Building engineering quality evaluation auxiliary system based on multi-modal large model

Abstract

The invention relates to the field of model analysis, in particular to a multi-modal large model-based building engineering quality evaluation auxiliary system, which is provided with a data acquisition module, a corpus analysis module, a variation recognition module, a corpus processing module and a large model module, the method comprises the following steps: dividing semantic variability of a user side, performing part-of-speech category labeling on historical sending information with strong semantic variability by a corpus analysis module, determining a corresponding variability tendency part-of-speech category, and processing the sending information based on the semantic variability of the user side by a corpus processing module, so that the sending information can be correctly understood by a large model. And the large model module receives the optimized sending information and outputs an output result of the large model. According to the method, a large model exclusively used for building engineering quality and safety evaluation is constructed, information in the field of building engineering is understood more accurately and quickly, and the retrieval accuracy and the message reply reliability are improved.

Description

Technical Field

[0001] The invention relates to the field of model analysis, and in particular to a construction engineering quality assessment auxiliary system based on a multi-modal large model. Background Art

[0002] Project quality issues run through the entire life cycle of project construction. Problems in any link will affect the quality and safety of the project, and may even lead to serious engineering accidents. In particular, due to the large differences in quality and safety management effects caused by non-structural and fragmented information, the informatization of project quality and safety assessment is urgent. With the rapid development of artificial intelligence semantic analysis large models (similar to ChatGPT), networked knowledge systems and update mechanisms will serve as the industry standard information artificial intelligence brain. Engineers only need to interact with the decision-making system, and the system will visually display the best solutions and search results to engineers.

[0003] Chinese patent publication number: CN118735732A, discloses an intelligent construction safety management system and method based on a large multimodal language model, the system includes: a data acquisition module, a scene recognition module, a safety behavior analysis module, an early warning broadcast module and a human-computer interaction module; the image data of the construction site is collected in real time through the mobile inspection device and the fixed monitoring device of the data acquisition module, and the data is uploaded to the edge computing node or the cloud server; the pre-trained large multimodal language model is introduced through the scene recognition module to perform semantic understanding and scene classification on the collected image data, and accurate recognition is achieved in combination with the scene information library; at the same time, an active learning mechanism is introduced to continuously optimize the model performance; the large multimodal language model is introduced to significantly improve the understanding and analysis capabilities of complex construction scenes, and realize intelligent identification and early warning of safety hazards; effectively lower the usage threshold and improve management efficiency.

[0004] However, in the prior art, there are still the following problems:

[0005] When interacting with the big model, there are cases where the semantics of the information sent are unclear, especially for special information, such as inverted information and typos. At this time, the information received by the big model will be wrong, and the understanding of the information will be biased, resulting in errors in the retrieval direction of the big model, inaccurate content, and incorrect or fabricated information, affecting the reliability of message replies. Summary of the invention

[0006] To this end, the present invention provides a construction project quality assessment auxiliary system based on a multimodal large model, which is used to solve the problem of unclear semantics of sent information when interacting with the large model, especially for special information, such as inverted information and typos. At this time, the information received by the large model will be erroneous, and the understanding of the information will be biased, resulting in errors in the retrieval direction of the large model, inaccurate content, and incorrect or fabricated information, affecting the reliability of message replies.

[0007] To achieve the above object, the present invention provides a construction engineering quality assessment auxiliary system based on a multi-modal large model, which comprises:

[0008] A data collection module is used to obtain some historical interaction data between the user terminal and the big model, including the historical sending information of the user terminal and the historical feedback information of the big model;

[0009] A corpus analysis module connected to the data acquisition module, for extracting interactive behavior features based on the historical interactive data, for determining semantic variation representation parameters of the user terminal based on the interactive behavior features, and for classifying the semantic variation of the user terminal;

[0010] An anomaly identification module, which is connected to the data collection module and the corpus analysis module, is used to mark the part-of-speech categories of the word segments after word segmentation processing of the historical sent information of the user terminal with strong semantic anomaly, and determine the corresponding part-of-speech category of the anomaly tendency according to the semantic collocation degree between the word segments in each historical sent information;

[0011] A corpus processing module is connected to the anomaly recognition module and is used to receive the current information sent by the user terminal and process the information based on the semantic anomaly of the user terminal, including:

[0012] Mark the part-of-speech category of each segment in the current sent information, generate several replacement segments to replace the segment corresponding to the part-of-speech category of the mutation tendency, optimize the sent information according to the replacement results, and record the optimized current sent information;

[0013] Or, record the current sending information;

[0014] The large model module is connected to the corpus processing module, and is used to receive the current sending information recorded by the corpus processing module and input it into the preconfigured large model, and output the output result of the large model.

[0015] Furthermore, the process of extracting interaction behavior features based on the historical interaction data by the corpus analysis module includes:

[0016] To determine the semantic collocation of each historically sent message and record the average semantic collocation;

[0017] To determine the repeated search behavior based on each historical sent information to determine the repeated search rate;

[0018] Among them, the repeated search behavior includes that the semantic matching degree between sentences of the information continuously sent by the user terminal is greater than a predetermined semantic matching degree threshold, and the repeated search rate is the proportion of the information corresponding to the repeated search behavior in the total amount of information sent.

[0019] Furthermore, the process of the corpus analysis module determining the semantic variation representation parameter of the user terminal based on the interactive behavior characteristics includes:

[0020] Determine the ratio of the semantic collocation threshold to the semantic collocation mean as the semantic collocation influencing factor;

[0021] Determine the ratio of the repeated search rate to the repeated search rate threshold as the repeated search impact factor;

[0022] The weighted sum of the semantic collocation influencing factor and the repeated search influencing factor is determined as the semantic variation representation parameter.

[0023] Furthermore, the corpus analysis module is used to classify the semantic variability of the user terminal, wherein:

[0024] If the variation representation parameter is greater than or equal to the reference variation representation parameter, the semantic variation is classified as strong semantic variation;

[0025] If the variance representation parameter is less than the reference variance representation parameter, the semantic variance is classified as weak semantic variance.

[0026] Furthermore, the process of the variation recognition module for marking the part-of-speech category of the segmented word includes:

[0027] Used to segment each historical sent information to obtain segmented segments;

[0028] Used to mark the part of speech category of each of the word segments.

[0029] Furthermore, the process of the variation identification module determining the corresponding variation tendency part-of-speech category according to the semantic collocation degree between the word segments in each historical sent information includes:

[0030] To determine the semantic collocation degree between the word segments in each historical sent message and the remaining word segments;

[0031] Based on the semantic collocation degree, the variant word segments in each historical sent information are identified, and the part-of-speech categories corresponding to the variant word segments are marked;

[0032] To count the occurrence probability of the marked part-of-speech categories in the historical sent information, and determine the part-of-speech categories with an occurrence probability greater than a predetermined probability threshold as the part-of-speech categories with a tendency to change;

[0033] If the semantic collocation degree between the word segments is less than a predetermined word segment semantic collocation degree threshold, each of the word segments is determined to be a variant word segment.

[0034] Furthermore, the corpus processing module is used to process the sent information based on the semantic variability of the user end, wherein:

[0035] If the semantic variability is strong semantic variability, the part-of-speech category of each word segment in the current sent information is marked, a number of replacement segments are generated to replace the word segments corresponding to the part-of-speech category of the variability tendency, the sent information is optimized according to the replacement results, and the optimized current sent information is recorded;

[0036] If the semantic variability is weak semantic variability, the currently sent information is recorded.

[0037] Furthermore, the corpus processing module is used to determine the word segment corresponding to the part-of-speech category of the variation tendency, including:

[0038] To determine whether the part-of-speech category of each word segment is a part-of-speech category with a tendency to change;

[0039] If the part-of-speech category of the segmented word is a part-of-speech category with a tendency to change, it is determined that the segmented word needs to be replaced.

[0040] Furthermore, the process of the corpus processing module generating a plurality of replacement segments to replace the segmented segments corresponding to the part-of-speech category of the variation tendency includes:

[0041] Determine a number of synonyms based on the word segment and use the synonyms as replacement segments;

[0042] Determining a number of strongly associated words based on the word segment, and using the strongly associated words as replacement segments;

[0043] Among them, the strongly associated words must satisfy the requirement that the semantic collocation degree between the words and the word segments is greater than a predetermined strong association threshold.

[0044] Furthermore, the process of optimizing the sent information according to the replacement result by the corpus processing module includes:

[0045] Used to replace the word segment and generate several sending messages;

[0046] To determine the semantic matching degree of the information sent after each replacement;

[0047] The replaced sent information corresponding to the highest semantic matching degree is determined as the optimized current sent information.

[0048] Compared with the prior art, the present invention is provided with a data acquisition module, a corpus analysis module, an anomaly identification module, a corpus processing module and a large model module. The anomaly identification module analyzes the historical transmission information and divides the semantic anomaly of the user terminal. The corpus analysis module performs part-of-speech category marking on the historical transmission information with strong semantic anomaly and determines the corresponding part-of-speech category of the anomaly tendency. The corpus processing module processes the transmission information based on the semantic anomaly of the user terminal so that the transmission information can be correctly understood by the large model. The large model module receives the optimized transmission information and outputs the output result of the large model. The present invention optimizes the abnormal information for the interaction problem in the field of construction engineering, so that the large model can understand the information more accurately and quickly, and improves the accuracy of retrieval and the reliability of message reply.

[0049] In particular, the present invention calculates semantic variation characterization parameters to divide semantic variability. When the model performs semantic analysis on interactive content in the field of construction engineering, due to differences in individual semantic behavior habits, the information sent by the user end is likely to have abnormal text habits, such as inverted sentences, missing characters, wrong characters, etc. At this time, the model is likely to be unable to accurately identify the sent information. Based on this, the present invention integrates the historical sent information of clients with abnormal text habits, filters out their abnormal sent information, and divides the semantic variability of the historical sent information, which is convenient for the subsequent division of the semantic variability of the user end, and subsequently adaptively processes the sent information sent by different user ends, so that the large model can understand the text more accurately and quickly, and improves the accuracy of retrieval and the reliability of message replies.

[0050] In particular, the present invention determines the part-of-speech categories with a tendency to mutate of the word segments of historical sent information, and accurately searches for sent information with strong semantic variability. In actual situations, there are often certain regularities in the semantic variability of the user end. For example, the wording habits are different from others, and the dialect-specific habits all have certain regularities. Therefore, the present invention considers finding the part-of-speech categories with a tendency to mutate, and then characterizing the parts of the wording habits that are prone to mutation, so as to facilitate subsequent targeted optimization of the sent information, so that the large model can understand the text more accurately and quickly, and improve the accuracy of retrieval and the reliability of message replies.

[0051] In particular, the present invention performs dual-line processing on currently sent information with different semantic variability. For currently sent information with weak semantic variability, it is directly recorded for retrieval and reply. For categories with strong semantic variability, in actual situations, if the currently sent information is still directly retrieved and replied, it will cause inaccurate replies given by the large model. Based on this, the present invention performs part-of-speech category judgment on the currently sent information sent by the user end with strong semantic variability to optimize the part-of-speech category with variability tendency in a targeted manner, so that the large model can understand the text more accurately and quickly, and improve the accuracy of retrieval and the reliability of message replies. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A schematic diagram of the structure of a construction engineering quality assessment auxiliary system based on a multi-modal large model according to an embodiment of the invention;

[0053] Figure 2 A logical block diagram of dividing the semantic variability of the user terminal according to an embodiment of the invention;

[0054] Figure 3 The logical block diagram for determining the corresponding part-of-speech category of the variation tendency according to an embodiment of the invention. DETAILED DESCRIPTION

[0055] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0056] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the protection scope of the present invention.

[0057] It should be noted that in the description of the present invention, unless otherwise clearly specified and limited, the term "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0058] See also Figure 1-Figure 3 As shown, Figure 1 The schematic diagram of the structure of the multi-modal large-scale construction project quality assessment auxiliary system of the embodiment of the invention is as follows: Figure 2 A logical block diagram of dividing the semantic variability of the user terminal according to an embodiment of the invention, Figure 3 The logic block diagram of determining the corresponding abnormal tendency part-of-speech category of the embodiment of the invention is as follows: The multi-modal large model construction project quality assessment auxiliary system of the embodiment of the invention includes:

[0059] A data collection module is used to obtain some historical interaction data between the user terminal and the big model, including the historical sending information of the user terminal and the historical feedback information of the big model;

[0060] It is understandable that the acquisition of interactive data requires the authorization of the user end.

[0061] A corpus analysis module connected to the data acquisition module, for extracting interactive behavior features based on the historical interactive data, for determining semantic variation representation parameters of the user terminal based on the interactive behavior features, and for classifying the semantic variation of the user terminal;

[0062] An anomaly identification module, which is connected to the data collection module and the corpus analysis module, is used to mark the part-of-speech categories of the word segments after word segmentation processing of the historical sent information of the user terminal with strong semantic anomaly, and determine the corresponding part-of-speech category of the anomaly tendency according to the semantic collocation degree between the word segments in each historical sent information;

[0063] A corpus processing module is connected to the anomaly recognition module and is used to receive the current information sent by the user terminal and process the information based on the semantic anomaly of the user terminal, including:

[0064] Mark the part-of-speech category of each segment in the current sent information, generate several replacement segments to replace the segment corresponding to the part-of-speech category of the mutation tendency, optimize the sent information according to the replacement results, and record the optimized current sent information;

[0065] Or, record the current sending information;

[0066] The large model module is connected to the corpus processing module, and is used to receive the current sending information recorded by the corpus processing module and input it into the preconfigured large model, and output the output result of the large model.

[0067] Specifically, for the pre-configured large model, its training data needs to be pre-processed, which includes the following steps:

[0068] S01, data collection: collect standard documents on construction and annotate the data. It can be understood that standard documents on construction include national relevant engineering quality standard documents and company-owned standard documents. It can be understood that national relevant engineering quality standard documents and company-owned standard documents contain engineering requirement index data, component material information, quality inspection standards, construction standards, accident handling standards and other multi-dimensional information, thereby providing comprehensive data support in the field of construction engineering.

[0069] S02, data cleaning, cleaning and archiving the collected data. It can be understood that data cleaning is to eliminate erroneous data and invalid data. There is no limitation on the specific method of archiving. Those skilled in the art can divide it according to the needs for easy classification and storage. For example, for the construction industry, text descriptions of quality standards are classified into one category, and text descriptions of error examples are classified into another category. Of course, other methods may also be included. Those skilled in the art can choose according to the needs, which will not be elaborated here.

[0070] S03, data storage, can store the archived data to build a corpus, which will not be described in detail.

[0071] Specifically, the pre-configured large model is applied to the field of construction engineering and can answer relevant professional questions in the field of construction engineering. There is no specific limitation on its training process. It can be understood that the existing natural language model can be trained through a pre-built corpus dedicated to the construction industry. Of course, there is no specific limitation on the architecture of the natural language model, and technical personnel in this field can choose it according to their needs.

[0072] Specifically, in some possible implementations, the large model adopts a multi-model collaboration approach. It is understandable that different models have their own advantages and disadvantages in different tasks. Therefore, a multi-model collaboration approach is preferred. For example, a management engine is used to manage and train different models.

[0073] In practical applications, when a multi-model collaboration approach is adopted, various pre-archived corpora can be used to train models suitable for different tasks, for example, models for answering questions related to building quality and models for answering questions related to error examples. It is worth noting that the above examples are only illustrative and are not intended to limit the scope of protection. Those skilled in the art can train models for different tasks according to their needs, which will not be elaborated here.

[0074] Specifically, for the multi-model collaboration mode, a management engine is used for routing selection. Routing selection is to dynamically assign tasks to the most appropriate model by analyzing the input features, ensuring that each model only processes the content it is best at. There is no limitation on the routing selection method. For example, the data features of the currently sent information can be determined by a data feature classifier, and the most appropriate model can be matched based on the data features. Of course, the data feature classifier also needs to be trained to ensure that it can accurately distinguish the currently sent information and transmit it to different models for output. This will not be repeated.

[0075] Specifically, in some possible implementations, when multiple models collaborate, an image processing model can be introduced to analyze on-site images and provide risk warnings. For example, an image processing model that can identify workers' illegal actions can be pre-trained, and there is no specific limitation on the structure of the image processing model. For example, an image processing model with a neural network architecture can be used to implement the corresponding functions through pre-training. By combining image and text multimodal information, an AI large model that can process multidimensional data can be constructed, thereby enabling accurate responses to text information, data analysis of the actual situation at the construction site, and providing risk warnings and other functions, thereby improving the applicability of large models in the field of construction engineering.

[0076] Specifically, the process of extracting interactive behavior features based on the historical interactive data by the corpus analysis module includes:

[0077] To determine the semantic collocation of each historically sent message and record the average semantic collocation;

[0078] To determine the repeated search behavior based on each historical sent information to determine the repeated search rate;

[0079] Among them, the repeated search behavior includes that the semantic matching degree between sentences of the information continuously sent by the user terminal is greater than a predetermined semantic matching degree threshold, and the repeated search rate is the proportion of the information corresponding to the repeated search behavior in the total amount of information sent.

[0080] It is understandable that in actual situations, the information sent is mostly a single sentence or a single paragraph of information. Therefore, the semantic collocation is for a single message sent, and the specific calculation method of the semantic collocation is not limited. For example, a pre-trained word vector model (such as Word2Vec, G loVe, FastText) can be used to obtain the vector representation of the keywords in a single message sent, and then calculate the cosine similarity between them. The mean cosine similarity between the keywords contained in the sentence can be determined as the semantic collocation, which will not be repeated here.

[0081] It is understandable that the semantic collocation between sentences between messages continuously sent by the user end refers to the messages sent each time, and there is no limitation on the specific method of calculating the semantic collocation between sentences. For example, a pre-trained language model such as a BERT model can be used. The BERT model captures the context information in the text through the encoder structure, generates a vector representation of the sentence, and then calculates the cosine similarity between sentences to represent the semantic collocation between sentences. Of course, other methods can also be used, which will not be repeated here.

[0082] Specifically, it is understandable that in actual situations, if the user's usage habits are rather special, there may be a situation where the input information cannot be effectively answered, and the question may be simply changed and input multiple times. Therefore, if cosine similarity is used to represent the semantic matching between sentences, in order to represent the situation where the semantics are relatively close, the semantic matching threshold between sentences is selected in the range of [0.75, 0.85].

[0083] Specifically, the process of the corpus analysis module determining the semantic variation representation parameter of the user terminal based on the interactive behavior characteristics includes:

[0084] Determine the ratio of the semantic collocation threshold to the semantic collocation mean as the semantic collocation influencing factor;

[0085] Determine the ratio of the repeated search rate to the repeated search rate threshold as the repeated search impact factor;

[0086] The weighted sum of the semantic collocation influencing factor and the repeated search influencing factor is determined as the semantic variation representation parameter.

[0087] Specifically, the weight coefficient of the semantic collocation impact factor is 0.6, and the weight coefficient of the repeated search impact factor is 0.4.

[0088] Specifically, the semantic matching degree threshold and the repeated search rate threshold are both obtained by cosine setting, wherein historical interaction data of several user terminals are obtained in advance, the semantic matching degree means and repeated search rates of the user terminals are recorded, the average value of the semantic matching degree means corresponding to each user terminal is solved, the semantic matching degree threshold is set to the product of the average value and a first precision coefficient, the repeated search rate mean is solved, and the repeated search rate threshold is set to the product of the repeated search rate mean and a second precision coefficient, the first precision coefficient is selected in the interval [0.75, 0.85], and the second precision coefficient is selected in the interval (0.85, 0.95].

[0089] Specifically, the present invention calculates semantic variation characterization parameters to divide semantic variability. When the model performs semantic analysis on interactive content in the field of construction engineering, due to differences in individual semantic behavior habits, the information sent by the user terminal is likely to have abnormal text habits, such as inverted sentences, missing words, wrong words, etc. At this time, the model is likely to be unable to accurately identify the sent information. Based on this, the present invention integrates the historical sending information of the client with abnormal text habits, filters out its abnormal sending information, and divides the semantic variability of the historical sending information, which is convenient for the subsequent division of the semantic variability of the user terminal. The sending information sent by different user terminals is adaptively processed subsequently, so that the large model can understand the text more accurately and quickly, and the accuracy of retrieval and the reliability of message replies are improved.

[0090] Specifically, the corpus analysis module is used to classify the semantic variability of the user terminal, wherein:

[0091] If the variation representation parameter is greater than or equal to the reference variation representation parameter, the semantic variation is classified as strong semantic variation;

[0092] If the variance representation parameter is less than the reference variance representation parameter, the semantic variance is classified as weak semantic variance.

[0093] Specifically, the benchmark anomaly characterization parameter is pre-calculated, a number of anomaly characterization parameters of information sent by the user end are obtained, the average value of the anomaly characterization parameters is calculated, and 1.24 times the average value of the anomaly characterization parameters is determined as the benchmark anomaly characterization parameter.

[0094] Specifically, the process of annotating the part-of-speech categories of the word segments by the variation recognition module includes:

[0095] Used to segment each historical sent information to obtain segmented segments;

[0096] Used to mark the part of speech category of each of the word segments.

[0097] Specifically, there is no limitation on the word segmentation method, and a word segmentation tool can be used to segment words to obtain word segments, which can be keywords. There is no limitation on the tool for marking part-of-speech categories, and technical personnel in this field can use any part-of-speech category marking tool in the prior art for marking, which will not be repeated here.

[0098] Specifically, the process of the variation recognition module determining the corresponding variation tendency part-of-speech category according to the semantic collocation degree between the word segments in each historical sent information includes:

[0099] To determine the semantic collocation degree between the word segments in each historical sent message and the remaining word segments;

[0100] Based on the semantic collocation degree, the variant word segments in each historical sent information are identified, and the part-of-speech categories corresponding to the variant word segments are marked;

[0101] To count the occurrence probability of the marked part-of-speech categories in the historical sent information, and determine the part-of-speech categories with an occurrence probability greater than a predetermined probability threshold as the part-of-speech categories with a tendency to change;

[0102] If the semantic collocation degree between the word segments is less than a predetermined word segment semantic collocation degree threshold, each of the word segments is determined to be a variant word segment.

[0103] Specifically, it can be understood that for the semantic similarity between two segmented words, the vector representations of the segmented words can be obtained, and then the cosine similarity between them can be calculated, and the cosine similarity can be determined as the semantic similarity between the segmented words.

[0104] Specifically, the word segmentation semantic collocation threshold is pre-set, wherein the semantic collocation between each word segment in a number of single-time sent information is pre-acquired, the mean of the semantic collocation is solved, and 0.65 to 0.85 times the mean of the semantic collocation is set as the word segmentation semantic collocation threshold.

[0105] Specifically, the purpose of setting a predetermined probability threshold is to identify parts of speech with higher probabilities. Therefore, in order to ensure data representation and exclude occasional situations, the probability threshold needs to be set to be greater than 10%.

[0106] Specifically, the present invention determines the part-of-speech categories with a tendency to mutate of the word segments of historical sent information, and accurately searches for sent information with strong semantic variability. In actual situations, there are often certain regularities in the semantic variability of the user end. For example, the wording habits are different from others, and the dialect-specific habits all have certain regularities. Therefore, the present invention considers finding the part-of-speech categories with a tendency to mutate, and then characterizing the parts of the wording habits that are prone to mutation, so as to facilitate subsequent targeted optimization of the sent information, so that the large model can understand the text more accurately and quickly, and improve the accuracy of retrieval and the reliability of message replies.

[0107] Specifically, the corpus processing module is used to process the sent information based on the semantic variability of the user end, wherein:

[0108] If the semantic variability is strong semantic variability, the part-of-speech category of each word segment in the current sent information is marked, a number of replacement segments are generated to replace the word segments corresponding to the part-of-speech category of the variability tendency, the sent information is optimized according to the replacement results, and the optimized current sent information is recorded;

[0109] If the semantic variability is weak semantic variability, the currently sent information is recorded.

[0110] Specifically, the present invention performs dual-line processing on the currently sent information with different semantic variability. For the currently sent information with weak semantic variability, it is directly recorded for retrieval and reply. For the categories with strong semantic variability, in actual situations, if the currently sent information is still directly retrieved and replied, it will cause the reply given by the large model to be inaccurate. Based on this, the present invention performs part-of-speech category judgment on the currently sent information with strong semantic variability to optimize it, determines the part-of-speech category with strong variability tendency in the currently sent information, and replaces it with synonyms or strongly associated words, so that the large model can understand the text more accurately and quickly, and improves the accuracy of retrieval and the reliability of message replies.

[0111] Specifically, the corpus processing module is used to determine the word segments corresponding to the part-of-speech categories of the variation tendency, including:

[0112] To determine whether the part-of-speech category of each word segment is a part-of-speech category with a tendency to change;

[0113] If the part-of-speech category of the segmented word is a part-of-speech category with a tendency to change, it is determined that the segmented word needs to be replaced.

[0114] Specifically, the process of the corpus processing module generating a plurality of replacement segments to replace the segmented segments corresponding to the part-of-speech categories of the variation tendency includes:

[0115] Determine a number of synonyms based on the word segment and use the synonyms as replacement segments;

[0116] Determining a number of strongly associated words based on the word segment, and using the strongly associated words as replacement segments;

[0117] Among them, the strongly associated words must satisfy the requirement that the semantic collocation degree between the words and the word segments is greater than a predetermined strong association threshold.

[0118] Specifically, the strong association threshold is pre-set, wherein the semantic collocation degree between each word segment in a number of single-time sent information is obtained in advance, the mean semantic collocation degree is solved, and 1.45 to 1.85 times the mean semantic collocation degree is set as the strong association threshold. Of course, the strong association threshold has an upper limit, and the upper limit value can be set to 0.95. If the calculated strong association threshold is the upper limit value, the strong association threshold is set to the upper limit value.

[0119] Specifically, the process of optimizing the sent information according to the replacement result by the corpus processing module includes:

[0120] Used to replace the word segment and generate several sending messages;

[0121] To determine the semantic matching degree of the information sent after each replacement;

[0122] The replaced sent information corresponding to the highest semantic matching degree is determined as the optimized current sent information.

[0123] Specifically, for a sent message, it may have several alternative words. By replacing several alternative words into the sent message respectively, there will be several replaced sent messages. The sent message with the highest semantic matching degree is determined as the optimized sent message, which can enable the large model module to better complete the auxiliary work of construction project quality assessment.

[0124] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

Claims

1. A construction engineering quality assessment auxiliary system based on a multi-modal large model, characterized in that: include: A data collection module is used to obtain some historical interaction data between the user terminal and the big model, including the historical sending information of the user terminal and the historical feedback information of the big model; A corpus analysis module connected to the data acquisition module, for extracting interactive behavior features based on the historical interactive data, for determining semantic variation representation parameters of the user terminal based on the interactive behavior features, and for classifying the semantic variation of the user terminal; An anomaly identification module, which is connected to the data collection module and the corpus analysis module, is used to mark the part-of-speech categories of the word segments after word segmentation processing of the historical sent information of the user terminal with strong semantic anomaly, and determine the corresponding part-of-speech category of the anomaly tendency according to the semantic collocation degree between the word segments in each historical sent information; A corpus processing module is connected to the anomaly recognition module and is used to receive the current information sent by the user terminal and process the information based on the semantic anomaly of the user terminal, including: Mark the part-of-speech category of each segment in the current sent information, generate several replacement segments to replace the segment corresponding to the part-of-speech category of the mutation tendency, optimize the sent information according to the replacement results, and record the optimized current sent information; Or, record the current sending information; The large model module is connected to the corpus processing module, and is used to receive the current sending information recorded by the corpus processing module and input it into the preconfigured large model, and output the output result of the large model.

2. The construction engineering quality assessment auxiliary system based on multimodal large model according to claim 1 is characterized in that: The process of extracting interaction behavior features based on the historical interaction data by the corpus analysis module includes: To determine the semantic collocation of each historically sent message and record the average semantic collocation; To determine the repeated search behavior based on each historical sent information to determine the repeated search rate; Among them, the repeated search behavior includes that the semantic matching degree between sentences of the information continuously sent by the user terminal is greater than a predetermined semantic matching degree threshold, and the repeated search rate is the proportion of the information corresponding to the repeated search behavior in the total amount of information sent.

3. The construction engineering quality assessment auxiliary system based on multimodal large model according to claim 1 is characterized in that: The process of the corpus analysis module determining the semantic variation representation parameter of the user terminal based on the interactive behavior characteristics includes: Determine the ratio of the semantic collocation threshold to the semantic collocation mean as the semantic collocation influencing factor; Determine the ratio of the repeated search rate to the repeated search rate threshold as the repeated search impact factor; The weighted sum of the semantic collocation influencing factor and the repeated search influencing factor is determined as the semantic variation representation parameter.

4. The construction engineering quality assessment auxiliary system based on multimodal large model according to claim 1 is characterized in that: The corpus analysis module is used to classify the semantic variability of the user terminal, wherein: If the variation representation parameter is greater than or equal to the reference variation representation parameter, the semantic variation is classified as strong semantic variation; If the variance representation parameter is less than the reference variance representation parameter, the semantic variance is classified as weak semantic variance.

5. The construction engineering quality assessment auxiliary system based on multimodal large model according to claim 1 is characterized in that: The process of the variation recognition module for marking the part-of-speech category of the segmented word segment includes: Used to segment each historical sent information to obtain segmented segments; Used to mark the part of speech category of each of the word segments.

6. The construction engineering quality assessment auxiliary system based on multimodal large model according to claim 1 is characterized in that: The process of the variation identification module determining the corresponding variation tendency part-of-speech category according to the semantic collocation degree between the word segments in each historical sent information includes: To determine the semantic collocation degree between the word segments in each historical sent message and the remaining word segments; for identifying variant word segments in each historical sent information based on the semantic collocation degree, and marking the part of speech category corresponding to the variant word segments; To count the occurrence probability of the marked part-of-speech categories in the historical sent information, and determine the part-of-speech categories with an occurrence probability greater than a predetermined probability threshold as the part-of-speech categories with a tendency to change; If the semantic collocation degree between the word segments is less than a predetermined word segment semantic collocation degree threshold, each of the word segments is determined to be a variant word segment.

7. The construction engineering quality assessment auxiliary system based on multimodal large model according to claim 1 is characterized in that: The corpus processing module is used to process the sent information based on the semantic variability of the user terminal, wherein: If the semantic variability is strong semantic variability, the part-of-speech category of each word segment in the current sent information is marked, a number of replacement segments are generated to replace the word segments corresponding to the part-of-speech category of the variability tendency, the sent information is optimized according to the replacement results, and the optimized current sent information is recorded; If the semantic variability is weak semantic variability, the currently sent information is recorded.

8. The construction engineering quality assessment auxiliary system based on multimodal large model according to claim 1 is characterized in that: The corpus processing module is used to determine the word segments corresponding to the part-of-speech categories of the variation tendency, including: To determine whether the part-of-speech category of each word segment is a part-of-speech category with a tendency to change; If the part-of-speech category of the segmented word is a part-of-speech category with a tendency to change, it is determined that the segmented word needs to be replaced.

9. The construction engineering quality assessment auxiliary system based on multimodal large model according to claim 1 is characterized in that: The process of generating a plurality of replacement segments to replace the segmented segments corresponding to the part-of-speech categories of the variation tendency by the corpus processing module includes: Determine a plurality of synonyms based on the word segment and use the synonyms as replacement segments; Determining a number of strongly associated words based on the word segment, and using the strongly associated words as replacement segments; Among them, the strongly associated words must satisfy the requirement that the semantic collocation degree between the words and the word segments is greater than a predetermined strong association threshold.

10. The construction engineering quality assessment auxiliary system based on multi-modal large model according to claim 1 is characterized in that: The process of the corpus processing module optimizing the sent information according to the replacement result includes: Used to replace the word segment and generate several sending messages; To determine the semantic matching degree of the information sent after each replacement; The replaced sent information corresponding to the highest semantic matching degree is determined as the optimized current sent information.

Images