A Product Review Detection Method Based on Reflective Chain Reasoning

Abstract

A product review detection method based on reflective chain reasoning belongs to the technical field of natural language processing (NLP). By structuring relevant background information into background feature representations and combining a "reflective chain" reasoning mechanism, the reasoning conclusion is gradually refined. Based on few-shot reasoning, this method can effectively improve the model's detection ability for low-scoring background-dependent cases through the generation and screening evaluation of feature combinations. At the same time, the screening and scoring mechanism of feature combinations can quickly filter out irrelevant information, reduce redundant calculations, and improve detection efficiency. By introducing termination conditions, this method can timely terminate the exploration of irrelevant paths during the reasoning process, thereby further optimizing the use of computing resources. The present invention exhibits stronger accuracy and context adaptation ability, can understand complex semantics more deeply, and significantly improves the ability to capture metaphorical expressions.

Description

Technical Field

[0001] The present invention belongs to the technical field of natural language processing (NLP), and particularly relates to a product review detection method based on reflective chain reasoning. Background Art

[0002] With the development of e-commerce and social media, a large amount of information about product performance, usage experience, and emotional feedback is contained in user-generated content (UGC), which has important commercial value. Accurately identifying users' tendency towards specific products has become a key technical link for enterprises in product optimization, user operation, and market strategy formulation.

[0003] In recent years, large language models (LLMs) have made significant progress in tasks such as sentiment analysis and text classification. However, in the actual analysis of social media product reviews, the existing methods generally have the following deficiencies: First, relying on sentiment lexicons or shallow semantic features, it is difficult to identify the deep emotional tendencies hidden in the reviews; second, lacking the ability to model context information such as product background knowledge, technical terms, and usage scenarios, resulting in limited accuracy of sentiment judgment in complex contexts.

[0004] In addition, although the long-chain reasoning based on large language models can improve semantic understanding ability to a certain extent, its reasoning path often lacks flexibility, the reasoning process is time-consuming, and it is easy to introduce noise in multi-round thinking, reducing the detection efficiency. Therefore, how to combine product background knowledge, mine the deep tendency in user reviews, and improve the reasoning efficiency at the same time is a technical problem that urgently needs to be solved in the current sentiment analysis field. Summary of the Invention

[0005] In order to more efficiently detect the deep emotions hidden in user reviews and comprehensively consider the background knowledge related to user remarks, the present invention proposes an innovative detection method. This method innovatively structures product-related background knowledge (such as parameter configuration, applicable scenarios, user groups, etc.) into background features, and combines them with emotional features to form structured feature combination items, enhancing the judgment ability of the large model. In addition, the existing methods in the field of large model text processing often rely on long reasoning chains with fixed thinking routes, which are unstable and time-consuming. To address this problem, the present invention designs a new mechanism that uses combined features to guide the large model to perform multiple rounds of rejudgment (reflective chain) on the original conclusion, without being restricted to a fixed thinking route, and preferentially uses high-value feature combination items for rejudgment, and immediately ends the reasoning after detecting the text emotion, improving the detection efficiency. Generally speaking, the present invention strengthens the ability to mine deep emotions while taking into account the detection efficiency.

[0006] The object of the present invention is as follows: There are obvious deficiencies in the existing product review detection methods for social media user comments. On the one hand, the language expressions on social media are usually implicit and complex, making it difficult to be accurately captured by traditional methods. On the other hand, such remarks are often closely related to the product background information, but the existing sentiment detection technologies are difficult to effectively utilize this specialized knowledge. In addition, although the long inference step method based on large language models can capture complex semantics, the extension of the inference chain will lead to a significant increase in the inference time consumption. To address these problems, the present invention proposes a detection method based on large language models, which innovatively structures product information as background features. Subsequently, the background features and sentiment features are selected and combined to form "feature combination items" for enhancing the large language model's understanding ability of background-dependent information. At the same time, a "reflection chain" multi-round inference mechanism based on feature combination items is designed. In each round, the most valuable feature combination item is selectively used to guide the large model to rejudge the original conclusion, significantly improving the detection ability of text sentiment and helping to increase the low-score detection efficiency. This method provides a new technical path and implementation means for efficient and accurate user feedback understanding.

[0007] The technical solution of the present invention is as follows:

[0008] A social media product review detection method based on reflection chain inference, comprising the following steps:

[0009] Step S1: Extract product-related information from the public corpus as background features, and standardize the background features through source credibility evaluation and timeliness filtering to obtain a background feature set;

[0010] Step S2: Determine whether the short text to be detected mentions the target product. If so, extract multi-dimensional sentiment features in the short text to be detected, otherwise terminate the detection;

[0011] Step S3: Use the large language model for preliminary sentiment classification to obtain an initial sentiment label;

[0012] Step S4: Based on the short text to be detected, sentiment features, and background features, use the large language model to match the sentiment features with the sentiment features to generate m feature combination items, and score and rank each feature combination item;

[0013] Step S5: Iteratively verify according to the priority based on the reflection chain mechanism until the confidence threshold is reached or all feature combination items are traversed;

[0014] Step S6: Output the final detection result.

[0015] Further, the standardization includes source filtering, length trimming, and knowledge weight calculation. The knowledge weight The calculation process is as follows:

[0016] ;

[0017] Among them, is the credibility of the information source, is the matching degree between texts, is the timeliness index of the retrieved information; , and are weighting coefficients, .

[0018] Furthermore, the acquisition of the emotional features described in step S2 specifically uses a large language model to analyze the short text to be detected, and extracts the emotional features of the short text to be detected in multiple dimensions. The formulaic description is as follows:

[0019] ;

[0020] Among them, is the short text to be detected; is the emotional feature of the j-th feature dimension extracted; is the probabilistic representation in the output process of the large model, j is the index of the feature dimension, and N is the number of feature dimensions.

[0021] Furthermore, the specific steps of step S4 are as follows:

[0022] The criteria for generating feature combination items include: one feature combination item contains one background feature and several emotional features; the selection of features depends on the large language model. Set the number of feature combination items to m, and the formulaic description for generating feature combination items is as follows:

[0023] ;

[0024] ;

[0025] Among them, represents the set of m feature combination items, is the i-th feature combination item, which contains the product background feature and several emotional features; is the product name mentioned in the text; is the background feature set; represents the background feature in the -th feature combination item; is the number of emotional features in the -th feature combination item, and the number is determined by the large model itself; represents a certain subset in the set of emotional features ; is the The levels of the feature combination items include extremely high, relatively high, medium, relatively low, and extremely low;

[0026] The calculation formula for obtaining the probability value of the feature combination item from the level is as follows:

[0027] ;

[0028] Where: is the probability value corresponding to the probability level assigned by the large language model to the th feature combination item; The calculation formula for scoring each feature combination item is as follows:

[0029] ;

[0030] Among them, is the score of the th feature combination item; is the weight of the product background feature in the th feature combination item; represents the rectified linear unit function; is the set filtering threshold;

[0031] Finally, the remaining feature combination items are sorted from high to low according to the scores.

[0032] Furthermore, the introspection chain mechanism is specifically as follows:

[0033] First, take the item with the highest score among the remaining feature combination items, connect the preliminary conclusion obtained in step S3, the original text, and the feature combination item to construct a prompt word, and input it into the large model for re-verification. If there is a target detection sentiment or all combination items are exhausted, the introspection is terminated. The formulaic description is as follows:

[0034] ;

[0035] Where, is the result of the th round of re-verification; represents the initial conclusion obtained in step S3; is the total number of re-verification times.

[0036] Furthermore, it includes lexical features, artistic technique features, viewpoint features, tone features, context features, and emotion features.

[0037] Furthermore, the final detection result includes a classification binary label and an explanatory conclusion.

[0038] Compared with the existing technology, the beneficial effects of the present invention are:

[0039] 1. Stronger sentiment analysis capabilities. By introducing structured background knowledge representation, the present invention makes full use of the background information of Internet products to assist in the detection of positive or negative comments on products. Compared with traditional methods that rely on surface text features, this method significantly improves the large language model's understanding and application capabilities of background knowledge by designing a generation and evaluation method for "feature combination items". Especially in situations strongly related to the background of the target product, it shows stronger accuracy and context adaptation capabilities, and can understand complex semantics more deeply. In addition, the "introspection chain" reasoning mechanism imitates the human introspection process, and further improves the detection ability of implicit emotions by gradually guiding the model to conduct multiple rounds of rejudgment.

[0040] 2. Optimized detection efficiency. During the detection process, the processes of selecting, scoring, sorting, and filtering feature combination items can quickly filter out irrelevant information, and preferentially use high-value combination items to enter the rejudgment stage. This strategy not only reduces redundant calculations, but also significantly improves the detection efficiency for positive samples.

[0041] 3. Scalability of application scenarios. The present invention not only demonstrates excellent performance in the current sentiment detection task, but its core technical ideas (such as "feature combination items" and "introspection chain") can also be further applied to a wider range of complex semantic analysis tasks, such as metaphor detection, stance detection, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a detailed flowchart of the introspection chain reasoning process.

[0043] Figure 2 It is a schematic diagram of the working mechanism of the introspection chain.

[0044] Figure 3 It is a structural example diagram of background features in the proposed background feature construction method.

[0045] Figure 4 It is a schematic diagram of the feature combination item construction process. DETAILED DESCRIPTION OF THE INVENTION

[0046] To better understand the purpose, structure, and function of the present invention, the following further describes in detail a method for detecting social media product comments based on introspection chain reasoning of the present invention with reference to the accompanying drawings.

[0047] This embodiment provides a social media product review detection method based on introspective chain reasoning, which can efficiently and accurately identify negative metaphorical expressions by combining product background knowledge and multi-round reasoning mechanism. Traditional sentiment detection methods mainly rely on text surface features or simple context semantic matching, but these methods often appear to be powerless when dealing with obscure and complex negative expressions. In addition, existing technologies usually ignore the close correlation between social media speech and specific background information, resulting in insufficient detection capabilities for implicit negative emotions.

[0048] In response to such problems, this embodiment proposes a comment sentiment detection scheme with "feature combination items" as the core, which structures the relevant information retrieved from encyclopedia entries and authoritative media into background features, combines them with sentiment features into "feature combination items", and combines them with the "reflection chain" reasoning mechanism to gradually refine the reasoning conclusions. This method can effectively improve the model's detection ability for background-dependent speech through the generation, screening and evaluation of feature combination items. At the same time, the screening and scoring mechanism of feature combination items can quickly filter out irrelevant information, reduce redundant calculations, and improve detection efficiency. In addition, by introducing termination conditions, this method can terminate the exploration of irrelevant paths in a timely manner during the reasoning process, thereby further optimizing the use of computing resources. This embodiment focuses on solving the problems of insufficient use of background knowledge and poor stability of the reasoning chain in existing sentiment detection technologies, especially in product object-related scenarios, showing significant detection performance improvements and efficiency advantages.

[0049] A social media product review detection method based on introspection chain reasoning, such as Figure 1 As shown, the following steps are included:

[0050] Step S1: extract product-related information from the public corpus as background features, standardize the background features through source credibility evaluation and timeliness filtering, and obtain a background feature set;

[0051] Retrieve information published on multiple encyclopedia platforms, shopping platforms, and certified media through online search tools;

[0052] Standardize each retrieved information item. Cut the length of each item to ensure that it does not exceed the upper limit. Filter the source to ensure the authority and authenticity of the information. Calculate the weight for the subsequent scoring process. The higher the weight, the more important the item. The combination of background knowledge items and weights constitutes the so-called "background features", such as Figure 3 shown.

[0053] Furthermore, the standardization includes information source filtering, length clipping and weight calculation. The calculation method is as follows:

[0054]

[0055] where is the preset source credibility, is the matching degree between texts, is the timeliness index of the retrieved information. , and are the preset weighting coefficients, and there is .

[0056] Step S2: Determine whether the short text to be detected mentions the target product. If so, extract the sentiment features in the short text to be detected; otherwise, abandon the subsequent detection; the target product name is extracted based on the named entity recognition model;

[0057] Specifically, the sentiment feature acquisition uses a large language model to analyze the short text to be detected, and extracts the features of the speech to be analyzed from multiple dimensions, including six categories: lexical features, artistic technique features, viewpoint features, tone features, context features, and emotion features. The formulaic description is as follows:

[0058]

[0059] where: is the given short comment to be detected; is the sentiment feature of the j-th feature dimension extracted; is the probabilistic representation in the output process of the large model, and j is the index of 6 feature dimensions;

[0060] The prompt template that may be used for sentiment feature acquisition can specifically be: Given the short comment to be detected [Sentence], the purpose is to detect low scores (or low scores, or neutral evaluations, the following is described based on low scores), please extract six features including lexical features, artistic technique features, viewpoint features, tone features, context features, and sentiment features from the text, and the output format: [Feature category]{Content}. For example, [Artistic technique feature]{This text uses exaggeration…}

[0061] In a comment detection, highlight the sentiment orientation of one aspect to help determine the reflection direction in the subsequent reflection chain reasoning.

[0062] Step S3: Use the few-shot two-step large model thinking chain to guide the model to infer a preliminary conclusion from shallow to deep, and this part uses the basic ability of the large language model for simple judgment.

[0063] The prompt template for obtaining the preliminary conclusion can be: Given the short comment to be detected: [Sentence], the name of the product being commented on: [Target], it is necessary to detect whether the text contains a low rating for the target product. The format of the final conclusion: [Yes / No constitutes a low rating] {A summary of the view in one paragraph}; The formulaic description is as follows:

[0064]

[0065] Among them, is the preliminary conclusion output, which is a combination of a label and an explanatory conclusion; is the product name mentioned in the text;

[0066] Step S4: Based on the original short text to be detected, the extracted sentiment features, and the background features, use the large language model to match the sentiment features with the background features to generate multiple feature combination items, and score and rank each combination, as Figure 4 shown.

[0067] The criteria for generating feature combination items include: one feature combination item contains one background feature and several sentiment features; the selection of features depends entirely on the large language model, and only need to guide the large language model to select several groups of combinations that are most likely to be helpful for detection from the background features and the sentiment features extracted in the previous steps; the number of combination items needs to be determined manually, for example, fixed at 5 items, and the formulaic description is as follows:

[0068]

[0069]

[0070] Among them, represents the set of 5 feature combination items, is the i-th feature combination item, which contains a background feature and several sentiment features, and is selected by the large model from the background features and sentiment features; is the product name mentioned in the text; is the set of background features; represents the background feature in the i-th combination item; is the number of sentiment features in the i-th combination item, which is determined by the large model itself; represents the sentiment feature a certain subset in this set . is the rating of the i-th combination item, including five levels: extremely high, relatively high, medium, relatively low, and extremely low.

[0071] The following are the prompt templates that may be used when generating feature combination items: Given the short comment to be detected: [Sentence], the product mentioned: [Target], it is necessary to detect whether the text has a low score. Select the five background features that are most helpful for this sentiment detection from the following background feature set, and match several sentiment features that are helpful for detection to each of them. Finally, give the probability level of a low score, including extremely high, relatively high, medium, relatively low, and extremely low. For each combination, the output format is: [background feature, sentiment feature 1,..., sentiment feature n, probability level]. For example: [12, 2, 5, relatively high]. Background feature set: [Knowledge]. Sentiment feature set: [Feature].

[0072] Obtain the grading result from the large model; guide the large model to output a level instead of a numerical value because the level output is usually more interpretable than the probability value. The linear piecewise provides an intuitive mapping relationship and to a certain extent conforms to the cognitive habits of humans. The numerical value in the 0-1 interval can be naturally interpreted as a probability. Therefore, the calculation formula for obtaining the probability value of the feature combination item from the level is as follows:

[0073]

[0074] where, is the probability value corresponding to the probability level assigned by the large language model to the th feature combination item.

[0075] The calculation formula for scoring each feature combination item is as follows:

[0076]

[0077] where, is the score of the th feature combination item; is the weight of the background feature in the th feature combination item; represents the rectified linear unit function, which is used to suppress the score below the threshold; is the artificially set filtering threshold, and the results below this threshold will be suppressed by the rectified linear unit function. To improve the detection efficiency, the feature combination items with a score of zero are excluded, and the remaining feature combination items are sorted from high to low according to the score.

[0078] Step S5: Based on the introspection chain mechanism and the feature combination items, perform multiple rounds of cyclic rejudgment, gradually refine the conclusion from top to bottom until it is detected as a low score or all combinations are traversed. The process is as Figure 2 shown.

[0079] The process of rejudging (reflecting) using combined item pairs includes: First, select the item with the highest score from the remaining feature combination items, and then construct a prompt by connecting the preliminary conclusion obtained in step S3, the original short text to be detected, and the feature combination item, and input it into the large model for rejudging. This process guides the large language model to "reflect" on the original conclusion and find possible overlooked clues from the prompts of the feature combination items. If there is low-scoring information or all combination items are exhausted, terminate the reflection; the formulaic description is as follows:

[0080]

[0081] Where: is the result of the round of rejudging; represents the initial conclusion obtained in step S3; is the total number of rejudging times.

[0082] The prompt template that may be used for each rejudging is as follows: Given the short comment to be detected: [Sentence], the target: [Target], it is necessary to detect whether the text contains a low score for the target product. Some people think that [Opinion]. The following gives a set of materials for this task, involving sentiment features and background features, which may supplement the above view. Please analyze again and draw a new conclusion [Yes / No constitutes a low score]{Viewpoint}, the materials are as follows: [Combination]

[0083] Step S6: Output the final detection result; the output result includes two parts: a classification binary label and an explanatory conclusion.

[0084] It can be understood that the present invention is described through some embodiments. Those skilled in the art know that without departing from the spirit and scope of the present invention, various changes or equivalent replacements can be made to these features and embodiments. In addition, under the teaching of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application belong to the scope protected by the present invention.

Claims

1. A product review detection method based on introspective chain reasoning, characterized in that, It includes the following steps: Step S1: Extract product-related information from the public corpus as background features, and standardize the background features through source credibility evaluation and timeliness filtering to obtain a background feature set; the standardization includes source filtering, length trimming, and knowledge weight calculation. The knowledge weight ω of the background feature is calculated as follows: ω = α·Authority + β·Relevance + γ·Freshness; where Authority is the source credibility, Relevance is the matching degree between texts, and Freshness is the timeliness index of the retrieved information; α, β, and γ are weighting coefficients, and α + β + γ = 1; Step S2: Determine whether the short text to be detected mentions the target product. If so, extract the multi-dimensional sentiment features in the short text to be detected; otherwise, terminate the detection. The acquisition of the sentiment features specifically uses a large language model to analyze the short text to be detected and extract the sentiment features of the short text to be detected multi-dimensionally. The formulaic description is as follows: Among them, S is the short text to be detected; is the sentiment feature of the j-th feature dimension extracted; argmaxp() is the probabilistic representation of the output process of the large model, j is the index of the feature dimension, and N is the feature dimension; Step S3: Use a large language model to perform preliminary sentiment classification to obtain an initial sentiment label; Step S4: Based on the short text to be detected, sentiment features, and background features, use a large language model to match the sentiment features with the sentiment features to generate m feature combination items, and score and rank each feature combination item; the criteria for generating feature combination items include: one feature combination item contains one background feature and several sentiment features; the selection of features depends on the large language model. Set the number of feature combination items to m, and the formulaic description for generating feature combination items is as follows: Among them, represents a set of m feature combination items, M i is the i-th feature combination item, including product background features and several sentiment features; t is the product name mentioned in the text; K is the background feature set; represents the background feature in the i-th feature combination item; k i is the number of sentiment features in the i-th feature combination item, which is determined by the large model itself; represents the sentiment feature a certain subset in this set G i is the level of the i-th feature combination item, including extremely high, relatively high, medium, relatively low, extremely low; The formula for obtaining the probability value of the feature combination item from the hierarchy is as follows: where: ε i is the probability value corresponding to the probability level assigned by the large language model to the I-th feature combination item; the calculation formula for scoring each feature combination item is as follows: Score i = ReLU(ω i ε i -δ o ); Among them, Score i is the score of the i-th feature combination item; ω i is the weight of the product background feature in the i-th feature combination item; ReLU represents the rectified linear unit; δ0 is the set filtering threshold; Finally, sort the remaining feature combination items in descending order of scores; Step S5: Iteratively verify based on the reflection chain mechanism according to the priority until the confidence threshold is reached or all feature combination items are traversed; the reflection chain mechanism is specifically as follows: First, take the item with the highest score among the remaining feature combination items, connect the preliminary conclusion obtained in Step S3, the original text, and the feature combination item to construct a prompt, and input it into the large model for re-verification. If the target detection sentiment exists or all combination items are exhausted, terminate the reflection. The formulaic description is as follows: O a = argmax p(O a | S, M i , O0, t), a = 1, 2, …, N a ; Among them, O a is the result of the a-th round of rejudgment; O0 represents the initial conclusion obtained in step S3; N a is the total number of rejudgments; Step S6: Output the final detection result.

2. The product review detection method based on introspective chain reasoning according to claim 1, characterized in that The feature dimensions are 6 types, including lexical features, artistic technique features, opinion features, tone features, context features, and emotion features.

3. The product review detection method based on introspective chain reasoning according to claim 2, wherein, The final detection result includes a classification binary label and an explanatory conclusion.

Images