A non-standard speech question and answer evaluation method and system based on semantic anchor point feature matching

Abstract

The application discloses a non-standard voice question and answer evaluation method and system based on semantic anchor point feature matching, relates to the technical field of intelligent question and answer evaluation, and comprises the following processes: step 1, dimension reduction of a standard answer and construction of a semantic anchor point; step 2, real-time noise removal and end-side segmentation of a front-end voice stream; step 3, multi-dimensional feature matching based on dynamic time warping and cosine similarity; and step 4, a comprehensive scoring engine fusing a logical coherence penalty factor. The application converts a business long text into a high-dimensional vector set that can be quantitatively calculated through extraction and vectorization modeling of a standard answer semantic anchor point, combines a spoken language text matching algorithm based on end-side noise removal and segmentation and cosine similarity calculation, solves the fault-tolerant matching problem of synonyms and inverted sentences, and simultaneously fuses a comprehensive scoring mechanism of a scoring point hit rate and a logical coherence penalty factor, so that the multidimensionality and logical rigor of the evaluation result are ensured.

Description

Technical Field

[0001] This invention relates to the field of intelligent question-answering assessment technology, specifically to a non-standard speech question-answering assessment method and system based on semantic anchor feature matching. Background Technology

[0002] In corporate training "question-based" practice scenarios, systems often need to automatically score users' oral responses. Currently, the mainstream intelligent question-and-answer assessment methods in the industry mainly rely on the following two solutions, but they both have obvious limitations in practical implementation:

[0003] Option 1: Evaluation based on pure speech features mainly extracts acoustic indicators such as speech rate, pauses and pitch for scoring. Its disadvantage is that it only extracts acoustic features such as speech rate, pauses and volume for evaluation, and cannot judge whether the user's knowledge points are correct. It can only assess the fluency of expression, and is not suitable for scenarios of professional knowledge assessment.

[0004] Option 2: Hard matching based on text keywords. First, convert the speech to text, and then use regular expressions or string comparison algorithms to check if the user has mentioned the "keywords" in the standard answer. However, in real spoken expression, users are likely to use inverted sentences, synonym substitutions, or a lot of interjections such as "um" and "ah".

[0005] Traditional keyword matching has a low error tolerance. Even if the meaning is completely correct, the system will still misjudge and deduct points if the user rephrases the wording slightly. In addition, when faced with long and complex business answers, relying on just a few discrete keywords cannot cover the complete logic.

[0006] Existing solutions cannot truly understand non-standardized spoken texts and cannot achieve accurate scoring at a deeper semantic level. Summary of the Invention

[0007] This invention provides a non-standard speech question-answering evaluation method and system based on semantic anchor feature matching, which can effectively solve the problems mentioned in the background art.

[0008] To achieve the above objectives, this invention provides the following technical solution: a non-standard speech question-answering evaluation method based on semantic anchor feature matching, which does not rely on fixed keywords, and uses natural language processing technology to reduce the dimensionality and vectorize the standard answer and the user's spoken response, respectively, including the following process:

[0009] Step 1: Dimensionality reduction and semantic anchor construction of standard answers;

[0010] Step 2: Real-time noise reduction and edge segmentation of the front-end audio stream;

[0011] Step 3: Multidimensional feature matching based on dynamic time warping and cosine similarity;

[0012] Step 4: A comprehensive scoring engine that integrates logical coherence penalty factors;

[0013] Step 1 specifically includes the following steps:

[0014] Step 101: Receiving and preprocessing the standard answer;

[0015] Step 102: Semantic anchor extraction based on NER and dependency parsing;

[0016] Step 103: Vectorization of semantic anchors and construction of feature library;

[0017] Step 2 specifically includes the following steps:

[0018] Step 201: Real-time audio stream acquisition and buffering;

[0019] Step 202: Intelligent voice segmentation based on VAD;

[0020] Step 203: ASR conversion and text denoising of audio segments;

[0021] Step 3 specifically includes the following steps:

[0022] Step 301: Vectorization of user response text;

[0023] Step 302: Calculate the cosine similarity matrix;

[0024] Step 303: Dynamic Time Warping Matching Algorithm;

[0025] Step 304: Threshold determination and hit result generation;

[0026] Step 4 specifically includes the following steps:

[0027] Step 401: Calculation of the hit rate of the sampling point;

[0028] Step 402: Logical coherence assessment;

[0029] Step 403: Negative lexical detection and negation intent recognition;

[0030] Step 404: Calculate the overall score.

[0031] According to the above technical solution, the 101 refers to the input of long text standard answers. ,right Text cleaning was performed to obtain the cleaned standard answer text. ;

[0032] Further Word segmentation is performed using a BERT-based Chinese word segmenter to generate word sequences. ;

[0033] The 102 method utilizes the Named Entity Recognition (NER) model to analyze word sequences. Perform entity recognition and extract key entity sets. ;

[0034] Using dependency parsing models Perform syntactic analysis and construct dependency trees It can identify subject-verb-object structures and modifying relationships;

[0035] Based on dependency tree and entity set The semantic unit extraction algorithm is used to automatically break down long standard answers into several core "intent sentences" or "knowledge nodes", which are called "semantic anchors".

[0036] Finally, a set of semantic anchors is generated. ;

[0037] The 103 method utilizes a pre-trained text vectorization model to transform each extracted "semantic anchor point" into a high-dimensional vector.

[0038] For each semantic anchor Its vector representation is calculated using the Embedding model:

[0039] ;

[0040] in, , The vector dimension is typically 768 or 1024.

[0041] Generate a standard high-dimensional vector set ;

[0042] Set of semantic anchors sum vector set Stored in the feature library.

[0043] According to the above technical solution, step 201 involves collecting the real-time audio stream of the user's answers on the client side.

[0044] A circular buffer is used to buffer the audio stream, with the buffer size set to 3 seconds of audio data;

[0045] The second point, 202, introduces an end-side silence detection algorithm for VAD-based intelligent speech segmentation, employing a dual detection mechanism based on energy threshold and zero-crossing rate. The specific implementation steps are as follows:

[0046] Calculate the energy of the current audio frame ,in, For audio sampling points, Frame length;

[0047] Calculate the zero-crossing rate of the current audio frame. ;

[0048] when and When this occurs, it is determined to be a silent segment;

[0049] When the duration of continuous silence exceeds a preset threshold, At that time, it automatically segments a continuous long speech stream into multiple short audio segments;

[0050] In step 203, the ASR conversion and text denoising of the audio segments will... Send to the ASR service in real time via WebSocket;

[0051] ASR service returns recognition results. It includes confidence level information;

[0052] The front end uses regular expressions to filter out redundant interjections in the text that have no actual semantic meaning;

[0053] The denoising algorithm employs a multi-pattern matching strategy, including exact matching and fuzzy matching;

[0054] Generate a clean sequence of user response text , of which each The denoised text corresponding to an audio segment.

[0055] According to the above technical solution, step 301 involves denoising the user's response text sequence. Similarly, it is transformed into a high-dimensional vector sequence through the Embedding model. ;

[0056] A batch vectorization strategy is adopted to input multiple texts into the model at once;

[0057] Each user vector With the corresponding text Establish a mapping relationship to form a set of vector-text pairs, as follows:

[0058] ;

[0059] The 302 section calculates the user vector. With each semantic anchor vector in the standard library

[0060] The cosine similarity between them is calculated using the following formula:

[0061] ;

[0062] Construct a similarity matrix ,in, .

[0063] According to the above technical solution, the 303 introduces the Dynamic Time Warping (DTW) algorithm;

[0064] The DTW algorithm finds the optimal alignment path that maximizes the cumulative similarity between the user vector sequence and the standard vector sequence;

[0065] The DTW recursive formula is as follows:

[0066] ;

[0067] in, Indicates the user sequence before Before the vector and the standard sequence The maximum cumulative similarity of vectors;

[0068] By backtracking the optimal path, each semantic anchor point is determined. Whether the answer was matched by the user;

[0069] The 304th step sets a similarity threshold. This threshold can be dynamically adjusted according to the difficulty of the questions, which are divided into easy, medium and hard questions.

[0070] For each semantic anchor If their maximum similarity If so, then the semantic anchor point is determined to have been hit;

[0071] Synonym and Inverted Sentence Compatibility Mechanism:

[0072] Because it uses high-dimensional semantic vector calculation, even if the user uses synonyms or reverses the word order, their distance in the vector space is still close.

[0073] According to the above technical solution, step 401 calculates the number of hit semantic anchors. Total number of semantic anchors ;

[0074] The weighted hit rate is calculated as follows:

[0075] ;

[0076] in, The set of anchor points that were hit. anchor point The weights;

[0077] In response to question 402, the system introduces a "logical coherence penalty factor" to assess whether the logical order of the user's answer is reasonable. The specific assessment steps are as follows:

[0078] Extract the sequence of user-hit anchor points ;

[0079] Extracting the logical order of anchor points from standard answers ;

[0080] Sequence similarity is calculated using the LCS algorithm:

[0081] ;

[0082] Calculate the logical coherence score:

[0083] ;

[0084] like If this occurs, a logical coherence penalty will be triggered.

[0085] According to the above technical solution, step 403 detects negative word elements in the user's text sequence that contradict the core intent of the standard answer;

[0086] Employ a context-sensitive negation detection algorithm:

[0087] Identify negation words and their scope of application, and determine whether negation words apply to key entities and core concepts;

[0088] If the negation word applies to a core concept, it triggers a penalty for negation intent;

[0089] Negative Intention Punishment Factor

[0090] The calculation formula is as follows:

[0091] ;

[0092] in, The number of key anchor points that affect the negated word;

[0093] The 404 score is calculated by integrating scores from multiple dimensions, as detailed below:

[0094] ;

[0095] To prevent extreme cases, boundary limits are imposed on the final score:

[0096] .

[0097] A non-standard speech question-answering evaluation system based on semantic anchor feature matching includes a semantic anchor construction module, a front-end speech stream processing module, a multi-dimensional feature matching module, and a comprehensive scoring engine module.

[0098] According to the above technical solution, the semantic anchor construction module performs dimensionality reduction processing on the standard answer of the long text, extracts semantic anchors and stores them in a quantized manner, including a receiving and preprocessing unit, a semantic anchor extraction unit, and a vectorization and feature library construction unit;

[0099] The receiving and preprocessing unit performs text cleaning on the standard answers and uses a BERT-based Chinese word segmenter for word segmentation.

[0100] The semantic anchor extraction unit includes NER entity recognition, dependency parsing, and semantic unit extraction algorithms;

[0101] The vectorization and feature library construction unit uses a pre-trained Embedding model to transform semantic anchors into high-dimensional vectors and stores the set of semantic anchors and the set of vectors into the feature library;

[0102] The front-end speech stream processing module performs user speech acquisition, noise reduction, and intelligent segmentation on the client side, including a speech stream acquisition and buffering unit, an intelligent speech segmentation unit, and an ASR conversion and text denoising unit.

[0103] The voice stream acquisition and buffering unit acquires the real-time voice stream of the user's answers and uses a ring buffer to cope with network fluctuations;

[0104] The intelligent voice segmentation unit performs silence detection based on a dual detection mechanism of energy threshold and zero crossing rate. When continuous silence exceeds 800ms, it automatically segments the speech into multiple short audio segments.

[0105] The ASR conversion and text denoising unit sends audio segments to the ASR service via WebSocket, and uses regular expressions and multi-pattern matching strategies to filter out interjections.

[0106] According to the above technical solution, the multidimensional feature matching module performs semantic vector matching between the user's answer and the standard answer, and calculates the hit result, including a user text vectorization unit, a cosine similarity calculation unit, a DTW dynamic time warping matching unit, and a threshold determination and hit result generation unit.

[0107] The user text vectorization unit transforms the denoised user response text into a high-dimensional vector through the Embedding model, using a batch vectorization strategy.

[0108] The cosine similarity calculation unit calculates the cosine similarity between the user vector and the semantic anchor vector, and constructs a similarity matrix;

[0109] The DTW Dynamic Time Warping Matching Unit introduces the DTW algorithm to solve the problem of inconsistency between user answers and standard answers in terms of time sequence. It maximizes the cumulative similarity by finding the optimal alignment path.

[0110] The threshold determination and hit result generation unit sets a similarity threshold to determine whether the semantic anchor point is hit, and supports fault-tolerant matching of synonyms and inverted sentences.

[0111] The comprehensive scoring engine module integrates evaluation factors from multiple dimensions to calculate the final comprehensive score, including a scoring point hit rate calculation unit, a logical coherence evaluation unit, a negative word element detection and negation intent recognition unit, and a comprehensive score calculation unit.

[0112] The sampling point hit rate calculation unit calculates the weighted hit rate, taking into account the anchor point weight;

[0113] The logical coherence assessment unit extracts the order of the user's hit anchor points and the logical order of the standard answer, and uses the LCS algorithm to calculate the sequence similarity. When the similarity is lower than the threshold, a logical coherence penalty is triggered.

[0114] The negative lexical detection and negation intent recognition unit detects negative lexical units and uses a context-sensitive negation detection algorithm to determine whether the negation word applies to the core concept and triggers negation intent penalty.

[0115] The comprehensive score calculation unit integrates the hit rate of scoring points, the logical coherence score, and the negative intent penalty factor to impose boundary restrictions on the final score.

[0116] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0117] This invention is based on multi-level semantic understanding, breaking the limitations of traditional "keyword matching". Even if users use highly colloquial expressions, synonym substitutions, or inverted sentences, as long as the core "semantic anchor" is successfully matched in the vector space, a score can be obtained, which greatly improves the accuracy of intelligent grading. Moreover, from NER entity recognition to dependency parsing and then to high-dimensional vector space matching, it achieves a comprehensive understanding from lexical, syntactic and semantic levels.

[0118] Through intelligent speech processing, the VAD-based intelligent segmentation and context-aware denoising algorithm effectively handle redundant information in colloquial expressions. At the same time, based on a robust matching algorithm, combined with DTW dynamic time warping and cosine similarity, it solves the problem of inconsistent user response order, while also being compatible with synonyms and inverted sentences.

[0119] Furthermore, by extracting and vectorizing the semantic anchor points of the standard answer, the long business text is transformed into a set of high-dimensional vectors that can be quantified and calculated. Combined with the colloquial text matching algorithm of edge-side denoising and segmentation and cosine similarity calculation, the error-tolerant matching problem of synonyms and inverted sentences is solved. At the same time, the comprehensive scoring mechanism of the scoring point hit rate and logical coherence penalty factor is integrated to ensure the multidimensionality and logical rigor of the evaluation results. Attached Figure Description

[0120] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0121] In the attached diagram: Figure 1 This is a flowchart of the evaluation method of the present invention; Figure 2 This is a flowchart of the overall process of the present invention; Figure 3 This is the logic diagram of the semantic anchor feature matching algorithm of this invention; Figure 4 This is a structural block diagram of the evaluation system of this invention; Figure 5 This is the architecture deployment diagram of the evaluation system of this invention. Detailed Implementation

[0127] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0128] Example: Figure 1-3 As shown, this invention provides a technical solution: a non-standard speech question-and-answer assessment method based on semantic anchor feature matching. It does not rely on fixed keywords and uses natural language processing technology to reduce the dimensionality and vectorize the standard answer and the user's spoken response, respectively, to achieve intelligent grading. The method includes the following process:

[0129] Step 1: Dimensionality reduction and semantic anchor construction of standard answers;

[0130] Step 2: Real-time noise reduction and edge segmentation of the front-end audio stream;

[0131] Step 3: Multidimensional feature matching based on dynamic time warping and cosine similarity;

[0132] Step 4: A comprehensive scoring engine that integrates logical coherence penalty factors;

[0133] Step 1 specifically includes the following steps:

[0134] Step 101: Receiving and preprocessing the standard answer;

[0135] Step 102: Semantic anchor extraction based on NER and dependency parsing;

[0136] Step 103: Vectorization of semantic anchors and construction of feature library;

[0137] Step 2 specifically includes the following steps:

[0138] Step 201: Real-time audio stream acquisition and buffering;

[0139] Step 202: Intelligent voice segmentation based on VAD;

[0140] Step 203: ASR conversion and text denoising of audio segments;

[0141] Step 3 specifically includes the following steps:

[0142] Step 301: Vectorization of user response text;

[0143] Step 302: Calculate the cosine similarity matrix;

[0144] Step 303: Dynamic Time Warping Matching Algorithm;

[0145] Step 304: Threshold determination and hit result generation;

[0146] Step 4 specifically includes the following steps:

[0147] Step 401: Calculation of the hit rate of the sampling point;

[0148] Step 402: Logical coherence assessment;

[0149] Step 403: Negative lexical detection and negation intent recognition;

[0150] Step 404: Calculate the overall score.

[0151] Based on the above technical solution, 101, the entered long text standard answer. ,right

[0152] Text cleaning involves removing HTML tags, special characters, and extra spaces to obtain the cleaned standard answer text. ;

[0153] Further Word segmentation is performed using a BERT-based Chinese word segmenter to generate word sequences. ;

[0154] 102. Using the Named Entity Recognition (NER) model to analyze word sequences. Perform entity recognition and extract key entity sets. This includes product names, disease names, and technical terms;

[0155] Using dependency parsing models Perform syntactic analysis and construct dependency trees

[0156] It identifies grammatical information, including subject-verb-object structure and modification relationships;

[0157] Based on dependency tree and entity set The semantic unit extraction algorithm is used to automatically break down long standard answers into several core "intent phrases" and "knowledge nodes", which are called "semantic anchors".

[0158] Finally, a set of semantic anchors is generated. ;

[0159] 103. Using a pre-trained text vectorization model, the extracted "semantic anchors" are transformed into high-dimensional vectors;

[0160] For each semantic anchor Its vector representation is calculated using the Embedding model:

[0161] ;

[0162] in, , The vector dimension is typically 768;

[0163] Generate a standard high-dimensional vector set ;

[0164] Set of semantic anchors sum vector set Stored in the feature library.

[0165] Based on the above technical solution, 201, on the client side, the real-time audio stream of the user's answers is collected;

[0166] A circular buffer is used to cache the audio stream, with the buffer size set to 3 seconds of audio data to ensure that there is enough data for segmentation when the network fluctuates;

[0167] 202. The intelligent speech segmentation based on VAD introduces an end-side silence detection algorithm, which adopts a dual detection mechanism based on energy threshold and zero-crossing rate. The specific implementation steps are as follows:

[0168] Calculate the energy of the current audio frame ,in, For audio sampling points, Frame length;

[0169] Calculate the zero-crossing rate of the current audio frame. ;

[0170] when and When this occurs, it is determined to be a silent segment;

[0171] When the duration of continuous silence exceeds a preset threshold, At that time, it automatically segments a continuous long speech stream into multiple short audio segments;

[0172] 203, ASR Conversion and Text Denoising of Audio Segments: The segmented audio segments...

[0173] Send to the ASR service in real time via WebSocket;

[0174] ASR service returns recognition results. It includes confidence level information;

[0175] The front end uses regular expressions to filter out redundant interjections in the text that have no actual semantic meaning;

[0176] The dictionary of redundant modal particles includes, but is not limited to: uh, um, ah, this, that;

[0177] The denoising algorithm employs a multi-pattern matching strategy, including exact matching and fuzzy matching;

[0178] Exact match directly deletes the complete word from the dictionary;

[0179] Fuzzy matching removes substrings containing dictionary terms;

[0180] Generate a clean sequence of user response text , of which each The denoised text corresponding to an audio segment.

[0181] Based on the above technical solution, 301, the denoised user response text sequence is... Similarly, it is transformed into a high-dimensional vector sequence through the Embedding model. ;

[0182] To improve matching efficiency, a batch vectorization strategy is adopted, which inputs multiple texts into the model at once to reduce the number of model calls;

[0183] Each user vector With the corresponding text Establish a mapping relationship to form a set of vector-text pairs, as follows:

[0184] ;

[0185] 302, Calculate the user vector With each semantic anchor vector in the standard library The cosine similarity between them is calculated using the following formula:

[0186] ;

[0187] Construct a similarity matrix ,in, .

[0188] Based on the above technical solution, 303, in order to solve the problem of inconsistency between user answers and standard answers in terms of time sequence, the Dynamic Time Warping (DTW) algorithm is introduced. DTW: Dynamic Time Warping;

[0189] The DTW algorithm finds the optimal alignment path that maximizes the cumulative similarity between the user vector sequence and the standard vector sequence;

[0190] The DTW recursive formula is as follows:

[0191] ;

[0192] in, Indicates the user sequence before Before the vector and the standard sequence The maximum cumulative similarity of vectors;

[0193] By backtracking the optimal path, each semantic anchor point is determined. Whether the answer was matched by the user;

[0194] 304, set a similarity threshold This threshold can be dynamically adjusted based on the difficulty of the questions, which are categorized into easy, medium, and difficult levels, as detailed below:

[0195] Easy question: Medium difficulty level: Difficult question: ;

[0196] For each semantic anchor If their maximum similarity If so, then the semantic anchor point is determined to have been hit;

[0197] Synonym and Inverted Sentence Compatibility Mechanism:

[0198] Because it uses high-dimensional semantic vector calculation, even if the user uses synonyms and reverses the word order, their distance in the vector space is still close. Specifically, the synonym is: saying "cheap" as "low price".

[0199] Specifically, the vector similarity between "this product is very cheap" and "this product is very inexpensive" can reach over 0.92, thus effectively solving the problem of traditional keyword matching failure.

[0200] Based on the above technical solution, 401, calculate the number of hit semantic anchors. Total number of semantic anchors ;

[0201] The weighted hit rate is calculated as follows:

[0202] ;

[0203] in, The set of anchor points that were hit. anchor point The weights;

[0204] 402. The system introduces a "logical coherence penalty factor" to evaluate whether the logical order of the user's answer is reasonable. The specific evaluation steps are as follows:

[0205] Extract the sequence of user-hit anchor points ;

[0206] Extracting the logical order of anchor points from standard answers ;

[0207] Sequence similarity is calculated using the LCS algorithm:

[0208] ;

[0209] Calculate the logical coherence score:

[0210] ;

[0211] like If this occurs, a logical coherence penalty will be triggered.

[0212] Based on the above technical solution, 403, the user's text sequence contains negative word elements that contradict the core intent of the standard answer;

[0213] The negative lexicon includes: cannot, is not, does not, wrong, error, failure, unable;

[0214] Employ a context-sensitive negation detection algorithm:

[0215] Identify the negation word and its scope of application, which is usually the 2-3 words following the negation word;

[0216] Determine whether the negation term applies to key entities and core concepts;

[0217] If the negation word applies to a core concept, it triggers a penalty for negation intent;

[0218] Negative Intention Punishment Factor The calculation formula is as follows:

[0219] ;

[0220] in, The number of key anchor points that affect the negated word;

[0221] 404. The final comprehensive score is calculated by integrating scores from multiple dimensions, as detailed below:

[0222] ;

[0223] To prevent extreme cases, boundary limits are imposed on the final score:

[0224] .

[0225] like Figure 4-5 As shown, a non-standard voice question-answering evaluation system based on semantic anchor feature matching includes a semantic anchor construction module, a front-end voice stream processing module, a multi-dimensional feature matching module, and a comprehensive scoring engine module, realizing complete evaluation processing from standard answer processing, user voice processing, semantic matching to comprehensive scoring.

[0226] Based on the above technical solution, the semantic anchor construction module performs dimensionality reduction on the standard answer of the long text, extracts semantic anchors and stores them in a quantized manner, including a receiving and preprocessing unit, a semantic anchor extraction unit, and a vectorization and feature library construction unit.

[0227] The receiving and preprocessing unit performs text cleaning on the standard answers, including removing HTML tags, special characters, and extra spaces, and uses a BERT-based Chinese word segmenter for word segmentation.

[0228] The semantic anchor extraction unit includes NER entity recognition, dependency parsing, and semantic unit extraction algorithms;

[0229] NER Entity Recognition: Extracts key entities using a named entity recognition model, including product names, disease names, and technical terms;

[0230] Dependency parsing: Constructs dependency trees to identify grammatical information, including subject-verb-object structures and modification relations;

[0231] Semantic unit extraction algorithm: Decompose long standard answers into "intent sentences" or "knowledge nodes", i.e. semantic anchors;

[0232] The vectorization and feature library construction unit uses a pre-trained Embedding model to transform semantic anchors into high-dimensional vectors with a dimension of 768, and stores the set of semantic anchors and the set of vectors into the feature library;

[0233] The front-end speech stream processing module performs voice acquisition, noise reduction, and intelligent segmentation on the client side, including a speech stream acquisition and buffering unit, a VAD intelligent speech segmentation unit, and an ASR conversion and text denoising unit.

[0234] The voice stream acquisition and buffering unit acquires the real-time voice stream of the user's answers and uses a circular buffer to cope with network fluctuations. The circular buffer contains 3 seconds of audio data.

[0235] The intelligent voice segmentation unit (VAD) performs silence detection based on a dual detection mechanism of energy threshold and zero-crossing rate. When continuous silence exceeds 800ms, it automatically segments the audio into multiple short sentence segments.

[0236] The ASR conversion and text denoising unit sends audio segments to the ASR service via WebSocket, and uses regular expressions and multi-pattern matching strategies to filter interjections. The matching strategies include exact matching and fuzzy matching, and the interjections include "um," "ah," and "this."

[0237] Based on the above technical solution, the multi-dimensional feature matching module performs semantic vector matching between user answers and standard answers to calculate the hit result, including user text vectorization unit, cosine similarity calculation unit, DTW dynamic time warping matching unit, threshold determination and hit result generation unit.

[0238] The user text vectorization unit transforms the denoised user response text into a high-dimensional vector through the Embedding model, and adopts a batch vectorization strategy to improve efficiency.

[0239] The cosine similarity calculation unit calculates the cosine similarity between the user vector and the semantic anchor vector, and constructs a similarity matrix;

[0240] The DTW Dynamic Time Warping Matching Unit introduces the DTW algorithm to solve the problem of inconsistency between user answers and standard answers in terms of time sequence. It maximizes the cumulative similarity by finding the optimal alignment path.

[0241] The threshold determination and hit result generation unit sets similarity thresholds, specifically 0.90 for easy questions, 0.85 for medium questions, and 0.80 for difficult questions, to determine whether semantic anchors are hit, and supports fault-tolerant matching of synonyms and inverted sentences;

[0242] The comprehensive scoring engine module integrates evaluation factors from multiple dimensions to calculate the final comprehensive score, including a scoring point hit rate calculation unit, a logical coherence evaluation unit, a negative word element detection and negation intent recognition unit, and a comprehensive score calculation unit.

[0243] The sampling point hit rate calculation unit calculates the weighted hit rate, taking into account the anchor point weight;

[0244] The logical coherence assessment unit extracts the order of the user's hit anchor points and the logical order of the standard answer, and uses the LCS algorithm to calculate the sequence similarity. When the similarity is lower than the threshold, a logical coherence penalty is triggered.

[0245] The negative lexical detection and negation intent recognition unit detects negative lexical units, including "cannot," "not," and "no." The context-sensitive negation detection algorithm determines whether the negation word applies to the core concept and triggers the negation intent penalty.

[0246] The comprehensive score calculation unit integrates the hit rate of scoring points, the logical coherence score, and the negative intent penalty factor to impose boundary restrictions on the final score, ranging from 0 to 100.

[0247] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A non-standard speech question-answering evaluation method based on semantic anchor feature matching, characterized in that: Without relying on fixed keywords, this method uses natural language processing technology to reduce the dimensionality of standard answers and vectorize user spoken responses, including the following process: Step 1: Dimensionality reduction and semantic anchor construction of standard answers; Step 2: Real-time noise reduction and edge segmentation of the front-end audio stream; Step 3: Multidimensional feature matching based on dynamic time warping and cosine similarity; Step 4: A comprehensive scoring engine that integrates logical coherence penalty factors; Step 1 specifically includes the following steps: Step 101: Receiving and preprocessing the standard answer; Step 102: Semantic anchor extraction based on NER and dependency parsing; Step 103: Vectorization of semantic anchors and construction of feature library; Step 2 specifically includes the following steps: Step 201: Real-time audio stream acquisition and buffering; Step 202: Intelligent voice segmentation based on VAD; Step 203: ASR conversion and text denoising of audio segments; Step 3 specifically includes the following steps: Step 301: Vectorization of user response text; Step 302: Calculate the cosine similarity matrix; Step 303: Dynamic Time Warping Matching Algorithm; Step 304: Threshold determination and hit result generation; Step 4 specifically includes the following steps: Step 401: Calculation of the hit rate of the sampling point; Step 402: Logical coherence assessment; Step 403: Negative lexical detection and negation intent recognition; Step 404: Calculate the overall score.

2. The non-standard speech question-answering evaluation method based on semantic anchor feature matching according to claim 1, characterized in that: The above 101 refers to the long text standard answer that was entered. ,right Text cleaning was performed to obtain the cleaned standard answer text. ; Further Word segmentation is performed using a BERT-based Chinese word segmenter to generate word sequences. ; The 102 method utilizes the Named Entity Recognition (NER) model to analyze word sequences. Perform entity recognition and extract key entity sets. ; Using dependency parsing models Perform syntactic analysis and construct dependency trees It can identify subject-verb-object structures and modifying relationships; Based on dependency tree and entity set The semantic unit extraction algorithm is used to automatically break down long standard answers into several core "intent phrases" or "knowledge nodes", which are called "semantic anchors". Finally, a set of semantic anchors is generated. ; The 103 method utilizes a pre-trained text vectorization model to transform each extracted "semantic anchor point" into a high-dimensional vector. For each semantic anchor Its vector representation is calculated using the Embedding model: ； in, , The vector dimension is typically 768 or 1024. Generate a standard high-dimensional vector set ; Set of semantic anchors sum vector set Stored in the feature library.

3. The non-standard speech question-answering evaluation method based on semantic anchor feature matching according to claim 1, characterized in that: In step 201, on the client side, the real-time audio stream of the user's answers is collected; A circular buffer is used to buffer the audio stream, with the buffer size set to 3 seconds of audio data; The second point, 202, introduces an end-side silence detection algorithm for VAD-based intelligent speech segmentation, employing a dual detection mechanism based on energy threshold and zero-crossing rate. The specific implementation steps are as follows: Calculate the energy of the current audio frame ,in, For audio sampling points, Frame length; Calculate the zero-crossing rate of the current audio frame. ; when and When this occurs, it is determined to be a silent segment; When the duration of continuous silence exceeds a preset threshold, At that time, it automatically segments a continuous long speech stream into multiple short audio segments; In step 203, the ASR conversion and text denoising of the audio segments will... Send to the ASR service in real time via WebSocket; ASR service returns recognition results. It includes confidence level information; The front end uses regular expressions to filter out redundant interjections in the text that have no actual semantic meaning; The denoising algorithm employs a multi-pattern matching strategy, including exact matching and fuzzy matching; Generate a clean sequence of user response text , of which each The denoised text corresponding to an audio segment.

4. The non-standard speech question-answering evaluation method based on semantic anchor feature matching according to claim 1, characterized in that: The 301 step involves denoising the user response text sequence. Similarly, it is transformed into a high-dimensional vector sequence through the Embedding model. ; A batch vectorization strategy is adopted to input multiple texts into the model at once; Each user vector With the corresponding text Establish a mapping relationship to form a set of vector-text pairs, as follows: ； The 302 section calculates the user vector. With each semantic anchor vector in the standard library The cosine similarity between them is calculated using the following formula: ； Construct a similarity matrix ,in, .

5. The non-standard speech question-answering evaluation method based on semantic anchor feature matching according to claim 4, characterized in that: The 303 method introduces the Dynamic Time Warping (DTW) algorithm. The DTW algorithm finds the optimal alignment path that maximizes the cumulative similarity between the user vector sequence and the standard vector sequence; The DTW recursive formula is as follows: ； in, Indicates the user sequence before Before the vector and the standard sequence The maximum cumulative similarity of vectors; By backtracking the optimal path, each semantic anchor point is determined. Whether the user's answer was correct; The 304th step sets a similarity threshold. This threshold can be dynamically adjusted according to the difficulty of the questions, which are divided into easy, medium and hard questions. For each semantic anchor If their maximum similarity If so, then the semantic anchor point is determined to have been hit; Synonym and Inverted Sentence Compatibility Mechanism: Because it uses high-dimensional semantic vector calculation, even if the user uses synonyms or reverses the word order, their distance in the vector space is still close.

6. The non-standard speech question-answering evaluation method based on semantic anchor feature matching according to claim 1, characterized in that: In step 401, the number of hit semantic anchors is calculated. Total number of semantic anchors ; The weighted hit rate is calculated as follows: ； in, The set of anchor points that were hit. anchor point The weights; In response to question 402, the system introduces a "logical coherence penalty factor" to assess whether the logical order of the user's answer is reasonable. The specific assessment steps are as follows: Extract the sequence of user-hit anchor points ; Extracting the logical order of anchor points from standard answers ; Sequence similarity is calculated using the LCS algorithm: ； Calculate the logical coherence score: ； like If this occurs, a logical coherence penalty will be triggered.

7. A non-standard speech question-answering evaluation method based on semantic anchor feature matching according to claim 6, characterized in that: The 403 method detects negative word elements in the user's text sequence that contradict the core intent of the standard answer. Employ a context-sensitive negation detection algorithm: Identify negation words and their scope of application, and determine whether negation words apply to key entities and core concepts; If the negation word applies to a core concept, it triggers a penalty for negation intent; Negative Intention Punishment Factor The calculation formula is as follows: ； in, The number of key anchor points that affect the negated word; The 404 score is calculated by integrating scores from multiple dimensions, as detailed below: ； To prevent extreme cases, boundary limits are imposed on the final score: 。 8. A non-standard speech question-answering evaluation system based on semantic anchor feature matching, characterized in that: It includes a semantic anchor construction module, a front-end speech stream processing module, a multi-dimensional feature matching module, and a comprehensive scoring engine module.

9. A non-standard speech question-answering evaluation system based on semantic anchor feature matching according to claim 8, characterized in that: The semantic anchor construction module performs dimensionality reduction on the standard answer of the long text, extracts semantic anchors and stores them in a quantized manner, including a receiving and preprocessing unit, a semantic anchor extraction unit, and a vectorization and feature library construction unit; The receiving and preprocessing unit performs text cleaning on the standard answers and uses a BERT-based Chinese word segmenter for word segmentation. The semantic anchor extraction unit includes NER entity recognition, dependency parsing, and semantic unit extraction algorithms; The vectorization and feature library construction unit uses a pre-trained Embedding model to transform semantic anchors into high-dimensional vectors and stores the set of semantic anchors and the set of vectors into the feature library; The front-end speech stream processing module performs user speech acquisition, noise reduction, and intelligent segmentation on the client side, including a speech stream acquisition and buffering unit, an intelligent speech segmentation unit, and an ASR conversion and text denoising unit. The voice stream acquisition and buffering unit acquires the real-time voice stream of the user's answers and uses a ring buffer to cope with network fluctuations; The intelligent voice segmentation unit performs silence detection based on a dual detection mechanism of energy threshold and zero crossing rate. When continuous silence exceeds 800ms, it automatically segments the speech into multiple short audio segments. The ASR conversion and text denoising unit sends audio segments to the ASR service via WebSocket, and uses regular expressions and multi-pattern matching strategies to filter out interjections.

10. A non-standard speech question-answering evaluation system based on semantic anchor feature matching according to claim 8, characterized in that: The multidimensional feature matching module performs semantic vector matching between user answers and standard answers to calculate the hit result, including a user text vectorization unit, a cosine similarity calculation unit, a DTW dynamic time warping matching unit, and a threshold determination and hit result generation unit. The user text vectorization unit transforms the denoised user response text into a high-dimensional vector through the Embedding model, using a batch vectorization strategy. The cosine similarity calculation unit calculates the cosine similarity between the user vector and the semantic anchor vector, and constructs a similarity matrix; The DTW Dynamic Time Warping Matching Unit introduces the DTW algorithm to solve the problem of inconsistency between user answers and standard answers in terms of time sequence. It maximizes the cumulative similarity by finding the optimal alignment path. The threshold determination and hit result generation unit sets a similarity threshold to determine whether the semantic anchor point is hit, and supports fault-tolerant matching of synonyms and inverted sentences. The comprehensive scoring engine module integrates evaluation factors from multiple dimensions to calculate the final comprehensive score, including a scoring point hit rate calculation unit, a logical coherence evaluation unit, a negative word element detection and negation intent recognition unit, and a comprehensive score calculation unit. The sampling point hit rate calculation unit calculates the weighted hit rate, taking into account the anchor point weight; The logical coherence assessment unit extracts the order of the user's hit anchor points and the logical order of the standard answer, and uses the LCS algorithm to calculate the sequence similarity. When the similarity is lower than the threshold, a logical coherence penalty is triggered. The negative lexical detection and negation intent recognition unit detects negative lexical units and uses a context-sensitive negation detection algorithm to determine whether the negation word applies to the core concept and triggers the negation intent penalty. The comprehensive score calculation unit integrates the hit rate of scoring points, the logical coherence score, and the negative intent penalty factor to impose boundary restrictions on the final score.

Images