A live commodity propaganda and record information consistency intelligent checking method and system

Abstract

This invention relates to the field of compliance review technology for live-streaming e-commerce, and particularly to an intelligent verification method and system for consistency between live-streaming product promotion and registration information. The method includes: extracting keywords based on key registration fields, assigning baseline weights according to field type labels, constructing a dynamic hot word table, and injecting hot word bias coefficients into a speech recognition decoder; performing fuzzy matching between the recognized text and keywords, and calibrating the successfully matched recognized text using keywords; performing semantic structure analysis on the calibrated text based on preset positive and negative trigger word sets, extracting and outputting noun phrases and their claim types; and comparing the noun phrases and their claim types with the corresponding key registration fields to generate a consistency verification result. This invention can improve the accuracy of identifying key registration information in noisy live-streaming environments while maintaining seamless integration with existing compliance review systems.

Description

Technical Field

[0001] This invention relates to the field of compliance review technology for live-streaming e-commerce, and in particular to an intelligent verification method and system for consistency between live-streaming product promotion and filing information. Background Technology

[0002] In the live-streaming e-commerce sector, to ensure consistency between product promotional content and registration information, regulatory authorities require real-time monitoring of the content delivered by livestreamers. Mainstream compliance review systems typically employ a combination of speech recognition and text comparison: first, a general speech recognition engine transcribes the livestreamer's content into text; then, this text is compared with the product registration information (such as ingredients, efficacy, and target audience) to determine if exaggerated or false advertising exists.

[0003] General-purpose speech recognition technologies, such as end-to-end models based on Transformers, are relatively mature in quiet environments with standard pronunciation. However, when deployed directly in complex live-streaming scenarios, they face the following practical challenges: continuous background music, frequent interspersed interactive sound effects, fast-paced speech by the host often with regional accents, and the host frequently flipping through product packaging while introducing products, creating a natural time difference between hand gestures and vocalizations. These factors combined lead to a significant decrease in the accuracy of general-purpose speech recognition engines in recognizing key information. Summary of the Invention

[0004] The main objective of this invention is to provide an intelligent verification method and system for the consistency between live-stream product promotion and registration information, aiming to solve the technical problems mentioned in the background section.

[0005] This invention proposes an intelligent verification method for the consistency between live-stream product promotion and registration information, including: In response to the listing of live-streamed products, the system obtains the product's registration information, which includes key registration fields with field type tags. Key words are extracted based on the key fields of the filing, and a benchmark weight is assigned according to the field type label to construct a dynamic hot word table. Hot word bias coefficients are obtained based on the dynamic hot word table, and the hot word bias coefficients are injected into the speech recognition decoder so that the speech recognition decoder positively biases the key words in the dynamic hot word table during recognition. The audio stream from the live broadcast room is collected and sent to the speech recognition decoder for real-time decoding, outputting the recognized text and the corresponding audio segment. Acoustic feature representations are extracted from the audio segment and feature hash values ​​are generated. A cache record matching the feature hash value is queried in the preset cache. If it exists, the stored noun phrases and claim types are directly output according to the cache record. If it does not exist, the identified text is fuzzily matched with the keywords, and the identified text that successfully matches is calibrated using the keywords to obtain the calibrated text. Based on the preset set of affirmative and negative trigger words, the calibrated text is subjected to semantic structure parsing, and noun phrases and their claim types are extracted and output. The noun phrase and its claim type are stored in the cache using the feature hash value as the key; The noun phrases and their claim types are compared with the corresponding key fields of the filing to generate a consistency verification result.

[0006] The beneficial effects of this invention are as follows: By dynamically loading registration information and constructing a weighted hot word list when products are listed, combined with post-processing semantic structure parsing and incremental caching, this invention can improve the accuracy of identifying key registration information in noisy live streaming environments, while maintaining the ability to connect with existing compliance review systems. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of a method flow according to an embodiment of this application.

[0008] Figure 2 This is a schematic diagram of the system structure according to an embodiment of this application.

[0009] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0010] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0011] In the live-streaming e-commerce scenario, regulatory compliance review requires real-time judgment of whether the content of the anchor's promotion of the product is consistent with the information filed with the regulatory agency. To achieve this goal, it is necessary to accurately identify the key information of the filing in the noisy live-streaming environment and correctly interpret the semantics of the anchor's expression of the key information. Therefore, the present invention constructs the following method steps.

[0012] like Figure 1 As shown, this application provides an intelligent verification method for the consistency between live-stream product promotion and registration information, including: S1, in response to the live-stream product listing operation, obtain the product's registration information, which includes key registration fields with field type tags; S2, extract keywords based on the filing key fields, allocate benchmark weights according to the field type labels, construct a dynamic hot word table, obtain hot word bias coefficients based on the dynamic hot word table, and inject the hot word bias coefficients into the speech recognition decoder so that the speech recognition decoder positively biases the keywords in the dynamic hot word table during recognition. S3, acquire the audio stream from the live broadcast room, send the audio stream to the speech recognition decoder for real-time decoding, and output the recognized text and the corresponding audio segment; S4, extract acoustic feature representations based on the audio segment, generate feature hash values, and query whether a cache record matching the feature hash value exists in a preset cache: S41, if it exists, then directly output the stored noun phrases and claim types according to the cached records; S42, if it does not exist, perform fuzzy matching between the identified text and the keywords, and calibrate the successfully matched identified text using the keywords to obtain the calibrated text; S5. Based on the preset set of affirmative trigger words and set of negative trigger words, perform semantic structure parsing on the calibrated text, extract and output noun phrases and their claim types; S6, store the noun phrase and its claim type into the cache using the feature hash value as the key; S7. Compare the noun phrase and its claim type with the corresponding filing key fields to generate a consistency verification result.

[0013] In this embodiment, when a live-streamed product is listed on the control panel by an operator, the system first responds to the operation and obtains the product's registration information. This information is obtained by initiating an API call to the merchant's backend database or the regulatory registration database via the product identifier. The returned registration information contains multiple key registration fields with field type tags. These field type tags are used to distinguish fields of different natures within the registration information; for example, ingredient names are marked as ingredient categories, efficacy descriptions as efficacy categories, applicable populations as population categories, and warning statements as warning categories. These field type tags are used for differentiated weight allocation when subsequently constructing a dynamic hot word list, enabling the bias parameters to reflect the differences in difficulty of different words in speech recognition.

[0014] After obtaining the registration information, the system extracts keywords from the key fields of the registration and assigns a baseline weight to each keyword based on the field type label, constructing a dynamic hot word table. The structure of the dynamic hot word table associates each keyword with its baseline weight, providing a data foundation for subsequent calculation of differentiated hot word bias coefficients. Based on the dynamic hot word table, the system calculates the hot word bias coefficient for each word and injects it into the decoder in the form of a bias vector through the native hot word weighting interface of the speech recognition decoder. The injected hot word bias coefficient causes the decoder to apply a positive bias to the keywords in the dynamic hot word table during subsequent recognition. The bias strength is related to the field type of the word, giving stronger path protection to obscure terms such as ingredient names, followed by efficacy words, and then warning and target audience words. This differentiated bias mechanism makes the decoder more inclined to output the correct registration keywords under noise interference, improving the recognition accuracy from the source. The construction of the dynamic hot word table and the bias injection are completed after the product is listed and before the anchor begins to introduce the product.

[0015] After the hot word bias coefficients are injected into the decoder, the system continuously collects the audio stream from the live broadcast room and feeds it into the speech recognition decoder, which has been configured with bias, for real-time decoding. The decoder adopts an end-to-end model that supports hot word weighting and streaming decoding, processes the audio data frame by frame, applies path cost discounts to the keywords in the dynamic hot word list during the beam search decoding process, and outputs three parts of the result in real time: the recognized text, the timestamp information of the audio segment corresponding to the recognized text, and the data in real time.

[0016] After acquiring the identified text and its corresponding audio segment, the system first performs acoustic feature extraction and cache lookup to determine whether the audio segment's content has been processed in a previous live stream. Based on the audio segment, acoustic feature representations are extracted and feature hash values ​​are generated. Then, the system checks a preset cache for a matching feature hash value that has not expired. If a cache hit occurs, the system directly reads and outputs the stored noun phrases and claim types from the cache record, bypassing subsequent fuzzy matching and semantic structure parsing steps. A cache hit means the audio segment has been successfully processed previously, and its semantic parsing results have been fully verified and stored, allowing for direct reuse and saving computational overhead. If a cache miss occurs, it indicates that the audio segment is being processed for the first time or the cache record has expired, requiring the system to execute a complete post-processing procedure.

[0017] The first step in post-processing is to perform fuzzy matching calibration between the recognized text and the keywords extracted from the registration information. Fuzzy matching calibration addresses potential homophone or near-homophone recognition errors that may still remain even after the decoder has been aided by hot word bias. It calculates a joint matching score between each recognized word and each keyword based on both glyph and pronunciation dimensions. Recognized words with matching scores higher than a preset replacement threshold are replaced with their corresponding keywords, resulting in calibrated text. This calibrated text provides input that more closely approximates the standard word forms used in the registration process for subsequent semantic parsing steps.

[0018] The second step of post-processing is to perform semantic structure parsing on the calibrated text based on a pre-defined set of affirmative and negative trigger words. The parsing process employs a trigger-word-based state machine, sequentially scanning the word segmentation and part-of-speech tagging results of the calibrated text. When an affirmative or negative trigger word is encountered, the process enters the corresponding state, extracting the continuous noun phrases following the trigger word and labeling the claim type of these noun phrases as affirmative or negative. Semantic structure parsing transforms the linear recognition text into structured semantic units. Each semantic unit consists of a noun phrase and a claim type, providing directly usable structured input for the final consistency comparison.

[0019] After semantic structure parsing is completed, the system stores the processing results of the audio segment obtained from fuzzy matching calibration and semantic parsing into a cache for reuse in subsequent hits of the same or similar audio segments. The content stored in the cache includes the noun phrases, claim types, and recognition confidence scores corresponding to the noun phrases output by semantic parsing, stored using the same feature hash value generated previously. The cache uses a least recently used eviction policy to manage storage space, and each record has a valid lifetime.

[0020] After obtaining the noun phrases and claim types, the system compares them with the key fields of the filing information initially obtained in the steps, generating a consistency verification result. The comparison process comprehensively examines various types of violations, including making affirmative claims about ingredients or effects not listed in the filing (exceeding the scope of the claim), making negative claims about ingredients or effects explicitly required in the filing (denying the required ingredients), and omissions in the broadcast of an ingredient or effect that is not mentioned at all. The system constructs a filing knowledge graph centered on key field nodes, performing structured matching and quantitative comparison of the broadcaster's affirmative and negative claim sets with the filing nodes to generate a comprehensive verification result reflecting compliance risks. The verification result may include immediate warning information and quantified compliance deviation indicators, allowing compliance reviewers to quickly locate problems.

[0021] In one embodiment of the present invention, the steps of extracting keywords based on the filing key fields, assigning benchmark weights according to the field type labels to construct a dynamic hot word table, obtaining hot word bias coefficients based on the dynamic hot word table, and injecting the hot word bias coefficients into the speech recognition decoder so that the speech recognition decoder positively biases the keywords in the dynamic hot word table during recognition include: S21, determine the baseline weight corresponding to the keyword based on the field type label, associate and store the keyword and its corresponding baseline weight to construct a dynamic hot word table; S22, obtain the product category to which the current live-streamed product belongs, offline statistics of the frequency of occurrence of keywords of all registered products under this category, and calculate the rarity factor of each keyword; S23, During the live broadcast, the signal-to-noise ratio of the current audio segment is estimated in real time, and the noise factor is calculated based on the signal-to-noise ratio; S24, use harmonic averaging to fuse the rarity factor and the noise factor of the same keyword to obtain the bias enhancement index of the keyword at the current time. S25, calculate the current hot word bias coefficient of each keyword using an exponential function based on the bias enhancement index, the benchmark weight, and the preset bias constant base. S26, construct a bias vector from the hot word bias coefficients of all the keywords, and inject the bias vector into the speech recognition decoder all at once through the hot word weighting interface of the speech recognition decoder. S27, by applying a decoding path cost discount proportional to the hot word bias coefficient to each keyword during the beam search process of the speech recognition decoder, the keywords in the dynamic hot word list are positively biased.

[0022] This embodiment discloses how to allocate benchmark weights, calculate hot word bias coefficients, and inject them into the decoder based on the field type labels of the filing information during the dynamic hot word injection process, so as to solve the problem of insufficient accuracy in identifying key filing information in live streaming scenarios.

[0023] It's important to note that in the compliance review process for live-streaming e-commerce, the content delivered by the streamer must be compared with the product registration information. Achieving this relies on accurately transcribing the speech into text. However, general speech recognition engines exhibit varying recognition capabilities when faced with different types of words in the registration information. Ingredient names in the registration information, such as "arbutin," "nicotinamide," and "ascorbate glucoside," are chemical terms rarely appearing in the training corpus of general language models. Their prior probability is far lower than that of common words with similar pronunciations. Efficacy descriptions, such as "moisturizing," "whitening," and "firming," as well as terms related to target audiences and warnings, such as "sensitive skin" and "use with sunscreen," are relatively common in the general corpus, and the language model's prior probability does not show a significant disadvantage. In the non-stationary noise environment of a live-streaming room, the resolution of acoustic features decreases, and the decoder's beam search process relies more heavily on the language model's prior information for pruning decisions. This leads to obscure ingredient names being easily misidentified as everyday words with similar pronunciations. Existing hot word biasing schemes typically assign uniform, fixed weights to all hot words, failing to differentiate and enhance the words based on their varying recognition difficulty. This limits the concentration of bias resources on the words that truly need protection, thus hindering the improvement of overall recognition accuracy.

[0024] To address the aforementioned issues, the technical solution in this embodiment first utilizes the semantic structure inherent in the registration information to assign hierarchical baseline weights to words of different field types. When a live-streamed product is listed, the system obtains the product's identifier and registration information. The registration information includes key registration fields with field type tags. Field type tags at least include ingredient categories, efficacy categories, and warning categories (applicable population categories can be merged into the warning category). The system has a built-in baseline weight mapping table, where the baseline weight corresponding to the ingredient category is the first weight, the baseline weight corresponding to the efficacy category is the second weight, and the baseline weight corresponding to the warning category is the third weight, with the first weight > the second weight > the third weight. This mapping table is set based on the following: ingredient names are at a significant disadvantage in language models due to their obscurity and require the highest bias strength to counteract prior bias; efficacy words are next; and warning and population words are next. The hierarchical weights allow bias resources to be reasonably allocated according to recognition difficulty, avoiding the inefficiency of treating all words equally.

[0025] The system reads the corresponding baseline weights from the mapping table based on the field type labels, and associates and stores the keywords extracted from each key field of the product registration with the baseline weight, forming a dynamic hot word table. Taking the registration information of a product "serum" as an example, the ingredient categories "arbutin" and "niacinamide" are associated with the first weight of 1.0; the efficacy categories "help fade spots" and "moisturize" are associated with the second weight of 0.8; and the warning category "must be used with sunscreen" and the target audience category "non-sensitive skin" are associated with the third weight of 0.6. The parameters of the baseline weight mapping table can be verified and determined through the development set to ensure that the weight ratios are basically consistent with the recognition difficulty distribution of various words in actual live broadcast noise samples. The development set contains registration information words for at least multiple product categories, covering high-frequency categories in live broadcast e-commerce such as beauty, health products, and food, and superimposes multiple sets of live broadcast noise audio samples with different signal-to-noise ratios.

[0026] Based on the aforementioned baseline weights, this invention further addresses the issue of insufficient accuracy in identifying key registration information in live streaming scenarios by using rarity factors and noise factors. Firstly, regarding the rarity of words, the system offline calculates the frequency of keywords appearing in all registered products within the current live streaming product category, and then calculates a rarity factor for each keyword. In one implementation, the rarity factor is calculated as follows: ;

[0027] In the formula, Represents the rarity factor. This indicates the total number of registered products under the current live-streaming product category. Indicates the presence of keywords The number of products. This rarity factor ranges from 0 to 1, with a higher value indicating that the term is rarer in that category (i.e., more difficult to identify). When a term has never appeared in the product registration of the same category... =0, at this time =1. Among them, offline statistics form a statistical database, which is updated offline regularly as new products are listed, without occupying real-time computing resources. This embodiment uses the natural logarithm (base e), but the logarithm of any base does not affect the normalization result because the numerator and denominator use the same base. In actual implementation, the common logarithm can also be used, and the two are equivalent.

[0028] During the live broadcast, the signal-to-noise ratio of the current audio segment is estimated in real time, and the noise factor is calculated based on the signal-to-noise ratio. In one implementation, the noise factor is calculated as follows: ;

[0029] In the formula, Indicates the noise factor. This represents the real-time estimated signal-to-noise ratio (SNR) of the current audio segment (in dB). In this embodiment, a noise spectrum estimation algorithm based on minimum tracking (such as the Martin algorithm) is used. The noise power spectrum is updated every frame, and the SNR of the current frame is calculated in real time. This algorithm is robust in non-stationary noise environments, and the computational cost is less than 0.1 milliseconds per frame, which meets the real-time requirements. This represents the preset reference signal-to-noise ratio (30dB in this embodiment). This represents a small constant (to prevent division by zero) that indicates a noise factor greater than 0.5 when the signal-to-noise ratio (SNR) of the current audio segment is lower than the reference SNR, and that the noise factor is closer to 1 the lower the SNR. This estimation is triggered at a preset time interval (e.g., every 3 seconds) or when the SNR change exceeds a threshold.

[0030] The rarity factor and noise factor of the same word are fused using harmonic averaging to obtain the bias enhancement index of the keyword at the current time. : ;

[0031] In this embodiment, the characteristic of the harmonic average is that: only when and When both are relatively large, Only when the bias is significantly increased will the fusion result be limited to the smaller value. This aligns perfectly with the actual needs of live streaming scenarios, where a significant increase in bias is only necessary when the vocabulary is extremely rare and the environment is very noisy. For common vocabulary or quiet environments, the increase is automatically suppressed to avoid incorrect adjustment. Based on the bias enhancement index, the baseline weight, and the preset bias constant base, the current hot word bias coefficient for each keyword is calculated using an exponential function. : ;

[0032] in, Key words The corresponding hot word bias coefficient, Key words The corresponding benchmark weights, This represents the base value of the bias constant (pre-set, its value is related to the language model weight coefficients of the specific ASR engine, and is set to 5.0 in this embodiment). Its calibration method is as follows: fix the baseline weights on the development set, and... The value is gradually increased from 1.0 to 10.0 in increments of 0.5. The recall and false positive rates for hot words are recorded for each value, and the value at the inflection point of the recall gain curve is selected as the default value. (Different ASR engines can be calibrated independently.) This represents the preset enhanced sensitivity coefficient (1.0 in this embodiment, and γ ranges from [0, 2], also set to 1.0 in this embodiment). This value was determined through simulation experiments: on multiple test sets with different combinations of signal-to-noise ratio and word rarity, using the harmonic mean of recognition precision and recall as a comprehensive index, a grid search was performed on γ between 0.5 and 2.0, and the value corresponding to the highest point of the comprehensive index, 1.0, was selected. The larger the value, the stronger the exponential function. The amplification effect is more significant when it is close to 1, making it suitable for scenarios with extremely harsh noise. The smaller the value, the smoother the bias enhancement, which is suitable for general live streaming environments. This invention uses this formula to treat the bias enhancement exponent as the exponent of an exponential function, achieving superlinear growth. For example, when... When it increases from 0.5 to 0.9, The increase from approximately 1.414 to approximately 1.866 indicates that the enhancement magnitude is non-linearly amplified. In extremely noisy environments with highly obscure vocabulary, a bias far exceeding that of linear relationships is required to effectively correct recognition errors.

[0033] Taking "arbutin" as an example, if it rarely appears in the registration of similar products ( =0.9), and the current signal-to-noise ratio is very low ( =0.8), then =2×0.9×0.8 / (0.9+0.8)≈0.847, Hot word bias coefficient =1.0×5.0×(2) 0.847 ≈9.0; if the environment is quiet ( =0.1), then ≈0.18, ≈5.65, the enhancement rate automatically decreases.

[0034] It should be noted that, due to the relatively slow change in environmental noise, the system recalculates all hot word bias coefficients at preset time intervals (e.g., every 3 seconds) or when the bias enhancement index changes beyond a threshold, and dynamically injects the updated bias vector through the hot word weighting interface of the speech recognition decoder. This interface supports runtime hot word weight adjustment, modifying only the weight parameters in memory, so the injection time is in the millisecond range and will not cause any perceptible delay to the real-time decoding stream.

[0035] The bias vector is composed of the hotword bias coefficients of all keywords, and is injected into the decoder all at once through the native hotword weighting interface of the speech recognition decoder. The hotword weighting interface is a standard function provided by current mainstream speech recognition engines. It receives a list of hotwords and their bias parameters, without recompiling the decoding graph, only updating the weight parameters on the edges of the decoding path. The injection process takes milliseconds and is completed within the speech interval between the product listing operation and the broadcaster's start of the broadcast, without introducing perceptible latency.

[0036] It should be noted that the above mechanism works in conjunction with the online positive feedback steps described in subsequent steps S81-S83: this step provides a basic bias based on category statistics and real-time environment, and the positive feedback steps make incremental fine adjustments based on the individual pronunciation characteristics of the broadcaster. The two work together to improve the overall recognition accuracy of key information in the filing.

[0037] After the bias vector is injected, when the audio stream from the live broadcast is sent to the decoder for frame-by-frame bundle search decoding, for each candidate path in the search space, if the hypothesis word at the end of the path matches a keyword in the dynamic hot word list, the decoder automatically queries the hot word bias coefficient corresponding to that word and applies a discount proportional to the bias coefficient to the cumulative cost of the path. The larger the bias coefficient, the more cost is deducted, and the higher the path ranks in the bundle pruning sorted by cost, the higher the probability of it being retained to the next time step. Through this mechanism, component names receive the strongest path protection due to the largest bias coefficient, effectively offsetting the prior suppression of rare words by the language model; efficacy words receive moderate protection; and warning words receive basic protection. Thus, under the interference of live broadcast noise, the decoder tends to output the correct words that match the registration information rather than common words with similar pronunciations, significantly improving the overall recognition accuracy of key registration information. At the same time, because the bias strength is matched with the recognition difficulty of the words, the false positive rate is controlled within a reasonable range.

[0038] It should be noted that product registration information naturally includes semantic distinctions such as ingredients and efficacy, which correspond precisely to the prior probability distribution of words in the language model. This scenario characteristic is transformed into the basis for bias resource allocation in this solution. The hierarchical mapping of benchmark weights, the decoupled calculation of benchmark weights and bias constants, the vectorized injection using the standard hot word interface, and the synergistic effect of cost discounting within the decoder together constitute a closed-loop speech recognition targeted enhancement scheme, achieving a solution to the problem of identifying key registration information.

[0039] In one embodiment of the present invention, the step of performing fuzzy matching between the identified text and the keywords, and calibrating the successfully matched identified text using the keywords to obtain calibrated text includes: S421, Obtain each identified word in the identified text, and calculate the edit distance similarity between each identified word and the keyword; S422, Obtain the pinyin syllable sequence of the identified words and the pinyin syllable sequence of the key words, and calculate the pinyin similarity between the two; S423, the edit distance similarity and the pinyin similarity are weighted and summed to obtain the joint matching score: ;

[0040] in, Indicates the joint matching score. Indicates edit distance similarity. Indicates the similarity of pinyin. Indicates the edit distance similarity weight. This represents the weight of pinyin similarity. ; S424, determine whether the joint matching score is greater than a preset replacement threshold. If it is greater, replace the identified words with the corresponding keywords. If it is not greater, retain the identified words.

[0041] In this embodiment, it is disclosed how to perform fuzzy matching calibration between the recognized text output by the decoder and the keywords extracted from the filing information when the cache misses, so as to correct the residual recognition errors caused by live broadcast noise or pronunciation confusion, and obtain the calibrated text.

[0042] It should be noted that in the compliance review process of live-stream e-commerce, although the speech recognition decoder has positively emphasized the filed key information through the injected hot word bias coefficient, under the actual live-stream environment, due to the combined effect of acoustic interference factors such as background music, reverberation, the host's dialect accent and speech rate changes, some homophonic or near-homophonic errors may still remain in the recognition text output by the decoder. These errors mainly occur in the confusion between rare component names and common words in the filed key information, for example, "arbutin" is misrecognized as "bear fruit dry". If such residual errors are directly sent to the subsequent semantic structure analysis step without correction, the extracted noun phrases will not be the standard words in the filed information, which makes the final consistency check unable to be performed correctly, resulting in missed detection or false detection. To solve this problem, this embodiment uses the pre-constructed filed keyword vocabulary as the calibration reference standard, and performs joint fuzzy matching based on two dimensions of glyph and pronunciation for each recognized word in the recognition text. The first step of the matching process is to obtain each recognized word in the recognition text, and calculate the edit distance similarity between the recognized word and the keyword vocabulary. Specifically, the Levenshtein edit distance between the recognized word and each keyword is calculated, that is, the minimum number of operations required to convert the recognized word into the keyword through inserting, deleting, and replacing characters. Since the string lengths of the keyword and the recognized word may be different, directly using the original integer value of the edit distance cannot achieve a fair comparison between words of different lengths. Therefore, the edit distance is divided by the length of the longer string between the recognized word and the currently compared keyword to obtain a normalized edit distance value, and the edit distance similarity is set to one minus the normalized value. Thus, the value range of edit distance similarity is normalized to between 0 and 1, and a larger value indicates that the two strings are more similar in character composition. For example, there is one replacement operation between "arbutin" and "bear fruit dry", the length of the longer string is 3, the normalized value is 1 / 3, and the edit distance similarity is 2 / 3. This indicator can effectively capture recognition errors caused by similar glyphs, and is particularly suitable for confusion arising from sharing the same morpheme in Chinese texts.

[0043] In the second step of matching, the pinyin syllable sequence of the recognized word and the pinyin syllable sequence of the keyword are obtained, and the pinyin similarity between the two is calculated. The recognized word and the keyword are respectively converted into pinyin syllable sequences that do not contain tone information, and pinyin syllables are divided into combinations of initials and finals. For example, "xiong guo gan (熊果苷, arbutin)" is converted into "x-iong g-uo g-an". For two syllable sequences, the Jaccard similarity coefficient is used to measure the degree of pronunciation overlap, that is, the size of the intersection of two syllable sets divided by the size of the union gives the pinyin similarity, which also ranges from 0 to 1. A larger value indicates that the pronunciations are closer. This indicator is specifically aimed at acoustic confusion caused by dialect accents or fast speech flow in live streaming scenarios. Such confusion may have great differences in glyphs but is highly similar in pinyin syllable composition. For example, affected by dialects, "xiong guo gan" is pronounced similar to "hong guo gan", and the pinyin sequence "h-ong g-uo g-an" still partially overlaps with "x-iong g-uo g-an", and the Jaccard coefficient can quantify this partial overlap. Pinyin similarity does not rely on tones, because tones are often weakened in live streaming speech, which avoids losing the matching ability due to inaccurate tones.

[0044] After obtaining the edit distance similarity and the pinyin similarity, the third step of matching is to perform weighted summation of the two to obtain a joint matching score. The weighting coefficients are the edit distance similarity weight and the pinyin similarity weight respectively, and the sum of the two is 1. The specific setting of weighting coefficients can be determined through development set experiments: on multiple sets of labeled test sets containing known errors, with the goal of maximizing the recognition accuracy of calibrated recorded key information, grid search is performed on weight combinations, and the weight pair with the highest accuracy is selected as the preset value. In a setting of this embodiment, the edit distance similarity weight is 0.6, and the pinyin similarity weight is 0.4, which gives a slightly higher weight to the glyph dimension, because in the Chinese live streaming compliance scenario, glyph similarity is often the main source of recognition errors, while pronunciation similarity serves as an effective supplement for correcting homophonic or near-homophonic errors.

[0045] In the last step, it is determined whether the joint matching score is greater than a preset replacement threshold. For each recognized word, the candidate with the highest joint matching score among all keywords is selected. If the highest score is greater than the preset replacement threshold, the recognized word is replaced with the candidate keyword; otherwise, the original recognized word is retained. The setting of the replacement threshold needs to balance the calibration recall rate and the false replacement rate: too low a threshold will cause irrelevant words to be forcibly replaced with keywords, introducing false positives; too high a threshold cannot correct real recognition errors. In actual deployment, the threshold is selected by parameter tuning on the development set. For example, for live streaming of the beauty category, the preset replacement threshold can be set to 0.75. When the joint matching score of the recognized word "xiong guo gan (熊果干)" and the standard word "xiong guo gan (熊果苷, arbutin)" exceeds 0.75, the system replaces it with "熊果苷 (arbutin)" to complete the calibration.

[0046] The above steps reduce errors at the source by lowering the path cost of keywords during the decoding stage, while fuzzy matching calibration acts as a post-defense line, repairing residual errors that penetrate the protection of hot word bias. The combined effect of these two methods further improves the lexical accuracy of key information in the final calibrated text that enters semantic structure parsing. This synergistic effect is particularly evident in handling dialect accent scenarios: systematic pronunciation shifts caused by certain dialects are difficult to completely offset by hot word bias during the decoding stage because hot word bias mainly affects path cost and cannot directly correct the acoustic model's score for dialect variants; however, these pronunciation shifts can be detected and matched in the pinyin similarity dimension, thus fuzzy matching calibration can compensate for the shortcomings of hot word injection in specific pronunciation scenarios, forming a complementary relationship. Ultimately, in the entire processing chain of consistency verification of filing information, the noun phrases output after double correction have higher accuracy, providing a reliable foundation for subsequent compliance deviation calculations.

[0047] In one embodiment of the present invention, the step of performing semantic structure parsing on the calibrated text according to a preset set of affirmative trigger words and a set of negative trigger words, and extracting and outputting noun phrases and their claim types includes: S51, perform word segmentation and part-of-speech tagging on the calibrated text; S52, sequentially scan the results of the word segmentation. When any word in the set of affirmative trigger words is hit, enter the affirmative state, extract the subsequent continuous noun components as the noun phrase to be determined, and output the noun phrase and the affirmative claim type. S53, when any word in the set of negation trigger words is hit, enter the negation state, extract the subsequent continuous noun components as the noun phrase to be determined, and output the noun phrase and the negation claim type.

[0048] This embodiment discloses how to perform semantic structure parsing on the calibrated text, extract the noun phrases actually claimed by the anchor and determine their claim type, so as to provide structured semantic input for subsequent consistency verification.

[0049] It is important to note that in the compliance review scenario of live-streaming e-commerce, the key to verifying the consistency of registration information lies in accurately determining whether the host makes a positive or negative claim about a certain ingredient or efficacy. For example, if a host says, "This serum contains niacinamide, but does not contain alcohol," the compliance review system needs to correctly extract the positive claim corresponding to "niacinamide" and the negative claim corresponding to "alcohol" in order to make a meaningful comparison with the registration information. If a negative structure is misjudged as a positive one, false claims that should be detected will be missed; if a positive structure is misjudged as a negative one, false alarms may occur. However, the sentence structure of live-streaming announcements is varied, and the ways of expressing affirmation and negation are rich. The same ingredient may immediately follow a trigger word, or it may be separated by modifiers, or even multiple positive or negative structures may appear in a single sentence. General text comparison or simple keyword matching cannot accurately capture such structured semantic relationships. Therefore, to solve the above problems, this embodiment constructs a lightweight semantic state machine based on a set of positive trigger words and a set of negative trigger words to perform sequential scanning and structural parsing of the calibrated text. The entire parsing process does not rely on deep neural network language models, but instead utilizes the linguistic feature that affirmative and negative expressions in live broadcasts have clear trigger words to achieve efficient structured extraction.

[0050] The first step in the analysis is to feed the calibrated text into a word segmenter for word segmentation and part-of-speech tagging. The word segmenter's function is to divide continuous text into independent lexical units and label each unit with a part-of-speech tag, such as noun, verb, adjective, or adverb. Word segmentation and part-of-speech tagging are performed using general-purpose Chinese natural language processing tools, and the output is a sequence of words and their corresponding part-of-speech tags. This step provides a standardized lexical flow and grammatical information for the subsequent state machine, enabling the boundary recognition of noun phrases to be based on part-of-speech tagging.

[0051] The second step in the analysis is to sequentially scan the word segmentation results and maintain a finite state automaton to recognize the semantic structure. The state automaton contains at least three states: neutral, affirmative, and negative. At the start of the scan, the state automaton is in a neutral state. During the sequential scan, when the current word matches any word in the set of affirmative trigger words, the state automaton transitions from the current state to the affirmative state; when the current word matches any word in the set of negative trigger words, the state automaton transitions to the negative state.

[0052] The set of affirmative trigger words includes at least "contains," "adds," "rich in," and "includes." In actual deployment, this set can be expanded to include common equivalent variations of the above words, such as "possess," "have," and "implies," as well as colloquial expressions related to live-streaming e-commerce. The set of negative trigger words includes at least "does not contain," "does not have," "not added," and "0 added," and can also be expanded to include variations such as "no added," "zero added," and "refuse to add." The selection principle for the trigger word set is to cover frequently occurring affirmative and negative expressions in live-streaming e-commerce scenarios, while maintaining scalability in the system configuration to allow for updates based on newly added expressions in actual live-streaming data.

[0053] Once the state machine enters an affirmative or negative state, it continues to sequentially scan subsequent words. For consecutive sequences of words tagged as nouns or belonging to noun phrases, the state machine collects them as noun phrases to be determined, stopping collection when a non-noun phrase or sentence boundary is encountered. The collected noun phrases to be determined form a semantic association with the affirmative or negative trigger word that triggered the state transition: if the current state is affirmative, the noun phrase and its affirmative claim type are output; if the current state is negative, the noun phrase and its negative claim type are output. When multiple trigger words appear in a sentence, the state machine will overwrite the previous state with the next trigger word to ensure that each noun phrase corresponds to its nearest trigger word.

[0054] Taking the calibrated text "This serum contains niacinamide and arbutin, but does not contain alcohol" as an example, after word segmentation and part-of-speech tagging, the system scans sequentially: "This" is a non-trigger word, and its state remains neutral; "serum" is a noun, but its current state is neutral, so it is not extracted yet; "contains" hits the set of affirmative trigger words, and its state transitions to affirmative; "niacinamide" is a noun, and it is collected in an affirmative state. Then, it encounters "and," a connecting word, and continues to collect the next noun "arbutin." The noun sequence terminates after encountering a comma, and the output "niacinamide" and "arbutin" are both affirmative claim types; "but" is a non-trigger word, and its state remains affirmative; "does not contain" hits the set of negative trigger words, and its state transitions to negative; "alcohol" is a noun, and it is collected in a negative state. It terminates at the end of the sentence, and the output "alcohol" is a negative claim type.

[0055] This state machine parsing method does not limit the noun phrase to be judged to belong to the dynamic hot word list or the key fields of the filing. Instead, it extracts the noun components following the trigger word as completely as possible based on part-of-speech tagging. This design is based on the consideration that during a live stream, the host may claim ingredients or effects that do not exist in the filing information, which is a typical case of false advertising beyond the scope of the filing. If extraction is only performed within the hot word list, such illegal claims will be missed and cannot be detected in subsequent consistency verification steps. The open-ended noun phrase extraction ensures complete recall coverage in semantic parsing. False or excessive advertising is uniformly handled by the compliance deviation calculation logic in subsequent comparison steps, achieving a functional separation and clear responsibility architecture design.

[0056] In terms of performance, the state machine's processing logic is a single-pass linear scan, with time complexity linearly related to the number of words in the calibrated text, and a single sentence processing time of less than 5 milliseconds. The computational overhead of word segmentation and part-of-speech tagging can be optimized by using a lightweight word segmenter and a preloaded dictionary, and does not constitute a bottleneck for real-time processing overall. The state machine's rule logic is clear and interpretable. In the event of misparsing or missing parsing, it can be quickly corrected by supplementing the trigger word set or adjusting the word segmentation dictionary without retraining the model, thus meeting the maintainability and auditability requirements of compliance review systems.

[0057] After this step, the calibrated text is transformed into a series of structured semantic units, each consisting of a noun phrase and a claim type. These semantic units will serve as input for subsequent consistency verification steps, comparing them with key fields in the filing information to generate verification results. It should be noted that the noun phrases output in this step are still in calibrated text form, and their consistency with the standard vocabulary in the filing information has been ensured by the fuzzy matching calibration in the previous step. The accurate labeling of claim types is completed by the state machine semantic parsing in this step. The combination of these two aspects allows subsequent compliance deviation calculations to rely on both correct lexical form and accurate semantic polarity, thereby generating reliable consistency verification results.

[0058] In one embodiment of the present invention, the step of storing the noun phrase and its claim type in the cache using the feature hash value as the key includes: S61, extract the Mel frequency cepstral coefficient feature frame sequence of the audio segment, and calculate the mean vector of all frames as the acoustic feature representation; S62, the acoustic feature representation is encoded using a hash algorithm to obtain a hash value of a fixed number of bits, and the hash value is concatenated with the identifier of the current product as a cache key; S63, using the cache key as an index, store the noun phrase and its claim type association in a cache that adopts the least recently used eviction policy, and set an effective lifetime.

[0059] This embodiment discloses how to construct a cache key based on the acoustic features of an audio segment after semantic structure parsing is completed, and store the parsed noun phrases and claim types into the cache to achieve rapid retrieval of repetitive or highly similar audio segments.

[0060] It's important to note that in live-stream e-commerce compliance review scenarios, when introducing a product, the host often uses similar or identical wording to repeatedly emphasize the product's core selling points, ingredient efficacy, or deny the existence of certain ingredients. For example, in a live stream introducing a "serum," the host might repeatedly say the same or highly similar statements such as "Our serum contains niacinamide" or "It's alcohol-free." If the system independently performs a complete speech recognition, fuzzy matching calibration, and semantic structure parsing process for each live stream audio, it will result in a large amount of redundant computation, wasting computing resources. In compliance review platforms that typically need to process multiple live stream audio streams in parallel, this redundant computation will directly limit the platform's scalability and real-time processing capabilities. To address this issue, this embodiment employs an incremental caching mechanism based on acoustic feature hashing. The acoustic feature representation of the audio segment is used as the query index, and the semantic parsing results that have already undergone full-process processing are associated and stored. When subsequent audio segments with highly similar acoustic features appear, the system can directly retrieve the stored noun phrases and claim types from the cache, bypassing fuzzy matching and semantic structure parsing, thus achieving computational reuse.

[0061] The first step in cache writing is to extract the Mel-frequency cepstral coefficient (MFCC) feature frame sequence of the audio segment. MFCC is a widely used acoustic feature representation in speech recognition, and its extraction process simulates the human ear's perception of sound signals at different frequencies. First, the audio segment is divided into frames and windowed, with each frame typically lasting 25 milliseconds and a frame shift of 10 milliseconds. The power spectrum of each frame is calculated, and the power spectrum is passed through a set of Mel-scale triangular filters. After taking the logarithm, a discrete cosine transform is performed to obtain the MFCC feature vector for each frame, typically with 13 dimensions or higher. For an audio segment consisting of multiple frames, the mean vector of the MFCC feature vectors from all frames is calculated as the acoustic feature representation of that audio segment. The reason for using the mean vector instead of retaining the entire frame sequence is that the mean vector maps variable-length audio segments uniformly to vectors of fixed dimensions, while retaining the overall spectral envelope features of the audio segment, which can effectively distinguish audio segments with different pronunciation content; for the same anchor repeating the same sentence at a similar speed and tone, their mean vectors are in close regions in the feature space, so that subsequent hash encoding can classify them into the same or similar hash buckets.

[0062] The second step in cache writing is to encode the mean vector using a hash algorithm to obtain a fixed-bit hash value. In this embodiment, the Locality Sensitive Hashing (LSH) algorithm is used for this encoding. The core property of LSH is that two vectors that are close in distance in the original space will generate the same hash value or hash values ​​with only a few bits different after hash encoding; vectors that are far apart in the original space will generate significantly different hash values. This property makes LSH naturally suitable for near-repeating audio detection in this scenario. In one specific implementation, LSH generates multiple sets of random projection vectors, performs an inner product between the original high-dimensional vector and each set of projection vectors, and determines whether the hash bit is 0 or 1 based on the sign of the inner product result, thus generating a fixed-bit binary hash value. The hash value is typically 32 bits, striking a balance between collision probability and storage efficiency. Even if the audio segments of different products have the same content, their acoustic characteristics will naturally shift due to differences in broadcasters or recording environments. Therefore, concatenating the product identifier with the hash value as a cache key is a necessary isolation measure to ensure that the cache spaces of different products do not overlap.

[0063] The third step in cache writing is to concatenate the hash value with the current product identifier as the cache key. Using this cache key as an index, the noun phrases and claim types output from the semantic structure parsing step are associated and stored in the cache. The product identifier is obtained along with the product listing operation in step S1 and is a unique key that identifies the product corresponding to the current verification process. The necessity of including the product identifier in the cache key is that different products may have the exact same registered ingredient name (e.g., multiple products contain "nicotinamide"), and the anchor may say the same phrase fragment (e.g., "contains nicotinamide") in live broadcasts of different products, but the consistency verification results of the two products may differ due to different registration information. If the cache key is based solely on acoustic feature hashing without distinguishing between products, cross-product cache cross-hit may occur, resulting in the output of incorrect noun phrases and claim types. By concatenating the product identifier into the cache key, the cache records of different products are naturally isolated. Even if the acoustic feature hashes collide, as long as the product identifiers are different, the cache keys will be different, preventing incorrect hits.

[0064] The cache employs a Least Recently Used (LRU) eviction policy to manage storage space, with a preset maximum cache capacity of 5000 records. When the number of records stored in the cache reaches the limit, the system automatically evicts the least recently accessed record, freeing up space for new records. Each cached record has a valid lifespan, set to 5 minutes in this embodiment. This lifespan is set because the audio environment, the broadcaster's status, and the product description content can all dynamically change during a live stream. Outdated cached records may no longer match current audio characteristics due to acoustic drift, and retaining them for too long could introduce false positives. The 5-minute validity period covers the main time frame of a complete product introduction by the broadcaster, effectively blocking most duplicate content within the validity period.

[0065] After completing the cache writing steps described above, when the system processes subsequent input audio segments, it follows the query process corresponding to the writing process. First, it extracts the MFCC mean vector of the audio segment, generates a hash value using the same Locality Sensitive Hashing algorithm, and concatenates it with the current product identifier to form a query key. Then, it searches the cache for a matching cached record whose valid lifetime has not expired. If a match is found, the noun phrase and claim type stored in the cache are directly output, skipping the fuzzy matching and semantic structure parsing steps, thus avoiding redundant calculations of identical or highly similar content. In typical live streaming scenarios, because the host makes many repetitive statements when introducing core selling points, a higher cache hit rate means a significant proportion of computing resources can be saved.

[0066] This embodiment reduces recognition errors at the source through hot word injection, corrects residual errors through fuzzy calibration, and provides structured semantic output through semantic parsing. Caching reuses the high-quality parsing results, and as an optional optimization path, caching does not affect processing accuracy when a match is missed, while directly contributing to reduced computational load when a match is hit. In compliance review platforms that need to process dozens or even hundreds of live audio streams in parallel, this caching mechanism effectively reduces the system's average single-channel processing latency and CPU usage, improving the platform's overall throughput.

[0067] In one embodiment of the present invention, the step of comparing the noun phrase and claim type with the corresponding filing key fields to generate a consistency verification result includes: S71, construct a filing knowledge graph based on the current product filing information, extract the filing key field nodes that should be claimed to be filed to form a set of affirmative filing nodes, and associate each node with a weight. The weight is taken as the hot word bias coefficient of the keyword corresponding to the node in the dynamic hot word table. If it is not in the table, the preset default weight is taken. S72, construct an affirmative claim set based on noun phrases of the claim type being affirmative, and construct a negative claim set based on noun phrases of the claim type being negative; S73, if there are noun phrases in the set of affirmative claims that do not belong to the set of affirmative filing nodes, then generate false advertising warning information; S74, calculate the sum of the weights of the intersection of the set of negative claims and the set of positive filing nodes, as the first weighted sum; S75, calculate the sum of the weights of the nodes in the set of affirmative filing nodes that do not appear in the set of affirmative claims or the set of negative claims, and combine them with a preset omission risk coefficient to obtain a second weighted sum; S76, Based on the first addend, the second addend, and the sum of the weights of all nodes in the set of nodes with confirmed filings, the compliance deviation is obtained, using the following formula: ;

[0068] In the formula, Indicates the degree of compliance deviation. This indicates a set of nodes that are definitely registered. To indicate a definitive declaration of a collection, Indicates the set of negative claims. Represents a node The weight, This indicates the risk factor for omissions in claims (preset value, 0 < 0). ≤1 (default value is 1). This indicates a denial of the part of the claim that involves affirmative filing details. This indicates a confirmed filing node that did not appear in any claim; S77, When the compliance deviation exceeds the preset warning threshold, a publicity compliance warning message is generated.

[0069] This embodiment specifically illustrates how to compare the noun phrases and claim types obtained from semantic structure parsing with the key fields of the filing information, construct a filing knowledge graph, classify and detect various types of violations and calculate compliance deviation, and finally generate a consistency verification result.

[0070] It's important to note that in the compliance review of live-streaming e-commerce, the ultimate goal of consistency verification is to compare the content actually claimed by the streamer with the information registered and filed with the regulatory agency to determine whether the promotional content is false or illegal. Existing compliance review processes, with their simple keyword matching or binary judgment (match or non-match), cannot meet the refined needs of actual compliance reviews. The filing information may require a product to claim multiple ingredients and effects, while the streamer may only mention some of them in their broadcast, and may actively deny the existence of certain filed ingredients or claim additional ingredients that are not listed in the filing. These promotional behaviors have different natures and legal consequences in terms of compliance. Denying an ingredient that is indeed present in the filing (e.g., the filing includes "arbutin," but the streamer claims "no arbutin") constitutes a false statement about the product's ingredients; claiming an effect that is not present in the filing (e.g., the filing does not mention "anti-aging," but the streamer claims "anti-aging") constitutes exaggerated advertising beyond the scope of the filing; and omitting a required ingredient or effect from the filing (e.g., the filing requires "containing niacinamide," but the streamer does not mention it throughout the broadcast) is also non-compliant advertising, and its severity needs to be included in a comprehensive assessment. Existing consistency verification schemes typically only target one or two of the aforementioned violation types, lacking a unified quantitative framework to comprehensively measure these different types of violations. Therefore, this embodiment constructs a unified comparison and quantification framework, which structures the registration information into a queryable knowledge graph, maps various claims made by the broadcaster to matching states with nodes in the graph, and comprehensively quantifies the compliance risks of live broadcast promotional content through a configurable deviation index.

[0071] Specifically, this embodiment first constructs a registration knowledge graph based on the current product's registration information. The system reads the registration information corresponding to the product, extracts all the ingredient names and efficacy descriptions that should be claimed, and uses these ingredients and efficacy as positive registration nodes. Each positive registration node represents an item of information that should be included in the product's advertising according to the registration requirements. All positive registration nodes constitute a set of positive registration nodes. For example, the set of positive registration nodes for the product "serum" includes nodes such as "arbutin," "niacinamide," "help fade spots," "moisturizing," "non-sensitive skin," and "requires use with sunscreen." Each positive registration node is associated with a node weight, which is set according to the following rules: if the keyword corresponding to the node exists in the dynamic hot word table, the hot word bias coefficient corresponding to the keyword in the dynamic hot word table is used as its node weight; if the keyword corresponding to the node is not included in the dynamic hot word table (for example, because the registration information adds a field that was not extracted in the hot word construction stage), the preset default weight is used. The preset default weight is a system configuration value, which is 0.5 in this embodiment. This weighting rule ensures that components that are difficult to identify (and therefore have a higher hot word bias coefficient) also receive a higher structural weight in the filing graph during the compliance deviation calculation. Their contribution to the final deviation is correspondingly greater, ensuring that violations of key components are assigned a higher risk score.

[0072] The node weights of the filing knowledge graph inherit the hot word bias coefficients from the dynamic hot word list. This design is not a random reuse of numerical values, but rather based on the inherent logic of live-streaming compliance review scenarios: the more obscure and difficult-to-identify a component is, the greater its commercial promotional value tends to be, and the higher the compliance risk of the streamer making false or omitted claims about it. Therefore, it should be given a greater weight for deviation measurement. This design ensures that the weight distribution of the filing knowledge graph matches the actual risk structure of product promotion, providing a scenario-adapted risk measurement basis for subsequent deviation calculations.

[0073] Secondly, based on the noun phrases and their claim types output from the semantic structure parsing step, the system constructs two claim sets. Noun phrases with affirmative claim types are collected to form an affirmative claim set; noun phrases with negative claim types are collected to form a negative claim set. These two sets are a structured summary of all parsed claims made by the broadcaster during the current live broadcast period, serving as input for comparison with the registered knowledge graph.

[0074] After obtaining the set of positive filing nodes, the set of positive claims, and the set of negative claims, the system performs three types of violation detection, corresponding to three types of violations: advertising beyond the scope, denying the inclusion of required ingredients, and omitting claims.

[0075] The first type of detection is for overreaching claims. If a noun phrase exists in the set of affirmative claims that does not belong to the set of affirmative filing nodes, it means that the broadcaster has made a positive claim that does not exist in the filing information, and this behavior constitutes false advertising. In this case, the system directly generates a false advertising warning message, which includes the detected overreaching noun phrase, the corresponding timestamp, and the cache key or identifier of the audio segment, for subsequent manual review or automatic recording. Overreaching claims detection is independent of the subsequent compliance deviation calculation because overreaching claims are a clear type of violation that can be identified without weighted comparison with other filing nodes.

[0076] The second and third categories of detection are uniformly quantified through compliance deviation calculation. The second category of detection targets the intersection of the set of negative claims and the set of positive registration nodes: when a livestreamer denies a certain ingredient or efficacy in the set of positive registration nodes—that is, when the livestreamer claims that the product "does not contain" a certain ingredient required for registration—this term simultaneously belongs to both the set of negative claims and the set of positive registration nodes, constituting a violation in the negative claim involving a positive registration node. The system calculates the sum of the node weights of all nodes in this intersection as the first weighted sum. Taking the product "essence" as an example, if the livestreamer claims "does not contain arbutin," and "arbutin" belongs to a positive registration node with a weight of 5.0, then the first weighted sum is 5.0.

[0077] The third type of detection targets omitted claims: Nodes in the affirmative filing node set that are neither covered by the affirmative claim set nor the negative claim set indicate that the required filing content was not mentioned by the broadcaster throughout the entire broadcast. The system calculates the sum of the weights of these omitted nodes and multiplies this sum by a preset omission risk coefficient to obtain a second weighted sum. The omission risk coefficient is a preset parameter used to adjust the contribution of omitted claims to the overall compliance deviation. In this embodiment, A value of 1.0 indicates that omitted claims and denials of required elements are treated equally; in some scenarios with a higher tolerance for omitted claims, This can be set to a smaller value, such as 0.5, to reduce the contribution of omitted claims to the deviation score. Taking the product "serum" as an example, if the weight of the "moisturizing" claim in the registration node is 4.0, and the host neither affirms nor denies "moisturizing," then the weight of the omitted claim is 4.0. When the sum is 1.0, the second weighted sum is 4.0.

[0078] After obtaining the first and second weighted sums, the system calculates the compliance deviation. The compliance deviation equals the sum of the first and second weighted sums divided by the total weight of all nodes in the set of nodes required for filing. The physical meaning of this calculation is: the proportion of weights that are not covered by compliance due to the streamer's denial or omission of certain claims within the total weight of all content that should be claimed in the filing requirements. The compliance deviation value ranges from 0 to 1; a higher value indicates a more severe deviation between the streamer's promotional content and the filing requirements. This formula quantifies two different types of violations—denial of required content and omission of claims—into a single scalar indicator, making the compliance risks of different live streams, sessions, and products comparable, and providing a quantitative basis for the regulatory platform's risk classification and early warning priority ranking.

[0079] When the compliance deviation exceeds a preset warning threshold, the system generates a compliance warning message for the current product's advertising. The preset warning threshold is a configurable parameter; in this embodiment, it is set to 0.3, meaning a warning is triggered when approximately one-third of the filing nodes with the specified weight are rejected or omitted. This threshold can be flexibly adjusted according to the specific requirements of regulatory agencies and the risk levels of different product categories. The generated warning message includes at least the specific numerical value of the compliance deviation, a list of the violation nodes that led to the deviation exceeding the limit, and the corresponding timestamps, allowing compliance review personnel to quickly locate and verify the violation.

[0080] In one embodiment of the present invention, an online adaptive positive feedback step is further included: S81, obtain the recognition confidence of each recognized word in the recognized text; S82, when the recognition confidence is greater than the preset high confidence threshold and the noun phrase is in the dynamic hot word table, the hot word bias coefficient corresponding to the noun phrase is updated by increasing the preset step size, and the increased bias coefficient does not exceed the preset maximum bias limit. S83, the updated hot word bias coefficient is applied to the subsequent decoding process of the audio stream.

[0081] This embodiment can use high-confidence recognition results to incrementally fine-tune the hot word bias coefficient of corresponding words in the dynamic hot word list, so that the decoder can gradually adapt to the pronunciation characteristics of the current anchor during continuous live broadcast, and further improve the recognition accuracy of subsequent audio streams.

[0082] It should be noted that in real-world scenarios, there are significant individual pronunciation differences among different broadcasters. These differences lead to a systematic variation in the recognition confidence of the ASR decoder when the same registered keywords are spoken by different broadcasters. Some broadcasters, due to their heavy regional accents, exhibit a fixed deviation between their acoustic features and standard pronunciation when pronouncing specific component names. This results in the decoder outputting a relatively low recognition confidence for that word, even with the assistance of hot word bias. Therefore, this embodiment introduces an online adaptive positive feedback mechanism. This mechanism obtains the recognition confidence of each recognition word in the output recognition text. When the recognition confidence of a certain noun phrase is higher than a preset high confidence threshold, and the noun phrase belongs to the component or efficacy words in the dynamic hot word list, a positive fine-tuning of its hot word bias coefficient is triggered. For a noun phrase composed of multiple recognition words, its confidence is taken as the minimum value of the confidence of each recognition word in the constituent words. Because the minimum value represents the quality level of the least reliable part of the recognition in the phrase, taking a conservative estimate can reduce the risk of triggering misadjustment due to the inflated confidence of some words. The matching method between the noun phrase and the keywords in the dynamic hot word list is exact string matching, that is, the noun phrase as a whole is compared with the keywords in the dynamic hot word list. If there are completely identical keywords, it is determined to belong to the dynamic hot word list.

[0083] Setting a preset high-confidence threshold is the first safety barrier of this positive feedback mechanism. The high-confidence threshold is used to filter out recognition samples that the decoder itself is fully confident in regarding its output; only these samples are allowed for positive adjustment of the bias coefficient. If the threshold is set too low, low-confidence misidentification results may also trigger bias enhancement, leading to an inappropriate increase in the bias coefficient of the incorrect word, creating a vicious cycle. The decoder will then be more inclined to output the misidentified word, and the false positive rate will rise accordingly. If the threshold is set too high, the number of positive feedback samples will be insufficient, resulting in a weak adaptive effect. In this embodiment, the preset high-confidence threshold is set to 0.9. This value was determined through validation on the development set: various dialect accents and noise combinations were simulated on the development set, aiming to achieve the optimal comprehensive recognition accuracy of key information after positive feedback. Thresholds in the range of 0.8 to 0.95 were tested traversally, and the value corresponding to the highest accuracy point was selected.

[0084] Another constraint on the triggering condition is that the noun phrase must belong to the ingredient or efficacy terms in the dynamic hot word list. Limiting the scope to ingredient and efficacy terms, excluding warning and target audience terms, is based on considerations of actual identification difficulty and the importance of compliance review. Ingredient and efficacy terms are the core objects of consistency verification of filing information, and their identification accuracy has the greatest impact on compliance judgment; at the same time, these two types of terms contain the most obscure professional terms, making them the most difficult to identify and benefiting the most from positive feedback. Warning and target audience terms are relatively common and easier to identify, and their bias coefficients are already small; further fine-tuning them has limited marginal benefits, and excluding them from the positive feedback scope reduces unnecessary bias update operations.

[0085] When both of the above conditions are met, the system increases the hot word bias coefficient corresponding to the noun phrase. The increase is achieved by multiplying the original bias coefficient by a coefficient (1+δ), where δ is a preset step size factor, δ∈(0,0.1]. Its advantages are: the update magnitude is proportional to the current bias coefficient value; words with larger bias coefficients have larger absolute increments in a single update, matching the higher bias requirements of these words; simultaneously, the update formula is simple, and the computational cost is negligible. In this embodiment, the preset step size factor is 0.05, meaning the bias coefficient increases by 5% each time positive feedback is triggered. This step size strikes a balance between adjusting the response speed and system stability. An excessively large step size will cause the bias coefficient to rapidly climb to its upper limit after only a few triggers, making the system overly sensitive to occasional high-confidence samples; an excessively small step size will result in weak adaptive effects, making it difficult to produce observable improvements within the timescale of a single live broadcast. This value was determined through online experiments simulating differences in accents among different broadcasters, ensuring that the bias coefficients of common component words can stably converge to near their upper limit within a single live broadcast (approximately 30 minutes).

[0086] To prevent the bias coefficient from growing indefinitely during continuous positive feedback, the system sets a preset maximum bias upper limit as a constraint. When the calculated new bias coefficient exceeds this upper limit, the upper limit value is taken as the actual bias coefficient. In this embodiment, the preset maximum bias upper limit is set to twice the initial bias coefficient. When the bias coefficient exceeds a certain multiple of the initial value, the decoder's bias strength for that word becomes too large. Even if the acoustic model scores that word significantly lower, the decoder may still force the output of that word due to excessive path cost discounting, leading to a significant increase in the false positive rate. Therefore, setting the upper limit to twice the initial value can control the growth of the bias coefficient within a relatively safe range, maintaining a balance between improving the recognition rate and suppressing false positives.

[0087] The updated hot word bias coefficients take effect when subsequent audio frames reach the decoder. This is achieved by re-injecting the updated bias coefficients through the decoder's hot word weighting interface after the initial update, overwriting the original bias parameters for that word. Since updating the bias coefficient of a single word involves modifying only one record, the injection overhead is extremely low and does not affect the real-time performance of the decoding pipeline. The next time the broadcaster utters the same or similar-sounding word, the decoder, due to the increased bias coefficient of that word, experiences a greater path cost discount, ranks higher in the beam search, and thus improves the probability of correct recognition.

[0088] like Figure 2 As shown, this application also provides an intelligent verification system for the consistency between live-stream product promotion and registration information, including: The product listing trigger module is used to respond to the listing operation of live-streamed products and obtain the product's registration information, which includes key registration fields with field type tags. The dynamic hot word table construction module is used to extract keywords based on the filing key fields, allocate benchmark weights according to the field type labels, construct a dynamic hot word table, obtain hot word bias coefficients based on the dynamic hot word table, and inject the hot word bias coefficients into the speech recognition decoder so that the speech recognition decoder positively biases the keywords in the dynamic hot word table during recognition. The audio acquisition and decoding module is used to acquire the audio stream of the live broadcast room, send the audio stream to the speech recognition decoder for real-time decoding, and output the recognized text and the corresponding audio segment. The matching calibration module is used to extract acoustic feature representations based on the audio segment, generate feature hash values, and query whether there is a cache record in the preset cache that matches the feature hash value. If it exists, the stored noun phrases and claim types are directly output based on the cache record. If it does not exist, the identified text is fuzzily matched with the keywords, and the identified text that successfully matches the keywords is calibrated to obtain the calibrated text. The semantic structure parsing module is used to perform semantic structure parsing on the calibrated text based on a preset set of affirmative trigger words and a set of negative trigger words, and to extract and output noun phrases and their claim types; The cache management module is used to store the noun phrases and their claim types into the cache using the feature hash value as the key; The consistency verification module is used to compare the noun phrases and their claim types with the corresponding filing key fields and generate consistency verification results.

[0089] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of an intelligent verification method for the consistency between live-streamed product promotion and registration information.

[0090] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps of an intelligent verification method for the consistency between live-streamed product promotion and registration information.

[0091] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0092] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for intelligently verifying the consistency between live-stream product promotion and registration information, characterized in that, include: In response to the listing of live-streamed products, the system obtains the product's registration information, which includes key registration fields with field type tags. Key words are extracted based on the key fields of the filing, and a benchmark weight is assigned according to the field type label to construct a dynamic hot word table. Hot word bias coefficients are obtained based on the dynamic hot word table, and the hot word bias coefficients are injected into the speech recognition decoder so that the speech recognition decoder positively biases the key words in the dynamic hot word table during recognition. The audio stream from the live broadcast room is collected and sent to the speech recognition decoder for real-time decoding, outputting the recognized text and the corresponding audio segment. Acoustic feature representations are extracted from the audio segment and feature hash values ​​are generated. A cache record matching the feature hash value is queried in the preset cache. If it exists, the stored noun phrases and claim types are directly output according to the cache record. If it does not exist, the identified text is fuzzily matched with the keywords, and the identified text that successfully matches is calibrated using the keywords to obtain the calibrated text. Based on a preset set of affirmative and negative trigger words, the calibrated text is semantically structured and parsed to extract and output noun phrases and their claim types. Specifically, this includes: segmenting and tagging the calibrated text into words; sequentially scanning the segmentation results; when any word in the set of affirmative trigger words is hit, the text enters an affirmative state, and the subsequent continuous noun phrases are extracted as noun phrases to be determined, and the noun phrases and their affirmative claim types are output; when any word in the set of negative trigger words is hit, the text enters a negative state, and the subsequent continuous noun phrases are extracted as noun phrases to be determined, and the noun phrases and their negative claim types are output. The noun phrase and its claim type are stored in the cache using the feature hash value as the key; The noun phrases and their claim types are compared with the corresponding key fields of the filing to generate a consistency verification result.

2. The intelligent verification method for consistency between live-stream product promotion and registration information according to claim 1, characterized in that, The steps of extracting keywords based on the key fields of the filing, assigning benchmark weights according to the field type labels, constructing a dynamic hot word table, obtaining hot word bias coefficients based on the dynamic hot word table, and injecting the hot word bias coefficients into the speech recognition decoder so that the speech recognition decoder positively biases the keywords in the dynamic hot word table during recognition include: The baseline weights corresponding to the keywords are determined based on the field type labels, and the keywords and their corresponding baseline weights are associated and stored to construct a dynamic hot word table. Obtain the product category to which the current live-streamed product belongs, and offline count the frequency of the keywords of all registered products under that product category to calculate the rarity factor of each keyword; During the live broadcast, the signal-to-noise ratio of the current audio segment is estimated in real time, and the noise factor is calculated based on the signal-to-noise ratio. The rarity factor and noise factor of the same word are fused by harmonic averaging to obtain the bias enhancement index of the keyword at the current time. Based on the bias enhancement index, the baseline weight, and the preset bias constant base, the current hot word bias coefficient of each keyword is calculated through an exponential function; The hot word bias coefficients of all the keywords are used to form a bias vector, and the bias vector is injected into the speech recognition decoder all at once through the hot word weighting interface of the speech recognition decoder. The speech recognition decoder applies a decoding path cost discount proportional to the hot word bias coefficient to each keyword during the beam search process, thereby positively biasing the keywords in the dynamic hot word list.

3. The intelligent verification method for consistency between live-stream product promotion and registration information according to claim 1, characterized in that, The step of performing fuzzy matching between the identified text and the keywords, and calibrating the successfully matched identified text using the keywords to obtain the calibrated text includes: Obtain each identified word in the identified text, and calculate the edit distance similarity between each identified word and the key words; Obtain the pinyin syllable sequence of the identified words and the pinyin syllable sequence of the key words, and calculate the pinyin similarity between the two; The edit distance similarity and the pinyin similarity are weighted and summed to obtain the joint matching score; Determine whether the joint matching score is greater than a preset replacement threshold. If it is, replace the identified words with the corresponding keywords. If it is not, retain the identified words.

4. The intelligent verification method for consistency between live-stream product promotion and registration information according to claim 1, characterized in that, The step of storing the noun phrase and its claim type into the cache using the feature hash value as the key includes: Extract the Mel frequency cepstral coefficient feature frame sequence of the audio segment, and calculate the mean vector of all frames as the acoustic feature representation; The acoustic feature representation is encoded using a hash algorithm to obtain a hash value of a fixed number of bits, and the hash value is concatenated with the identifier of the current product as a cache key; Using the cache key as an index, the noun phrase and its claim type are associated and stored in a cache that employs a least recently used eviction policy, and a valid lifetime is set.

5. The intelligent verification method for consistency between live-stream product promotion and registration information according to claim 1, characterized in that, The step of comparing the noun phrase and its claim type with the corresponding filing key fields to generate a consistency verification result includes: Based on the current product registration information, a registration knowledge graph is constructed, and the key registration field nodes that should be included in the product promotion are extracted to form a set of confirmed registration nodes. Each node is associated with a weight, which is the hot word bias coefficient of the keyword corresponding to the node in the dynamic hot word table. If it is not in the table, a preset default weight is used. Construct a set of affirmative claims based on noun phrases of the claim type that are affirmative, and construct a set of negative claims based on noun phrases of the claim type that are negative; If there are noun phrases in the set of affirmative claims that do not belong to the set of affirmative filing nodes, then a false advertising warning message is generated; The sum of the weights of the intersection of the set of negative claims and the set of positive filing nodes is calculated as the first weighted sum; Calculate the sum of the weights of nodes in the set of affirmative filing nodes that do not appear in the set of affirmative claims or the set of negative claims, and combine this with a preset omission risk coefficient to obtain a second weighted sum; The compliance deviation is obtained by combining the first weighted sum, the second weighted sum, and the total weight of all nodes in the set of nodes with positive filing records. When the compliance deviation exceeds a preset warning threshold, a compliance warning message is generated.

6. The intelligent verification method for consistency between live-stream product promotion and registration information according to claim 1, characterized in that, It also includes an online adaptive positive feedback step: Obtain the recognition confidence score of each recognized word in the recognized text; When the recognition confidence is greater than the preset high confidence threshold and the noun phrase is in the dynamic hot word table, the hot word bias coefficient corresponding to the noun phrase is updated by increasing the preset step size, and the increased bias coefficient does not exceed the preset maximum bias limit. The updated hot word bias coefficients are used in the decoding process of the audio stream.

7. A smart verification system for the consistency between live-stream product promotion and registration information, characterized in that, include: The product listing trigger module is used to respond to the listing operation of live-streamed products and obtain the product's registration information, which includes key registration fields with field type tags. The dynamic hot word table construction module is used to extract keywords based on the filing key fields, allocate benchmark weights according to the field type labels, construct a dynamic hot word table, obtain hot word bias coefficients based on the dynamic hot word table, and inject the hot word bias coefficients into the speech recognition decoder so that the speech recognition decoder positively biases the keywords in the dynamic hot word table during recognition. The audio acquisition and decoding module is used to acquire the audio stream of the live broadcast room, send the audio stream to the speech recognition decoder for real-time decoding, and output the recognized text and the corresponding audio segment. The matching calibration module is used to extract acoustic feature representations based on the audio segment, generate feature hash values, and query whether there is a cache record in the preset cache that matches the feature hash value. If it exists, the stored noun phrases and claim types are directly output based on the cache record. If it does not exist, the identified text is fuzzily matched with the keywords, and the identified text that successfully matches the keywords is calibrated to obtain the calibrated text. The semantic structure parsing module is used to perform semantic structure parsing on the calibrated text based on a preset set of affirmative trigger words and a set of negative trigger words, extracting and outputting noun phrases and their claim types. Specifically, it includes: segmenting the calibrated text into words and tagging them with parts of speech; sequentially scanning the results of the word segmentation; when any word in the set of affirmative trigger words is hit, entering an affirmative state, and extracting the subsequent continuous noun components as noun phrases to be judged, and outputting the noun phrase and its affirmative claim type; when any word in the set of negative trigger words is hit, entering a negative state, and extracting the subsequent continuous noun components as noun phrases to be judged, and outputting the noun phrase and its negative claim type. The cache management module is used to store the noun phrases and their claim types into the cache using the feature hash value as the key; The consistency verification module is used to compare the noun phrases and their claim types with the corresponding filing key fields and generate consistency verification results.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images