WO2021031480A1 - 文本生成方法和装置 - Google Patents

文本生成方法和装置 Download PDF

Info

Publication number
WO2021031480A1
WO2021031480A1 PCT/CN2019/126797 CN2019126797W WO2021031480A1 WO 2021031480 A1 WO2021031480 A1 WO 2021031480A1 CN 2019126797 W CN2019126797 W CN 2019126797W WO 2021031480 A1 WO2021031480 A1 WO 2021031480A1
Authority
WO
WIPO (PCT)
Prior art keywords
vector
entity
attribute
text
knowledge graph
Prior art date
Application number
PCT/CN2019/126797
Other languages
English (en)
French (fr)
Inventor
吴智东
Original Assignee
广州视源电子科技股份有限公司
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 广州视源电子科技股份有限公司 filed Critical 广州视源电子科技股份有限公司
Publication of WO2021031480A1 publication Critical patent/WO2021031480A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/35—Clustering; Classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
    • G06F16/367—Ontology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Definitions

  • This application relates to the field of natural language processing, such as a text generation method and device.
  • Text generation technology is an important research direction in the field of Natural Language Processing (NLP). It aims to automatically generate sentences that conform to the laws of human language and have no grammatical errors through rules and algorithms.
  • NLP Natural Language Processing
  • the comments of the above methods are constructed manually, not generated by algorithms. Therefore, the above methods cannot generate different comments for each student in a batch, intelligent, and personalized manner.
  • this method since the comment is obtained by calculating the similarity between the student information and the comment template, this method only considers the information on the character surface, and does not consider the semantic information of the comment text.
  • the deep learning algorithm considers the statistical distribution of text in multiple dimensions and uses probability to generate comments.
  • deep learning algorithms lack knowledge information, lack the ability to learn about the potential relationship between a specific student’s daily behavior and comments, and lack the ability to generate personalized comments for specific students. Even the comments generated by the deep learning algorithm are consistent with the actual situation of the students. Performance matching is not high and imprecise.
  • This application provides a text generation method and device to at least solve the technical problem that the text information generated by only using deep learning algorithms in related technologies lacks personalized comments for students, which results in a poor match between text information and students’ actual performance .
  • the present application provides a text generation method, including: selecting a target knowledge graph of a target entity from a knowledge graph set, where the knowledge graph set is used to represent the attribute value of at least one entity on a preset attribute, and the target entity is the one to be evaluated Object; Determine the entity vector, attribute vector, and attribute value vector of the target entity based on the target knowledge graph.
  • the entity vector, attribute vector, and attribute value vector are represented by triple vectors; generate and match according to the entity vector, attribute vector and attribute value vector The text matched by the target entity.
  • the above method further includes: generating a knowledge graph set, wherein the step of generating the knowledge graph set includes: constructing a planning layer of the knowledge graph set, wherein the planning layer includes at least: entities Type, attribute type, and attribute value type; obtain record information, where the record information includes: attribute value of at least one entity on a preset attribute; input the record information into the planning layer to generate a knowledge graph set.
  • the above method further includes: preprocessing the record information to obtain the processed record information, wherein the preprocessing includes at least one of the following: entity extraction, attribute extraction, attribute value extraction, and entity extraction. Disambiguation.
  • Generating text matching the target entity based on entity vector, attribute vector, and attribute value vector includes: inputting entity vector, attribute vector, and attribute value vector into the text generation model.
  • the text generation model includes deep neural network models and deep neural networks. The network model is trained according to the triple sample and the text sample; the text matching the target entity is generated based on the text generation model.
  • the above method further includes: generating a text generation model, wherein the step of generating the text generation model includes: obtaining triple samples and text samples; using presets The algorithm converts the entity samples in the triple sample into Boolean vectors, and uses the preset model to convert the attribute samples and attribute value samples in the triple sample into high-dimensional numerical vectors to obtain the triple vector samples; The tuple vector sample and the text sample train the text generation model to obtain the trained text generation model.
  • a text generation model based on triple vector samples and text samples to obtain a trained text generation model including: using an encoder combined with attention mechanism to process triple vector samples and text samples to obtain context vectors; using combined attention
  • the decoder of the mechanism processes the context vector to obtain text information; based on the text information, the text generation model is trained to minimize the loss function.
  • the present application also provides a text generation method, including: receiving a selection instruction, wherein the selection instruction is used to select a target entity to be evaluated; and displaying text matching the target entity, wherein the text is determined based on the target knowledge graph of the target entity
  • the entity vector, attribute vector and attribute value vector of the target entity are generated.
  • the target knowledge map comes from the knowledge map set.
  • the knowledge map set is used to represent the attribute value of at least one entity on the preset attribute.
  • the entity vector, attribute vector and attribute value vector It is represented by a triple vector.
  • the present application also provides a text generation device, including: a selection module for selecting a target knowledge graph of a target entity from a knowledge graph set, wherein the knowledge graph set is used to represent the attribute value of at least one entity on a preset attribute, The target entity is the object to be evaluated; the determination module is used to determine the entity vector, attribute vector, and attribute value vector of the target entity based on the target knowledge graph, where the entity vector, attribute vector and attribute value vector are represented by triple vectors; text The generating module is used to generate text matching the target entity based on the entity vector, attribute vector and attribute value vector.
  • the present application also provides a storage medium, the storage medium includes a stored program, wherein the device where the storage medium is located is controlled to execute any of the above-mentioned text generation methods when the program is running.
  • the present application also provides a processor, which is used to run a program, where any one of the above text generation methods is executed when the program is running.
  • the target knowledge graph of the target entity is selected from the knowledge graph set, where the knowledge graph set is used to represent the attribute value of at least one entity on the preset attribute, and the target entity is the object to be evaluated; it is determined based on the target knowledge graph
  • the entity vector, attribute vector, and attribute value vector of the target entity are represented by triplet vectors; the text matching the target entity is generated according to the entity vector, attribute vector and attribute value vector.
  • this application uses the usual performance of multiple entities to build a knowledge graph set, and then extracts the triple vector of the target knowledge graph from it, and then combines the deep learning algorithm to generate comments.
  • This solution combines the knowledge graph and deep learning to connect the deep learning algorithm to all attributes of the entity, thereby solving the lack of personalized comments on the entity in the text information generated by the deep learning algorithm in related technologies, resulting in text information and
  • the technical problem that the actual performance of the entity is not highly matched has achieved the goal of generating comments that meet the usual performance of the entity to the greatest extent, and improved the matching of comments.
  • Fig. 1 is a flowchart of a text generation method according to Embodiment 1 of the present application
  • FIG. 2 is a basic principle block diagram of a comment generation method according to Embodiment 1 of the present application;
  • Figure 3 is a detailed schematic diagram based on the basic principles of the comment generation method shown in Figure 2;
  • FIG. 5 is a schematic structural diagram of a text generation device according to Embodiment 3 of the present application.
  • Fig. 6 is a schematic structural diagram of a text generation device according to Embodiment 4 of the present application.
  • an embodiment of a text generation method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and, although The logical sequence is shown in the flowchart, but in some cases, the steps shown or described may be performed in a different order than here.
  • Fig. 1 is a text generation method according to an embodiment of the present application. As shown in Fig. 1, the method includes the following steps:
  • Step S102 Select the target knowledge graph of the target entity from the knowledge graph set, where the knowledge graph set is used to represent the attribute value of at least one entity on the preset attribute, and the target entity is the object to be evaluated.
  • the above entity can be any object that needs evaluation, such as students, institutions, and company employees; for students, the above preset attributes can be classroom performance, self-image, social performance, emotional performance, weekly test results , Final grades, etc., the corresponding attribute values can be positive, tidy, active, stable, large fluctuations, excellent, etc.; for institutions, the above preset attributes can be brand image, number of authorized patents, annual profit, social welfare, etc., corresponding The attribute value of can be large impact, greater than 100 items, 200 million, active, etc.
  • Knowledge Graph as a new knowledge organization and retrieval technology in the era of big data, is used to describe concepts and their relationships in the physical world in the form of symbols.
  • the knowledge graph set brings together the knowledge graphs of multiple entities.
  • the knowledge graph of each entity records the daily behavior of the entity. Since each entity is an independent individual, the knowledge graph of each entity is naturally different. When it is necessary to evaluate a certain entity, that is, the target entity, select the target knowledge graph of the target entity from the knowledge graph set.
  • the knowledge graph of student A is extracted from the knowledge graph set.
  • the knowledge graph records the attribute values of student A on all attributes, that is, records all aspects of student A Daily behavior performance.
  • Step S104 Determine the entity vector, attribute vector, and attribute value vector of the target entity based on the target knowledge graph, where the entity vector, attribute vector and attribute value vector are represented by a triple vector.
  • the generated text can be greatly improved The matching degree.
  • triplet is a general representation form of the knowledge graph, and this embodiment takes the triplet as an example, which does not constitute a limitation to the application.
  • Step S106 Generate text matching the target entity according to the entity vector, attribute vector and attribute value vector.
  • the text generation model for generating text may be a deep neural network model.
  • Deep neural network is a comprehensive subject about the combination of mathematics and computer. Unlike machine learning, deep neural network can realize end-to-end data high-dimensional feature extraction and abstraction, and solve the problem of feature extraction in machine learning. For example, typical Seq2Seq model, generative adversarial network model, etc.
  • Seq2Seq is an Encoder-Deocder structure model.
  • the basic idea is to use two recurrent neural networks, one as an encoder and one as a decoder.
  • the encoder turns a variable-length input sequence into a fixed-length vector. This vector can Regarding the semantics of the sequence, the decoder decodes this fixed-length vector into a variable-length output sequence;
  • the generative adversar network (Generative Adversaria l, GAN) model includes at least two modules, one is a generative model, and one Confrontation model, the mutual game learning of two models produces quite good output. Therefore, the application of the above two deep neural network algorithms in the field of comment generation can achieve more accurate and robust effects than machine learning methods.
  • the triple vector consisting of entity vector, attribute vector and attribute value vector determined by the target knowledge graph is input into the deep neural network model to generate comments that match the daily behavior of the target entity text.
  • this application combines the knowledge graph and the deep neural network, takes into account the daily behavior of the target entity, and for different entities, it can automatically generate comments that match the actual performance of the entity, which improves the matching and accuracy of the comments.
  • Teachers need to write a summary comment for each student during the winter and summer vacations. Teachers can extract the knowledge graph of the student to be evaluated from the knowledge graph set by clicking the mouse.
  • the knowledge graph records the student's daily performance, such as classroom performance, self-image, social performance, emotional performance, final grades and other information.
  • the computer terminal executing the method of this embodiment determines the student's triple vector based on the student's knowledge graph, and inputs it into the deep neural network model.
  • the display interface of the computer terminal automatically generates comments matching the student's daily performance.
  • a target knowledge graph of a target entity is selected from a knowledge graph set, where the knowledge graph set is used to represent the attribute value of at least one entity on a preset attribute, and the target entity is an object to be evaluated; Determine the entity vector, attribute vector and attribute value vector of the target entity based on the target knowledge graph.
  • the entity vector, attribute vector and attribute value vector are represented by triple vector; generate and target entity based on entity vector, attribute vector and attribute value vector The matched text.
  • this application uses the usual performance of multiple entities to build a knowledge graph set, and then extracts the triple vector of the target knowledge graph from it, and then combines the deep learning algorithm to generate comments.
  • This solution combines the knowledge graph and deep learning to connect the deep learning algorithm to all attributes of the entity, thereby solving the lack of personalized comments on the entity in the text information generated by the deep learning algorithm in related technologies, resulting in text information and
  • the technical problem that the actual performance of the entity is not highly matched has achieved the goal of generating comments that meet the usual performance of the entity to the greatest extent, and improved the matching of comments.
  • the above method may further include step S101, generating a knowledge graph set, wherein the step of generating the knowledge graph set may specifically include the following steps:
  • Step S1012 Construct a planning layer of the knowledge graph set, where the planning layer includes at least entity type, attribute type, and attribute value type.
  • the above-mentioned planning layer can be edited by the ontology construction tool Protégé software.
  • Protégé software is an ontology editing and knowledge acquisition software developed based on the Java language. Users only need to construct an ontology model at the conceptual level, which is simple and easy to operate.
  • the planning layer is equivalent to the structure of the knowledge graph.
  • the planning layer includes at least entity types, attribute types, and attribute value types. Of course, it can also include information such as time.
  • Step S1014 Obtain record information, where the record information includes: an attribute value of at least one entity on a preset attribute.
  • the aforementioned record information may be manually input into the computer terminal that executes the method of this embodiment.
  • Li Ming showed positive class performance, good image, final grade A, etc.
  • Zhang Wei showed dozing off in class, not active social performance, final grade B, etc.
  • the daily behavior of the target entity can be fully considered to avoid missing features.
  • Step S1016 Input the record information into the planning layer to generate a knowledge graph set.
  • the entity information, attribute information, and attribute value information obtained in step S1014 are correspondingly filled into the entity type, attribute type, and attribute value type of the planning layer constructed in step S1012 to construct a knowledge graph set of all entities. And stored in the graph database Neo4j.
  • the above method may further include: step S1015, preprocessing the record information to obtain processed record information, wherein the preprocessing includes at least one of the following: Entity extraction, attribute extraction, attribute value extraction and entity disambiguation.
  • the aforementioned entity extraction, attribute extraction, and attribute value extraction may be entity recognition, attribute recognition, and attribute value recognition, including detection and classification of entities, attributes, and attribute values.
  • step S104 determines the entity vector, attribute vector, and attribute value vector of the target entity based on the target knowledge graph, which may specifically include the following steps:
  • Step S1042 Extract entity information, attribute information and attribute value information of the target entity in the target knowledge graph.
  • Step S1044 Use a preset algorithm to convert the entity information into a Boolean vector, and use a preset model to convert both the attribute information and the attribute value information into a high-dimensional numeric vector to obtain a triplet vector.
  • the foregoing preset algorithm may be a OneHot algorithm
  • the foregoing preset model may be a BERT model or a Word2Vector model.
  • the BERT model represented by the two-way encoder of Transformer, is suitable for the construction of the most advanced model for a wide range of tasks.
  • the entity information, attribute information and attribute value information are converted into numerical vectors that are easy to be processed by the neural network model.
  • the neural network model is connected to all the attributes of the target entity and can then be extracted High-latitude attribute vector features.
  • multiple triples (e i , p ij , v ij ) of the target entity in the target knowledge graph are extracted, where e i , p ij and v ij represent the information of the i-th entity and the information of the i-th entity, respectively
  • the j-th attribute information, the j-th attribute value information of the i-th entity, and then e i , p ij , and v ij are respectively represented as V ei , V pi , and V vi vectors.
  • the OneHot algorithm is used to represent the entity e i as a Boolean vector
  • the BERT model is used to represent the attribute p ij and the attribute value v ij as a high-latitude numerical vector, namely
  • t and s represent feature extraction functions, which are also mapping functions of a neural network structure.
  • step S106 generates text matching the target entity according to the entity vector, the attribute vector, and the attribute value vector, which may specifically include the following steps:
  • Step S1062 input the entity vector, the attribute vector and the attribute value vector into the text generation model, where the text generation model includes a deep neural network model, and the deep neural network model is obtained by training based on the triple sample and the text sample.
  • the aforementioned deep neural network model may be a Seq2Seq model, a generative confrontation network model, and so on.
  • Step S1064 Generate text matching the target entity based on the text generation model.
  • the entity vector V ei , the attribute vector V pi and the attribute value vector V vi are input into the text generation model to generate a summary comment text y * about the target entity.
  • the above-described Concluding Comments Text can be expressed as y * output sequence y 1, ... y T ', where y t' indicates the output character time t ', i.e.
  • arg max represents the text with the largest probability vector value among the candidate texts to be selected.
  • the above method may further include step S1061, generating a text generation model, wherein the step of generating the text generation model may include:
  • Step S10611 Obtain triad samples and text samples.
  • the above-mentioned triple sample and text sample can form an aligned corpus, expressed as ⁇ ((e,p,v),y)
  • Step S10612 Use a preset algorithm to convert the entity samples in the triple sample into a Boolean vector, and use the preset model to convert the attribute samples and attribute value samples in the triple sample into high-latitude numeric vectors to obtain a triple Set of vector samples.
  • the aforementioned preset algorithm can also be a one-hot algorithm, and the aforementioned preset model can also be a two-way encoder representation model.
  • the process of converting triplet samples into triplet vector samples is similar to step S1044, here No longer.
  • step S10613 the text generation model is trained based on the triple vector sample and the text sample to obtain a trained text generation model.
  • the text generation model collects the daily behavior performance data of all entities, and uses it as training corpus to train the text generation model, the above scheme can generate summary comments that conform to the entity’s daily behavior performance based on the specific entity’s daily behavior performance.
  • step S10613 trains the text generation model based on the triple vector sample and the text sample to obtain a trained text generation model, which may specifically include the following steps:
  • Step S106131 using an encoder combined with an attention mechanism to process the triple vector samples and text samples to obtain a context vector.
  • the encoder turns a variable-length input sequence into a fixed-length vector, which can be regarded as the sequence
  • the decoder decodes this fixed-length vector into a variable-length output sequence.
  • the context vector encoded by the encoder combined with the attention mechanism is:
  • f represents the coding function
  • h t , y t′-1 , s t′-1 , c t′ respectively represent the hidden layer output of the encoder t, the output of the decoder t′-1, and the decoder t
  • Step S106132 using a decoder combined with the attention mechanism to process the context vector to obtain text information.
  • the decoder output combined with the attention mechanism is:
  • g represents the decoding function
  • y t′ , y t′-1 , s t′ , and c t′ represent the output at time t′, the output at time t′-1, and the hidden layer state of the decoder at time t′, respectively ,
  • the context vector at time t' represents the decoding function
  • Step S106133 based on the text information, train a text generation model to minimize the loss function.
  • the goal of training the text generation model is to minimize the negative log-likelihood loss function of the text generation model:
  • x i and y i represent the i-th input text and output text respectively, i ⁇ 1,...,I ⁇ , and ⁇ is the model parameter.
  • the result of training is that the generated text is strongly correlated with the original text, and text grammatical errors are minimized.
  • the preset algorithm in step S1044 and step S10612 is a one-hot algorithm
  • the preset model is a BERT model or a Word2Vector model.
  • FIG. 2 is a basic principle block diagram of a comment generation method according to an embodiment of the present application.
  • first collect the teacher's record of each student's daily behavior data and then fill it into the designed knowledge map planning layer to construct the knowledge map set of all students.
  • the target knowledge graphs of the students to be evaluated are extracted from the knowledge graph set, and then input into the trained text generation model, and then the summary comments on the daily performance of the students are automatically output.
  • the detailed principle is shown in Figure 3.
  • the daily behavior data of students includes classroom performance, self-image, social performance, emotional performance, etc.
  • the planning layer of the knowledge graph includes entity types, attribute types, and attribute value types, which are in the construction of the knowledge graph set
  • the students’ daily behavior data is preprocessed by entity extraction, attribute extraction, attribute value extraction, entity disambiguation and other operations, and then filled into the corresponding planning layer.
  • entity extraction, attribute extraction, attribute value extraction, entity disambiguation and other operations are preprocessed by entity extraction, attribute extraction, attribute value extraction, entity disambiguation and other operations, and then filled into the corresponding planning layer.
  • the student ID When evaluating the student ID, first extract the knowledge subgraph of the student ID, then extract the triple information, convert it into a triple vector for representation, and finally input it into the trained text generation model to generate candidates Student comments, the teacher reconfirms whether the comment needs to be modified to get the final student comment.
  • the text generation model is trained by the Encoder-Deocder model combined with the attention mechanism of triple samples and comment samples.
  • the above embodiment of the present application selects the target knowledge graph of the target entity from the knowledge graph set, where the knowledge graph set is used to characterize the attribute value of at least one entity on the preset attribute, and the target entity is the object to be evaluated; Determine the entity vector, attribute vector and attribute value vector of the target entity based on the target knowledge graph.
  • the entity vector, attribute vector and attribute value vector are represented by triple vector; generate and target entity based on entity vector, attribute vector and attribute value vector The matched text.
  • this application uses the usual performance of multiple entities to establish a knowledge graph set, and then extracts the triple vector of the target knowledge graph from it, and then combines the deep learning algorithm to generate comments; by combining entity information, attribute information and attribute values
  • the information is converted into a numerical vector that is easy to be processed by the neural network model.
  • the neural network model is connected to all the attributes of the target entity, and then the high-latitude attribute vector features can be extracted; the Encoder-Deocder model combined with the attention mechanism can optimize the text output effect;
  • the text information generated by deep learning algorithms only lacks a personalized comment on the entity, which leads to the technical problem that the text information does not match the actual performance of the entity. It achieves the greatest possible generation of comments in line with the entity’s usual performance. Objective, to improve the matching degree of comments.
  • Fig. 4 is another text generation method according to an embodiment of the present application. As shown in Fig. 4, the method includes the following steps:
  • Step S402 Receive a selection instruction, where the selection instruction is used to select a target entity to be evaluated.
  • the above selection instruction can be triggered by a teacher through a mouse click or touch on a touch screen; in an optional solution, the above target entity can be any object to be evaluated, such as students, institutions, and company employees.
  • Step S404 Display the text matching the target entity, where the text is generated based on the entity vector, attribute vector and attribute value vector of the target entity determined by the target knowledge map of the target entity.
  • the target knowledge map comes from the knowledge map set, and the knowledge map set is used To represent the attribute value of at least one entity on the preset attribute, the entity vector, the attribute vector and the attribute value vector are represented by a triple vector.
  • the above-mentioned entity can be any object that needs evaluation, such as students, institutions, and company employees; for students, the above-mentioned preset attributes can be classroom performance, self-image, and social Performance, emotional performance, weekly test scores, final scores, etc., the corresponding attribute values can be positive, clean, active, stable, undulating, excellent, etc.; for institutions, the above-mentioned preset attributes can be brand image, number of authorized patents, For annual profit, social welfare, etc., the corresponding attribute value can be large impact, greater than 100 items, 200 million, active, etc.; the text generation model for the above text generation can be a deep neural network model.
  • Knowledge Graph as a new knowledge organization and retrieval technology in the era of big data, is used to describe concepts and their relationships in the physical world in the form of symbols.
  • the knowledge graph set brings together the knowledge graphs of multiple entities.
  • the knowledge graph of each entity records the daily behavior of the entity. Since each entity is an independent individual, the knowledge graph of each entity is naturally different. When it is necessary to evaluate a certain entity, that is, the target entity, select the target knowledge graph of the target entity from the knowledge graph set.
  • the matching degree of the generated text can be greatly improved.
  • deep neural network is a comprehensive subject about the combination of mathematics and computer. Unlike machine learning, deep neural network can achieve end-to-end data high-dimensional feature extraction and abstraction, and solve the problem of feature extraction in machine learning. problem. For example, typical Seq2Seq model, generative adversarial network model, etc.
  • Seq2Seq is an Encoder-Deocder structure model.
  • the basic idea is to use two recurrent neural networks, one as an encoder and one as a decoder.
  • the encoder turns a variable-length input sequence into a fixed-length vector.
  • This vector can Considering the semantics of the sequence, the decoder decodes this fixed-length vector into a variable-length output sequence;
  • the Generative Adversarial Networks (GAN) model includes at least two modules, a generative model and an adversarial model.
  • the mutual game learning of the two models produces quite good output. Therefore, the above two deep neural network algorithms are applied in the field of comment generation and can achieve more accurate and robust effects than machine learning methods.
  • the computer terminal After the computer terminal detects the selection instruction of clicking the target entity from the display interface, it will display the comment text matching the target entity on the display interface.
  • this application combines the knowledge graph and the deep neural network, takes into account the daily behavior of the target entity, and for different entities, it can automatically generate comments that match the actual performance of the entity, which improves the matching and accuracy of the comments.
  • a selection instruction is first received, wherein the selection instruction is used to select the target entity to be evaluated, and then display the text matching the target entity, where the text is determined based on the target knowledge graph of the target entity
  • the entity vector, attribute vector, and attribute value vector of the target entity are generated.
  • the target knowledge map comes from the knowledge map set.
  • the knowledge map set is used to represent the attribute value of at least one entity on the preset attribute.
  • the entity vector, attribute vector and attribute value vector are used Three-tuple vector representation. Compared with related technologies, this application uses the usual performance of multiple entities to build a knowledge graph set, and then extracts the triple vector of the target knowledge graph from it, and then combines the deep learning algorithm to generate comments.
  • This solution combines the knowledge graph and deep learning to connect the deep learning algorithm to all attributes of the entity, thereby solving the lack of personalized comments on the entity in the text information generated by the deep learning algorithm in related technologies, resulting in text information and
  • the technical problem that the actual performance of the entity is not highly matched can achieve the purpose of generating comments that meet the usual performance of the entity to the greatest extent, and improve the matching of comments.
  • the above method may further include step S403 of generating a knowledge graph set, wherein the step of generating the knowledge graph set may specifically include the following steps:
  • Step S4032 Construct a planning layer of the knowledge graph set, where the planning layer at least includes: entity type, attribute type, and attribute value type.
  • the above-mentioned planning layer can be edited by the ontology construction tool Protégé software.
  • Protégé software is an ontology editing and knowledge acquisition software developed based on the Java language. Users only need to construct an ontology model at the conceptual level, which is simple and easy to operate.
  • the planning layer is equivalent to the structure of the knowledge graph.
  • the planning layer includes at least entity types, attribute types, and attribute value types. Of course, it can also include information such as time.
  • Step S4034 Obtain record information, where the record information includes: attribute value of at least one entity on a preset attribute.
  • the aforementioned record information may be manually input into the computer terminal that executes the method of this embodiment.
  • Li Ming showed positive class performance, good image, final grade A, etc.
  • Zhang Wei showed dozing off in class, not active social performance, final grade B, etc.
  • the daily behavior of the target entity can be fully considered to avoid missing features.
  • Step S4036 Input the record information into the planning layer to generate a knowledge graph set.
  • the entity information, attribute information, and attribute value information are correspondingly filled into the entity type, attribute type, and attribute value type of the constructed planning layer to construct a knowledge graph set of all entities and store it in the graph database Neo4j in.
  • the above method may further include: step S4035, preprocessing the recording information to obtain processed recording information, wherein the preprocessing includes at least one of the following: Entity extraction, attribute extraction, attribute value extraction and entity disambiguation.
  • the aforementioned entity extraction, attribute extraction, and attribute value extraction may be entity recognition, attribute recognition, and attribute value recognition, including detection and classification of entities, attributes, and attribute values.
  • the entity vector, attribute vector, and attribute value vector of the target entity determined by the target knowledge graph in step S404 may specifically include the following steps:
  • Step S4041 Extract entity information, attribute information and attribute value information of the target entity in the target knowledge graph.
  • Step S4042 Use a preset algorithm to convert the entity information into a Boolean vector, and use a preset model to convert both the attribute information and the attribute value information into a high-dimensional numerical vector to obtain a triplet vector.
  • the foregoing preset algorithm may be a one-hot algorithm
  • the foregoing preset model may be a BERT model or a Word2Vector model.
  • the BERT model represented by the two-way encoder of Transformer, is suitable for the construction of the most advanced model for a wide range of tasks.
  • the entity information, attribute information and attribute value information are converted into numerical vectors that are easy to be processed by the neural network model.
  • the neural network model is connected to all the attributes of the target entity and can then be extracted High-latitude attribute vector features.
  • multiple triples (e i , p ij , v ij ) of the target entity in the target knowledge graph are extracted, where e i , p ij and v ij represent the information of the i-th entity and the information of the i-th entity, respectively
  • the j-th attribute information, the j-th attribute value information of the i-th entity, and then e i , p ij , and v ij are respectively represented as V ei , V pi , and V vi vectors.
  • the OneHot algorithm is used to represent the entity e i as a Boolean vector
  • the BERT model is used to represent the attribute p ij and the attribute value v ij as a high-latitude numerical vector, namely
  • t and s represent feature extraction functions, which are also mapping functions of a neural network structure.
  • step S404 may specifically include the following steps:
  • Step S4046 Input the entity vector, the attribute vector and the attribute value vector into the text generation model, where the text generation model includes a deep neural network model, which is obtained by training based on the triple sample and the text sample.
  • the aforementioned deep neural network model may be a Seq2Seq model, a generative adversarial network model, and so on.
  • Step S4047 Generate text matching the target entity based on the text generation model.
  • the entity vector V ei , the attribute vector V pi and the attribute value vector V vi are input into the text generation model to generate a summary comment text y * about the target entity.
  • the above-described Concluding Comments Text can be expressed as y * output sequence y 1, ... y T ', where y t' indicates the output character time t ', i.e.
  • arg max represents the text with the largest probability vector value among the candidate texts to be selected.
  • the above method may further include step S4045, generating a text generation model, wherein the step of generating the text generation model may include:
  • Step S40451 Obtain triplet samples and text samples.
  • the above-mentioned triple sample and text sample can form an aligned corpus, expressed as ⁇ ((e,p,v),y)
  • Step S40452 Use a preset algorithm to convert the entity samples in the triplet sample into a Boolean vector, and use the preset model to convert the attribute samples and attribute value samples in the triplet sample into high-latitude numeric vectors to obtain a triplet Set of vector samples.
  • the aforementioned preset algorithm can also be a one-hot algorithm, and the aforementioned preset model can also be a two-way encoder representation model.
  • the process of converting triplet samples into triplet vector samples is similar to step S1044, here No longer.
  • step S40453 the text generation model is trained based on the triple vector sample and the text sample to obtain a trained text generation model.
  • the text generation model collects the daily behavior performance data of all entities, and uses it as training corpus to train the text generation model, the above scheme can generate summary comments that conform to the entity’s daily behavior performance based on the specific entity’s daily behavior performance.
  • step S40453 trains the text generation model based on the triple vector sample and the text sample to obtain a trained text generation model, which may specifically include the following steps:
  • Step S404531 using an encoder combined with an attention mechanism to process the triple vector samples and text samples to obtain a context vector.
  • the encoder turns a variable-length input sequence into a fixed-length vector, which can be regarded as the sequence
  • the decoder decodes this fixed-length vector into a variable-length output sequence.
  • the context vector encoded by the encoder combined with the attention mechanism is:
  • f represents the coding function
  • h t , y t′-1 , s t′-1 , c t′ respectively represent the hidden layer output of the encoder t, the output of the decoder t′-1, and the decoder t
  • Step S404532 using a decoder combined with an attention mechanism to process the context vector to obtain text information.
  • the decoder output combined with the attention mechanism is:
  • g represents the decoding function
  • y t′ , y t′-1 , s t′ , and c t′ represent the output at time t′, the output at time t′-1, and the hidden layer state of the decoder at time t′, respectively ,
  • the context vector at time t' represents the decoding function
  • step S404533 based on the text information, a text generation model is trained to minimize the loss function.
  • the goal of training the text generation model is to minimize the negative log-likelihood loss function of the text generation model:
  • x i and y i represent the i-th input text and output text respectively, i ⁇ 1,...,I ⁇ , and ⁇ is the model parameter.
  • the result of training is that the generated text is strongly correlated with the original text, and text grammatical errors are minimized.
  • the preset algorithm in step S4042 and step S40452 is a one-hot algorithm
  • the preset model is a BERT model or a Word2Vector model.
  • the method according to the above embodiment can be implemented by means of software plus the necessary general hardware platform, of course, it can also be implemented by hardware, but in many cases the former is Better implementation.
  • the technical solution of this application essentially or the part that contributes to the related technology can be embodied in the form of a software product, the computer software product is stored in a storage medium (such as ROM/RAM, magnetic disk, optical disk) )
  • a storage medium such as ROM/RAM, magnetic disk, optical disk
  • a terminal device which can be a mobile phone, a computer, a server, or a network device, etc.
  • FIG. 5 is a schematic diagram of the text generation device according to an embodiment of the present application.
  • the device 500 includes a selection module 502, a determination module 504, and a text generation module 506.
  • the selection module 502 is used to select the target knowledge graph of the target entity from the knowledge graph set, where the knowledge graph set is used to represent the attribute value of at least one entity on the preset attribute, and the target entity is the object to be evaluated; the determining module 504, used to determine the entity vector, attribute vector, and attribute value vector of the target entity based on the target knowledge graph, wherein the entity vector, attribute vector, and attribute value vector are represented by triple vectors; the text generation module 506 is used to determine the entity vector , Attribute vector and attribute value vector to generate text matching the target entity.
  • the above-mentioned device may further include: a graph generation module, configured to generate a knowledge graph set before selecting the target knowledge graph of the target entity from the knowledge graph set, wherein the graph generation module includes: a building module for constructing a knowledge graph A set of planning layers, where the planning layer includes at least entity type, attribute type, and attribute value type; a first acquisition module for acquiring record information, where record information includes: attribute value of at least one entity on a preset attribute; Input the record information into the planning layer, and the graph generation sub-module is used to generate the knowledge graph set.
  • a graph generation module configured to generate a knowledge graph set before selecting the target knowledge graph of the target entity from the knowledge graph set
  • the graph generation module includes: a building module for constructing a knowledge graph A set of planning layers, where the planning layer includes at least entity type, attribute type, and attribute value type
  • a first acquisition module for acquiring record information, where record information includes: attribute value of at least one entity on a preset attribute
  • the above device may further include: a preprocessing module, configured to preprocess the record information before inputting the record information to the planning layer to obtain processed record information, wherein the preprocessing includes at least one of the following: Entity extraction, attribute extraction, attribute value extraction and entity disambiguation.
  • a preprocessing module configured to preprocess the record information before inputting the record information to the planning layer to obtain processed record information, wherein the preprocessing includes at least one of the following: Entity extraction, attribute extraction, attribute value extraction and entity disambiguation.
  • the determining module includes: an extraction module for extracting entity information, attribute information, and attribute value information of the target entity in the target knowledge graph; a first conversion module for converting the entity information into a Boolean vector using a preset algorithm , Use the preset model to convert both the attribute information and the attribute value information into high-dimensional numerical vectors to obtain triplet vectors.
  • the text generation module includes: an input module for inputting entity vectors, attribute vectors, and attribute value vectors into the text generation model, where the text generation model includes a deep neural network model, and the deep neural network model is based on triples Samples and text samples are trained; the text generation sub-module is used to generate text matching the target entity based on the text generation model.
  • the above-mentioned device may further include: a model generation module for generating a text generation model before the entity vector, attribute vector, and attribute value vector are input to the text generation model, wherein the model generation module includes: a second acquisition module , Used to obtain the triple sample and text sample; the second conversion module, used to use the preset algorithm to convert the entity sample in the triple sample into a Boolean vector, and use the preset model to convert the attribute sample in the triple sample , The attribute value samples are all converted into high-dimensional numerical vectors to obtain triple vector samples; the training module is used to train the text generation model based on the triple vector samples and text samples to obtain a trained text generation model.
  • the model generation module includes: a second acquisition module , Used to obtain the triple sample and text sample; the second conversion module, used to use the preset algorithm to convert the entity sample in the triple sample into a Boolean vector, and use the preset model to convert the attribute sample in the triple sample ,
  • the attribute value samples are all converted into high-dimensional numerical vector
  • the training module includes: an encoding module, which is used to process triple vector samples and text samples using an encoder combined with an attention mechanism, to obtain a context vector; a decoding module, used to process context using a decoder combined with an attention mechanism The vector is used to obtain text information; the training sub-module is used to train the text generation model based on the text information to minimize the loss function.
  • the foregoing preset algorithm is a one-hot algorithm
  • the preset model is a BERT model or a Word2Vector model.
  • selection module 502, determination module 504, and text generation module 506 correspond to steps S102 to S106 in Embodiment 1.
  • the three modules are the same as the examples and application scenarios implemented by the corresponding steps, but not It is limited to the content disclosed in Example 1 above.
  • FIG. 6 is a schematic diagram of the apparatus for generating text according to an embodiment of the present application.
  • the device 600 includes a receiving module 602 and a display module 604.
  • the receiving module 602 is used to receive a selection instruction, where the selection instruction is used to select the target entity to be evaluated; the display module 604 is used to display text matching the target entity, where the text is determined according to the target knowledge graph of the target entity Generate the entity vector, attribute vector and attribute value vector of the target entity.
  • the target knowledge map comes from the knowledge map set.
  • the knowledge map set is used to represent the attribute value of at least one entity on the preset attribute, the entity vector, attribute vector and attribute value
  • the vector is represented by a triple vector.
  • the above device may further include a graph generation module for generating a knowledge graph set before displaying the text matching the target entity, wherein the graph generation module may include: a building module for constructing a planning layer of the knowledge graph set , Where the planning layer includes at least: entity type, attribute type, and attribute value type; a first acquisition module for acquiring record information, where the record information includes: attribute value of at least one entity on a preset attribute; graph generation sub-module , Used to input record information into the planning layer to generate a knowledge graph set.
  • the above device may further include a preprocessing module, which is used to preprocess the record information before inputting the record information to the planning layer to obtain processed record information, wherein the preprocessing includes at least one of the following: entity Extraction, attribute extraction, attribute value extraction and entity disambiguation.
  • a preprocessing module which is used to preprocess the record information before inputting the record information to the planning layer to obtain processed record information, wherein the preprocessing includes at least one of the following: entity Extraction, attribute extraction, attribute value extraction and entity disambiguation.
  • the display module further includes a determining module for determining the entity vector, attribute vector, and attribute value vector of the target entity according to the target knowledge graph, wherein the determining module may include: an extraction module for extracting information in the target knowledge graph The entity information, attribute information, and attribute value information of the target entity; the first conversion module is used to convert the entity information into a Boolean vector using a preset algorithm, and use the preset model to convert both the attribute information and the attribute value information into high-dimensional numerical values Vector, get the triple vector.
  • the determining module may include: an extraction module for extracting information in the target knowledge graph The entity information, attribute information, and attribute value information of the target entity; the first conversion module is used to convert the entity information into a Boolean vector using a preset algorithm, and use the preset model to convert both the attribute information and the attribute value information into high-dimensional numerical values Vector, get the triple vector.
  • the display module also includes a text generation module for generating text based on the entity vector, attribute vector, and attribute value vector.
  • the text generation module may include: an input module for combining the entity vector, attribute vector, and attribute value.
  • the vector is input to the text generation model, where the text generation model includes a deep neural network model, which is trained based on triple samples and text samples; the text generation sub-module is used to generate matching target entities based on the text generation model text.
  • the above device may further include a model generation module for generating a text generation model before the entity vector, attribute vector, and attribute value vector are input into the text generation model
  • the model generation module may include: a second acquisition module , Used to obtain the triple sample and text sample; the second conversion module, used to use the preset algorithm to convert the entity sample in the triple sample into a Boolean vector, and use the preset model to convert the attribute sample in the triple sample ,
  • the attribute value samples are converted into high-dimensional numerical vectors to obtain triple vector samples; the training module is used to train the text generation model based on the triple vector samples and text samples to obtain a trained text generation model.
  • the training module may include: an encoding module for processing triple vector samples and text samples with an encoder combined with an attention mechanism to obtain context vectors; a decoding module for processing with a decoder combined with an attention mechanism The context vector obtains text information; the training sub-module is used to train the text generation model based on the text information to minimize the loss function.
  • the foregoing preset algorithm is a one-hot algorithm
  • the preset model is a BERT model or a Word2Vector model.
  • the above-mentioned receiving module 602 and the display module 604 correspond to steps S402 to S404 in Embodiment 2.
  • the examples and application scenarios implemented by these two modules are the same as the corresponding steps, but are not limited to the above-mentioned Embodiment 2. What is disclosed.
  • a storage medium includes a stored program, wherein the device where the storage medium is located is controlled to execute the text generation method in Embodiment 1 or 2 when the program is running.
  • a processor is provided, and the processor is configured to run a program, wherein the text generation method in Embodiment 1 or 2 is executed when the program is running.
  • the disclosed technical content can be implemented in other ways.
  • the device embodiments described above are merely illustrative.
  • the division of the units may be a logical function division, and there may be other divisions in actual implementation, for example, multiple units or components may be combined or may be Integrate into another system, or some features can be ignored or not implemented.
  • the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of units or modules, and may be in electrical or other forms.
  • 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 units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.
  • each unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist alone physically, or two or more units may be integrated into one unit.
  • the above-mentioned integrated unit can be implemented in the form of hardware or software functional unit.
  • the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium.
  • the technical solution of this application essentially or the part that contributes to the related technology or all or 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,
  • a number of instructions are included to enable a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application.
  • the aforementioned storage media include: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program code .

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Evolutionary Computation (AREA)
  • Databases & Information Systems (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Animal Behavior & Ethology (AREA)
  • Machine Translation (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种文本生成方法和装置。其中，该方法包括：从知识图谱集中选择目标实体的目标知识图谱（S102）；基于目标知识图谱确定目标实体的实体向量、属性向量和属性值向量（S104）；依据实体向量、属性向量和属性值向量生成与目标实体匹配的文本（S106）。

Description

文本生成方法和装置
本公开要求在2019年08月21日提交中国专利局、申请号为201910775353.1的中国专利申请的优先权，以上申请的全部内容通过引用结合在本公开中。
技术领域
本申请涉及自然语言处理领域，例如涉及一种文本生成方法和装置。
背景技术
文本生成技术，是自然语言处理(Natural Language Processing，NLP)领域的一个重要的研究方向，旨在通过规则、算法等自动生成符合人类语言规律、没有语法错误的句子。
文本生成技术的应用场合非常多。例如，在教育行业，每到学期结束时，教师需要根据学生的日常表现，写出一段关于学生表现的描述性、建议性的评语。传统生成每个学生的评语的方法大多依靠教师手动撰写，这样的方式不仅消耗教师大量的时间，而且，教师也不一定能准确记得所有学生的日常表现。因此，相关比较成熟的解决方式是，将输入的学生信息，与人工构造的评语模板进行相似度计算，选取相似度最高的模板作为生成的评语。
然而，上述方法的评语是人工构建出来的，而非通过算法生成的，因此，上述方法不能批量化、智能化、个性化地对每个学生生成不同的评语。另外，由于评语是通过计算学生信息与评语模板之间的相似度来获得的，这种方式仅考虑了字符表面的信息，没有考虑到评语文本的语义层信息。为解决这一问题，深度学习算法考虑了文本在多个维度上的统计分布，并使用概率的方式来生成评语。但是，深度学习算法缺乏知识性信息，对特定学生的日常行为表现与评语之间潜在关系的学习能力不足，缺乏对特定学生生成个性化评语的能力，以至深度学习算法生成的评语与学生的实际表现匹配度不高、不精确。
针对相关技术中仅利用深度学习算法生成的文本信息缺乏对实体的个性化评述，导致文本信息与实体的实际表现匹配度不高的技术问题，目前尚未提出有效的解决方案。
发明内容
本申请提供了一种文本生成方法和装置，以至少解决相关技术中仅利用深度学习算法生成的文本信息缺乏对学生的个性化评述，导致文本信息与学生的实际表现匹配度不高的技术问题。
本申请提供了一种文本生成方法，包括：从知识图谱集中选择目标实体的目标知识图谱，其中，知识图谱集用于表征至少一个实体在预设属性上的属性值，目标实体为待评价的对象；基于目标知识图谱确定目标实体的实体向量、属性向量和属性值向量，其中，实体向量、属性向量和属性值向量用三元组向量表征；依据实体向量、属性向量和属性值向量生成与目标实体匹配的文本。
在从知识图谱集中选择目标实体的目标知识图谱之前，上述方法还包括：生成知识图谱集，其中，生成知识图谱集的步骤包括：构建知识图谱集的规划层，其中，规划层至少包括：实体类型、属性类型和属性值类型；获取记录信息，其中记录信息包括：至少一个实体在预设属性上的属性值；将记录信息输入至规划层中，生成知识图谱集。
在将记录信息输入至规划层之前，上述方法还包括：对记录信息进行预处理，得到处理后的记录信息，其中，预处理包括如下至少之一：实体抽取、属性抽取、属性值抽取和实体消歧。
基于目标知识图谱确定目标实体的实体向量、属性向量和属性值向量，包括：提取目标知识图谱中的目标实体的实体信息、属性信息和属性值信息；使用预设算法将实体信息转换为布尔向量，使用预设模型将属性信息和属性值信息均转换为高纬数值型向量，得到三元组向量。
依据实体向量、属性向量和属性值向量生成与目标实体匹配的文本，包括：将实体向量、属性向量和属性值向量输入至文本生成模型中，其中，文本生成模型包括深度神经网络模型，深度神经网络模型根据三元组样本和文本样本训练得到；基于文本生成模型生成与目标实体匹配文本。
在将实体向量、属性向量和属性值向量输入至文本生成模型之前，上述方法还包括：生成文本生成模型，其中，生成文本生成模型的步骤包括：获取三元组样本和文本样本；使用预设算法将三元组样本中的实体样本转换为布尔向量，使用预设模型将三元组样本中的属性样本、属性值样本均转换为高纬数值型向量，得到三元组向量样本；基于三元组向量样本和文本样本训练文本生成模型，得到训练好的文本生成模型。
基于三元组向量样本和文本样本训练文本生成模型，得到训练好的文本生成模型，包括：利用结合注意力机制的编码器处理三元组向量样本和文本样本，得到上下文向量；利用结合注意力机制的解码器处理上下文向量，得到文本信息；基于文本信息，以最小化损失函数来训练文本生成模型。
本申请还提供了一种文本生成方法，包括：接收选中指令，其中，选中指令用于选中待评价的目标实体；显示与目标实体匹配的文本，其中，文本依据目标实体的目标知识图谱确定出的目标实体的实体向量、属性向量和属性值向量生成，目标知识图谱来自知识图谱集，知识图谱集用于表征至少一个实体在预设属性上的属性值，实体向量、属性向量和属性值向量用三元组向量表征。
本申请还提供了一种文本生成装置，包括：选择模块，用于从知识图谱集中选择目标实体的目标知识图谱，其中，知识图谱集用于表征至少一个实体在预设属性上的属性值，目标实体为待评价的对象；确定模块，用于基于目标知识图谱确定目标实体的实体向量、属性向量和属性值向量，其中，实体向量、属性向量和属性值向量用三元组向量表征；文本生成模块，用于依据实体向量、属性向量和属性值向量生成与目标实体匹配的文本。
本申请还提供了一种存储介质，存储介质包括存储的程序，其中，在程序运行时控制存储介质所在设备执行上述任意一种文本生成方法。
本申请还提供了一种处理器，处理器用于运行程序，其中，程序运行时执行上述任意一种文本生成方法。
在本申请中，从知识图谱集中选择目标实体的目标知识图谱，其中，知识图谱集用于表征至少一个实体在预设属性上的属性值，目标实体为待评价的对象；基于目标知识图谱确定目标实体的实体向量、属性向量和属性值向量，其中，实体向量、属性向量和属性值向量用三元组向量表征；依据实体向量、属性向量和属性值向量生成与目标实体匹配的文本。与相关技术相比，本申请利用多个实体的平时表现建立知识图谱集，然后从中提取目标知识图谱的三元组向量，继而结合深度学习算法生成评语。该方案通过将知识图谱和深度学习相结合，使得深度学习算法连接到实体的所有属性，进而解决了相关技术中仅利用深度学习算法生成的文本信息缺乏对实体的个性化评述，导致文本信息与实体的实际表现匹配度不高的技术问题，达到了最大程度地生成符合实体平时表现的评语的目的，提高了评语的匹配度。
附图说明
此处所说明的附图用来提供对本申请的进一步理解，构成本申请的一部分，本申请的示意性实施例及其说明用于解释本申请，并不构成对本申请的不当限定。在附图中：
图1是根据本申请实施例1的一种文本生成方法的流程图；
图2是根据本申请实施例1的一种评语生成方法的基本原理框图；
图3是基于图2所示评语生成方法基本原理的详细原理图；
图4是根据本申请实施例2的一种文本生成方法的流程图；
图5是根据本申请实施例3的一种文本生成装置的结构示意图；
图6是根据本申请实施例4的一种文本生成装置的结构示意图。
具体实施方式
为了使本技术领域的人员更好地理解本申请方案，下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本申请一部分的实施例，而不是全部的实施例。需要说明的是，本申请的说明书和权利要求书及上述附图中的术语“第一”、“第二”等是用于区别类似的对象，而不必用于描述特定的顺序或先后次序。应该理解这样使用的数据在适当情况下可以互换，以便这里描述的本申请的实施例能够以除了在这里图示或描述的那些以外的顺序实施。此外，术语“包括”和“具有”以及他们的任何变形，意图在于覆盖不排他的包含，例如，包含了一系列步骤或单元的过程、方法、系统、产品或设备不必限于清楚地列出的那些步骤或单元，而是可包括没有清楚地列出的或对于这些过程、方法、产品或设备固有的其它步骤或单元。
实施例1
根据本申请实施例，提供了一种文本生成方法的实施例，需要说明的是，在附图的流程图示出的步骤可以在诸如一组计算机可执行指令的计算机系统中执行，并且，虽然在流程图中示出了逻辑顺序，但是在某些情况下，可以以不同于此处的顺序执行所示出或描述的步骤。
图1是根据本申请实施例的文本生成方法，如图1所示，该方法包括如下步骤：
步骤S102，从知识图谱集中选择目标实体的目标知识图谱，其中，知识图谱集用于表征至少一个实体在预设属性上的属性值，目标实体为待评价的对象。
一种可选方案中，上述实体可以为学生、机构、公司员工等任何需要评价的对象；对于学生来讲，上述预设属性可以为课堂表现、自我形象、社交表现、情绪表现、周考成绩、期末成绩等，对应的属性值可以为积极、整洁、活跃、稳定、起伏较大、优良等；针对机构，上述预设属性可以为品牌形象、授权专利数量、年利润、社会公益等，对应的属性值可以为影响大、大于100件、2亿、活跃等。
知识图谱(Knowledge Graph，KG)作为大数据时代下新的知识组织与检索技术，用于以符号形式描述物理世界中的概念及其相互关系。知识图谱集汇集了多个实体的知识图谱，每个实体的知识图谱记录着该实体的日常行为表现，由于每个实体都是一个独立的个体，每个实体的知识图谱自然不一样。需要评价某一个实体，即目标实体时，从知识图谱集中选择目标实体的目标知识图谱即可。
以学生为例，在需要生成学生A的总结性评语时，从知识图谱集中抽取学生A的知识图谱，该知识图谱记载着学生A在所有属性上的属性值，即记录着学生A各个方面的日常行为表现。
步骤S104，基于目标知识图谱确定目标实体的实体向量、属性向量和属性值向量，其中，实体向量、属性向量和属性值向量用三元组向量表征。
上述步骤中，通过在目标知识图谱中提取目标实体的实体信息、属性信息和属性值信息，并将其转换为易于文本生成模型处理的实体向量、属性向量和属性值向量，可以大大提高生成文本的匹配度。
需要说明的是，三元组是知识图谱的一种通用表示形式，本实施例以三元组进行举例，并不构成对本申请的限制。
步骤S106，依据实体向量、属性向量和属性值向量生成与目标实体匹配的文本。
一种可选方案中，上述生成文本的文本生成模型可以为深度神经网络模型。
深度神经网络是一门关于数学和计算机结合的综合学科，与机器学习不同，深度神经网络能实现端对端的数据高纬特征提取、抽象，解决了机器学习中特征难以提取的问题。例如典型的Seq2Seq模型、生成对抗网络模型等。
Seq2Seq是一个Encoder-Deocder结构的模型，基本思想是利用两个循环神经网络，一个作为编码器，一个作为解码器，编码器将一个可变长度的输入序列变为固定长度的向量，这个向量可以看作该序列的语义，解码器将这个固定长度的向量解码成可变长度的输出序列；生成对抗网络(Generat ive Adversar ia l Networks，GAN)模型中至少包括两个模块，一个生成模型，一个对抗模型，两个模型的互相博弈学习产生相当好的输出。因此，上述两种深度神经网络算法应用在评语生成领域，能够达到比机器学习方法更加精确和鲁棒的效果。
上述步骤中，将通过目标知识图谱确定出的由实体向量、属性向量和属性值向量构成的三元组向量，输入到深度神经网络模型中，可以生成与目标实体的日常行为表现相匹配的评语文本。
容易注意到，在相关的文本生成领域，即使有基于知识图谱的文本生成，也不是完全使用知识图谱本省的实体信息、属性信息和属性值等信息，而是将知识图谱作为中介，再通过搜索，或计算相似性的方法查找合适的文本。然而，本申请将知识图谱和深度神经网络相结合，考虑了目标实体的日常行为表现，且对不同实体，能够自动生成符合该实体实际表现情况的评语，提高了评语的匹配度和准确度。
仍以学生为例，教师在寒暑假之际需要针对每位学生写一段总结性的评语。教师可以通过点击鼠标从知识图谱集中抽取待评价学生的知识图谱，该知识图谱记载着该学生的日常表现，例如课堂表现、自我形象、社交表现、情绪表现、期末成绩等信息。执行本实施例方法的计算机终端基于该学生的知识图谱确定出该学生的三元组向量，输入至深度神经网络模型中，计算机终端的显示界面会自动生成与该学生的日常表现匹配的评语。采用上述方案，大大节约了教师的时间和精力，避免了教师记不准或记不全学生的日常行为表现，导致评语与学生的匹配度不高的问题出现。
基于本申请上述实施例提供的方案，从知识图谱集中选择目标实体的目标知识图谱，其中，知识图谱集用于表征至少一个实体在预设属性上的属性值，目标实体为待评价的对象；基于目标知识图谱确定目标实体的实体向量、属性向量和属性值向量，其中，实体向量、属性向量和属性值向量用三元组向量表征；依据实体向量、属性向量和属性值向量生成与目标实体匹配的文本。与相关技术相比，本申请利用多个实体的平时表现建立知识图谱集，然后从中提取目标知识图谱的三元组向量，继而结合深度学习算法生成评语。该方案通过将知识图谱和深度学习相结合，使得深度学习算法连接到实体的所有属性，进而解决了相关技术中仅利用深度学习算法生成的文本信息缺乏对实体的个性化评述，导致文本信息与实体的实际表现匹配度不高的技术问题，达到了最大程度地生成符合实体平时表现的评语的目的，提高了评语的匹配度。
可选地，在执行步骤S102从知识图谱集中选择目标实体的目标知识图谱之前，上述方法还可以包括步骤S101，生成知识图谱集，其中，生成知识图谱集的步骤具体可以包括以下步骤：
步骤S1012，构建知识图谱集的规划层，其中，规划层至少包括：实体类型、属性类型和属性值类型。
一种可选方案中，上述规划层可以通过本体构建工具Protégé软件编辑。Protégé软件是基于Java语言开发的本体编辑和知识获取软件，用户只需要在概念层次上进行本体模型的构建即可，简单易操作。
规划层相当于知识图谱的架构，规划层中至少包括实体类型、属性类型和属性值类型，当然，也可以包括时间等信息。
步骤S1014，获取记录信息，其中记录信息包括：至少一个实体在预设属性上的属性值。
一种可选方案中，上述记录信息可以通过人工输入至执行本实施例方法的计算机终端中。例如，李明课堂表现积极、形象佳、期末成绩A等，张伟课堂表现爱打瞌睡、社交表现不积极、期末成绩B等。如此，在生成目标实体的文本时，可以全面考虑目标实体的日常行为表现，避免遗漏特征。
步骤S1016，将记录信息输入至规划层中，生成知识图谱集。
上述步骤中，将步骤S1014中获取的实体信息、属性信息、属性值信息对应填充到步骤S1012构建的 规划层的实体类型、属性类型和属性值类型中，以此构造所有实体的知识图谱集，并存储到图形数据库Neo4j中。
可选地，在执行步骤S1016将记录信息输入至规划层之前，上述方法还可以包括：步骤S1015，对记录信息进行预处理，得到处理后的记录信息，其中，预处理包括如下至少之一：实体抽取、属性抽取、属性值抽取和实体消歧。
一种可选方案中，上述实体抽取、属性抽取、属性值抽取可以为实体识别、属性识别、属性值识别，包括实体、属性、属性值的检测和分类。
需要说明的是，通过实体消歧处理，可以区分出两个不同的名字代表同一个实体，或一个相同的名字指代两个不同的实体的情况。
可选地，步骤S104基于目标知识图谱确定目标实体的实体向量、属性向量和属性值向量，具体可以包括以下步骤：
步骤S1042，提取目标知识图谱中的目标实体的实体信息、属性信息和属性值信息。
步骤S1044，使用预设算法将实体信息转换为布尔向量，使用预设模型将属性信息和属性值信息均转换为高纬数值型向量，得到三元组向量。
一种可选方案中，上述预设算法可以为独热(OneHot)算法，上述预设模型可以为BERT模型或Word2Vector模型。其中，BERT模型，用Transformer的双向编码器表示，适用于广泛任务的最先进模型的构建。
在对目标知识图谱中的三元组进行信息表示时，将实体信息、属性信息和属性值信息转换为易于神经网络模型处理的数值向量，神经网络模型连接到目标实体的所有属性，继而可以提取高纬属性向量特征。具体地，提取目标知识图谱中目标实体的多个三元组(e i，p ij，v ij)，其中，e i、p ij、v ij分别表示第i个实体信息、第i个实体的第j个属性信息、第i个实体的第j个属性值信息，然后将e i，p ij，v ij分别表征成V ei，V pi，V vi向量。
在一个可选的实施例中，使用OneHot算法将实体e i表征成布尔向量，使用BERT模型将属性p ij、属性值v ij表征成高纬数值型向量，即
Figure PCTCN2019126797-appb-000001
Figure PCTCN2019126797-appb-000002
Figure PCTCN2019126797-appb-000003
其中，t、s表示特征提取函数，也是一个神经网络结构的映射函数。
可选地，步骤S106依据实体向量、属性向量和属性值向量生成与目标实体匹配的文本，具体可以包括以下步骤：
步骤S1062，将实体向量、属性向量和属性值向量输入至文本生成模型中，其中，文本生成模型包括深度神经网络模型，深度神经网络模型根据三元组样本和文本样本训练得到。
如前所述，上述深度神经网络模型可以是Seq2Seq模型、生成对抗网络模型等。
步骤S1064，基于文本生成模型生成与目标实体匹配文本。
上述步骤中，将实体向量V ei、属性向量V pi和属性值向量V vi输入至文本生成模型中，即可生成关于目标实体的总结性评语文本y *。
一种可选方案中，上述总结性评语文本y *可表示为输出序列y 1,…y T’，其中y t’表示t’时刻的输出文 字，即
Figure PCTCN2019126797-appb-000004
本的概率向量，arg max表示选取生成的候选文本中，概率向量值最大的文本。
可选地，在执行步骤S1062将实体向量、属性向量和属性值向量输入至文本生成模型之前，上述方法还可以包括步骤S1061，生成文本生成模型，其中，生成文本生成模型的步骤可以包括：
步骤S10611，获取三元组样本和文本样本。
一种可选方案中，上述三元组样本和文本样本可以组成对齐语料，表示为{((e,p,v),y)|((e 1,p 1,v 1),y 1),…((e i,p i,v i),y i)}。
步骤S10612，使用预设算法将三元组样本中的实体样本转换为布尔向量，使用预设模型将三元组样本中的属性样本、属性值样本均转换为高纬数值型向量，得到三元组向量样本。
如前所述，上述预设算法也可以为独热算法，上述预设模型也可以为双向编码器表征模型，将三元组样本转换为三元组向量样本的过程和步骤S1044类似，在此不再赘述。
步骤S10613，基于三元组向量样本和文本样本训练文本生成模型，得到训练好的文本生成模型。
在构造好由三元组和评语的组成的对齐语料后，就可以基于构造的语料，使用深度神经网络的算法，训练文本生成模型。由于文本生成模型采集了所有实体的日常行为表现数据，并以此作为训练语料，训练文本生成模型，因此，上述方案能根据具体的实体的日常行为表现生成符合该实体的总结性评语。
在一个可选的实施例中，步骤S10613基于三元组向量样本和文本样本训练文本生成模型，得到训练好的文本生成模型，具体可以包括以下步骤：
步骤S106131，利用结合注意力机制的编码器处理三元组向量样本和文本样本，得到上下文向量。
在Encoder-Deocder结构的模型中，存在两个循环神经网络，一个作为编码器，一个作为解码器，编码器将一个可变长度的输入序列变为固定长度的向量，这个向量可以看做该序列的语义，解码器将这个固定长度的向量解码成可变长度的输出序列。然而，如果输入序列的长度很长，固定长度的向量效果未免不佳，而结合注意力(Attention)机制的编码器可以解决此效果不佳的问题。具体的，结合注意力机制的编码器编码的上下文向量为：
c t'＝f(h t，y t'-1，s t'-1，c t')
其中，f表示编码函数，h t、y t′-1、s t′-1、c t′分别表示编码器t时刻的隐含层输出、解码器t′-1时刻的输出、解码器t′-1时刻的隐含层状态、t′时刻的上下文向量。
步骤S106132，利用结合注意力机制的解码器处理上下文向量，得到文本信息。
考虑到编码器提取的最终上下文向量中，特征信息有限，且较难捕获到输入的局部特征，故需要结合编码器中注意力机制的输出结果，作为解码器的输入参数。具体地，结合注意力机制的解码器输出为：
P(y t'|y 1,...,y t'-1,c t')＝g(y t'-1,s t',c t')
其中，g表示解码函数，y t′、y t′-1、s t′、c t′分别表示t′时刻的输出、t′-1时刻的输出、解码器t′时刻的隐含层状态、t′时刻的上下文向量。
步骤S106133，基于文本信息，以最小化损失函数来训练文本生成模型。
需要说明的是，对文本生成模型训练的目标为：最小化文本生成模型的负对数似然损失函数：
Figure PCTCN2019126797-appb-000005
其中，x i、y i分别代表第i个输入文本、输出文本，i∈{1,…,I}，θ是模型参数。
训练的结果是生成的文本与原始文本强相关，且最小化文本语法错误。
可选地，步骤S1044和步骤S10612中的预设算法为独热算法，预设模型为BERT模型或Word2Vector模型。
仍以学生为例，图2是根据本申请实施例的一种评语生成方法的基本原理框图。如图2所示，首先采集教师对每个学生的日常行为数据的记录情况，然后将其填充到设计好的知识图谱规划层中，以此构造所有学生表现的知识图谱集。在需要生成待评价学生的评语时，从知识图谱集中抽取待评价学生的目标知识图谱，然后将其输入到训练好的文本生成模型中，进而自动输出关于学生日常表现的总结性评语。详细的原理如图3所示，学生的日常行为数据包括课堂表现、自我形象、社交表现、情绪表现等，知识图谱的规划层规划有实体类型、属性类型和属性值类型，在构造知识图谱集时，将学生的日常行为数据进过实体抽取、属性抽取、属性值抽取、实体消歧等操作进行预处理，然后填充到对应的规划层中即可。在评价学生ID时，首先抽取出学生ID的知识子图，然后提取出三元组信息，将其转换为三元组向量的形式进行表征，最后输入到训练好的文本生成模型中，生成候选学生评语，由教师再次确认是否需要对该评语进行修改，以得到最终的学生评语。其中，文本生成模型由三元组样本和评语样本对结合注意力机制的Encoder-Deocder模型训练得到。
由上可知，本申请上述实施例，从知识图谱集中选择目标实体的目标知识图谱，其中，知识图谱集用于表征至少一个实体在预设属性上的属性值，目标实体为待评价的对象；基于目标知识图谱确定目标实体的实体向量、属性向量和属性值向量，其中，实体向量、属性向量和属性值向量用三元组向量表征；依据实体向量、属性向量和属性值向量生成与目标实体匹配的文本。与相关技术相比，本申请利用多个实体的平时表现建立知识图谱集，然后从中提取目标知识图谱的三元组向量，继而结合深度学习算法生成评语；通过将实体信息、属性信息和属性值信息转换为易于神经网络模型处理的数值向量，神经网络模型连接到目标实体的所有属性，继而可以提取高纬属性向量特征；通过结合注意机制的Encoder-Deocder模型可以优化文本输出的效果；进而解决了相关技术中仅利用深度学习算法生成的文本信息缺乏对实体的个性化评述，导致文本信息与实体的实际表现匹配度不高的技术问题，达到了最大程度地生成符合实体平时表现的评语的目的，提高了评语的匹配度。
实施例2
根据本申请实施例，从显示界面的角度提供了另一种文本生成方法的实施例，需要说明的是，在附图的流程图示出的步骤可以在诸如一组计算机可执行指令的计算机系统中执行，并且，虽然在流程图中示出了逻辑顺序，但是在某些情况下，可以以不同于此处的顺序执行所示出或描述的步骤。
图4是根据本申请实施例的另一种文本生成的方法，如图4所示，该方法包括如下步骤：
步骤S402，接收选中指令，其中，选中指令用于选中待评价的目标实体。
一种可选方案中，上述选择指令可以由教师通过鼠标点击触发，也可以通过触摸屏触摸触发；一种可选方案中，上述目标实体可以为学生、机构、公司员工等任何待评价的对象。
步骤S404，显示与目标实体匹配的文本，其中，文本依据目标实体的目标知识图谱确定出的目标实体的实体向量、属性向量和属性值向量生成，目标知识图谱来自知识图谱集，知识图谱集用于表征至少一个实体在预设属性上的属性值，实体向量、属性向量和属性值向量用三元组向量表征。
一种可选方案中，上述一种可选方案中，上述实体可以为学生、机构、公司员工等任何需要评价的对象；对于学生来讲，上述预设属性可以为课堂表现、自我形象、社交表现、情绪表现、周考成绩、期末成绩等，对应的属性值可以为积极、整洁、活跃、稳定、起伏较大、优良等；针对机构，上述预设属性可以 为品牌形象、授权专利数量、年利润、社会公益等，对应的属性值可以为影响大、大于100件、2亿、活跃等；上述生成文本的文本生成模型可以为深度神经网络模型。
知识图谱(Knowledge Graph，KG)作为大数据时代下新的知识组织与检索技术，用于以符号形式描述物理世界中的概念及其相互关系。知识图谱集汇集了多个实体的知识图谱，每个实体的知识图谱记录着该实体的日常行为表现，由于每个实体都是一个独立的个体，每个实体的知识图谱自然不一样。需要评价某一个实体，即目标实体时，从知识图谱集中选择目标实体的目标知识图谱即可。
通过在目标知识图谱中提取目标实体的实体信息、属性信息和属性值信息，并将其转换为易于文本生成模型处理的实体向量、属性向量和属性值向量，可以大大提高生成文本的匹配度。
需要说明的是，深度神经网络是一门关于数学和计算机结合的综合学科，与机器学习不同，深度神经网络能实现端对端的数据高纬特征提取、抽象，解决了机器学习中特征难以提取的问题。例如典型的Seq2Seq模型、生成对抗网络模型等。
Seq2Seq是一个Encoder-Deocder结构的模型，基本思想是利用两个循环神经网络，一个作为编码器，一个作为解码器，编码器将一个可变长度的输入序列变为固定长度的向量，这个向量可以看做该序列的语义，解码器将这个固定长度的向量解码成可变长度的输出序列；生成对抗网络(Generative Adversarial Networks，GAN)模型中至少包括两个模块，一个生成模型，一个对抗模型，两个模型的互相博弈学习产生相当好的输出。因此，上述两种深度神经网络算法应用在评语生成领域，能够达到比机器学习方法更加精确和鲁棒的效果。
上述步骤中，计算机终端检测到来自显示界面的点击目标实体的选中指令后，就会在显示界面显示与目标实体匹配的评语文本。
容易注意到，在相关的文本生成领域，即使有基于知识图谱的文本生成，也不是完全使用知识图谱本省的实体信息、属性信息和属性值等信息，而是将知识图谱作为中介，再通过搜索，或计算相似性的方法查找合适的文本。然而，本申请将知识图谱和深度神经网络相结合，考虑了目标实体的日常行为表现，且对不同实体，能够自动生成符合该实体实际表现情况的评语，提高了评语的匹配度和准确度。
基于本申请上述实施例提供的方案，首先接收选中指令，其中，选中指令用于选中待评价的目标实体，然后显示与目标实体匹配的文本，其中，文本依据目标实体的目标知识图谱确定出的目标实体的实体向量、属性向量和属性值向量生成，目标知识图谱来自知识图谱集，知识图谱集用于表征至少一个实体在预设属性上的属性值，实体向量、属性向量和属性值向量用三元组向量表征。与相关技术相比，本申请利用多个实体的平时表现建立知识图谱集，然后从中提取目标知识图谱的三元组向量，继而结合深度学习算法生成评语。该方案通过将知识图谱和深度学习相结合，使得深度学习算法连接到实体的所有属性，进而解决了相关技术中仅利用深度学习算法生成的文本信息缺乏对实体的个性化评述，导致文本信息与实体的实际表现匹配度不高的技术问题，达到最大程度地生成符合实体平时表现的评语的目的，提高了评语的匹配度。
可选地，在执行步骤S404显示与目标实体匹配的文本之前，上述方法还可以包括步骤S403，生成知识图谱集，其中，生成知识图谱集的步骤具体可以包括以下步骤：
步骤S4032，构建知识图谱集的规划层，其中，规划层至少包括：实体类型、属性类型和属性值类型。
一种可选方案中，上述规划层可以通过本体构建工具Protégé软件编辑。Protégé软件是基于Java语言开发的本体编辑和知识获取软件，用户只需要在概念层次上进行本体模型的构建即可，简单易操作。
规划层相当于知识图谱的架构，规划层中至少包括实体类型、属性类型和属性值类型，当然，也可以包括时间等信息。
步骤S4034，获取记录信息，其中记录信息包括：至少一个实体在预设属性上的属性值。
一种可选方案中，上述记录信息可以通过人工输入至执行本实施例方法的计算机终端中。例如，李明课堂表现积极、形象佳、期末成绩A等，张伟课堂表现爱打瞌睡、社交表现不积极、期末成绩B等。如此，在生成目标实体的文本时，可以全面考虑目标实体的日常行为表现，避免遗漏特征。
步骤S4036，将记录信息输入至规划层中，生成知识图谱集。
上述步骤中，将实体信息、属性信息、属性值信息对应填充到构建好的规划层的实体类型、属性类型和属性值类型中，以此构造所有实体的知识图谱集，并存储到图形数据库Neo4j中。
可选地，在执行步骤S4036将记录信息输入至规划层之前，上述方法还可以包括：步骤S4035，对记录信息进行预处理，得到处理后的记录信息，其中，预处理包括如下至少之一：实体抽取、属性抽取、属性值抽取和实体消歧。
一种可选方案中，上述实体抽取、属性抽取、属性值抽取可以为实体识别、属性识别、属性值识别，包括实体、属性、属性值的检测和分类。
需要说明的是，通过实体消歧处理，可以区分出两个不同的名字代表同一个实体，或一个相同的名字指代两个不同的实体的情况。
可选地，步骤S404中目标知识图谱确定出的目标实体的实体向量、属性向量和属性值向量，具体可以包括以下步骤：
步骤S4041，提取目标知识图谱中的目标实体的实体信息、属性信息和属性值信息。
步骤S4042，使用预设算法将实体信息转换为布尔向量，使用预设模型将属性信息和属性值信息均转换为高纬数值型向量，得到三元组向量。
一种可选方案中，上述预设算法可以为独热算法，上述预设模型可以为BERT模型或Word2Vector模型。其中，BERT模型，用Transformer的双向编码器表示，适用于广泛任务的最先进模型的构建。
在对目标知识图谱中的三元组进行信息表示时，将实体信息、属性信息和属性值信息转换为易于神经网络模型处理的数值向量，神经网络模型连接到目标实体的所有属性，继而可以提取高纬属性向量特征。具体地，提取目标知识图谱中目标实体的多个三元组(e i，p ij，v ij)，其中，e i、p ij、v ij分别表示第i个实体信息、第i个实体的第j个属性信息、第i个实体的第j个属性值信息，然后将e i，p ij，v ij分别表征成V ei，V pi，V vi向量。
在一个可选的实施例中，使用OneHot算法将实体e i表征成布尔向量，使用BERT模型将属性p ij、属性值v ij表征成高纬数值型向量，即
Figure PCTCN2019126797-appb-000006
Figure PCTCN2019126797-appb-000007
Figure PCTCN2019126797-appb-000008
其中，t、s表示特征提取函数，也是一个神经网络结构的映射函数。
可选地，步骤S404中依据实体向量、属性向量和属性值向量生成文本的步骤具体可以包括以下步骤：
步骤S4046，将实体向量、属性向量和属性值向量输入至文本生成模型中，其中，文本生成模型包括深度神经网络模型，深度神经网络模型根据三元组样本和文本样本训练得到。
如前所述，上述深度神经网络模型可以是Seq2Seq模型、生成对抗网络模型等。
步骤S4047，基于文本生成模型生成与目标实体匹配文本。
上述步骤中，将实体向量V ei、属性向量V pi和属性值向量V vi输入至文本生成模型中，即可生成关于目标实体的总结性评语文本y *。
一种可选方案中，上述总结性评语文本y *可表示为输出序列y 1,…y T’，其中y t’表示t’时刻的输出文 字，即
Figure PCTCN2019126797-appb-000009
本的概率向量，arg max表示选取生成的候选文本中，概率向量值最大的文本。
可选地，在执行步骤S4046将实体向量、属性向量和属性值向量输入至文本生成模型之前，上述方法还可以包括步骤S4045，生成文本生成模型，其中，生成文本生成模型的步骤可以包括：
步骤S40451，获取三元组样本和文本样本。
一种可选方案中，上述三元组样本和文本样本可以组成对齐语料，表示为{((e,p,v),y)|((e 1,p 1,v 1),y 1),…((e i,p i,v i),y i)}。
步骤S40452，使用预设算法将三元组样本中的实体样本转换为布尔向量，使用预设模型将三元组样本中的属性样本、属性值样本均转换为高纬数值型向量，得到三元组向量样本。
如前所述，上述预设算法也可以为独热算法，上述预设模型也可以为双向编码器表征模型，将三元组样本转换为三元组向量样本的过程和步骤S1044类似，在此不再赘述。
步骤S40453，基于三元组向量样本和文本样本训练文本生成模型，得到训练好的文本生成模型。
在构造好由三元组和评语的组成的对齐语料后，就可以基于构造的语料，使用深度神经网络的算法，训练文本生成模型。由于文本生成模型采集了所有实体的日常行为表现数据，并以此作为训练语料，训练文本生成模型，因此，上述方案能根据具体的实体的日常行为表现生成符合该实体的总结性评语。
在一个可选的实施例中，步骤S40453基于三元组向量样本和文本样本训练文本生成模型，得到训练好的文本生成模型，具体可以包括以下步骤：
步骤S404531，利用结合注意力机制的编码器处理三元组向量样本和文本样本，得到上下文向量。
在Encoder-Deocder结构的模型中，存在两个循环神经网络，一个作为编码器，一个作为解码器，编码器将一个可变长度的输入序列变为固定长度的向量，这个向量可以看做该序列的语义，解码器将这个固定长度的向量解码成可变长度的输出序列。然而，如果输入序列的长度很长，固定长度的向量效果未免不佳，而结合注意力(Attention)机制的编码器可以解决此效果不佳的问题。具体的，结合注意力机制的编码器编码的上下文向量为：
c t'＝f(h t，y t'-1，s t'-1，c t')
其中，f表示编码函数，h t、y t′-1、s t′-1、c t′分别表示编码器t时刻的隐含层输出、解码器t′-1时刻的输出、解码器t′-1时刻的隐含层状态、t′时刻的上下文向量。
步骤S404532，利用结合注意力机制的解码器处理上下文向量，得到文本信息。
考虑到编码器提取的最终上下文向量中，特征信息有限，且较难捕获到输入的局部特征，故需要结合编码器中注意力机制的输出结果，作为解码器的输入参数。具体地，结合注意力机制的解码器输出为：
P(y t'|y 1,...,y t'-1,c t')＝g(y t'-1,s t',c t')
其中，g表示解码函数，y t′、y t′-1、s t′、c t′分别表示t′时刻的输出、t′-1时刻的输出、解码器t′时刻的隐含层状态、t′时刻的上下文向量。
步骤S404533，基于文本信息，以最小化损失函数来训练文本生成模型。
需要说明的是，对文本生成模型训练的目标为：最小化文本生成模型的负对数似然损失函数：
Figure PCTCN2019126797-appb-000010
其中，x i、y i分别代表第i个输入文本、输出文本，i∈{1,…,I}，θ是模型参数。
训练的结果是生成的文本与原始文本强相关，且最小化文本语法错误。
可选地，步骤S4042和步骤S40452中的预设算法为独热算法，预设模型为BERT模型或Word2Vector模型。
通过以上的实施方式的描述，本领域的技术人员可以清楚地了解到根据上述实施例的方法可借助软件加必需的通用硬件平台的方式来实现，当然也可以通过硬件，但很多情况下前者是更佳的实施方式。基于这样的理解，本申请的技术方案本质上或者说对相关技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质(如ROM/RAM、磁碟、光盘)中，包括若干指令用以使得一台终端设备(可以是手机，计算机，服务器，或者网络设备等)执行本申请各个实施例所述的方法。
实施例3
根据本申请实施例，提供了一种文本生成的装置，图5是根据本申请实施例的文本生成装置的示意图。如图5所示，该装置500包括选择模块502、确定模块504和文本生成模块506。
其中，选择模块502，用于从知识图谱集中选择目标实体的目标知识图谱，其中，知识图谱集用于表征至少一个实体在预设属性上的属性值，目标实体为待评价的对象；确定模块504，用于基于目标知识图谱确定目标实体的实体向量、属性向量和属性值向量，其中，实体向量、属性向量和属性值向量用三元组向量表征；文本生成模块506，用于依据实体向量、属性向量和属性值向量生成与目标实体匹配的文本。
可选地，上述装置还可以包括：图谱生成模块，用于在从知识图谱集中选择目标实体的目标知识图谱之前，生成知识图谱集，其中，图谱生成模块包括：构建模块，用于构建知识图谱集的规划层，其中，规划层至少包括：实体类型、属性类型和属性值类型；第一获取模块，用于获取记录信息，其中记录信息包括：至少一个实体在预设属性上的属性值；将记录信息输入至规划层中，图谱生成子模块，用于生成知识图谱集。
可选地，上述装置还可以包括：预处理模块，用于在将记录信息输入至规划层之前，对记录信息进行预处理，得到处理后的记录信息，其中，预处理包括如下至少之一：实体抽取、属性抽取、属性值抽取和实体消歧。
可选地，确定模块包括：提取模块，用于提取目标知识图谱中的目标实体的实体信息、属性信息和属性值信息；第一转换模块，用于使用预设算法将实体信息转换为布尔向量，使用预设模型将属性信息和属性值信息均转换为高维数值型向量，得到三元组向量。
可选地，文本生成模块包括：输入模块，用于将实体向量、属性向量和属性值向量输入至文本生成模型中，其中，文本生成模型包括深度神经网络模型，深度神经网络模型根据三元组样本和文本样本训练得到；文本生成子模块，用于基于文本生成模型生成与目标实体匹配文本。
可选地，上述装置还可以包括：模型生成模块，用于在将实体向量、属性向量和属性值向量输入至文本生成模型之前，生成文本生成模型，其中，模型生成模块包括：第二获取模块，用于获取三元组样本和文本样本；第二转换模块，用于使用预设算法将三元组样本中的实体样本转换为布尔向量，使用预设模型将三元组样本中的属性样本、属性值样本均转换为高维数值型向量，得到三元组向量样本；训练模块，用于基于三元组向量样本和文本样本训练文本生成模型，得到训练好的文本生成模型。
可选地，训练模块包括：编码模块，用于利用结合注意力机制的编码器处理三元组向量样本和文本样本，得到上下文向量；解码模块，用于利用结合注意力机制的解码器处理上下文向量，得到文本信息；训练子模块，用于基于文本信息，以最小化损失函数来训练文本生成模型。
可选地，上述预设算法为独热算法，预设模型为BERT模型或Word2Vector模型。
需要说明的是，上述选择模块502、确定模块504和文本生成模块506对应于实施例1中的步骤S102 至步骤S106，该三个模块与对应的步骤所实现的示例和应用场景相同，但不限于上述实施例1所公开的内容。
实施例4
根据本申请实施例，提供了另一种文本生成的装置，图6是根据本申请实施例的文本生成装置的示意图。如图6所示，该装置600包括接收模块602和显示模块604。
其中，接收模块602，用于接收选中指令，其中，选中指令用于选中待评价的目标实体；显示模块604，用于显示与目标实体匹配的文本，其中，文本依据目标实体的目标知识图谱确定出的目标实体的实体向量、属性向量和属性值向量生成，目标知识图谱来自知识图谱集，知识图谱集用于表征至少一个实体在预设属性上的属性值，实体向量、属性向量和属性值向量用三元组向量表征。
可选地，上述装置还可以包括图谱生成模块，用于在显示与目标实体匹配的文本之前，生成知识图谱集，其中，图谱生成模块可以包括：构建模块，用于构建知识图谱集的规划层，其中，规划层至少包括：实体类型、属性类型和属性值类型；第一获取模块，用于获取记录信息，其中记录信息包括：至少一个实体在预设属性上的属性值；图谱生成子模块，用于将记录信息输入至规划层中，生成知识图谱集。
可选地，上述装置还可以包括预处理模块，用于在将记录信息输入至规划层之前，对记录信息进行预处理，得到处理后的记录信息，其中，预处理包括如下至少之一：实体抽取、属性抽取、属性值抽取和实体消歧。
可选地，显示模块中还包括确定模块，用于根据目标知识图谱确定目标实体的实体向量、属性向量和属性值向量，其中，确定模块可以包括：提取模块，用于提取目标知识图谱中的目标实体的实体信息、属性信息和属性值信息；第一转换模块，用于使用预设算法将实体信息转换为布尔向量，使用预设模型将属性信息和属性值信息均转换为高纬数值型向量，得到三元组向量。
可选地，显示模块中还包括文本生成模块，用于依据实体向量、属性向量和属性值向量生成文本，其中，文本生成模块可以包括：输入模块，用于将实体向量、属性向量和属性值向量输入至文本生成模型中，其中，文本生成模型包括深度神经网络模型，深度神经网络模型根据三元组样本和文本样本训练得到；文本生成子模块，用于基于文本生成模型生成与目标实体匹配文本。
可选地，上述装置还可以包括模型生成模块，用于在将实体向量、属性向量和属性值向量输入至文本生成模型之前，生成文本生成模型，其中，模型生成模块可以包括：第二获取模块，用于获取三元组样本和文本样本；第二转换模块，用于使用预设算法将三元组样本中的实体样本转换为布尔向量，使用预设模型将三元组样本中的属性样本、属性值样本均转换为高纬数值型向量，得到三元组向量样本；训练模块，用于基于三元组向量样本和文本样本训练文本生成模型，得到训练好的文本生成模型。
可选地，训练模块可以包括：编码模块，用于利用结合注意力机制的编码器处理三元组向量样本和文本样本，得到上下文向量；解码模块，用于利用结合注意力机制的解码器处理上下文向量，得到文本信息；训练子模块，用于基于文本信息，以最小化损失函数来训练文本生成模型。
可选地，上述预设算法为独热算法，预设模型为BERT模型或Word2Vector模型。
需要说明的是，上述接收模块602和显示模块604对应于实施例2中的步骤S402至步骤S404，该两个模块与对应的步骤所实现的示例和应用场景相同，但不限于上述实施例2所公开的内容。
实施例5
根据本申请实施例，提供了一种存储介质，存储介质包括存储的程序，其中，在程序运行时控制存储介质所在设备执行实施例1或2中的文本生成方法。
实施例6
根据本申请实施例，提供了一种处理器，处理器用于运行程序，其中，在程序运行时执行实施例1或2中的文本生成方法。
上述本申请实施例序号仅仅为了描述，不代表实施例的优劣。
在本申请的上述实施例中，对各个实施例的描述都各有侧重，某个实施例中没有详述的部分，可以参见其他实施例的相关描述。
在本申请所提供的几个实施例中，应该理解到，所揭露的技术内容，可通过其它的方式实现。其中，以上所描述的装置实施例仅仅是示意性的，例如所述单元的划分，可以为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如多个单元或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些接口，单元或模块的间接耦合或通信连接，可以是电性或其它的形式。
所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
另外，在本申请各个实施例中的各功能单元可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中。上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。
所述集成的单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请的技术方案本质上或者说对相关技术做出贡献的部分或者该技术方案的全部或部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可为个人计算机、服务器或者网络设备等)执行本申请各个实施例所述方法的全部或部分步骤。而前述的存储介质包括：U盘、只读存储器(ROM，Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、移动硬盘、磁碟或者光盘等各种可以存储程序代码的介质。

Claims (11)

  1. 一种文本生成方法，包括：
    从知识图谱集中选择目标实体的目标知识图谱，其中，所述知识图谱集用于表征至少一个实体在预设属性上的属性值，所述目标实体为待评价的对象；
    基于所述目标知识图谱确定所述目标实体的实体向量、属性向量和属性值向量，其中，所述实体向量、属性向量和属性值向量用三元组向量表征；
    依据所述实体向量、属性向量和属性值向量生成与所述目标实体匹配的文本。
  2. 根据权利要求1所述的方法，其中，在从知识图谱集中选择目标实体的目标知识图谱之前，所述方法还包括：生成所述知识图谱集，其中，生成所述知识图谱集的步骤包括：
    构建所述知识图谱集的规划层，其中，所述规划层至少包括：实体类型、属性类型和属性值类型；
    获取记录信息，其中所述记录信息包括：至少一个实体在预设属性上的属性值；
    将所述记录信息输入至所述规划层中，生成所述知识图谱集。
  3. 根据权利要求2所述的方法，其中，在将所述记录信息输入至所述规划层之前，所述方法还包括：
    对所述记录信息进行预处理，得到处理后的记录信息，其中，所述预处理包括如下至少之一：实体抽取、属性抽取、属性值抽取和实体消歧。
  4. 根据权利要求1所述的方法，其中，基于所述目标知识图谱确定所述目标实体的实体向量、属性向量和属性值向量，包括：
    提取所述目标知识图谱中的所述目标实体的实体信息、属性信息和属性值信息；
    使用预设算法将所述实体信息转换为布尔向量，使用预设模型将所述属性信息和所述属性值信息均转换为高纬数值型向量，得到所述三元组向量。
  5. 根据权利要求1所述的方法，其中，依据所述实体向量、属性向量和属性值向量生成与所述目标实体匹配的文本，包括：
    将所述实体向量、属性向量和属性值向量输入至文本生成模型中，其中，所述文本生成模型包括深度神经网络模型，所述深度神经网络模型根据三元组样本和文本样本训练得到；
    基于所述文本生成模型生成与所述目标实体匹配的文本。
  6. 根据权利要求5所述的方法，其中，在将所述实体向量、属性向量和属性值向量输入至文本生成模型之前，所述方法还包括：生成所述文本生成模型，其中，生成所述文本生成模型的步骤包括：
    获取所述三元组样本和文本样本；
    使用预设算法将所述三元组样本中的实体样本转换为布尔向量，使用预设模型将所述三元组样本中的属性样本、属性值样本均转换为高纬数值型向量，得到三元组向量样本；
    基于所述三元组向量样本和所述文本样本训练所述文本生成模型，得到训练好的文本生成模型。
  7. 根据权利要求5所述的方法，其中，基于所述三元组向量样本和所述文本样本训练所述文本生成模型，得到训练好的文本生成模型，包括：
    利用结合注意力机制的编码器处理所述三元组向量样本和所述文本样本，得到上下文向量；
    利用结合注意力机制的解码器处理所述上下文向量，得到文本信息；
    基于所述文本信息，以最小化损失函数来训练所述文本生成模型。
  8. 一种文本生成方法，包括：
    接收选中指令，其中，所述选中指令用于选中待评价的目标实体；
    显示与所述目标实体匹配的文本，其中，所述文本依据所述目标实体的目标知识图谱确定出的所述目标实体的实体向量、属性向量和属性值向量生成，所述目标知识图谱来自知识图谱集，所述知识图谱集用于表征至少一个实体在预设属性上的属性值，所述实体向量、属性向量和属性值向量用三元组向量表征。
  9. 一种文本生成装置，包括：
    选择模块，用于从知识图谱集中选择目标实体的目标知识图谱，其中，所述知识图谱集用于表征至少一个实体在预设属性上的属性值，所述目标实体为待评价的对象；
    确定模块，用于基于所述目标知识图谱确定所述目标实体的实体向量、属性向量和属性值向量，其中，所述实体向量、属性向量和属性值向量用三元组向量表征；
    文本生成模块，用于依据所述实体向量、属性向量和属性值向量生成与所述目标实体匹配的文本。
  10. 一种存储介质，其中，所述存储介质包括存储的程序，在所述程序运行时控制所述存储介质所在设备执行权利要求1或8的文本生成方法。
  11. 一种处理器，其中，所述处理器用于运行程序，所述程序运行时执行权利要求1或8的文本生成方法。
PCT/CN2019/126797 2019-08-21 2019-12-20 文本生成方法和装置 WO2021031480A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910775353.1A CN110489755A (zh) 2019-08-21 2019-08-21 文本生成方法和装置
CN201910775353.1 2019-08-21

Publications (1)

Publication Number Publication Date
WO2021031480A1 true WO2021031480A1 (zh) 2021-02-25

Family

ID=68552697

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/126797 WO2021031480A1 (zh) 2019-08-21 2019-12-20 文本生成方法和装置

Country Status (2)

Country Link
CN (1) CN110489755A (zh)
WO (1) WO2021031480A1 (zh)

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113158189A (zh) * 2021-04-28 2021-07-23 绿盟科技集团股份有限公司 一种恶意软件分析报告生成方法、装置、设备和介质
CN113239203A (zh) * 2021-06-02 2021-08-10 北京金山数字娱乐科技有限公司 一种基于知识图谱的筛选方法及装置
CN113609291A (zh) * 2021-07-27 2021-11-05 科大讯飞(苏州)科技有限公司 实体分类方法、装置、电子设备和存储介质
CN113761167A (zh) * 2021-09-09 2021-12-07 上海明略人工智能(集团)有限公司 一种会话信息抽取方法、系统、电子设备及存储介质
CN116150929A (zh) * 2023-04-17 2023-05-23 中南大学 一种铁路选线知识图谱的构建方法
CN116306925A (zh) * 2023-03-14 2023-06-23 中国人民解放军总医院 一种生成端到端实体链接方法和系统
CN116452072A (zh) * 2023-06-19 2023-07-18 华南师范大学 一种教学评价方法、系统、设备和可读存储介质
CN117332282A (zh) * 2023-11-29 2024-01-02 之江实验室 一种基于知识图谱的事件匹配的方法及装置

Families Citing this family (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110489755A (zh) * 2019-08-21 2019-11-22 广州视源电子科技股份有限公司 文本生成方法和装置
CN111061152B (zh) * 2019-12-23 2021-04-27 深圳供电局有限公司 基于深度神经网络的攻击识别方法及智慧能源功率控制装置
CN111209389B (zh) * 2019-12-31 2023-08-11 天津外国语大学 一种电影故事生成方法
CN111897955B (zh) * 2020-07-13 2024-04-09 广州视源电子科技股份有限公司 基于编解码的评语生成方法、装置、设备和存储介质
CN111930959B (zh) * 2020-07-14 2024-02-09 上海明略人工智能(集团)有限公司 用于图谱知识生成文本的方法与装置
CN112036146A (zh) * 2020-08-25 2020-12-04 广州视源电子科技股份有限公司 一种评语生成方法、装置、终端设备及存储介质
CN112069781B (zh) * 2020-08-27 2024-01-02 广州视源电子科技股份有限公司 一种评语生成方法、装置、终端设备及存储介质
CN113157941B (zh) * 2021-04-08 2023-04-18 支付宝(杭州)信息技术有限公司 业务特征数据处理、文本生成方法、装置及电子设备
CN113111188B (zh) * 2021-04-14 2022-08-09 清华大学 一种文本生成方法及系统
CN113488165B (zh) * 2021-07-26 2023-08-22 平安科技（深圳）有限公司 基于知识图谱的文本匹配方法、装置、设备以及存储介质
CN113569554B (zh) * 2021-09-24 2021-12-28 北京明略软件系统有限公司 一种数据库中实体对匹配方法、装置、电子设备及存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108345690A (zh) * 2018-03-09 2018-07-31 广州杰赛科技股份有限公司 智能问答方法与系统
CN109189944A (zh) * 2018-09-27 2019-01-11 桂林电子科技大学 基于用户正负反馈画像编码的个性化景点推荐方法及系统
CN109684394A (zh) * 2018-12-13 2019-04-26 北京百度网讯科技有限公司 文本生成方法、装置、设备和存储介质
US20190251182A1 (en) * 2018-02-12 2019-08-15 International Business Machines Corporation Extraction of information and smart annotation of relevant information within complex documents
CN110489755A (zh) * 2019-08-21 2019-11-22 广州视源电子科技股份有限公司 文本生成方法和装置

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10783179B2 (en) * 2015-08-03 2020-09-22 International Business Machines Corporation Automated article summarization, visualization and analysis using cognitive services
CN108763336A (zh) * 2018-05-12 2018-11-06 北京无忧创新科技有限公司 一种签证自助服务系统

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190251182A1 (en) * 2018-02-12 2019-08-15 International Business Machines Corporation Extraction of information and smart annotation of relevant information within complex documents
CN108345690A (zh) * 2018-03-09 2018-07-31 广州杰赛科技股份有限公司 智能问答方法与系统
CN109189944A (zh) * 2018-09-27 2019-01-11 桂林电子科技大学 基于用户正负反馈画像编码的个性化景点推荐方法及系统
CN109684394A (zh) * 2018-12-13 2019-04-26 北京百度网讯科技有限公司 文本生成方法、装置、设备和存储介质
CN110489755A (zh) * 2019-08-21 2019-11-22 广州视源电子科技股份有限公司 文本生成方法和装置

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113158189B (zh) * 2021-04-28 2023-12-26 绿盟科技集团股份有限公司 一种恶意软件分析报告生成方法、装置、设备和介质
CN113158189A (zh) * 2021-04-28 2021-07-23 绿盟科技集团股份有限公司 一种恶意软件分析报告生成方法、装置、设备和介质
CN113239203A (zh) * 2021-06-02 2021-08-10 北京金山数字娱乐科技有限公司 一种基于知识图谱的筛选方法及装置
CN113609291A (zh) * 2021-07-27 2021-11-05 科大讯飞(苏州)科技有限公司 实体分类方法、装置、电子设备和存储介质
CN113761167A (zh) * 2021-09-09 2021-12-07 上海明略人工智能(集团)有限公司 一种会话信息抽取方法、系统、电子设备及存储介质
CN113761167B (zh) * 2021-09-09 2023-10-20 上海明略人工智能(集团)有限公司 一种会话信息抽取方法、系统、电子设备及存储介质
CN116306925A (zh) * 2023-03-14 2023-06-23 中国人民解放军总医院 一种生成端到端实体链接方法和系统
CN116306925B (zh) * 2023-03-14 2024-05-03 中国人民解放军总医院 一种生成端到端实体链接方法和系统
CN116150929A (zh) * 2023-04-17 2023-05-23 中南大学 一种铁路选线知识图谱的构建方法
CN116452072A (zh) * 2023-06-19 2023-07-18 华南师范大学 一种教学评价方法、系统、设备和可读存储介质
CN116452072B (zh) * 2023-06-19 2023-08-29 华南师范大学 一种教学评价方法、系统、设备和可读存储介质
CN117332282A (zh) * 2023-11-29 2024-01-02 之江实验室 一种基于知识图谱的事件匹配的方法及装置
CN117332282B (zh) * 2023-11-29 2024-03-08 之江实验室 一种基于知识图谱的事件匹配的方法及装置

Also Published As

Publication number Publication date
CN110489755A (zh) 2019-11-22

Similar Documents

Publication Publication Date Title
WO2021031480A1 (zh) 文本生成方法和装置
CN109766427B (zh) 一种基于协同注意力的虚拟学习环境智能问答方法
CN104050160B (zh) 一种机器与人工翻译相融合的口语翻译方法和装置
WO2023035610A1 (zh) 基于关键词感知的多模态注意力视频问答方法与系统
WO2021218028A1 (zh) 基于人工智能的面试内容精炼方法、装置、设备及介质
CN115599899B (zh) 基于飞行器知识图谱的智能问答方法、系统、设备及介质
CN114595306B (zh) 基于距离感知自注意力机制和多角度建模的文本相似度计算系统及方法
CN114722839A (zh) 人机协同对话交互系统及方法
CN115146975A (zh) 基于深度学习的面向师-机-生教学成效评估方法和系统
CN114818717A (zh) 融合词汇和句法信息的中文命名实体识别方法及系统
CN112231491A (zh) 基于知识结构的相似试题识别方法
CN115775349A (zh) 基于多模态融合的假新闻检测方法和装置
CN112069781A (zh) 一种评语生成方法、装置、终端设备及存储介质
CN115408488A (zh) 用于小说场景文本的分割方法及系统
CN114880307A (zh) 一种开放教育领域知识的结构化建模方法
CN113486174A (zh) 模型训练、阅读理解方法、装置、电子设备及存储介质
CN112749566B (zh) 一种面向英文写作辅助的语义匹配方法及装置
CN117235261A (zh) 一种多模态方面级情感分析方法、装置、设备及存储介质
CN117235347A (zh) 一种基于大语言模型的青少年算法代码辅助学习系统及方法
CN116258147A (zh) 一种基于异构图卷积的多模态评论情感分析方法及系统
CN116127959A (zh) 基于深度学习的图像意境挖掘与意境转换中国古诗的方法
CN112818212B (zh) 语料数据采集方法、装置、计算机设备和存储介质
CN115116444A (zh) 一种语音识别文本的处理方法、装置、设备及存储介质
CN116661940B (zh) 组件识别方法、装置、计算机设备和存储介质
CN117453895B (zh) 一种智能客服应答方法、装置、设备及可读存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19942467

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19942467

Country of ref document: EP

Kind code of ref document: A1