Education consultation service method based on artificial intelligence technology assistance

Abstract

The invention belongs to the technical field of artificial intelligence, and discloses an artificial intelligence technology assistance-based education consultation service method, which comprises the steps of obtaining original voice data in an education consultation scene, and removing background noise through time-frequency masking processing to obtain high-quality de-noised voice data; performing character transcription on the de-noised voice data, performing adaptive text segmentation by using a pre-constructed term index database, and identifying and replacing terms which are wrongly transcribed; the method comprises the following steps of: normalizing oral statements in a text through anaphora resolution based on a context memory mechanism and an omission and completion technology based on dependency syntactic analysis; generating a preliminary scheme framework containing problem diagnosis, core suggestion and implementation steps by utilizing a large language model based on the standardized text; according to the method, the accuracy of voice recognition and the continuity of semantic understanding in education consultation can be effectively improved, and automatic and structured output of professional consultation schemes is realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method for providing educational consulting services based on artificial intelligence technology. Background Technology

[0002] With the deep integration of artificial intelligence technology and the education industry, educational consulting services have become an important means to help parents and students plan their learning paths and resolve confusion regarding college entrance examinations. Traditional educational consulting services mainly rely on human advisors, which suffers from problems such as high costs, slow response times, and difficulty in standardizing service quality. In recent years, intelligent consulting systems based on Large Language Models (LLM) have gradually emerged, attempting to provide consulting services through automated means.

[0003] However, in practical educational consulting applications, existing technical solutions still face many challenges, mainly in the following aspects: Speech recognition accuracy is low in complex acoustic environments. Educational consultation scenarios often occur in open environments such as offices, where background noise is complex and variable. Existing speech front-end processing technologies typically use fixed noise reduction algorithms, lacking adaptive processing of the internal characteristics of the speech signal. For example, traditional noise reduction methods often fail to adequately distinguish between unvoiced and voiced segments in speech, resulting in damage to unvoiced components (high-frequency components) or residual noise in voiced segments while removing background noise. This signal distortion directly leads to a high error rate in subsequent automatic speech recognition (ASR) systems when processing consultation speech, severely impacting the foundation of semantic understanding.

[0004] Difficulties in identifying specialized terminology and chaotic text segmentation. The field of educational consulting contains a large number of specific specialized terms. General speech recognition models lack domain knowledge and are prone to transcribing these specialized terms into similar-sounding general words (e.g., recognizing "IELTS" as "yasi"). In addition, existing text processing workflows often directly process the transcribed long text without an adaptive segmentation mechanism based on domain knowledge. This leads to key terms being incorrectly truncated or confused with the context, making it difficult for subsequent semantic analysis to capture core information.

[0005] Semantic comprehension biases caused by colloquial expressions. In consultation dialogues, users' language expressions exhibit significant colloquial characteristics, mainly manifested in unclear referents (frequent use of pronouns such as "this," "that," and "he") and omissions of components (such as omitting subjects or predicates when answering questions). Existing technical solutions mostly directly input transcribed text into large language models, lacking specialized standardization processing steps for colloquial texts (such as referential resolution and omission completion). This leads to the model easily developing illusions when processing multi-turn dialogues due to loss of contextual memory or unclear referents, failing to accurately understand the user's true intentions. In view of this, this invention proposes an AI-assisted educational consultation service method to address the above problems. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a method for providing educational consulting services based on artificial intelligence technology, comprising: The original speech data in the educational consultation scenario is obtained, and the background noise of the original speech data is removed by time-frequency masking to obtain denoised speech data; The denoised speech data is transcribed into text to obtain transcribed text data. Based on a pre-built term index library, the transcribed text data is adaptively segmented to generate an initial text unit sequence. The pre-built term index library is used to identify and replace transcribed terms in the initial text unit sequence to obtain preliminary text data; By using reference resolution based on context memory mechanism and omission completion based on dependency parsing, the sentences in the preliminary text data are processed to obtain normalized text; Based on the standardized text, a preliminary solution framework containing problem diagnosis, core recommendations, and implementation steps is generated using a large language model, and the preliminary solution framework is sent to the display terminal.

[0007] Furthermore, methods for removing background noise from the original speech data through time-frequency masking include: The original speech data is subjected to frame-segmentation and windowing processing to obtain several speech frames; The speech frame is converted to the frequency domain by short-time Fourier transform to obtain the complex spectrum; The complex spectrum is input into a pre-trained deep complex convolutional network to generate a time-frequency mask matrix; The time-frequency mask matrix is ​​multiplied point-by-point with the complex spectrum to obtain the enhanced complex spectrum; The enhanced complex spectrum is subjected to inverse short-time Fourier transform and overlap-addition processing to obtain denoised speech data.

[0008] Furthermore, the frame-by-frame windowing process employs an adaptive windowing method, including: Identify speech segments in the original speech data and determine voicing / unvoicing within each speech segment; wherein the method for determining voicing / unvoicing is as follows: The speech segment is pre-framed using a preset initial window length to obtain pre-framed units; The acoustic features of the pre-framing unit are extracted, including the signal zero-crossing rate, short-time energy, spectral centroid, and autocorrelation function peak value. The extracted acoustic features are combined into a feature vector and input into a pre-trained voiced / unvoiced sound classification model; Based on the output of the voiced / unvoiced sound classification model, the voiced / unvoiced sound attributes of the pre-framing unit are determined; When a speech segment is determined to be unvoiced, the first window length is used to re-frame the speech segment. When a speech segment is determined to be voiced, the second window length is used to re-frame the speech segment. The length of the first window is shorter than the length of the second window.

[0009] Furthermore, the training methods for the deep complex convolutional network include: Construct a training dataset, which includes multiple sets of noisy speech data collected in educational consultation scenarios and their corresponding clean speech data; The scale-invariant signal-to-noise ratio between the denoised speech data output during forward propagation and the clean speech data is used as the loss function, and the network weight parameters are updated through the error backpropagation algorithm. The deep complex convolutional network adopts a U-Net structure of encoder-bottleneck layer-decoder, where the encoder is used for downsampling to extract high-level abstract features and the decoder is used for upsampling to recover high-resolution time-frequency masks.

[0010] Furthermore, the methods for generating the initial text unit sequence include: The denoised speech data is transcribed into text using an automatic speech recognition model to obtain transcribed text data; the transcribed text data includes a character sequence and the corresponding character code and timestamp. The character encoding is traversed to determine the positions of ASCII character segments and Chinese character segments, and a segmentation mark is added at the junction of the two character types; The character sequence is segmented based on the segmentation marker to generate a pre-segmented sequence; The pre-built terminology index library is used to perform a forward scan on the pre-segmented sequence. Based on the maximum length matching principle, the matched technical terms are locked as indivisible atomic units. The timestamp of the atomic unit is defined as the time from the start time of the first character to the end time of the last character it contains. The atomic unit is treated as a character. The timestamp difference between adjacent characters is calculated. If the timestamp difference exceeds a preset silence threshold, the adjacent characters are segmented to obtain the initial text unit sequence.

[0011] Furthermore, methods for identifying and replacing transcribed terms in the initial text unit sequence include: The initial text unit sequence is tagged with part-of-speech tags to extract text segments with noun attributes, and these text segments are used as the word units to be detected. Based on the word unit to be detected, a fuzzy search is performed in a pre-built term index library to obtain multiple standard candidate terms with similar pronunciations; Obtain the pinyin sequence of the word unit to be checked and the pinyin sequence of the standard candidate terms; The pinyin sequence to be tested is aligned with the pinyin sequence of the term, and the longest common subsequence of the two is extracted. Calculate the ratio of the length of the longest common subsequence to the length of the term pinyin sequence, and record this ratio as the speech matching degree; When the speech matching degree exceeds the preset confidence threshold and the text of the word to be tested is inconsistent with the text of the standard candidate term, it is determined to be a transcription error, and the standard candidate term with the highest speech matching degree replaces the word to be tested, generating preliminary text data.

[0012] Furthermore, the reference resolution based on the context memory mechanism includes the following steps: Construct and maintain an entity cache stack that follows the last-in-first-out principle. The entity cache stack is used to store named entities and their type attributes that appear in historical dialogues. Identify the current named entity and indicator pronoun from the preliminary text data; Extract the governing words that govern the demonstrative pronouns based on dependency parsing; Calculate the semantic compatibility score between each historical named entity in the entity cache stack and the current indicator pronoun and its governing word, the semantic compatibility score being calculated based on the semantic preference of the governing word; The historical named entity with the highest semantic compatibility score is selected to replace the indicator pronoun, and the current named entity and its type attribute are stored in the entity cache stack.

[0013] Furthermore, the semantic compatibility score is calculated in the following manner: Based on a pre-built verb argument collocation dictionary, the semantic preference of the governing word for the argument component is obtained; A matching factor is obtained based on the degree of matching between this semantic preference and the attributes of historical named entity types; The time decay factor is obtained based on the difference in the number of dialogue turns between historical named entities and the current statement; The product of the matching factor and the time decay factor is calculated as the semantic compatibility score.

[0014] Furthermore, the ellipsis completion based on dependency parsing includes the following steps: Perform dependency parsing on the current statement to check for missing core arguments; If a core argument is missing, retrieve the previous statement adjacent to the current statement from the dialogue history of the current session. Dependency parsing is performed on the previous round of statements. Based on the question-and-answer interaction logic and the principle of syntactic consistency, components with the same syntactic role as the missing core arguments are extracted. The extracted components are filled into the missing positions of the current sentence to obtain the normalized text.

[0015] Furthermore, the problem diagnosis includes the user's core educational needs and the root cause of the problem; the core recommendations include the key points of the solution to the root cause of the problem; and the implementation steps include breaking down the core recommendations into specific, executable operational steps.

[0016] The technical effects and advantages of this invention are as follows: This invention introduces adaptive frame-segmentation and windowing based on voicing / unvoicing judgment and a deep complex convolutional network to suppress background noise while maintaining time-frequency resolution, preserving high-frequency details of unvoiced sounds, and improving the signal-to-noise ratio and integrity of speech data. It combines a pre-built terminology index library for adaptive text segmentation and terminology correction based on speech matching degree to correct transcriptional biases in educational terminology. Utilizing context-based memory-based referential resolution and dependency parsing-based omission completion techniques, it restores missing semantic components and referents in spoken dialogue, transforming unstructured spoken language into semantically complete standard text. This, in turn, drives a large language model to generate structured solutions containing problem diagnosis, core suggestions, and implementation steps, improving the accuracy and feasibility of educational consulting services. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of an educational consulting service method based on artificial intelligence technology according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 Please see Figure 1 As shown in this embodiment, an educational consulting service method based on artificial intelligence technology includes: The original speech data in the educational consultation scenario is obtained, and the background noise of the original speech data is removed by time-frequency masking to obtain denoised speech data; The denoised speech data is transcribed into text to obtain transcribed text data. Based on a pre-built term index library, the transcribed text data is adaptively segmented to generate an initial text unit sequence. The pre-built term index library is used to identify and replace transcribed terms in the initial text unit sequence to obtain preliminary text data; By using reference resolution based on context memory mechanism and omission completion based on dependency parsing, the sentences in the preliminary text data are processed to obtain normalized text; Based on the standardized text, a preliminary solution framework containing problem diagnosis, core recommendations, and implementation steps is generated using a large language model, and the preliminary solution framework is sent to the display terminal.

[0020] Furthermore, methods for removing background noise from the original speech data through time-frequency masking include: Raw voice data in an educational consultation scenario is received via a microphone array or communication interface; the raw voice data is then processed by frame segmentation and windowing to obtain several voice frames; the frame segmentation and windowing process employs an adaptive windowing method to balance the requirements of temporal and frequency domain resolution; specifically, it includes the following sub-steps: Valid speech segments are identified in the original speech data using a short-time energy-based endpoint detection algorithm. The identified speech segments are pre-framed using a preset initial window length (e.g., 300ms) to obtain a series of pre-framed units for subsequent analysis.

[0021] For each pre-framing unit, a set of acoustic features capable of distinguishing between voiced and unvoiced sounds are extracted, the acoustic features including: Zero-crossing rate: Calculated by counting the number of times a signal crosses zero. The zero-crossing rate of unvoiced consonants (such as the consonants / f / and / s / ) is typically significantly higher than that of voiced consonants (such as vowels). The formula for calculating the zero-crossing rate is: In the formula, This represents the zero-crossing rate of the signal in the i-th pre-frame unit; Indicates the frame length (number of sampling points) of the pre-framing unit; i represents the sequence number of the pre-framing unit; This represents the amplitude of the nth sampling point in the i-th pre-framing unit; This represents the amplitude of the (n-1)th sampling point in the i-th pre-framing unit; This is a sign function. When the value is greater than or equal to zero, the value is 1; when the value is less than zero, the value is -1.

[0022] Short-time energy: Calculated by summing the squares of the signal amplitudes within a frame; voiced sounds typically have significantly higher energy than unvoiced sounds. Spectral centroid: used to reflect the central position of the spectral energy distribution. The energy of unvoiced sounds is concentrated in the high frequency, so the spectral centroid is relatively high. The spectral centroid is calculated by first performing a Discrete Fourier Transform (DFT) on the pre-framing unit to obtain the frequency domain amplitude, and then calculating the spectral centroid based on the frequency domain amplitude.

[0023] The peak value of the autocorrelation function reflects the periodicity of the signal. Voiced sounds exhibit obvious periodicity; the autocorrelation function shows a significant peak value at the period of the voiced sound. Unvoiced sounds, on the other hand, resemble random noise and have no obvious peak value. The peak value of the autocorrelation function is calculated as follows: Calculate the short-time autocorrelation function of the i-th pre-segmented frame unit. The short-time autocorrelation function is expressed as follows: In the formula, This represents the delay, ranging from 100 to 200, and the maximum value within this range is taken as the peak value of the autocorrelation function. The extracted acoustic features are normalized and combined into a feature vector, which is then input into a pre-trained voiced / unvoiced sound classification model. This model is a binary classifier based on Support Vector Machine (SVM), and its construction method is as follows: A training set containing a large number of speech samples is constructed. The four acoustic features (zero-crossing rate, short-time energy, spectral centroid, and autocorrelation function peak) of each frame of speech in the samples are extracted to form the input feature vector. Based on the phoneme alignment information of phonetics, the samples corresponding to vowels and voiced consonants are labeled as voiced consonants, and the samples corresponding to unvoiced consonants and silences are labeled as unvoiced consonants. The labeled feature vector data is used to supervise the training of the support vector machine (SVM) until the model converges. After the training is completed, the model parameters are saved, thus obtaining a classification model that can automatically determine the voiced / unvoiced attributes based on acoustic features.

[0024] The obtained speech frames are subjected to a short-time Fourier transform; specifically, each frame signal is subjected to a fast Fourier transform to convert it from the time domain to the frequency domain, resulting in a complex spectrum containing real and imaginary parts; the complex spectrum preserves the phase information of the speech signal.

[0025] The obtained complex spectrum is used as input features and fed into a pre-trained deep complex convolutional network. The deep complex convolutional network employs a U-Net architecture of encoder-bottleneck-decoder. The encoder, constructed from multiple complex convolutional layers (ComplexConv2d), performs downsampling to extract high-level abstract features of the speech signal in the time-frequency domain (such as formant structure and harmonic texture). The bottleneck layer, located in the middle of the network, captures the deepest contextual information. The decoder, composed of multiple complex transposed convolutional layers (ComplexTransposedConv2d), performs upsampling. Each layer of the decoder is concatenated with the output of the corresponding layer of the encoder via skip connections, thus fusing shallow detail features with deep semantic features to ultimately reconstruct a time-frequency mask matrix of the same size as the input spectrogram. This mask matrix is ​​a complex matrix containing amplitude and phase mask information, representing the ratio of speech components to noise components at each time-frequency point.

[0026] The training process for deep complex convolutional networks is as follows: First, a targeted training dataset is constructed by collecting a large number of real-world speech recordings from educational consultation scenarios. The original recordings are pre-processed to remove noise and obtain clean speech samples. Then, various common noise interferences, such as air conditioner noise, keyboard typing, and distant ambient noise, are selected and manually mixed with these noises at different signal-to-noise ratios (SNR) to generate noisy speech data covering multiple noise scenarios and SNR conditions, ensuring the diversity and practicality of the dataset. Next, the model optimization objective is determined. Scale-invariant signal-to-noise ratio (SI-SNR) is used as the loss function. This function quantifies the difference between the output of the noisy speech data after model processing and the target clean speech data, improving the model's robustness to changes in speech waveform amplitude. Finally, model parameter optimization is performed. Based on the above loss function, the gradients of the network layer parameters are calculated. An error backpropagation algorithm (such as the Adam optimizer) is used to iteratively update the network weight parameters to continuously reduce the loss value until the model performance stabilizes and the network converges, completing the entire training process.

[0027] The time-frequency mask matrix is ​​multiplied point-by-point with the original noisy complex spectrum to suppress the amplitude of the noise band and correct the phase, thus obtaining the enhanced complex spectrum. An inverse short-time Fourier transform is then performed on the enhanced complex spectrum to convert the signal back from the frequency domain to the time domain. During the transformation, an overlapping addition method is used to handle the overlapping parts between frames to eliminate the discontinuities caused by frame splitting, ultimately outputting denoised speech data with a continuous waveform and a clean background.

[0028] The aforementioned denoising method directly processes the complex spectrum using a deep complex convolutional network, effectively preserving the phase information of the speech signal and avoiding phase distortion and artifacts caused by traditional amplitude spectrum enhancement methods. The adaptive windowing mechanism based on acoustic features uses a short window for unvoiced sounds to capture transient changes and a long window for voiced sounds to preserve harmonic structure, resolving the contradiction that fixed window length cannot simultaneously achieve time-frequency resolution and optimizing feature representation from the signal source. Training with scale-invariant signal-to-noise ratio as the loss function achieves accurate suppression of complex background noise in the consultation room environment, thereby significantly improving the accuracy of subsequent speech recognition.

[0029] Furthermore, the methods for generating the initial text unit sequence include: The denoised speech data processed in the previous steps is transcribed into text using an Automatic Speech Recognition (ASR) model to obtain transcribed text data. The transcribed text data includes character sequences and corresponding character codes and timestamps. The character sequence is the recognized complete string (e.g., "Inquiry about the IB course..."), the character code is the computer code (such as UTF-8 or Unicode) corresponding to each character, used to distinguish Chinese characters from ASCII characters (English), and the timestamp is the start and end time (accurate to milliseconds) of each character in the original speech stream.

[0030] The character encoding is traversed, and the specific positions of ASCII character segments and Chinese character segments are determined by detecting the encoding attributes of adjacent characters. When it is detected that the current character and the next character belong to different types (i.e., the junction of "Chinese-ASCII" or "ASCII-Chinese"), a segmentation mark is added at the junction (e.g., ...). <sep>(Symbols). Based on the segmentation markers, the character sequence is segmented to generate a pre-segmented sequence. This operation can effectively avoid the adhesion of English terms and Chinese auxiliary words, and ensure that the boundaries of different language types of segments are clear. For example, when a user expresses "My GPA is 3.5", after the above steps, the pre-segmented result of "my" "GPA" "is 3.5" can be obtained.

[0031] Based on the pre-segmented sequence, a pre-built terminology index (stored as a Trie tree or AC automaton structure) is invoked. This terminology index is generated based on a knowledge graph in the education field and contains information such as the full names of all institutions (e.g., UCLA), abbreviations (e.g., UCLA), professional terms (e.g., Computer Science), and exam names. The terminology index is used to perform a forward scan of the pre-segmented sequence, following the maximum length matching principle. If a string in the pre-segmented sequence is detected to completely match a professional term in the terminology index, that string is locked as an indivisible atomic unit. The timestamp of the atomic unit is the time from the start time of its first character to the end time of its last character, ensuring the integrity of the terminology in the time dimension. For example, when scanning "...University of California, Los Angeles...", if "University of California, Los Angeles" is detected as an entry in the terminology index, it is locked as an atomic unit.

[0032] Treating each atomic unit as a character, the processed sequence (including individual characters and atomic units) is traversed, and the timestamp difference between adjacent characters (i.e., the difference between the end time of the preceding character and the start time of the following character) is calculated. A silence threshold is preset (e.g., 500ms, which can be dynamically adjusted according to the actual speech rate). If the calculated timestamp difference exceeds the preset silence threshold, it indicates that the speaker has made a significant pause at this point, signifying a semantic boundary transition. In this case, segmentation is performed between the adjacent characters. After the above steps, a clear initial text unit sequence with well-defined semantic boundaries is obtained, effectively solving the segmentation problem caused by the mixing of Chinese and English, numerous terminologies, and non-fluency of spoken language in educational consultation scenarios.

[0033] Furthermore, methods for identifying and replacing transcribed terms in the initial text unit sequence include: Perform part-of-speech tagging on the initial text unit sequence using a part-of-speech tagging model (such as a Chinese part-of-speech tagging model based on the BiLSTM-CRF architecture), classify the characters in each text unit according to their grammatical attributes, extract the text segments with noun attributes, and determine such text segments as the to-be-detected word units. By part-of-speech screening, the error detection range can be narrowed and the processing efficiency can be improved; for example, for the initial text unit sequence "我想了解雅斯考试的报名时间” (I want to know the registration time of the IELTS exam), after part-of-speech tagging, "雅斯” (IELTS), "考试” (exam), and "报名时间” (registration time) are extracted as the to-be-detected word units.

[0034] Based on the to-be-detected word units, perform fuzzy retrieval in a pre-constructed term index library. During the retrieval process, multiple standard candidate terms with similar pronunciations can be obtained; for example, for the to-be-detected word unit "雅斯” (IELTS), after fuzzy retrieval, standard candidate terms such as "雅思” (IELTS) and "亚斯” (Yas) can be obtained; for the to-be-detected word unit "托幅” (TOEFL), standard candidate terms such as "托福” (TOEFL) and "托业” (TOEIC) can be obtained.

[0035] Obtain the to-be-detected pinyin sequence of the to-be-detected word unit and the term pinyin sequence of the standard candidate term, where the generation of the pinyin sequence needs to include initials, finals, and tone information; use the dynamic programming algorithm to perform sequence alignment processing on the to-be-detected pinyin sequence and the term pinyin sequence, and extract their longest common subsequence; calculate the ratio of the length of the longest common subsequence to the length of the term pinyin sequence, and record this ratio as the speech matching degree. This ratio intuitively reflects the pronunciation similarity between the to-be-detected word unit and the standard candidate term. For example, the length of the longest common subsequence of the pinyin sequence "tuō-fú” of the to-be-detected word unit "托幅” (TOEFL) and the pinyin sequence "tuō-fú” of the standard candidate term "托福” (TOEFL) is 2, the length of the term pinyin sequence is 2, and the speech matching degree is 1.0, while the length of the longest common subsequence with "托业” (TOEIC) (pinyin sequence "tuō-yè”) is 1, and the speech matching degree is 0.5.

[0036] Set a confidence threshold (such as 0.8, which can be dynamically adjusted according to the pronunciation characteristics of terms in the education field), and compare the calculated speech matching degree with this threshold; when the speech matching degree exceeds the preset confidence threshold and the text of the to-be-detected word unit is inconsistent with the standard candidate term, it is determined as a transcription error; at this time, select the term with the highest speech matching degree from multiple standard candidate terms as the replacement term to replace this to-be-detected word unit. After all to-be-detected word units are processed, preliminary text data is generated. For example, the speech matching degree between the to-be-detected word unit "雅斯” (IELTS) and the standard candidate term "雅思” (IELTS) is 1.0 (exceeding the preset confidence threshold of 0.8), and the text is inconsistent, so "雅斯” (IELTS) is replaced with "雅思” (IELTS); the speech matching degree between the to-be-detected word unit "托幅” (TOEFL) and "托福” (TOEFL) is 1.0, and after replacement, "托福” (TOEFL) is obtained. The finally corrected preliminary text data is "我想了解雅思考试的报名时间” (I want to know the registration time of the IELTS exam).

[0037] Through the above process, transcription errors of terms caused by similar pronunciations in educational consultation scenarios can be accurately identified; accurate replacement based on the similarity matching of pinyin sequences can improve the accuracy of text data and provide high-quality input data for subsequent semantic processing steps such as referential resolution and omission completion.

[0038] Furthermore, by using context-based memory-based referential resolution and dependency parsing-based ellipsis completion, the sentences in the preliminary text data are processed to obtain normalized text, including: The core of context-based referential resolution is the construction and maintenance of an entity cache stack. This entity cache stack is a data structure that follows the Last-In-First-Out (LIFO) principle. Each stack element records the named entities that appeared in the historical dialogue and their type attributes, such as PERSON (person), ORGANIZATION (institution / school), COURSE (course), and other common types in educational consulting scenarios. The entity cache stack operates as follows: the most recently mentioned named entity is pushed onto the top of the stack, and the earliest mentioned named entity is at the bottom. When the stack is full, the named entity at the bottom is removed, thus simulating the human tendency to prioritize recent information in conversation. From the current sentence in the initial text data, newly emerging named entities (such as technical terms) are identified using a named entity recognition tool. At the same time, all demonstrative pronouns to be resolved (such as "he", "she", "it", "this scheme") are identified. Dependency parsing is then performed on the current sentence to construct a syntactic tree. Then, the governing word that governs each demonstrative pronoun is found. This governing word is usually a predicate verb. For example, in the sentence "We should adopt it", the verb "adopt" is the governing word of the demonstrative pronoun "it".

[0039] Traverse each historical named entity in the entity cache stack and calculate the semantic compatibility score between the historical named entity and the current demonstrative pronoun and its governing word. The calculation process for this score is as follows: Based on a pre-built verb argument collocation dictionary, the semantic preference of the current governing word for its argument components (subject or object) is queried. This verb argument collocation dictionary records the expected argument types of verbs in educational consultation scenarios. For example, the semantic preference of the object of the verb "admit" is {student, person}, and the semantic preference of the object of the verb "read" is {book, document}. Then, the matching degree between the semantic preference and the type attribute of the historical named entity is calculated. If the type of the historical entity belongs to the semantic preference set of the governing word, the matching factor is set to 1.0; otherwise, it is set to 0.1 (a lower penalty value). Then, the difference Δt between the number of dialogue turns between the historical named entity and the current statement is calculated to obtain the time decay factor. λ is the decay coefficient (preferably between 0.6 and 0.8). The larger Δt is, the smaller the time decay factor is, thus reflecting the forgetting pattern of named entities over time. Finally, the matching factor and the time decay factor are multiplied to obtain the final semantic compatibility score, i.e., semantic compatibility score = matching factor × time decay factor.

[0040] The semantic compatibility scores of all historical named entities in the entity cache stack are compared. The historical named entity with the highest score is selected to replace the demonstrative pronoun in the current statement. At the same time, the newly identified current named entity and its type attribute are pushed onto the top of the entity cache stack, and the context memory resources are updated. For example, if the current statement is "How much does it cost to register?", the entity cache stack stores two historical entities: "IELTS course (course type)" and "TOEFL training (course type)". The semantic preference of the dominant word "is" is "course / service type". If the number of turns for "IELTS course" is 1, the time decay factor is 0.8, the matching factor is 1.0, and the semantic compatibility score is 0.8. If the number of turns for "TOEFL training" is 3, the matching factor is 1.0, and the semantic compatibility score is 0.512, then "IELTS course" is selected to replace "it", and the corrected statement is "How much does it cost to register for an IELTS course?".

[0041] After resolving the pronoun reference, dependency parsing is performed on the processed current statement to check for missing core arguments. Core arguments include the subject and object, judged from two dimensions: missing elements and verb type. From the missing elements dimension, if the predicate verb lacks a dependency node with a nominal subject relationship, it is considered a missing subject; if a transitive verb lacks a dependency node with a direct object relationship, it is considered a missing object. That is, if the verb is transitive but lacks a corresponding direct object node in the syntax tree, or if the verb is intransitive but lacks a necessary nominal subject node, it is considered a missing core argument. For example, in educational consulting scenarios, missing core arguments often occur when the subject is omitted (e.g., "I want to know the registration process" omits "I") or the object is omitted (e.g., "I plan to enroll" omits "IELTS course").

[0042] When a core argument is detected as missing, the previous statement adjacent to the current statement (usually the consultant's question or the previous interaction content between both parties) is retrieved from the conversation history of the current session. Dependency syntactic analysis is performed on the previous statement. According to the question-and-answer interaction logic and the syntactic consistency principle, components with the same syntactic role as the missing core argument are extracted. If the current statement is missing a subject and the subject of the previous statement is a user-related expression (such as "What do you want to know"), the user perspective subject logic of the previous statement (such as adding "I") is inherited. If the current statement is missing an object and the previous statement contains an interrogative pronoun corresponding to the missing object (such as "Which course do you want to sign up for" in "Which course"), the entity component in the previous statement that semantically matches the interrogative pronoun is extracted. The extracted component is filled into the missing position of the syntactic tree of the current statement to generate a semantically complete normalized text. For example, the current statement "want to sign up" corresponds to the previous statement "Do you want to sign up for the TOEFL intensive course". After analysis, the object is missing (and the subject is also missing implicitly). The "TOEFL intensive course" in the previous statement is extracted as the object to fill in, and the subject "I" is added. The corrected statement is "I want to sign up for the TOEFL intensive course".

[0043] The above processing provides high-quality input with zero ambiguity for the subsequent large language model, significantly reducing the model inference difficulty, effectively suppressing the hallucination phenomenon caused by context missing or unclear reference, and ensuring the accuracy and professionalism of the educational consulting plan finally generated by the large language model.

[0044] Using the normalized text as input, an input sequence for the large language model (LLM) is constructed in combination with a preset system prompt. The system prompt includes a clear role setting and output format constraint. The role setting focuses on the educational consulting scenario. For example, "You are a senior educational expert with more than 5 years of experience in study abroad planning and domestic升学指导 (it seems there is a wrong Chinese here, assuming it should be something like college entrance guidance) and are good at giving accurate plans by combining user needs, academic performance background and budget". The output format constraint clearly requires the large language model to generate content in three fixed sections: "Problem Diagnosis, Core Suggestions, Implementation Steps", and each section needs to fit the actual needs of the educational consulting scenario. This structured prompt design can guide the large language model to focus on the core task, reduce the generation of irrelevant information, and ensure the adaptability of the output content to the educational consulting scenario.

[0045] The large language model is a general large language model (such as GPT-4, Wenxin Yiyan) or a dedicated large language model fine-tuned with educational domain corpus. The fine-tuning corpus includes educational consulting conversation records, college application guides, exam preparation plans, etc., to improve the model's understanding ability of professional terms and scenario requirements in the educational domain.

[0046] It should be noted that there seems to be an incorrect Chinese expression "升学指导" in the role setting part of ID=6, and it is translated according to the general meaning. You may need to correct it according to the actual situation.The completed input sequence is fed into a large language model, which then extracts the user's core intent, background information (such as grades, budget, and target schools) and constraints to generate a preliminary solution framework containing three core dimensions: problem diagnosis, core recommendations, and implementation steps. The problem diagnosis includes the user's core educational needs and the root cause of the problem; the core recommendations include the key points of the solution to the root cause of the problem; and the implementation steps include decomposing the core recommendations into specific, executable operational steps. For example, if the core need is to pass the postgraduate entrance examination for computer science in China, the real obstacle is a weak English foundation (failed the CET-4), and there are only 8 months left to prepare, lacking a systematic plan. The core advice given is to focus on postgraduate entrance examination preparation, adopt a strategy of reviewing CET-4 / 6 and postgraduate entrance examination English in conjunction, prioritize vocabulary and reading comprehension, and combine this with a tiered review plan for basic computer science courses. The implementation steps are broken down into actionable actions with time nodes, and clear stage goals. From January to March, memorize core postgraduate entrance examination vocabulary, complete CET-4 past papers, and review data structure basics simultaneously. From April to June, focus on postgraduate entrance examination English reading and writing, and master operating system and computer network knowledge points. From July to August, conduct full-set mock exams, identify and fill gaps in knowledge, and prepare professional skills related to the re-examination.

[0047] After generating the initial solution framework, it is encapsulated into a structured data packet (such as JSON format) and sent to the display terminal via network communication. After receiving the data packet, the display terminal parses and displays it according to the preset interface logic. Based on the displayed content, the consultant adds personalized details (such as recommending specific training institutions or reminding users of application precautions) or fine-tunes the solution based on real-time user feedback, thereby improving the efficiency and professionalism of the consulting service and reducing communication costs.

[0048] The above process utilizes the semantic understanding and content generation capabilities of large language models to efficiently transform standardized dialogue data into well-organized, focused, and actionable educational consulting solutions. This achieves intelligent and standardized output of educational consulting services, ensuring the stability of service quality, providing professional support for consultants, and enhancing the overall consulting service experience.

[0049] The above description is merely a preferred embodiment of the present invention and is 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.

[0050] 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, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0051] In the description of this invention, it should be understood that the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.

[0052] In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0053] In the description of this invention, "several" means one or more, and "a large number" means two or more.

[0054] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0055] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0056] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.< / sep>

Claims

1. An education consultation service method based on artificial intelligence technology assistance, characterized by, The method comprises the following steps: Obtain original speech data in an education consultation scenario, remove background noise of the original speech data through time-frequency masking processing, and obtain denoised speech data; Transcribe the denoised speech data into text data, and perform adaptive text segmentation on the transcribed text data based on a pre-constructed term index library to generate an initial text unit sequence; Identify and replace the transcribed error terms in the initial text unit sequence using the pre-constructed term index library to obtain preliminary text data; Process the sentences in the preliminary text data through context memory mechanism-based reference resolution and dependency syntax analysis-based omission completion to obtain standardized text; Based on the standardized text, use a large language model to generate a preliminary scheme framework containing problem diagnosis, core suggestions and implementation steps, and send the preliminary scheme framework to a display terminal.

2. The education counseling service method based on artificial intelligence technology assistance according to claim 1, characterized in that, The method for removing background noise of the original speech data through time-frequency masking processing comprises: Frame and window the original speech data to obtain a plurality of speech frames; Convert the speech frames to the frequency domain through short-time Fourier transform to obtain complex spectra; Input the complex spectra into a pre-trained deep complex convolutional network to generate a time-frequency mask matrix; Multiply the time-frequency mask matrix and the complex spectra point by point to obtain enhanced complex spectra; Perform inverse short-time Fourier transform and overlap-add processing on the enhanced complex spectra to obtain denoised speech data.

3. The education counseling service method based on artificial intelligence technology assistance according to claim 2, characterized in that, The frame and window processing adopts an adaptive windowing method, which comprises: Identify speech segments in the original speech data and perform voice judgment within the speech segments; wherein the voice judgment method is: Pre-frame the speech segments using a preset initial window length to obtain pre-framing units; Extract acoustic features of the pre-framing units, including signal zero-crossing rate, short-time energy, spectral centroid and autocorrelation function peak value; Combine the extracted acoustic features into a feature vector and input it into a pre-trained voice classification model; Determine the voice attribute of the pre-framing unit according to the output of the voice classification model; When it is determined to be a clear voice, re-frame the speech segment using a first window length; When it is determined to be a voiced voice, re-frame the speech segment using a second window length; Wherein the first window length is shorter than the second window length.

4. The education counseling service method based on artificial intelligence technology assistance according to claim 3, characterized in that, The training method of the deep complex convolutional network comprises: Construct a training data set containing multiple groups of noisy speech data collected in the education consultation scenario and their corresponding clean speech data; Use the scale-invariant signal-to-noise ratio between the denoised speech data output in the forward propagation process and the clean speech data as a loss function, and update the network weight parameters through an error back propagation algorithm; The deep complex convolutional network adopts a U-Net structure of encoder-bottleneck layer-decoder, wherein the encoder is used for downsampling to extract high-level abstract features, and the decoder is used for upsampling to restore high-resolution time-frequency masks.

5. The education counseling service method based on artificial intelligence technology assistance according to claim 4, characterized in that, The method for generating an initial text unit sequence comprises: Transcribe the denoised voice data into text data by using an automatic speech recognition model, wherein the text data comprises a character sequence, character codes corresponding to the character sequence, and time stamps; Iterate through the character codes to determine the positions of ASCII character segments and Chinese character segments, and add a split marker at the junction of the two types of characters; Split the character sequence based on the split marker to generate a pre-split sequence; Scan the pre-split sequence forward based on the pre-constructed term index library, and lock the hit professional terms as indivisible atomic units based on the maximum length matching principle, wherein the time stamp of the atomic unit is defined as the start time of the first character to the end time of the last character contained in the atomic unit; Treat the atomic unit as a character, calculate the time stamp difference between adjacent characters, and if the time stamp difference exceeds a preset silence threshold, split the adjacent characters to obtain an initial text unit sequence.

6. The education counseling service method based on artificial intelligence technology assistance according to claim 5, characterized in that, The way of identifying and replacing the transcription error terms in the initial text unit sequence includes: Part-of-speech tagging is performed on the initial text unit sequence to extract text segments with noun attributes, and the text segments are used as word units to be checked; Based on the word units to be checked, fuzzy search is performed in the pre-constructed term index library to obtain a plurality of standard candidate terms with similar pronunciation; Obtain the pinyin sequence of the word unit to be checked and the term pinyin sequence of the standard candidate term; Perform sequence alignment processing on the pinyin sequence to be checked and the term pinyin sequence to extract the longest common subsequence thereof; Calculate the ratio of the length of the longest common subsequence to the length of the term pinyin sequence, and record this ratio as the voice matching degree; When the voice matching degree exceeds a preset confidence threshold and the text of the word unit to be checked is inconsistent with the standard candidate term, it is determined that there is a transcription error, and the standard candidate term with the highest voice matching degree is used to replace the word unit to be checked to generate preliminary text data.

7. The education counseling service method based on artificial intelligence technology assistance according to claim 6, characterized in that, The reference resolution based on the context memory mechanism includes the following steps: An entity cache stack following the last-in-first-out principle is constructed and maintained, which is used to store the named entities and their type attributes appearing in the historical dialogue; The current named entity and the demonstrative pronoun are identified from the preliminary text data; The governing word governing the demonstrative pronoun is extracted based on the dependency syntax analysis; The semantic compatibility score between each historical named entity in the entity cache stack and the current demonstrative pronoun and its governing word is calculated, and the semantic compatibility score is calculated based on the semantic preference of the governing word; The historical named entity with the highest semantic compatibility score is selected to replace the demonstrative pronoun, and the current named entity and its type attribute are stored in the entity cache stack.

8. The education counseling service method based on artificial intelligence technology assistance according to claim 7, characterized in that, The semantic compatibility score is calculated by the following method: Based on the pre-constructed verb argument collocation dictionary, the semantic preference of the argument component of the governing word is obtained; Based on the matching degree of the semantic preference and the type attribute of the historical named entity, a matching factor is obtained; Based on the difference in the number of dialogue rounds between the historical named entity and the current sentence, a time decay factor is obtained; The product of the matching factor and the time decay factor is calculated as the semantic compatibility score.

9. The education counseling service method based on artificial intelligence technology assistance according to claim 8, characterized in that, The omission completion based on the dependency syntax analysis comprises the following steps: dependency syntax analysis is performed on the current sentence to check whether there is a missing core argument; if there is a missing core argument, the last round of sentence adjacent to the current sentence is obtained from the dialogue history of the current session; dependency syntax analysis is performed on the last round of sentence to extract components having the same syntactic role as the missing core argument according to the question and answer interaction logic and the syntactic consistency principle; the extracted components are filled into the missing position of the current sentence to obtain the normalized text.

10. The education consultation service method based on artificial intelligence technology assistance according to claim 9, characterized in that, The content of the problem diagnosis includes the core education needs and the problem root of the user; the content of the core suggestion includes the core points of the solution to the problem root; and the content of the implementation step includes decomposing the core suggestion into executable specific operation steps.

Images