BERT-based media information viewpoint extraction method, apparatus and device, and storage medium

Abstract

The invention discloses a BERT-based media information viewpoint extraction method, device and equipment and a storage medium, and the method comprises the steps: S1, inputting a to-be-processed media corpus, and recognizing various types of viewpoint holders in corpus information, the various types of viewpoint holders comprising named entities, pronouns and anaphora phrases; s2, performing semantic association processing on the recognized viewpoint holder and corpus information to obtain associated input data; s3, inputting the associated input data into a trained viewpoint extraction model, wherein the viewpoint extraction model is constructed based on a BERT model, an attention mechanism layer and a viewpoint classification layer; the method has the beneficial effects that the comprehensiveness of viewpoint holder recognition is greatly improved by recognizing various types of viewpoint holders, and the semantic relation between the viewpoint holders and corpus information can be deeply mined by adopting a word vector representation method and a semantic similarity calculation method, so that the viewpoint information is extracted more accurately.

Description

Technical Field

[0001] This invention relates to the field of opinion extraction technology, specifically to a method, apparatus, device, and storage medium for media information opinion extraction based on BERT. Background Art

[0002] In today's digital information age, media information is experiencing explosive growth. Various news reports, social media updates, forum discussions, and other media content contain a wealth of viewpoints, which are of crucial value to many fields, including market research, public opinion analysis, policy making, and business decision-making. Accurately and efficiently extracting viewpoints from media information can help relevant personnel quickly grasp public attitudes, market trends, and social hotspots, thereby making more scientific and rational decisions.

[0003] Opinion extraction refers to extracting textual content that expresses personal opinions, feelings, or beliefs from a given text. The extracted opinion information includes at least: the opinion holder, the topic, and the statement. Specifically, the opinion holder expresses an opinion on a particular topic, and this opinion may contain a certain degree of emotional bias.

[0004] Traditional opinion extraction methods primarily rely on manual annotation and rule matching. While manual annotation ensures high accuracy, it requires significant manpower and time, and the results are easily influenced by the subjective factors of the annotators, making it difficult to meet the real-time requirements of large-scale media information processing. Rule matching methods identify opinion information based on pre-defined grammatical rules or keyword lists. This method is effective for simple and standardized opinion expressions, but it falls short for complex, varied, and implicit opinion expressions, easily leading to misjudgments and omissions. With the rapid development of deep learning technology, neural network-based opinion extraction methods have gradually become a research hotspot. Among them, opinion extraction methods based on the BERT (Bidirectional Encoder Representations from Transformers) model have attracted much attention due to their powerful language representation capabilities and contextual understanding abilities.

[0005] In existing technologies, some BERT-based media information opinion extraction methods mainly focus on identifying named entities in corpus information and treating these named entities as opinion holders. The combination of named entities and corpus information is then input into a trained opinion extraction model. This model extracts opinion information corresponding to named entities from the corpus information based on the BERT model and a softmax layer. This method reduces the workload of manual opinion extraction to some extent and improves the accuracy of opinion extraction.

[0006] However, existing technologies still have many shortcomings. First, in terms of opinion holder identification, only named entities are considered as opinion holders, ignoring other subjects that may express opinions, such as pronouns ("he," "she," "it," etc.) and referential phrases ("the company," "the team," etc.). In actual media information, these pronouns and referential phrases are often used to refer to the aforementioned entities and express opinions. This limitation of existing technologies can lead to the omission of some opinion holder information, thus affecting the comprehensiveness of opinion extraction.

[0007] Secondly, regarding the data input method, simply combining named entities and corpus information into the model fails to fully consider the semantic relationships and contextual information between the two. Opinion expression in media information often relies on specific contexts and semantic relationships. The current simple combination method may result in the model's insufficient understanding of opinion information, failing to accurately capture the intrinsic connection between the opinion holder and the opinion, thus affecting the accuracy of opinion extraction.

[0008] Furthermore, existing technologies have relatively simple model structures, relying solely on the BERT model and softmax layers to construct opinion extraction models, lacking the ability to classify and process opinion information in a more detailed manner. In practical applications, opinion information has various types and forms of expression, such as positive, negative, and neutral opinions, as well as opinions at different granularities (e.g., opinions on product features, opinions on brand image, etc.). The simple model structures of existing technologies are insufficient to handle these complex and varied forms of opinion expression, and cannot conduct deeper mining and analysis of opinion information.

[0009] Furthermore, existing technologies do not adequately consider the contextual information in the corpus. Opinion expressions in media information are often ambiguous; the same word or sentence may convey different meanings in different contexts. Existing technologies lack effective modeling of contextual information, and for these ambiguous opinion expressions, they may be unable to accurately determine their true meaning, leading to errors in opinion extraction.

[0010] Existing technologies also suffer from insufficient processing capabilities for long text corpora. Long texts contain a large amount of information, and key viewpoints may be scattered throughout the text. Existing technologies may not be able to effectively capture this key information, easily leading to information loss or confusion, thus affecting the effectiveness of viewpoint extraction.

[0011] Meanwhile, existing technologies lack domain adaptability. Media information from different domains has different language styles, technical terms, and ways of expressing opinions. Existing technologies have not been specifically optimized for media information from different domains, and the effectiveness of opinion extraction may vary significantly across different domains, failing to meet the needs of cross-domain applications.

[0012] Furthermore, existing technologies lack dynamic update mechanisms. As media information is constantly updated and changed, new forms of opinion expression and linguistic phenomena emerge continuously. Once the models in existing technologies are trained, they cannot be dynamically updated based on new corpus information or forms of opinion expression, leading to a gradual decline in model performance and an inability to adapt to the rapidly changing media environment.

[0013] Furthermore, existing technologies only extract opinion information without analyzing the sentiment bias of those opinions. In practical applications, understanding the sentiment bias (positive, negative, neutral) of opinions is crucial for gaining a deeper understanding of public attitudes and market trends. This limitation of existing technologies fails to meet the needs for more in-depth sentiment analysis.

[0014] In summary, existing BERT-based media information opinion extraction methods have many shortcomings in terms of opinion holder identification, data input methods, model structure, contextual understanding, long text processing, domain adaptability, dynamic updates, and sentiment analysis. They are difficult to meet the requirements of comprehensiveness, accuracy, and efficiency in opinion extraction in practical applications. Therefore, we propose a BERT-based media information opinion extraction method, device, equipment, and storage medium. Summary of the Invention

[0015] The purpose of this invention is to provide a method, apparatus, device, and storage medium for extracting media information viewpoints based on BERT, in order to solve the problems mentioned in the background art.

[0016] To achieve the above objectives, the BERT-based media information opinion extraction method includes the following steps:

[0017] S1. Input the media corpus to be processed, and identify multiple types of opinion holders in the corpus information. These multiple types of opinion holders include named entities, pronouns, and referential phrases.

[0018] S2. Perform semantic association processing on the identified opinion holders and corpus information to obtain the associated input data;

[0019] S3. Input the associated input data into the trained opinion extraction model. The opinion extraction model is built based on the BERT model, the attention mechanism layer, and the opinion classification layer.

[0020] S4. The opinion extraction model extracts opinion information corresponding to opinion holders from the corpus information based on the associated input data, and judges the sentiment tendency of the opinion information.

[0021] Preferably, the step of semantically associating the identified opinion holders with the corpus information to obtain associated input data includes: converting the opinion holders and corpus information into vector form using a word vector representation method; calculating the semantic similarity between the opinion holder vector and the corpus information vector using a semantic similarity calculation method; and associating the opinion holders and corpus information based on the semantic similarity to obtain associated input data.

[0022] Preferably, the training process of the opinion extraction model is as follows: collecting a media information corpus containing opinion holders, opinion information, and sentiment tags; preprocessing the data in the media information corpus according to the semantic association processing method in step S2 to obtain training data; using the training data to train the initial model built based on the BERT model, attention mechanism layer, and opinion classification layer to obtain the trained opinion extraction model.

[0023] Preferably, the sentiment judgment of opinion information includes: adding a sentiment analysis layer after the opinion classification layer; the sentiment analysis layer judges the sentiment of opinion information based on a trained sentiment classification model, and the sentiment classification model is trained using a corpus of opinion information containing sentiment labels.

[0024] As a preferred embodiment, the method further includes: establishing a dynamic update mechanism, regularly collecting new media information corpus, and preprocessing it according to the semantic association processing method in step S2; and using the preprocessed new corpus to incrementally train the opinion extraction model.

[0025] As a preferred approach, for long text corpora, a segmentation and attention mechanism is used to extract viewpoints. Specifically, this includes: segmenting the long text corpus into segments of a certain length; identifying viewpoint holders and performing semantic association processing on each segment; and using an attention mechanism to fuse the processing results of each segment to focus on key viewpoint information.

[0026] As a preferred approach, the opinion extraction model is optimized using a domain-adaptive optimization method for media information in different domains. Specifically, this includes: collecting media information corpora from different domains; and fine-tuning the opinion extraction model using domain-specific corpora.

[0027] A media information opinion extraction device based on BERT includes an opinion holder identification module, a semantic association processing module, and an opinion extraction model module.

[0028] The opinion holder identification module is used to identify multiple types of opinion holders in the corpus information, including named entities, pronouns, and referential phrases;

[0029] The semantic association processing module is used to perform semantic association processing on the identified opinion holders and corpus information to obtain the associated input data.

[0030] The opinion extraction model module is used to receive the associated input data, extract opinion information corresponding to opinion holders based on the BERT model, attention mechanism layer and opinion classification layer, and make sentiment judgment on the opinion information.

[0031] A BERT-based media information viewpoint extraction device includes a processor and a memory for storing a computer program, which, when executed by the processor, implements the BERT-based media information viewpoint extraction method described above.

[0032] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the BERT-based media information viewpoint extraction method described above.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. This invention significantly improves the comprehensiveness of opinion holder identification by recognizing various types of opinion holders, including named entities, pronouns, and referential phrases. In previous media information, many opinions were not directly expressed by named entities, but rather by pronouns or referential phrases that referred to the entities mentioned above and elaborated on the opinions. This invention can accurately identify these pronouns and referential phrases as opinion holders, ensuring that more opinion information can be effectively extracted and avoiding the loss of opinion information due to the omission of opinion holders, thus providing a richer data foundation for subsequent opinion analysis and decision-making.

[0035] 2. This invention, by employing word vector representation and semantic similarity calculation methods, can deeply explore the semantic connections between opinion holders and corpus information. In complex media contexts, the relationship between opinion holders and opinions is not a simple combination of words, but rather a deep semantic association. For example, in the sentence "This expert pointed out that the technology has great potential, but there are still some challenges," through semantic association processing, the model can more accurately understand the semantic relationship between "this expert" and subsequent opinions, thereby extracting opinion information more precisely. This semantic association processing method enables the model to better understand the context, reduces opinion extraction errors caused by semantic ambiguity, and improves the accuracy and reliability of opinion extraction.

[0036] 3. This invention presents an opinion extraction model based on the BERT model, an attention mechanism layer, and an opinion classification layer. Compared to the simpler model structures in existing technologies, this model possesses significantly enhanced opinion information processing capabilities. The attention mechanism layer enables the model to focus on key parts of the corpus information, highlighting information relevant to opinion expression while ignoring irrelevant information. For example, in long texts, the attention mechanism can automatically identify key sentences and words related to the opinion holder, thereby improving the efficiency of opinion extraction. The opinion classification layer can perform more detailed classification of the extracted opinion information, such as distinguishing between positive, negative, and neutral opinions, and classifying opinions at different granularities. This improved model structure can better handle complex and varied forms of opinion expression, enabling deeper mining and analysis of opinion information, and providing richer and more valuable opinion information for applications in different fields.

[0037] 4. The dynamic update mechanism of this invention can periodically collect new media information corpora and incrementally train the model, enabling it to adapt to new forms of opinion expression and linguistic phenomena in a timely manner. In a rapidly changing media environment, new vocabulary, expressions, and viewpoints are constantly emerging, and the dynamic update mechanism ensures that the model maintains high performance. The domain-adaptive optimization method is specifically optimized for media information in different domains, improving the model's viewpoint extraction performance in different domains.

[0038] 5. This invention performs sentiment analysis on opinion information and employs techniques such as model compression and distributed computing to optimize computational resource consumption, meeting diverse user needs. Sentiment analysis can further uncover the emotional attitudes behind opinion information, providing users with deeper emotional insights. For example, in market research, understanding consumers' positive or negative emotional tendencies towards products helps companies develop more effective marketing strategies. Simultaneously, by optimizing computational resource consumption, the cost of processing large-scale corpora is reduced, improving processing efficiency. When processing massive amounts of media information, computational resource optimization ensures that the model completes the opinion extraction task within a limited time, meeting the real-time requirements of practical applications. Attached Figure Description

[0039] Figure 1 This is a flowchart of the method of the present invention;

[0040] Figure 2 This is a structural diagram of the media information opinion extraction device of the present invention;

[0041] Figure 3 This is a structural diagram of the media information opinion extraction device of the present invention. DETAILED DESCRIPTION

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0043] Please see Figure 1 This invention provides a technical solution: a media information opinion extraction method based on BERT, comprising the following steps:

[0044] S1. Input the media corpus to be processed, and identify various types of opinion holders in the corpus information. These various types of opinion holders include named entities, pronouns, and referential phrases.

[0045] It should be noted that users input the media corpus to be processed into the system. This corpus can come from various channels such as news websites, social media platforms, and forums. The system will perform a preliminary check on the input corpus to ensure that the format is correct and readable.

[0046] S2. Semantically associate the identified opinion holders with the corpus information to obtain the associated input data.

[0047] It should be noted that existing named entity recognition tools (such as StanfordNER and spaCy) are used to perform named entity recognition on the corpus information, identifying named entities such as person names, place names, and organization names as potential opinion holders. For example, in the sentence "Apple launched a new iPhone, which received widespread attention from consumers," "Apple" is identified as a named entity. A combination of rule-based and machine learning methods is used to identify pronouns (such as "he," "she," "it," "they," etc.) and referential phrases (such as "the company," "the team," "the aforementioned products," etc.). By analyzing the context, the entities referred to by pronouns and referential phrases are determined. For example, in the sentence "Zhang San is a programmer, and he recently completed an important project," "he" is identified as referring to "Zhang San."

[0048] S3. Input the associated input data into the trained opinion extraction model. The opinion extraction model is built based on the BERT model, the attention mechanism layer, and the opinion classification layer. The training process of the opinion extraction model is as follows: collect a media information corpus containing opinion holders, opinion information, and sentiment labels; preprocess the data in the media information corpus according to the semantic association processing method in step S2 to obtain training data; use the training data to train the initial model built based on the BERT model, the attention mechanism layer, and the opinion classification layer to obtain the trained opinion extraction model.

[0049] It's important to note that pre-trained word vector models (such as Word2Vec and GloVe) are used to convert the identified opinion holders and corpus information into vector form. Each word is mapped to a fixed-dimensional vector, which captures the semantic information of the word. For example, "Apple Inc." and "new iPhone" are converted into corresponding word vectors. Semantic similarity calculation methods, such as cosine similarity, are used to calculate the semantic similarity between the opinion holder vector and the corpus information vector. By comparing the angle between the vectors, the degree of semantic association between the opinion holder and different parts of the corpus information is determined. For example, the cosine similarity between the "Apple Inc." vector and the vectors of each word in the sentence is calculated to find the part most semantically relevant to "Apple Inc." Based on the semantic similarity calculation results, the opinion holder and the part most semantically related to the corpus information are combined to generate the associated input data. For example, "Apple Inc." is combined with sentence fragments containing related opinions.

[0050] S4. The opinion extraction model extracts opinion information corresponding to opinion holders from the corpus information based on the associated input data, and performs sentiment judgment on the opinion information. The sentiment judgment of opinion information includes: adding a sentiment analysis layer after the opinion classification layer; the sentiment analysis layer performs sentiment judgment on opinion information based on the trained sentiment classification model, and the sentiment classification model is trained using a corpus of opinion information containing sentiment labels.

[0051] It should be noted that when building the model, a suitable pre-trained BERT model (such as BERT-base, BERT-large, etc.) is selected as the base model. An attention mechanism layer is added after the BERT model to focus on key information related to the expression of opinions. The attention mechanism layer can automatically learn the importance weights of different parts, making the model pay more attention to important words and sentences. After the attention mechanism layer, an opinion classification layer is added to classify the extracted opinion information. The opinion classification layer can use a fully connected neural network and a softmax function to divide the opinion information into different categories, such as positive opinions, negative opinions, and neutral opinions. After the opinion classification layer, a sentiment analysis layer is added to judge the sentiment of the opinion information. The sentiment analysis layer can use a pre-trained sentiment classification model, which can be trained using a corpus of opinion information containing sentiment labels.

[0052] During model training, a media information corpus containing opinion holders, opinion information, and sentiment labels is collected. The data in the collected corpus is preprocessed according to the semantic association processing method in the data preprocessing module to obtain training data. The preprocessed data is divided into training set, validation set, and test set for model training, validation, and testing. The initial model is trained using the training data. During training, the backpropagation algorithm and optimizer (such as the Adam optimizer) are used to adjust the model parameters so that the model can minimize the loss function. Through multiple iterations of training, the model's performance on the validation set reaches its optimum. The trained model is evaluated using the test set, and the model's accuracy, recall, F1 score, and other metrics are calculated to evaluate the model's opinion extraction and sentiment judgment performance. If the model's performance does not meet the requirements, the model structure, hyperparameters, etc., can be adjusted, and the training and evaluation can be repeated.

[0053] S5. Establish a dynamic update mechanism to regularly collect new media information corpus and preprocess it according to the semantic association processing method in step S2; use the preprocessed new corpus to incrementally train the opinion extraction model.

[0054] It's important to note that regularly collecting new media information corpora is crucial. These corpora reflect the latest forms of opinion expression and linguistic phenomena. For example, with the development of social media, new online buzzwords and expressions may emerge, requiring timely collection of relevant corpora. The newly collected corpora are preprocessed according to the semantic association processing methods in the data preprocessing module to obtain incremental training data. The preprocessed new corpora are then used to incrementally train the opinion extraction model, updating its parameters. Incremental training allows the model to adapt to new data without retraining the entire model, maintaining its performance.

[0055] The identified opinion holders and corpus information are semantically associated to obtain associated input data, including: converting opinion holders and corpus information into vector form using word vector representation; calculating the semantic similarity between opinion holder vectors and corpus information vectors using semantic similarity calculation methods; and associating opinion holders and corpus information based on semantic similarity to obtain associated input data.

[0056] It should be noted that in semantic association processing, pre-trained word vector models (such as Word2Vec, GloVe, etc.) are used to convert the identified opinion holders and corpus information into vector form. Each word is mapped to a fixed-dimensional vector, and these vectors can capture the semantic information of the word. Semantic similarity calculation methods such as cosine similarity are used to calculate the semantic similarity between the opinion holder vector and the corpus information vector. By comparing the angle between the vectors, the degree of semantic association between the opinion holder and different parts of the corpus information is determined. Based on the semantic similarity calculation results, the parts of the opinion holder and the corpus information with the closest semantic association are combined to generate the associated input data.

[0057] For long text corpora, a segmentation and attention mechanism is used to extract viewpoints. Specifically, the long text corpora are segmented into segments of a certain length; viewpoint holder identification and semantic association are performed on each segment; and the processing results of each segment are fused using an attention mechanism to focus on key viewpoint information.

[0058] It should be noted that for long text corpora, segmentation is performed based on text length or semantic coherence. During segmentation, tagged information is added to each segment to record its position and order in the original text, so that the original text's context can be restored during subsequent processing and result integration. For each segment, an opinion holder identification module is used to identify named entities (such as personal names, place names, and organization names), pronouns (such as "he," "she," "it," "they," etc.), and referential phrases (such as "the company," "the team," "the aforementioned products," etc.). By analyzing the context, the entities referred to by pronouns and referential phrases are determined. Finally, a semantic association processing module is used to process each segment... The process involves semantically associating the identified opinion holders with the corpus information within each segment. Word vectors are used to convert both the opinion holders and the corpus information into vector form. Then, semantic similarity calculations, such as cosine similarity, are employed to identify the most semantically related parts between the opinion holders and the corpus information, generating the associated segmented input data. This associated input data is then fed into an opinion extraction model that includes an attention mechanism layer. The attention mechanism layer automatically learns the importance weights of each segment, focusing on segments containing key opinion information. Based on the weights of each segment, the key opinion information from each segment is fused to obtain the opinion extraction result for the entire long text.

[0059] For media information from different fields, a domain-adaptive optimization method is used to optimize the opinion extraction model. Specifically, this includes: collecting media information corpora from different fields; and fine-tuning the opinion extraction model using domain-specific corpora.

[0060] It should be noted that the process involves collecting media information corpora from different fields such as finance, healthcare, technology, entertainment, and education; labeling the collected domain corpora with opinions, including opinion holders, opinion information, and sentiment tags; fine-tuning the BERT-based opinion extraction model using domain-specific corpora; evaluating the fine-tuned model using methods such as cross-validation; and continuously adjusting the fine-tuning strategy and parameters based on the evaluation results until the model achieves optimal performance on the domain-specific corpora.

[0061] Please see Figure 2 A media information opinion extraction device based on BERT includes an opinion holder identification module, a semantic association processing module, and an opinion extraction model module.

[0062] The opinion holder identification module is used to identify various types of opinion holders in the corpus information, including named entities, pronouns, and referential phrases; the semantic association processing module is used to perform semantic association processing between the identified opinion holders and the corpus information to obtain the associated input data; the opinion extraction model module is used to receive the associated input data and extract the opinion information corresponding to the opinion holders based on the BERT model, attention mechanism layer, and opinion classification layer, as well as to judge the sentiment tendency of the opinion information.

[0063] It should be noted that the opinion holder identification module uses named entity recognition tools (such as StanfordNER, spaCy, etc.) and a pronoun and pronoun phrase recognition method that combines rule-based and machine learning to accurately identify various types of opinion holders in the corpus information; it provides an interface with the semantic association processing module to pass the identified opinion holder information to the semantic association processing module for subsequent processing; it uses pre-trained word vector models (such as Word2Vec, GloVe, etc.) to convert opinion holders and corpus information into vector form, and performs semantic association processing through semantic similarity calculation methods (such as cosine similarity) to obtain the associated input data; the associated input data is stored in an appropriate data structure so that the opinion extraction model module can read and process it efficiently.

[0064] Please see Figure 3 A BERT-based media information view extraction device includes a processor and a memory for storing a computer program, which, when executed by the processor, implements any of the above-mentioned BERT-based media information view extraction methods.

[0065] It should be noted that BERT-based media information opinion extraction devices require high-performance processors, such as Intel Xeon series or AMD EPYC series server processors, to ensure rapid processing of large-scale media information corpora; large-capacity storage, including high-speed cache (such as DDR4 memory) and large-capacity hard disks (such as SSDs or HDDs), for storing computer programs, corpora, and model parameters; and the installation of a stable operating system, such as Linux (such as Ubuntu, CentOS) or Windows Server, to provide a basic environment for the computer program to run. The computer program for the BERT-based media information opinion extraction method is stored in the memory, and the processor executes the program to implement functions such as long text segmentation processing, domain-adaptive optimization, opinion holder identification, semantic association processing, opinion extraction, and sentiment judgment.

[0066] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements any of the above-mentioned BERT-based media information viewpoint extraction methods.

[0067] It should be noted that the computer-readable storage medium can be a physical medium such as a hard drive, optical disc, USB flash drive, or solid-state drive (SSD), or it can be a virtual storage space in a network storage device (such as NAS or SAN). A computer program for extracting media information opinions based on BERT is written and stored on the computer-readable storage medium. When the processor accesses the storage medium and executes the computer program therein, it implements functions such as long text segmentation, domain-adaptive optimization, opinion holder identification, semantic association processing, opinion extraction, and sentiment judgment, providing users with media information opinion extraction services.

[0068] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A media information opinion extraction method based on BERT, characterized in that, The steps include: S1. Input the media corpus to be processed, and identify multiple types of opinion holders in the corpus information. These multiple types of opinion holders include named entities, pronouns, and referential phrases. S2. Perform semantic association processing on the identified opinion holders and corpus information to obtain the associated input data; S3. Input the associated input data into the trained opinion extraction model. The opinion extraction model is built based on the BERT model, the attention mechanism layer, and the opinion classification layer. S4. The opinion extraction model extracts opinion information corresponding to opinion holders from the corpus information based on the associated input data, and makes a sentiment judgment on the opinion information.

2. The media information opinion extraction method based on BERT according to claim 1, characterized in that, The step of semantically associating the identified opinion holders with the corpus information to obtain associated input data includes: converting the opinion holders and corpus information into vector form using word vector representation; calculating the semantic similarity between the opinion holder vector and the corpus information vector using semantic similarity calculation method; and associating the opinion holders and corpus information based on the semantic similarity to obtain associated input data.

3. The media information opinion extraction method based on BERT according to claim 1, characterized in that, The training process of the opinion extraction model is as follows: collecting a media information corpus containing opinion holders, opinion information, and sentiment tags; preprocessing the data in the media information corpus according to the semantic association processing method in step S2 to obtain training data; The training data is used to train the initial model built on the BERT model, the attention mechanism layer, and the opinion classification layer to obtain the trained opinion extraction model.

4. The media information opinion extraction method based on BERT according to claim 1, characterized in that, The process of judging the sentiment orientation of opinion information includes: adding a sentiment orientation analysis layer after the opinion classification layer; the sentiment orientation analysis layer judges the sentiment orientation of the opinion information based on a trained sentiment classification model, and the sentiment classification model is trained using a corpus of opinion information containing sentiment orientation labels.

5. The media information opinion extraction method based on BERT according to claim 1, characterized in that, It also includes: establishing a dynamic update mechanism, regularly collecting new media information corpus, and preprocessing it according to the semantic association processing method in step S2; using the preprocessed new corpus to incrementally train the opinion extraction model.

6. The media information opinion extraction method based on BERT according to claim 1, characterized in that, For long text corpora, a segmentation and attention mechanism is used to extract viewpoints. Specifically, the long text corpora are segmented into segments of a certain length; viewpoint holder identification and semantic association are performed on each segment; and the processing results of each segment are fused using an attention mechanism to focus on key viewpoint information.

7. The media information opinion extraction method based on BERT according to claim 1, characterized in that, For media information from different fields, a domain-adaptive optimization method is used to optimize the opinion extraction model. Specifically, this includes: collecting media information corpora from different fields; and fine-tuning the opinion extraction model using domain-specific corpora.

8. A media information opinion extraction device based on BERT, characterized in that, This includes a viewpoint holder identification module, a semantic association processing module, and a viewpoint extraction model module; The opinion holder identification module is used to identify multiple types of opinion holders in the corpus information, including named entities, pronouns, and referential phrases; The semantic association processing module is used to perform semantic association processing on the identified opinion holders and corpus information to obtain the associated input data. The opinion extraction model module is used to receive the associated input data, extract opinion information corresponding to opinion holders based on the BERT model, attention mechanism layer and opinion classification layer, and make sentiment judgment on the opinion information.

9. A media information opinion extraction device based on BERT, characterized in that, It includes a processor and a memory for storing a computer program, which, when executed by the processor, implements the BERT-based media information viewpoint extraction method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the BERT-based media information viewpoint extraction method as described in any one of claims 1-7.

Images