A method and system for in-depth mining of job-seeking dialogue intentions applied to college student employment guidance

Abstract

This invention provides a method and system for deep mining of job-seeking dialogue intent applied to college student employment guidance. First, it acquires a set of dialogue data of target users in a job-seeking guidance scenario. This set of data contains multiple dialogue sequences arranged chronologically. Next, it extracts features from the dialogue data set to generate semantic association features and interactive behavior features. Then, based on a preset intent recognition network, it performs intent recognition on the semantic association features and interactive behavior features, generating a set of intent classification tags to indicate the distribution priority of users' potential job-seeking needs. Finally, it constructs an intent optimization strategy based on the intent classification tag set and feeds it back to the dialogue service module of the job-seeking guidance system to adjust the response content generation rules, thereby improving the service quality of the guidance system.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method and system for in-depth mining of job-seeking dialogue intent applied to college students' employment guidance. Background Technology

[0002] In the field of college student career guidance, traditional job-seeking guidance systems have many shortcomings in handling dialogues with students. Currently, most systems simply perform keyword matching on students' job-seeking inquiries and then generate response content based on preset templates. This approach lacks a deep understanding of students' dialogue intentions and fails to accurately grasp their potential job-seeking needs.

[0003] On the one hand, traditional systems fail to consider the semantic coherence between user job-seeking inquiries and the system's responses, leading to semantic disconnects in the dialogue and making it difficult for students to receive coherent and targeted guidance. On the other hand, they also neglect the intent triggering conditions and response time constraints corresponding to user job-seeking inquiries, causing the system's responses to potentially fail to meet students' actual needs and time expectations, thus reducing the effectiveness and efficiency of guidance.

[0004] Furthermore, existing systems cannot effectively mine and analyze students' dialogue data, nor can they determine the priority of students' potential job-seeking needs, thus hindering the provision of personalized and precise guidance services. Therefore, a new method is needed to deeply mine the intentions of college students in job-seeking dialogues in order to improve the performance and service quality of job-seeking guidance systems. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for in-depth mining of job-seeking dialogue intent applied to college student employment guidance, the method comprising:

[0006] Acquire a set of dialogue data of target users in a job search coaching scenario. The set of dialogue data contains multiple dialogue sequences arranged in chronological order. Each dialogue sequence consists of user job consultation statements and corresponding coaching system response statements.

[0007] Feature extraction is performed on the dialogue dataset to generate semantic association features and interactive behavior features for each dialogue sequence; wherein, the semantic association features characterize the semantic coherence between the user's employment consultation statement and the tutoring system's response statement, and the interactive behavior features characterize the intent triggering conditions and response timeliness constraints corresponding to the user's employment consultation statement.

[0008] Based on a preset intent recognition network, the semantic association features and the interaction behavior features are used to perform intent recognition, and an intent classification label set for the dialogue sequence is generated. The intent classification label set is used to indicate the distribution priority of the user's potential job search needs.

[0009] An intent optimization strategy is constructed based on the intent classification tag set, and the intent optimization strategy is fed back to the dialogue service module of the job search coaching system to adjust the response content generation rules.

[0010] In another aspect, embodiments of the present invention also provide a job-seeking dialogue service system, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to run the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0011] Based on the above, this embodiment of the invention comprehensively processes the dialogue data set of target users in job-seeking guidance scenarios, extracts features from multiple dimensions such as semantic coherence and interactive behavior characteristics, and accurately generates a set of intent classification labels for dialogue sequences using a preset intent recognition network. This determines the distribution priority of users' potential job-seeking needs, and on this basis, constructs an intent optimization strategy and feeds it back to the dialogue service module of the job-seeking guidance system. This effectively adjusts the response content generation rules, enabling the job-seeking guidance system to better understand students' potential job-seeking needs, provide more coherent, accurate, and personalized guidance services, comprehensively improve the system's performance and service quality, and enhance students' experience and effectiveness in the job-seeking guidance process. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the execution flow of the job-seeking dialogue intent deep mining method applied to college student employment guidance provided in an embodiment of the present invention.

[0013] Figure 2 This is a schematic diagram of the hardware architecture of the job search dialogue service system provided in an embodiment of the present invention. Detailed Implementation

[0014] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a method for in-depth mining of job-seeking dialogue intent applied to college student employment guidance, provided by an embodiment of the present invention. The following is a detailed description of this method for in-depth mining of job-seeking dialogue intent applied to college student employment guidance.

[0015] Step S110: Obtain a set of dialogue data of the target user in the job search guidance scenario. The set of dialogue data contains multiple dialogue sequences arranged in chronological order. Each dialogue sequence consists of the user's job consultation statement and the corresponding response statement from the guidance system.

[0016] In this embodiment, in the scenario of college student employment guidance, the source of dialogue data can be multifaceted, such as the message record database of the job search guidance system. This message record database continuously records the interaction content between the target user and the guidance system. Assume that the target user has had multiple interactions with the guidance system over a relatively long period, and the content of these interactions covers various aspects such as career planning, job skill requirements, and interview techniques.

[0017] For example, a user might ask a career advice question like, "For someone from a different background entering the field of artificial intelligence, besides self-studying relevant courses, are there more efficient ways to improve skills and gain recognition? Also, how should one weigh the development potential and stability of companies of different sizes in this field?" The tutoring system would then provide corresponding responses, such as, "For someone from a different background entering the field of artificial intelligence, besides self-studying courses, participating in professional training programs and obtaining relevant certifications is a relatively efficient approach. Regarding companies of different sizes, large companies offer high stability but are highly competitive, while small companies offer greater development potential but also carry relatively higher risks. It's necessary to weigh these factors based on your own career goals and risk tolerance."

[0018] Each user's job search inquiry and the coaching system's response is combined into a dialogue sequence, which is then arranged chronologically to form a dialogue dataset. This dataset records the various questions and answers received by the target user during the job search coaching process.

[0019] It is worth noting that when acquiring the dialogue data set of target users in the job search coaching scenario, we strictly followed relevant laws and regulations to ensure that we obtained the users' permission and authorization.

[0020] For example, before interacting with target users for job coaching, clearly state the purpose and use of the data collection through a statement or agreement. The statement should detail how the collected dialogue data is intended to optimize job coaching services, including but not limited to more accurately understanding users' job-seeking needs, providing personalized coaching suggestions, and improving the coaching system's response content generation rules. Simultaneously, inform users of the scope and methods of data use, ensuring they clearly understand how their data will be utilized.

[0021] For example, the data collection statement should be prominently displayed on the registration page or in the user agreement of the job coaching system. The statement would specifically mention: "We will collect your conversation data with the coaching system to analyze your job search needs and provide you with more tailored career planning advice, job recommendations, and interview skills coaching services. This data will only be used to optimize the system's services and will not be used for other unrelated commercial purposes."

[0022] To ensure that users' true intentions are obtained, an explicit authorization method is adopted. Before users begin using the job search coaching service, they will be required to confirm the data collection statement. This can be done by having users check an agreement box or click a confirmation button, allowing them to clearly indicate their consent to the system collecting and using their conversation data.

[0023] For example, during user registration, an authorization confirmation page will pop up, displaying the detailed content of the data collection statement and providing a checkbox indicating "I have read and agree to the above data collection statement." Data collection will only begin after the user checks the box and clicks the confirmation button.

[0024] As job search coaching services continue to develop and improve, the purpose and use of data collection may change. To ensure that users always understand and agree to the use of their data, the data collection statement will be updated periodically, and users will be required to reconfirm their authorization.

[0025] For example, every so often (e.g., every six months), an updated data collection statement is pushed to users, along with a link to reconfirm authorization. After clicking the link, users will see the updated statement and confirm their authorization again by checking an agreement box.

[0026] Furthermore, the system respects users' right to control their personal data and allows them to withdraw their authorization for data collection at any time. If a user decides to no longer allow the system to collect and use their conversation data, they can withdraw their authorization through the system's interface or by contacting customer service.

[0027] Once a user withdraws their authorization, the collection of new conversation data from that user will immediately cease, and the already collected data will be properly handled in accordance with relevant laws, regulations, and internal rules. For example, conversation data related to that user will be anonymized or deleted to ensure it will not be used for service optimization or other related purposes.

[0028] By taking the above measures, we ensure that we obtain the user's permission and authorization when collecting dialogue data sets, fully protect the user's privacy and legitimate rights and interests, and provide a legal and reliable data foundation for subsequent intent mining and service optimization.

[0029] Step S120: Extract features from the dialogue data set to generate semantic association features and interactive behavior features for each dialogue sequence; wherein, the semantic association features characterize the semantic coherence between the user's employment consultation statement and the tutoring system's response statement, and the interactive behavior features characterize the intent triggering conditions and response timeliness constraints corresponding to the user's employment consultation statement.

[0030] In this embodiment, to deeply analyze the information in the dialogue dataset, feature extraction is required to obtain semantic association features and interaction behavior features. Semantic association features help understand whether the user's questions and the tutoring system's answers are semantically coherent and matched; interaction behavior features help grasp the underlying intent triggers of the user's questions and the requirements for timely responses.

[0031] Step S121: Perform semantic segmentation on the user employment consultation statements in the dialogue sequence to obtain multiple semantic unit sets, each semantic unit set containing at least one lexical fragment with independent semantic expression capabilities.

[0032] In this embodiment, for user employment consultation statements in a dialogue sequence, the semantic segmentation process involves breaking down the statement into individual parts with independent semantic expression capabilities. Take, for example, the previously mentioned user employment consultation statement: "For those entering the field of artificial intelligence from a different major, besides self-studying relevant courses, are there more efficient ways to improve skills and gain recognition, while also considering how to balance the development potential and stability of companies of different sizes in this field in the future."

[0033] Segmentation can be performed using methods based on lexical and syntactic analysis. First, basic vocabulary is identified based on lexical rules, such as "cross-disciplinary," "enter," "artificial intelligence field," "self-study," "related courses," "more efficient," "skill improvement," "gain recognition," "pathways," "future," "companies of different sizes," "development potential," "stability," and "trade-offs." Then, based on syntactic structure and semantic relationships, these vocabulary words are combined into units with independent semantic meaning. For example, "cross-disciplinary entry into the field of artificial intelligence" can be considered a semantic unit, expressing the user's desired approach to entering this specific field; "self-studying related courses" is also a semantic unit, indicating the user's current learning methods; "more efficient ways to improve skills and gain recognition" is one of the core concerns of the user, constituting a semantic unit; and "the trade-off between the future development potential and stability of companies of different sizes in this field" is also an important semantic unit, reflecting the user's in-depth thinking about career development. Through this segmentation, multiple sets of semantic units are obtained, providing a foundation for subsequent encoding processing.

[0034] Step S122: Call the pre-trained context encoding model to perform serialization encoding on the semantic unit set to generate a semantic vector set of the user employment consultation statement, wherein each semantic vector in the semantic vector set corresponds to the context semantic representation of a semantic unit.

[0035] In this embodiment, after obtaining the set of semantic units, it is necessary to use a pre-trained context encoding model to serialize and encode them. The pre-trained context encoding model can be a model like BERT, which has been pre-trained on large-scale text data and can learn general semantic information of language.

[0036] Step S1221: Input the set of semantic units into the embedding layer of the context coding model for vector mapping processing to obtain an initial set of semantic vectors.

[0037] In this embodiment, the set of semantic units is first input into the embedding layer of the context coding model. The role of the embedding layer is to map each semantic unit to a low-dimensional vector space, so that semantically similar units are closer together in the vector space. Taking the semantic units obtained from the previous segmentation as an example, such as "entering the field of artificial intelligence from a different major" and "self-studying related courses", the embedding layer will generate a corresponding vector for each semantic unit. Assuming that the output dimension of the embedding layer is d, then each semantic unit will be converted into a d-dimensional vector. For the semantic unit "entering the field of artificial intelligence from a different major", after processing by the embedding layer, a d-dimensional vector v1 will be obtained; for "self-studying related courses", another d-dimensional vector v2 will be obtained. And so on, after performing such mapping processing on all semantic units, an initial set of semantic vectors is obtained, which contains the initial vectors corresponding to all semantic units.

[0038] Step S1222: Input the initial semantic vector set into the bidirectional long short-term memory network layer of the context encoding model for context dependency modeling to obtain an intermediate semantic vector set; wherein, the bidirectional long short-term memory network layer includes a forward propagation subnetwork and a backward propagation subnetwork, the forward propagation subnetwork is used to capture the positive dependency relationship between the current semantic unit and the historical semantic unit, and the backward propagation subnetwork is used to capture the negative dependency relationship between the current semantic unit and the future semantic unit.

[0039] In this embodiment, after obtaining the initial set of semantic vectors, it is input into the Bidirectional Long Short-Term Memory (Bi-LSTM) layer of the context encoding model. The Bi-LSTM layer consists of a forward propagation subnetwork and a backward propagation subnetwork, which can capture the dependencies between semantic units from different directions.

[0040] The forward propagation subnetwork starts from the beginning of the sequence and processes the vector of each semantic unit sequentially. For the current semantic unit vector, it incorporates information from previously processed semantic unit vectors to update the representation of the current vector. For example, when processing the vector of the semantic unit "more efficient ways to improve skills and gain recognition," the forward propagation subnetwork considers the information carried by previous semantic unit vectors such as "entering the field of artificial intelligence from a different major" and "self-studying relevant courses," thereby better understanding the contextual meaning of this semantic unit in the entire sentence.

[0041] The backpropagation subnetwork starts from the end of the sequence and processes the semantic unit vectors in reverse. It can capture the dependencies between the current semantic unit and subsequent semantic units. For example, when processing the semantic unit vector "entering the field of artificial intelligence from another major", the backpropagation subnetwork will refer to information from subsequent semantic unit vectors such as "self-studying relevant courses" and "more efficient ways to improve skills and gain recognition".

[0042] Through the combined action of the forward and backward propagation subnetworks, the vector of each semantic unit can be integrated with its contextual information to obtain an updated vector representation. Combining the vectors processed by all semantic units through the Bi-LSTM layer yields the intermediate semantic vector set.

[0043] Step S1223: Perform multi-head self-attention processing on the intermediate semantic vector set to generate an attention weight matrix. Then, perform a weighted summation on the intermediate semantic vector set based on the attention weight matrix to obtain an enhanced semantic vector set.

[0044] In this embodiment, after obtaining the intermediate semantic vector set, it needs to be subjected to multi-head self-attention processing. The multi-head self-attention mechanism allows the model to focus on different relationships between semantic units in different representation subspaces.

[0045] First, for each vector in the intermediate semantic vector set, its query, key, and value vectors are calculated separately. Assuming the intermediate semantic vector set is {v1, v2, ..., vn}, for vector vi, a linear transformation is performed to obtain the corresponding query vector Qi, key vector Ki, and value vector Vi. Then, the similarity between the query vector Qi and all key vectors Kj (j=1, 2, ..., n) is calculated, commonly using dot product operations. After obtaining the similarity matrix, it is normalized to obtain the attention weight matrix.

[0046] The attention weight matrix reflects the degree of attention each semantic unit vector pays to other semantic unit vectors. For example, when considering the semantic unit "more efficient ways to improve skills and gain recognition", the attention weight matrix might show its attention to semantic units such as "self-studying relevant courses" and "cross-disciplinary entry into the field of artificial intelligence".

[0047] Based on the attention weight matrix, a weighted sum is performed on the intermediate semantic vector set. For each semantic unit vector, it is multiplied by the corresponding weight in the attention weight matrix, and then all results are summed to obtain the enhanced vector representation. Combining all the enhanced vectors forms the enhanced semantic vector set. The vectors in this enhanced semantic vector set highlight the important relationships between semantic units and contain richer contextual information.

[0048] Step S1224: Input the enhanced semantic vector set into the normalization layer of the context encoding model for feature scaling processing to obtain the semantic vector set of the user's employment consultation statement.

[0049] In this embodiment, after obtaining the enhanced semantic vector set, it is input into the normalization layer of the context coding model. The normalization layer performs feature scaling on the vectors in the enhanced semantic vector set, so that the feature values ​​of the vectors are within a suitable range, which helps the model training and convergence.

[0050] A commonly used normalization method is layer normalization. For each vector in the augmented semantic vector set, layer normalization calculates its mean and variance, and then standardizes the vector based on the mean and variance. Specifically, for vector v, its mean μ and variance σ² are first calculated, and then normalized using the formula... The vector is normalized, where ε is a very small constant used to avoid cases where the denominator is zero.

[0051] After normalization, a set of semantic vectors representing user employment consultation statements is obtained. The vectors in this set exhibit better numerical stability and can be used more effectively for subsequent semantic association analysis.

[0052] Step S123: Perform bidirectional attention matching processing based on the semantic vector set of the user's employment consultation statement and the semantic vector set of the tutoring system's response statement to generate a semantic association matrix between the user's employment consultation statement and the tutoring system's response statement.

[0053] In this embodiment, after obtaining the semantic vector sets of the user's employment consultation statements and the semantic vector sets of the tutoring system's response statements, bidirectional attention matching is required to generate a semantic association matrix. Bidirectional attention matching allows us to understand the degree of semantic correlation between the user's questions and the tutoring system's responses.

[0054] Step S1231: Use the semantic vector set of the user's employment consultation statement as the query vector set, and use the semantic vector set of the tutoring system's answer statement as the key vector set.

[0055] In this embodiment, for bidirectional attention matching, the semantic vector set of the user's employment consultation statement is defined as the query vector set, denoted by Q; the semantic vector set of the tutoring system's response statement is defined as the key vector set, denoted by K. For example, the semantic vector set of the user's employment consultation statement is Q = {q1, q2, ..., qm}, where qi represents the semantic vector of the i-th semantic unit in the user's employment consultation statement; the semantic vector set of the tutoring system's response statement is K = {k1, k2, ..., kn}, where kj represents the semantic vector of the j-th semantic unit in the tutoring system's response statement.

[0056] Step S1232: Calculate the cosine similarity between each query vector in the query vector set and each key vector in the key vector set, and generate an initial similarity matrix.

[0057] In this embodiment, for each query vector qi in the query vector set Q and each key vector kj in the key vector set K, the cosine similarity between them is calculated. The cosine similarity is obtained by dividing the dot product of the two vectors by the product of their magnitudes. Specifically, for vectors qi and kj, the cosine similarity sim(qi, kj) = (qi·kj) / (||qi||||kj||), where · represents the dot product operation, and ||·|| represents the magnitude of the vector.

[0058] Combining the cosine similarity calculation results between all query vectors and key vectors yields the initial similarity matrix S. The element Sij in the i-th row and j-th column of matrix S represents the cosine similarity between query vector qi and key vector kj. This initial similarity matrix reflects the preliminary degree of similarity between semantic units in the user's employment consultation statement and semantic units in the tutoring system's response statement.

[0059] Step S1233: Perform row direction normalization on the initial similarity matrix to obtain the query attention weight matrix.

[0060] In this embodiment, after obtaining the initial similarity matrix S, row direction normalization is performed on it. The purpose of row direction normalization is to make the sum of the elements in each row equal to 1, so that each row element can be converted into a weight value, representing the degree of attention of the query vector to the key vector.

[0061] For the i-th row of the initial similarity matrix S, calculate the sum of the elements in that row, Si_sum = ΣjSij. Then, divide each element Sij in that row by Si_sum to obtain the normalized element Aij = Sij / Si_sum. Performing this normalization process on all rows yields the query attention weight matrix A. The i-th row of the query attention weight matrix A represents the attention allocation of the query vector qi to each key vector in the key vector set.

[0062] Step S1234: Perform column direction normalization on the initial similarity matrix to obtain the key attention weight matrix.

[0063] In this embodiment, similar to row-direction normalization, column-direction normalization is performed on the initial similarity matrix S. Column-direction normalization is performed to ensure that the sum of the elements in each column is 1, thus obtaining the key attention weight matrix B.

[0064] For the j-th column of the initial similarity matrix S, calculate the sum of the elements in that column, Sj_sum = ΣiSij. Then, divide each element Sij in that column by Sj_sum to obtain the normalized element Bij = Sij / Sj_sum. Performing this normalization process on all columns yields the key attention weight matrix B. The j-th column of the key attention weight matrix B represents the attention allocation of the key vector kj to each query vector in the query vector set.

[0065] Step S1235: Multiply the query attention weight matrix and the key attention weight matrix element by element to generate a bidirectional attention weight matrix.

[0066] In this embodiment, after obtaining the query attention weight matrix A and the key attention weight matrix B, they are multiplied element-wise. For corresponding elements Aij and Bij in matrices A and B, multiplication yields a new element Cij = AijBij. This multiplication process is repeated for all corresponding elements to obtain the bidirectional attention weight matrix C. The bidirectional attention weight matrix integrates the attention of the query vector to the key vector and the attention of the key vector to the query vector, thus more comprehensively reflecting the semantic relationship between the user's employment consultation statement and the tutoring system's response statement.

[0067] Step S1236: Perform weighted aggregation processing on the key vector set according to the bidirectional attention weight matrix to generate a context-aware vector set corresponding to the query vector set.

[0068] In this embodiment, a weighted aggregation process is performed on the key vector set K using a bidirectional attention weight matrix C. For each query vector qi in the query vector set Q, the key vectors in the key vector set K are weighted and summed according to the i-th row element {Cij} of the bidirectional attention weight matrix C. Specifically, for the query vector qi, the corresponding context-aware vector ci = ΣjCijkj.

[0069] By combining the context-aware vectors corresponding to all query vectors, we obtain the context-aware vector set C' corresponding to the query vector set. Each vector in the context-aware vector set incorporates information from relevant semantic units in the tutoring system's response statement, reflecting the representation of the user's employment consultation statement within the context of the tutoring system's response statement.

[0070] Step S1237: Concatenate the context-aware vector set with the query vector set to obtain an enhanced context vector set.

[0071] In this embodiment, the context-aware vector set C' and the query vector set Q are concatenated. The concatenation is performed by linking the context-aware vectors and their corresponding query vectors sequentially. For example, for the query vector qi and its corresponding context-aware vector ci, they are concatenated into a new vector ei = [qi; ci], where ; denotes the vector concatenation operation.

[0072] Combining all the concatenated vectors yields the enhanced context vector set E. The vectors in this set contain both the semantic information of the user's employment consultation statements and the contextual information of the tutoring system's responses, providing richer features for the subsequent generation of the semantic association matrix.

[0073] Step S1238: Perform a fully connected transformation on the enhanced context vector set to generate the semantic association matrix.

[0074] In this embodiment, after obtaining the enhanced context vector set E, a fully connected transformation is performed on it. The fully connected transformation maps each vector in the enhanced context vector set to a new vector space through a linear transformation.

[0075] Suppose that the vector dimension in the augmented context vector set E is d1, and the output dimension of the fully connected layer is d2. For each vector ei in the augmented context vector set E, a new vector ri = Wei + b is obtained by linearly transforming it through a weight matrix W (dimension d2 × d1) and a bias vector b (dimension d2).

[0076] Combining all the transformed vectors forms the semantic association matrix R. The semantic association matrix R reflects the degree of semantic association between the user's employment consultation statement and the counseling system's response, providing important feature information for subsequent intent recognition.

[0077] Step S124: Perform max pooling on the semantic association matrix and extract the semantic vector corresponding to the maximum response association value in the semantic association matrix as the semantic association feature.

[0078] In this embodiment, after obtaining the semantic association matrix R, max pooling is performed on it. The purpose of max pooling is to extract the most representative information from the semantic association matrix.

[0079] For each row of the semantic association matrix R, find the maximum value in that row and its corresponding column index. Assume the maximum value in the i-th row is Rimax, and the corresponding column index is jmax. Then, extract the semantic vector containing the element in the jmax-th column of the i-th row of the semantic association matrix R as the semantic vector corresponding to the maximum response association value.

[0080] By combining the semantic vectors corresponding to the maximum response correlation values ​​of all rows, we obtain the semantic correlation features. These features highlight the closest semantic connections between the user's employment consultation statements and the tutoring system's responses, helping to more accurately identify the user's intent.

[0081] Step S125: Perform behavioral pattern analysis on the user employment consultation statements in the dialogue sequence, identify the interaction triggering conditions and response timeliness constraint vectors of the user employment consultation statements, and concatenate the interaction triggering conditions and response timeliness constraint vectors to generate the interaction behavior features.

[0082] In this embodiment, in order to fully understand the intent behind the user's employment consultation statements and the requirements for the response, it is necessary to perform behavioral pattern analysis and processing, extract the interaction trigger conditions and response timeliness constraint vectors, and concatenate them into interactive behavior features.

[0083] For example, step S1251: perform word-by-word matching processing on the user's employment consultation statement according to the preset interactive trigger word library, and generate a set of trigger word existence tags, wherein each trigger word existence tag indicates the existence status of the corresponding trigger word in the user's employment consultation statement.

[0084] In this embodiment, the preset interaction trigger word library is a set of words that can trigger specific interaction intentions. For a user's employment consultation statement, each word needs to be matched against this interaction trigger word library. Assume the interaction trigger word library is T, which contains multiple trigger words t1, t2, ..., tn. For each word in the user's employment consultation statement, check whether it exists in the interaction trigger word library T. If it exists, the corresponding trigger word is marked as a state, such as "exists"; if it does not exist, it is marked as another state, such as "does not exist".

[0085] Taking the user's job consultation statement, "For those entering the field of artificial intelligence from a different major, besides self-studying relevant courses, are there more efficient ways to improve skills and gain recognition? Also, considering the future development potential and stability of companies of different sizes in this field, how should one weigh these factors?" as an example, assuming the interaction trigger word library T contains trigger words such as "from a different major," "improve skills," and "development potential," when performing word-by-word matching on this statement, if "from a different major" is encountered and found to exist in the interaction trigger word library T, then the trigger word for "from a different major" is marked as "existent." However, for the word "for," it is not in the interaction trigger word library T, so its corresponding trigger word is marked as "not existing." Combining the existence marks corresponding to all trigger words generates a set of trigger word existence marks. This set of trigger word existence marks can intuitively reflect which preset interaction trigger words are contained in the user's job consultation statement, providing a basis for subsequently determining the interaction trigger conditions.

[0086] Step S1252: Extract the set of time-sensitive words from the user's employment consultation statement, and calculate the time sensitivity attenuation coefficient based on the position information of the set of time-sensitive words in the dialogue sequence.

[0087] In this embodiment, time-sensitive words refer to those words that reflect the user's requirements or concerns about time. These time-sensitive words are extracted from the user's job consultation statements to form a time-sensitive word set. For example, time-sensitive words such as "as soon as possible," "soon," and "for some time in the future" may appear in the user's job consultation statements, and these are extracted to form a set S.

[0088] Then, a time sensitivity decay coefficient is calculated based on the positional information of these time-sensitive words in the dialogue sequence. Positional information can include the sequential position of the time-sensitive words within the sentence or the time interval between them and the start of the dialogue. It is assumed that the earlier a time-sensitive word appears in the dialogue sequence, the higher the user's time sensitivity; as time progresses or its position in the sentence increases, the time sensitivity gradually decreases.

[0089] The time sensitivity decay coefficient can be calculated in a relative way. For example, the dialogue sequence can be viewed as a linear sequence, and a weight can be assigned based on the position of the time-sensitive word within the sequence. Let the position of the time-sensitive word in the dialogue sequence be p, and the length of the entire dialogue sequence be L. A decay function f(p, L) can be defined to calculate the time sensitivity decay coefficient. For example, when the time-sensitive word is at the beginning of the dialogue sequence, p=1, and the time sensitivity decay coefficient is relatively high; when p is close to L, the time sensitivity decay coefficient is relatively low. This calculation quantifies the user's sensitivity to time, providing key parameters for subsequently generating response timeliness constraint vectors.

[0090] Step S1253: Perform a weighted summation on the set of trigger word presence markers to generate an interactive trigger condition vector, wherein the weight of each trigger word presence marker is determined based on a predefined trigger word priority table.

[0091] In this embodiment, after obtaining the set of trigger word existence tags, a weighted summation process is required to generate an interaction trigger condition vector. A predefined trigger word priority table assigns different weights to each trigger word, reflecting the importance of the trigger word in triggering the interaction intent.

[0092] Suppose the set of trigger word existence markers is M, which contains multiple trigger word existence markers m1, m2, ..., mn, and their corresponding predefined weights are w1, w2, ..., wn. For each trigger word existence marker, if its state is "existent", the corresponding weight is added during the weighted summation; if its state is "not present", the corresponding weight is not added.

[0093] For example, given a trigger word marker m1, if it indicates the presence of the trigger word "cross-disciplinary" and its predefined weight is w1, then w1 is added to the sum during weighted summation. If m2 indicates the absence of a trigger word, its corresponding weight w2 is not included in the summation. Summing all the weights involved in the summation yields the interaction trigger condition vector. This vector comprehensively considers the importance of different trigger words, more accurately reflecting the interaction conditions triggered by the user's employment consultation statements.

[0094] Step S1254: Perform linear transformation processing based on the time sensitivity attenuation coefficient and the predefined time sensitivity benchmark value to generate a response timeliness constraint vector.

[0095] In this embodiment, the time sensitivity attenuation coefficient has been calculated, and a predefined time sensitivity benchmark value exists. These two parameters are combined through a linear transformation to generate a response timeliness constraint vector.

[0096] Let the time sensitivity attenuation coefficient be *a*, and the predefined time sensitivity baseline value be *b*. A linear transformation can be achieved using a linear function, assuming the linear function is g(a, b) = ka + cb, where k and c are predefined coefficients. This linear function is used to combine and calculate the time sensitivity attenuation coefficient and the time sensitivity baseline value.

[0097] For example, when the time sensitivity decay coefficient 'a' is high, it indicates that the user is more sensitive to time. After linear transformation, the value of the generated response timeliness constraint vector will also be larger, indicating a need for a more timely response. Conversely, when 'a' is low, the value of the response timeliness constraint vector is smaller, indicating a relatively relaxed requirement for response time. In this way, the generated response timeliness constraint vector can accurately reflect the user's constraint requirements for response timeliness.

[0098] Step S1255: Concatenate the interaction trigger condition vector and the response timeliness constraint vector to generate the interaction behavior feature.

[0099] In this embodiment, after obtaining the interaction trigger condition vector and the response timeliness constraint vector, they are concatenated to obtain the interaction behavior features. Feature concatenation involves joining two vectors in a predetermined order to form a new vector.

[0100] Assume the interaction trigger condition vector is V1 with dimension d1, and the response timeliness constraint vector is V2 with dimension d2. Concatenating them yields an interaction behavior feature vector V with dimension d1 + d2. For example, if V1 = [v11, v12, ..., v1d1] and V2 = [v21, v22, ..., v2d2], then the concatenated interaction behavior feature vector V = [v11, v12, ..., v1d1, v21, v22, ..., v2d2]. This interaction behavior feature vector integrates the interaction trigger conditions and response timeliness constraints of the user's employment consultation statement, providing a more comprehensive feature basis for subsequent intent recognition.

[0101] Step S130: Based on a preset intent recognition network, perform intent recognition on the semantic association features and the interaction behavior features to generate an intent classification label set for the dialogue sequence. The intent classification label set is used to indicate the distribution priority of the user's potential job search needs.

[0102] In this embodiment, the preset intent recognition network is a trained neural network model used to process semantic association features and interaction behavior features, thereby identifying the user's intent and generating a set of intent classification labels. This set of intent classification labels can reflect the distribution priority of the user's potential job search needs, helping the job search coaching system to better understand user needs.

[0103] Step S131: Input the semantic association features into the feature fusion layer of the intent recognition network for linear transformation processing to generate the first transformed features.

[0104] In this embodiment, semantic association features are input into the feature fusion layer of the intent recognition network. The main function of the feature fusion layer is to perform a linear transformation on the input features. Assuming the semantic association feature is F1, the feature fusion layer has a weight matrix W1 and a bias vector b1. The semantic association feature is transformed using the linear transformation formula F1_transformed=W1F1+b1.

[0105] The linear transformation here maps semantic association features to a new feature space, making the features more conducive to subsequent processing and analysis. After the linear transformation, the first transformed feature F1_transformed is obtained. This first transformed feature contains adjusted and optimized semantic association information, providing a better foundation for subsequent feature fusion and intent recognition.

[0106] Step S132: Input the interactive behavior features into the feature fusion layer for nonlinear transformation processing to generate the second transformation feature.

[0107] In this embodiment, the interactive behavior features are input into the feature fusion layer of the intent recognition network for nonlinear transformation processing. Unlike linear transformation, nonlinear transformation can introduce more nonlinear relationships, enhancing the model's ability to express complex features.

[0108] Assuming the interaction behavior feature is F2, the feature fusion layer uses a non-linear function, such as the ReLU function (Rectified Linear Unit). First, the feature fusion layer has a weight matrix W2 and a bias vector b2, which are linearly combined to obtain the intermediate result z = W2F2 + b2. Then, the intermediate result z is input into the ReLU function, i.e., F2_transformed = ReLU(z). Through this non-linear transformation, the interaction behavior feature is converted into a second transformed feature F2_transformed. The second transformed feature can capture the non-linear information in the interaction behavior feature, better reflecting the complex intentions in the user's employment consultation statements.

[0109] Step S133: Perform gated weight calculation processing on the first transformation feature and the second transformation feature to generate dynamic fusion weight coefficients; wherein, the gated weight calculation processing includes: inputting the first transformation feature and the second transformation feature into the sigmoid function to generate a weight allocation ratio.

[0110] In this embodiment, in order to reasonably fuse the first transformation feature and the second transformation feature, gating weight calculation is required to generate dynamic fusion weight coefficients. The gating weight calculation uses the sigmoid function to generate the weight allocation ratio.

[0111] First, the first transform feature F1_transformed and the second transform feature F2_transformed are combined to form a new input vector. Assume they are concatenated to obtain the vector F_combined = [F1_transformed; F2_transformed]. Then, a linear transformation maps F_combined to a one-dimensional space. Let the weight vector be W_gate and the bias value be b_gate, resulting in the intermediate result z_gate = W_gateF_combined + b_gate.

[0112] Next, the intermediate result `z_gate` is input into the `sigmoid` function. The `sigmoid` function maps the input value to a range between 0 and 1, generating a weight allocation ratio. Let the generated weight allocation ratio be α, i.e., α = sigmoid(z_gate). This weight allocation ratio α is the dynamic fusion weight coefficient, which represents the weight of the first transform feature when fusing the first and second transform features. Therefore, the weight of the second transform feature is 1 - α. Through this dynamic weight allocation, the fusion ratio can be automatically adjusted according to different input features, making the fused features more accurately reflect the user's intent.

[0113] Step S134: The first transformation feature and the second transformation feature are weighted and summed according to the dynamic fusion weight coefficient to generate the fusion intention feature.

[0114] In this embodiment, after obtaining the dynamic fusion weight coefficient α, the first transformation feature F1_transformed and the second transformation feature F2_transformed are weighted and summed according to it to generate the fusion intention feature.

[0115] The formula for calculating the fused intent feature F_fused is F_fused = αF1_transformed + (1-α)F2_transformed. This weighted summation process effectively fuses semantic association features and interactive behavior features. A larger α indicates a greater weighting of semantic association features in the fused intent feature; a smaller α indicates a greater influence from interactive behavior features. The resulting fused intent feature combines the advantages of both features, more comprehensively reflecting the intent information contained in the user's employment consultation statement.

[0116] Step S135: Input the fused intent features into the classifier layer of the intent recognition network for multi-label classification processing to generate the intent classification label set, where each intent classification label corresponds to a confidence score for a job search demand type.

[0117] In this embodiment, the fused intent features are input into the classifier layer of the intent recognition network for multi-label classification. The classifier layer maps the fused intent features to different job requirement types and generates a confidence score for each type, thereby obtaining a set of intent classification labels.

[0118] Step S1351: Divide the fused intent features into multiple feature subspaces, each feature subspace corresponding to a preset job search intent category.

[0119] In this embodiment, to perform multi-label classification, the fused intent features need to be segmented into multiple feature subspaces. The preset job-seeking intent categories are some predefined job-seeking demand types, such as career planning, job skills improvement, salary and benefits consultation, etc.

[0120] Assume the fusion intent feature is F_fused, with dimension d. Based on the preset number of job intention categories n, F_fused is divided into n feature subspaces. A linear partitioning method can be used to allocate the dimensions of F_fused to each feature subspace according to a set rule. For example, the first d1 dimensions of F_fused are assigned to the first feature subspace, corresponding to the first job intention category; the next d2 dimensions are assigned to the second feature subspace, corresponding to the second job intention category, and so on. In this way, each feature subspace contains feature information related to a specific job intention category.

[0121] Step S1352: Perform independent fully connected transformation processing on each feature subspace to generate an initial prediction score for each job application intention category.

[0122] In this embodiment, each segmented feature subspace undergoes an independent fully connected transformation. The fully connected transformation maps the feature subspace to a one-dimensional score value through a linear transformation.

[0123] Suppose the i-th feature subspace is F_subspace_i, with dimension di. The fully connected layer has a weight matrix W_i and a bias vector b_i. Using the linear transformation formula score_i = W_iF_subspace_i + b_i, the i-th feature subspace is converted into an initial predicted score score_i. This process is performed on all feature subspaces to obtain the initial predicted score for each job intention category. These initial predicted scores reflect the preliminary prediction results of the fused intention features across each job intention category.

[0124] Step S1353: Sort the initial prediction scores to obtain a prediction score sequence.

[0125] In this embodiment, after obtaining the initial predicted score for each job intention category, these scores are sorted. The purpose of sorting is to clearly understand the relative importance of each job intention category.

[0126] All initial predicted scores, score_1, score_2, ..., score_n, are sorted in descending order to obtain a predicted score sequence. For example, the sorted predicted score sequence is score_sorted_1>= score_sorted_2>= ...>= score_sorted_n. This predicted score sequence visually displays the order of predicted scores for each job intention category, providing a foundation for subsequent filtering and labeling.

[0127] Step S1354: Binarize the predicted score sequence according to a preset score threshold, and mark the job intention category corresponding to the predicted score that exceeds the score threshold as a valid intention label.

[0128] In this embodiment, a preset scoring threshold is used to filter out valid job intention categories. The predicted score sequence is binarized to determine whether each predicted score exceeds the scoring threshold.

[0129] Assuming a score threshold of "threshold", for each score "score_sorted_i" in the predicted score sequence, if "score_sorted_i" > "threshold", its corresponding job intention category is marked as a valid intent label; if "score_sorted_i" <= "threshold", it is not marked. This binarization process filters out job intention categories with higher predicted scores that are more likely to represent the user's true intent, reducing unnecessary interference.

[0130] Step S1355: Normalize the confidence scores of all valid intent labels to generate the intent classification label set.

[0131] In this embodiment, after obtaining valid intent labels, their confidence scores need to be normalized. The purpose of normalization is to adjust the confidence scores to a uniform range, facilitating comparison and analysis.

[0132] Suppose the set of confidence scores for valid intent tags is S = {s1, s2, ..., sm}, where m is the number of valid intent tags. Normalization can be achieved by calculating the proportion of each confidence score in the total confidence scores of all valid intent tags. Let the set of normalized confidence scores be S_normalized. For each confidence score si, its normalized score s_normalized_i = si / Σjsj (j from 1 to m). Combining all normalized confidence scores with their corresponding job search intent categories generates a set of intent category tags. Each intent category tag in this set corresponds to a confidence score for a type of job search need, and these confidence scores are between 0 and 1, accurately reflecting the distribution priority of users' potential job search needs.

[0133] Step S140: Construct an intent optimization strategy based on the intent classification tag set, and feed the intent optimization strategy back to the dialogue service module of the job search coaching system to adjust the response content generation rules.

[0134] In this embodiment, after obtaining the intent classification tag set, an intent optimization strategy needs to be constructed based on it, and the strategy is fed back to the dialogue service module of the job search coaching system to adjust the response content generation rules of the dialogue service module in order to better meet the user's job search needs.

[0135] Step S141: Extract the target intent label with the highest confidence score from the intent classification label set, and identify the job search demand type corresponding to the target intent label.

[0136] In this embodiment, the intent classification label set includes multiple intent classification labels and their corresponding confidence scores. By comparing these confidence scores, the intent label corresponding to the highest score is identified and used as the target intent label.

[0137] Assume the set of intent category labels is L = {l1, l2, ..., ln}, and the corresponding set of confidence scores is C = {c1, c2, ..., cn}. By traversing the set of confidence scores C, we find the maximum value cmax and its corresponding index i, then the target intent label is li. Then, based on a predefined mapping relationship, we identify the job-seeking need type corresponding to the target intent label li, such as career planning or job skills enhancement. This job-seeking need type represents the user's current most pressing job-seeking need, providing crucial information for subsequent matching of coaching resources.

[0138] Step S142: Match the preset tutoring resource template library according to the job search demand type to obtain a set of tutoring content fragments associated with the job search demand type.

[0139] In this embodiment, the preset tutoring resource template library is a collection containing various job-seeking tutoring content, which is categorized and stored according to different job-seeking need types. A matching operation is performed within the tutoring resource template library based on the identified job-seeking need type. Assume the preset tutoring resource template library is R, containing multiple tutoring resource categories R1, R2, ..., Rk, each corresponding to a job-seeking need type. Once the job-seeking need type is determined, the corresponding tutoring resource category Ri is found. This category Ri contains multiple tutoring content fragments, which are detailed tutoring information prepared specifically for that job-seeking need type. For example, if the job-seeking need type is career planning, the corresponding tutoring resource category might contain tutoring content fragments on how to conduct self-assessment, how to set career goals, and the development trends of different industries. Through this matching operation, a set of tutoring content fragments associated with that job-seeking need type is obtained, denoted as F. Each tutoring content fragment in this set F is related to the user's current most pressing job-seeking need, providing rich material for subsequent content recommendations.

[0140] Step S143: Perform priority sorting on the set of tutoring content segments to generate a content recommendation sequence; wherein, the priority sorting process includes: calculating a comprehensive weight value based on the historical click-through rate and user feedback rating of the tutoring content segments and performing priority sorting based on the comprehensive weight value.

[0141] In this embodiment, after obtaining the set F of tutoring content fragments, it is necessary to prioritize them to generate a content recommendation sequence. The ranking is based on the historical click-through rate and user feedback rating of the tutoring content fragments. First, for each tutoring content fragment fj in set F, there is a corresponding historical click-through rate hj and user feedback rating sj. The historical click-through rate hj reflects the frequency with which the tutoring content fragment has been clicked and viewed by users in the past, while the user feedback rating sj ​​reflects the user's satisfaction level with the tutoring content fragment.

[0142] To comprehensively consider historical click-through rates and user feedback ratings, a comprehensive weight value (wj) needs to be calculated for each tutoring content segment. This comprehensive weight value can be calculated using a weighted combination method, for example, wj = αhj + βsj, where α and β are pre-set weight coefficients, and α + β = 1. The values ​​of α and β are adjusted according to specific circumstances to balance the importance of historical click-through rates and user feedback ratings in the comprehensive weight value.

[0143] After calculating the overall weight value of each tutoring content segment, the set of tutoring content segments F is sorted according to these overall weight values. Tutoring content segments with higher overall weight values ​​are ranked first, and those with lower overall weight values ​​are ranked last, thus generating a content recommendation sequence P. This content recommendation sequence arranges the tutoring content segments according to priority, which can better meet the needs of users and prioritize recommending more popular and valuable tutoring content to users.

[0144] Step S144: Perform an adaptation verification process on the semantic association features between the content recommendation sequence and the current dialogue sequence, remove tutoring content fragments that do not match the current dialogue topic, and generate an optimized content recommendation sequence.

[0145] In this embodiment, after obtaining the content recommendation sequence P, it is necessary to perform an adaptation verification process between it and the semantic association features of the current dialogue sequence. The semantic association features can reflect the topic and core content of the current dialogue. Verification can ensure that the recommended tutoring content is relevant to the user's current problems and concerns.

[0146] For each tutoring content fragment pj in the content recommendation sequence P, it is compared with the semantic association features of the current dialogue sequence. A match can be determined by calculating the similarity between the semantic vector of the tutoring content fragment and the semantic association feature vector. Assuming the semantic vector of the tutoring content fragment pj is vpj and the semantic association feature vector is vf, the similarity calculation method can use cosine similarity, etc. If the calculated similarity is lower than a preset similarity threshold, it indicates that the tutoring content fragment does not match the current dialogue topic, and it is removed from the content recommendation sequence P.

[0147] After this adaptation verification process, the optimized content recommendation sequence P' is obtained. This optimized content recommendation sequence only contains tutoring content fragments that match the current dialogue topic, which can more accurately meet the user's needs and improve the quality and relevance of the recommended content.

[0148] Step S145: Encapsulate the optimized content recommendation sequence into the intent optimization strategy, wherein the intent optimization strategy includes recommended content triggering conditions and response timing constraints.

[0149] In this embodiment, after obtaining the optimized content recommendation sequence P', it needs to be encapsulated into an intent optimization strategy. The intent optimization strategy includes not only the recommended guidance content, but also the recommended content triggering conditions and response timing constraints.

[0150] Step S1451: Analyze the semantic topic and keyword set of each tutoring content segment in the optimized content recommendation sequence.

[0151] In this embodiment, for each tutoring content segment pj in the optimized content recommendation sequence P', it is necessary to parse its semantic theme and keyword set. The semantic theme is a summary of the core content of the tutoring content segment, while the keyword set is a set of words that can represent the key information of the tutoring content segment.

[0152] Natural language processing (NLP) techniques can be used for parsing. For example, part-of-speech tagging and named entity recognition can be performed on tutoring content fragments to extract important words and entities. Then, based on the semantic relationships between these words and entities, the semantic theme of the tutoring content fragment can be determined. Simultaneously, representative and important words are grouped into a keyword set kj. Through this parsing process, a deeper understanding of the core information of each tutoring content fragment can be achieved, providing a foundation for generating subsequent content matching rules.

[0153] Step S1452: Generate content matching rules based on the semantic topic and keyword set. The content matching rules are used to detect whether the user's employment consultation statement contains the target keywords in the keyword set in subsequent dialogues.

[0154] In this embodiment, after parsing the semantic topic and keyword set of each tutoring content segment, content matching rules are generated based on this information. The role of the content matching rules is to detect whether the user's employment consultation statements contain target keywords from the keyword set in subsequent dialogues.

[0155] For each tutoring content segment pj and its corresponding keyword set kj, the generated content matching rule can be expressed as follows: if a user's employment consultation statement contains any target keyword from the keyword set kj, then the tutoring content segment is considered relevant to the user's consultation. For example, if the keyword set kj = {"career planning", "industry trends", "skills enhancement"}, when the target keyword "career planning" appears in the user's employment consultation statement, according to the content matching rule, it can be determined that the tutoring content segment pj is relevant to the user's consultation. Such a content matching rule can accurately identify user needs and provide a basis for recommending suitable tutoring content.

[0156] Step S1453: Generate a time-triggered rule based on the response timing constraint, wherein the time-triggered rule is used to limit the maximum delayed response time of the tutoring content segment during the dialogue process.

[0157] In this embodiment, the response timing constraints specify when a tutoring content segment should respond during the dialogue. Based on these response timing constraints, time-triggered rules are generated.

[0158] Time-triggered rules can be set according to different situations. For example, if the response timing constraint requires the tutoring content segment to respond as quickly as possible, then the time-triggered rule can be set to respond within a short period of time after the user raises the relevant inquiry; if the response timing constraint allows for a certain delay, then a maximum delay response time can be set. Assuming the maximum delay response time is t, the time-triggered rule can be expressed as: the tutoring content segment must respond within time t after the user raises the relevant inquiry, otherwise it is considered an invalid response. Such a time-triggered rule can ensure that tutoring content is provided to the user at the appropriate time, improving the user experience.

[0159] Step S1454: Logically combine the content matching rules and the time triggering rules to generate strategy execution conditions.

[0160] In this embodiment, after generating the content matching rules and time-triggered rules, they need to be logically combined to generate the strategy execution conditions. This logical combination can use an AND operation, meaning that the corresponding tutoring content recommendation strategy is only executed when both the content matching rules and the time-triggered rules are satisfied.

[0161] For example, if the content matching rule is "the user's employment consultation statement contains the target keyword from the keyword set kj," and the time triggering rule is "the tutoring content segment responds within time t after the user submits the relevant consultation," then the strategy execution condition can be expressed as: when the user's employment consultation statement contains the target keyword from the keyword set kj, and the tutoring content segment responds within time t after the user submits the relevant consultation, the recommendation strategy for that tutoring content segment is executed. This logical combination ensures that the recommended tutoring content is both relevant to the user's needs and provided to the user at the appropriate time.

[0162] Step S1455: Bind the strategy execution conditions with the optimized content recommendation sequence to generate the intent optimization strategy.

[0163] In this embodiment, the generated strategy execution conditions are bound to the optimized content recommendation sequence P'. This binding process associates the strategy execution conditions with each tutoring content segment, ensuring that each tutoring content segment has a corresponding execution condition.

[0164] For each tutoring content fragment pj in the optimized content recommendation sequence P', it is bound to the corresponding policy execution condition cj. The bound result can be represented as (pj, cj), where pj is the tutoring content fragment and cj is the corresponding policy execution condition. Combining all the bound results generates the intent optimization strategy. This intent optimization strategy clarifies the recommendation conditions and timing for each tutoring content fragment, providing detailed guidance for the dialogue service module of the job search tutoring system.

[0165] Step S150: Feed back the intent optimization strategy to the dialogue service module of the job search coaching system to adjust the response content generation rules.

[0166] In this embodiment, after generating the intent optimization strategy, it needs to be fed back to the dialogue service module of the job search coaching system to adjust the response content generation rules of the module, so as to better meet the user's job search needs.

[0167] Step S151: The intent optimization strategy is divided into multiple strategy execution units, each strategy execution unit corresponding to a response rule for a job search intent.

[0168] In this embodiment, the intent optimization strategy is a comprehensive set of strategies, including multiple tutoring content fragments and their corresponding execution conditions. For ease of management and execution, the intent optimization strategy needs to be broken down into multiple strategy execution units.

[0169] For each coaching content segment and its corresponding strategy execution condition in the intent optimization strategy, they are combined into a strategy execution unit. Assuming the intent optimization strategy contains n coaching content segments p1, p2, ..., pn, and corresponding strategy execution conditions c1, c2, ..., cn, then we can obtain n strategy execution units (p1, c1), (p2, c2), ..., (pn, cn). Each strategy execution unit corresponds to a response rule for a job-seeking intent, specifying under what conditions which coaching content segment should be recommended.

[0170] Step S152: Perform conflict detection processing on the policy execution unit to identify policy execution units with overlapping rules or logical contradictions.

[0171] In this embodiment, after obtaining multiple policy execution units, conflict detection processing is required. Since different policy execution units may have overlapping rules or logical contradictions, this can lead to confusion during execution. Therefore, it is necessary to identify and handle these conflicts promptly.

[0172] Conflict detection can be achieved by comparing the policy execution conditions of policy execution units. For example, for two policy execution units (p1, c1) and (p2, c2), check whether their policy execution conditions c1 and c2 overlap or contradict each other. If such a situation exists, it indicates that the two policy execution units conflict. For example, policy execution condition c1 requires a response within time t1 after the user submits an inquiry, while policy execution condition c2 requires a response within time t2 after the user submits an inquiry, and t1 and t2 conflict. In this case, further processing is required for these two policy execution units.

[0173] Step S153: Perform priority reset processing on conflicting strategy execution units, and reallocate execution priorities according to the confidence scores in the intent classification label set.

[0174] In this embodiment, after identifying conflicting policy execution units, their priorities need to be reset. The priority reset is based on the confidence score in the intent classification label set.

[0175] For conflicting strategy execution units, determine the confidence score of their corresponding job-seeking intentions within the intent category label set. A higher confidence score indicates a more important job-seeking intention, and the corresponding strategy execution unit should have a higher execution priority. For example, if strategy execution units (p1, c1) and (p2, c2) conflict, and their corresponding job-seeking intentions have confidence scores of s1 and s2 respectively within the intent category label set, then if s1 > s2, the execution priority of strategy execution unit (p1, c1) is set higher than that of (p2, c2). This priority resetting process ensures that, in the event of conflict, the more important job-seeking intention receives priority in response.

[0176] Step S154: Compile the processed policy execution unit into an executable code fragment, and deploy the executable code fragment to the policy engine of the dialogue service module.

[0177] In this embodiment, after performing conflict detection and priority reset processing on the policy execution unit, the processed policy execution unit needs to be compiled into an executable code fragment. An executable code fragment is a form of code that a computer can understand and execute; it transforms the logic and rules of the policy execution unit into a concrete code implementation.

[0178] The compilation process can be tailored to the programming language and architecture of the dialogue service module. For example, if the dialogue service module is developed using Python, the logic of the policy execution unit can be implemented using Python code. After compiling each policy execution unit into a corresponding executable code snippet, these snippets are deployed to the policy engine of the dialogue service module. The policy engine is the core component of the dialogue service module, responsible for executing these executable code snippets and dynamically recommending appropriate tutoring content based on the user's inquiry and policy execution conditions.

[0179] Step S155: Establish a policy version management mechanism in the policy engine to record the effective time and coverage of different versions of the intended optimization policy.

[0180] In this embodiment, to better manage intent optimization strategies, a strategy version management mechanism needs to be established in the strategy engine. This mechanism can record the effective time and coverage of different versions of intent optimization strategies, facilitating strategy tracking and adjustment.

[0181] Step S1551: Assign a unique version identifier to each executable code fragment and record the policy effective time window corresponding to the version identifier.

[0182] In this embodiment, each executable code snippet deployed to the policy engine is assigned a unique version identifier. The version identifier can be a string or a number, used to uniquely identify the intent optimization policy version corresponding to that executable code snippet.

[0183] Simultaneously, the effective time window for the policy corresponding to each version identifier is recorded. The effective time window specifies when the intent optimization policy for that version begins to take effect and when it becomes ineffective. For example, the effective time window for the executable code snippet with version identifier v1 is from time t_start1 to time t_end1, indicating that the intent optimization policy for that version is effective within this time period. By recording the version identifier and the effective time window, the time range of intent optimization policies for different versions can be clearly understood.

[0184] Step S1552: Set a policy switching trigger in the dialogue service module to automatically activate the corresponding version of the executable code fragment when the current time is detected to have reached the policy effective time window.

[0185] In this embodiment, to enable automatic switching between different versions of intent optimization strategies, a strategy switching trigger is set in the dialogue service module. The strategy switching trigger monitors the current time in real time and compares it with the recorded policy effective time window.

[0186] When the current time reaches the effective time window of a certain version of the policy, the policy switching trigger automatically activates the corresponding version of the executable code snippet. For example, when the current time reaches the effective time t_start1 of the executable code snippet with version identifier v1, the policy switching trigger will switch the execution logic in the policy engine to the executable code snippet of that version, thus enabling the intent optimization policy of that version to take effect. This automatic switching mechanism ensures that the appropriate intent optimization policy is used in different time periods.

[0187] Step S1553: Establish a strategy rollback log to record the timestamp and strategy execution performance metrics for each strategy switching operation.

[0188] In this embodiment, a policy rollback log is established to track and evaluate policy switching operations. The policy rollback log records the timestamp and policy execution performance metrics for each policy switching operation.

[0189] The timestamp records the specific time when the strategy switching operation occurred, facilitating the analysis of the order and frequency of strategy switching. Strategy performance metrics can include user click-through rates and satisfaction ratings, used to evaluate the effectiveness of each version of the intent optimization strategy. For example, after a strategy switch, the timestamp of the switch and the user click-through rate and satisfaction ratings for the recommended guidance content within a certain period after the switch are recorded. By analyzing these records, the strengths and weaknesses of different versions of the intent optimization strategy can be understood, providing a basis for subsequent strategy adjustments.

[0190] Step S1554: When the execution effect index of the current activation strategy is detected to be lower than the preset threshold, a strategy rollback operation is triggered to restore the response rules of the dialogue service module to the executable code fragment of the previous version.

[0191] In this embodiment, a threshold for the execution performance metric is preset to determine whether the currently activated intent optimization strategy is effective. When the execution performance metric of the currently activated strategy is detected to be lower than the preset threshold, it indicates that there may be a problem with this version of the intent optimization strategy, and a strategy rollback operation needs to be triggered.

[0192] The policy rollback operation restores the response rules of the dialogue service module to the executable code snippet of the previous version. For example, if the performance metrics of the currently active executable code snippet with version identifier v2 are lower than a preset threshold, the policy rollback operation will switch the execution logic in the policy engine back to the executable code snippet with version identifier v1, thus restoring the intent optimization policy of the previous version to its original effect. This policy rollback mechanism allows for the timely correction of ineffective policies, ensuring the response quality of the dialogue service module.

[0193] By following the steps above, the entire process from acquiring the dialogue data set to adjusting the response content generation rules of the job search coaching system is completed. This enables us to more accurately uncover users' job search dialogue intentions and provide them with job search coaching services that better meet their needs.

[0194] Furthermore, in the above embodiments, the intent recognition network mainly includes a feature fusion layer and a classifier layer.

[0195] The feature fusion layer receives semantic association features and interaction behavior features as input. The semantic association features undergo a linear transformation to generate a first transformed feature, and the interaction behavior features undergo a non-linear transformation to generate a second transformed feature. Then, through gated weight calculation, the first and second transformed features are weighted and fused to generate a fused intent feature. Specifically, the gated weight calculation part within the feature fusion layer combines the first and second transformed features, performs a linear transformation, and then uses a sigmoid function to generate dynamic fusion weight coefficients to determine the proportion of the two features in the weighted summation.

[0196] The classifier layer receives the fused intent features as input. In the classifier layer, the fused intent features are first segmented into multiple feature subspaces, each corresponding to a predefined job-seeking intent category. Then, each feature subspace undergoes an independent fully connected transformation to obtain an initial prediction score for each job-seeking intent category. These initial prediction scores are then sorted and binarized. Finally, the confidence scores of the valid intent labels are normalized to generate a set of intent classification labels.

[0197] In terms of connectivity, the output of the feature fusion layer is connected to the input of the classifier layer, forming a sequential data stream. The linear and nonlinear transformation parts of the feature fusion layer process semantic association features and interaction behavior features respectively, while the gating weight calculation part fuses these two processing results. The various sub-steps of the classifier layer, such as feature subspace segmentation, fully connected transformation, sorting, binarization, and normalization, are also sequentially connected, gradually completing the classification processing of the fused intent features.

[0198] During training, a large dataset of dialogues is first collected, containing a rich collection of user job-seeking inquiries and corresponding responses from the tutoring system. This data undergoes preprocessing, including cleaning and annotation. During annotation, each dialogue sequence is labeled with its corresponding job-seeking intent category, forming the training dataset.

[0199] The training dataset is divided into a training set, a validation set, and a test set. The training set is used to train the model, the validation set is used to tune the model's hyperparameters during training, and the test set is used to evaluate the model's final performance.

[0200] Next, initialize the parameters of the intent recognition network, including the weight matrix and bias vector of the feature fusion layer, and the weight matrix and bias vector of the classifier layer. These parameters can be initialized to small random values ​​using a random initialization method.

[0201] The dialogue sequences in the training set are input into the intent recognition network, and processed sequentially through feature extraction, feature fusion layers, and a classifier layer to obtain a set of predicted intent classification labels. A loss function is used to calculate the loss between the predicted intent classification label set and the true labels. A commonly used loss function is the cross-entropy loss function, which measures the difference between the predicted and true results.

[0202] Based on the gradient of the loss function, the model parameters are updated using an optimization algorithm (such as stochastic gradient descent). The optimization algorithm adjusts the model's weight matrix and bias vector according to the direction and magnitude of the gradient, so that the value of the loss function gradually decreases.

[0203] Repeat the forward propagation, loss calculation, and backpropagation steps described above, continuously updating the model parameters until the model's performance reaches a satisfactory level. During training, a validation set can be used to monitor the model's performance and prevent overfitting.

[0204] During training, several hyperparameters need to be adjusted, such as the learning rate, batch size, and number of training epochs. The learning rate controls the step size at which the optimization algorithm updates parameters, the batch size determines the number of data samples used in each training iteration, and the number of training epochs indicates the number of times the entire training dataset is used. Optimal combinations of hyperparameters can be found using methods such as grid search and random search.

[0205] Finally, the trained model is evaluated using a test set, and metrics such as accuracy, recall, and F1 score are calculated to assess its performance. If the model's performance is unsatisfactory, its structure or hyperparameters can be adjusted, and the model can be retrained.

[0206] In the context of college student career guidance, the input data for the intent recognition network consists of semantic association features and interactive behavior features. Semantic association features are obtained through feature extraction and processing of user career consultation statements and guidance system responses, reflecting the semantic coherence between the two. Interactive behavior features are obtained through behavioral pattern analysis of user career consultation statements, reflecting the intent triggering conditions and response timeliness constraints corresponding to the user's career consultation statements.

[0207] Specifically, the generation process of semantic association features includes semantic segmentation, serialization encoding, bidirectional attention matching, and max pooling of the dialogue sequence. The generation process of interaction behavior features includes identifying interaction triggering conditions and response timeliness constraint vectors, and then concatenating these features. These processing steps ensure that the input data accurately reflects the semantic and behavioral information of the dialogue sequence, providing rich features for intent recognition.

[0208] The output data of the intent recognition network is a set of intent classification labels, with each label corresponding to a confidence score for a type of job demand. This output data is closely related to the scenario of college student career counseling, and it can indicate the distribution priority of users' potential job demand.

[0209] In job coaching systems, intent optimization strategies can be built based on intent classification tag sets to provide users with more precise job coaching content. For example, if the intent classification tag set shows that a user has a high confidence score for their need for career planning, then the job coaching system can prioritize recommending coaching content related to career planning, such as career planning courses and career assessment tools.

[0210] There is a close intrinsic relationship between the input and output data. Semantic association features and interactive behavior features serve as inputs, reflecting the user's dialogue information and behavioral patterns in a job-seeking coaching scenario. The intent recognition network processes and analyzes this input data to uncover the user's potential job-seeking intent and generates a corresponding set of intent classification labels as output.

[0211] This inherent correlation enables job coaching systems to accurately identify users' job-seeking needs based on their dialogue content and behavioral patterns, and to provide personalized coaching services. For example, when a user's inquiry reveals a strong need for job skills improvement, semantic association features and interactive behavior features will reflect this information. The intent recognition network will then generate intent classification labels related to job skills improvement with high confidence scores. Based on this output, the job coaching system can recommend job skills training courses, skills improvement strategies, and other coaching content to the user.

[0212] Furthermore, the data collection process may involve sensitive data, such as users' personal contact information and educational background. To protect this sensitive data, it must be encrypted. Symmetric encryption algorithms (such as AES) or asymmetric encryption algorithms (such as RSA) can be used to encrypt the data, ensuring its security during transmission and storage. Before using sensitive data for training and analysis, the data should be anonymized. For example, users' real names can be replaced with anonymous identifiers, and specific school names can be replaced with school categories. This allows for model training and analysis without disclosing user privacy. Strict control must be exercised over access to sensitive data. Only authorized personnel should have access to this data, and the access process should be recorded. Role-based access control (RBAC) and other methods can be used to manage user access permissions.

[0213] When displaying and sharing data, sensitive data should be anonymized. For example, only statistical information should be shown, not specific personal data. This ensures data availability while protecting user privacy.

[0214] Figure 2 This diagram illustrates the hardware structure of a job-seeking dialogue service system 100 provided in an embodiment of the present invention for implementing the above-described method for deep mining of job-seeking dialogue intent applied to college student employment guidance. Figure 2 As shown, the job search dialogue service system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0215] In one possible design, the job-seeking dialogue service system 100 can be a single server or a group of servers. The server group can be centralized or distributed (e.g., the job-seeking dialogue service system 100 can be a distributed system). In some embodiments, the job-seeking dialogue service system 100 can be local or remote. For example, the job-seeking dialogue service system 100 can access information and / or data stored in machine-readable storage medium 120 via a network. Alternatively, the job-seeking dialogue service system 100 can directly connect to machine-readable storage medium 120 to access stored information and / or data. In some embodiments, the job-seeking dialogue service system 100 can be implemented on a job-seeking dialogue service system. By way of example only, the job-seeking dialogue service system can include a private cloud, public cloud, hybrid cloud, community cloud, distributed cloud, internal cloud, multi-tiered cloud, etc., or any aggregation thereof.

[0216] Machine-readable storage medium 120 may store data and / or instructions. In some embodiments, machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, machine-readable storage medium 120 may store data and / or instructions used by job search dialogue service system 100 to perform or use in order to accomplish the exemplary methods described in this invention.

[0217] In the specific implementation process, one or more processors 110 execute computer-executable instructions stored in machine-readable storage medium 120, so that processor 110 can execute the job-seeking dialogue intent deep mining method applied to college student employment guidance as described in the above method embodiment. The processor 110, machine-readable storage medium 120 and communication unit 140 are connected through bus 130. The processor 110 can be used to control the sending and receiving actions of communication unit 140.

[0218] The specific implementation process of processor 110 can be found in the various method embodiments executed by the job search dialogue service system 100 described above. The implementation principle and technical effect are similar, and will not be repeated here.

[0219] Furthermore, embodiments of the present invention also provide a readable storage medium containing computer-executable instructions. When the processor executes the computer-executable instructions, the above-mentioned method for in-depth mining of job-seeking dialogue intent applied to college student employment guidance is implemented.

[0220] It should be noted that, in order to simplify the description of this invention and thus aid in the understanding of one or more embodiments, the foregoing description of the embodiments of this invention sometimes combines multiple features into a single embodiment, drawing, or description thereof. Similarly, it should be noted that, in order to simplify the description of this invention and thus aid in the understanding of one or more embodiments, the foregoing description of the embodiments of this invention sometimes combines multiple features into a single embodiment, drawing, or description thereof.

Claims

1. A job-hunting dialogue intent deep mining method applied to university student employment guidance, characterized in that, The method comprises: obtaining a conversation data set of a target user in a job counseling scene, the conversation data set comprising a plurality of conversation sequences arranged in chronological order, each conversation sequence being composed of a user employment counseling sentence and a corresponding counseling system answer sentence; performing feature extraction on the conversation data set to generate semantic association features and interaction behavior features of each conversation sequence; wherein the semantic association features represent the semantic coherence between the user employment counseling sentence and the counseling system answer sentence, and the interaction behavior features represent the intent trigger condition and response timeliness constraint of the user employment counseling sentence; performing intent recognition on the semantic association features and the interaction behavior features based on a preset intent recognition network to generate an intent classification label set of the conversation sequence, the intent classification label set being used to indicate the distribution priority of the user's potential job-seeking demand; constructing an intent optimization strategy according to the intent classification label set, and feeding back the intent optimization strategy to a dialogue service module of the job counseling system to adjust the response content generation rule.

2. The method of claim 1, wherein, The feature extraction on the conversation data set to generate semantic association features and interaction behavior features of each conversation sequence comprises: performing semantic segmentation processing on the user employment counseling sentence in the conversation sequence to obtain a plurality of semantic unit sets, each semantic unit set comprising at least one vocabulary fragment having independent semantic expression capability; calling a pre-trained context encoding model to perform sequence encoding processing on the semantic unit set to generate a semantic vector set of the user employment counseling sentence, wherein each semantic vector in the semantic vector set corresponds to a context semantic representation of a semantic unit; performing bidirectional attention matching processing on the semantic vector set of the user employment counseling sentence and the semantic vector set of the counseling system answer sentence to generate a semantic association matrix between the user employment counseling sentence and the counseling system answer sentence; performing maximum pooling processing on the semantic association matrix to extract the semantic vector corresponding to the maximum response association value in the semantic association matrix as the semantic association feature; performing behavior pattern analysis processing on the user employment counseling sentence in the conversation sequence to identify the interaction trigger condition and response timeliness constraint vector of the user employment counseling sentence, and performing feature splicing on the interaction trigger condition and the response timeliness constraint vector to generate the interaction behavior feature.

3. The method of claim 2, wherein, The calling of the pre-trained context encoding model to perform sequence encoding processing on the semantic unit set to generate the semantic vector set of the user employment counseling sentence comprises: inputting the semantic unit set into an embedding layer of the context encoding model for vector mapping processing to obtain an initial semantic vector set; The initial semantic vector set is input into a bidirectional long short-term memory network layer of the context encoding model for context-dependent modeling processing to obtain an intermediate semantic vector set; wherein the bidirectional long short-term memory network layer comprises a forward propagation subnetwork and a backward propagation subnetwork, the forward propagation subnetwork is used to capture a forward dependency relationship between a current semantic unit and a historical semantic unit, and the backward propagation subnetwork is used to capture a reverse dependency relationship between the current semantic unit and a future semantic unit; The intermediate semantic vector set is subjected to multi-head self-attention processing to generate an attention weight matrix, and the intermediate semantic vector set is subjected to weighted summation according to the attention weight matrix to obtain an enhanced semantic vector set; The enhanced semantic vector set is input into a normalization layer of the context encoding model for feature scaling processing to obtain a semantic vector set of the user employment counseling sentence.

4. The method of claim 2, wherein, The bidirectional attention matching processing is performed according to the semantic vector set of the user employment counseling sentence and the semantic vector set of the tutoring system answer sentence to generate a semantic association matrix between the user employment counseling sentence and the tutoring system answer sentence, which comprises: The semantic vector set of the user employment counseling sentence is taken as a query vector set, and the semantic vector set of the tutoring system answer sentence is taken as a key vector set; The cosine similarity between each query vector in the query vector set and each key vector in the key vector set is calculated to generate an initial similarity matrix; The initial similarity matrix is subjected to row direction normalization processing to obtain a query attention weight matrix; The initial similarity matrix is subjected to column direction normalization processing to obtain a key attention weight matrix; The query attention weight matrix and the key attention weight matrix are subjected to element-by-element multiplication processing to generate a bidirectional attention weight matrix; The key vector set is subjected to weighted aggregation processing according to the bidirectional attention weight matrix to generate a context-aware vector set corresponding to the query vector set; The context-aware vector set and the query vector set are subjected to splicing processing to obtain an enhanced context vector set; The enhanced context vector set is subjected to fully connected transformation processing to generate the semantic association matrix.

5. The method of claim 1, wherein, The intent recognition network is used for intent recognition of the semantic association feature and the interaction behavior feature to generate an intent classification label set of the dialogue sequence, which comprises: The semantic association feature is input into a feature fusion layer of the intent recognition network for linear transformation processing to generate a first transformed feature; The interaction behavior feature is input into the feature fusion layer for nonlinear transformation processing to generate a second transformed feature; The first transformed feature and the second transformed feature are subjected to gate weight calculation processing to generate a dynamic fusion weight coefficient; wherein the gate weight calculation processing comprises inputting the first transformed feature and the second transformed feature into a sigmoid function to generate a weight distribution ratio; The first transformed feature and the second transformed feature are subjected to weighted summation according to the dynamic fusion weight coefficient to generate a fused intent feature; Input the fusion intention feature into a classifier layer of an intention recognition network for multi-label classification processing to generate the intention classification label set, each intention classification label corresponding to a confidence score of one job-seeking demand type.

6. The method of claim 5, wherein, The inputting of the fusion intention feature into the classifier layer of the intention recognition network for multi-label classification processing to generate the intention classification label set comprises: segmenting the fusion intention feature into a plurality of feature subspaces, each feature subspace corresponding to one pre-set job-seeking intention category; performing independent fully connected transformation processing on each feature subspace to generate an initial prediction score of each job-seeking intention category; performing sorting processing on the initial prediction score to obtain a prediction score sequence; performing binary processing on the prediction score sequence according to a pre-set score threshold, and marking the job-seeking intention category corresponding to the prediction score exceeding the score threshold as an effective intention label; performing normalization processing on the confidence scores of all effective intention labels to generate the intention classification label set.

7. The method of claim 1, wherein, The construction of the intention optimization strategy according to the intention classification label set comprises: extracting a target intention label with the highest confidence score from the intention classification label set, and identifying a job-seeking demand type corresponding to the target intention label; matching a pre-set coaching resource template library according to the job-seeking demand type to obtain a set of coaching content segments associated with the job-seeking demand type; performing priority sorting processing on the set of coaching content segments to generate a content recommendation sequence; wherein the priority sorting processing comprises calculating a comprehensive weight value according to a historical click rate and a user feedback score of a coaching content segment and performing priority sorting processing based on the comprehensive weight value; performing adaptability checking processing on the content recommendation sequence and semantic association features of a current dialogue sequence to remove coaching content segments that do not match the current dialogue theme, and generating an optimized content recommendation sequence; encapsulating the optimized content recommendation sequence as the intention optimization strategy, wherein the intention optimization strategy contains a recommended content trigger condition and a response timing constraint condition.

8. The method of claim 7, wherein, The encapsulation of the optimized content recommendation sequence as the intention optimization strategy comprises: analyzing the semantic theme and keyword set of each coaching content segment in the optimized content recommendation sequence; generating a content matching rule according to the semantic theme and keyword set, the content matching rule being used to detect whether a user's employment consulting statement contains a target keyword in the keyword set in a subsequent dialogue; generating a time trigger rule according to the response timing constraint condition, the time trigger rule being used to limit the maximum delay response time of the coaching content segment in the dialogue process; performing logical combination processing on the content matching rule and the time trigger rule to generate a strategy execution condition; binding the strategy execution condition and the optimized content recommendation sequence to generate the intention optimization strategy.

9. The method of claim 1, wherein, The feedback of the intention optimization strategy to the dialogue service module of the job coaching system to adjust the response content generation rule comprises: Split the intention optimization strategy into multiple policy execution units, each corresponding to a response rule for a job-seeking intention; Perform conflict detection processing on the policy execution units to identify policy execution units with overlapping rules or logical contradictions; Perform priority resetting processing on the policy execution units with conflicts, and reassign execution priorities based on the confidence scores in the intention classification tag set; Compile the processed policy execution units into executable code snippets and deploy them to the policy engine of the dialog service module; Establish a policy version management mechanism in the policy engine to record the effective time and coverage range of different versions of the intention optimization strategy; The policy version management mechanism in the policy engine includes: Assign a unique version identifier to each executable code snippet and record the policy effective time window corresponding to the version identifier; Set a policy switching trigger in the dialog service module to automatically activate the executable code snippet of the corresponding version when it is detected that the current time has reached the policy effective time window; Establish a policy rollback log to record the timestamp and policy execution effect indicator of each policy switching operation; When it is detected that the execution effect indicator of the currently activated policy is lower than the preset threshold, trigger a policy rollback operation to restore the response rules of the dialog service module to the executable code snippet of the previous version.

10. A job search conversation service system characterized by comprising: The job-seeking dialog service system includes a processor and a memory, the memory and the processor are connected, the memory is used to store programs, instructions or codes, and the processor is used to run the programs, instructions or codes in the memory to implement the job-seeking dialog intention deep mining method for university student employment guidance according to any one of claims 1-9.

Images