Social media information niche field term normalization and sentiment classification decoupling method and system

Abstract

The application discloses a social media information niche field term normalization and emotion classification decoupling method and system, comprising the following steps: receiving original social media information text, outputting structured metadata conforming to a frozen mark protocol through a niche field term identification engine, and keeping the original text character sequence unchanged; and inputting the original text and the metadata as joint input into an emotion classification model for discrimination. The frozen mark protocol serves as a contract layer to keep the data format stable, realize independent iterative update of the niche field term identification engine without triggering retraining of the emotion classification model. The application belongs to the technical field of natural language processing, and can realize independent iteration of the niche field term identification layer and the emotion classification layer through the frozen mark protocol as the contract layer.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method and system for sentiment analysis of niche online terms in social media information scenarios. Background Technology

[0002] Social media platforms (Weibo, Douban, Xiaohongshu, etc.) have become the core arena for public social media information expression. Vertical communities such as entertainment, e-sports, and animation have formed highly segmented expression systems. Niche terms such as fan circle slang, homophonic aliases, and emoji code words have emerged explosively on a weekly or even daily basis.

[0003] Social media information monitoring systems need to perform real-time sentiment analysis on these veiled expressions in order to identify potential negative social media information risks.

[0004] However, the frequent evolution of niche terms (or slang) in social media poses a serious challenge to the sustainability of existing natural language processing systems.

[0005] Current mainstream practice in the engineering field is to use a three-layer pipeline to handle sentiment analysis of niche domain terms: The first layer constructs a dictionary or vector library of niche domain terms, and matches known niche domain terms; The second layer utilizes a large language model (LLM) to provide real-time interpretation of unmatched niche domain terms, translating dynamic niche domain terms into standard semantic expressions; The third layer feeds the normalized standard text into pre-trained language models such as RoBERTa for sentiment classification.

[0006] In addition, knowledge augmentation schemes (such as EKBSA) enhance the model's capabilities by injecting external sentiment knowledge graphs; the Add-onLexicon method interpolates dictionary polarity scores into the feature vector as weighted features.

[0007] These methods alleviate the OOV problem to some extent, but they do not design service-level decoupling mechanisms for the monthly or even weekly evolution rate of niche terms, making them difficult to apply directly to high-frequency social media information monitoring scenarios.

[0008] The aforementioned existing technical solutions have the following structural defects: 1. Performance degradation caused by distribution coupling: Any iteration of the normalization layer, such as dictionary update, LLM strategy adjustment, or injection of new niche domain terms, will forcibly change the input distribution of the downstream classification model. Under the monthly niche domain term evolution rate, the classifier performance with frozen parameters will undergo uncontrollable linear degradation, which in engineering terms manifests as an 8 to 15 percentage point drop in F1 score.

[0009] 2. Semantic whitening: Normalization operations inevitably result in the loss of emotional polarity information when compressing the semantics of niche domain terms. When niche domain terms with context-dependent polarity, such as irony and metaphor, are uniformly translated into standard expressions, the emotional signals are systematically whitened.

[0010] 3. Unsustainable maintenance: To fix performance degradation, the downstream classification model needs to be retrained synchronously after each normalization layer iteration. In the scenario of daily evolution of niche terms in social media information, this retraining cost makes system maintenance unsustainable.

[0011] 4. Lack of service-level decoupling: The knowledge injection in existing knowledge enhancement schemes occurs during the forward propagation stage within the model, lacking service-level decoupling capabilities and failing to achieve independent iteration between the niche domain terminology understanding layer and the sentiment classification layer.

[0012] In summary, existing technologies have failed to address the coupling issue between the normalization of niche domain terminology and sentiment classification at the architectural design level, leading to uncontrollable performance degradation of computer systems in social media information monitoring scenarios due to input distribution drift.

[0013] Therefore, there is an urgent need to develop methods and systems for decoupling terminology normalization and sentiment classification in niche social media information fields to solve the problems in existing technologies. Summary of the Invention

[0014] The purpose of this invention is to provide a method and system for decoupling the normalization of niche domain terms and sentiment classification in social media information. It can achieve independent iteration of the niche domain terminology recognition layer and the sentiment classification layer by using a frozen tagging protocol as a contract layer. The structure is simple and easy to use, thereby solving the problem of uncontrollable performance degradation of computer systems in social media information monitoring scenarios caused by input distribution drift, as mentioned in the background art.

[0015] To achieve the above objectives, the present invention provides the following technical solution: A method and system for decoupling terminology normalization and sentiment classification in niche social media information fields, including the following steps: S1: Receive the original social media message text; S2: Perform niche domain terminology recognition on the original social media information text and output structured metadata that conforms to the freeze tagging protocol; the niche domain terminology recognition engine performs niche domain terminology recognition on the original social media information text and outputs structured metadata that conforms to the freeze tagging protocol; the niche domain terminology recognition engine keeps the character sequence of the original social media information text unchanged during the processing and does not perform text replacement operation; S3: Take the original social media information text and the structured metadata as joint inputs, feed them into the sentiment classification model for sentiment discrimination, and output sentiment tags; The freeze-marking protocol is a predefined data format contract that maintains data format stability throughout the system's lifecycle. It serves as a contract layer between the niche domain terminology recognition engine and the sentiment classification model, isolating the impact of the iterative updates of the niche domain terminology recognition engine on the input distribution of the sentiment classification model.

[0016] By adopting the above technical solution, the recognition of niche domain terms and sentiment classification are completely decoupled at the architectural level. The input distribution of the classification model is stabilized by the frozen labeling protocol, which fundamentally eliminates the risk of model performance degradation caused by iteration of the niche domain term layer.

[0017] As a further aspect of the present invention: the niche domain terminology recognition engine in S2 employs a three-level short-circuit routing mechanism for processing, including: S21: Dictionary matching of known niche domain terms based on automata; S22: If no match is found, semantic similarity retrieval is performed on the missing segments based on the vector library; S23: If the match is still not found, a large language model will be called for fallback identification.

[0018] By adopting the above technical solution, the specific dimensions and formats of structured metadata are defined, providing a standardized output specification for the niche domain terminology recognition layer, and ensuring the accuracy and consistency of semantic information transmission between different modules.

[0019] As a further aspect of the present invention: the three-level short-circuit routing mechanism is connected in series through confidence gates. When a high-confidence result is hit, it is returned in advance. When a low-confidence result appears, it is marked as uncertain and sent directly downstream. At the same time, a large language model is triggered to perform asynchronous verification.

[0020] By adopting the above technical solution, a lightweight model input modification scheme is provided. Through special marking, the semantics of niche domain terms are directly embedded into the original text context, enabling the model to intuitively perceive the position and attributes of niche domain terms.

[0021] As a further aspect of the present invention: the structured metadata includes at least four fields: niche domain term fragment location, domain category, polarity hint, and confidence level; the freeze tagging protocol specifies the enumeration set of domain category and polarity hint, and the numerical range of confidence level.

[0022] By adopting the above technical solution, a deep feature fusion scheme is provided, which enhances the model's ability to absorb features of niche domain terms through vector concatenation and improves the accuracy of sentiment discrimination in complex contexts.

[0023] As a further aspect of the present invention: In step S3, the original social media information text and the structured metadata are used as joint input, including: S31: Convert the structured metadata into a special token, and embed the special token into the corresponding niche term fragment in the original social media information text to form injected text; S32: Input the injected text into the pre-trained language model for encoding and classification.

[0024] By adopting the above technical solutions, a progressive and complementary identification mechanism was constructed, which takes into account both the millisecond-level response of known niche domain terms and the intelligent identification of unknown niche domain terms, thus ensuring the high availability and high coverage of the identification service.

[0025] As a further aspect of the present invention: In step S3, the original social media information text and the structured metadata are used as joint input, including: S31: Encode the original social media information text using a pre-trained language model to obtain a text feature vector; Construct a fixed-dimensional feature vector of niche domain terms based on the structured metadata; The text feature vector is concatenated with the niche domain terminology feature vector to obtain a concatenated vector, which is then input into the classification head for sentiment determination.

[0026] By adopting the above technical solution, an intelligent flow control mechanism is introduced, which significantly reduces the frequency of high-latency model calls and optimizes the overall inference efficiency of the system while ensuring recognition accuracy.

[0027] This invention also discloses a system for decoupling terminology normalization and sentiment classification in niche social media information domains, comprising: The niche domain terminology recognition module is configured to recognize the received original social media information text, output structured metadata that conforms to the freeze tag protocol, and keep the character sequence of the original social media information text unchanged. The sentiment classification module is configured to receive the original social media information text and the structured metadata, perform joint reasoning, and output the sentiment discrimination result; The contract layer module is configured to store and maintain the freeze mark protocol as a data format contract between the niche domain terminology recognition module and the sentiment classification module. The niche domain terminology recognition module supports independent iteration, and the contract layer module isolates the impact of the iteration of the niche domain terminology recognition module on the input distribution of the sentiment classification module through a stable data format.

[0028] By adopting the above technical solutions, a modular and independently evolving system architecture was built, providing an engineering implementation platform for the social media information monitoring system with low-cost maintenance and high-frequency iteration capabilities.

[0029] As a further aspect of the present invention: the niche terminology recognition module includes: The dictionary matching submodule is used for querying known niche domain terms based on automata; The vector library recall submodule is used to perform semantic similarity retrieval of missing segments based on the vector library; The fallback recognition submodule is used to call a large language model to identify new memes; The routing control submodule is used to connect the dictionary matching submodule, the vector library recall submodule, and the fallback identification submodule in series through a confidence gate.

[0030] By adopting the above technical solution, the internal logic of the niche domain terminology recognition module was refined, achieving a smooth transition from rule matching to model inference and ensuring the robustness of the recognition service.

[0031] As a further aspect of the present invention: the emotion classification module includes: The text encoding submodule is used to encode the original text based on a pre-trained language model; The feature processing submodule is used to encode the structured metadata into niche domain terminology feature vectors; The classification output submodule is used to concatenate the text encoding results with the feature vectors of niche domain terms and output sentiment tags.

[0032] By adopting the above technical solutions, the data processing link of the sentiment classification module has been improved, and the deep integration of the original text semantics and the structured niche domain terminology features has been achieved, thereby enhancing the system's generalization ability.

[0033] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the above-described method for decoupling the normalization of niche terms in social media information from sentiment classification.

[0034] By adopting the above technical solution, a storage medium for computer program products is provided, enabling the above decoupling method to be executed and distributed in a standardized manner in digital computing devices.

[0035] Compared with the prior art, the beneficial effects of the present invention are: This invention maintains the original text unchanged and introduces a frozen labeling protocol, ensuring that the input distribution of the classification layer is completely consistent during the training and inference phases, thus completely eliminating normalization-classification coupling drift (NCCD). Any iteration of the niche domain terminology recognition layer, such as dictionary updates, LLM strategy adjustments, or injection of new niche domain terms, will not trigger input distribution drift in the downstream classification model.

[0036] The niche domain terminology recognition layer in this invention does not perform text replacement, but only outputs structured semantic tags. Context-dependent emotional polarity information in the original text, such as irony and metaphor, is fully preserved and passed to the classification layer, thus avoiding semantic bleaching and preserving context-dependent emotional information.

[0037] In this invention, the niche domain terminology recognition layer supports daily or even higher frequency independent iterations, such as dictionary hot loading, incremental vector library updates, and LLM prompt tuning, while the sentiment classification layer does not require retraining. Under the same social media information monitoring requirements, the number of annual model retraining times is reduced from 6 to 12 times to 0 times, significantly reducing system maintenance costs.

[0038] Under the premise of a stable labeling protocol, this invention achieves zero discrimination risk drift for known niche domain terms. The introduction of new niche domain terms only affects newly added samples, and the overall risk drift is strictly limited to a linear function of the label difference probability, thus avoiding cumulative degradation of historical performance.

[0039] This invention provides two implementation paths: Special Token injection and Feature bypass splicing, which are adapted to the needs of lightweight and rapid deployment and high-precision scenarios, respectively. The microservice-based three-level short-circuit routing design ensures accuracy while controlling the end-to-end latency within the range of 10 to 500ms, meeting the latency requirements of real-time social media information monitoring, and has strong engineering feasibility.

[0040] Other features and advantages of the present invention will be disclosed in detail in the following detailed description and accompanying drawings. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the steps of a method for decoupling the normalization of niche social media information terminology and sentiment classification in an embodiment of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] This invention provides a method and system for decoupling the normalization of niche domain terminology and sentiment classification in social media information. It aims to solve the problems of normalization-classification coupling drift (NCCD), semantic bleaching, and high maintenance costs caused by the coupling architecture of "niche domain terminology normalization → sentiment classification" in the prior art. The core of this invention is to propose a decoupled contextual niche domain terminology architecture, which uses a frozen tagging protocol as a contract layer to achieve independent iteration of the niche domain terminology recognition layer and the sentiment classification layer.

[0044] In this embodiment of the invention, a method and system for decoupling social media information niche domain terminology normalization and sentiment classification are described below. Figure 1 As shown, it includes the following: Let the original social media message text be x, the niche domain terminology labeling function be m, and the sentiment classification model be c. Then the decoupled reasoning process is formally represented as:

[0045] Where x remains the original text sequence, without any translation or rewriting; m is a tagging function for niche domain terms, and its output m(x) is a set of structured metadata conforming to the freeze tagging protocol. ; Each label item It should contain at least four fields: the niche term fragment location information span, the domain category, the polarity hint, and the confidence level; The sentiment classification model c performs joint inference based on the original text x and the structured metadata m(x) to output the sentiment judgment result.

[0046] Example 1: Decoupling Method Based on Special Token Injection This embodiment is suitable for lightweight deployment scenarios. By modifying the vocabulary of the model input layer, semantic injection of metadata for niche domain terms is achieved.

[0047] Step S1: Receive the original social media message text.

[0048] The system receives raw social media text messages from external data sources. For example, the input text might be: "This drama is really based on a thesis defense. Can the screenwriter please stop writing?"

[0049] Step S2: Use a niche domain terminology recognition engine to identify the original social media information text and output structured metadata that conforms to the freeze tag protocol.

[0050] The niche terminology recognition engine is internally encapsulated as an independent microservice and uses a three-level short-circuit routing mechanism for processing: The first-level routing is dictionary matching: a niche domain terminology dictionary is built based on the Aho-Corasick automaton, and multi-pattern matching is performed on the input text. If a known niche domain term is matched and the matching length is greater than a preset threshold, the result is returned directly. For example, if the dictionary contains the entry "relying on defense", an automaton can complete the matching in O(N) time complexity.

[0051] The second level of routing is vector library recall: For the missing parts, the Chroma (open source vector database) vector library is called to perform semantic similarity retrieval, calculate the cosine similarity between the text fragment vector and the niche domain term vector, and if the similarity score is higher than the threshold of 0.82, the result is recalled.

[0052] The third-level routing is an asynchronous fallback for LLM (Large Language Model): For low-confidence results, open-source models such as Qwen-14B are called through LiteLLM Proxy (LiteLLM proxy server), and new meme identification is performed using preset prompts.

[0053] In this step, the niche domain terminology recognition engine outputs structured metadata that conforms to the frozen tagging protocol. For example, for the text fragment "relying on defense", the output result is as follows: m(x)={span:[5,9],domain:"general",polarity_hint:"strong_neg",confidence:0.98} Among them, the span field indicates the start and end positions of niche domain terms in the original text, the domain field indicates that the domain is general, the polarity_hint field indicates that the polarity is strongly negative, and the confidence field indicates the confidence level.

[0054] Step S3: Input the original social media message text and structured metadata as joint inputs into the sentiment classification model for sentiment determination.

[0055] This embodiment uses the Special Token injection path: S31: Add special tokens to the vocabulary of the pre-trained language model, including the start token. <slang> End marker< / slang> and attribute tags<slang_p=strong_neg_d=general> .

[0056] S32: During the text preprocessing stage, based on the location information in the structured metadata, the Special Token is embedded into the corresponding position in the original text to form the injected text; for example, the injected text generated after the above example processing is: "This drama is really..."<slang_p=strong_neg_d=general> "Based on the defense, can screenwriters just stop writing?"

[0057] S33: Encode and classify the injected text input into the pre-trained language model; The model structure adopts the Chinese-roberta-wwm-ext (Chinese RoBERTa base model), and is fine-tuned by LoRA (Low-Rank Adaptation) technique. The fine-tuning parameters are set to rank=8, scaling factor alpha=16, and the target module is the q_proj and v_proj layers in the attention mechanism. During the training phase, the model learns the semantic association between Special Tokens and context. Since the vocabulary of Special Tokens is fixed, the iteration of the niche domain terminology recognition layer will not introduce new Tokens, thus ensuring a stable input distribution.

[0058] Example 2 The technical difference between this embodiment and Embodiment 1 is that this embodiment is applicable to high-precision scenarios, does not require modification of the model vocabulary, and achieves decoupling through feature vector fusion.

[0059] Steps S1 and S2 are the same as in Example 1; It receives the raw text and outputs structured metadata through a niche domain terminology recognition engine, keeping the raw text unchanged.

[0060] Step S3: Input the original social media message text and structured metadata as joint inputs into the sentiment classification model for sentiment determination.

[0061] This embodiment uses a Feature bypass splicing path: First, the original text is encoded using a pre-trained language model, and the CLS (classification label) vector from the last hidden state is taken as the text feature vector. The dimension d=768; CLS vectors are typically used to represent the semantics of an entire sentence.

[0062] Subsequently, a fixed-dimensional niche domain terminology feature vector f is constructed based on the structured metadata, with dimension k=8; the construction logic of the feature vector f is as follows:

[0063] in, This is a polarity indicator; it is 1 if the corresponding polarity is contained in the metadata, and 0 otherwise. The coverage rate of niche terminology is calculated by dividing the number of characters in the niche terminology by the total number of characters. and These represent the maximum confidence level and the average confidence level in the metadata, respectively. The number of niche domain terminology fragments identified.

[0064] Next, the text feature vector is concatenated with the feature vector of niche domain terms to obtain the concatenated vector. The calculation formula is as follows:

[0065] After concatenation, the vector dimension is 768 + 8 = 776.

[0066] Finally, the concatenated vector is input into the classification head for sentiment determination. The classification head adopts a two-layer MLP (Multilayer Perceptron) structure. The first layer maps 776 dimensions to 256 dimensions with ReLU activation function. The second layer maps 256 dimensions to 3 dimensions (positive, neutral, negative) and outputs sentiment labels.

[0067] System Architecture and Contract Layer Description This invention also provides a system for decoupling the normalization of niche terminology and sentiment classification in social media information, including a niche terminology recognition module, a sentiment classification module, and a contract layer module.

[0068] The niche domain terminology recognition module is configured to execute the above three-level short-circuit routing logic, supports dictionary hot loading and LLM Prompt tuning, and its internal implementation evolution does not affect the output format. The contract layer module stores and maintains the frozen tag protocol, which specifies the domain enumeration, such as entertainment, e-sports, animation, and general, and the polarity hint enumeration, such as the value range of strongly negative, mildly negative, ironic, and neutral, as a data format contract between modules. The sentiment classification module receives the same raw social media text during the training and inference phases, and reads niche domain terminology features only according to the format defined by the contract layer, so that no retraining is required when the niche domain terminology recognition module iterates independently.

[0069] This invention provides a method and system for decoupling the normalization of niche terminology and sentiment classification in social media information. It can achieve independent iteration of the niche terminology recognition layer and the sentiment classification layer by using a frozen tagging protocol as a contract layer.

[0070] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0071] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for decoupling terminology normalization and sentiment classification in niche social media information fields, characterized in that, Includes the following steps: S1: Receive the original social media message text; S2: Recognize niche domain terms in the original social media information text and output structured metadata that conforms to the freeze tagging protocol; The original social media information text is processed by a niche domain terminology recognition engine to identify niche domain terms and output structured metadata that conforms to the freeze tagging protocol. The niche terminology recognition engine maintains the character sequence of the original social media information text unchanged during processing and does not perform text replacement operations; S3: Take the original social media information text and the structured metadata as joint inputs, feed them into the sentiment classification model for sentiment discrimination, and output sentiment tags; The freeze-marking protocol is a predefined data format contract that maintains data format stability throughout the system's lifecycle. It serves as a contract layer between the niche domain terminology recognition engine and the sentiment classification model, isolating the impact of the iterative updates of the niche domain terminology recognition engine on the input distribution of the sentiment classification model.

2. The method for decoupling social media information niche domain terminology normalization and sentiment classification according to claim 1, characterized in that, The niche domain terminology recognition engine in S2 employs a three-level short-circuit routing mechanism, including: S21: Dictionary matching of known niche domain terms based on automata; S22: If no match is found, semantic similarity retrieval is performed on the missing segments based on the vector library; S23: If the match is still not found, a large language model will be called for fallback identification.

3. The method for decoupling social media information niche domain terminology normalization and sentiment classification according to claim 2, characterized in that, The three-level short-circuit routing mechanism is connected in series through confidence gates. When a high-confidence result is hit, it returns in advance. When a low-confidence result appears, it is marked as uncertain and sent directly downstream. At the same time, it triggers a large language model to perform asynchronous verification.

4. The method for decoupling social media information niche domain terminology normalization and sentiment classification according to claim 1, characterized in that, The structured metadata includes at least four fields: niche domain term fragment location, domain category, polarity hint, and confidence level; the freeze tagging protocol specifies the enumeration set of domain category and polarity hint, and the numerical range of confidence level.

5. The method for decoupling social media information niche domain terminology normalization and sentiment classification according to claim 1, characterized in that, In step S3, the original social media information text and the structured metadata are used as joint inputs, including: S31: Convert the structured metadata into a special token, and embed the special token into the corresponding niche term fragment in the original social media information text to form injected text; S32: Input the injected text into the pre-trained language model for encoding and classification.

6. The method for decoupling social media information niche domain terminology normalization and sentiment classification according to claim 1, characterized in that, In step S3, the original social media information text and the structured metadata are used as joint inputs, including: S31: Encode the original social media information text using a pre-trained language model to obtain a text feature vector; Construct a fixed-dimensional feature vector of niche domain terms based on the structured metadata; The text feature vector is concatenated with the niche domain terminology feature vector to obtain a concatenated vector, which is then input into the classification head for sentiment determination.

7. A system for decoupling terminology normalization and sentiment classification in niche social media information domains, characterized in that, include: The niche domain terminology recognition module is configured to recognize the received original social media information text, output structured metadata that conforms to the freeze tag protocol, and keep the character sequence of the original social media information text unchanged. The sentiment classification module is configured to receive the original social media information text and the structured metadata, perform joint reasoning, and output the sentiment discrimination result; The contract layer module is configured to store and maintain the freeze mark protocol as a data format contract between the niche domain terminology recognition module and the sentiment classification module. The niche domain terminology recognition module supports independent iteration, and the contract layer module isolates the impact of the iteration of the niche domain terminology recognition module on the input distribution of the sentiment classification module through a stable data format.

8. The social media information niche domain terminology normalization and sentiment classification decoupling system according to claim 7, characterized in that, The niche terminology recognition module includes: The dictionary matching submodule is used for querying known niche domain terms based on automata; The vector library recall submodule is used to perform semantic similarity retrieval of missing segments based on the vector library; The fallback recognition submodule is used to call a large language model to identify new memes; The routing control submodule is used to connect the dictionary matching submodule, the vector library recall submodule, and the fallback identification submodule in series through a confidence gate.

9. The social media information niche domain terminology normalization and sentiment classification decoupling system according to claim 7, characterized in that, The emotion classification module includes: The text encoding submodule is used to encode the original text based on a pre-trained language model; The feature processing submodule is used to encode the structured metadata into niche domain terminology feature vectors; The classification output submodule is used to concatenate the text encoding results with the feature vectors of niche domain terms and output sentiment tags.

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 method for decoupling social media information niche domain terminology normalization and sentiment classification as described in any one of claims 1 to 6.

Images