Language analysis method and system based on fast thinking-evaluation-slow thinking aspect-level sentiment analysis framework

Abstract

The invention relates to the technical field of computer data processing, in particular to a language analysis method and system based on a fast thinking-evaluation-slow thinking aspect level sentiment analysis framework. The framework comprises a large reasoning model and a large evaluation model, the large reasoning model is responsible for aspect emotion reasoning, and the large evaluation model is responsible for evaluating reasoning correctness to prevent cognitive deviation. The large reasoning model firstly generates a preliminary emotion and reasoning process in a target aspect in a fast thinking stage. Then, the large evaluation model evaluates the reasoning. If inference is considered to be correct, the emotion result is adopted. Otherwise, the reasoning large model enters a slow thinking stage, the reasoning evaluation reasonability is reflected, the original reasoning is selectively corrected or reserved, and dynamic correction is achieved. The method can effectively enhance the understanding ability of a large model for complex text sentiment, and significantly improve the accuracy of sentiment judgment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer data processing, and particularly relates to a language analysis method and system based on an aspect-level sentiment analysis framework of fast thinking-evaluation-slow thinking. BACKGROUND

[0002] In recent years, with the rapid development of large language models (LLM), their application in aspect-level sentiment analysis has been increasingly widely concerned. And with the increasing demand for fine-grained sentiment analysis, aspect-based sentiment analysis (ABSA) has become an important research direction in the field of sentiment computing.

[0003] Aspect-based sentiment analysis, as an important branch of sentiment analysis, aims to analyze the sentiment polarity related to specific aspects in a sentence. Unlike traditional sentence-level or document-level sentiment analysis, aspect-based sentiment analysis aims to analyze the sentiment polarity related to specific aspects in a sentence. For example, in "the food in this restaurant is very good, but the service is too slow", the food and service are aspect terms, expressing positive and negative sentiment polarity respectively. Through aspect-based sentiment analysis, the true attitude of users in different aspects can be accurately described, providing more explanatory basis for product improvement, service optimization or user experience research. Before the rise of large models, ABSA research mainly focused on deep neural network models under the supervised learning paradigm. Early mainstream methods relied on recurrent neural networks (RNN) or convolutional neural networks (CNN) to capture the sequential features of the context. To more accurately establish the semantic association between aspect words and sentiment words, researchers used attention mechanisms to give the model the ability to focus on key information in the sentence. In addition, considering the importance of syntactic structure to sentiment analysis, graph convolutional neural networks based on dependency trees were also used to capture long-distance syntactic dependencies. Although these methods have made significant progress in feature modeling, they are essentially limited by the limited labeled data and pre-defined feature interaction patterns under the supervised learning paradigm, making it difficult to break through the constraints of the closed feature space. This leads to a significant decline in the generalization ability of traditional deep models when faced with out-of-distribution data samples.

[0004] In recent years, the emergence of large models has provided a new paradigm for ABSA research. With the rich semantic knowledge learned on large-scale corpus, large models have shown strong generalization and transfer ability in fine-grained sentiment recognition tasks. However, existing large models still face serious challenges in dealing with complex sentiment expressions containing implicit sentiment, irony, or high dependence on context. Large models often struggle to fully understand the underlying sentiment in such texts, and are prone to misjudgment of the target aspect sentiment polarity. This is because large models tend to rely on capturing prominent feature words about aspects in the text for quick judgment. While this shallow pattern matching is effective in simple samples, it is easily misled by surface clues in complex sentiment contexts, leading to misjudgment. Existing reasoning patterns are unidirectional and lack evaluation and reflection correction mechanisms for generated results. Once the large model makes an incorrect judgment based on local clues, the initial sentiment judgment error cannot be identified and corrected in time. How to endow large models with a "reflective" deep reasoning ability similar to humans, enabling them to identify their own incorrect judgments and make effective corrections, is a problem that needs to be addressed.

[0005] Traditional neural network models significantly improve the extraction and representation of sentiment features by modeling the semantic dependency between aspects and context. Wang et al. use LSTM to capture the dependency between aspects and context, achieving effective modeling of basic sentiment features. Gu et al. propose a position-aware bidirectional attention network based on bidirectional GRU, which not only considers the relationship between aspects and context, but also incorporates the influence of position on sentiment classification. Wang et al. propose a model based on relationship graph attention, which effectively captures the semantic relationship between aspect words and context by modeling aspect-oriented dependency syntax structures to improve sentiment classification performance. However, it relies on fixed network architecture and pre-defined feature interaction patterns, which are essentially limited to the fixed feature representation space learned during model training.

[0006] The development of large models provides a new approach to aspect-level sentiment analysis tasks. With strong semantic understanding and reasoning ability, large models can achieve sentiment recognition with zero or few samples, and through prompt engineering and chain-of-thought (CoT) reasoning, they exhibit stronger reasoning ability and interpretability. Zhang et al. pointed out that although large models have potential in aspect-level sentiment recognition, they still struggle to accurately predict in complex semantic scenarios. Fei et al. demonstrated that through chain-of-thought reasoning, models can improve their ability to recognize sentiment in complex scenarios. However, chain-of-thought reasoning is a unidirectional and self-inspection-lacking reasoning pattern, making it difficult to fully address complex sentiment expressions.

[0007] In the dual-processing theory of fast thinking and slow thinking, the dual-processing theory provides the core theoretical basis for the fast thinking-evaluation-slow thinking cooperation framework (FESCR). The theory divides human cognition into two complementary modes: the fast thinking of system one is responsible for rapid and intuitive preliminary judgment, and the slow thinking of system two is responsible for careful and deep reasoning analysis. The ideal human decision-making is the result of the efficient cooperation of the two systems: system one is responsible for handling daily trivial matters and providing preliminary judgments; when system one detects abnormalities or cannot solve problems, it will seek help from system two. This cognitive mechanism directly inspires the "fast thinking-evaluation-slow thinking" cooperation framework (FESCR). Compared with the traditional one-way and lack of self-examination reasoning mode, the FESCR framework can evaluate and reflect the preliminary results, enhance the understanding of complex text sentiment, and improve the accuracy of aspect-level sentiment analysis.

[0008] In-context Learning (ICL) is an important paradigm in the application of large models in recent years. Unlike the traditional Fine-tuning method that relies on parameter updates, ICL activates the model's potential reasoning and knowledge expression capabilities by structuring the context information in the input prompt without changing the model parameters. Among them, prompt engineering guides the model to more accurately understand the task goal and output logic through explicit task instructions and context prompts. However, existing methods show obvious one-way nature, lacking dynamic correction mechanism for incorrect sentiment judgments.

[0009] Therefore, it is necessary to study a new language analysis method and system based on the aspect-level sentiment analysis framework of fast thinking-evaluation-slow thinking. SUMMARY

[0010] The main purpose of the present application is to provide a language analysis method and system based on the aspect-level sentiment analysis framework of fast thinking-evaluation-slow thinking. The FESCR proposed in the present application further combines the dual-processing theory to construct an analysis framework that can evaluate and reflect the initial results, providing a new cognitive heuristic framework for aspect-level sentiment analysis tasks.

[0011] To solve the above technical problems.

[0012] In order to achieve the above purpose, the present application provides a language analysis method based on the aspect-level sentiment analysis framework of fast thinking-evaluation-slow thinking, comprising the steps of: S1, definition of aspect-level sentiment analysis: for a given sentence X and the aspect term A contained therein, construct the target task of judging the sentiment polarity of aspect A in the sentence X ∈{positive sentiment, negative sentiment, neutral sentiment}; S2, fast thinking stage: guided by fast thinking prompt words, reasoning large model is dominant execution, simulating the system one of dual processing theory in cognitive psychology, defining the objective function of fast thinking stage (c , R quick ) = LLM reasoning ([S, A], Prompt quick ); wherein LLM reasoning represents reasoning large model, S represents input sentence, A is target aspect term, Prompt quick is the prompt word of fast thinking stage, the output result is composed of preliminary sentiment prediction and corresponding reasoning process R quick ; S3, evaluation stage: guided by evaluation prompt words, evaluation large model is dominant execution, and injects domain label in prompt words, builds specific domain cognitive context for evaluation large model, improves evaluation accuracy; define the objective function of evaluation stage: (c eval ,E eval ) = LLM eval ([S, A, R quick , domain], Prompt evl ); wherein, S represents input sentence, A is target aspect, R quick is the reasoning process of reasoning large model in fast thinking stage, domain is domain label, Prompt evl is the prompt word of evaluation stage, c evl is the correctness evaluation result, E eval is the evaluation process; S4, slow thinking stage: simulating system two in dual processing theory, guided by slow thinking prompt words, reasoning large model is dominant execution, when the judgment in evaluation stage is incorrect, reasoning large model enters slow thinking stage, reflects the evaluation rationality and selectively corrects or retains the original reasoning, realizes dynamic correction; the objective function of slow thinking stage is: (c ,R slow ) = LLM reasoning ([S, A, domain, P quick ,E evl ], Prompt slow ); wherein, S represents input sentence, A is target aspect, domain is domain label, P quick ,E evl are the output processes of fast thinking stage and evaluation stage respectively, Prompt slow is the prompt word of slow thinking stage, is the final confirmed sentiment label after deep reflection, and is used as the final prediction result of the task, R slow is the final reasoning process.

[0013] Preferably, the fast thinking and slow thinking stages are responsible for the Qwen-Turbo large language model.

[0014] Preferably, in the fast thinking and slow thinking stages, the corresponding emotional prediction results and the corresponding reasons are output, and the temperature parameter is set to 0 to ensure the stability of the output.

[0015] Preferably, the evaluation stage adopts the Moonshot-Kimi-K2-Instruct model to perform the evaluation task.

[0016] The application also provides a language analysis system based on the aspect-level sentiment analysis framework of fast thinking-evaluation-slow thinking, which comprises a memory, a processor, a computer program stored in the memory and executable on the processor, and the processor executes the computer program to realize the steps of the language analysis method based on the aspect-level sentiment analysis framework of fast thinking-evaluation-slow thinking according to any one of the above.

[0017] The application also provides a computer readable storage medium storing a computer program, which is executed by a processor to realize the steps of the language analysis method based on the aspect-level sentiment analysis framework of fast thinking-evaluation-slow thinking according to any one of the above.

[0018] The application innovatively proposes a fast thinking-evaluation-slow thinking collaborative framework (FESCR) by combining the dual processing theory of fast thinking-slow thinking in cognitive psychology. The framework includes a reasoning large model and an evaluation large model. The former is responsible for aspect sentiment reasoning, and the latter is responsible for evaluating the correctness of reasoning to prevent cognitive bias. The reasoning large model first generates the preliminary emotion and reasoning process of the target aspect in the fast thinking stage. Then, the evaluation large model evaluates the reasoning. If the reasoning is correct, the emotional result is adopted. Otherwise, the reasoning large model enters the slow thinking stage, reflects on the evaluation rationality and selectively corrects or retains the original reasoning, realizing dynamic correction. The application can effectively enhance the understanding ability of the large model for complex text emotion and significantly improve the accuracy of emotional judgment. It also provides an innovative path based on cognitive science for building a large model reasoning mechanism with more cognitive rationality and reliability. Compared with the prior art, the application has the following advantages: The dual processing theory of fast thinking-slow thinking in cognitive psychology is first used in the aspect-level sentiment analysis task of the large model, providing a new cognitive heuristic framework for the aspect-level sentiment analysis task. An innovative fast thinking-evaluation-slow thinking collaborative framework is proposed for the aspect-level sentiment analysis task, which can effectively improve the understanding ability of the model for complex emotional text. Experiments on multiple mainstream ABSA benchmark datasets show that the framework significantly improves the accuracy of sentiment judgment. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A logic schematic diagram of an aspect-level sentiment analysis framework based on fast thinking-evaluation-slow thinking is provided for the embodiment.

[0020] Figure 2 A case quantity diagram of the language analysis method of the aspect-level sentiment analysis framework based on fast thinking-evaluation-slow thinking through two different paths of the framework is provided for the embodiment.

[0021] Figure 3 A flowchart of the language analysis method of the aspect-level sentiment analysis framework based on fast thinking-evaluation-slow thinking is provided for the embodiment.

[0022] Figure 4 A language analysis system hardware structure schematic diagram of the aspect-level sentiment analysis framework based on fast thinking-evaluation-slow thinking is provided for the embodiment. DETAILED DESCRIPTION

[0023] In order to better explain the present application, so as to be understood, the present application is described in detail by specific embodiments in combination with the accompanying drawings.

[0024] Based on the dual processing theory in cognitive psychology, the present application proposes a fast thinking-evaluation-slow thinking coordination framework, aiming to transform the large model from a single, static responder into an intelligent analysis system that can simulate the dynamic cognitive decision-making process of human beings through multi-model collaboration. The overall framework is shown in Figure 1 Through the organic cooperation of the three stages, accurate judgment of emotional expression is realized. The specific process is as follows, a given input (X and the aspect terms A contained therein) first enters the fast thinking stage, which generates a preliminary sentiment judgment and reasoning process under the guidance of fast thinking cues. Then, in the evaluation stage, the evaluation model evaluates the correctness of the reasoning process of fast thinking under the guidance of evaluation cues and forms an evaluation report, and dynamically decides whether to activate the slow thinking stage according to the evaluation result. In the slow thinking stage, the reasoning large model makes the final reflection and modification of the reasoning process identified as incorrect under the guidance of slow thinking cues, referring to the evaluation report.

[0025] As shown in Figure 2 A language analysis method of an aspect-level sentiment analysis framework based on fast thinking-evaluation-slow thinking is provided in the embodiment, which comprises steps S1-S4.

[0026] S1, Definition of Aspect-level Sentiment Analysis: For a given sentence X and its contained aspect term A, construct the sentiment polarity judgment for aspect A in the sentence X The target task is to determine the sentiment polarity of the aspect term A in the given sentence X.

[0027] S2, Fast Thinking Stage: Guided by fast thinking cues, the reasoning large model is mainly executed, simulating the system one of the dual processing theory in cognitive psychology. The objective function of the fast thinking stage is defined as , quick ) = LLM reasoning ([S, A], Prompt quick ); where LLM reasoning represents the reasoning large model, S represents the input sentence, A is the target aspect term, Prompt quick is the cue for the fast thinking stage, and the output result consists of the preliminary sentiment prediction and the corresponding reasoning process R quick .

[0028] Specifically, the fast thinking stage in step S2 is the initial link of the framework, which is mainly executed by the reasoning large model. Its design aims to reproduce the intuitive and automated cognitive characteristics of system one in the dual processing theory. In this stage, the model, guided by task instructions, focuses on the prominent sentiment words in the input text about the target aspect, such as "good", "bad", etc., and forms a preliminary sentiment judgment and reasoning process based on such surface semantic clues.

[0029] When facing texts with clear emotional expression and simple structure, this intuitive mechanism can efficiently and accurately complete the judgment, embodying the automated processing advantage similar to the first human reaction. However, when encountering implicit emotions, irony, or highly context-dependent expressions, this intuitive reasoning is easily misled, resulting in sentiment judgment bias, such as misunderstanding sarcastic positive words as real praise.

[0030] Due to the lack of internal evaluation and reflection ability in this stage, such erroneous intuitive reasoning is often directly output as the final prediction, making it difficult to self-correct. Therefore, the output of this stage is defined as an intuitive hypothesis that needs to be externally verified. The output process of the reasoning large model in the fast thinking stage is formally represented as: , quick ) = LLM reasoning ([S, A], Prompt quick ); where LLM reasoning represents the reasoning large model, S represents the input sentence, A is the target aspect term, Prompt quick is the cue for the fast thinking stage, and the output result consists of the preliminary sentiment prediction and the corresponding reasoning process Rquick Composition.

[0031] S3, evaluation stage: the evaluation of the large model is guided by the evaluation prompt word, and the domain label is injected into the prompt word, the cognitive context of the evaluation large model is constructed, and the objective function of the evaluation stage is defined: (c eval ,E eval ) = LLM eval ([S, A, R quick , domain], Prompt evl ); wherein S represents an input sentence, A is a target aspect, R quick is the reasoning process of the reasoning large model in the fast thinking stage, domain is the domain label, Prompt evl is the prompt word of the evaluation stage, c evl is the correctness evaluation result, E eval is the evaluation process.

[0032] Specifically, the evaluation stage in step S3 is a bridge connecting fast thinking and slow thinking. Since the reasoning large model may produce erroneous intuition in the fast thinking stage, this stage uses the evaluation large model to construct an objective external perspective and gives it the core responsibility of cognitive decision maker.

[0033] In order to ensure the accuracy of the evaluation, we use the contextualized evidence constraint strategy in the prompt word, that is, we explicitly inject domain labels (Domain Label) such as Laptop and Restaurant, to construct the cognitive context of the evaluation large model in a specific domain. Through this contextualized independent perspective and different knowledge application mode, the evaluation large model can effectively break the thinking inertia of fast thinking, combine the field context to sharply find the errors in the reasoning process of the reasoning large model, such as cross-domain emotional ambiguity, and dynamically decide whether to activate the slow thinking stage according to the evaluation result.

[0034] Specifically, this stage uses the evaluation large model to strictly evaluate the reasoning process in the fast thinking stage in combination with domain background knowledge, and uses it as the basis for switching between fast and slow thinking stages. When the emotional reasoning process of fast thinking is evaluated as correct, the framework considers that the current task belongs to a simple and intuitive aspect-level sentiment analysis task, and the intuitive judgment of fast thinking is reliable enough, so its answer is directly adopted without activating the slow thinking stage. On the contrary, when the evaluation finds that the reasoning is wrong, it is determined that the current task involves complex emotional expression, and the fast thinking stage cannot reliably handle it. At this time, the system will activate the slow thinking stage, and the corresponding evaluation process will be reported to the slow thinking stage for reference. The output process of the evaluation large model in the evaluation stage is formally expressed as: (c eval ,E eval ) = LLM eval ([S, A, Rquick , Prompt evl ); wherein S represents an input sentence, A is a target aspect, R quick is the reasoning process of the reasoning large model in the fast thinking stage, domain is a domain label, Prompt evl is a prompt word in the evaluation stage, c evl is a correctness evaluation result, E eval is an evaluation process. S4, slow thinking stage: simulating system two in the dual processing theory, guided by the slow thinking prompt word, the reasoning large model is mainly executed, when the evaluation stage judges that it is incorrect, the reasoning large model enters the slow thinking stage, the objective function of the slow thinking stage is , R slow ) = LLM reasoning ([S, A, domain, R quick , E evl ], Prompt slow ); wherein S represents an input sentence, A is a target aspect, domain is a domain label, R quick , E evl are the output processes of the fast thinking stage and the evaluation stage respectively, Prompt slow is a prompt word in the slow thinking stage, is the final confirmed emotional label after deep reflection, and is used as the final prediction result of the task, R slow is the final reasoning process.

[0035] Specifically, in step S4, the slow thinking stage simulates the reflective reasoning and correction process of system two in the dual processing theory, which is careful and requires cognitive effort. This stage will be formally activated only when the evaluation stage sends an explicit trigger signal, that is, the evaluation thinks that the reasoning is wrong.

[0036] In this stage, the reasoning large model analyzes the evaluation report of the evaluation large model in the context of a specific field. In this stage, the reasoning large model does not blindly believe the evaluation report, but first evaluates the evaluation report itself to determine its rationality. This simulates the process of humans thinking deeply, which will think about the initial thinking itself. If the slow thinking stage thinks that the problem provided by the evaluation stage is unreasonable, it will maintain the emotional reasoning process of the fast thinking stage, which reflects the recognition of the effective work of system 2 to system 1. If the slow thinking confirms the problem pointed out by the evaluation stage, it will combine this problem and then output a revised and more reliable emotional judgment and reasoning process. The whole slow thinking process is a deep reasoning analysis under sufficient information support, with the ability to reflect and correct errors. The output process of the reasoning large model in the slow thinking stage is formally expressed as: , Rslow ) = LLM reasoning ([S, A, domain, P quick , E evl ], Prompt slow ); wherein S represents an input sentence, A is a target aspect, domain is a domain label, P quick , E evl are output processes of a fast thinking stage and an evaluation stage respectively, Prompt slow is a prompt word of a slow thinking stage, is a final confirmed sentiment label after deep reflection and serves as the final prediction result of the task, and R slow is a final reasoning process Preferably, a Qwen-Turbo large language model is used for the fast thinking and slow thinking stages.

[0037] Preferably, in the fast thinking and slow thinking stages, the corresponding sentiment prediction results and the corresponding reasons are output, and the temperature parameter is set to 0 to ensure the stability of the output.

[0038] Preferably, the evaluation stage uses a Moonshot-Kimi-K2-Instruct model to perform the evaluation task.

[0039] Preferably, the specific prompt words used in the fast thinking stage, the evaluation stage and the slow thinking stage are as follows: A specific experimental process is provided below to introduce the language analysis method and system of the aspect-level sentiment analysis framework based on fast thinking-evaluation-slow thinking of the application.

[0040] The proposed method is evaluated on three commonly used ABSA benchmark datasets, including Lap14, Rest14 and Rest15. These datasets are all from the SemEval competition and cover two fields of laptops (Lap14) and restaurants (Rest14, Rest15). Each sample is composed of a review text, a specified aspect and a sentiment label (positive, neutral, negative). The detailed statistics are shown in Table 1.

[0041] Table 1 To evaluate the effectiveness of the model, Accuracy and Macro-F1 are used as evaluation indicators. The accuracy quantifies the proportion of samples that the model predicts correctly as a whole. Macro-F1 provides a more balanced evaluation perspective of the model performance by calculating the average of all class F1 scores, especially focusing on the balance of the performance of each class.

[0042] To comprehensively verify the effectiveness of the proposed method, the following representative baseline models are selected for comparison: IAN model: uses two interactive attention networks to learn the representations of context and target, focusing on the relevant parts of context and target by utilizing mutual attention mechanism, respectively generating specific aspect representations for context and target; TD-LSTM model: uses two target-dependent LSTM networks to capture the dependency between specific aspects and left and right context; ATAE-LSTM uses attention-based LSTM model to calculate the attention score of specific aspects, so that the model can focus on the key context information around different aspects in the sentence; ASGCN: simplifies ASGCN by replacing two GCN layers in ASGCN with two CNN layers; BIGCN: proposes a novel hierarchical structure of words and syntax graphs. It uses a global word-level graph to encode the co-occurrence information of words, and distinguishes various types of dependency or word pair relationships through independent hierarchical syntax; SenticGCN: builds a graph neural network by integrating SenticNet sentiment knowledge, effectively enhancing the dependency graph structure of the sentence. This innovative sentiment-enhanced graph model not only accurately captures different dimensions of sentiment features, but also comprehensively reveals the relevance between specific dimensions and context information; Standard Prompting: through the general language understanding ability obtained by pre-training, the input sentence and the target aspect are directly encoded together, and the corresponding sentiment polarity judgment is output; CoT: Chain-of-Thought (CoT) is a prompting technique that guides the model to explicitly generate a series of intermediate reasoning steps before generating the final answer. This step-by-step derivation mechanism significantly optimizes the model's ability to handle complex, multi-step reasoning tasks; SEDC-GCN: proposes a structure-enhanced dual-channel graph convolutional network (SEDC-GCN), which constructs text sequence graphs and enhanced dependency graphs and designs a dual-channel graph encoder, combining aspect-specific semantic and structural attention mechanisms, effectively fusing multi-perspective information to improve aspect-level sentiment classification performance.

[0043] In the model framework implementation, the Qwen-Turbo large language model is responsible for fast thinking and slow thinking stages, and outputs the corresponding sentiment prediction results and corresponding reasons. The temperature parameter is set to 0 to ensure the stability of the output. In the evaluation stage, a Moonshot-Kimi-K2-Instruct model different from the fast / slow thinking stage is selected to perform the evaluation task, avoiding homogenization cognitive bias, providing diversified analysis perspectives for the reasoning verification process, and improving the reliability of the evaluation link.

[0044] Experiments were conducted on three classic aspect-based sentiment analysis datasets, Laptop14, Restaurant14 and Restaurant15, to verify the effectiveness of the proposed framework. According to the experimental results shown in Table 2, it can be observed that compared with the comparative model, the FESCR framework proposed in the present application has achieved stable performance improvement on the three benchmark datasets.

[0045] Table 2 Specifically, the thought chain method outperforms the standard prompt method in accuracy and F1 score on the three datasets. The performance improvement is mainly due to the step-by-step reasoning guided by the thought chain method, which requires the model to demonstrate a complete reasoning process. This not only enhances the explainability of the analysis, but also reduces errors caused by jumping judgments through a structured reasoning process. However, the thought chain method still has limitations in performance improvement, and the FESCR framework proposed in this paper has achieved superior performance compared to the thought chain method, verifying the superiority of the FESCR architecture design. The FESCR framework proposed in this paper has achieved superior performance compared to the thought chain method, verifying the superiority of the FESCR architecture design. FESCR introduces an independent "evaluation-slow thinking" closed-loop mechanism, which can make efficient fast thinking intuitive judgments, and then identify potential errors through strict evaluation in the evaluation stage, and reflect and correct in the slow thinking stage. This dynamic correction capability effectively makes up for the defect that a single reasoning model cannot self-check and correct.

[0046] The FESCR framework significantly outperforms all traditional neural network comparison models. First, the performance of traditional neural network methods is limited by the limited labeled data under the supervised learning paradigm, while large language models learn rich language knowledge and world knowledge through pre-training on super large-scale corpus, have stronger generalization ability, and can better handle unseen expression methods and complex language phenomena. The FESCR framework not only fully releases the reasoning potential of large models, but also can further enhance the understanding of text through evaluation and slow thinking, and improve the performance of aspect sentiment analysis.

[0047] Through the above analysis, the FESCR framework integrates fast thinking and slow thinking into the aspect-level sentiment analysis process, and its accuracy exceeds that of traditional neural network methods and existing prompting methods, fully verifying the effectiveness of the framework.

[0048] To verify the effectiveness of each stage in the proposed FESCR framework, a series of ablation experiments were conducted, which tested the fast thinking stage (FT) alone, the fast thinking + slow thinking two-stage (FST), and compared them with the complete three-stage collaborative model (FESCR).

[0049] The results are shown in Table 3. First, the fast thinking stage (FT) alone has shown strong performance on each dataset, which proves that the thought chain prompt can effectively improve the model's sentiment analysis ability. However, the two-stage framework containing fast and slow thinking (FST) is slightly lower than the model with only fast thinking stage in multiple indicators. We believe that this may be due to the lack of evaluation model guidance, the slow thinking stage cannot accurately identify the cases that need to be corrected, and even may make unnecessary modifications to the correct fast thinking results, resulting in performance degradation. The complete three-stage framework achieves the best accuracy on all datasets, which fully proves the effectiveness of the framework. The evaluation model judges whether the reasoning process of fast thinking is correct or not, and generates a well-reasoned and clear conclusion report, so that slow thinking can improve the incorrect reasoning steps in the initial reasoning process, thereby improving the performance.

[0050] Table 3 In this experiment, the number of cases that only go through the fast thinking and evaluation stages (FET) and the complete three stages of fast thinking, evaluation and slow thinking (FEST) are counted to explore the role of evaluation and slow thinking stages in improving the performance of the framework. As shown in Figure 2 In the Lap4, Rest14 and Rest15 datasets, only a small number of samples trigger the slow thinking stage. First, the evaluation model's assessment of the correctness of the fast thinking stage's reasoning process is a prerequisite for the activation of the slow thinking stage. However, if the evaluation model's discrimination ability is insufficient, it may either miss the cases that truly require slow thinking, resulting in errors that cannot be corrected, or it may overactivate the slow thinking stage, thereby reducing system efficiency and even modifying correct reasoning processes into incorrect ones. Therefore, the accuracy of the evaluation model greatly affects the accuracy of the entire framework. Slow thinking plays a role in improving the performance of the framework by conducting in-depth reasoning analysis when the evaluation model considers it incorrect. In the above three datasets, it is fully demonstrated that the evaluation stage of the framework can effectively identify errors in the fast thinking reasoning process, and the slow thinking stage can effectively reflect and correct them, thereby verifying the important role of the evaluation and slow thinking stages.

[0051] Table 4 To more intuitively demonstrate the working mechanism of the FESCR framework, two representative cases are selected for in-depth analysis to effectively understand complex emotions and make correct emotional judgments. The specific process is shown in Table 4. In the first case, the input sentence is "great food", and the analysis target is the aspect "food". In the fast thinking stage, the reasoning large model accurately identifies the adjective "great" as a positive sentiment word directly modifying the target aspect, and judges the emotional tendency to be positive. In the evaluation stage, the evaluation large model confirms that the reasoning process is reasonable after evaluation, and outputs the correct reason, directly adopting the fast thinking result. This case demonstrates the high processing capacity of the framework in simple and clear emotional expression. When the reasoning of fast thinking is sufficiently reliable, the evaluation verifies it as correct, avoiding unnecessary deep reflection, and also reflects the intelligence of the framework in resource allocation.

[0052] The aspect in the case "this place has ruined me for neighborhood sushi" is "sushi", which demonstrates the value of complete three-stage cooperation. The reasoning large model first interprets the sentence at the surface level, judges the emotion to be "negative", and believes that the sentence indicates a strong negative experience. "Ruined me" implies that the sushi at this store has a negative impact on the speaker's ability or willingness to enjoy other neighborhood sushi, thus inferring dissatisfaction with the sushi. The evaluation large model evaluates this reasoning process and believes that the reasoning is incorrect, pointing out that "ruined me for neighborhood sushi" is an idiom expression that actually conveys a very positive emotion, meaning that the sushi at this store is so excellent that other neighborhood sushi pales in comparison. The fast thinking incorrectly interprets "ruined me" as dissatisfaction, but in this context it actually represents extreme satisfaction with the quality of the sushi. The reasoning large model in the slow thinking stage refers to the problem pointed out by the evaluation and re-reasoning, correcting the emotional judgment to "positive" and analyzing that the sentence uses an idiom expression, "ruined me" does not express negative emotions here, but emphasizes the outstanding quality of the current restaurant's sushi by comparison, reflecting the speaker's high satisfaction rather than dissatisfaction.

[0053] As Figure 4As shown, the present application also provides a language analysis system based on the fast thinking-evaluation-slow thinking aspect-level sentiment analysis framework, comprising a memory, a processor, a computer program stored in the memory and executable on the processor, and the processor executes the computer program to implement the steps of the language analysis method based on the fast thinking-evaluation-slow thinking aspect-level sentiment analysis framework according to any one of the above embodiments.

[0054] Figure 4 is a hardware structure schematic diagram for implementing the language analysis method based on the fast thinking-evaluation-slow thinking aspect-level sentiment analysis framework provided by an embodiment of the present application. As shown in Figure 4 the embodiment / computer 6 comprises a processor 60, a memory 61, and a computer program 62 stored in the memory 61 and executable on the processor 60, such as a program for implementing the language analysis method based on the fast thinking-evaluation-slow thinking aspect-level sentiment analysis framework. The processor 60 executes the computer program 62 to implement the steps in each of the above embodiments of the language analysis method based on the fast thinking-evaluation-slow thinking aspect-level sentiment analysis framework. Alternatively, the processor 60 executes the computer program 62 to implement the functions of each module / unit in each of the above device embodiments.

[0055] For example, the computer program 62 can be divided into one or more modules / units stored in the memory 61 and executed by the processor 60 to complete the present application. The one or more modules / units can be a series of computer program instruction segments capable of completing a specific function, which are used to describe the execution process of the computer program 62 in the computer 6.

[0056] The computer 6 can be a desktop computer, a notebook, a palm computer, and a cloud server, etc. The computer 6 device can include, but is not limited to, the processor 60, the memory 61. Those skilled in the art can understand, Figure 4 The computer 6 is only an example and does not constitute a limitation on the computer 6, which can include more or fewer components than shown, or combine certain components, or different components, for example, the computer 6 can also include an input / output device, a network access device, a bus, etc.

[0057] The processor 60 can be a central processing unit (CPU), and can also be other general-purpose processors, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0058] The memory 61 can be an internal storage unit of the computer 6, such as a hard disk or a memory of the computer 6. The memory 61 can also be an external storage device of the computer 6, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Further, the memory 61 can also include both the internal storage unit and the external storage device of the computer 6. The memory 61 is used to store the computer program and other programs and data required by the terminal device. The memory 61 can also be used to temporarily store data that has been output or will be output.

[0059] It can be clearly understood by those skilled in the art that, for the convenience and brevity of description, only the division of the above functional units and modules is exemplified, and in actual application, the above functions can be completed by different functional units and modules according to needs, that is, the internal structure of the apparatus is divided into different functional units or modules to complete all or part of the above-described functions. Each functional unit and module in the embodiment can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit, and the integrated unit can be realized in the form of hardware or in the form of a software functional unit. In addition, the specific names of each functional unit and module are only for the purpose of mutual distinction, and do not limit the protection scope of the present application. The specific working process of the units and modules in the above system can refer to the corresponding process in the foregoing method embodiments, and will not be described here.

[0060] In the above embodiments, the description of each embodiment has its own emphasis, and the parts not described or recorded in detail in a certain embodiment can refer to the relevant description of other embodiments.

[0061] Those skilled in the art can understand that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware or in a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0062] In the embodiments provided by the present application, it should be understood that the disclosed apparatus / terminal device and method can be implemented in other ways. For example, the apparatus / terminal device embodiments described above are merely schematic, and the division of the modules or units is merely a logical function division, and there can be another division manner in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between the units can be indirect couplings or communication connections through some interfaces, devices or units, and can be electrical, mechanical or in other forms.

[0063] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e. can be located in one place, or can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0064] In addition, each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can exist physically independently, or two or more units can be integrated into one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0065] The integrated module / unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on such understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer-readable storage medium. When the processor executes the computer program, the steps of the above-mentioned various method embodiments can be implemented. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate forms, etc. The computer-readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content included in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0066] The above is only a specific application example of the present application, and does not constitute any limitation on the protection scope of the present application. In addition to the above-mentioned embodiments, the present application can also have other implementation manners. Any technical solution formed by equivalent replacement or equivalent transformation falls within the scope of the present application.

Claims

1. A language analysis method based on an aspect-level sentiment analysis framework of fast thinking-evaluation-slow thinking, characterized in that, Includes the following steps: S1, Definition of Aspect-Level Sentiment Analysis: Given a sentence X and its contained aspect terms A, construct a system to determine the sentiment polarity of the sentence X for aspect A. Target tasks ∈ {positive emotions, negative emotions, neutral emotions}; S2, Fast Thinking Stage: Guided by fast thinking cues, the reasoning model is primarily executed, simulating System One of the dual-process theory in cognitive psychology, and defining the objective function of the fast thinking stage ( quick ,R quick =LLM reasoning ([S,A],Prompt quick ); where LLM reasoning This represents a large-scale inference model, where S represents the input sentence, A represents the target aspect term, and Prompt... quick The prompts for the rapid thinking stage are used to determine the output based on preliminary sentiment prediction. quick and the corresponding reasoning process R quick composition; S3, Evaluation Phase: The evaluation phase is guided by evaluation prompts and led by the large-scale evaluation model. Domain labels are injected into the prompts to construct a domain-specific cognitive context for the large-scale evaluation model. The objective function for this phase is defined as: (c eval E eval =LLM eval ([S,A,R quick ,domain],Prompt evl ); where S represents the input sentence, A represents the target aspect, and R represents the target aspect. quick To illustrate the reasoning process of a large reasoning model during the fast thinking phase, `domain` represents the domain label, and `Prompt` represents the reasoning process of the model. evl c is a cue word for the assessment phase. evl For the correctness assessment results, E eval For the evaluation process; S4, Slow Thinking Phase: Simulating System 2 in dual-process theory, the reasoning model is guided by slow thinking cues and primarily executes the process. When the evaluation phase determines the reasoning model to be incorrect, the large-scale reasoning model enters the slow thinking phase. The objective function of the slow thinking phase is ( slow ,R slow =LLM reasoning ([S,A,domain,R quick E evl Prompt slow ); where S represents the input sentence, A represents the target aspect, domain represents the domain label, and R represents the target area. quick E evl The output processes for the quick thinking and evaluation phases are respectively, Prompt. slow These are cue words for the slow thinking phase. slow The sentiment labels, ultimately confirmed after deep reflection, serve as the final prediction results for the task. slow This is the final reasoning process.

2. The language analysis method based on the aspect-level sentiment analysis framework of fast thinking-evaluation-slow thinking as described in claim 1, characterized in that, The Qwen-Turbo large language model is used to handle the fast thinking and slow thinking stages.

3. The language analysis method based on the aspect-level sentiment analysis framework of fast thinking-evaluation-slow thinking according to claim 2, characterized in that, During the fast and slow thinking phases, the corresponding sentiment prediction results and reasons are output, and the temperature parameter is set to 0 to ensure output stability.

4. The language analysis method based on the aspect-level sentiment analysis framework of fast thinking-evaluation-slow thinking as described in claim 1, characterized in that, The evaluation phase employs the Moonshot-Kimi-K2-Instruct model to perform the evaluation task.

5. A language analysis system based on a fast-thinking-evaluation-slow-thinking aspect-level sentiment analysis framework, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the language analysis method based on the aspect-level sentiment analysis framework of fast thinking-evaluation-slow thinking as described in any one of claims 1 to 4.

6. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the language analysis method based on the aspect-level sentiment analysis framework of fast thinking-evaluation-slow thinking as described in any one of claims 1 to 4.

Images