A customer service dialogue interaction optimization method and system based on sequence perception

Abstract

The present invention provides a customer service dialogue interaction optimization method and system based on sequence perception, including obtaining a demand data set and a medical guidance related data set; generating a context-related information list of the queried demand question; extracting the user query input behavior characteristics of the demand data set to obtain an aggregated feature vector; using a model trained based on sequence perception to obtain effective user query input behavior data, and obtain the final user intention prediction. By establishing a calibration process for recommended content based on the recommended content obtained when a query behavior occurs, setting a method of updating the system model associated with all questions in the context of the user's session according to each question input by the user, the present invention obtains context information with context memory and correct recommended content, which can provide relevant medical answers even when the keyword information is incomplete and the system model is limited.

Description

Technical Field

[0001] The present invention relates to the technical field of customer service dialogue optimization recommendation, and specifically to a customer service dialogue interaction optimization method and system based on sequence perception. Background Art

[0002] Currently, customer service systems are usually optimized for specific tasks or problems. Due to the colloquial and fragmented nature of user input and the limitations of the system model, the keyword information that the customer service system can obtain is incomplete, unable to generate contextual associations, and lacks the ability to respond to and distinguish unexpected user queries, resulting in irrelevant answers.

[0003] For example, patent document CN118609565A discloses a real-time voice intelligent customer service system and method, which mentions that the converted text is semantically analyzed and intent recognized through a natural language processing module to understand the meaning and purpose of the customer's question. It still has disadvantages. The semantic analysis and intent recognition capabilities of the natural language processing module are insufficient, and it cannot accurately understand complex or unexpected customer questions. This disadvantage is reflected in the fact that the current natural language processing module cannot accurately recognize that the user's intention is to ask whether the symptoms are related to a specific disease, but mistakenly recognizes it as the user asking other questions, such as our common language model dialogue "I feel pain in the big toe joint when walking recently. Is this a symptom of hallux valgus? What medicine should I take to relieve it relatively well?" It is recognized as an answer about the symptoms of various foot problems, such as "Hallux valgus is a foot disease, which is usually related to improper wearing of shoes. In order to prevent hallux valgus, it is recommended that you wear loose and comfortable shoes and avoid high heels." Instead of answering questions about symptoms; at the same time, the business knowledge and FAQs stored in its knowledge base module cannot cover all customer query needs, especially some specific or uncommon questions, which makes the system unable to provide relevant answers. The common problem is the lack of effective management of context and context memory, which leads to the incoherence of the conversation. As a result, the system cannot correctly associate the relationship between "big toe joint pain" and "hallux valgus", and ignores the user's real intention to ask for drug recommendations. It is unable to use previous conversation information to understand the current user's query and form an effective contextual association, that is, it provides an answer that is irrelevant to the question and gives general advice on how to prevent hallux valgus, rather than targeting the user's current pain and treatment needs.

[0004] For example, patent document CN118377863A discloses an intelligent customer service response method, apparatus, computer equipment and storage medium, which proposes to segment consultation request information to obtain keyword information, screen matching results corresponding to the keyword information in a preset knowledge base, score the matching results, generate intelligent answer information based on the keyword information and a preset intelligent large language model, and send the intelligent answer information to the user, providing the user with intelligent answer information, further optimizing the customer service response method, and improving the efficiency of the intelligent customer service system and user experience.

[0005] The system still has the disadvantages of relying on keyword matching and lacking understanding of the deep semantics of the query. For example, when a user asks "I am pregnant, what medicine should I take to relieve bunion pain?", the system proposes to input keyword information into the category large language model for keyword analysis based on the intention analysis subunit to obtain keyword intent. Based on this, it can only recognize the two keywords "bunion pain" and "medication", but cannot understand that the user actually needs specific drug recommendations for bunion pain symptoms in pregnant patients. At the same time, the system is preset with an intelligent large language model from the beginning, resulting in the defect of judging the right and wrong of the model output. At the same time, the system's scoring of matching results depends on the preset threshold, resulting in the problem of being unable to accurately judge the correctness of the output and reflect the quality of the answer. Therefore, the system will select an answer because of a high-matching keyword, but the answer does not fully meet the user's complex query needs. Summary of the invention

[0006] In view of the deficiencies in the prior art, the purpose of the present invention is to provide a customer service dialogue interaction optimization method and system based on sequence perception to solve the problems raised in the above background technology.

[0007] In order to achieve the above-mentioned purpose, the present invention is implemented through the following technical solutions: A customer service dialogue interaction optimization method based on sequence perception, comprising the steps of: S1, pre-dialogue, generating a pre-recommended context-related information list according to the dialogue content, and forming an aggregated feature vector, which performs the following operations: S1-1, initial conversation background, in which the queried demand problems generated by the user during the conversation based on the user's query input behavior are continuously obtained, and are aggregated into a demand data set S data and medical guidance related dataset G data ; S1-2, respectively from the demand data set S data and medical guidance related dataset G dataIn the process of constructing key feature vectors for characterizing user query input behaviors at different levels, before the user performs a second query input behavior, the key feature vectors are tagged with parts of speech and recognized by natural language processing tools to generate initial contextual association information with recommended concept attributes; S1-3, evaluating the known recommended content R in the system according to the initial contextual association information con , and calibrate the erroneous knowledge, inject correct knowledge, enhance the system's understanding of complex query behavior, and then generate a context-related information list of the queried demand problem based on attribute matching sorting, wherein the context-related information list includes the user's current query intention and all context information related to the intention; extract the parent user query input behavior feature P of the demand data set fea And the sub-level user query input behavior characteristics C of the medical guidance association dataset fea , we get the aggregate feature vector H that represents the user query intent and the contextual information associated with the intent in the previous session results agg ; S2, simulate the user's query input behavior to obtain the user's real conversation intention, which performs the following operations: use the aggregated feature vector H agg The classification model is trained to generate a user intention recognition sequence, and then an LSTM model is constructed to determine its authenticity. Finally, a recall test is performed to screen out the expected user query behavior data as the output result, and S3 is executed; S3, post-dialogue, visualizes the medical information related to the queried demand problem.

[0008] As a second aspect of the present invention, a sequence-awareness-based customer service conversation interaction optimization system is proposed, comprising a memory and a processor, wherein the memory comprises a sequence-awareness-based customer service conversation interaction optimization program, and when the sequence-awareness-based customer service conversation interaction optimization program is executed by the processor, the above-mentioned sequence-awareness-based customer service conversation interaction optimization method is implemented.

[0009] Compared with the prior art, the present invention has the following beneficial effects:

[0010] The present invention first constructs a key feature vector containing a set of parameters for describing the user's query keywords, text content keywords, and browsing time node information, and extracts keywords from fragmented input information; then, in order to enable the system to identify and extract multi-level information related to the user's query, the present invention matches entities or keywords in the data in the user's input question with categories in a predefined database based on the attributes of the parameter set, forming a comprehensive "associative" data structure that not only contains the user's query information, but also contains the system's understanding and processing of these information;

[0011] Next, to ensure that the user's query needs are understood, the present invention establishes a calibration process for recommended content based on the recommended content obtained when a query behavior occurs, sets a system model associated with all questions in the context of the user's session according to each question input by the user, and obtains context information with context memory and correct recommended content, and then performs matching analysis on the list of context-related information in the session context, and first identifies the recommended content that best meets the current user's needs;

[0012] Finally, by constructing a time series model and aggregating feature vectors, we capture the timing information and contextual association information of different users regarding the same conversation content, thereby predicting the true intentions of users in colloquial and fragmented questions, and conducting further true intention questions, so as to quickly locate and judge the user's medical problems, and provide relevant medical answers even when the keyword information entered by the user is incomplete due to lack of medical knowledge. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The disclosure of the present invention is described with reference to the accompanying drawings. It should be understood that the drawings are only for illustrative purposes and are not intended to limit the scope of protection of the present invention. In the drawings, the same reference numerals are used to refer to the same components. Among them:

[0014] Figure 1 This is a schematic diagram of the overall architecture of a customer service dialogue interaction optimization method based on sequence perception proposed in one embodiment of the present invention;

[0015] Figure 2 The invention is proposed in an embodiment, describing the processing of user's query about "hallux valgus" treatment information, and from the known recommended content R con Schematic diagram of a flow chart for selecting and displaying the most relevant recommendations;

[0016] Figure 3 A schematic diagram of a process of forming an aggregate vector in a pre-dialogue to solve the problem of user input being colloquialized, according to an embodiment of the present invention;

[0017] Figure 4 A schematic diagram of a process for perceiving, understanding and predicting user query intentions based on time series analysis proposed in one embodiment of the present invention;

[0018] Figure 5 This is a schematic diagram of a process for enhancing retrieval using a large language model in the medical field, as proposed in one embodiment of the present invention;

[0019] Figure 6This is a flowchart of the entire process from evaluating erroneous knowledge to injecting correct knowledge and calibrating erroneous knowledge, proposed in one embodiment of the present invention, in order to solve the defects of the existing language model in outputting right and wrong judgments. DETAILED DESCRIPTION

[0020] It is easy to understand that according to the technical solution of the present invention, without changing the essential spirit of the present invention, a person skilled in the art can propose a variety of interchangeable structural modes and implementation modes. Therefore, the following specific implementation modes and drawings are only exemplary descriptions of the technical solution of the present invention, and should not be regarded as the whole of the present invention or as a limitation or restriction to the technical solution of the present invention.

[0021] The present invention is further described in detail below with reference to the accompanying drawings, but is not intended to limit the present invention.

[0022] As an understanding of the technical concept and implementation principle of the present invention, the main purpose of the customer service dialogue interaction optimization method and system based on sequence perception proposed by the present invention is to solve the colloquialism, fragmentation and system model prediction limitations of user input problems, which lead to incomplete keyword information that the customer service system can obtain, and cannot generate context association, resulting in irrelevant questions. Therefore, the present invention uses natural language processing tools and attribute matching calculations to initially generate a context association information, and sets a calibration process for the recommended content in the system based on the obtained context association information in this process, and sets to update the system model associated with all questions in the user's conversation background according to each user input question. Different from the prior art, a list of context association information with context memory and correct recommended content is obtained. This list is used to guide the system on how to provide the next accurate recommendation according to the user's conversation content, and then train a classification model to identify the user's real query intention, and generate a user intention identification sequence for subsequent authenticity judgment, so as to guide the user to answer his true intention, so as to achieve the effect of quickly locating and judging the user's medical problem, and in the case of incomplete keyword information input by the user due to lack of medical knowledge, the relevant medical answer can also be provided.

[0023] As a first aspect of the present invention, Figure 1As shown in the figure, a customer service dialogue interaction optimization method based on sequence perception is proposed, which includes the following steps: S1, forming an aggregated feature vector based on the previous dialogue. It can be understood that the purpose of this step is to solve the colloquialization of user input problems. Through the association of each user query input behavior, the keywords or context-related information of the queried problem in the user query input behavior are extracted, and similarity aggregation is performed to form a comprehensive "association" data structure that not only contains the user's query information, but also includes the system's understanding and processing of this information. Then, based on the formed context-related information list, the subsequent interaction path is determined to improve the accuracy and relevance of the response. At the same time, the data (or input vector) associated with the user's query input behavior is constructed in the early stage, and then the time series model and the aggregated feature vector H are constructed later. agg , capturing the timing information and contextual association information of user conversations, so as to predict and determine the true intention of users in spoken and fragmented input.

[0024] In one embodiment of the present invention, the step of executing S1 includes:

[0025] S1-1, initial session context, in which the queried demand questions generated by the user based on the user's query input behavior during the session are continuously obtained and compiled into a demand dataset S data And the required data set S data Medical guidance related dataset G data It is understandable that the above-mentioned demand data set S data Contains the user's query requirements in the session, and each question is composed of query content feature data C data and user characteristic data U data Composition, medical guidance related dataset G data Contains a medical guidance recommendation demand data set that is approximately associated with the user query behavior that occurred, such as historically associated approximate case data. To facilitate the understanding of subsequent solutions, it is preferred to define user feature data as including: age U age 、Gender gen , Past medical history Family medical history ill 、Contact U tel , Living habits rul ; The query content feature data is defined as including: a brief description of the current user's symptoms C sym , the time when the current user's symptoms occurred C tims 、Query requirements C req .

[0026] S1-2. Extract the query input behavior feature P of the parent user of the demand data set fea And the sub-level user query input behavior characteristics C of the medical guidance association dataset feaThe specific implementation steps are as follows:

[0027] S1-21, use the matching function to select the required data set S data and medical guidance related dataset G data The key feature vectors of different levels are constructed to characterize the user's query input behavior, with the aim of predicting the user's next possible query. The process is as follows:

[0028] Assume that the required data set S data There are N queried demand questions, each of which is composed of user feature data representing user personal information and query content feature data representing user session content. Suppose the medical guidance related data set G data It contains M medical guidance recommendation demand data;

[0029] According to the matching function f of the set user query input behavior mat According to the following formula, from the demand data set S data Identify and extract the parent key feature vector : , where Represented as the i-th queried demand problem, e par Represented as the set user query input behavior matching function f mat The set of all parent parameters about the parent feature matching strategy in;

[0030] At the same time, according to the formula, from the medical guidance association data set G data Identify and extract the sub-level key feature vector : , where It is represented as the jth medical guidance recommendation demand data that is approximately associated with the user's query input behavior, e chi Represents the set user query input behavior matching function g mat The sub-level parameter set of all sub-level feature matching strategies in , M represents the medical guidance association dataset G data The number of medical guidance recommendation demand data generated in, where the number M is greater than, equal to or less than N. In specific implementation, if the demand data set S dataIf there is at least one generated medical guidance recommendation demand data in each queried demand question, then M≥N. Otherwise, if the generated medical guidance recommendation demand data is not a direct reflection of the user's query demand question after the query input behavior occurs in the session context, then M<N. At the same time, the parent and child parameter sets of the parent and child feature matching strategies are used to describe and identify the query keywords, browsing time information, medical guidance recommendation demand data, etc. generated after the user's query input behavior. And when processing the subsequent matching problem of user query demand problems and medical guidance data sets, the matching function f is used mat and matching function g mat , preferably using the support vector machine (SVM) supervised learning algorithm in machine learning.

[0031] S1-22, generate a list of context-related information of the queried demand problem to identify the context-related information in the queried demand problem relative to the demand data set S data and medical guidance related dataset G data It can be understood that this step is completed based on the above parameter set, and the purpose is to extract the key information that best represents the user's query behavior from the user's query demand question, and then form the context summary data of this key information through matching. Before the user enters the query again, the context-related information list is used to determine the subsequent interaction path, improve the accuracy and relevance of the response, and optimize the user experience. The specific implementation steps are as follows:

[0032] S1-221, use natural language processing (NLP) tools to analyze the key feature vectors of the parent and the child key feature vector The context-related information representing the queried demand question is used to identify noun attributes and non-noun attributes, specifically including part-of-speech tagging and entity recognition of the context-related information to determine the recommended concept attributes to which the context-related information belongs, so as to facilitate subsequent attribute judgment and recommendation.

[0033] For ease of understanding, Figure 2 As shown, the present invention takes the query question "hallux valgus, big toe bone" by a user (pregnant woman) as an example:

[0034] The initial conversation context is: "hallux valgus, big toe bone" query context;

[0035] The user's query question during the session is: "Ballux valgus" (context-related information);

[0036] The part-of-speech tagging of the context-related information obtained during attribute recognition is: "single-legged (adjective)", "hallux valgus (noun)"; in this example, "hallux valgus" is a proper noun related to medical guidance that is approximately related to the query demand question, so it is identified as an entity and obtained: "hallux valgus" (medical guidance related data). Further, the recommended content R related to the query demand question is obtained. con , R con ={("single-legged", "adjective", " / "),("hallux valgus", "noun", "medical guidance related data")}. At this point, the context-related information "hallux valgus" in the query demand question is used as an entity of the medical guidance related data, and subsequent attribute judgment and recommendation can be performed.

[0037] S1-222, custom calculation of the function of the number of noun entities a1 appearing in the returned query demand problem, and set the number threshold a2 for defining the number of noun entities, compare it with the number of noun entities a1 appearing in the returned query, if the number of noun entities a1 appearing is greater than the number threshold a2, then it is considered that the contextual information in the queried demand problem has the recommended concept attribute, otherwise it does not. It can be understood that "hallux valgus" has been obtained as an entity of medical guidance related data in S1-221, so the queried demand problem "hallux valgus on one foot" has the recommended concept attribute, this recommended concept attribute is reflected in the attribute matching process, which is the link given by the system to better understand the recommended content of the queried demand problem. In specific implementation, the higher the attribute matching, the closer the concept attribute in the contextual information in the queried demand problem is to the concept attribute of the recommended content.

[0038] S1-223, using the identified concept attributes, based on the contextual information and known recommended content R con The parent key feature vector is obtained by calculating the similarity and the child key feature vector The attribute matching degree of each context-related information in the system is calculated, and then a list of context-related information representing the queried demand problems of the user during the session is generated according to the sorting of the attribute matching degrees. Based on the above technical concept, it should be noted that based on S1-221, it has been known that the queried demand problem "hallux valgus on one foot" has the conceptual attributes required for recommendation. At this point, the system selects the content with the highest matching degree with the user's current context-related information in terms of conceptual attributes for recommendation, and generates a list of context-related information representing the queried demand problems of the user during the session.

[0039] Example: Assume that there is already a known recommended content R containing "hallux valgus" as medical guidance related data con , which is represented by R con ={rcon1 , r con2 , ..., r conS}, and for the user's query problem "hallux valgus", the following formula is used to calculate its key feature vector at the parent level and the child key feature vector Attribute matching in: (1)

[0040] (2),

[0041] Where P i and C i Represented as the parent key feature vector and the child key feature vector The feature vector representing the contextual information, r cons Represented as known recommended content R con The sth feature vector containing the medical guidance related data entity, Represented as the above eigenvector r cons The Euclidean length, max s Denoted as represents all known recommended content R con Among the feature vectors, select the key feature vectors of the current and parent and the child key feature vector Based on this, we can calculate the attribute matching degree and recommend the known content R con Sort and recommend content with high matching degree first. As shown in the following table, a list of contextual information representing the user's query needs during the session is generated:

[0042]

[0043] As shown in the above lists list.1 to list.4, the recommended content has been approximately completed, but it is still unknown whether it is the user's original query information or derived from the system based on the query information. Therefore, it is necessary to perform S1-23 to compare the sequence number of each parameter in the sub-level parameter set with the context-related information list to determine whether these parameters are located in the original query unit (the user's original query information) or the derived query unit (information further expanded based on the original query).

[0044] Example: The original query unit is the user's initial question or request, such as "Do you want to consult about hallux valgus and big toe?", which identifies the main focus in the conversation context, that is, the user wants to obtain demand questions about "halux valgus" and "big toe". The derived query unit is a further inquiry or answer based on the original query unit, such as "the feature vector P2 in list.2: "How old are you this year" and "more than 20" are derived query units, which are the part of the conversation that further obtains the user's age information based on the original query", and the feature vector in list.4: "Is it congenital or caused by wearing shoes later?", "more than 20" and "acquired" are derived query units, the purpose of which is to help users refine their demand problem status.

[0045] It should be noted that before the user performs a second query input behavior, the following S1-23 needs to be executed to locate the parameter set in the key feature vector used to describe and identify the user's query input behavior, optimize the context-related information list, learn the user's possible next query based on the parameter set, and prepare corresponding answers or recommendations, so that the system can prepare more accurate answers or recommendations in advance, thereby giving "many patients suffer from hallux valgus foot deformity and bunion pain due to external factors such as wearing shoes. Once it occurs, it will become more and more serious, easily causing protrusion next to the big toe, overlapping toes, increased pain, etc. Severe patients will also be affected in walking and cannot heal themselves." The specific implementation steps are:

[0046] S1-231, use a programming language to initialize and define a data structure, requiring that the data structure includes: an empty list L for storing all context-related information in the session context empty , and the empty list L empty Each element of corresponds to each query demand question and its corresponding answer in the context of the conversation; and a dictionary Q storing the original query unit and the derived query unit list , Q list ={"key value 1": NO, "key value 2": empty list}, and the dictionary Q list Each key value corresponds to each queried demand question and the corresponding answer in the conversation context.

[0047] S1-232, collect context information, define the first queried question of the user in the context of the session as the original query unit, and add it to the dictionary Q list At the same time, traverse each queried demand question and its corresponding answer in all interaction data in the session background, and add it as an element to the empty list L empty .

[0048] S1-233, starting from the second queried demand question in the conversation context, check whether each question and answer is a direct answer to the original query or a further inquiry, and classify the subsequent questions and answers directly related to the original query into derived query units, and at the same time, store the information of the derived query units in the dictionary Q list The corresponding key value of .

[0049] S1-234, Determine the parameter position: (3), where i represents the value of each parameter in the sub-level parameter set in the empty list L empty The serial number in k ori Represented as the original query unit in the dictionary Q list The corresponding key value in k deri Represented as a derived query unit in the dictionary Q list The corresponding key value in , it can be understood that, if i=1, it means that each parameter in the sub-parameter set is small, and it is classified into the original query unit to reflect the user's initial query intention; otherwise, it is classified into the derived query unit to reflect that it is further refined or expanded information based on the original query.

[0050] S1-235. For each derived query unit, determine the correlation between it and the original query unit, and update the empty list L according to the correlation. empty and dictionary Q list to reflect the association between the parameter sets, and then use the updated parameter sets to optimize the context-related information list of the user's query input behavior.

[0051] S1-24. Extract the parent user query input behavior feature P of the demand data set fea And the sub-level user query input behavior characteristics C of the medical guidance association dataset fea In the specific implementation, according to the parent user query input behavior feature P during the user query input behavior process fea The matching degree between the query content feature data and the user feature data in MS qi And the correlation strength CS corresponding to each matching degree qi The identified extracted parent user query input behavior feature P fea ; According to the process of user query input behavior, each time the user performs a query input behavior q i The corresponding medical guidance case data, as well as the matching degree MC reflecting the correlation between the user query content feature data and the user feature data and the medical guidance case data qi The identified extracted sub-level user query input behavior features C fea .

[0052] In one embodiment of the present invention, the matching degree MSqi The calculation formula is: (4), where S rcon Represented as known recommended content R con The number of elements in the medical guidance related data, sim(q i ,r con ) represents each query input node n in the process of user query input behavior i Query input behavior q occurs at i With r con Cosine similarity function of; association strength CS qi The calculation formula is: (5), where rel(q i ,r con ) represents the input node n for each query i Query input behavior q occurs at i With r con Correlation function; matching degree MC qi The calculation formula is: (6), where Represented as medical guidance case data r data,con Quantity, sim(q i ,r data,con ) represents the input node n for each query i Query input behavior q occurs at i Medical guidance case data data,con The cosine similarity function.

[0053] Based on the above technical concept, it should be noted that the attribute matching degree refers to the key feature vector of the parent and child key feature matching vector The specific features and known recommended content R con The similarity or correlation measure between the features in the attribute matching is calculated by MS qi Based on the attribute matching, the matching degree provides the MS qi The basic data, and the matching degree MS qi This basic data is converted into an actual ranking score. The higher the attribute matching degree, the more similar the user query input behavior is to the feature vector of a certain recommended content. In this embodiment, the correlation function used is an evaluation experience constant defined by medical experts in the field when the system is built. It is preferred to adopt a value scoring system. For example, if the query is about "hallux valgus on one foot" and the recommended content mentions "hallux valgus on one foot", according to this rule, the output is the user query input behavior and r con The correlation is high.

[0054] S1-25. Synthesize parent user query input behavior feature P feaAnd the query input behavior characteristics of sub-level users C fea , we get the aggregate feature vector H that represents the user query intent and the contextual information associated with the intent in the previous session results agg ;H agg =[(MS qi ,CS qi ,MC qi )|i=1,2,...,N], where i represents the current aggregated feature vector H agg The i-th query input behavior feature set in , N represents the current aggregate feature vector H agg There are N query input behavior feature sets in MS qi ,CS qi Represented as parent matching degree and association strength, MC qi Expressed as child matching degree.

[0055] At this point, it can be understood that due to the colloquial and fragmented nature of user input questions, the customer service system lacks the ability to respond to unexpected user queries. The present invention forms a comprehensive "association" data structure based on the attributes of the parameter set to identify the recommended content that best meets the current user needs.

[0056] However, since the existing system model is based on a preset intelligent large language model to achieve prediction or recommendation, as shown in the above table list.5, the system selects an answer ("hospital location") because of a highly matched keyword ("location"), there will be a defect that the answer does not fully meet the user's complex query needs ("where is the thumb sac").

[0057] Based on this, in one embodiment of the present invention, a calibration process for establishing recommended content based on the recommended content obtained when a query behavior occurs is proposed, and a system model associated with all questions in the context of the user session is set to be updated according to each question input by the user. In this way, the recommended content that best meets the current user's needs is identified again. "Many patients suffer from hallux valgus foot deformity and bunion pain due to external factors such as wearing shoes. The bunion is located at the bottom of the big toe joint and is a structure composed of soft tissue. Its main function is to absorb and disperse the pressure of the foot. Once it occurs, it will become more and more serious, and it is easy to cause protrusions next to the big toe, overlapping toes, increased pain, etc. Severe patients will also be affected in walking and cannot heal themselves." Figure 6 As shown, the specific implementation steps include:

[0058] S1-3, set each time the context association information is generated, the known recommended content Rcon in the system is calibrated to improve the accuracy of the system recommendation. The specific process is:

[0059] S1-31. Define a set of correct fact instances based on the medical guidance data based on the consensus of medical research and clinical practice, denoted as {(E, P, C)}, where E is the set of medical guidance data noun entities that appear in the process of returning the query demand question. In the medical guidance data, common medical entities include: diseases such as colds and diabetes, drugs such as aspirin and insulin, treatment methods such as surgical treatment and drug treatment, and symptoms such as fever and cough. P is a correct relationship set. In the medical guidance data, common correct relationships include: treatment, which means that a certain drug or treatment method is effective for a certain disease, side effects, which means the adverse reactions that may occur after the use of a certain drug, diagnosis, which means the process of determining the disease through certain symptoms or test results, and prevention, which means the method of avoiding the occurrence of the disease through certain measures. C is an object set, which represents the specific object set related to the noun entity set E and the correct relationship P. If there exists an entity set E={disease, drug} and a relationship set P={treatment, side effect}, for the relationship "treatment", the object set C contains specific "treatment methods" or "treatment drugs"; for the relationship "side effect", the object set C contains specific "side effect descriptions" or "side effect types", so the object set C is expressed as: C={treatment methods, treatment drugs, side effect descriptions, side effect types}, and the resulting {(E, P, C)}={(disease, treatment, treatment methods), (disease, side effect, side effect description), (drug, treatment, treatment drugs), (drug, side effect, side effect type)}.

[0060] S1-32, the medical data that contradicts the correct fact instance is defined as the set of error relations, denoted as {(E,P e ,C e )}, E is the set of medical guidance related data noun entities that appear in the process of returning the query demand question, P e is the wrong relation set, C e A set of objects that contradicts the correct set of concrete objects.

[0061] Example: Still in the dialogue shown in list.5 above, the correct relationship: the user asks "Where is the bunion?" and expects to get information about the bunion's position in the human body. Incorrect relationship: the answer given is the address of the hospital, such as "Our hospital is at No. 123, zz Street, yy District, xx City", which does not meet the actual intention of the user's inquiry.

[0062] Therefore, S1-33 ​​is executed to construct a calculation index to evaluate the system's mastery of the correct fact instance of "the location of the thumb", so as to achieve the ability of the recommendation system to predict the object "location information" under the given entity "thumbnail" and the correct relationship "location". That is, by calculating the conditional probability Q predicted by the pre-trained model in the recommendation systemT The expected probability F of the conditional relationship between (C|E,P) and the error predicted by the model Pe [(C|E,P e )], the incorrect knowledge in the recommendation system is identified. It is understandable that the model of the original recommendation system incorrectly answers the question of "the location of the thumb" as the address of the hospital, that is, it incorrectly associates the wrong relation "hospital address" with the entity "thumbuncle".

[0063] Example: The user asks "Where is the bunion?" and wants to evaluate the system's mastery of this correct fact instance. Then K T (E,P,C)=(Q T (C|E,P)+b) / (F Pe [(C|E,P e )]+b)(7), where b is a smoothing factor to prevent division by zero. In practical applications, the value of b is usually selected based on experience to ensure the stability of the model. T (C|E,P) is the probability of predicting “position in the human body” when the recommendation system is given the relationship between “thumbone” and “position”, Q T (C|E,P)=Q T (location in the human body|buncle,location), F Pe [(C|E,P e )] is the expected probability of predicting “hospital address” when the recommendation system is given “thumbnail” and the wrong relation “address”, F Pe [(C|E,P e )]=F Pe [(hospital location|bunion,address)]. Using the above probability, calculate K T (buncle, position, position in the human body), and set the threshold y based on experience. If K T If (thumbone, position, position in the human body) is less than the threshold y, it is considered that there is wrong knowledge in the recommendation system and calibration is required. Otherwise, it is considered that the recommendation system has mastered this correct fact instance.

[0064] S1-34, calibrate the wrong knowledge and inject correct knowledge. The specific process is as follows:

[0065] S1-341. Add correct fact instances to the pre-trained model in the recommendation system, such as adding data about "the location of the thumb in the human body" to obtain an updated training data set, D new =D old ∪{(bunion, position, position in the human body)}.

[0066] S1-342, adjust the loss function of the pre-trained model by increasing the weight of the wrong predictions, and retrain the model using the updated training dataset and loss function,

[0067] θ new =θ old -η×((e -kT (y T )+b)×L old +(1-(e -kT (y T )+b))×L new )(8), where b is 0.01, θ new is the parameter of the retrained model, θ old is the parameter of the current training model, η is the learning rate, e -kT (y T )+b is the added error prediction weight, between 0 and 1, L old is the loss function value in the current iteration or training state, L new is the loss function value in the new state or iteration after updating the training model parameters, where K T When (bunion, position, position in the human body) is less than the threshold y, the weight of the wrong prediction is close to 0, which means that the model needs to add more correct fact instances. Otherwise, it means that the correct fact instances are mastered and the newly injected knowledge has little impact.

[0068] S1-343. During the training process, freeze all layers of the original model, update only the layers related to the wrong prediction weights, and use the updated parameters to train the knowledge calibration module. Finally, integrate the correct facts into the model in the form of a knowledge graph, let domain experts review the output of the model and provide feedback to improve the model.

[0069] Example: Since the model incorrectly answers the question "location of the bunion" as the address of the hospital, that is, it incorrectly associates the wrong relation "hospital address" with the entity "bunion", the result is K T (Bunion, location, bottom of big toe joint) is less than the threshold y, which means that the model does not have enough knowledge of the fact instance of "the location of the bunion". Therefore, the loss function of the model is adjusted by increasing the weight of the wrong prediction, and the correct fact instance is added to the pre-trained model. For example, if the fact "the location of the bunion in the human body is at the bottom of the big toe joint" is known, this information is added as a new data point to the training dataset, and D is obtained. new =D old ∪{(bunion, location, position in the human body)}=D old∪{(bunion, location, base of big toe joint)}. Retrain the model using the updated training dataset and loss function so that the model can better grasp the correct relationship. During this update process, freeze all layers of the original model and only update the layers related to the wrong prediction weights. With the updated parameters, only the knowledge calibration module needs to be trained to enable the model to better grasp the correct fact instance of "the location of the bunion". Finally, the correct fact instance of "the location of the bunion in the human body is at the base of the big toe joint" is integrated into the model in the form of a knowledge graph, and invite domain experts to review the output of the model to ensure that the model can now correctly answer the question "Where is the bunion?" and provide feedback to further improve the model.

[0070] like Figure 5 As shown, in one embodiment of the present invention, after the correct fact instance is integrated into the model in the form of a knowledge graph, the rapid retrieval of the subsequent recommendation system can be achieved by constructing a knowledge graph retrieval method, thereby enhancing the model's understanding of complex queries.

[0071] The specific implementation steps include: first, identifying the entities that need to be added to the knowledge graph (such as "bunion") and the relationship between them (such as "position"), and creating a node for each added entity in the knowledge graph (such as "bunion"), and at the same time creating edges for each relationship between the added entities, connecting the corresponding nodes (such as "position" connects "bunion" and "the bottom of the big toe joint"). Secondly, calling the KEYBERT tool in the Python library to process the queried demand question ("Where is the thumb?"), searching in the constructed knowledge graph according to the extracted keywords, and finding the relevant nodes (such as "bunion") representing the keywords of the queried demand question and the edges connected to it (such as "position"), as well as the nodes pointed to by the edges (such as "the bottom of the big toe joint") to generate document fragments and summaries (such as "the thumb is located at the bottom of the big toe joint of the human body"). Thirdly, the similarity between the generated document fragments and summaries and the queried demand question is calculated by formula, and the weighted recall similarity exceeds the set threshold y th The document fragment, (9), where TF(t i ,d) is the word t i The frequency in the document doc, IDF(t i ) is the word t i In the inverse document doc frequency, n is the total number of words, and the threshold y thIt is used to determine whether the similarity between the document fragment and the user's query question is high enough to be recalled. Finally, the recalled document fragment is integrated with the user's query question and keywords to form complete context information, and the appropriate interface is selected to input into the trained language model in the recommendation system. For example: the user asks "Where is the bunion?", and the recalled fragment is: "The bunion is located at the bottom of the big toe joint of the human body". The integration result: the user's question and the recalled fragment are directly integrated together to form a complete context, and the result is "The bunion is located at the bottom of the big toe joint of the human body". At the same time, when inviting domain experts to review the output of the model on this basis, the improvement can be obtained as "The bunion is a structure composed of soft tissue. It is located at the bottom of the big toe joint of the human body. Its main function is to absorb and disperse the pressure of the foot."

[0072] Example: Still the dialogue shown in List.6 above, an example of providing safe drug recommendations for pregnant users, solving the defect that the existing system cannot understand the complex query requirements of users. It is understandable that the user asked two questions: ① Drug recommendations for bunion pain, ② She is pregnant and hopes to get relatively safer drug recommendations, so the keywords obtained are "bunion pain" and "pregnant women". Then, firstly, according to the processing logic of step S1-31, for the correct fact instance of bunion pain, there is an entity set E={bunion pain, pregnant women}, a relationship set P={treatment, contraindications}, and an object set C={specific drug recommendations, safe drugs for pregnant women}. Secondly, add nodes and relationships to the above entities in the knowledge graph, then there are node 1 (bunion pain), node 2 (ibuprofen and acetaminophen), edge 1 (treatment connects bunion pain and ibuprofen), edge 2 (treatment connects bunion pain and acetaminophen), and edge 3 (pregnant women contraindications connect ibuprofen and pregnant women). Again, the KEYBERT tool in the Python library is called to process the user query question "Which medicine is better for bunion pain during pregnancy?" According to the extracted keywords (bunion pain, pregnant women, medicine), the relevant nodes and edges are retrieved in the knowledge graph (the system queries the built-in drug database to obtain information about the safety of ibuprofen, naproxen and acetaminophen during pregnancy). The recalled document fragments are integrated with the user's query demand question and keywords through the similarity calculation formula (9) to form complete context information. The final integration result is: "For bunion pain, it is generally recommended to choose over-the-counter anti-inflammatory drugs such as ibuprofen or naproxen, as well as acetaminophen to relieve pain, but since you are pregnant, ibuprofen or naproxen is not recommended because they may pose a risk to the fetus. Acetaminophen can be considered, but it is best used under the guidance of a doctor. I suggest you go to a local doctor to obtain professional treatment advice."

[0073] like Figure 3-Figure 4As shown, we will then use the method of building a time series model, taking the recommended content that best meets the current user needs as the input sample representing the true intent, and all the interaction data generated during the time period when the user's query behavior occurs as the full sample data, to predict the true intent of the user under the problems of colloquial and fragmented input.

[0074] The customer service dialogue interaction optimization method based on sequence perception also includes: S2, simulating the user's query input behavior to obtain the user's real dialogue intention. In one embodiment of the present invention, the step of executing S2 includes:

[0075] S2-1. Using the aggregated feature vector H agg Train a classification model to generate a user intent recognition sequence for user query input behavior. The specific implementation steps are:

[0076] S2-11. First, based on the aggregated feature vector H agg , for each training sample h i , determine the user's authenticity intention label through manual annotation, and then, i and its corresponding user authenticity intention label to form a new training sample pair (h i , labels), again, using all training sample pairs (h i , label) to train the classification model. At the same time, a loss function L is defined in this process to evaluate the difference between the classification model prediction and the actual label, where , where N is the total number of training sample pairs, l(h i , label) is expressed as the loss of a single training sample pair. Finally, the model parameters are updated using gradient descent to minimize the loss function L.

[0077] S2-12, based on the user intention recognition model, the user intention recognition sequence of the user query input behavior is calculated, which performs the following operations: First, the trained classification model and the aggregated feature vector H are used agg Predict the user query input behavior; secondly, for each sample h i , calculate the predicted label y output by the classification model i , and finally combine them to form the user intention recognition sequence y list ,y list =[ý1,ý2,...,ý N ,], which represents the user’s real query intention in the previous conversation.

[0078] S2-2, using the LSTM deep learning model trained based on sequence perception to judge the authenticity of the user intention recognition sequence, obtain valid user query input behavior data, and obtain the final user intention prediction. The specific implementation steps are:

[0079] S2-21. Retrieve all interaction data of different users with similar semantics in different time periods during the session context from the system log. The interaction data needs to include at least the user's query input and the answer of the online customer service system, and carefully annotate the interaction data in each session context, and then combine the processed interaction data with the corresponding annotated labels to construct the final training sample library. It should be noted that after obtaining the interaction data, it is preferred to use regular expressions to remove all irrelevant information in the interaction data, such as timestamps, user IDs, special characters and punctuation marks, before annotating.

[0080] S2-22. Use time series analysis methods to retain the timing information of interaction data. At the same time, for each time point ti in the time series, extract the user historical query features that contain the query records, interaction histories, and preference settings associated with different users when they have similar semantic queries during the session. and the manually annotated user historical query features that are manually marked according to the set rules, including the query topics and query prices associated with the query demand problems with similar semantics , construct the input feature vector It should be noted that extracting user historical query features It helps to better predict the user's real query intention.

[0081] S2-23. Formulate the number of layers, number of hidden units, and structural parameters of the input and output layers of the LSTM model architecture. Input to the LSTM model and start training modeling, h (t) (l) =f(h (t) (l-1) ,X i ), where h (t) (l) It is represented as the hidden state of the lth LSTM layer at time point ti, expressed as a vector, and f is represented as the recursive function of LSTM.

[0082] S2-24, the authenticity of the user intention recognition sequence is judged, and the process is as follows:

[0083] First, according to the formula (10), calculate the user intention recognition sequence y list The output y at each time point titi and the LSTM model outputs the behavior sequence y list The output y at each time point ti ti The difference value of

[0084] Secondly, the training sample pairs are re-divided according to the difference value: a group of sample data with difference values ​​less than the threshold is obtained through a predefined threshold, and another group of sample data with difference values ​​greater than the threshold. For the first group of sample data, it is considered that the user intention recognition sequence and the behavior sequence output by the LSTM model are close enough, and the authenticity of the sample data is high. On the contrary, for the second group of data, further manual review is required to confirm its authenticity;

[0085] Finally, the hidden vector h of the LSTM model is (t) (l) Connect to the classifier for classification and obtain the final user intent prediction.

[0086] S2-25. Perform a recall test on the user query input behavior data, take the behavior data that meets the expected goal as the output result, execute the following S3, post-dialogue, input the user's real dialogue intention into the system background, and visualize the medical information related to the queried demand question.

[0087] In one embodiment of the present invention, when S3 is implemented, the user's real conversation intention is input into the system background open source model database (medical information database), and medical related information matching the query requirements during the user conversation is retrieved from the database, and the retrieved medical information related to the user's requirements, such as plans, prices, and time periods, is output to the user.

[0088] Post-session example: real user query input behavior data 1: "I want to know the plan that has the least impact on me before I get treatment";

[0089] System: "It is recommended to use HD-ultra-minimally invasive hallux valgus correction surgery, which causes little damage, recovers quickly, does not recur, does not leave sequelae, and allows you to walk immediately after the surgery without hospitalization. You can walk after the surgery without affecting your normal life";

[0090] Real user query input behavior data 2: "What is the price?"

[0091] System: "The cost of treatment is approximately XXXX yuan, and the specific cost may vary depending on your specific situation. I can help you make an appointment for a detailed consultation, and our doctor will give you a more accurate estimate at that time."

[0092] As an embodiment of the present invention, the pre-trained model used in S1-3 protection is a large language model in the medical field, preferably the BianQue health model.

[0093] As a second aspect of the present invention, a customer service conversation interaction optimization system based on sequence perception is proposed, comprising a memory and a processor, wherein the memory comprises a customer service conversation interaction optimization program based on sequence perception, and when the customer service conversation interaction optimization program based on sequence perception is executed by the processor, the above-mentioned customer service conversation interaction optimization method based on sequence perception is implemented.

[0094] The technical scope of the present invention is not limited to the contents in the above description. Those skilled in the art can make various deformations and modifications to the above embodiments without departing from the technical idea of ​​the present invention, and these deformations and modifications should all fall within the protection scope of the present invention.

Claims

1. A customer service dialogue interaction optimization method based on sequence perception, characterized in that: It includes step S1, a pre-conversation, generating a pre-recommended context-related information list according to the conversation content, and forming an aggregated feature vector, which performs the following operations: S1-1, an initial conversation context, in which the queried demand questions generated by the user during the conversation based on the user's query input behavior are continuously obtained, and are aggregated into a demand data set S data and medical guidance related dataset G data ; S1-2, respectively from the demand data set S data and medical guidance related dataset G data In the process of constructing key feature vectors for characterizing user query input behaviors at different levels, natural language processing tools are used to perform part-of-speech tagging and entity recognition on the key feature vectors before the user performs a second query input behavior, thereby generating initial contextual association information with recommended concept attributes. S1-3, evaluating the known recommended content R in the system according to the initial contextual information con , and calibrate the erroneous knowledge, inject correct knowledge, enhance the system's understanding of complex query behavior, and then generate a context-related information list of the queried demand problem based on attribute matching sorting, wherein the context-related information list includes the user's current query intention and all context information related to the intention; extract the parent user query input behavior feature P of the demand data set fea And the sub-level user query input behavior characteristics C of the medical guidance association dataset fea , we get the aggregate feature vector H that represents the user query intent and the contextual information associated with the intent in the previous session results agg ; S2, simulate the user's query input behavior to obtain the user's real conversation intention, which performs the following operations: use the aggregated feature vector H agg The classification model is trained to generate a user intent recognition sequence, and then an LSTM model is built to determine its authenticity. Finally, a recall test is performed to select the expected user query behavior data as the output result, and S3 is executed. S3, post-dialogue, visualization of medical information related to the queried question.

2. The customer service dialogue interaction optimization method based on sequence perception according to claim 1 is characterized by: In S1-3, it is assumed that each time contextual information is generated, the known recommended content R in the system con A recommendation calibration is performed to improve the accuracy of the system recommendation. The specific process is as follows: First, a set of correct fact instances is defined based on the medical guidance related data based on the consensus of medical research and clinical practice, denoted as {(E, P, C)}, where E is the medical guidance related data noun entity set that appears in the process of the system returning the query demand question, P is the correct relationship set, and C is the specific object set related to the noun entity set E and the correct relationship P; the medical data that contradicts the correct fact instance is defined as the wrong relationship set, denoted as {(E, P e ,C e )},P e is the wrong relation set, C e is a set of objects that contradicts the correct set of specific objects; secondly, the conditional probability Q predicted by the pre-trained model in the calculation system T The expected probability F of the conditional relationship between (C|E,P) and the error predicted by the model Pe [(C|E,P e )], identify the wrong knowledge in the system; then, add the correct fact instances to the pre-trained model in the system to obtain the updated training data set, adjust the loss function of the pre-trained model by adding the wrong prediction weight, and retrain the model using the updated training data set and loss function, θ new =θ old -η×((e -kT (y T )+b)×L old +(1-(e -kT (y T )+b))×L new ), where b is 0.01, θ new is the parameter of the retrained model, θ old is the parameter of the current training model, η is the learning rate, e -kT (y T )+b is the added error prediction weight; L old is the loss function value in the current iteration or training state, L new In order to obtain the loss function value under the new state or iteration after updating the model parameters, finally, the correct fact instances are integrated into the model in the form of a knowledge graph, allowing domain experts to review the output of the model and provide feedback to improve the model, completing the calibration injection.

3. The customer service dialogue interaction optimization method based on sequence perception according to claim 2 is characterized in that: After the correct fact instance is integrated into the model in the form of a knowledge graph, it is also necessary to implement rapid retrieval in the subsequent system by building a knowledge graph retrieval method. The specific process is as follows: first, identify the knowledge graph entities that need to be added and the relationships between them, and create a node for each added entity in the knowledge graph, and at the same time create an edge for the relationship between each added entity to connect the corresponding nodes; Secondly, the KEYBERT tool in the Python library is called to process the queried demand question. According to the extracted keywords, the constructed knowledge graph is searched to find the relevant nodes and edges connected to the keywords representing the queried demand question, as well as the nodes pointed to by the edges, to generate document fragments and summaries; thirdly, the similarity between the generated document fragments and summaries and the queried demand question is calculated, and the weighted recall similarity exceeds the set threshold y th The document fragment, , where TF(t i ,d) is the word t i The frequency in the document doc, IDF(t i ) is the word t i In the inverse document doc frequency, n is the total number of words; finally, the recalled document fragments are integrated with the user's query questions and keywords to form complete context information, and the appropriate interface is selected to input into the trained model in the recommendation system.

4. The customer service dialogue interaction optimization method based on sequence perception according to claim 1 is characterized in that: In S1-2, the matching function is used to construct the key feature vector. The specific process is as follows: Suppose the required data set S data There are N queried demand questions, each of which is composed of user feature data representing user personal information and query content feature data representing user session content. Suppose the medical guidance related data set G data contains M medical guidance recommendation demand data; according to the formula, from the demand data set S data Identify and extract the parent key feature vector : , where Represented as the i-th queried demand problem, e par Represented as the set user query input behavior matching function f mat The set of all parent parameters about the parent feature matching strategy in; At the same time, according to the formula, from the medical guidance association data set G data Identify and extract the sub-level key feature vector : , where It is represented as the jth medical guidance recommendation demand data that is approximately associated with the user's query input behavior, e chi Represents the set user query input behavior matching function g mat The sub-level parameter set of all sub-level feature matching strategies in , M represents the medical guidance association dataset G data The amount of medical guidance recommendation demand data generated.

5. The customer service dialogue interaction optimization method based on sequence perception according to claim 1 or 4, characterized in that: The specific process of generating a list of contextual information is as follows: First, use the natural language processing (NLP) tool to perform a search on the parent key feature vectors. and the child key feature vector The context-related information representing the queried demand question is subjected to part-of-speech tagging and entity recognition; secondly, a function for customizing the number a1 of noun entities appearing in the process of returning the queried demand question is calculated, and a quantity threshold a2 for defining the number of noun entities is set, and the function is compared with the number a1 of noun entities appearing in the returned query. If the number a1 of noun entities appearing is greater than the quantity threshold a2, it is considered that the context-related information in the queried demand question has the attribute of recommendation concept, otherwise it is not. Thirdly, using the identified concept attributes, we recommend content R based on contextual information and known content. con The similarity of the parent key feature vector and the child key feature vector is calculated to obtain the attribute matching degree of each context-related information, so as to generate a list of context-related information representing the queried demand problem of the user during the session according to the sorting of the attribute matching degrees.

6. The customer service dialogue interaction optimization method based on sequence perception according to claim 1 is characterized by: In S2, the aggregated feature vector H is used agg The specific process of training a classification model is as follows: First, based on the aggregated feature vector H agg , for each training sample h i , determine the user's authenticity intention label through manual annotation; Secondly, the training sample h i and its corresponding user authenticity intention label to form a new training sample pair (h i , labels); again, using all training sample pairs (h i , label) to train the classification model. At the same time, a loss function L is defined in the process to evaluate the difference between the classification model prediction and the actual label; finally, the model parameters are updated using gradient descent to minimize the loss function L.

7. The customer service dialogue interaction optimization method based on sequence perception according to claim 1 or 6, characterized in that: The specific process of generating the user intention recognition sequence of the user query input behavior is as follows: First, the trained classification model and the aggregated feature vector H are used to agg Predict the user query input behavior; secondly, for each sample h i , calculate the predicted label y output by the classification model i ; Finally, combine them to form the user intention recognition sequence y list ,y list =[ý1,ý2,...,ý N ,], which represents the user’s real query intention in the previous conversation.

8. The customer service dialogue interaction optimization method based on sequence perception according to claim 1 is characterized by: In S2, the specific process of training the LSTM model is as follows: first, retrieve all interaction data with similar semantics to the queried demand question in different time periods during the session context of different users from the system log, and perform detailed category labeling on the interaction data in each session context, then combine the processed interaction data with the corresponding labeled labels to construct the final training sample library; Secondly, the time series analysis method is used to retain the time series information of the interaction data, and at the same time, for each time point ti in the time series, the user historical query features including query records, interaction history, and preference settings associated with the queried demand problem are extracted respectively. and the historical query features of users that are manually labeled according to the set rules, including the query topics and query prices associated with the query demand problem. , construct the input feature vector Finally, the number of layers, number of hidden units, and structural parameters of the input and output layers of the LSTM model architecture are determined, and the input feature vector Input into the LSTM model and start training modeling, h (t) (l) =f(h (t) (l-1) ,X i ), where h (t) (l) It is represented as the hidden state of the lth LSTM layer at time point ti, expressed as a vector, and f is represented as the recursive function of LSTM.

9. The customer service dialogue interaction optimization method based on sequence perception according to claim 1, characterized in that: In S2, the specific process of judging the authenticity of the user intention recognition sequence is as follows: first, according to the formula , calculate the user intention recognition sequence y list The output y at each time point ti ti and the LSTM model outputs the behavior sequence y list The output y at each time point ti ti ; secondly, the training sample pairs are re-divided according to the difference value: a group of sample data with difference values ​​less than the threshold is obtained through a predefined threshold, and another group of sample data with difference values ​​greater than the threshold is obtained. For the first group of sample data, it is considered that the user intention recognition sequence and the behavior sequence output by the LSTM model are close enough, and the authenticity of the sample data is high. On the contrary, for the second group of data, further manual review is required to confirm its authenticity; finally, the hidden vector h of the LSTM model is (t) (l) Connect to the classifier for classification and obtain the final user intent prediction.

10. A customer service dialogue interaction optimization system based on sequence perception, characterized in that: include: A memory and a processor, wherein the memory includes a sequence-awareness-based customer service conversation interaction optimization program, and the sequence-awareness-based customer service conversation interaction optimization program, when executed by the processor, implements the customer service conversation interaction optimization method described in any one of claims 1 to 9.

Images