Method, device, electronic device and storage medium for identifying media referencing type

Abstract

The present invention discloses a method, device, electronic device and storage medium for identifying media citation types, and belongs to the field of natural language processing. The method includes: inputting source information published by source media and quoted information published by quoted media into a pre-trained language model to obtain chapter representation vectors and sentence representation vectors of the source information and quoted information respectively; performing bidirectional interactive semantic information learning on each sentence representation vector of the source information and quoted information to obtain a bidirectional interactive semantic chapter representation vector; fusing the chapter representation vector and the interactive semantic chapter representation vector obtained based on the pre-trained language model to obtain an enhanced semantic representation vector; performing media citation type identification according to the enhanced semantic representation vector to obtain a media citation type. Through the present invention, the semantic representation capability of the chapter is enhanced in a multi-level and fine-grained manner, and the recognition performance of the media citation type is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to a method, device, electronic device and storage medium for identifying media referencing types. Background Art

[0002] The identification of media citation types aims to analyze the citation methods between information pairs with citation relationships. Accurately identifying citation relationships helps to characterize the communication model from multiple perspectives, gain in-depth insights into the dynamic evolution of information dissemination, and provide decision-making support for management departments to proactively perceive social hot spots.

[0003] In existing research, (Yao R, Li Q, Wang L, et al. A novel neural approach for news reprint prediction [C] / / 2019 International Joint Conference on Neural Networks (IJCNN). IEEE, 2019: 1-8.) and (Li Q, Liu H, Yao R, et al. Entity Matters in News: An Association Network-enhanced Method for News Reprint Prediction [J]. IEEE Intelligent Systems, 2021.) studied the mining of citation relationships by integrating citation association networks and semantic content; the pre-trained language model BERT (Kenton JDMWC, Toutanova L K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding [C] / / Proceedings of NAACL-HLT.2019:4171-4186.) With its powerful deep semantic encoding ability, it improves the performance of multiple downstream tasks such as sentiment analysis and text classification in the field of natural language processing, and provides a technical solution for enhancing the ability to identify media citation types; (Wang L, Yao R, Li Q, et al. A Multi-level Semantic Fusion Approach for News Reprint Pattern Detection [C] / / 2021 International Joint Conference on Neural Networks (IJCNN). IEEE, 2021: 1-8.) The long short-term memory neural network (LSTM) is used to model the multi-level interactive semantics of media information pairs to identify citation types. However, this method only realizes the single-direction mining of sentence association information of "source information → quoted information", but ignores the reverse semantic association. Summary of the invention

[0004] The embodiments of the present invention provide a method, device, electronic device and storage medium for identifying media citation types, so as to solve the technical problem that it is difficult to deeply model the deep semantic features contained in different citation types in the prior art.

[0005] According to one aspect of an embodiment of the present application, a method for identifying media citation types is provided, including: inputting source information published by a source media and citation information published by a citation media into a pre-trained language model to obtain a first chapter representation vector of the source information and a first sentence representation vector set of the source information, a second chapter representation vector of the citation information, and a second sentence representation vector set of the citation information, wherein the first sentence representation vector set includes sentence representation vectors of each sentence in the source information, and the second sentence representation vector set includes sentence representation vectors of each sentence in the citation information; performing bidirectional interactive semantic information learning based on the first sentence representation vector set and the second sentence representation vector set to obtain a first interactive semantic chapter representation vector from the source information to the citation information and a second interactive semantic chapter representation vector from the citation information to the source information; fusing the first chapter representation vector and the first interactive semantic chapter representation vector, the second chapter representation vector and the second interactive semantic chapter representation vector to obtain a first enhanced semantic representation vector of the source information and a second enhanced semantic representation vector of the citation information; performing media citation type identification based on the first enhanced semantic representation vector and the second enhanced semantic representation vector to obtain a media citation type.

[0006] Further, the method of inputting the source information published by the source media and the quoted information published by the quoted media into the pre-trained language model to obtain the first chapter representation vector of the source information and the first sentence representation vector set of the source information, the second chapter representation vector of the quoted information and the second sentence representation vector set of the quoted information includes: constructing a media quoted data sequence pair from the source information published by the source media and the quoted information published by the quoted media; inputting the media quoted data sequence pair into the pre-trained language model to output a first perceptual feature representation of the source information and a second perceptual feature representation of the quoted information; obtaining a sentence representation vector of each sentence of the source information and a sentence representation vector of each sentence of the quoted information according to the sequence position latent vector of each sentence in the first perceptual feature representation and the second perceptual feature representation; and outputting the latent vector of the first position of the chapter in the first perceptual feature representation and the second perceptual feature representation as the first chapter representation vector of the source information and the second chapter representation vector of the quoted information.

[0007] Furthermore, the method of constructing a media citation data sequence pair from the source information published by the source media and the cited information published by the cited media includes: constructing a media citation information pair based on the source information and the cited information; adding interval identifiers to the beginning position and inter-sentence position of the source information and the cited information in the media citation information pair to obtain a media citation data sequence pair.

[0008] Furthermore, the method of obtaining the sentence representation vector of each sentence of the source information and the sentence representation vector of each sentence of the quoted information according to the sequence position latent vector of each sentence in the first perceptual feature representation and the second perceptual feature representation includes: calculating a first average value of the sequence position latent vector of each sentence in the first perceptual feature representation; calculating a second average value of the sequence position latent vector of each sentence in the second perceptual feature representation; determining the first average value as the sentence representation vector of each sentence of the source information, and determining the second average value as the sentence representation vector of each sentence of the quoted information.

[0009] Furthermore, the method performs bidirectional interactive semantic information learning based on the first sentence representation vector set and the second sentence representation vector set to obtain a first interactive semantic chapter representation vector from source information to quoted information and a second interactive semantic chapter representation vector from quoted information to source information, including: constructing a first sentence semantic representation pair from source media to quoted media and a second sentence semantic representation pair from quoted media to source media based on the first sentence representation vector set and the second sentence representation vector set; inputting the first sentence semantic representation pair into a fully connected neural network to obtain a first sentence pair semantic association score matrix from source media to quoted media; inputting the second sentence semantic representation pair into a fully connected neural network to obtain a second sentence pair semantic association score matrix from quoted media to source media; and inputting the first sentence semantic representation pair into a fully connected neural network to obtain a second sentence pair semantic association score matrix from quoted media to source media. The sentence pair semantic association score matrix is ​​subjected to maximum pooling processing to obtain the first maximum association vector between the sentence pairs in the direction of the source media to the quoted media; the second sentence pair semantic association score matrix is ​​subjected to maximum pooling processing to obtain the second maximum association vector between the sentence pairs in the direction of the quoted media to the source media; the first maximum association vector and the second maximum association vector are normalized respectively to obtain the first set of weights and the second set of weights of the contribution of each clause in the source information and the quoted information to the text semantics; the first set of weights is used to perform weighted calculation on the first sentence representation vector set to obtain the first interactive semantic chapter representation vector from the source information to the quoted information; the second set of weights is used to perform weighted calculation on the second sentence representation vector set to obtain the second interactive semantic chapter representation vector from the quoted information to the source information.

[0010] Furthermore, the media recitation type is identified based on the first enhanced semantic representation vector and the second enhanced semantic representation vector to obtain the media recitation type, including: taking the first enhanced semantic representation vector and the second enhanced semantic representation vector as input data, inputting into a fully connected neural network with softmax as a non-linear activation function, and outputting the media recitation type.

[0011] Furthermore, outputting the media citation type includes: outputting a media citation type tag; wherein the media citation type tag is used to characterize the citation type, and the citation type includes: word variation, sentence rewriting, and paragraph restatement.

[0012] According to another aspect of the embodiment of the present application, there is also provided a device for identifying the type of media citation, including: a text encoding module, used to input source information published by source media and quoted information published by quoted media into a pre-trained language model, to obtain a first chapter representation vector of the source information and a first sentence representation vector set of the source information, a second chapter representation vector of the quoted information and a second sentence representation vector set of the quoted information, wherein the first sentence representation vector set includes sentence representation vectors of each sentence in the source information, and the second sentence representation vector set includes sentence representation vectors of each sentence in the quoted information; a semantic information learning module, used to classify the source information according to the first sentence representation vector set; and the second sentence representation vector set to perform bidirectional interactive semantic information learning to obtain a first interactive semantic chapter representation vector from source information to quoted information and a second interactive semantic chapter representation vector from quoted information to source information; a semantic fusion module, used to fuse the first chapter representation vector and the first interactive semantic chapter representation vector, the second chapter representation vector and the second interactive semantic chapter representation vector to obtain a first enhanced semantic representation vector of source information and a second enhanced semantic representation vector of quoted information; a citation type identification module, used to identify media citation types according to the first enhanced semantic representation vector and the second enhanced semantic representation vector to obtain media citation types.

[0013] Furthermore, the text encoding module includes: a sequence pair construction unit, which is used to construct a media citation data sequence pair using the source information published by the source media and the cited information published by the cited media; a perceptual feature output unit, which is used to input the media citation data sequence pair into a pre-trained language model, and output a first perceptual feature representation of the source information and a second perceptual feature representation of the cited information; a sentence representation vector processing unit, which is used to obtain a sentence representation vector of each sentence of the source information and a sentence representation vector of each sentence of the cited information based on the sequence position latent vector of each sentence in the first perceptual feature representation and the second perceptual feature representation; and a chapter representation vector processing unit, which is used to output the latent vector of the first position of the chapter in the first perceptual feature representation and the second perceptual feature representation as the first chapter representation vector of the source information and the second chapter representation vector of the cited information.

[0014] Furthermore, the sequence pair construction unit includes: an information pair construction unit, which is used to construct a media reference information pair based on the source information and the reference information; an identifier adding unit, which is used to add interval identifiers to the beginning position and inter-sentence position of the source information and the reference information in the media reference information pair to obtain a media reference data sequence pair.

[0015] Furthermore, the sentence representation vector processing unit includes: a first calculation unit, used to calculate a first average value of the sequence position latent vector of each sentence in the first perceptual feature representation; and determine the first average value as the sentence representation vector of each sentence of the source information; a second calculation unit, used to calculate a second average value of the sequence position latent vector of each sentence in the second perceptual feature representation; and determine the second average value as the sentence representation vector of each sentence of the quoted information.

[0016] Furthermore, the semantic information learning module includes: a representation pair construction unit, which is used to construct a first sentence semantic representation pair from the source media to the redirected media direction and a second sentence semantic representation pair from the redirected media to the source media direction according to the first sentence representation vector set and the second sentence representation vector set; an association score processing unit, which is used to input the first sentence semantic representation pair into a fully connected neural network to obtain a first sentence pair semantic association score matrix from the source media to the redirected media direction; input the second sentence semantic representation pair into a fully connected neural network to obtain a second sentence pair semantic association score matrix from the redirected media to the source media direction; and a pooling processing unit, which is used to perform maximum pooling processing on the first sentence pair semantic association score matrix to obtain a second sentence pair semantic association score matrix from the source media to the redirected media direction. a maximum correlation vector; performing maximum pooling processing on the semantic correlation score matrix of the second sentence pair to obtain the second maximum correlation vector between the sentence pairs in the direction of the quoted media to the source media; a normalization processing unit, used to perform normalization processing on the first maximum correlation vector and the second maximum correlation vector respectively to obtain a first set of weights and a second set of weights for the contribution of each clause in the source information and the quoted information to the text semantics; a weighted processing unit, used to perform weighted calculation on the first sentence representation vector set using the first set of weights to obtain a first interactive semantic chapter representation vector from the source information to the quoted information; and perform weighted calculation on the second sentence representation vector set using the second set of weights to obtain a second interactive semantic chapter representation vector from the quoted information to the source information.

[0017] Furthermore, the citation type identification module includes: a media citation type output unit, which is used to take the first enhanced semantic representation vector and the second enhanced semantic representation vector as input data, input them into a fully connected neural network with softmax as a nonlinear activation function, and output the media citation type.

[0018] Furthermore, the media citation type output module includes: a first output unit, used to output a media citation type label; wherein the media citation type label is used to characterize the citation type, and the citation type includes: word variation, sentence rewriting, and paragraph restatement.

[0019] According to another aspect of an embodiment of the present application, a storage medium is further provided, which includes a stored program, and the above steps are executed when the program is run.

[0020] According to another aspect of an embodiment of the present application, there is also provided an electronic device, including a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; wherein: the memory is used to store computer programs; and the processor is used to execute the steps in the above method by running the program stored in the memory.

[0021] The embodiment of the present application also provides a computer program product including instructions, which, when executed on a computer, enables the computer to execute the steps in the above method.

[0022] Through the present invention, the information released by the source media and the information released by the quoted media are subjected to text encoding, two-way interactive semantic information learning, and semantic fusion enhancement steps based on a pre-trained language model, and the obtained enhanced semantic representation is predicted, so as to identify the citation type between the media. This scheme is encoded by a pre-trained language model, and can generate a chapter semantic representation vector and a semantic representation vector of each sentence of the media release information suitable for the field of media citation. The two-way interactive semantic information learning can model sentence-level semantic association information from the two directions of "source information → quoted information" and "quoted information → source information". By multi-level and fine-grained modeling of media semantic features that can reflect the potential characteristics of the citation type, the semantic representation ability of the chapter is enhanced, and the recognition performance of the media citation type is effectively improved, so as to realize the citation type identification with more diversified information interaction, deeper semantic representation, and more multi-level feature representation. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0024] Figure 1 is a hardware structure block diagram of a computer according to an embodiment of the present invention;

[0025] Figure 2 is a flow chart of a method for identifying a media referencing type according to an embodiment of the present invention;

[0026] Figure 3 It is a structural block diagram of a device for identifying media redirection type according to an embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only embodiments of a part of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present application. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0028] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0029] The method embodiment provided in the first embodiment of the present application can be executed in a mobile phone, a computer, a tablet or a similar computing device. Taking running on a computer as an example, Figure 1 is a hardware structure block diagram of a computer according to an embodiment of the present invention. Figure 1 As shown, the computer may include one or more ( Figure 1 Only one is shown in the figure) a processor 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA) and a memory 104 for storing data. Optionally, the computer may also include a transmission device 106 and an input / output device 108 for communication functions. It can be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the above-mentioned computer. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations are shown.

[0030] The memory 104 can be used to store computer programs, for example, software programs and modules of application software, such as a computer program corresponding to a method for identifying a type of media transfer in an embodiment of the present invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, to implement the above method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the computer via a network. Examples of the above network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0031] The transmission device 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the computer. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, referred to as NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0032] In this embodiment, a method for identifying a media transfer type is provided. Figure 2 FIG. 4 is a flow chart of a method for identifying a media transfer type according to an embodiment of the present invention. Figure 2 As shown, the process includes the following steps:

[0033] S100, inputting source information published by source media and quoted information published by quoted media into a pre-trained language model, obtaining a first chapter representation vector of the source information and a first sentence representation vector set of the source information, a second chapter representation vector of the quoted information and a second sentence representation vector set of the quoted information, wherein the first sentence representation vector set includes sentence representation vectors of each sentence in the source information, and the second sentence representation vector set includes sentence representation vectors of each sentence in the quoted information;

[0034] In this embodiment, the source information published by the source media and the quoted information published by the quoted media are text information, both of which contain several sentences. In order for the computer to process natural language, the input text needs to be represented by a vector. Pre-trained language models (PLMs) are an application of transfer learning. They use almost unlimited text to learn the context-related representation of each member of the input sentence. It implicitly learns general grammatical semantic knowledge and can embed contextual semantics through a large amount of expected unsupervised pre-training. Common pre-trained language models include: ELMo, GPT, BERT, Roberta, and ERNIE, SpanBERT, RoBERTa, ALBERT models based on BERT improvements, etc.

[0035] In this embodiment, the source information is n a , the quoted information is n b ,in, Indicates that it contains m sentences Indicates that it contains k sentences By inputting source information published by source media containing several sentences and quoted information published by quoted media into the pre-trained language model, the chapter semantic representation vectors of the source information and the quoted information and the semantic representation vectors of each sentence are obtained.

[0036] S200, performing bidirectional interactive semantic information learning based on the first sentence representation vector set and the second sentence representation vector set to obtain a first interactive semantic chapter representation vector from source information to quoted information and a second interactive semantic chapter representation vector from quoted information to source information;

[0037] In order to mine the fine-grained semantic associations between media source information and quoted information pairs, this embodiment learns the interactive semantic associations between each sentence pair from two directions: "source information→quoted information" and "quoted information→source information".

[0038] S300, fusing the first chapter representation vector and the first interactive semantic chapter representation vector, the second chapter representation vector and the second interactive semantic chapter representation vector to obtain a first enhanced semantic representation vector of source information and a second enhanced semantic representation vector of quoted information;

[0039] In this embodiment, the chapter representation vector obtained by encoding based on the pre-trained language model is fused with the interactive semantic chapter representation vector obtained after bidirectional interactive semantic information learning to obtain enhanced semantic representation vectors of source information and quoted information.

[0040] S400: Perform media recitation type identification according to the first enhanced semantic representation vector and the second enhanced semantic representation vector to obtain a media recitation type.

[0041] After obtaining the enhanced semantic representation vectors of the source information and the referenced information, the embodiment of the present invention performs reference type identification based on the fused enhanced semantic representation vector of the source information and the enhanced semantic representation vector of the referenced information.

[0042] The embodiment of the present invention proposes a method for identifying media citation types by bidirectional interactive semantic information enhanced modeling. The input data is source information released by large-scale source media and citation information released by citation media. After text encoding based on a pre-trained language model, bidirectional interactive semantic information learning, and semantic fusion enhancement steps, the obtained enhanced semantic representation is predicted and output, so as to identify the citation types between media. This scheme uses a pre-trained language model for encoding, and can generate a chapter semantic representation vector and a semantic representation vector of each sentence of media release information suitable for the field of media citation. The bidirectional interactive semantic information learning can model sentence-level semantic association information from two directions of "source information → citation information" and "citation information → source information". By multi-level and fine-grained modeling of media semantic features that can reflect the potential characteristics of citation types, the semantic representation ability of the chapter is enhanced, and the recognition performance of media citation types is effectively improved, so as to achieve citation type identification with more diversified information interaction, deeper semantic representation, and more multi-level feature representation.

[0043] In one implementation of the embodiment of the present invention, source information published by source media and quoted information published by quoted media are input into a pre-trained language model to obtain a first chapter representation vector of the source information and a first sentence representation vector set of the source information, a second chapter representation vector of the quoted information and a second sentence representation vector set of the quoted information, S100, including:

[0044] S101, constructing a media referencing data sequence pair from source information published by source media and referencing information published by referencing media;

[0045] In this embodiment, the pre-trained language model is a BERT model. BERT stands for Bidirectional Encoder Representation from Transformers, which means a bidirectional encoder based on the Transformer model. It is a pre-trained language representation model that can generate deep bidirectional language representation.

[0046] S102, inputting the media reference data sequence into a pre-trained language model, and outputting a first perceptual feature representation of the source information and a second perceptual feature representation of the referenced information;

[0047] The source information published by the source media and the quoted information published by the quoted media are constructed into a media quoted data sequence pair that meets the format required by the BERT model input. The sequence pair is input into the BERT model and encoded through the self-attention mechanism, and the perceptual feature representation of each sequence position that takes into account the full text information is output. Among them, after the source information and the quoted information are input into the BERT model, the output perceptual feature representation contains the latent vector representation of each sequence position of each sentence of the source information and the quoted information, as well as the latent vector representation of the first position of the article; among them, the perceptual feature representation of the source information is H a , the perceptual feature of the quoted information is represented by H b :

[0048]

[0049]

[0050] in, is the source information n a BERT input sequence; For citation information b BERT input sequence; Indicates source media release information n a The i-th sentence The hidden vector representation of each sequence position of is, and the sequence length is Release information for media quotations b The jth sentence The hidden vector representation of each sequence position of is, and the sequence length is is the hidden vector output of the first position of the source information; is the latent vector output of the first position of the quoted information; each sequence of the sentence can be the individual words or words or single characters that constitute the sentence, and each latent vector represents the dimension d hidden Set to the default value of 768 for the BERT model.

[0051] S103, obtaining a sentence representation vector of each sentence of the source information and a sentence representation vector of each sentence of the quoted information according to the sequence position latent vector of each sentence in the first perceptual feature representation and the second perceptual feature representation;

[0052] S104, outputting the latent vector of the first position of the chapter in the first perceptual feature representation and the second perceptual feature representation as the first chapter representation vector of the source information and the second chapter representation vector of the quoted information.

[0053] To model the semantics of media release information, the embodiment of the present invention uses the latent vector output of the first position of [CLS] of the BERT encoder and As its representation, the representation vector encodes the full text through the self-attention mechanism, so that it can effectively represent the deep semantic information of the full text.

[0054] In an embodiment of the present invention, the vectors calculated based on the latent vectors of each sequence position of each sentence in the perceptual feature representation are used as the sentence representation vectors of the source information and the quoted information, and the latent vectors of the first position of the chapter in the perceptual feature representation are used as the chapter representation vectors of the source information and the quoted information.

[0055] In the embodiment of the present invention, the pre-trained language model BERT is used as a text encoder. Its powerful semantic encoding ability can generate chapter semantic representations and semantic representation vectors of each sentence of media release information that are suitable for the field of media citation.

[0056] In one embodiment, source information published by a source media and cited information published by a cited media are used to construct a media citation data sequence pair, S101, including: constructing a media citation information pair according to the source information and the cited information; adding interval identifiers to the beginning position and inter-sentence position of the source information and the cited information in the media citation information pair to obtain a media citation data sequence pair.

[0057] For text classification tasks, the BERT model inserts a [CLS] symbol before the text, and uses the output vector corresponding to this symbol as the semantic representation of the entire text, which can be used for text classification. It can be understood that compared with other words / characters already in the text, this symbol without obvious semantic information will more "fairly" integrate the semantic information of each word / character in the text. For sentence pair classification tasks, in addition to adding the [CLS] symbol and using the corresponding output as the semantic representation of the text, the BERT model also uses a [SEP] symbol to separate the input sentences, and attaches two different text vectors to the two sentences to distinguish them.

[0058] In this embodiment, in order to adapt to the encoding format of the BERT model, spacing identifiers are added at the first position and between sentences of the source information and the quoted information text. Specifically, a [CLS] symbol is inserted at the first position of the source information and the quoted information text to connect the full text information, and a [SEP] symbol is inserted after each clause of the source information and the quoted information text as a sentence separator, thereby generating a media quoted data input sequence pair.

[0059] Among them, the media quoted information (n a ,n b ) BERT input sequence pair after adding interval identifier operation It can be expressed as:

[0060]

[0061]

[0062] [CLS] is the beginning position identifier, and [SEP] is the sentence separator.

[0063] In one embodiment, according to the sequence position latent vector of each sentence in the first perceptual feature representation and the second perceptual feature representation, a sentence representation vector of each sentence in the source information and a sentence representation vector of each sentence in the quoted information are obtained, S103, including:

[0064] Calculate a first average value of the sequence position latent vector of each sentence in the first perceptual feature representation; calculate a second average value of the sequence position latent vector of each sentence in the second perceptual feature representation; determine the first average value as the sentence representation vector of each sentence of the source information, and determine the second average value as the sentence representation vector of each sentence of the quoted information.

[0065] In the case of inputting the media transfer data of the source information and the transfer information into the sequence pair (n a ,n b ) The input BERT model is encoded through the self-attention mechanism, and the model outputs the perceptual feature representation H of each sequence position that takes into account the full text information. a and H b ,

[0066]

[0067]

[0068] Calculate the source media release information n a The i-th sentence The hidden vector of each sequence position The average value of the source media release information n a The i-th sentence The sentence representation vector Among them, the sequence length of each sentence in the source information is Calculate the number of cited media releases n b The jth sentence The hidden vector of each sequence position The average value of the cited media release information n b The jth sentence The sentence representation vector Among them, the sequence length of each sentence in the quoted information is Right now:

[0069]

[0070]

[0071] This embodiment calculates the average value of the latent vectors of the sequence position of each sentence in the perceptual feature representation, and uses the average value of the latent vectors of each sequence position of each sentence as the sentence representation vector of the sentence, which is used to represent the semantic representation of the sentence, and takes into account the full-text information in the text information.

[0072] In one embodiment, bidirectional interactive semantic information learning is performed based on the first sentence representation vector set and the second sentence representation vector set to obtain a first interactive semantic chapter representation vector from source information to quoted information and a second interactive semantic chapter representation vector from quoted information to source information, S200, including:

[0073] S201, constructing a first sentence semantic representation pair from source media to redirected media and a second sentence semantic representation pair from redirected media to source media according to the first sentence representation vector set and the second sentence representation vector set;

[0074] In order to mine the fine-grained semantic associations between media quoted information pairs, the embodiments of the present invention learn the interactive semantic associations between each sentence pair from two directions: "source information → quoted information" (s → r) and "quoted information → source information" (r → s), and a The semantic representation of each sentence Release information with quoted media respectively b The semantic representation of each sentence The first sentence semantic representation pair that constitutes the direction from source media to quoted media is Information released by the media b The semantic representation of each sentence Publish information with source media separately a The semantic representation of each sentence The second sentence semantic representation pair that constitutes the direction from the quoted media to the source media is

[0075] S202, inputting the first sentence semantic representation pair into a fully connected neural network to obtain a first sentence pair semantic association score matrix from source media to redirected media; inputting the second sentence semantic representation pair into a fully connected neural network to obtain a second sentence pair semantic association score matrix from redirected media to source media;

[0076] In the embodiment of the present invention, the first sentence semantic representation pair is Input the fully connected neural network to obtain the sentence pair semantic association score in the (s→r) direction The semantic representation of the second sentence is Input the fully connected neural network to obtain the sentence pair semantic association score in the (r→s) direction The calculation process is as follows:

[0077]

[0078]

[0079] in, is the semantic association score of the first sentence pair between the i-th sentence in the source information and the j-th sentence in the quoted information; is the semantic association score of the second sentence pair between the jth sentence in the quoted information and the ith sentence in the source information; is the sentence representation vector of each sentence of the source information; is the sentence representation vector of each sentence of the quoted information; W1, b1 are the learnable parameters in the fully connected network.

[0080] After obtaining the semantic association scores between all sentence pairs, the semantic association score matrix of the first sentence pair from the source media to the referenced media and the semantic association score matrix of the second sentence pair from the referenced media to the source media can be constructed according to the semantic association scores of each sentence pair. The first sentence with the quoted information The association score is the first row and first column of the semantic association score matrix for the first sentence. The second sentence with the quoted information The association score of is the first row and second column of the semantic association score matrix of the second sentence pair. Similarly, the semantic association score matrix S between all sentence pairs in the direction from source media to quoted media can be constructed. s→r ∈R m×k The semantic correlation score matrix S between all sentence pairs from the source media to the source media r→s ∈R k×m .

[0081] S203, performing maximum pooling processing on the semantic association score matrix of the first sentence pair to obtain a first maximum association vector between sentence pairs in the direction of source media to redirected media; performing maximum pooling processing on the semantic association score matrix of the second sentence pair to obtain a second maximum association vector between sentence pairs in the direction of redirected media to source media;

[0082] Based on the two-way sentence-level interactive semantic association score matrix S s→r ∈R m×k and S r→s ∈R k×m , the embodiment of the present invention respectively applies the maximum pooling operation to obtain the bidirectional maximum correlation vector V between the sentence pairs. s→r ∈R m and V r→s ∈R k ,in,

[0083] V s→r =max_pooling(S s→r );

[0084] V r→s =max_pooling(S r→s );

[0085] For the maximum correlation vector V in the direction of "source information → quoted information" s→r , each dimension value represents the source media release information n a Each sentence in the quote is compared to the media release information n b The maximum correlation mined from all sentences in . The maximum correlation vector V in the opposite direction r→s Each dimension value represents n b Each sentence in n a The maximum relevance of all sentences in .

[0086] S204, respectively normalizing the first maximum relevance vector and the second maximum relevance vector to obtain a first set of weights and a second set of weights of contributions of each clause in the source information and the quoted information to the text semantics;

[0087] In the embodiment of the present invention, the bidirectional maximum relevance vector is normalized by a softmax layer as the contribution weight of each clause to the text semantics, and the first group of weights a of the contribution of each clause in the source information and the quoted information to the text semantics is obtained. s→r and the second set of weights a r→s ;in,

[0088] a s→r =softmax(V s→r );

[0089] a r→s =softmax(V r→s );

[0090] S205, using the first set of weights to perform weighted calculation on the first set of sentence representation vectors to obtain a first interactive semantic chapter representation vector from source information to quoted information; using the second set of weights to perform weighted calculation on the second set of sentence representation vectors to obtain a second interactive semantic chapter representation vector from quoted information to source information.

[0091] The embodiment of the present invention adopts the first group of weights a s→r Represent the vector set for the first sentence Perform weighted calculation to obtain the first interactive semantic chapter representation vector from source information to quoted information Using the second set of weights a r→sRepresent the vector set for the second sentence Perform weighted calculation to obtain the second interactive semantic chapter representation vector of the quoted information to the source information The calculation process is as follows:

[0092]

[0093]

[0094] The embodiment of the present invention is based on the semantic representation of each sentence of the media citation information pair, and uses a fully connected neural network to learn the interactive semantic association scores between each sentence pair from the two directions of "source information→quoted information" and "quoted information→source information", thereby constructing a sentence-level interactive semantic association matrix in two directions, and based on the bidirectional sentence-level interactive association matrix, it is used as the source of the attention mechanism to measure the semantic contribution of each sentence to the text in which it is located, thereby calculating the weighted sum of all sentences to obtain the interactive semantic chapter representation of the source information and the citation information, which reflects the media semantic features of the potential characteristics of the citation type in a more fine-grained manner.

[0095] In one embodiment, the first chapter representation vector and the first interactive semantic chapter representation vector, the second chapter representation vector and the second interactive semantic chapter representation vector are merged to obtain a first enhanced semantic representation vector of source information and a second enhanced semantic representation vector of quoted information, S300, including:

[0096] The first chapter representation vector and the first interactive semantic chapter representation vector, the second chapter representation vector and the second interactive semantic chapter representation vector are fused according to the following formula to obtain a first enhanced semantic representation vector of the source information and a second enhanced semantic representation vector of the quoted information:

[0097]

[0098]

[0099] Among them, f a is the first enhanced semantic representation vector; f b is the second enhanced semantic representation vector; Represent the vector for the first chapter; Represent vectors for the second chapter; A first interactive semantic chapter representation vector from source information to quoted information; The second interactive semantic chapter representation vector for referencing information to source information.

[0100] In one implementation, the performing of media recitation type identification according to the first enhanced semantic representation vector and the second enhanced semantic representation vector to obtain the media recitation type, S400, includes:

[0101] S401, taking the first enhanced semantic representation vector and the second enhanced semantic representation vector as input data, inputting them into a fully connected neural network with softmax as a nonlinear activation function, and outputting a media citation type.

[0102] The embodiment of the present invention is based on the obtained bidirectional interactive semantic information enhanced representation vector f a and f b , is input to another fully connected neural network with softmax as nonlinear activation function, so as to identify the citation pattern between media news pairs. The calculation process is:

[0103]

[0104] To minimize the training set The cross entropy loss function is the training target;

[0105] The loss function is formally defined as follows:

[0106]

[0107] Among them, f a is the first enhanced semantic representation vector; f b is the second enhanced semantic representation vector; N and c represent the number of training set samples and citation types respectively; y i and are the true and predicted label vectors of the i-th sample data respectively.

[0108] In one embodiment, outputting the media quotation type includes: outputting a media quotation type tag; wherein the media quotation type tag is used to characterize the quotation type, and the quotation type includes: word variation, sentence rewriting, and paragraph restatement.

[0109] In this embodiment, the input data is input into a fully connected neural network, and the output is a media citation type label, which is used to characterize the citation type, and the citation type includes: word variation, sentence rewriting, and chapter restatement. For example, the output media citation type label can be 0, 1, or 2, representing word variation, sentence rewriting, and chapter restatement, respectively, where word variation can be understood as rewriting at the word level; sentence rewriting involves the reorganization of the word order at the sentence level, the conversion of the writing style such as sentence pattern, etc.; chapter restatement focuses on the reorganization and rewriting around the theme at the paragraph level.

[0110] The following is a complete explanation and description of the solution of the embodiment of the present invention through a specific implementation. In this implementation, the method for identifying the media transfer type includes:

[0111] S1, construct media citation information pairs according to the source information published by the source media and the citation information published by the citation media (n a ,n b );

[0112] in, Indicates that it contains m sentences Source information; Indicates that it contains k sentences citation information;

[0113] S2, in order to adapt the encoding format of the BERT model, an interval identifier is added to the beginning position of the source information and the quoted information in the media quoted information pair and the position between sentences to obtain a media quoted data sequence pair

[0114] The adding of the separator identifier to the text beginning position and the sentence position of the source information and the quoted information in the media quoted information pair includes: inserting a [CLS] symbol at the beginning position of the input source information and the quoted information text to connect the full text information, and inserting a [SEP] symbol after each clause as a sentence separator, and adding the separator identifier to the media quoted information pair (n a ,n b ) BERT input sequence pair after adding interval identifier operation It can be expressed as:

[0115]

[0116]

[0117] [CLS] is the position identifier of the beginning of the text, and [SEP] is the sentence separator;

[0118] S3, the media transfer data sequence is Input the pre-trained semantic BERT model, and output the perceptual feature representation H of each sequence position that takes into account the full text information. a and H b ,in,

[0119]

[0120]

[0121] H a is the first perceptual feature representation of the source information; H b It is the second perceptual characteristic representation of the quoted information; Indicates source media release information n a The i-th sentence The hidden vector representation of each sequence position of is, and the sequence length is Release information for media quotations b The jth sentence The hidden vector representation of each sequence position of is, and the sequence length is is the hidden vector output of the first position of the source information; is the hidden vector output of the first position of the quoted information; each hidden vector represents the dimension d hidden Set to the default value of 768 for the BERT model;

[0122] S4, taking the average of the latent vectors of the sequence positions of each sentence in the first perceptual feature representation and the second perceptual feature representation, the obtained vector represents the semantic representation of the sentence, and obtaining the sentence representation vector of each sentence of the source information and the sentence representation vector of each sentence of the quoted information

[0123]

[0124]

[0125] S5, to model the semantics of media release information, use the latent vector output of the first position of [CLS] of the BERT encoder and As its chapter semantic representation, that is, according to the latent vector output of the first position of the chapter in the first perceptual feature representation and the second perceptual feature representation The first chapter as source information represents the vector and the second chapter as quoted information represents the vector;

[0126] S6, constructing a sentence semantic representation pair according to the sentence representation vector of each sentence in the source information and the sentence representation vector of each sentence in the quoted information

[0127] S7, using a fully connected neural network according to the sentence semantic representation pair, obtaining a semantic association score of the first sentence pair from the source media to the referenced media and, the semantic relevance score of the second sentence pair from the source media to the source media

[0128]

[0129]

[0130] in, is the semantic association score of the first sentence pair between the i-th sentence in the source information and the j-th sentence in the quoted information; is the semantic association score of the second sentence pair between the jth sentence in the quoted information and the ith sentence in the source information; is the sentence representation vector of each sentence of the source information; is the sentence representation vector of each sentence of the quoted information; W1, b1 are the learnable parameters in the fully connected network;

[0131] S8, constructing a first sentence pair semantic association score matrix S according to the semantic association scores of each sentence pair s→r ∈R m×k and the second sentence pair semantic association score matrix S r→s ∈R k×m ;

[0132] S9, respectively, the semantic association score matrix S of the first sentence pair s→r ∈R m×k and the second sentence pair semantic association score matrix S r→s ∈R k×m Perform maximum pooling to obtain the bidirectional maximum correlation vector between sentence pairs, and obtain the first maximum correlation vector V between sentence pairs. s→r ∈R m and the second maximum relevance vector V r→s ∈R k ;

[0133] Among them, V s→r =max_pooling(S s→r );V r→s =max_pooling(S r→s );

[0134] For the maximum correlation vector V in the direction of "source information → quoted information" s→r , each dimension value represents the source media release information n a Each sentence in the quote is compared to the media release information n b The maximum correlation mined from all sentences in . Similarly, the maximum correlation vector V in the opposite direction r→s Each dimension value in represents the value of n b Each sentence is compared to n a The maximum relevance of all sentences in .

[0135] S10, respectively normalizing the first maximum relevance vector and the second maximum relevance vector through a softmax layer, and using the normalization as the contribution weight of each clause to the text semantics, to obtain a first set of weights a of the contribution of each clause in the source information and the quoted information to the text semantics s→r and the second set of weights a r→s ;

[0136] Among them, a s→r =softmax(V s→r );a r→s=softmax(V r→s );

[0137] S11, using the first set of weights a s→r The vector set representing the first sentence Perform weighted calculation to obtain the first interactive semantic chapter representation vector from source information to quoted information Using the second set of weights a r→s The vector set for the second sentence is Perform weighted calculation to obtain the second interactive semantic chapter representation vector of the quoted information to the source information

[0138] in,

[0139]

[0140] S12, according to the following formula, the first chapter representation vector and the first interactive semantic chapter representation vector, the second chapter representation vector and the second interactive semantic chapter representation vector are merged to obtain a first enhanced semantic representation vector of the source information and a second enhanced semantic representation vector of the quoted information:

[0141]

[0142]

[0143] Among them, f a is the first enhanced semantic representation vector; f b is the second enhanced semantic representation vector; Represent the vector for the first chapter; Represent vectors for the second chapter; Represents the vector for the first interactive semantic chapter; Represent the vector for the second interactive semantic chapter;

[0144] S13, inputting the first enhanced semantic representation vector and the second enhanced semantic representation vector into a fully connected neural network with softmax as a nonlinear activation function to perform media recitation type identification to obtain a media recitation type;

[0145] The calculation process is

[0146] The training goal is to minimize the training set The cross entropy loss function of

[0147] The loss function is

[0148] f ais the first enhanced semantic representation vector; f b is the second enhanced semantic representation vector; N and c represent the number of training set samples and citation types respectively; y i and are the true and predicted label vectors of the i-th sample data respectively.

[0149] This solution uses the pre-trained language model BERT as a text encoder. Its strong semantic encoding ability can generate the semantic representation of the media release information and the semantic representation vector of each sentence that is suitable for the field of media citation. Sentence-level semantic association information is modeled from the two directions of "source information → citation information" and "citation information → source information". Multi-level and fine-grained modeling can reflect the media semantic features of the potential characteristics of the citation type, thereby enhancing the semantic representation ability of the chapter and effectively improving the recognition performance of the media citation type.

[0150] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0151] In this embodiment, a device for identifying the type of media transfer is also provided, which is used to implement the above-mentioned embodiments and preferred implementation modes, and the descriptions that have been made will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceivable.

[0152] Figure 3 is a device for identifying media referencing types according to an embodiment of the present invention, such as Figure 3 As shown, the device includes: a text encoding module 1000, a semantic information learning module 2000, a semantic fusion module 3000, and a citation type identification module 4000, wherein:

[0153] The text encoding module 1000 is used to input the source information published by the source media and the quoted information published by the quoted media into the pre-trained language model to obtain the first chapter representation vector of the source information and the first sentence representation vector set of the source information, the second chapter representation vector of the quoted information and the second sentence representation vector set of the quoted information, wherein the first sentence representation vector set includes the sentence representation vectors of each sentence in the source information, and the second sentence representation vector set includes the sentence representation vectors of each sentence in the quoted information;

[0154] A semantic information learning module 2000 is used to perform bidirectional interactive semantic information learning based on the first sentence representation vector set and the second sentence representation vector set to obtain a first interactive semantic chapter representation vector from source information to quoted information and a second interactive semantic chapter representation vector from quoted information to source information;

[0155] A semantic fusion module 3000 is used to fuse the first chapter representation vector and the first interactive semantic chapter representation vector, and the second chapter representation vector and the second interactive semantic chapter representation vector to obtain a first enhanced semantic representation vector of source information and a second enhanced semantic representation vector of quoted information;

[0156] The reference type identification module 4000 is used to identify the media reference type according to the first enhanced semantic representation vector and the second enhanced semantic representation vector to obtain the media reference type.

[0157] Optionally, the text encoding module includes: a sequence pair construction unit, used to construct a media citation data sequence pair using the source information published by the source media and the cited information published by the cited media; a perceptual feature output unit, used to input the media citation data sequence pair into a pre-trained language model, and output a first perceptual feature representation of the source information and a second perceptual feature representation of the cited information; a sentence representation vector processing unit, used to obtain a sentence representation vector of each sentence of the source information and a sentence representation vector of each sentence of the cited information based on the sequence position latent vector of each sentence in the first perceptual feature representation and the second perceptual feature representation; a chapter representation vector processing unit, used to output the latent vector of the first position of the chapter in the first perceptual feature representation and the second perceptual feature representation as the first chapter representation vector of the source information and the second chapter representation vector of the cited information.

[0158] Optionally, the sequence pair construction unit includes: an information pair construction unit, used to construct a media reference information pair based on the source information and the reference information; an identifier adding unit, used to add interval identifiers to the beginning position and inter-sentence position of the source information and the reference information in the media reference information pair to obtain a media reference data sequence pair.

[0159] Optionally, the sentence representation vector processing unit includes: a first calculation unit, used to calculate a first average value of the serial position latent vector of each sentence in the first perceptual feature representation; and determine the first average value as the sentence representation vector of each sentence of the source information; a second calculation unit, used to calculate a second average value of the serial position latent vector of each sentence in the second perceptual feature representation; and determine the second average value as the sentence representation vector of each sentence of the quoted information.

[0160] Optionally, the semantic information learning module includes: a representation pair construction unit, which is used to construct a first sentence semantic representation pair from the source media to the redirected media direction and a second sentence semantic representation pair from the redirected media to the source media direction according to the first sentence representation vector set and the second sentence representation vector set; an association score processing unit, which is used to input the first sentence semantic representation pair into a fully connected neural network to obtain a first sentence pair semantic association score matrix from the source media to the redirected media direction; input the second sentence semantic representation pair into a fully connected neural network to obtain a second sentence pair semantic association score matrix from the redirected media to the source media direction; and a pooling processing unit, which is used to perform maximum pooling processing on the first sentence pair semantic association score matrix to obtain a second sentence pair semantic association score matrix from the source media to the redirected media direction. a maximum correlation vector; performing maximum pooling processing on the semantic correlation score matrix of the second sentence pair to obtain the second maximum correlation vector between the sentence pairs in the direction of the quoted media to the source media; a normalization processing unit, used to perform normalization processing on the first maximum correlation vector and the second maximum correlation vector respectively to obtain a first set of weights and a second set of weights for the contribution of each clause in the source information and the quoted information to the text semantics; a weighted processing unit, used to perform weighted calculation on the first sentence representation vector set using the first set of weights to obtain a first interactive semantic chapter representation vector from the source information to the quoted information; and perform weighted calculation on the second sentence representation vector set using the second set of weights to obtain a second interactive semantic chapter representation vector from the quoted information to the source information.

[0161] Optionally, the citation type identification module includes: a media citation type output unit, used to take the first enhanced semantic representation vector and the second enhanced semantic representation vector as input data, input them into a fully connected neural network with softmax as a nonlinear activation function, and output the media citation type.

[0162] Optionally, the media citation type output module includes: a first output unit, used to output a media citation type label; wherein the media citation type label is used to characterize the citation type, and the citation type includes: word variation, sentence rewriting, and paragraph restatement.

[0163] It should be noted that the above modules can be implemented by software or hardware. For the latter, it can be implemented in the following ways, but not limited to: the above modules are all located in the same processor; or the above modules are located in different processors in any combination.

[0164] An embodiment of the present invention further provides a storage medium, in which a computer program is stored, wherein the computer program is configured to execute the steps of any of the above method embodiments when running.

[0165] Optionally, in this embodiment, the above-mentioned storage medium may include but is not limited to: a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and other media that can store computer programs.

[0166] An embodiment of the present invention further provides an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0167] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0168] Optionally, in this embodiment, the processor may be configured to execute the steps in any of the above method embodiments through a computer program.

[0169] Optionally, the specific examples in this embodiment may refer to the examples described in the above embodiments and optional implementation modes, and this embodiment will not be described in detail here.

[0170] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0171] In the above embodiments of the present application, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0172] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0173] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0174] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0175] If 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. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art 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, including a number of instructions to enable a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or optical disk and other media that can store program codes.

[0176] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for identifying media citation types, characterized in that: The method comprises: Inputting the source information published by the source media and the quoted information published by the quoted media into the pre-trained language model, obtaining a first chapter representation vector of the source information and a first sentence representation vector set of the source information, a second chapter representation vector of the quoted information and a second sentence representation vector set of the quoted information, wherein the first sentence representation vector set includes sentence representation vectors of each sentence in the source information, and the second sentence representation vector set includes sentence representation vectors of each sentence in the quoted information; Performing bidirectional interactive semantic information learning based on the first sentence representation vector set and the second sentence representation vector set to obtain a first interactive semantic chapter representation vector from source information to quoted information and a second interactive semantic chapter representation vector from quoted information to source information; The first chapter representation vector and the first interactive semantic chapter representation vector are merged, and the second chapter representation vector and the second interactive semantic chapter representation vector are merged to obtain a first enhanced semantic representation vector of source information and a second enhanced semantic representation vector of quoted information; Perform media transfer type identification according to the first enhanced semantic representation vector and the second enhanced semantic representation vector to obtain a media transfer type; The method of performing bidirectional interactive semantic information learning based on the first sentence representation vector set and the second sentence representation vector set to obtain a first interactive semantic chapter representation vector from source information to quoted information and a second interactive semantic chapter representation vector from quoted information to source information includes: Constructing a first sentence semantic representation pair from source media to redirected media and a second sentence semantic representation pair from redirected media to source media according to the first sentence representation vector set and the second sentence representation vector set; Input the first sentence semantic representation pair into a fully connected neural network to obtain a first sentence pair semantic association score matrix from source media to redirected media; input the second sentence semantic representation pair into a fully connected neural network to obtain a second sentence pair semantic association score matrix from redirected media to source media; Performing maximum pooling processing on the semantic association score matrix of the first sentence pair to obtain a first maximum association vector between the sentence pairs in the direction of the source media to the transferred media; performing maximum pooling processing on the semantic association score matrix of the second sentence pair to obtain a second maximum association vector between the sentence pairs in the direction of the transferred media to the source media; Normalizing the first maximum relevance vector and the second maximum relevance vector respectively to obtain a first set of weights and a second set of weights of contributions of each clause in the source information and the quoted information to the text semantics; The first set of weights is used to perform weighted calculation on the first set of sentence representation vectors to obtain a first interactive semantic chapter representation vector from source information to quoted information; the second set of weights is used to perform weighted calculation on the second set of sentence representation vectors to obtain a second interactive semantic chapter representation vector from quoted information to source information.

2. The method for identifying media referencing types according to claim 1, wherein: The method of inputting the source information published by the source media and the quoted information published by the quoted media into the pre-trained language model to obtain the first chapter representation vector of the source information and the first sentence representation vector set of the source information, the second chapter representation vector of the quoted information and the second sentence representation vector set of the quoted information includes: Construct a media reference data sequence pair using source information published by source media and reference information published by reference media; Inputting the media citation data sequence into a pre-trained language model, and outputting a first perceptual feature representation of the source information and a second perceptual feature representation of the citation information; Obtaining sentence representation vectors of each sentence of source information and sentence representation vectors of each sentence of quoted information according to the sequence position latent vectors of each sentence in the first perceptual feature representation and the second perceptual feature representation; The latent vector output of the first position of the chapter in the first perceptual feature representation and the second perceptual feature representation is used as the first chapter representation vector of the source information and the second chapter representation vector of the quoted information.

3. The method for identifying media referencing types according to claim 2, wherein: The method of constructing a media referencing data sequence pair from the source information published by the source media and the referencing information published by the referencing media includes: Constructing a media referencing information pair according to the source information and the referencing information; An interval identifier is added to the text beginning position and the sentence position of the source information and the quoted information in the media quoted information pair to obtain a media quoted data sequence pair.

4. The method for identifying media referencing types according to claim 2, wherein: The step of obtaining the sentence representation vector of each sentence of the source information and the sentence representation vector of each sentence of the quoted information according to the sequence position latent vector of each sentence in the first perceptual feature representation and the second perceptual feature representation includes: Calculating a first average value of the sequence position latent vectors of each sentence in the first perceptual feature representation; calculating a second average value of the sequence position latent vectors of each sentence in the second perceptual feature representation; The first average value is determined as a sentence representation vector of each sentence of the source information, and the second average value is determined as a sentence representation vector of each sentence of the quoted information.

5. The method for identifying media referencing types according to claim 1, wherein: The step of identifying the media referencing type according to the first enhanced semantic representation vector and the second enhanced semantic representation vector to obtain the media referencing type includes: The first enhanced semantic representation vector and the second enhanced semantic representation vector are used as input data, input into a fully connected neural network with softmax as a nonlinear activation function, and the media citation type is output.

6. The method for identifying media referencing types according to claim 5, characterized in that: Output media citation types include: Output media citation type label; The media quotation type tag is used to characterize the quotation type, and the quotation type includes: word variation, sentence rewriting, and paragraph restatement.

7. A device for identifying media referencing types, characterized in that: include: A text encoding module, used to input the source information published by the source media and the quoted information published by the quoted media into a pre-trained language model, to obtain a first chapter representation vector of the source information and a first sentence representation vector set of the source information, a second chapter representation vector of the quoted information and a second sentence representation vector set of the quoted information, wherein the first sentence representation vector set includes sentence representation vectors of each sentence in the source information, and the second sentence representation vector set includes sentence representation vectors of each sentence in the quoted information; A semantic information learning module, configured to perform bidirectional interactive semantic information learning based on the first sentence representation vector set and the second sentence representation vector set to obtain a first interactive semantic chapter representation vector from source information to quoted information and a second interactive semantic chapter representation vector from quoted information to source information; A semantic fusion module, used for fusing the first chapter representation vector and the first interactive semantic chapter representation vector, the second chapter representation vector and the second interactive semantic chapter representation vector, to obtain a first enhanced semantic representation vector of source information and a second enhanced semantic representation vector of quoted information; a reference type identification module, configured to identify a media reference type according to the first enhanced semantic representation vector and the second enhanced semantic representation vector to obtain a media reference type; The semantic information learning module includes: a representation pair construction unit, which is used to construct a first sentence semantic representation pair from the source media to the redirected media and a second sentence semantic representation pair from the redirected media to the source media according to the first sentence representation vector set and the second sentence representation vector set; an association score processing unit, which is used to input the first sentence semantic representation pair into a fully connected neural network to obtain a first sentence pair semantic association score matrix from the source media to the redirected media; input the second sentence semantic representation pair into a fully connected neural network to obtain a second sentence pair semantic association score matrix from the redirected media to the source media; and a pooling processing unit, which is used to perform maximum pooling processing on the first sentence pair semantic association score matrix to obtain a first maximum score matrix between sentence pairs from the source media to the redirected media. A maximum correlation vector; performing maximum pooling processing on the semantic correlation score matrix of the second sentence pair to obtain the second maximum correlation vector between the sentence pairs in the direction of the quoted media to the source media; a normalization processing unit, used to perform normalization processing on the first maximum correlation vector and the second maximum correlation vector respectively to obtain a first set of weights and a second set of weights for the contribution of each clause in the source information and the quoted information to the text semantics; a weighted processing unit, used to use the first set of weights to perform weighted calculation on the first sentence representation vector set to obtain a first interactive semantic chapter representation vector from the source information to the quoted information; and use the second set of weights to perform weighted calculation on the second sentence representation vector set to obtain a second interactive semantic chapter representation vector from the quoted information to the source information.

8. An electronic device, characterized in that: The electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method for identifying the media transfer type as claimed in any one of claims 1 to 6.

9. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for identifying the media transfer type according to any one of claims 1 to 6.

Images