Whole-process engineering consultation project information interaction method and system

Abstract

The invention discloses a whole-process engineering consultation project information interaction method and system, and belongs to the technical field of information interaction, and the method specifically comprises the steps: recognizing a user role in an engineering consultation project, collecting the historical data of a target role in the project for the target role, constructing a corresponding role corpus, and carrying out the recognition of the user role based on a universal pre-training language model. Performing directional parameter adjustment by utilizing the role corpus, generating a role customization interaction model, and when a user initiates an interaction request, calling the role customization interaction model based on a role identity to generate response content and outputting the response content; according to the method, the professionality, the semantic matching degree and the multi-role adaptive capacity in the information interaction process are remarkably improved, and the problems of information understanding deviation and low communication efficiency caused by multiple participants and large context difference in an engineering consultation project are effectively solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of information interaction, and in particular relates to a method and system for information interaction of a whole-process engineering consulting project. Background Art

[0002] Full-process engineering consulting refers to providing owners or project managers with integrated, comprehensive consulting services throughout the lifecycle of a construction project, encompassing feasibility studies, survey and design, bidding and procurement, construction implementation, final acceptance, and operations and maintenance. This model involves numerous stakeholders, typically including the owner, design firm, construction company, supervision firm, cost consulting firm, and operations and maintenance management team.

[0003] In actual engineering projects, different participants have significant differences in their responsibilities, task boundaries, and information needs. With the development of artificial intelligence technology, natural language processing technology, and large language models, some engineering management systems have begun to introduce AI-based question-and-answer interaction modules to assist project participants in obtaining information. However, the AI ​​question-and-answer or knowledge graph systems commonly used in existing technologies are mostly based on general question-and-answer engines or unified corpus models, which make it difficult to accurately cover the professional terminology, expression styles, and problem backgrounds of various roles. As a result, the answers often deviate from role requirements, are unprofessional, or are inconsistent with the context.

[0004] Therefore, there is an urgent need for an information interaction method for the entire engineering consulting project. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the present invention proposes a whole-process engineering consulting project information interaction method and system.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The information exchange method for the whole process engineering consulting project includes:

[0008] Identify user roles in engineering consulting projects;

[0009] For the target role, collect its historical data in the project and build the corresponding role corpus;

[0010] Based on a general pre-trained language model, the character corpus is used to perform parameter-oriented adjustment to generate a character-customized interaction model;

[0011] When a user initiates an interaction request, the role-customized interaction model is called based on the user's role identity to generate and output the response content.

[0012] Specifically, for the target role, historical data of the target role in the project is collected to build a corresponding role corpus, including:

[0013] Collect historical data of engineering projects related to each role;

[0014] Perform semantic segmentation, named entity recognition, and intent classification on the collected historical data of engineering projects to extract data related to roles and characters;

[0015] Based on cluster analysis, the extracted results are classified by subject to generate a preliminary corpus collection;

[0016] Based on the expert knowledge graph, the preliminary corpus collection is semantically enhanced and expanded with hyponyms and hyponyms to construct a structured role corpus.

[0017] Specifically, the method of performing parameter-oriented adjustment based on the general pre-trained language model and utilizing the role corpus to generate a role-customized interaction model includes:

[0018] Based on the role corpus, a role task guidance set is constructed, wherein the task guidance set includes questions, scenarios, semantic information paths, and semantic labels;

[0019] Adapt the roles and assign dynamic weights to training samples during the language model training process based on the semantic labels in the role task guidance set.

[0020] The language model is tested using evaluation samples containing character identity prompts to detect how well the output matches the character.

[0021] The language model is subjected to parameter-directed adjustment according to the matching degree to obtain a trained language model, namely, a role-customized interaction model.

[0022] Specifically, based on the role corpus, a role task guidance set is constructed, including:

[0023] Perform task clustering on the role corpus and divide it into task clusters;

[0024] Combined with the engineering industry knowledge graph, it matches the semantic path relationship between entity words in the role corpus and task nodes;

[0025] Extracting corpus segments whose path weights in semantic path relationships are higher than a preset weight threshold, wherein the corpus segments contain entity words and task node content;

[0026] The extracted corpus fragments are role-copied and task intent annotated to generate a structured role-task guidance set.

[0027] Specifically, the adaptation of the roles, according to the semantic labels in the role task guidance set, assigns dynamic weights to the training samples during the language model training process, including:

[0028] According to the corpus characteristics of the target role, the role language style is extracted from the role task guidance set;

[0029] Match each sentence of the training sample with a predefined semantic label to construct a label matching vector. The predefined semantic labels include task type label, character language style label, term density label, and context complexity label.

[0030] Perform a dot product operation on the label matching vector and the predefined label weight level vector to obtain the initial semantic importance score of each training sample;

[0031] Calculate the overlap between the semantic path in the training sample and the semantic path in the role-task guidance set, and make a multiplicative adjustment to the initial semantic importance score to obtain a weighted semantic importance score;

[0032] During the language model training process, the training sample weights are dynamically modified based on the real-time loss convergence speed or gradient change trend and combined with the weighted semantic importance score to obtain the dynamic weights of the training samples.

[0033] Specifically, the language model is tested using an evaluation sample containing role identity prompt words to detect its matching degree with the role in the output content, and the language model parameter is adjusted according to the matching degree, including:

[0034] Constructing an evaluation sample, wherein each input sentence in the evaluation sample contains a clear role identity prompt word;

[0035] Using the trained language model, input the evaluation sample for inference and generate response text;

[0036] Perform role semantic matching analysis on the generated response text;

[0037] The matching analysis results are converted into a role consistency score and compared with the preset role semantic baseline. If the score is lower than the threshold, corpus adjustment or incremental training is performed.

[0038] Specifically, when a user initiates an interaction request, the role-customized interaction model is called based on the user's role identity to generate and output response content, including:

[0039] Receive user interaction requests and parse their identity information, determine the user's role category based on the identity information or historical behavior characteristics, and generate a role identification code;

[0040] Input the user input request and role identification code into the role customization interaction model to generate the current user's response text;

[0041] Perform role consistency review on the response text, and based on the role semantic review results, output the response text to the user terminal or perform response correction processing, and output the corrected response text to the user terminal.

[0042] Specifically, identifying user roles in an engineering consulting project includes:

[0043] Receive a user access request and extract identity information associated with the user access request, including user account information, access terminal information, or historical behavior records;

[0044] Analyzing the matching relationship between the identity information and the predefined role set based on the preset role identification rules;

[0045] Users are divided into role categories corresponding to their identity information, and role labels representing the role categories are generated.

[0046] The whole-process engineering consulting project information interaction system is used to implement the above-mentioned whole-process engineering consulting project information interaction method, including: a role identification module, a corpus construction module, a model generation module and a content output module;

[0047] The role identification module is used to identify user roles in engineering consulting projects;

[0048] The corpus building module is used to collect historical data of the target role in the project and build a corresponding role corpus;

[0049] The model generation module is used to perform parameter-oriented adjustment based on the general pre-trained language model using the role corpus to generate a role-customized interaction model;

[0050] The content output module is used to call the role-customized interaction model based on the role identity of the user to generate and output response content when the user initiates an interaction request.

[0051] Specifically, the model generation module includes: a task guidance set construction unit, a weight allocation unit, and a training and orientation adjustment unit;

[0052] The task guidance set construction unit is used to construct a role task guidance set based on the role corpus;

[0053] The weight allocation unit is used to adapt the role and assign dynamic weights to training samples during the language model training process according to the semantic labels in the role task guidance set;

[0054] The training and directional adjustment unit is used to assign dynamic weights to training samples and perform training, use the output results of the evaluation samples to detect the matching degree, and perform parameter directional adjustment on the language model according to the matching degree.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] The present invention proposes a method and system for information interaction in the entire engineering consulting project process. The method determines the role category based on user identity information, constructs a role task guidance set containing semantic labels such as task type and language style, uses sample weighted training and dynamic parameter adjustment to perform role-customized fine-tuning on the general language model, and introduces role consistency assessment feedback to optimize the model, thereby significantly improving the professionalism, semantic matching and multi-role adaptability in the information interaction process, and effectively solving the problems of information comprehension deviation and low communication efficiency caused by the large number of participants and large context differences in engineering consulting projects. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A flow chart of the information interaction method for the entire engineering consulting project provided by the present invention;

[0058] Figure 2 Generate a flow chart for the role customization interaction model provided by the present invention;

[0059] Figure 3 A flowchart of text generation provided by the present invention;

[0060] Figure 4 This is the architecture diagram of the full-process engineering consulting project information interaction system provided by the present invention. DETAILED DESCRIPTION

[0061] The present application is described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present application, but are not intended to limit the present application in any form. It should be noted that those skilled in the art may make several variations and improvements without departing from the scope of the present application. These all fall within the scope of protection of the present application.

[0062] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0063] It should be noted that, if there is no conflict, the various features in the embodiments of the present application can be combined with each other and are all within the scope of protection of the present application. In addition, although the functional modules are divided in the device schematic and the logical order is shown in the flow chart, in some cases, the steps shown or described can be performed in a different order than the module division in the device or the order in the flow chart. In addition, the words "first", "second", "third", etc. used in this application do not limit the data and execution order, but only distinguish between the same items or similar items with basically the same functions and effects.

[0064] Unless otherwise defined, all technical and scientific terms used in this specification have the same meanings as those commonly understood by those skilled in the art to which this application belongs. The terms used in this specification and in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application. The term "and / or" as used in this specification includes any and all combinations of one or more of the relevant listed items.

[0065] Example 1

[0066] See also Figure 1-Figure 3 The present invention provides an embodiment: a method for interacting with information on a full-process engineering consulting project, comprising the following specific steps:

[0067] Step S1: Identify user roles in the engineering consulting project.

[0068] BIM (Building Information Model) is one of the engineering data sources. By extracting structured information such as component attributes, construction progress, spatial relationships, etc. contained in the BIM model into semantic fragments and mapping them to corresponding task nodes and terminology labels in the role task guidance set, it serves as an important supplement to the construction of role corpus. For example, for the construction unit role, information such as floor-component-construction status is extracted from the BIM model to form query expressions related to on-site management and resource scheduling, thereby enhancing the model's ability to understand spatial semantics and task status, and improving the practicality and response accuracy of the role-customized interaction model in engineering scenarios.

[0069] In actual deployment, when a user initiates an information request through the engineering management platform, the access interceptor module or the front-end access gateway is first triggered to obtain the basic information bound to the request.

[0070] The specific steps of step S1 are:

[0071] Step S101: receiving a user access request, and extracting identity information associated with the user access request, including user account information, access terminal information, or historical behavior records.

[0072] In this embodiment, by parsing the user's Token or Session information, extracting its account ID, and then calling the user information table, the user's basic identity information, such as role code, department, project ID, etc., is quickly located. If the account information is not sufficient to clarify the role attribution, the terminal features and user's behavior trajectory are further extracted to construct a set of behavioral feature vectors.

[0073] Step S102: Based on a preset role recognition rule, the matching relationship between the identity information and the predefined role set is analyzed.

[0074] In this embodiment, a role mapping rule table is preset, and based on the extracted user identity information, it is compared with the above rule triggering conditions item by item, and the comparison adopts methods such as Boolean matching, interval matching or fuzzy matching.

[0075] Step S103: Categorize users into role categories corresponding to their identity information, and generate role tags representing the role categories.

[0076] In this embodiment, if a user has multiple features that match multiple roles at the same time, a weighted scoring mechanism is used to assign different weight scores to each rule item. The final summary score is used as the role fit, and the role with the highest score is identified and classified as the user's main role. The role judgment result is converted into a structured label format.

[0077] Step S2: For the target role, collect its historical data in the project and build a corresponding role corpus.

[0078] The specific steps of step S2 are:

[0079] Step S201: Collect historical data of engineering projects related to each role.

[0080] Step S202: performing semantic segmentation, named entity recognition, and intent classification on the collected historical data of the engineering project, and extracting data related to the characters.

[0081] In this embodiment, cost business data is used as one of the key role corpus sources throughout the project approval, design, bidding, construction to final settlement stages. By extracting information such as the bill of quantities, quota data, change records, payment nodes and settlement logic, an association between semantic paths and task nodes is established and included in the role task guidance set. At the same time, combined with semantic tags such as cost control, budget deviation, change pricing, etc., the model's ability to understand cost-related tasks is enhanced, and deep penetration of cost corpus is achieved throughout the entire process, thereby improving the professional response effect of roles such as owners and cost consultants in cost management questions and answers.

[0082] Specifically, semantic segmentation is first performed using sentence splitting rules and engineering terminology segmentation algorithms. Based on the engineering context dictionary, patterns such as "Construction status is as follows" and "Acceptance opinion is as follows" are identified to divide the text into logical language blocks. Each paragraph of text is processed by a pre-trained NER model to identify entities related to the engineering task. The divided logical language blocks are then sent to the intent classifier for label judgment. By comparing entity types, intent categories, and role responsibilities, the logical language blocks are annotated with role labels, and only paragraphs associated with roles are retained as training samples.

[0083] Step S203: Based on cluster analysis, the extraction results are classified by subject to generate a preliminary corpus set.

[0084] In this embodiment, the extracted data related to the role is converted into a semantic vector through an embedding model. The vector reflects the contextual meaning, tone, entity words and other information of the data related to the role. Clustering is performed based on the Euclidean distance or cosine similarity between the semantic vectors, and a label is generated for each cluster center or cluster center, such as risk warning, personnel scheduling, design change response, etc. The label is used as the semantic subject classification mark of the corpus, and each type of data and its cluster label are packaged and stored to obtain a preliminary corpus set, whose fields include {role ID, semantic topic, original data, intent classification, named entity, etc.}.

[0085] Step S204: Based on the expert knowledge graph, semantic enhancement and hyponym expansion are performed on the preliminary corpus set to construct a structured role corpus.

[0086] In this embodiment, the preliminary corpus set has been clustered and has semantic topic labels, but the following problems still exist: the granularity of terms in the corpus varies, some are very detailed, and some are too broad, such as cold-drawn threaded steel bars and structural construction; the same corpus block may not be clustered into the same topic due to different expressions.

[0087] Specifically, all professional terms appearing in the preliminary corpus are first scanned, and through term segmentation and entity recognition, they are mapped to nodes in the expert knowledge graph, such as steel section welding is mapped to the welding process node; if a term is a subordinate node, its superordinate term is retrieved upward through the graph and inserted into the original corpus as a supplementary field, and the original fragment is added to the graph supplementary content to form a structural field, such as {original text, main term, superordinate term, subordinate term, role, task type, graph path}, thereby promoting the corpus from a fragment text to a semantic object.

[0088] Step S3: Based on the general pre-trained language model, the character corpus is used to perform parameter-oriented adjustment to generate a character-customized interaction model.

[0089] The specific steps of step S3 are:

[0090] Step S301: constructing a role task guidance set based on the role corpus, wherein the task guidance set includes questions, scenarios, semantic information paths and semantic tags.

[0091] The specific steps of step S301 are:

[0092] Step S3011: Perform task clustering on the role corpus and divide it into task clusters.

[0093] In this embodiment, each piece of corpus, i.e., a sentence or a paragraph, is converted into a semantic vector through a universal language encoder, and the semantic distance between different corpora is quantified. An unsupervised clustering algorithm is used to cluster the corpora into several groups based on the similarity between the semantic vectors. Each group is a task cluster, representing a specific role business task semantic domain. Each cluster automatically generates a task label by analyzing the high-frequency keywords, entity categories and intent types in its central corpus.

[0094] Step S3012: Combined with the engineering industry knowledge graph, match the semantic path relationship between entity words in the role corpus and task nodes.

[0095] In this example, representative professional terms are extracted from the entities in each character corpus. Node matching is performed within the knowledge graph to identify synonymous or near-synonymous nodes. For each graph node matched to an entity term, the graph path is traversed upward or downward to find connections to predefined task nodes. For example, formwork support can be traced back to concrete pouring preparation and then to the structural construction phase. The number of hops, relationship type, and node semantic similarity of each path are used to calculate the path semantic weight.

[0096] Step S3013: extracting corpus segments whose path weights in the semantic path relationship are higher than a preset weight threshold, wherein the corpus segments include entity words and task node contents.

[0097] In this embodiment, several semantic path scoring rules are set. For example, if the path consists of only one hop, it is set to a high weight; if the path is more than three hops and involves abstract nodes in the middle, it is set to a low weight. In addition, the relationship types involved in the path are also weighted, and finally a weight score between 0 and 1 is generated for each semantic path; a weight threshold is set, and only the corpus fragments corresponding to the semantic paths with weights greater than this threshold are retained.

[0098] Specifically, the preset weight threshold is determined by those skilled in the art through cross-validation or simulation experiments.

[0099] Step S3014: Perform role copying and task intent annotation on the extracted corpus segments to generate a structured role-task guidance set.

[0100] In this embodiment, the original fragments extracted from the corpus are often expressed from the perspective of only one role. A rule library is introduced to copy the extracted corpus fragments into semantically equivalent expressions for other roles. For example, the construction party's suggestions for rectification are converted into a rectification notice issued by the supervision party or the owner's attention to current construction hazards; combined with the graph path, keywords and template library in the previous stage, the extracted corpus fragments are automatically labeled with task intent, and each generated corpus is converted into a structured record to obtain a structured role task guidance set.

[0101] Step S302: adapt the role and assign dynamic weights to training samples in the process of training the language model according to the semantic labels in the role task guidance set.

[0102] The specific steps of step S302 are:

[0103] Step S3021: extracting the character's language style from the character task guidance set based on the corpus features of the target character.

[0104] In this embodiment, language feature analysis is performed on each corpus to extract: keyword preference, sentence pattern, professional term distribution, and tone intensity and tendency. The above multi-dimensional features are converted into a unified style vector to represent the average expression characteristics of a role's language. For example, the construction party's expression style is: verb-dominant + operational terminology + clear time plan, while the supervision party tends to use a judgment + risk warning + suggestion tone; a general expression template is generated based on the clustering results of the role style vectors.

[0105] Step S3022: Match each sentence of the training sample with a predefined semantic label to construct a label matching vector. The predefined semantic labels include task type labels, role language style labels, term density labels, and context complexity labels.

[0106] In this embodiment, a predefined list of semantic tags is set by an expert or a rule base, such as a task type tag set, a style tag set, a term density level, and a context complexity level; each sentence of the training sample is matched according to the predefined semantic tags, and each corpus is mapped into a semantic tag vector.

[0107] Step S3023: performing a dot product operation on the label matching vector and the predefined label weight level vector to obtain an initial semantic importance score of each training sample.

[0108] In this embodiment, for each training sample, there is already a label matching vector. Those skilled in the art set a label weight level vector based on the training objective, multiply the two vectors by the corresponding dimensions and then sum them to obtain the initial semantic importance score of the training sample.

[0109] Step S3024: Calculate the overlap between the semantic path in the training sample and the semantic path in the role task guidance set, and make a multiplication adjustment on the initial semantic importance score to obtain a weighted semantic importance score.

[0110] In this embodiment, for each training sample, the main entity words and context task verbs are identified, and the semantic path of the training sample is constructed, for example: construction elevator failure → transportation interruption → impact on progress; it is compared with the typical path set in the role task guidance set. For example, the construction elevator-related path already has construction elevator → temporary vertical transportation → operation coordination task, and two nodes overlap. The path matching algorithm is used to obtain an overlap score of 0-1, and the overlap score is multiplied by the original importance score to obtain a weighted semantic importance score.

[0111] Step S3025: During the language model training process, based on the real-time loss convergence speed or gradient change trend, combined with the weighted semantic importance score, the training sample weight is dynamically modified to obtain the dynamic weight of the training sample.

[0112] In this embodiment, the loss decline curve of each training sample in the training iteration is recorded. If the decline slows down or fluctuates repeatedly, it means that the model has not fully learned the training sample. By tracking the amplitude of the parameter gradient change, it is judged whether the sample continues to affect the model. If the gradient tends to zero for a long time, it means that the training sample has been learned, and a dynamic weight adjustment logic is constructed: high semantic importance + poor learning → increase the weight; high semantic importance + too fast learning → moderate weakening; low semantic importance + good learning → weight is set to a low value, and finally the dynamic weight of the training sample is output.

[0113] The benefits of this step are: by extracting the role language style from the role task guidance set, and combining multi-dimensional semantic labels such as task type, language style, term density and context complexity to construct a label matching vector, the semantic features of each training sample can be accurately identified; then the initial importance score is obtained by performing a dot product operation with the label weight level, and the structural overlap between the sample and the role task semantic path is combined to further realize the semantic enhancement evaluation of the sample value. On this basis, the loss convergence speed and gradient change trend in the training process are introduced to realize the dynamic adjustment of the training sample weights, so that the training process pays more attention to samples that are difficult to master but strongly related to the role, effectively improving the model's performance in terms of role language expression, task context, etc.

[0114] Step S303: Using an evaluation sample containing role identity prompt words to test the language model, and detecting its matching degree to the role in the output content.

[0115] The specific steps of step S303 are:

[0116] Step S3031: constructing an evaluation sample, wherein each input sentence in the evaluation sample contains a clear role identity prompt word;

[0117] In this embodiment, key participating roles are identified from the full-process engineering consulting scenario, such as owners, construction parties, design units, supervision units, third-party auditors, etc. Based on the corpus and semantic task templates collected in the early stage, the issues or expressions that each type of role often pays attention to in actual engineering are extracted, and several language templates are designed for each type of role to ensure that the role identity is clearly indicated in the input. The above-mentioned prompt words are embedded in typical expressions to form evaluation samples.

[0118] Step S3032: Using the trained language model, input the evaluation sample for inference and generate a response text;

[0119] Step S3033: Perform role semantic matching analysis on the generated response text.

[0120] It should be noted that the matching degree is calculated using a weighted scoring method.

[0121] The benefits of this step are: by embedding role identity cues in the evaluation samples, it is possible to accurately test whether the model can understand and reflect the target role's language style, terminology habits, and task focus during the generation process; combined with role semantic matching analysis, the model output is compared with the role's preset semantic baseline and quantified as a role consistency score.

[0122] Step S304: performing parameter-oriented adjustment on the language model according to the matching degree to obtain a trained language model, i.e., a role-customized interaction model;

[0123] The matching analysis results are converted into a role consistency score and compared with the preset role semantic baseline. If the score is lower than the threshold, corpus adjustment or incremental training is performed.

[0124] In this embodiment, by comparing the vocabulary, sentence structure, intention and other dimensions in the output sentence, a score is given to determine whether the output is consistent with the input role, which is used as the role consistency score; each role corresponds to a set of semantic baseline vectors, including common terms, task focus, and expression style. If these elements are lacking in the output, or other role-specific language appears, the consistency score is low; if it is detected that the consistency score of a certain role is consistently low, analyze whether there are problems such as scarcity of role samples and incomplete terminology coverage in the corpus, and adjust the corpus or conduct incremental training in the direction of the role.

[0125] The benefits of this step are: in the process of building a role-customized interaction model based on a general pre-trained language model, by introducing a role corpus and a task guidance set, it not only achieves deep modeling of different engineering roles (such as owners, construction, supervision, etc.) in terms of language expression, task focus, and terminology usage, but also realizes sample differentiation learning and knowledge focus during the model training process through semantic labels and dynamic weight mechanisms, significantly improving the model's sensitivity to specific role task contexts. At the same time, by designing evaluation samples with role identity prompt words, it is possible to quantify the accuracy of the model's role matching, and to perform targeted fine-tuning of the model parameters based on matching feedback, so that the final generated role-customized interaction model has the ability to recognize roles and adapt to language styles.

[0126] Step S4: When a user initiates an interaction request, the role-customized interaction model is called based on the user's role identity to generate and output response content.

[0127] The specific steps of step S4 are:

[0128] Step S401: receiving a user interaction request and parsing the user's identity information, determining the user's role category based on the identity information or historical behavior characteristics, and generating a role identification code.

[0129] Step S402: Input the user input request and the role identification code into the role customization interaction model to generate a response text of the current user.

[0130] Step S403: Perform role consistency review on the response text, and based on the role semantic review result, output the response text to the user terminal or perform response correction processing, and output the corrected response text to the user terminal.

[0131] The benefits of this step are: by combining user identity recognition, role-customized model calling and response consistency review, personalized and professional question and answer output for multiple parties can be achieved. It can accurately identify user roles such as owners, designers, construction, supervisors, etc., and dynamically call the corresponding language model to ensure that the output content is consistent with the role requirements in terms of terminology usage, task focus, language style, etc. At the same time, role consistency review is introduced to avoid generating answers that are inconsistent with the role identity, thereby improving the accuracy of the interaction and role adaptability.

[0132] Example 2

[0133] See also Figure 4 ,Another embodiment provided by the present invention: a full-process engineering consulting project information interaction system, comprising: a role identification module, a corpus construction module, a model generation module and a content output module;

[0134] The role identification module is used to identify user roles in engineering consulting projects;

[0135] The corpus building module is used to collect historical data of the target role in the project and build a corresponding role corpus;

[0136] The model generation module is used to perform parameter-oriented adjustment based on the general pre-trained language model using the role corpus to generate a role-customized interaction model;

[0137] The content output module is used to call the role-customized interaction model based on the role identity of the user to generate and output response content when the user initiates an interaction request.

[0138] Model generation module, including: task guidance set construction unit, weight distribution unit, training and orientation adjustment unit;

[0139] The task guidance set construction unit is used to construct a role task guidance set based on the role corpus;

[0140] The weight allocation unit is used to adapt the role and assign dynamic weights to training samples during the language model training process according to the semantic labels in the role task guidance set;

[0141] The training and directional adjustment unit is used to assign dynamic weights to training samples and perform training, use the output results of the evaluation samples to detect the matching degree, and perform parameter directional adjustment on the language model according to the matching degree.

[0142] In addition, the parts of the above technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0143] The above-described specific embodiments further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above description is merely a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. The whole process engineering consulting project information interaction method is characterized by: include: Identify user roles in engineering consulting projects; For the target role, collect its historical data in the project and build the corresponding role corpus; Based on a general pre-trained language model, the character corpus is used to perform parameter-oriented adjustment to generate a character-customized interaction model, wherein the parameter-oriented adjustment is based on the matching degree of the output content to the character, and corpus adjustment or incremental training is performed; When a user initiates an interaction request, the role-customized interaction model is called based on the user's role identity to generate and output the response content.

2. The whole-process engineering consulting project information interaction method according to claim 1, characterized in that: For the target role, historical data of the target role in the project is collected to build a corresponding role corpus, including: Collect historical data of engineering projects related to each role; Perform semantic segmentation, named entity recognition, and intent classification on the collected historical data of engineering projects to extract data related to roles and characters; Based on cluster analysis, the extracted results are classified by subject to generate a preliminary corpus collection; Based on the expert knowledge graph, the preliminary corpus collection is semantically enhanced and expanded with hyponyms and hyponyms to construct a structured role corpus.

3. The whole-process engineering consulting project information interaction method according to claim 2, characterized in that: The method of performing parameter-oriented adjustment based on the general pre-trained language model and utilizing the role corpus to generate a role-customized interaction model includes: Based on the role corpus, a role task guidance set is constructed, wherein the task guidance set includes questions, scenarios, semantic information paths, and semantic labels; Adapt the roles and assign dynamic weights to training samples during language model training based on the semantic labels in the role-task guidance set; The language model is tested using evaluation samples containing character identity prompts to detect how well the output matches the character. The language model is subjected to parameter-directed adjustment according to the matching degree to obtain a trained language model, namely, a role-customized interaction model.

4. The whole-process engineering consulting project information interaction method according to claim 3 is characterized in that: Based on the role corpus, a role task guidance set is constructed, including: Perform task clustering on the role corpus and divide it into task clusters; Combined with the engineering industry knowledge graph, it matches the semantic path relationship between entity words in the role corpus and task nodes; Extracting corpus segments whose path weights in semantic path relationships are higher than a preset weight threshold, wherein the corpus segments contain entity words and task node content; The extracted corpus fragments are role-copied and task intent annotated to generate a structured role-task guidance set.

5. The whole-process engineering consulting project information interaction method according to claim 4 is characterized in that: The adaptation of the roles, according to the semantic labels in the role task guidance set, assigns dynamic weights to the training samples during the language model training process, including: According to the corpus characteristics of the target role, the role language style is extracted from the role task guidance set; Match each sentence of the training sample with a predefined semantic label to construct a label matching vector. The predefined semantic labels include task type label, character language style label, term density label, and context complexity label. Perform a dot product operation on the label matching vector and the predefined label weight level vector to obtain the initial semantic importance score of each training sample; Calculate the overlap between the semantic path in the training sample and the semantic path in the role-task guidance set, and make a multiplicative adjustment to the initial semantic importance score to obtain a weighted semantic importance score; During the language model training process, the training sample weights are dynamically modified based on the real-time loss convergence speed or gradient change trend and combined with the weighted semantic importance score to obtain the dynamic weights of the training samples.

6. The whole-process engineering consulting project information interaction method according to claim 5, characterized in that: The method of testing the language model using an evaluation sample containing role identity prompt words to detect the matching degree of the language model to the role in the output content, and adjusting the parameters of the language model according to the matching degree includes: Constructing an evaluation sample, wherein each input sentence in the evaluation sample contains a clear role identity prompt word; Using the trained language model, input the evaluation sample for inference and generate response text; Perform role semantic matching analysis on the generated response text; The matching analysis results are converted into a role consistency score and compared with the preset role semantic baseline. If the score is lower than the threshold, corpus adjustment or incremental training is performed.

7. The whole-process engineering consulting project information interaction method according to claim 6 is characterized in that: When a user initiates an interaction request, the role-customized interaction model is called based on the user's role identity to generate and output response content, including: Receive user interaction requests and parse their identity information, determine the user's role category based on the identity information or historical behavior characteristics, and generate a role identification code; Input the user input request and role identification code into the role customization interaction model to generate the current user's response text; Perform role consistency review on the response text, and based on the role semantic review results, output the response text to the user terminal or perform response correction processing, and output the corrected response text to the user terminal.

8. The whole-process engineering consulting project information interaction method according to claim 7, characterized in that: The identification of user roles in engineering consulting projects includes: Receive a user access request and extract identity information associated with the user access request, including user account information, access terminal information, or historical behavior records; Analyzing the matching relationship between the identity information and the predefined role set based on the preset role identification rules; Users are divided into role categories corresponding to their identity information, and role labels representing the role categories are generated.

9. A full-process engineering consulting project information interaction system, used to implement the full-process engineering consulting project information interaction method according to any one of claims 1 to 8, characterized in that: include: Role identification module, corpus construction module, model generation module and content output module; The role identification module is used to identify user roles in engineering consulting projects; The corpus building module is used to collect historical data of the target role in the project and build a corresponding role corpus; The model generation module is used to perform parameter-oriented adjustment based on the general pre-trained language model using the role corpus to generate a role-customized interaction model; The content output module is used to call the role-customized interaction model based on the role identity of the user to generate and output response content when the user initiates an interaction request.

10. The full-process engineering consulting project information interaction system according to claim 9, characterized in that: The model generation module includes: a task guidance set construction unit, a weight distribution unit, and a training and orientation adjustment unit; The task guidance set construction unit is used to construct a role task guidance set based on the role corpus; The weight allocation unit is used to adapt the role and assign dynamic weights to training samples during the language model training process according to the semantic labels in the role task guidance set; The training and directional adjustment unit is used to assign dynamic weights to training samples and perform training, use the output results of the evaluation samples to detect the matching degree, and perform parameter directional adjustment on the language model according to the matching degree.

Images