knowledge based collection of illustrative plates | |
| CN102246164B (en) | Information search method and information providing method based on user view | |
| CN104462064B (en) | A kind of method and system of information of mobile terminal communication prompt input content | |
| CN102567509B (en) | Method and system for instant messaging with visual messaging assistance | |
| CN104142995B (en) | The social event recognition methods of view-based access control model attribute | |
| US7689527B2 (en) | Attribute extraction using limited training data | |
| CN107729309A (en) | A kind of method and device of the Chinese semantic analysis based on deep learning | |
| CN106874397B (en) | Automatic semantic annotation method for Internet of things equipment | |
| CN109271537B (en) | Text-to-image generation method and system based on distillation learning | |
| CN107515934A (en) | A kind of film semanteme personalized labels optimization method based on big data | |
| CN110502621A (en) | Answering method, question and answer system, computer equipment and storage medium | |
| CN111026842A (en) | Natural language processing method, natural language processing device and intelligent question-answering system | |
| CN111046133A (en) | Question-answering method, question-answering equipment, storage medium and device based on atlas knowledge base | |
| CN110851599A (en) | Automatic scoring method and teaching and assisting system for Chinese composition | |
| CN106777040A (en) | A kind of across media microblogging the analysis of public opinion methods based on feeling polarities perception algorithm | |
| CN107862561A (en) | A kind of method and apparatus that user-interest library is established based on picture attribute extraction | |
| CN111223014A (en) | Method and system for online generating subdivided scene teaching courses from large amount of subdivided teaching contents | |
| Pal et al. | Putting semantic information extraction on the map: Noisy label models for fact extraction | |
| CN106202338A (en) | Image search method based on the many relations of multiple features | |
| CN112800184B (en) | Short text comment emotion analysis method based on Target-Aspect-Opinion joint extraction | |
| CN110008307B (en) | Method and device for identifying deformed entity based on rules and statistical learning | |
| CN113076476B (en) | User portrait construction method of microblog heterogeneous information | |
| CN106897274B (en) | Cross-language comment replying method | |
| CN111125387A (en) | Multimedia list generation and naming method and device, electronic equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |