Method and system for structured generation of interview feedback based on natural language processing

Abstract

The application provides an interview feedback structuring generation method and system based on natural language processing, which comprises the following steps: obtaining an interview dialogue text sequence set with a speaking timestamp marker, containing interviewee question and answer text segments; processing the text sequence by a dialogue flow state analyzer to output a dialogue behavior state transition sequence containing dialogue stage attribution labels and state transition path identifiers; calling a feedback mapping activation network to jointly process the state transition sequence and the text sequence to generate an initial feedback item mapping set carrying activation intensity markers and feedback category guide markers; using a feedback item hierarchical growth network to perform multi-level clustering and branch growth on the feedback items to construct a feedback item hierarchical structure tree from a root topic to specific statement items; and rendering the document according to the hierarchical structure tree to output a structured interview feedback document conforming to a preset format specification. The method realizes the full-process automatic generation of interview evaluation from unstructured dialogue to standardized feedback document.

Description

Technical Field

[0001] This application relates to the field of natural language processing, and in particular to a method and system for generating structured interview feedback based on natural language processing. Background Technology

[0002] With the widespread application of online recruitment and digital interviews, automatically generating structured feedback reports based on interview dialogue content has become an important requirement for improving recruitment efficiency. Structured interview feedback generation methods refer to the technical means of analyzing and processing interview dialogue text and outputting structured feedback text that can be used as a reference for recruitment decisions. Currently, a common approach is to extract relevant fragments from the dialogue text using preset rule templates or keyword matching and fill them into a fixed format to form a feedback document. However, when faced with the unique interactive dynamism and semantic complexity of interview dialogues, the generated feedback content often presents a fragmented distribution, making it difficult to form a complete feedback system with logical hierarchy and internal coherence. Furthermore, the feedback statements are prone to deviating from the actual context of different stages of the interview process, resulting in insufficient professionalism and systematicity in the final output feedback document. Summary of the Invention

[0003] This invention provides a method and system for generating structured interview feedback based on natural language processing.

[0004] In a first aspect, embodiments of the present invention provide a method for generating structured interview feedback based on natural language processing, the method comprising: Obtain a set of interview dialogue text sequences with speaking timestamps after the interview dialogue audio stream has been converted by speech recognition. The set of interview dialogue text sequences contains interviewer question text fragments and applicant response text fragments with a sequential relationship. The set of interview dialogue text sequences is input into a pre-built dialogue flow state parser to perform dialogue behavior state transition deduction processing, and outputs a dialogue behavior state transition sequence corresponding to each timestamp node in the set of interview dialogue text sequences. The dialogue behavior state transition sequence includes the dialogue stage attribution label at each timestamp node and the dialogue state transition path identifier between adjacent timestamp nodes. The preset feedback mapping activation network is invoked to perform joint feedback item mapping processing on the dialogue behavior state transition sequence and the interview dialogue text sequence set, generating an initial feedback item mapping set corresponding to each semantic unit in the candidate's response text fragment. Each initial feedback item in the initial feedback item mapping set is associated with an activation intensity marker and a feedback category guidance marker in the feedback knowledge space. The initial set of feedback item mappings is input into the feedback item hierarchical growth network for multi-level clustering and branching growth processing, generating a feedback item hierarchical structure tree that starts with the feedback root topic and ends with the specific feedback statement. There is a parent-child relationship between nodes in any adjacent level of the feedback item hierarchical structure tree. The initial feedback item mapping set is document-rendered based on the feedback item hierarchy structure tree, and a structured interview feedback document conforming to the preset document format specifications is output. Each independent feedback paragraph in the structured interview feedback document has a content mapping correspondence with a branch node in the feedback item hierarchy structure tree.

[0005] Secondly, embodiments of the present invention provide a computer system, including: a memory for storing computer-executable instructions or computer programs; and a processor for executing the computer-executable instructions or computer programs stored in the memory to implement the above-described structured generation method for interview feedback based on natural language processing.

[0006] The embodiments of this application have the following beneficial effects: This invention inputs interview dialogue text into a dialogue flow state parser to deduce the dialogue behavior state transitions, outputting a dialogue behavior state transition sequence containing stage affiliation tags and migration path identifiers, explicitly capturing the stage switching patterns and leadership transfer patterns in the interview process. A feedback mapping activation network is invoked to jointly map the state transition sequence and the text sequence, generating an initial feedback item mapping set corresponding to each semantic unit. This establishes a mapping association between the response semantic unit and the standardized feedback knowledge space, marked with activation intensity and category guidance tags, ensuring that the feedback statements are context-appropriate and possess knowledge system relevance. The initial feedback item set is then input into a hierarchical growth network for multi-level clustering and branching growth, generating a hierarchical structure tree from the root topic to specific statements, realizing the transformation from discrete feedback items to a logically hierarchical organizational form. Finally, based on this structure tree, document rendering is performed, outputting a structured feedback document that conforms to the format specifications. Each paragraph in the document has a content mapping correspondence with the branch nodes of the structure tree, ensuring a strict correspondence and traceability between the feedback content and the interview analysis results. This method enables end-to-end automated generation of structured professional feedback documents from unstructured interview dialogues, improving the systematic organization of feedback content, consistency of context matching, and standardization of document generation. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the architecture of the application scenario provided in the embodiments of this application; Figure 2 This is a schematic diagram of the structure of the computer system provided in the embodiments of this application; Figure 3This is a flowchart illustrating the structured generation method for interview feedback based on natural language processing provided in this application embodiment. Detailed Implementation

[0008] 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.

[0009] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0010] See Figure 1 , Figure 1 This is an architectural diagram of an application scenario provided in this application embodiment. The terminal 400 is connected to the computer system 200 through the network 300. The network 300 can be a wide area network or a local area network, or a combination of the two, and data transmission is achieved using wireless or wired links.

[0011] In some embodiments, the computer system 200 may be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The terminal 400 may be a smartphone, tablet, laptop, desktop computer, etc., but is not limited to these. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited in this embodiment.

[0012] The following describes a computer system that implements the natural language processing-based structured interview feedback generation method provided in the embodiments of this application. See also Figure 2 , Figure 2 This is a schematic diagram of the structure of the computer system provided in the embodiments of this application. Figure 2The computer system shown includes at least one processor 210, memory 250, at least one network interface 220, and an external interface 230. The various components in the computer system 200 are coupled together via a bus system 240. It is understood that the bus system 240 is used to implement communication between these components. In addition to a data bus, the bus system 240 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 2 The general labeled all buses as Bus System 240.

[0013] Processor 210 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0014] External interface 230 may include, for example, one or more speakers and / or one or more visual displays. External interface 230 may also include one or more input devices 232, such as a keyboard, mouse, microphone, touch screen display, camera, etc.

[0015] The memory 250 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. The memory 250 may optionally include one or more storage devices physically located away from the processor 210.

[0016] The memory 250 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 250 described in this application embodiment is intended to include any suitable type of memory.

[0017] In some embodiments, memory 250 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0018] Operating system 251 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks; The network communication module 252 is used to reach the terminal via one or more (wired or wireless) network interfaces 220, exemplary network interfaces 220 include: Bluetooth, WiFi, and Universal Serial Bus (USB), etc. Presentation module 253 is configured to enable the display of information (e.g., external interface for operating peripheral devices and displaying content and information) via one or more output devices 231 (e.g., display screen, speaker, etc.) associated with external interface 230; The input processing module 254 is used to detect and translate one or more user inputs or interactions from one or more input devices 232.

[0019] Based on the above description of the application scenarios and computer systems provided in the embodiments of this application, the following describes the structured interview feedback generation method based on natural language processing provided in the embodiments of this application. In actual implementation, the structured interview feedback generation method based on natural language processing provided in the embodiments of this application can be implemented by a computer system. See also Figure 3 , Figure 3 This is a flowchart illustrating the structured interview feedback generation method based on natural language processing provided in this application embodiment. Next, we will combine... Figure 3 The steps shown are explained.

[0020] Step S100: Obtain a set of interview dialogue text sequences with speaking timestamps after the interview dialogue voice stream has been converted by speech recognition. The set of interview dialogue text sequences includes interviewer question text fragments and applicant response text fragments with a sequential relationship.

[0021] The acquisition of interview dialogue audio streams can be accomplished using various audio capture devices. For example, in online video interview scenarios, the real-time audio and video communication channel established between the interviewer's terminal and the applicant's terminal can simultaneously provide two independent audio streams: one carrying the interviewer's voice input and the other carrying the applicant's voice input. The two audio streams can be mixed and stored on the server or local client, or stored separately. As another example, in offline interview scenarios, a microphone array or recording device deployed in the interview room can record the entire interview in single-track or multi-track format, generating a mixed audio file containing the statements of both interviewers. The embodiments of this application do not specifically limit the source, sampling rate, encoding format, or channel configuration of the interview dialogue audio stream.

[0022] After acquiring the interview dialogue audio stream, it is input into a pre-defined automatic speech recognition engine for speech-to-text conversion. The specific automatic speech recognition engine can employ an end-to-end speech recognition model based on deep neural networks. For example, the automatic speech recognition engine can use a Transformer-Transducer architecture trained based on connection-time classification criteria. This architecture consists of an audio encoder, a label encoder, and a joint network. The audio encoder receives log-Mel filter bank features as input and extracts acoustic hidden layer representations through a multi-layer self-attention mechanism. The label encoder models the output text label sequence at the language model level. The joint network fuses the outputs of the audio encoder and the label encoder, and then outputs the vocabulary probability distribution corresponding to the current time step via a Softmax layer. During frame-by-frame decoding of the interview dialogue audio stream, the automatic speech recognition engine synchronously generates start and end timestamps corresponding to each word or sub-word unit in the recognized text. The precision of the timestamps can be set to the millisecond level.

[0023] After the automatic speech recognition engine completes the speech recognition and conversion, the output interview dialogue text sequence set consists of one or more time-ordered text record sequences. Each text record contains at least three fields: a speech content field, used to store the recognized text segment; a speech start timestamp field, used to mark the start time of the corresponding speech stream; and a speech end timestamp field, used to mark the end time of the corresponding speech stream. The interview dialogue text sequence set is organized according to the order of the speech start timestamps, thus naturally preserving the alternating round relationship between the interviewer's questions and the candidate's responses during the interview process. Furthermore, the interview dialogue text sequence set can also use a preset speaker separation module to perform cluster analysis on different voiceprint features in the mixed audio stream, assigning speaker identity labels to each text segment, such as "Speaker A" or "Speaker B," and distinguishing the text segments into interviewer question text segments and candidate response text segments based on the speaker identity labels. The speaker separation module can be implemented using a speaker clustering algorithm based on probabilistic linear discriminant analysis. First, Mel-frequency cepstral coefficients or deep speaker embedding vectors are extracted from the audio stream as voiceprint features. Then, a bottom-up agglomerative hierarchical clustering method is used to merge speech segments belonging to the same speaker into the same cluster, ultimately generating speech segment boundary information with speaker identifiers. Through the above processing, step S100 outputs a set of interview dialogue text sequences with speaking timestamps, explicitly including interviewer question text segments and applicant response text segments with a chronological relationship.

[0024] Step S200: Input the set of interview dialogue text sequences into the pre-built dialogue flow state parser to perform dialogue behavior state transition deduction processing, and output the dialogue behavior state transition sequence corresponding to each timestamp node in the set of interview dialogue text sequences. The dialogue behavior state transition sequence includes the dialogue stage attribution label at each timestamp node and the dialogue state transition path identifier between adjacent timestamp nodes.

[0025] In one implementation, step S200 may specifically include the following steps S210 to S260: Step S210: Obtain the interviewer's question text fragments and the applicant's response text fragments from the interview dialogue text sequence set, and construct a round-pairing relationship mapping table between the interviewer's question text fragments and the applicant's response text fragments based on the speaking timestamp.

[0026] Specifically, after receiving the set of interview dialogue text sequences, the dialogue flow state parser first separates and extracts each text fragment according to the speaker's identity label. All text fragments labeled as "interviewer" are grouped into the "interviewer question text fragment set," and all text fragments labeled as "applicant" are grouped into the "applicant response text fragment set." Subsequently, based on the start and end timestamps of each text fragment, the parser performs alignment matching on the timeline, pairing and binding time-adjacent "interviewer question text fragments" with "applicant response text fragments" to form several rounds of dialogue pairs. The specific data structure of the round pairing mapping table can be implemented as a key-value pair mapping table, where the key field stores the round sequence identifier, and the value field stores a composite data structure containing information such as the content of the interviewer question text fragment, the content of the applicant response text fragment, the question start timestamp, and the response start timestamp. If, during an interview, the interviewer asks questions continuously without giving the candidate a chance to respond, or if the candidate's response spans multiple speech segments in a single response, the round-pairing logic inside the dialogue flow state parser will merge or segment the text segments based on a preset silence interval threshold and semantic integrity judgment rules, ensuring that each round-pair dialogue contains a relatively complete question semantic unit and a relatively complete response semantic unit.

[0027] Step S220: Input the text fragments of the interviewer's questions and the text fragments of the applicant's responses with the round-pairing relationship mapping table into the text understanding unit of the dialogue flow state parser. Perform joint context semantic absorption processing on each pair of text fragments with the round-pairing relationship mapping table to generate a round-pairing dialogue representation vector carrying bidirectional dialogue context awareness.

[0028] In one implementation, step S220 may specifically include the following steps S221 to S226: Step S221: Perform word segmentation on the interviewer's question text fragment with round pairing relationship mapping table to obtain the question word sequence, and inject the relative position indicator of each word in the question word sequence into the interviewer's question text fragment to generate the initial word representation set of the question side.

[0029] The text understanding unit first invokes a pre-built lexical analyzer to perform word segmentation on the interviewer's question text fragment. The lexical analyzer can use a byte-pair encoding-based sub-word segmentation algorithm or a dictionary-based maximum matching word segmentation algorithm to segment the continuous text string into a discrete sequence of word units. Taking the interviewer's question text fragment "Please describe the most challenging project you were responsible for at your previous company" as an example, the question word sequence obtained after word segmentation might be: ["please", "describe", "a", "you", "at", "previous", "company", "responsible for", "of", "most", "with", "challenging", "of", "project"]. After obtaining the question word sequence, the text understanding unit assigns a relative position indicator to each word in the sequence. The relative position indicator is used to identify the word's position within the interviewer's question text fragment. For example, the position indicator can be calculated using a sine-cosine position encoding formula. Subsequently, the text understanding unit obtains the word embedding vector corresponding to each word through a lookup table, and adds the word embedding vector to the relative position indicator vector element-wise to generate an initial word representation set for the question side that integrates word semantic information and position information. The initial vocabulary on the question side indicates that each element in the set is a real vector of a fixed dimension, which can be set to 512 or 768 dimensions.

[0030] Step S222: Perform word segmentation on the applicant's response text fragment with round pairing relationship mapping table to obtain response word sequence, and inject the relative position indicator of each word in the response word sequence within the applicant's response text fragment to generate the initial word representation set of the response side.

[0031] Corresponding to the processing flow in step S221, the text understanding unit also performs word segmentation and positional information injection operations on the candidate's response text fragment. Taking the candidate's response text fragment as "I previously led a project at Company A to build a recommendation system from scratch, and the biggest challenge was the data sparsity problem during the cold start phase," the response word sequence after word segmentation will be assigned relative position indicators in sequence and fused with its corresponding word embedding vectors to generate the initial word representation set for the response side. The initial word representation set for the response side and the initial word representation set for the question side share the same set of word embedding matrices and positional encoding functions to ensure that the word representations of both sides are in the same vector semantic space.

[0032] Step S223: Input the initial vocabulary representation set of the question side into the first directional encoding subunit of the text understanding unit, and perform sequential semantic absorption processing along the direction from the start position of the question to the end position of the question to generate the forward semantic absorption result sequence of the question side.

[0033] The first directional encoding subunit is structurally a stacked unidirectional Transformer encoder layer structure that processes the input sequence along the forward temporal direction. Internally, it consists of alternating stacked multi-head self-attention sublayers and feedforward fully connected sublayers, with residual connections and layer normalization operations between layers. The first directional encoding subunit receives the initial set of lexical representations from the questioner as the input sequence. In each encoder layer, the multi-head self-attention sublayer first calculates the attention weight distribution for each pair of lexical representations at all positions in the input sequence. To ensure temporal causality, the multi-head self-attention sublayer applies an upper triangular mask matrix, ensuring that each lexical position can only focus on its own representation and the lexical representations of its preceding positions, but not the lexical information of its following positions. Through progressive abstraction and context integration via multi-layer sequential encoding, the first directional encoding subunit finally outputs the forward semantic absorption result sequence from the questioner. Each position vector in this sequence incorporates accumulated semantic information from the question's starting position to the current position, fully reflecting the layered logical progression of the interviewer's questioning intent.

[0034] Step S224: Input the initial vocabulary representation set of the response side into the second direction encoding subunit of the text understanding unit, and perform reverse semantic absorption processing along the direction from the end position of the response to the start position of the response to generate the backward semantic absorption result sequence of the response side.

[0035] The network architecture of the second-direction encoding subunit is similar to that of the first-direction encoding subunit, also employing a Transformer-based unidirectional encoder layer stacking structure. However, it forms a mirror image of the first-direction encoding subunit in the processing direction. After receiving the initial vocabulary representation set from the response side, the second-direction encoding subunit first reverses the input sequence, making words that were originally at the end of the response text appear at the beginning of the sequence, and words that were originally at the beginning of the response text appear at the end of the sequence. Subsequently, the second-direction encoding subunit performs the same forward sequential encoding processing as in step S223 on the reversed sequence, gradually absorbing contextual information in each layer through a multi-head self-attention mechanism with causal masks. After completing multi-layer encoding, the second-direction encoding subunit reverses the output sequence again, restoring the original vocabulary order, to obtain the response side backward semantic absorption result sequence. Because it adopts a reverse absorption direction from the end of the response to the beginning of the response, each word in the backward semantic absorption result sequence of the response side will focus on aggregating the semantic information after its position until the end of the response. This processing method helps to highlight the concluding statements and summary opinions given by the applicant at the end of the response, and makes it easier to focus on the part of the response that answers the key points of the question when matching the semantics with the question side later.

[0036] Step S225: Input the forward semantic absorption result sequence of the question side and the backward semantic absorption result sequence of the response side into the interactive perception subunit of the text understanding unit, perform cross-sequence semantic association perception processing, and generate a semantic association response intensity distribution map between question words and response words.

[0037] In one implementation, step S225 may specifically include the following steps S2251 to S2256: Step S2251: Perform spatial mapping transformation on the query-side forward semantic absorption result sequence, mapping each element in the query-side forward semantic absorption result sequence to the query representation space, and generating the query-side query representation sequence.

[0038] In this step, the interaction-aware subunit first performs a linear space mapping transformation on the query-side forward semantic absorption result sequence. Specifically, the interaction-aware subunit internally maintains a trainable query mapping matrix W. Q The size of the matrix is ​​d model ×d q , where d model d represents the original dimension of each vector in the sequence of forward semantic absorption results from the query side. q This represents the spatial dimension of the mapped query representation. For the i-th element h in the query-side forward semantic absorption result sequence... i q Through matrix multiplication operation q i =hi q ·W Q The element is linearly projected from the original semantic space to the query representation space to obtain the corresponding query representation vector q on the query side. i After mapping each element in the sequence one by one, a query-side representation sequence Q={q1,q2,...,q} is formed. m}, where m is the length of the sequence of words to be asked.

[0039] Step S2252: Perform spatial mapping transformation on the response-side backward semantic absorption result sequence, mapping each element in the response-side backward semantic absorption result sequence to the key representation space and the value representation space respectively, generating the response-side key representation sequence and the response-side value representation sequence.

[0040] Corresponding to step S2251, the interactive sensing subunit performs two independent linear space mapping transformations on the response-side backward semantic absorption result sequence. The interactive sensing subunit internally maintains two trainable mapping matrices, namely the key mapping matrix W. K Sum value mapping matrix W V Both matrices have a size of d. model ×d kv , where d kv This represents the spatial dimensions of the mapped key and value representations. The j-th element h in the sequence of backward semantic absorption results from the response side... j a Through matrix multiplication operation k j =h j a ·W K Obtain the response sidekey representation vector k j Through matrix multiplication operation v j =h j a ·W V Obtain the response side value representation vector v j After mapping all elements in the sequence, a response sidekey representation sequence K={k1,k2,...,k} is generated. n The response side value represents the sequence V={v1,v2,...,v}. n}, where n is the length of the response word sequence.

[0041] Step S2253: Perform correlation response measurement processing on each query representation in the query representation sequence on the question side and each key representation in the key representation sequence on the response side to obtain the original correlation response matrix used to characterize the degree of correlation between the question words and the response words.

[0042] After obtaining the query representation sequence Q, the key representation sequence K, and the value representation sequence V on the query side, the interaction-aware subunit uses a scaled dot product attention mechanism to quantify the degree of semantic association across sequences. Specifically, for the i-th query representation q in the query representation sequence... i The response side key represents the j-th key in the sequence, which represents k. j Calculate their dot product similarity score(i,j)=q i ·k j T The dot product similarity reflects the consistency of vector direction between the i-th question word and the j-th response word in the mapped semantic space. To eliminate the influence of vector dimension on the magnitude of the dot product, the calculated dot product similarity is divided by a scaling factor. The scaled relevance scores are obtained. Iterating through all m query representations in the query-side query representation sequence and all n key representations in the response-side key representation sequence, the scaled relevance scores between each pair of words are calculated. The results are organized into an m x n two-dimensional matrix, which is the original relevance-response matrix S. raw The element S in the i-th row and j-th column raw(i,j) =(q i ·k j T ) / .

[0043] Step S2254: Perform distribution normalization on the original association response matrix along the response word dimension, so that all association response values ​​corresponding to each question word in the original association response matrix form a normalized distribution, and generate a normalized association response matrix.

[0044] After obtaining the original correlation response matrix S raw Next, the interactive perception subunit independently performs Softmax normalization on each row of the matrix. After Softmax normalization, the sum of all elements in the i-th row of the matrix equals 1. The element P(i,j) in the i-th row and j-th column can be interpreted as the associative attention weight obtained by the j-th response word given the i-th question word. Softmax normalization is then applied row by row to the entire original associative response matrix to generate a normalized associative response matrix P. The size of the normalized associative response matrix P is the same as that of S. raw The same applies; it still consists of m rows and n columns.

[0045] Step S2255: Using the normalized correlation response matrix as a guiding signal, selectively aggregate the response side value representation sequence to generate a response side context aggregation representation sequence for each question word.

[0046] The interactive perception subunit uses the normalized association response matrix P as the basis for attention weight allocation and performs a weighted convergence operation on the response-side value representation sequence V. For the i-th question word, its context-gathered representation c in the response-side semantic space... i The result is calculated by the weighted sum of the vectors representing all values ​​in the response side value representation sequence: c i =Σ j=1 n P(i,j)·v j Vector c i It integrates the value representation information of all words in the response text, but the proportion of each word's information included is regulated by the corresponding attention weight P(i,j) in the normalized association response matrix. The response word that is more semantically related to the i-th question word has a higher value representation in c. i The higher the contribution ratio, the better. The above weighted aggregation operation is performed sequentially on each position in the query representation sequence on the query side to generate the response side context aggregation representation sequence C={c1,c2,...,c...}. m}

[0047] Step S2256: Perform channel merging processing on the response-side context aggregation representation sequence and the question-side forward semantic absorption result sequence, update the semantic representation content of the question-side forward semantic absorption result sequence, and adjust the correlation response intensity distribution at the corresponding position in the semantic correlation response intensity distribution map according to the updated semantic representation content.

[0048] After obtaining the response-side context aggregation representation sequence C, the interaction-aware subunit combines it with the original question-side forward semantic absorption result sequence H. q ={h1 q h2 q ,...,h m q A concatenation operation is performed along the feature channel dimension to generate an enhanced question-side semantic representation sequence H. q '={[h1 q ;c1],[h2 q ;c2],...,[h m q ;c m]}, where [;] denotes vector concatenation. The enhanced semantic representation sequence of the question side retains the original semantic content of the question text itself and incorporates relevant contextual information from the response side that has been filtered by the attention mechanism. Subsequently, the interaction-aware subunit optionally feeds the enhanced semantic representation sequence of the question side into a feedforward fully connected layer for feature fusion and dimensionality reduction. The activation function of the feedforward fully connected layer uses a Gaussian error linear unit, and the output dimension is restored to the same dimension as the original question-side forward semantic absorption result sequence. The updated semantic representation content will be used to refresh the relevant records in the semantic association response intensity distribution map. Specifically, the association intensity distribution value corresponding to the i-th question word and each response word in the semantic association response intensity distribution map will be updated to the attention weight distribution recalculated based on the enhanced representation, or the original distribution value in the normalized association response matrix P will be retained and iteratively fine-tuned according to the feedback gradient of the subsequent network. The semantic association response intensity distribution map finally resides in the output buffer of the interaction-aware subunit in the form of a two-dimensional tensor.

[0049] Step S226: Based on the semantic association response intensity distribution map, reorganize the forward semantic absorption result sequence of the question side and the backward semantic absorption result sequence of the response side to generate a round dialogue representation vector that simultaneously integrates the semantic information of the question side and the semantic information of the response side.

[0050] The text understanding unit aggregates and merges the intermediate representations generated in steps S221 to S225 to generate the final vector representation of this round of dialogue. The text understanding unit extracts several correspondences with the highest correlation strength between question and response words from the semantic association response intensity distribution map. Based on these high-correlation correspondences, it locates the start and end positions of key semantic segments in the question-side forward semantic absorption result sequence and the response-side backward semantic absorption result sequence, respectively. For example, if the semantic association response intensity distribution map shows a significant correlation response peak between the question word "challenging" and consecutive words such as "maximum," "challenge," "lies in," "cold start," and "data sparsity" in the response word sequence, the text understanding unit merges the vector representation corresponding to "challenging" in the question-side forward semantic absorption result sequence with the vector representation corresponding to the consecutive segment "the greatest challenge lies in cold start data sparsity" in the response-side backward semantic absorption result sequence. During the fusion process, the identified key question-side vector representations and key response-side vector representations are subjected to average pooling or max pooling, respectively, to obtain the question-side aggregated vector and the response-side aggregated vector. These two aggregated vectors are then input into a gated fusion unit, which uses a learnable gate vector g = σ(W). g ·[v q ;v a ]+b gThe system dynamically adjusts the fusion ratio of question-side and response-side information, and outputs the fused vector representation v. fused =g⊙v q +(1-g)⊙v a Where ⊙ represents element-wise multiplication, σ represents the Sigmoid activation function, and W g and b g These are trainable parameters. Furthermore, the text understanding unit can also incorporate the first and last position vectors of the question-side forward semantic absorption result sequence, and the first and last position vectors of the response-side backward semantic absorption result sequence, into the fusion scope. These vectors are concatenated along the feature dimension and compressed to a preset turn-based dialogue representation vector dimension through a fully connected layer, thereby generating a turn-based dialogue representation vector carrying bidirectional dialogue context awareness. This turn-based dialogue representation vector, as a compact semantic encoding of the turn-based dialogue, will be passed to subsequent units of the dialogue flow state parser for state classification and transition derivation.

[0051] Step S230: Input the round dialogue representation vector into the state classification unit of the dialogue flow state parser for dialogue stage type determination processing, and assign a dialogue stage belonging label to each round dialogue representation vector. The dialogue stage belonging label indicates the interview stage type in which the current dialogue round is located.

[0052] The state classification unit is structurally a multilayer perceptron classifier. The number of neurons in its input layer matches the dimension of the dialogue round representation vector. The hidden layer contains two fully connected layers with 1024 and 512 neurons respectively. Batch normalization layers and random deactivation layers are inserted between the fully connected layers to prevent overfitting. The probability of random deactivation is set to 0.1, and the activation function is a modified linear unit. The number of neurons in the output layer equals the total number of preset interview dialogue stage categories, followed by a Softmax activation function to output the predicted probability distribution for each category. The preset set of interview dialogue stage category labels can, for example, include but is not limited to: "Opening Greetings," "Self-Introduction," "Project Experience Exploration," "Technical Ability Assessment," "Behavioral Question Assessment," "Career Planning Inquiry," "Salary Expectation Communication," "Reverse Questioning," and "Closing Acknowledgments." After receiving the dialogue round representation vector generated in step S220, the state classification unit forward-propagates the vector through the layers of the multilayer perceptron, obtaining the confidence scores for each preset interview stage category at the output layer. The state classification unit selects the category with the highest confidence score as the dialogue stage affiliation label for that round of dialogue, and records this label along with the confidence score. The assignment result of the dialogue stage affiliation label reflects the judgment conclusion of the dialogue flow state parser on the macro-level interview stage of the current interview round.

[0053] Step S240: Input the dialogue stage attribution label of adjacent dialogue rounds into the transition derivation unit of the dialogue flow state parser for state transition path derivation processing. Based on the combination pattern of the dialogue stage attribution label of the previous dialogue round and the dialogue stage attribution label of the next dialogue round, determine the dialogue state transition path identifier between adjacent timestamp nodes.

[0054] The transition derivation unit internally maintains a preset dialogue state transition rule table. This rule table defines all legal dialogue stage transition paths and their corresponding identifiers during the interview process in the form of a two-dimensional matrix or a finite state automaton. For example, the rule table records the path identifier for the transition from "Self-introduction stage" to "Project experience exploration stage" as "SELF_INTRO_TO_PROJECT_DEEP_DIVE", the path identifier for the transition from "Project experience exploration stage" to "Technical ability assessment stage" as "PROJECT_TO_TECH_ASSESS", the path identifier for the transition from "Technical ability assessment stage" back to "Technical ability assessment stage" as "TECH_ASSESS_CONTINUE", and the path identifier for the transition from "Salary expectation communication stage" to "Closing thanks stage" as "COMPENSATION_TO_CLOSING", etc. Furthermore, the rule table also includes rules for identifying abnormal transition paths. For example, if the label of the previous dialogue round is "opening greeting stage" and the label of the next dialogue round jumps directly to "salary expectation communication stage," the transition derivation unit will determine that this transition is a jump transition in the interview process based on the rule table and assign a corresponding jump transition path identifier. During the actual execution of state transition path derivation, the transition derivation unit reads the state classification results of adjacent rounds one by one, using the label of the previous dialogue stage as the source state and the label of the next dialogue stage as the target state. A key-value query operation is performed in the rule table. If the corresponding preset transition rule is matched, a predefined dialogue state transition path identifier is returned. If not, an identifier representing an undefined transition path is generated according to the default processing logic in the rule table, and log recording may be triggered optionally. The determination of the dialogue state transition path identifier allows for an explicit representation of the macro-rhythm and process transition characteristics of the interview process.

[0055] Step S250: Construct a dialogue state transition link in time order based on the dialogue stage attribution label and the dialogue state transition path identifier, and mark the timestamp information of each node in the dialogue state transition link to maintain the temporal alignment with the interview dialogue text sequence set.

[0056] After classifying the states of all rounds and deducing the transition paths between adjacent rounds, the dialogue flow state parser assembles the scattered dialogue stage attribution labels and dialogue state transition path identifiers into a complete dialogue state transition link. The dialogue state transition link is implemented in the data structure as a doubly linked list or a sequential list. Each node object in the list contains the following attribute fields: a unique node identifier, the associated round number, a dialogue stage attribution label string, a dialogue state transition path identifier string pointing to the next node, a node effective start timestamp, and a node effective end timestamp. The node effective start timestamp is taken from the start timestamp of the interviewer's question text segment in that round of dialogue, and the node effective end timestamp is taken from the end timestamp of the applicant's response text segment in that round of dialogue. Nodes are connected sequentially in ascending order of round number, and the dialogue state transition path identifier field within each node points to the transition path from the previous node to the current node. Through the above node construction operations, the dialogue flow state parser strings together discrete round-level state information into a dialogue state transition link that evolves continuously in the time dimension. Since each state node is directly bound to the specific timestamp range of the question and answer text fragments in the corresponding round, the dialogue state transition link naturally maintains a precise temporal alignment with the set of interview dialogue text sequences output in step S100. Subsequent processing steps can perform bidirectional retrieval and association between the state transition link and the original text sequence through timestamp index.

[0057] Step S260: Output the dialogue state transition links that have completed temporal alignment as a dialogue behavior state transition sequence. Each transition node in the dialogue behavior state transition sequence contains the current dialogue stage affiliation label and the dialogue state transition path identifier pointing to the previous node.

[0058] Based on the dialogue state transition chain constructed in step S250, the dialogue flow state parser serializes and encapsulates it to generate a dialogue behavior state transition sequence as the final output of step S200. The dialogue behavior state transition sequence can be expressed in JSON or Structured Data Exchange (SDE) format when output. Each element in the sequence corresponds to a node in the transition chain, and each element outputs at least the following information: node number, dialogue stage affiliation label, dialogue state transition path identifier, start timestamp value, and end timestamp value. Optionally, the dialogue behavior state transition sequence can also include auxiliary information such as the confidence score of each node and the dimensionality-reduced visualization coordinates of the dialogue representation vector of the corresponding round in the feature space. The output of the dialogue behavior state transition sequence marks the completion of the dialogue flow state parsing processing stage. This sequence will serve as one of the core input signals for the feedback mapping activation network in the subsequent step S300, providing key dialogue stage state constraints for feedback item mapping.

[0059] Step S300: Invoke the preset feedback mapping activation network to perform joint feedback item mapping processing on the dialogue behavior state transition sequence and the interview dialogue text sequence set, and generate an initial feedback item mapping set corresponding to each semantic unit in the candidate's response text fragment. Each initial feedback item in the initial feedback item mapping set is associated with an activation intensity label and a feedback category guidance label in the feedback knowledge space.

[0060] In one implementation, step S300 may specifically include the following steps S310 to S360: Step S310: Input the dialogue behavior state transition sequence into the state awareness unit of the feedback mapping activation network, extract the dialogue stage affiliation label and dialogue state transition path identifier corresponding to each timestamp node in the dialogue behavior state transition sequence, and generate dialogue stage evolution trajectory information.

[0061] The State Aware Unit (SAU) is the pre-context parsing module in the Feedback Mapping Activation Network (FPNA) and contains a sequence encoder and a trajectory aggregator. The sequence encoder employs a unidirectional Long Short-Term Memory (LSTM) network structure. The calculation formulas for the input, forget, and output gates follow the standard LSTM gating mechanism, and the hidden layer dimension is set to 256. The sequence encoder sequentially reads the embedded representation vectors of the dialogue stage attribution labels and the embedded representation vectors of the dialogue state transition path identifiers for each node in the dialogue behavior state transition sequence. The label embeddings and path identifier embeddings are obtained through two independent embedding matrices, each with an embedding dimension of 128. At each time step, the sequence encoder outputs a hidden state vector, which incorporates the cumulative state evolution information from the sequence start point to the current time step. The trajectory aggregator collects the hidden state vectors output by the sequence encoder at all time steps, performs temporal average pooling and max pooling on them, concatenates the pooling results, and generates a fixed-dimensional global trajectory summary vector through a fully connected layer. The global trajectory summary vector and the hidden state vectors at each time step together constitute the dialogue stage evolution trajectory information. The information on the evolution trajectory of the dialogue phase includes both the local state context at each time step and a general description of the overall evolution trend of the entire interview dialogue from the initial stage to the current stage.

[0062] Step S320: Input the candidate's response text fragments from the interview dialogue text sequence set into the response parsing unit of the feedback mapping activation network, perform semantic role decomposition processing on the candidate's response text fragments, extract the core semantic role units and the modifying semantic role units surrounding the core semantic role units in the candidate's response text fragments, and generate response semantic skeleton information.

[0063] The response parsing unit integrates a semantic role labeling module based on a fine-tuned pre-trained language model. The basic framework of this module uses the RoBERTa-large model architecture, which contains a 24-layer Transformer encoder with 1024 hidden layers and 16 self-attention heads. A linear classification layer and a conditional random field sequence labeling layer are layered on top of this pre-trained language model to predict the semantic role label for each word in the input word sequence. The label set for semantic role labels follows a propositional semantic role labeling system, including but not limited to: core agent label, core patient label, core action label, time modifier label, location modifier label, manner modifier label, degree modifier label, cause modifier label, purpose modifier label, condition modifier label, and concession modifier label. When performing semantic role decomposition on the applicant's response text fragment, the response parsing unit first inputs the response text fragment into the RoBERTa-large model and extracts the context representation vector for each word output by the model's last layer. Subsequently, the linear classification layer maps the context representation vector of each word to the label space dimension, obtaining the emission probability of each word for each semantic role label. The conditional random field layer then introduces label transition constraints and uses the Viterbi decoding algorithm to solve for the globally optimal label sequence. After decoding, the response parsing unit identifies continuous core agent word segments, core action word segments, core patient word segments, and other various modifier word segments based on the label sequence. It extracts the text spans assigned semantic role labels as semantic role units. The response parsing unit further determines the dependency connections between core semantic role units and modifier semantic role units using a semantic dependency analysis algorithm. Core semantic role units are connected through core argument relations edges, and modifier semantic role units are attached to the core semantic role units they modify or upper-level modifier semantic role units through modification relations edges. This constructs a directed acyclic graph with semantic role units as nodes and semantic dependency relations as edges; this graph structure constitutes the response semantic skeleton information. The semantic skeleton information of the response is described at the data level by a list of nodes and a list of edges. Each node records the original text span it covers, the semantic role type label, and the initial weight field. Each edge records its starting node index, ending node index, and dependency relationship type.

[0064] Step S330: Input the dialogue stage evolution trajectory information and response semantic skeleton information into the conditional mapping unit of the feedback mapping activation network, and perform association strength adjustment processing on the semantic role units in the response semantic skeleton information according to the dialogue stage attribution label in the dialogue stage evolution trajectory information to generate modulated semantic skeleton information constrained by the dialogue stage state conditions.

[0065] In one implementation, step S330 may specifically include the following steps S331 to S336: Step S331: Extract the dialogue stage attribution label at each timestamp node from the dialogue stage evolution trajectory information, and determine the relative stage sequence indicator of the dialogue stage attribution label in the preset interview stage sequence.

[0066] The conditional mapping unit maintains a pre-defined interview stage sequence configuration file. This file arranges all possible dialogue stage affiliation tags in the interview process into an ordered list according to the chronological order of a typical interview process, and assigns a relative stage ordinal indicator, starting from 0 and incrementing by an integer, to each stage tag in the list. For example, the ordinal indicator for "Opening Greetings Stage" is 0, for "Self-Introduction Stage" it is 1, for "Project Experience Exploration Stage" it is 2, for "Technical Ability Assessment Stage" it is 3, and so on. The conditional mapping unit iterates through the dialogue stage evolution trajectory information, performing an exact string matching query in the pre-defined interview stage sequence configuration file for each timestamp corresponding to the dialogue stage affiliation tag. If a match is found, the relative stage ordinal indicator associated with that tag is read; if no match is found, the relative stage ordinal indicator is set to a default value or estimated by interpolation based on the ordinal position of adjacent known tags. The extraction of the relative stage ordinal indicator transforms discrete dialogue stage category tags into continuous or ordered numerical variables, facilitating their use as conditional inputs in subsequent adjustment processing.

[0067] Step S332: Query the preset stage-semantic role association strength configuration table according to the relative stage sequence indicator to obtain the baseline association strength adjustment indicator signal of different semantic role types under the current dialogue stage affiliation label conditions.

[0068] The conditional mapping unit internally contains a pre-defined stage-semantic role association strength configuration table. This table is a two-dimensional lookup table. The row index corresponds to the relative stage ordinal indicator, and the column index corresponds to various semantic role type labels. Each cell in the table stores a real-valued adjustment indicator signal. The adjustment indicator signal value is set within the range [0.2, 2.0]. A value greater than 1.0 indicates that the semantic role type should receive increased attention in the current interview stage, a value less than 1.0 indicates that the semantic role type should receive decreased attention in the current interview stage, and a value equal to 1.0 indicates that the original strength should be maintained without adjustment. The specific parameter values ​​in the configuration table can be manually set by domain experts or learned through statistical analysis of historical high-performing interview feedback samples. For example, in the row corresponding to the "Project Experience Deepening Stage" of the relative stage sequence indicator, the adjustment indicator signal values ​​of the "Core Action Label" column and the "Method Modification Label" column are set to 1.5 and 1.3, respectively, to strengthen the focus on the applicant's behavior and execution method, while the adjustment indicator signal value of the "Time Modification Label" column is set to 0.8 to appropriately weaken the weight of time information. In the row corresponding to the "Technical Ability Assessment Stage" of the relative stage sequence indicator, the adjustment indicator signal values ​​of the "Core Action Label" column and the "Degree Modification Label" column are set to 1.6 and 1.4, respectively. The condition mapping unit uses the relative stage sequence indicator obtained in step S331 as the row index and each semantic role type label as the column index, and performs a cell value query operation column by column in the configuration table to obtain the set of baseline association strength adjustment indicator signals for each semantic role type under the current dialogue stage attribution label.

[0069] Step S333: Extract each semantic role unit contained in the response semantic skeleton information, identify the semantic role type label of each semantic role unit, and use the semantic role type label to distinguish the functional positioning of the semantic role unit in semantic expression.

[0070] The condition mapping unit iterates through the node list in the response semantic skeleton information, reading the semantic role type label field stored in each semantic role unit node object. This semantic role type label field has been assigned a value in the semantic role decomposition processing stage of step S320, and the specific name of the label is consistent with the column index of the stage-semantic role association strength configuration table, ensuring the namespace consistency of subsequent matching operations.

[0071] Step S334: Match the semantic role type label of each semantic role unit with the baseline association strength adjustment indication signal to determine the individual association strength adjustment indication component of each semantic role unit under the constraint of the attribution label in the current dialogue stage.

[0072] For the k-th semantic role unit in the response semantic skeleton information, the condition mapping unit obtains its semantic role type tag string. k The key to retrieve the reference correlation strength adjustment indication signal set obtained in step S332 is tag. k The adjustment indicator signal value is then assigned. If the retrieval is successful, this adjustment indicator signal value is directly assigned to the semantic role unit as its individual association strength adjustment indicator component (adj). k If the search fails, it indicates that the specific semantic role type is not covered in the current adjustment configuration table. In this case, the individual association strength adjustment indicator component (adj) will be adjusted. k Setting it to the default neutral value of 1.0 means that the semantic role unit is neither enhanced nor suppressed.

[0073] Step S335: Apply the individual association strength adjustment indicator component to the original connection weight field of the corresponding semantic role unit in the response semantic skeleton information, and perform update and replacement processing on the original connection weight field to obtain the semantic role unit after updating the connection weight field.

[0074] Each semantic role unit node in the response semantic skeleton information has an initial connection weight field. This field is assigned an initial value during the response semantic skeleton information initialization phase in step S320. The initial value can be uniformly set to 1.0, or it can be calculated based on the node centrality index of the semantic role unit in the dependency graph. For example, the normalized result of the PageRank score of the node in the dependency graph can be used as the initial connection weight. The condition mapping unit reads the initial connection weight field value w of the k-th semantic role unit. k_raw Adjust the indicator component adj with the individual association strength determined in step S334. k Perform element-wise multiplication or scalar multiplication to obtain the updated connection weight field value w. k_new =w k_raw ×adj k Conditional mapping unit uses w k_new The original connection weight fields in the semantic role unit node objects are overwritten to complete the weight update operation. As a result, during the interview stage where interviewers focus on project experience, the connection weights of semantic role units related to behaviors and implementation methods are systematically increased, gaining greater influence in subsequent semantic matching retrieval, while the connection weights of semantic role units with lower relevance to the current stage's focus are relatively decreased.

[0075] Step S336: Based on all semantic role units after updating the connection weight fields and their interdependent connection edges, reconstruct the semantic skeleton topology to generate modulated semantic skeleton information constrained by the state conditions of the dialogue stage.

[0076] After updating the connection weight fields of all semantic role units in the response semantic skeleton information, the conditional mapping unit maintains the original dependency connection edges between semantic role units, only replacing the original node set with the updated semantic role unit node set, and reassembling it into a semantic skeleton graph with the same structure but changed node weight attributes. The conditional mapping unit serializes the reassembled semantic skeleton graph into the same data exchange format as the response semantic skeleton information in step S320, and outputs modulated semantic skeleton information constrained by the state conditions of the dialogue stage. While maintaining the original semantic content and dependency structure, the modulated semantic skeleton information explicitly encodes the differentiated attention to different semantic components in the applicant's response at the current interview stage through differentiated adjustment of node weights, providing accurate semantic guidance signals for subsequent conditional feedback item retrieval in the feedback knowledge space.

[0077] Step S340: Based on the combination relationship between semantic role units in the modulated semantic skeleton information, perform semantic matching retrieval processing in the preset feedback knowledge space, activate candidate feedback item nodes that have a semantic correspondence with the modulated semantic skeleton information, and obtain the set of activated candidate feedback item nodes.

[0078] In one implementation, step S340 may specifically include the following steps S341 to S346: Step S341: Perform connected subgraph segmentation on the modulated semantic skeleton information to identify the set of independent semantic connected subgraphs composed of semantic role units in the modulated semantic skeleton information. Each independent semantic connected subgraph represents a relatively complete semantic expression segment.

[0079] After receiving the modulated semantic skeleton information, the Feedback Mapping Activation Network (FMAP) first performs weak connectivity analysis on the semantic skeleton graph structure. The analysis algorithm employs a depth-first search-based graph traversal strategy, starting from any unvisited node in the graph and recursively visiting all reachable nodes along the undirected direction of dependency edges. The set of nodes covered by a single traversal and the edges between them constitute an independent semantically connected subgraph. This traversal process is repeated until all nodes in the graph have been visited and assigned to a connected subgraph. For larger connected subgraphs containing a large number of semantic role units, the FMAP can optionally perform further spectral clustering segmentation. A Laplacian matrix is ​​constructed based on the updated connection weight field values ​​of each node and the dependency relationship type between nodes. The first k eigenvectors of the Laplacian matrix are calculated, and k-means clustering is applied to the nodes in the low-dimensional embedding space spanned by these eigenvectors. The clustering results are used as the basis for finer-grained subgraph partitioning. After completing the connected subgraph segmentation process, an independent semantic connected subgraph set is obtained. Each subgraph in the set is a local topological structure of the original semantic skeleton graph. The semantic role units inside the subgraph are tightly coupled through dependency edges, and together express a relatively independent and complete semantic proposition, such as a specific project experience described by the applicant, the process of overcoming a technical difficulty, or a team collaboration case.

[0080] Step S342: Perform linear serialization on the semantic role units in each independent semantically connected subgraph according to the semantic dependency direction to generate a semantic role unit sequence representation. The semantic role unit sequence representation preserves the sequential relationship between semantic role units.

[0081] For the p-th independent semantically connected subgraph in the set of independent semantically connected subgraphs, the Feedback Mapping Activation Network (FPNA) needs to convert it from a graph structure into a linear sequence for input into subsequent semantic encoding modules. The linear sequence arrangement follows the principle of semantic dependency direction priority and original text order as an aid. First, the FPNA identifies the root node with zero in-degree in the subgraph; the root node is typically a core action semantic role unit (SRU) or a core agent semantic role unit (ARU). Starting from the root node, a breadth-first traversal is performed along the out-degree direction of the directed dependency edges, adding the SRU nodes visited sequentially during the traversal to a temporary queue in the order of visit. If multiple child nodes appear at the same level during the traversal, they are reordered based on their starting positions in the original applicant response text fragment, with nodes appearing earlier being added to the queue first. After the traversal, the sequentially arranged sequence of SRU node objects in the temporary queue represents the SRU sequence of the independent semantically connected subgraph. Each element in the sequence carries the node description text of the corresponding SRU, a semantic role type label, and an updated connection weight field value.

[0082] Step S343: Convert the semantic role unit sequence representation into a retrieval semantic representation that can be searched in the feedback knowledge space. The feedback knowledge space pre-stores a large number of standard feedback item nodes and the index semantic representation of each standard feedback item node.

[0083] The feedback mapping activation network invokes a sequence encoder to encode the semantic role unit sequence representation into a fixed-length dense vector, which serves as the retrieval semantic representation of the independent semantically connected subgraph. This sequence encoder employs the same Sentence-BERT model architecture as the feedback knowledge space construction phase and loads the same model weight parameters as those used in constructing the feedback knowledge space index, ensuring semantic space consistency. The specific encoding process is as follows: The node description text of each semantic role unit in the sequence representation is extracted sequentially. The node description text represents the text span covered by that semantic role unit in the original applicant's response text fragment. The node description texts of each node are concatenated into a long text string in sequence, with node texts separated by a special delimiter. The concatenated long text string is input into the Sentence-BERT model, and the sentence vector output by the model is taken as the retrieval semantic representation of the independent semantically connected subgraph. Since the Sentence-BERT model has compressed variable-length sequences into fixed-dimensional semantic vectors through pooling layers during the pre-training stage, this retrieval semantic representation can compactly summarize the semantic content carried by independent semantically connected subgraphs, and is in the same vector space as the index semantic representation of standard feedback item nodes in the feedback knowledge space.

[0084] Step S344: Perform semantic similarity matching between the retrieved semantic representation and the index semantic representation of all standard feedback item nodes in the feedback knowledge space, and select standard feedback item nodes whose semantic similarity matching results meet the preset activation conditions as candidate feedback item nodes.

[0085] The feedback mapping activation network uses the retrieval semantic representation vector q generated in step S343 as the query vector and performs approximate nearest neighbor retrieval in the FAISS index of the feedback knowledge space. The retrieval uses inner product similarity as a semantic similarity metric, calculated as sim(q, v). i )=q·v i T , where v iThis is the index semantic representation vector of the i-th standard feedback item node in the feedback knowledge space. The FAISS index returns the top-K standard feedback item nodes with the highest inner product similarity values ​​as the initial recall results, where K is set to 50 or 100. After obtaining the initial recall results, the feedback mapping activation network performs preset activation condition filtering on the recalled nodes. The preset activation conditions are, for example, that the inner product similarity value is not lower than a preset similarity threshold, which is set to 0.65 for example; and that the compatibility between the feedback category region to which the standard feedback item node belongs and the label to which the current dialogue stage belongs satisfies a preset compatibility rule. The preset compatibility rule is defined through a stage-category mapping table, for example, the feedback category regions compatible with the "technical ability assessment stage" include "technical ability evaluation", "knowledge mastery evaluation", "problem-solving evaluation", etc. Recalled nodes that simultaneously meet the similarity threshold condition and the category compatibility condition are retained, while nodes that do not meet either condition are removed. The retained standard feedback item nodes are the activated candidate feedback item nodes.

[0086] Step S345: Organize all the selected candidate feedback item nodes into a set according to their storage location in the feedback knowledge space to form an active candidate feedback item node set. Each node in the active candidate feedback item node set inherits its node identifier and node description content in the feedback knowledge space.

[0087] The feedback mapping activation network completely extracts each candidate feedback item node object selected in step S344 from the feedback knowledge space and organizes them into a list structure. This list is the set of activated candidate feedback item nodes. Each element in the set is a structure that contains at least the following fields: a node identifier field, directly copied from the unique identifier of the corresponding node in the feedback knowledge space; a node description content field, directly copied from the node description content text of the corresponding node in the feedback knowledge space; a feedback category region identifier field, directly copied from the category information of the corresponding node in the feedback knowledge space; and a semantic similarity score field, recording the inner product similarity value obtained by the node in the semantic similarity matching process in step S344.

[0088] Step S346: Record the subgraph identifier of the independent semantic connectivity subgraph on which each candidate feedback item node is activated, and establish the source association mapping relationship between the candidate feedback item node and the specific semantic fragment in the applicant's response text fragment.

[0089] The feedback mapping activation network assigns a globally unique subgraph identifier to each independent semantically connected subgraph generated in step S341. The subgraph identifier can be a universally unique identifier string or an incrementing integer sequence number. During the construction of the activated candidate feedback item node set in step S345, for each successfully activated candidate feedback item node, the feedback mapping activation network adds a source subgraph identifier field to the structure corresponding to that node while adding it to the set. The value of this field is the subgraph identifier of the independent semantically connected subgraph that triggered the activation. By recording the source subgraph identifier, a direct source-tracing association mapping relationship is established between the candidate feedback item node and a specific semantic fragment in the candidate's response text. Subsequent processing steps can use the subgraph identifier to trace back to the specific text fragment in the candidate's original response that generated the feedback item, providing factual support for generating an interpretable structured interview feedback document.

[0090] Step S350: Determine the activation intensity marker of each candidate feedback item node in the set of activated candidate feedback item nodes in the feedback knowledge space. The activation intensity marker is used to characterize the degree of matching between the candidate feedback item node and the modulated semantic skeleton information.

[0091] The feedback mapping activation network calculates the activation strength label for each candidate feedback node in the set of activated candidate feedback node nodes. The activation strength label is a continuous real value between 0 and 1, and its calculation formula integrates multi-dimensional matching signals. The primary matching signal is the semantic similarity score, and the feedback mapping activation network uses the inner product similarity value sim recorded in step S344. raw This is mapped to the activation intensity dimension through a scaling and translation operation, with the mapping function s. sem =σ(a·sim raw +b), where σ is the Sigmoid function, and a and b are adjustable scaling and offset parameters, taking values ​​of 4.0 and -2.0 respectively, so that sim raw When s equals the preset similarity threshold of 0.65 sem Approximately 0.5. The secondary matching signal is the semantic role weight consistency score. The feedback mapping activation network calculates the average value w of the updated connection weight field of all semantic role units that participate in constituting the source-independent semantic connected subgraph in the modulated semantic skeleton information. avg The average value is compared with the predefined expected weight distribution of the candidate feedback item node in the feedback knowledge space. The expected weight distribution is the weight vector pre-labeled for the candidate feedback item node during the knowledge graph construction stage based on its associated standard evaluation dimensions. The consistency score s weightThe activation intensity label is calculated using the cosine similarity function. The final value of the activation intensity label is a weighted harmonic sum of the semantic similarity score and the semantic role weight consistency score, with weight coefficients set to 0.7 and 0.3, respectively. After calculation, the feedback mapping activation network writes the activation intensity label value into the activation intensity label field of the corresponding candidate feedback item node.

[0092] Step S360: Based on the feedback category region to which each candidate feedback item node belongs in the feedback knowledge space, assign a feedback category guidance tag to each candidate feedback item node, and generate an initial feedback item mapping set that carries both an activation intensity tag and a feedback category guidance tag.

[0093] The feedback mapping activation network reads the feedback category region identifier field from the structure of each candidate feedback item node. This field has already been assigned a value during the feedback knowledge space construction phase. The feedback category regions in the feedback knowledge space are divided according to the conventional dimensions of interview evaluation, including but not limited to the following top-level categories: "Communication and Expression Ability Evaluation", "Professional Skills Mastery Evaluation", "Project Experience Matching Evaluation", "Problem Solving Approach Evaluation", "Teamwork Awareness Evaluation", "Learning Adaptability Evaluation", "Career Motivation Stability Evaluation", and "Comprehensive Quality Impression Evaluation". Each top-level category can be further subdivided into several subcategories. The feedback mapping activation network directly uses the feedback category region identifier string to which the candidate feedback item node belongs as the feedback category guide marker for that node. Finally, the feedback mapping activation network packages all nodes in the activated candidate feedback item node set together with their activation strength markers and feedback category guide markers to generate an initial feedback item mapping set. The initial feedback item mapping set, as the output of step S300, will be completely passed to the feedback item hierarchical growth network in step S400 for subsequent structured organization and hierarchical growth processing.

[0094] Step S400: Input the initial feedback item mapping set into the feedback item hierarchical growth network for multi-level clustering and branching growth processing to generate a feedback item hierarchical structure tree that starts with the feedback root topic and ends with the specific feedback statement. There is a parent-child relationship between nodes in any adjacent level of the feedback item hierarchical structure tree.

[0095] In one implementation, step S400 may specifically include the following steps S410 to S460: Step S410: Obtain the feedback category guidance mark carried by each initial feedback item in the initial feedback item mapping set, and perform initial clustering processing on the initial feedback item mapping set according to the feedback category guidance mark to generate a primary feedback item cluster set corresponding to different feedback category guidance marks.

[0096] The feedback item hierarchical growth network first traverses all initial feedback item records in the initial feedback item mapping set, reading the feedback category guide flag field from each record. The network maintains a hash map with the feedback category guide flag string as the key and the initial feedback item list as the value. For each initial feedback item, if its feedback category guide flag already has a corresponding key in the hash map, the initial feedback item is appended to the end of the list corresponding to that key; otherwise, an empty list is created with the feedback category guide flag as the key, and the initial feedback item is added to the list. After traversal, each key-value pair in the hash map constitutes a primary feedback item cluster, with the key being the cluster's feedback category guide flag and the value being the set of all initial feedback items belonging to that feedback category. The network then extracts each primary feedback item cluster from the hash map one by one, packaging them into primary feedback item cluster objects. Each object contains three attribute fields: cluster identifier, feedback category guide flag, and initial feedback item list. The set of all primary feedback item cluster objects constitutes the primary feedback item cluster set.

[0097] Step S420: Input each primary feedback item cluster in the primary feedback item cluster set into the knowledge anchoring layer of the feedback item hierarchy growth network, and perform category matching and alignment processing with the standard feedback category nodes in the preset feedback knowledge hierarchy framework to determine the anchoring position of each primary feedback item cluster in the feedback knowledge hierarchy framework.

[0098] In one implementation, step S420 may specifically include the following steps S421 to S426: Step S421: Perform intra-cluster semantic aggregation representation processing on the initial feedback items contained in the primary feedback item cluster, and extract the cluster centroid semantic representation information that can represent the overall semantic direction of the primary feedback item cluster. The cluster centroid semantic representation information integrates the semantic content of each initial feedback item in the cluster.

[0099] The knowledge anchoring layer first traverses the list of initial feedback items in the target primary feedback item cluster, extracting the node description text for each initial feedback item. The node description text is a feedback statement in natural language form, such as "The applicant has a deep understanding of microservice architecture and can clearly explain the service decomposition principles," or "The candidate lacks quantitative indicators when describing project difficulties, making their statements less persuasive." The knowledge anchoring layer collects the node description texts of all initial feedback items within the cluster into a text list, and calls the same Sentence-BERT sentence embedding model as in step S343 to encode each node description text one by one, generating the corresponding feedback item semantic representation vector. After obtaining the semantic representation vectors of all feedback items within the cluster, the knowledge anchoring layer performs an average pooling operation on these vectors, using the average vector obtained from the pooling as the cluster centroid semantic representation information of the primary feedback item cluster. The cluster centroid semantic representation information is a real vector with dimension 768. This vector is located near the center of each feedback item vector within the cluster in the semantic space, and can well represent the aggregate semantic theme of the entire primary feedback item cluster.

[0100] Step S422: Extract the category definition description information of each standard feedback category node from the preset feedback knowledge hierarchy framework, and perform semantic representation transformation on the category definition description information to generate standard category semantic representation information corresponding to each standard feedback category node.

[0101] The knowledge anchoring layer performs a traversal operation on all standard feedback category nodes in the feedback knowledge hierarchy framework. For each standard feedback category node encountered, it reads the stored category definition description information text. The knowledge anchoring layer calls the same Sentence-BERT sentence embedding model as in step S421, taking the category definition description information text as input and encoding it into a 768-dimensional semantic representation vector. This vector is the standard category semantic representation information corresponding to the standard feedback category node. Since the cluster centroid semantic representation information of the primary feedback item cluster and the standard category semantic representation information of the standard feedback category node are both encoded by the exact same Sentence-BERT model, they reside in the same semantic vector space, and vector similarity can be directly calculated.

[0102] Step S423: Perform semantic alignment measurement on the cluster centroid semantic representation information and the standard category semantic representation information of each standard feedback category node to obtain the semantic alignment measurement result that represents the semantic fit between the primary feedback item cluster and each standard feedback category node.

[0103] For the currently processed primary feedback item cluster, the knowledge anchoring layer calculates the cosine similarity between its cluster centroid semantic representation vector and the standard category semantic representation vector of each standard feedback category node generated in step S422. The cosine similarity value ranges from -1 to 1. The closer the value is to 1, the higher the semantic fit between the overall semantic orientation of the primary feedback item cluster and the definition description of the standard feedback category node. The knowledge anchoring layer uses the calculated cosine similarity value as a measure of the semantic alignment degree corresponding to the standard feedback category node.

[0104] Step S424: Sort all standard feedback category nodes in descending order according to the semantic alignment degree measurement results, select the standard feedback category node with the highest semantic alignment degree measurement result as the initial candidate anchor node of the primary feedback item cluster, and record the highest semantic alignment degree measurement result as the anchor reliability label.

[0105] The semantic alignment measurement results between the primary feedback item cluster and all standard feedback category nodes, calculated in step S423 of the knowledge anchoring layer, are used to construct a candidate list consisting of standard feedback category node objects and their corresponding cosine similarity scores. The knowledge anchoring layer uses a quicksort algorithm to sort the candidate list in descending order of cosine similarity scores. After sorting, the knowledge anchoring layer extracts the first element of the list; the standard feedback category node in this element is the initial candidate anchor node, and its corresponding cosine similarity score is the anchoring reliability label.

[0106] Step S425: If the anchoring reliability mark is not lower than the preset anchoring acceptance threshold, then the hierarchical path information and node coordinate information of the preliminary candidate anchoring node in the feedback knowledge hierarchy framework are assigned to the primary feedback item cluster as the anchoring mounting position of the primary feedback item cluster.

[0107] The knowledge anchoring layer has a pre-defined anchoring acceptance threshold of 0.55. The knowledge anchoring layer compares the anchoring reliability flag obtained in step S424 with the anchoring acceptance threshold. If the anchoring reliability flag is greater than or equal to 0.55, it indicates a high semantic consistency between the semantic content of the primary feedback item cluster and the definition description of the preliminary candidate anchoring node, making the anchoring result reliable. At this point, the knowledge anchoring layer locates the preliminary candidate anchoring node in the feedback knowledge hierarchy framework, traces back along its parent node pointer to the root node, and records the node identifier sequence of all standard feedback category nodes traversed along this backtracking path; this sequence is the hierarchical path information. Simultaneously, the knowledge anchoring layer records the internal index number of the preliminary candidate anchoring node in the tree-like data structure of the feedback knowledge hierarchy framework as node coordinate information. The knowledge anchoring layer encapsulates the hierarchical path information and node coordinate information into an anchor mounting location object and appends this object to the anchor mounting location attribute field of the primary feedback item cluster object.

[0108] Step S426: If the anchoring reliability mark is lower than the preset anchoring acceptance threshold, the primary feedback item cluster is split into multiple smaller-granularity sub-feedback item clusters, and the category matching alignment process is re-executed for each sub-feedback item cluster until all sub-feedback item clusters obtain anchoring positions that meet the anchoring acceptance threshold.

[0109] If the anchoring reliability score is less than 0.55, it indicates that the internal semantics of the current primary feedback item cluster are relatively scattered and difficult to be covered by a single feedback knowledge hierarchy framework standard category node. Forcibly anchoring the entire cluster to the highest similarity node may lead to overly coarse classification granularity. In this case, the knowledge anchoring layer triggers a cluster splitting process. The cluster splitting process adopts an adaptive splitting algorithm based on hierarchical clustering. First, it calculates the cosine distance between each pair of the semantic representation vectors of the feedback items of all initial feedback items within the cluster, constructing a distance matrix. Subsequently, using the distance matrix as input, it performs the agglomerative clustering process in hierarchical clustering, using the Ward minimum variance method as the inter-cluster distance metric, generating a binary tree diagram describing the gradual merging process of feedback items within the cluster. The knowledge anchoring layer performs a horizontal cut at a suitable height in this tree diagram. The selection of the cut height is automatically determined based on the principle of maximizing the silhouette coefficient. The initial feedback items contained in each branch formed after the cut constitute a sub-feedback item cluster. For each sub-feedback cluster generated by the split, the knowledge anchoring layer treats it as a new independent primary feedback cluster and recursively executes the processing flow from steps S421 to S426. This involves re-extracting the cluster centroid semantic representation information, recalculating the semantic alignment degree measurement results with each standard feedback category node, re-determining preliminary candidate anchoring nodes, and evaluating the anchoring reliability label, until each sub-feedback cluster successfully obtains a valid anchoring position that meets the anchoring acceptance threshold. During the recursive process, if an extreme case occurs where a sub-feedback cluster contains only a single initial feedback item and the anchoring reliability label is still below the threshold value, the knowledge anchoring layer will forcibly attach the sub-feedback cluster to the standard feedback category node with the highest semantic alignment degree measurement result and record the forced anchoring event in the log.

[0110] Step S430: Perform a hierarchical growth operation on the primary feedback item clusters according to the anchor mounting position, and aggregate the primary feedback item clusters with similar anchor mounting positions and the same upper-level category affiliation into intermediate-level feedback branch nodes. The intermediate-level feedback branch nodes aggregate the common feedback attribute features in their subordinate primary feedback item clusters.

[0111] After anchoring the knowledge of all primary feedback item clusters, the feedback item hierarchical growth network initiates a bottom-up hierarchical growth operation. First, the network groups all primary feedback item clusters according to the hierarchical path information in their anchored locations. Specifically, primary feedback item clusters with identical hierarchical path information are grouped into the same cluster to be aggregated. For each cluster to be aggregated, the network creates a new intermediate-level feedback branch node object, with its node type identifier set to "intermediate-level feedback branch node". The network collects all initial feedback items contained in all primary feedback item clusters within the cluster and processes the node description text of these initial feedback items using a text summarization module. This module employs a generative summarization algorithm based on the pre-trained language model BART to integrate multiple feedback statement texts into one or two concise branch semantic topic description texts, which serve as the branch semantic topic description text field content for the intermediate-level feedback branch node. For example, if a cluster to be aggregated contains multiple feedback statements about the applicant's communication skills, the text summarization module might generate "The applicant can clearly and logically articulate technical viewpoints, but speaks too fast and some details are not fully explained" as the aggregated branch semantic topic description text. Furthermore, intermediate-level feedback branch nodes also aggregate statistical attribute features such as the distribution of feedback category guidance labels and average activation intensity labels of the initial feedback items in their subordinate primary feedback item clusters. A parent-child hierarchical connection is established between intermediate-level feedback branch nodes and their subordinate primary feedback item clusters.

[0112] Step S440: Extract the branch semantic topic description text of each intermediate-level feedback branch node, and perform topic aggregation processing on the branch semantic topic description text of all intermediate-level feedback branch nodes. Further aggregate the intermediate-level feedback branch nodes with semantically compatible branch semantic topic description texts to form the feedback root topic node.

[0113] After generating all intermediate-level feedback branch nodes, the feedback item hierarchical growth network uses these nodes as input for a new round of aggregation. The network extracts the semantic topic description text for each intermediate-level feedback branch node and merges these descriptions into a long document. It then calls a Latent Dirichlet Allocation (LDA) model to model the topic of this long document, setting the number of topics to one and forcing the model to extract the single most core topic from the entire document set. The LDA outputs the topic word distribution for this core topic. Based on the top few most probable topics in this distribution, combined with a predefined document title template, the network generates the node description text for the root feedback topic node, such as "Comprehensive Ability Interview Feedback Evaluation Report for Applicant Zhang San". This root feedback topic node is the root node of the entire feedback item hierarchical structure tree. The network establishes parent-child relationships between the root feedback topic node and all intermediate-level feedback branch nodes, with the intermediate-level feedback branch nodes serving as direct children of the root feedback topic node. If the number of intermediate-level feedback branch nodes is large, the feedback term hierarchical growth network can optionally insert one or two additional layers of intermediate convergence nodes between the feedback root topic node and the intermediate-level feedback branch nodes to construct a more balanced tree structure. The generation of intermediate convergence nodes is also based on semantic clustering of the semantic topic description text of the lower-level node branches. The clustering algorithm uses k-means, and the k value is automatically selected based on the silhouette coefficient.

[0114] Step S450: Construct a tree-like connection structure from the root topic node of the feedback down through the intermediate level feedback branch nodes to the leaf node corresponding to each initial feedback item in the primary feedback item cluster. In the tree-like connection structure, the parent node and the child node are connected by directed belonging edges.

[0115] After completing bottom-up node aggregation and top-down node generation, the feedback item hierarchical growth network begins to construct a tree-like connection structure linking all hierarchical nodes. For each intermediate-level feedback branch node, the feedback item hierarchical growth network creates directed belonging edges from itself to each of its subordinate primary feedback item clusters, based on the aggregation relationship recorded in step S430. The edge type is identified as "containing". For the feedback root topic node, the feedback item hierarchical growth network creates directed belonging edges from itself to each of its subordinate intermediate-level feedback branch nodes. For each primary feedback item cluster, the feedback item hierarchical growth network traverses its internal list of initial feedback items, creating a leaf node object for each initial feedback item. The node type identifier of the leaf node object is set to "leaf node", the leaf node content field is copied from the node description text of the initial feedback item, and the activation intensity flag field is copied from the activation intensity flag of the initial feedback item. The feedback item hierarchical growth network creates directed belonging edges from the primary feedback item cluster to each of its subordinate leaf nodes. Through the edge connection operations described above, a complete directed acyclic tree diagram extending from the root node to all leaf nodes is constructed.

[0116] Step S460: Generate a feedback item hierarchy structure tree containing node identifiers, node level depth information, and parent-child node connection path information based on the tree connection relationship structure. Each leaf node in the feedback item hierarchy structure tree is a specific feedback statement.

[0117] The feedback item hierarchical growth network performs a breadth-first traversal on the tree-like connection structure constructed in step S450. During the traversal, a globally unique node identifier is assigned to each node. Node identifiers can be in hierarchical encoding form, for example, the root node identifier is "R", its first child node identifier is "R-1", the second child node identifier is "R-2", the grandchild node identifier is "R-1-1", and so on. The node identifier itself reflects the node's hierarchical path in the tree. The feedback item hierarchical growth network calculates the node's hierarchical depth information for each node. The root node's hierarchical depth is 0, the root node's direct child node's hierarchical depth is 1, and so on, with leaf nodes having the largest hierarchical depth. For each directed belonging edge, the feedback item hierarchical growth network records the node identifier of its starting parent node and the node identifier of its ending child node, forming a list of parent-child node connection path information. The feedback item hierarchical growth network serializes and packages the data structure containing all node objects, the node identifier mapping table, the node hierarchical depth information, and the parent-child node connection path information list into a unified output, generating the final feedback item hierarchical structure tree. Each leaf node in the feedback item hierarchy tree corresponds to a specific feedback statement. The content text of the specific feedback statement is stored in the leaf node content field of the leaf node, which can be directly extracted and rendered into the document body by step S500.

[0118] Step S500: Perform document rendering processing on the initial feedback item mapping set according to the feedback item hierarchy structure tree, and output a structured interview feedback document that conforms to the preset document format specifications. Each independent feedback paragraph in the structured interview feedback document has a content mapping correspondence with a branch node in the feedback item hierarchy structure tree.

[0119] In one implementation, step S500 may specifically include the following steps S510-S560: Step S510: Extract the node description text content of the feedback root topic node from the feedback item hierarchy structure tree, and fill the node description text content of the feedback root topic node into the document main title field position of the preset document format specification.

[0120] The document rendering engine first locates the root node with a node level depth of 0 in the feedback item hierarchy tree. It then reads the node description text content field stored in this node object. This field represents the main title of the interview feedback document generated in step S440, such as "Comprehensive Interview Evaluation Report on Applicant Li Si". The document rendering engine initializes a blank document object, which follows the page setup parameters defined in the preset document format specification. The document rendering engine writes the read node description text content into the document object's main title field and applies the "main title style" defined in the preset document format specification. The main title style typically includes: bold font, font size of 1 or 1, center alignment, and 1.5 line spacing after paragraphs.

[0121] Step S520: Following the hierarchical traversal order from the root topic node to the leaf node in the feedback item hierarchy structure tree, extract the branch name description text content of each intermediate level feedback branch node in sequence, and fill the branch name description text content of the intermediate level feedback branch node into the chapter title field position of the preset document format specification as the chapter title.

[0122] In one implementation, step S520 may specifically include the following steps S521 to S526: Step S521: Perform a breadth-first traversal operation starting from the root node on the feedback item hierarchy structure tree, visit each node in the feedback item hierarchy structure tree layer by layer, and record the node hierarchy depth label and node type identifier of each visited node.

[0123] The document rendering engine internally maintains a first-in, first-out (FIFO) traversal queue, adding the root node of the feedback item hierarchy tree as the starting point for traversal. The engine then performs the following operations in a loop until the queue is empty: It retrieves a node from the head of the queue, recording its node hierarchy depth and type identifier; it then traverses all of the node's direct child nodes, adding them to the tail of the queue in the order they appear in the tree structure. Through breadth-first traversal, the engine ensures that all nodes are processed in a hierarchical, top-down, and sequentially ordered manner, consistent with the top-down, layer-by-layer organization logic of structured documents.

[0124] Step S522: When a target node with the node type identifier of intermediate level feedback branch node is accessed, read the branch name description text content field stored in the target node and obtain the node level depth mark of the target node.

[0125] During the breadth-first traversal, whenever the node type identifier field of the current node retrieved from the queue is "intermediate-level feedback branch node", the document rendering engine pauses the regular traversal process and enters the intermediate-level node processing sub-process. The document rendering engine reads the value of the branch name description text content field from the target node's data structure, and simultaneously reads the value of the node level depth marker field of the target node, where the node level depth marker is an integer.

[0126] Step S523: Based on the node hierarchy depth marker of the target node, find the chapter level indicator corresponding to the target node in the preset hierarchy depth and chapter level mapping relationship. The chapter level indicator is used to determine the heading level of the branch name description text content in the document.

[0127] The document rendering engine internally includes a pre-defined mapping table between hierarchy depth and chapter level. This table defines the correspondence between integer values ​​of node hierarchy depth markers and chapter level indicators. For example, hierarchy depth 1 maps to first-level chapter titles, hierarchy depth 2 maps to second-level chapter titles, hierarchy depth 3 maps to third-level chapter titles, and hierarchy depth 4 maps to fourth-level chapter titles. The document rendering engine uses the node hierarchy depth marker of the target node as the lookup key to perform a key-value lookup in this mapping table to obtain the corresponding chapter level indicator.

[0128] Step S524: Combine the branch name description text content field with the chapter level indicator to generate a title formatted text string that conforms to the corresponding level title format requirements in the preset document format specification. The title formatted text string contains the corresponding level title number prefix.

[0129] The document rendering engine determines the numbering format of the current chapter title based on the chapter level indicator. In the preset document format specification, the numbering prefix format for the first-level chapter title is a sequence of Chinese numerals such as "一、", "二、", "三、", etc.; the numbering prefix format for the second-level chapter title is "(一)", "(二)", "(三)", etc.; the numbering prefix format for the third-level chapter title is a sequence of Arabic numerals with a dot, such as "1.", "2.", "3.", etc.; the numbering prefix format for the fourth-level chapter title is "(1)", "(2)", "(3)", etc. The document rendering engine internally maintains the current counter for each level of chapters. Each time a chapter title of a certain level is generated, the value of the counter at that level is converted into a number prefix string according to the specified numbering format. The document rendering engine concatenates the generated number prefix string with the text content field value of the branch name description of the target node, separated by a space in the middle, to obtain the complete formatted title text string, such as "二、Analysis of the Matching Degree of Project Experience and the Fit with the Position".

[0130] Step S525: Fill the formatted title text string into the chapter title field position specified by the chapter level indicator in the preset document format specification, and ensure that the font attributes and paragraph attributes of the formatted title text string are consistent with the corresponding definitions in the preset document format specification.

[0131] The document rendering engine inserts the formatted title text string at the write cursor position of the current document object and appends a line break after it. Subsequently, the document rendering engine applies a style to the paragraph object where the just-inserted formatted title text string is located. The style application operation includes: retrieving the corresponding chapter title style definition in the style sheet of the preset document format specification according to the chapter level indicator. The style definition includes property items such as font name, font size, whether it is bold, space before paragraph, space after paragraph, line spacing, outline level, etc. The document rendering engine sets the format attributes of the paragraph object to the corresponding values defined in the style sheet item by item. For example, the style definition for the first-level chapter title is: font is boldface, font size is三号, bold, space before paragraph is 0.5 lines, space after paragraph is 0.5 lines, alignment is left alignment, and outline level is level 1. The document rendering engine calls the property setting interface of the document object model to complete the above format application.

[0132] Step S526: During the traversal of the feedback item hierarchy tree, if it is detected that there is a jump in the node hierarchy depth marker between the currently accessed target node and the previous processed target node, corresponding empty chapter placeholder markers are automatically generated in the preset document format specification to maintain the continuity of the document structure hierarchy.

[0133] During the process of breadth-first traversal of the intermediate-level feedback branch nodes, the document rendering engine continuously tracks the node hierarchy depth marker last of the previous processed target node depthWhen processing the current target node, the document rendering engine compares the node's node hierarchy depth marker (cur) with the current node's node level depth marker. depth with last depth If cur depth equal to last depth +1 indicates that the document structure is normally downgraded and no additional processing is required. If cur depth Less than or equal to last depth This indicates that the current chapter level jumps back to the previous or same level, and the document rendering engine will then normally end the main text area of ​​the current chapter. If cur depth Greater than last depth +1, for example, jumping directly from level depth 1 to level depth 3, indicates that the intermediate levels of the document lack explicit chapter titles. To maintain the continuity of the document's hierarchical structure, the document rendering engine automatically performs empty chapter placeholder generation. The document rendering engine starts from the last... depth Starting with +1, generate chapter titles corresponding to the missing level depth sequentially. The text content of the chapter titles is filled with preset empty chapter placeholder text, such as "(To be supplemented)", until cur is generated. depth Up to level -1, then process cur normally. depth The target node at each level. Automatically generated empty chapter placeholders also apply the corresponding chapter title style. Through this mechanism, even if there are hierarchical jumps in the feedback item hierarchy tree, the final generated structured interview feedback document still maintains a complete and continuous chapter structure.

[0134] Step S530: For each intermediate-level feedback branch node, based on the parent-child connection relationship in the feedback item hierarchy tree, obtain the statement text content of the specific feedback statement corresponding to all leaf nodes under that intermediate-level feedback branch node.

[0135] In one implementation, step S530 may specifically include the following steps S531 to S536: Step S531: In the feedback item hierarchy structure tree, take the currently processed intermediate level feedback branch node as the search starting point, and perform a depth-first traversal search along the direction of the directed belonging edge pointing to the child node, and mark all nodes passed on the search path.

[0136] The document rendering engine creates a recursive search function that takes a node object as input. Internally, the function first adds the input node object to a set of visited nodes. Then, it retrieves a list of all direct child nodes of the input node object and recursively calls the search function for each child node in the list. The document rendering engine passes the currently processed intermediate-level feedback branch node as an argument to this recursive search function, initiating a depth-first search. Depth-first traversal ensures that the entire subtree rooted at the intermediate-level feedback branch node is completely traversed, without missing any leaf nodes on any branch.

[0137] Step S532: During the search process, identify the node type identifier of the node on the search path. When the node type identifier indicates that the current node is a leaf node, stop further searching of the branch of that node and add the leaf node to the leaf node collection set.

[0138] Internally, the recursive search function checks the node type identifier of the input node before adding it to the set of visited nodes. If the value of the node type identifier field is equal to "leaf node," it indicates that the current search path has reached the end of the tree structure, and the node no longer has any child nodes. In this case, the recursive search function adds the leaf node object to a predefined global leaf node collection list and then returns directly, without executing the subsequent steps of obtaining the child node list and recursive calls. If the node type identifier is not equal to "leaf node," the process of obtaining the child node list and recursive calls is executed normally.

[0139] Step S533: Continue to traverse and search other branches on the search path until the complete search of the subtree rooted at the currently processed intermediate level feedback branch node is completed, and a leaf node collection set containing all leaf nodes in the subtree is obtained.

[0140] Once the recursive search function starts from the currently processed intermediate-level feedback branch node, it automatically traverses all nodes and branches in the subtree in depth-first order. When the recursive search function finishes execution and returns to the initial call point, the global leaf node collection list has already collected all leaf node objects under that subtree. The document rendering engine outputs this leaf node collection list as the search results.

[0141] Step S534: Traverse each leaf node in the leaf node collection set, and obtain the leaf node content field stored in the feedback item hierarchy structure tree for each leaf node. The leaf node content field is the text content of the specific feedback statement.

[0142] The document rendering engine iterates through the leaf node collection list obtained in step S533, reading the leaf node content field for each leaf node object in the list. The leaf node content field stores a complete feedback statement, such as "able to clearly explain the project background and one's role," "has a comprehensive understanding of the system architecture design, considering scalability in high-concurrency scenarios," or "lacks specific data support when describing the process of overcoming technical difficulties." The document rendering engine temporarily stores the read statement text in a string list.

[0143] Step S535: Associate and store all the obtained statement text content with the corresponding leaf node identifier to form a mapping record table between leaf node identifiers and statement text content. The mapping record table is used to support the rapid tracing and location of specific feedback statements.

[0144] While traversing the leaf node collection, the document rendering engine simultaneously reads the node identifier of each leaf node. The engine creates a hash map object, using the leaf node identifier as the key and the text content of the statement read from that leaf node as the value, storing each pair of leaf node identifiers and statement text content in the hash map. This hash map is temporarily stored in memory as an intermediate data structure. After the document rendering is complete, this hash map can be serialized and saved to the document's attribute extension fields, or used to generate a reference index at the end of the document. This allows readers to quickly jump to or highlight the source node in the feedback item hierarchy structure tree when clicking on a specific feedback statement in the feedback document.

[0145] Step S536: Output the set of statement text content in the mapping record table as the statement text content of all specific feedback statements under the current intermediate-level feedback branch node.

[0146] The document rendering engine extracts all value objects from the mapping record table constructed in step S535, forming a list of statement text content. This list is the collection of statement text content for all specific feedback statements under the currently processed intermediate-level feedback branch node, and will be delivered to the subsequent display order arrangement processing steps.

[0147] Step S540: For the text content of all specific feedback statements under each intermediate level feedback branch node, arrange them in the order of display according to the activation intensity mark corresponding to the specific feedback statement in the initial feedback statement mapping set, and arrange the text content of the statement with higher activation intensity mark in the front area of ​​the paragraph.

[0148] After obtaining the set of text content of the statements under the current intermediate level feedback branch node, the document rendering engine does not directly output the leaf nodes according to their storage order in the tree. Instead, it optimizes and sorts the output order according to the activation intensity flag of each statement to ensure that the feedback statements that best match the current interview dialogue context, are the most important, or are the most prominent are presented to the interviewer first, thereby improving the reading efficiency and information delivery effect of the interview feedback document.

[0149] As one implementation method, step S540 involves arranging the text content of all specific feedback statements under each intermediate-level feedback branch node according to the activation intensity markers corresponding to the specific feedback statements in the initial feedback statement mapping set, placing the text content of statements with higher activation intensity markers in the beginning area of ​​the paragraph. This may specifically include the following steps S541~S546: Step S541: Retrieve the initial feedback item record of each leaf node corresponding to the specific feedback statement under the intermediate level feedback branch node currently being processed from the initial feedback item mapping set, and read the activation intensity flag from the initial feedback item record.

[0150] The document rendering engine obtains a list of node identifiers for all leaf nodes under the current intermediate-level feedback branch node through the mapping record table generated in step S535. For each leaf node identifier in the list, the document rendering engine performs a precise search in the initial feedback item mapping set output in step S300. Each initial feedback item record in the initial feedback item mapping set contains a node identifier field, which corresponds one-to-one with the leaf node identifier. After finding the corresponding initial feedback item record based on the leaf node identifier, the document rendering engine reads the value of the activation intensity flag field in that record. The activation intensity flag is a floating-point number between 0 and 1.

[0151] Step S542: Construct a temporary queue of binary tuples containing statement text content and corresponding activation intensity tags. Each element in the temporary queue of binary tuples consists of a statement text content and its associated activation intensity tag.

[0152] The document rendering engine initializes an empty list as a temporary queue for binary tuples. For each specific feedback statement under the current intermediate-level feedback branch node, the document rendering engine retrieves its statement text content string from the statement text content set in step S536, and retrieves its activation intensity flag value from the corresponding initial feedback item record read in step S541. The two are packaged into a binary data structure, and the document rendering engine appends the binary tuple to the tail of the temporary queue for binary tuples.

[0153] Step S543: Perform a queue rearrangement operation based on the activation intensity mark on the tuple temporary queue, so that the order of elements in the tuple temporary queue is consistent with the order of the activation intensity marks, wherein the elements with larger activation intensity marks are arranged at the front of the queue.

[0154] The document rendering engine uses a sorting algorithm to rearrange the temporary queue of binary tuples. The sorting algorithm is an introspective sort, where the comparison function compares the activation intensity marker values ​​of the elements in two binary tuples, requiring the element with the larger activation intensity marker value to be placed first in the sequence, achieving a descending order. After the sorting algorithm completes, the first element in the temporary queue is the concrete feedback statement with the highest activation intensity marker, and the last element is the concrete feedback statement with the lowest activation intensity marker.

[0155] Step S544: After completing the queue rearrangement operation, perform secondary order adjustment processing on elements with the same activation intensity marker in the binary temporary queue. The secondary order adjustment processing determines the order based on the order of the timestamps of the statement text content in the original applicant's response text fragment.

[0156] Based on the descending sorting result in step S543, the document rendering engine further scans the queue for consecutive groups of elements with completely equal activation intensity marker values. For each group of elements with equal activation intensity markers, the document rendering engine needs to determine their internal relative order based on the appearance time of the specific feedback statements corresponding to these elements in the original interview dialogue. The document rendering engine traces back to the source association mapping relationship between the candidate feedback item nodes and semantic segments in the candidate's response text fragment established in step S346 through leaf node identifiers and initial feedback item mapping records, obtains the source subgraph identifier corresponding to each specific feedback statement, and then locates its starting character offset or starting timestamp in the candidate's response text fragment based on the subgraph identifier. The document rendering engine performs an ascending stable sorting of the groups of elements with equal activation intensity markers based on the order of these starting timestamps, that is, statements with earlier timestamps are placed first, and statements with later timestamps are placed later. The secondary order adjustment process ensures that even when the activation intensity markers cannot distinguish priorities, the presentation order of the feedback statements still conforms to the natural chronological order of the interview dialogue.

[0157] Step S545: Extract the text content of the statements in the binary temporary queue after the queue rearrangement operation and secondary order adjustment to form a list of arranged text content of statements.

[0158] The document rendering engine iterates through the temporary queue of binary tuples after the sorting processes in steps S543 and S544, retrieving the text content string of each statement from each binary tuple element. The retrieved strings are then appended sequentially to a new list, which is the arranged list of statement text content. The order of the strings in the arranged list of statement text content determines the final rendering order when written to the document's main text area.

[0159] Step S546: Map each statement text content in the arranged statement text content list sequentially to the independent statement entry in the body paragraph area below the corresponding chapter title field in the preset document format specification.

[0160] Following the chapter title corresponding to the current intermediate-level feedback branch node, the document rendering engine creates a main text paragraph area. This paragraph area is organized using an item list format. The engine iterates through the arranged list of statement text content. For each statement text content in the list, it creates an independent paragraph object, fills the paragraph object with the statement text content, and adds an item number identifier to the beginning of the paragraph object. The item number identifier uses the text item numbering style defined in the preset document format specification, such as solid dots, hyphens, or Arabic numerals with periods. The engine then sequentially appends each statement paragraph with an item number identifier to the document object, completing the filling of the current chapter's main text area.

[0161] Step S550: Write the text content of the statements after the display order is arranged into the body paragraph area below the corresponding chapter title field in the preset document format specification. Each statement text content occupies an independent statement entry, and an entry number identifier is assigned to each independent statement entry.

[0162] Based on step S546, the document rendering engine applies the "body item style" defined in the preset document format specification to each independent statement item written into the body paragraph area. The body item style includes: font is SimSun or Dengxian, font size is 12pt or 12pt, first-line indent is 0 characters (using hanging indentation to align with the item symbol), left indent is 2 characters, paragraph spacing is 0 lines before and after, and line spacing is 1.5 lines. The document rendering engine sets the indentation and font attributes of paragraph objects through the document object model to ensure they conform to the specification requirements. The allocation of item number identifiers is implemented using an automatic numbering field. The document rendering engine calls the automatic numbering function of the document processing library to set a numbering format template for the item paragraphs, with the starting value starting from 1 and automatically incrementing for each new item.

[0163] Step S560: Perform format consistency verification and layout style application processing on the document with all content filled in, and generate a structured interview feedback document containing the specified font style, paragraph indentation style and item number style.

[0164] After completing the writing of chapter titles and the filling of text entries for all intermediate-level feedback branch nodes in the feedback item hierarchy structure tree, the document rendering engine performs final format consistency checks and global layout style applications on the entire document object. Format consistency checks include: checking whether the numbering sequence of all chapter titles is continuous, and automatically correcting any skipped or repeated numbers; checking whether the numbering sequence of all text entries is independent and continuous within the chapter; checking for violations of orphan control, and adjusting the pagination position of paragraphs spanning multiple pages; and checking whether the anchoring positions of floating objects such as images and tables (if present) are valid. Layout style applications include: applying a preset font embedding scheme to all levels of headings and text in the document to ensure consistent font display when the document is opened on different devices; setting the page footer content, which can include the document title abbreviation, page numbers, and generation date; and generating and inserting a table of contents, which is generated based on the outline level and outline text of each level of heading, with page numbers automatically calculated using field codes. After all the above processing steps, the document rendering engine finally outputs a structured interview feedback document that fully conforms to the preset document format specifications. The document format is either a portable document format or an open XML format for office documents. Each independent feedback paragraph in the structured interview feedback document maintains a content mapping relationship with a branch node in the feedback item hierarchy tree. Readers can clearly understand the comprehensive evaluation and specific feedback on the candidate's various ability dimensions by following the document's chapter hierarchy.

[0165] This application provides a computer-readable storage medium storing computer-executable instructions or a computer program. When the computer-executable instructions or the computer program are executed by a processor, the processor will execute the interview feedback structure generation method based on natural language processing provided in this application. For example, ... Figure 3 The method for generating structured interview feedback based on natural language processing is shown.

[0166] In some embodiments, the computer-readable storage medium may be a read-only memory (ROM), random access memory (RAM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory, magnetic surface memory, optical disk, or CD-ROM, etc.; or it may be a device that includes one or any combination of the above-mentioned memories.

[0167] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A method for generating structured interview feedback based on natural language processing, characterized in that, The method includes: A set of interview dialogue text sequences with speaking timestamps is obtained after the interview dialogue audio stream is converted by speech recognition. The set of interview dialogue text sequences includes interviewer question text fragments and applicant response text fragments with a sequential relationship. The set of interview dialogue text sequences is input into a pre-built dialogue flow state parser for dialogue behavior state transition deduction processing, and the output is a dialogue behavior state transition sequence corresponding to each timestamp node in the set of interview dialogue text sequences. The dialogue behavior state transition sequence includes the dialogue stage attribution label at each timestamp node and the dialogue state transition path identifier between adjacent timestamp nodes. A preset feedback mapping activation network is invoked to perform joint feedback item mapping processing on the dialogue behavior state transition sequence and the interview dialogue text sequence set, generating an initial feedback item mapping set corresponding to each semantic unit in the candidate's response text fragment. Each initial feedback item in the initial feedback item mapping set is associated with an activation intensity marker and a feedback category guidance marker in the feedback knowledge space. The initial feedback item mapping set is input into the feedback item hierarchical growth network for multi-level clustering and branching growth processing to generate a feedback item hierarchical structure tree that starts with the feedback root topic and ends with the specific feedback statement. In the feedback item hierarchical structure tree, there is a parent-child relationship between nodes at any adjacent level. The initial feedback item mapping set is document-rendered based on the feedback item hierarchy structure tree, and a structured interview feedback document conforming to the preset document format specification is output. Each independent feedback paragraph in the structured interview feedback document has a content mapping correspondence with a branch node in the feedback item hierarchy structure tree.

2. The method according to claim 1, characterized in that, The step of inputting the set of interview dialogue text sequences into a pre-built dialogue flow state parser for dialogue behavior state transition deduction processing, and outputting a dialogue behavior state transition sequence corresponding to each timestamp node in the set of interview dialogue text sequences, includes: Obtain the interviewer's question text fragments and the applicant's response text fragments from the interview dialogue text sequence set, and construct a round-pairing relationship mapping table between the interviewer's question text fragments and the applicant's response text fragments based on the speaking timestamp markers; The text fragments of the interviewer's questions and the applicant's responses, which have a round-pairing relationship mapping table, are input into the text understanding unit of the dialogue flow state parser. For each pair of text fragments with a round-pairing relationship mapping table, joint context semantic absorption processing is performed to generate a round-pairing dialogue representation vector carrying bidirectional dialogue context awareness. The dialogue round representation vector is input into the state classification unit of the dialogue flow state parser for dialogue stage type determination processing. The dialogue stage attribution label is assigned to each dialogue round representation vector, and the dialogue stage attribution label indicates the interview stage type of the current dialogue round. The dialogue stage attribution labels of adjacent dialogue rounds are input into the transition derivation unit of the dialogue flow state parser for state transition path derivation processing. Based on the combination pattern of the dialogue stage attribution labels of the previous dialogue round and the dialogue stage attribution labels of the next dialogue round, the dialogue state transition path identifier between the adjacent timestamp nodes is determined. Construct a time-ordered dialogue state transition link based on the dialogue stage attribution label and the dialogue state transition path identifier, and mark the timestamp information of each node in the dialogue state transition link to maintain the temporal alignment with the interview dialogue text sequence set. The dialogue state transition links that have completed temporal alignment are output as the dialogue behavior state transition sequence. Each transition node in the dialogue behavior state transition sequence contains a label indicating the current dialogue stage and a dialogue state transition path identifier pointing to the previous node.

3. The method according to claim 2, characterized in that, The text understanding unit of the dialogue flow state parser inputs the interviewer's question text fragments and the applicant's response text fragments, which have a round-pairing relationship mapping table, into the dialogue flow state parser. It then performs joint contextual semantic absorption processing on each pair of text fragments with the round-pairing relationship mapping table to generate a round-pairing dialogue representation vector carrying bidirectional dialogue context awareness, including: The interviewer's question text fragment with the round-pairing relationship mapping table is subjected to word segmentation to obtain a question word sequence, and each word in the question word sequence is injected with its relative position indicator in the interviewer's question text fragment to generate an initial word representation set for the question side. The candidate's response text fragment with the round-pairing relationship mapping table is subjected to word segmentation to obtain a response word sequence. Each word in the response word sequence is injected with its relative position indicator in the candidate's response text fragment to generate an initial word representation set for the response side. The initial vocabulary representation set of the question side is input into the first directional encoding subunit of the text understanding unit, and sequential semantic absorption processing is performed along the direction from the question start position to the question end position to generate a question side forward semantic absorption result sequence. The initial vocabulary representation set of the response side is input into the second direction encoding subunit of the text understanding unit, and reverse semantic absorption processing is performed along the direction from the end position of the response to the start position of the response to generate the backward semantic absorption result sequence of the response side. The forward semantic absorption result sequence of the question side and the backward semantic absorption result sequence of the response side are input into the interactive perception subunit of the text understanding unit to perform cross-sequence semantic association perception processing and generate a semantic association response intensity distribution map between question words and response words. Based on the semantic association response intensity distribution map, the forward semantic absorption result sequence of the questioning side and the backward semantic absorption result sequence of the response side are reorganized and processed to generate the round dialogue representation vector that simultaneously integrates the semantic information of the questioning side and the semantic information of the response side.

4. The method according to claim 3, characterized in that, The step of inputting the forward semantic absorption result sequence from the question side and the backward semantic absorption result sequence from the response side into the interactive perception subunit of the text understanding unit, performing cross-sequence semantic association perception processing, and generating a semantic association response intensity distribution map between question words and response words includes: The query-side forward semantic absorption result sequence is subjected to spatial mapping transformation processing, and each element in the query-side forward semantic absorption result sequence is mapped to the query representation space to generate the query-side query representation sequence. The response-side backward semantic absorption result sequence is subjected to spatial mapping transformation processing, and each element of the response-side backward semantic absorption result sequence is mapped to the key representation space and the value representation space respectively, generating the response-side key representation sequence and the response-side value representation sequence; Each query representation in the query representation sequence on the question side and each key representation in the key representation sequence on the response side are subjected to an association response metric to obtain an original association response matrix that characterizes the degree of association between the question words and the response words. The original association response matrix is ​​subjected to distribution normalization along the response vocabulary dimension, so that all association response values ​​corresponding to each question word in the original association response matrix form a normalized distribution, thereby generating a normalized association response matrix. Using the normalized correlation response matrix as a guiding signal, the response side value representation sequence is selectively aggregated to generate a response side context aggregation representation sequence for each question word; The response-side context aggregation representation sequence and the question-side forward semantic absorption result sequence are merged into channels to update the semantic representation content of the question-side forward semantic absorption result sequence. The correlation response intensity distribution at the corresponding position in the semantic correlation response intensity distribution map is then adjusted according to the updated semantic representation content.

5. The method according to claim 1, characterized in that, The method invokes a preset feedback mapping activation network to perform joint feedback item mapping processing on the dialogue behavior state transition sequence and the interview dialogue text sequence set, generating an initial feedback item mapping set corresponding to each semantic unit in the applicant's response text fragment, including: The dialogue behavior state transition sequence is input into the state awareness unit of the feedback mapping activation network, and the dialogue stage attribution label and dialogue state transition path identifier corresponding to each timestamp node in the dialogue behavior state transition sequence are extracted to generate dialogue stage evolution trajectory information. The candidate's response text fragments from the interview dialogue text sequence set are input into the response parsing unit of the feedback mapping activation network. The semantic role decomposition process is performed on the candidate's response text fragments, and the core semantic role units and the modifying semantic role units surrounding the core semantic role units are extracted to generate response semantic skeleton information. The dialogue stage evolution trajectory information and the response semantic skeleton information are input into the conditional mapping unit of the feedback mapping activation network. Based on the dialogue stage attribution label in the dialogue stage evolution trajectory information, the semantic role unit in the response semantic skeleton information is subjected to association strength adjustment processing to generate modulated semantic skeleton information constrained by the dialogue stage state conditions. Based on the combination relationship between semantic role units in the modulated semantic skeleton information, semantic matching retrieval processing is performed in the preset feedback knowledge space to activate candidate feedback item nodes that have a semantic correspondence with the modulated semantic skeleton information, thereby obtaining a set of activated candidate feedback item nodes. Determine the activation intensity marker of each candidate feedback item node in the set of activated candidate feedback item nodes in the feedback knowledge space. The activation intensity marker is used to characterize the degree of matching between the candidate feedback item node and the modulated semantic skeleton information. Based on the feedback category region to which each candidate feedback item node belongs in the feedback knowledge space, a feedback category guidance tag is assigned to each candidate feedback item node, generating the initial feedback item mapping set that simultaneously carries an activation intensity tag and a feedback category guidance tag.

6. The method according to claim 5, characterized in that, The step of inputting the dialogue stage evolution trajectory information and the response semantic skeleton information into the conditional mapping unit of the feedback mapping activation network, and performing association strength adjustment processing on the semantic role units in the response semantic skeleton information according to the dialogue stage attribution label in the dialogue stage evolution trajectory information to generate modulated semantic skeleton information constrained by the dialogue stage state conditions includes: Extract the dialogue stage attribution label at each timestamp node from the dialogue stage evolution trajectory information, and determine the relative stage sequence indicator of the dialogue stage attribution label in the preset interview stage sequence. The relative stage sequence indicator is used to query the preset stage-semantic role association strength configuration table to obtain the baseline association strength adjustment indication signal of different semantic role types under the current dialogue stage affiliation label condition. Extract each semantic role unit contained in the response semantic skeleton information, identify the semantic role type label of each semantic role unit, and the semantic role type label is used to distinguish the functional positioning of the semantic role unit in semantic expression; The semantic role type label of each semantic role unit is matched with the baseline association strength adjustment indication signal to determine the individual association strength adjustment indication component of each semantic role unit under the constraint of the attribution label in the current dialogue stage. The individual association strength adjustment indicator component is applied to the original connection weight field of the corresponding semantic role unit in the response semantic skeleton information, and the original connection weight field is updated and replaced to obtain the semantic role unit after updating the connection weight field. Based on all the semantic role units after updating the connection weight fields and their interdependent connection edges, the semantic skeleton topology is reconstructed to generate the modulated semantic skeleton information constrained by the dialogue stage state conditions.

7. The method according to claim 5, characterized in that, Based on the combination relationship between semantic role units in the modulated semantic skeleton information, semantic matching retrieval processing is performed in a preset feedback knowledge space to activate candidate feedback item nodes that have a semantic correspondence with the modulated semantic skeleton information, thereby obtaining a set of activated candidate feedback item nodes, including: The modulated semantic skeleton information is subjected to connected subgraph segmentation processing to identify the set of independent semantic connected subgraphs composed of semantic role units in the modulated semantic skeleton information. Each independent semantic connected subgraph represents a relatively complete semantic expression segment. The semantic role units in each independent semantically connected subgraph are linearly serialized and arranged according to the semantic dependency direction to generate a semantic role unit sequence representation, which preserves the sequential relationship between semantic role units. The semantic role unit sequence representation is converted into a retrieval semantic representation that can be searched in the feedback knowledge space, which pre-stores a large number of standard feedback item nodes and the index semantic representation of each standard feedback item node; The semantic similarity of the retrieval semantic representation is matched with the index semantic representation of all standard feedback item nodes in the feedback knowledge space, and the standard feedback item nodes whose semantic similarity matching results meet the preset activation conditions are selected as the candidate feedback item nodes. All the selected candidate feedback item nodes are organized in a set according to their storage location in the feedback knowledge space to form the set of activated candidate feedback item nodes. Each node in the set of activated candidate feedback item nodes inherits its node identifier and node description content in the feedback knowledge space. Record the subgraph identifier of the independent semantic connectivity subgraph on which each candidate feedback item node is activated, and establish a source association mapping relationship between the candidate feedback item node and the specific semantic fragment in the applicant's response text fragment.

8. The method according to claim 1, characterized in that, The process of inputting the initial feedback item mapping set into the feedback item hierarchical growth network for multi-level clustering and branching growth to generate a feedback item hierarchical structure tree starting with the feedback root topic and ending with specific feedback statements includes: Obtain the feedback category guidance tag carried by each initial feedback item in the initial feedback item mapping set, and perform initial clustering processing on the initial feedback item mapping set according to the feedback category guidance tag to generate a primary feedback item cluster set corresponding to different feedback category guidance tags; Each primary feedback item cluster in the primary feedback item cluster set is input into the knowledge anchoring layer of the feedback item hierarchical growth network, and is matched and aligned with the standard feedback category nodes in the preset feedback knowledge hierarchy framework to determine the anchoring position of each primary feedback item cluster in the feedback knowledge hierarchy framework. Based on the anchor mounting position, a hierarchical growth operation is performed on the primary feedback item clusters. Primary feedback item clusters with similar anchor mounting positions and sharing the same upper-level category are aggregated into intermediate-level feedback branch nodes. The intermediate-level feedback branch nodes aggregate the common feedback attribute features in their subordinate primary feedback item clusters. Extract the branch semantic topic description text of each intermediate level feedback branch node, and perform topic aggregation processing on the branch semantic topic description text of all intermediate level feedback branch nodes. Further aggregate the intermediate level feedback branch nodes with semantically compatible branch semantic topic description text to form the feedback root topic node. Construct a tree-like connection structure from the root feedback node down through the intermediate level feedback branch nodes to the leaf node corresponding to each initial feedback item in the primary feedback item cluster. In the tree-like connection structure, parent nodes and child nodes are connected by directed belonging edges. Based on the tree-like connection structure, a feedback item hierarchy structure tree is generated, which includes node identifiers, node level depth information, and parent-child node connection path information. Each leaf node in the feedback item hierarchy structure tree is a specific feedback statement.

9. The method according to claim 8, characterized in that, The step of inputting each primary feedback item cluster in the primary feedback item cluster set into the knowledge anchoring layer of the feedback item hierarchy growth network, and performing category matching and alignment processing with the standard feedback category nodes in the preset feedback knowledge hierarchy framework to determine the anchoring position of each primary feedback item cluster in the feedback knowledge hierarchy framework includes: The initial feedback items contained in the primary feedback item cluster are subjected to intra-cluster semantic aggregation representation processing to extract the cluster centroid semantic representation information that can represent the overall semantic direction of the primary feedback item cluster. The cluster centroid semantic representation information integrates the semantic content of each initial feedback item within the cluster. Extract the category definition description information of each standard feedback category node from the preset feedback knowledge hierarchy framework, and perform semantic representation transformation on the category definition description information to generate standard category semantic representation information corresponding to each standard feedback category node; The semantic alignment degree measurement process is performed between the cluster centroid semantic representation information and the standard category semantic representation information of each standard feedback category node to obtain the semantic alignment degree measurement result that characterizes the semantic fit between the primary feedback item cluster and each standard feedback category node. Based on the semantic alignment degree measurement result, all standard feedback category nodes are sorted in descending order, and the standard feedback category node with the highest semantic alignment degree measurement result is selected as the preliminary candidate anchor node of the primary feedback item cluster, and the highest semantic alignment degree measurement result is recorded as the anchoring reliability mark. If the anchoring reliability mark is not lower than the preset anchoring acceptance threshold, then the hierarchical path information and node coordinate information of the preliminary candidate anchoring node in the feedback knowledge hierarchy framework are assigned to the primary feedback item cluster as the anchoring mounting position of the primary feedback item cluster. If the anchoring reliability mark is lower than the preset anchoring acceptance threshold, the primary feedback item cluster is split into multiple smaller-granularity sub-feedback item clusters, and the category matching alignment process is re-executed for each sub-feedback item cluster until all sub-feedback item clusters obtain anchoring positions that meet the anchoring acceptance threshold.

10. A computer system, characterized in that, include: Memory is used to store executable instructions or computer programs. A processor, when executing computer-executable instructions or computer programs stored in the memory, implements the structured generation method for interview feedback based on natural language processing as described in any one of claims 1 to 9.

Images