Data collaboration and aid decision-making system and method based on deep learning

Abstract

The invention discloses a deep learning-based data collaboration and auxiliary decision-making system and method, and relates to the technical field of computer data processing. Firstly, keywords and word frequencies, grammar structure type distribution and conflict event processing behavior records are extracted from a network collaboration office scene; the method comprises the following steps of: constructing three types of feature vectors of vocabularies, sentence pattern structures and conflict judgment, splicing into an original high-dimensional feature vector, obtaining a user behavior fingerprint vector through dimension reduction mapping, extracting key semantic vocabularies, combining with a timestamp and a user identity coding vector to form a document operation vector, obtaining a context relationship vector through time sequence feature extraction and weighted summation, and obtaining a user behavior fingerprint vector. Calculating a semantic deviation degree between the normalized dimension space and the context relation vector in the normalized dimension space; and finally, extracting three types of feature vectors corresponding to the current input, performing splicing and dimension reduction to obtain a real-time behavior representation vector, calculating the difference degree between the real-time behavior representation vector and a fingerprint vector, performing judgment through a comprehensive risk assessment function, and if a result is greater than a threshold value, triggering an early warning prompt.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer data processing, in particular to a data collaboration and auxiliary decision-making system and method based on deep learning. BACKGROUND

[0002] With the continuous evolution of computer and Internet technology, network-based multi-person collaborative office systems have been widely used in modern enterprises and organizations, and have become a key means to improve work efficiency and information synchronization capability. Especially in high-concurrency, multi-role collaborative scenarios, user input content not only needs to meet individual expression intent, but also needs to maintain logical compatibility with the overall document semantics, which puts higher requirements on the system's conflict perception, semantic understanding and risk prediction capabilities. The existing technology usually uses the time sequence features of input content or the preset weight of the operator to solve conflicts. Such a way can only solve the problem of conflict, but cannot eliminate the adverse effects of incorrect content on the context.

[0003] With the intelligent and dynamic evolution of collaborative office scenarios, the inherent limitations of the above traditional conflict processing paradigm have become increasingly prominent. For example, it can only be arbitrated after the conflict occurs, and cannot pre-empt incorrect input behavior, which may still be misread, quoted or used as a basis for subsequent reasoning by other collaborators during its short existence, thereby implanting noise that is difficult to trace in the semantic flow of the document. Therefore, it is necessary to build a collaborative auxiliary decision-making mechanism with semantic risk warning and behavior anomaly detection capabilities, which effectively blocks the erosion of incorrect information on the collaborative knowledge body without sacrificing the smoothness of collaboration, which has become a key technical problem currently faced by those skilled in the art. SUMMARY

[0004] The purpose of the present application is to provide a data collaboration and auxiliary decision-making system and method based on deep learning to solve the problems in the prior art.

[0005] To achieve the above purpose, the present application provides the following technical scheme: a data collaboration and auxiliary decision-making method based on deep learning: Step 1: Identify the behavior characteristics of the target user and establish the user's behavior fingerprint vector. Take the target user's complete historical operation log in the network collaborative office scenario as input data, extract the keywords and their corresponding word frequency information, syntax structure type distribution and user's handling behavior records for conflict events, construct a vocabulary feature vector, a sentence structure feature vector and a conflict judgment vector respectively, and sequentially concatenate to form an original high-dimensional feature vector. After dimensionality reduction mapping, the fingerprint vector is obtained. Step 2: Obtain the text input content and corresponding timestamp information of all participating users within a unit time interval, extract the key semantic words in the text content, convert the feature words into word vector representation using the word embedding model, and combine the timestamp and user identity code vector of this input into a document operation vector. Perform time series feature extraction on the continuous document operation vector, and calculate the context relationship vector using weighted summation method; Step 3: Evaluate the semantic deviation of the target user's current input content relative to the global context, collect the target user's new input text content within a unit time, and convert it into a new document operation vector. In the normalized dimension space, calculate the difference between the new document operation vector and the context relationship vector; Step 4: Extract the word feature vector, sentence structure feature vector, and conflict judgment vector corresponding to the target user's current input content. After concatenating these three vectors, perform dimension reduction to obtain the real-time behavior representation vector in the current input state. Calculate the difference between the real-time behavior representation vector and the fingerprint vector. By constructing a comprehensive risk assessment function, when the value obtained by the comprehensive risk assessment function after a certain input is greater than the threshold value, trigger an early warning prompt.

[0006] Further, step 1 includes: Step 1-1: Take a user in the remote collaborative office scenario as the target user, collect the target user's complete historical operation log in the network collaborative office scenario, collect the key words and word frequency corresponding to the key words in the operation log, establish the word feature vector, collect the syntax structure features of the target user's input content record, and establish the sentence structure feature vector; Step 1-2: Collect the target user's handling records when there are conflict operations in the network collaborative office scenario, classify the target user's handling records into three labels: acceptance, non-acceptance, and abandonment, collect the proportion of each label, and establish the conflict judgment vector; Step 1-3: Concatenate the word feature vector, sentence structure feature vector, and conflict judgment vector to obtain the target user's original feature vector, perform dimension reduction through a fully connected neural network layer, and output the target user's behavior fingerprint vector.

[0007] First step: Construct the behavior fingerprint vector of the target user; Let the target user be , where is the set of all users participating in the collaborative office system. The system continuously collects the target user 's complete historical operation log in the collaborative platform, which includes but is not limited to text input records, editing operation types, conflict handling behaviors, and corresponding timestamp information.

[0008] 1 From the target user Lexical features are extracted from all historical text input records. Specifically, after word segmentation of all input text, user-defined features are formed. Exclusive vocabulary list Count the j-th keyword Frequency of occurrence After normalizing the frequency vector, the vocabulary feature vector lv is obtained. u Its j-th dimension element is defined as: ; By segmenting and tagging user input, features are extracted from sentence components, such as subject-verb agreement, passive voice warnings, conditional statements, and transitional statements. One-hot encoding is used to encode words or phrases to obtain feature codes for different sentence structures. Alternatively, open-source tools such as NLTK are used to analyze sentence structure and aggregate the probability distributions of different sentence structures to obtain sentence structure feature vectors (rv). u ; During collaborative editing, when content conflicts occur, the system records the user's handling of the conflict. Suppose the target user encounters N conflicts in their history, and their handling results are divided into three categories: acceptance, rejection, and abandonment. The corresponding frequency of each of the three handling results is counted and denoted as na, nb, and nc, respectively, along with the proportion, generating a conflict judgment vector cv. u , ; The three vectors above are concatenated to obtain the original vector x. u The fingerprint vector f of the target user u is obtained by performing dimensionality reduction mapping on the original vector through a fully connected layer. u , , where ReLU represents the ReLU activation function, W represents the weight matrix, and b represents the bias vector.

[0009] Furthermore, step 2 includes: Step 2-1: In the operation record of the current shared document, obtain the text content entered by all users within a unit of time, the timestamp of each input, obtain the identity information of all users who participated in the operation of the current shared document within a unit of time, and set the identity identifier for each user; Step 2-2: Obtain the text content input by any user each time, extract the feature words in the text content, convert the feature words into word vectors, concatenate the word vectors, timestamps and identity identifiers in a single input process to obtain document operation vectors, sort the document operation vectors according to the time sequence of input to obtain the document operation sequence; Steps 2-3: Input the document operation sequence into the bidirectional long short-term memory network encoder to obtain the hidden state sequence, where the i-th hidden state in the hidden state sequence is denoted as hi, and the attention weight αi corresponding to the i-th hidden state is calculated. ,in, W represents the weight matrix, and Q represents the total number of hidden states in the hidden state sequence. Steps 2-4: Calculate the context relation vector E. .

[0010] Assume the current collaborative operation occurs within a time window. ,in The preset time window length is used to record all user operations on the shared document within this window.

[0011] Get Time Window The text content s of the i-th user input i Corresponding timestamp t i and user identification u i ; For each input text s i The feature words are extracted and mapped to word vectors using a pre-trained word embedding model, such as Word2Vec or GloVe. The semantic representation m of the input is obtained by averaging all word vectors of the i-th input. i : Normalized semantic representation and normalized timestamp and the embedded vector ID of the user's identity. i The vectors are concatenated to form the document operation vector yi for the i-th operation. , Arrange all document operation vectors in chronological order to form a document operation sequence. Q represents the total number of document operations within the time window.

[0012] The sequence Y is input into a bidirectional long short-term memory (BiLSTM) encoder. This encoder consists of a forward LSTM and a backward LSTM, which output the hidden state h, respectively. f,i and h b,i After splicing, we obtain the complete hidden state of step i: h i =[h f,i :h b,i ]; This yields the hidden state sequence. .

[0013] An attention mechanism is introduced to aggregate key contextual information, obtain a learnable query vector q, and compute the i-th hidden state h. i Attention score: , Then calculate the normalized attention weights. : , Finally, the context relation vector E: .

[0014] Furthermore, step 3 includes: Step 3-1: Collect the text content input by the target user within a unit of time. Using the same word embedding model as in Step 2, obtain the document operation vector of the text content. Normalize the document operation vector and the context relation vector respectively, and then map them to the same dimension. Step 3-2: Calculate the difference D between the document operation vector and the context relation vector in the same dimension.

[0015] Get target user u in the current time window The latest input text is processed using the same word embedding model and average pooling strategy as in the second step to obtain its semantic vector sn. To eliminate the influence of vector scaling, this vector is L2 normalized to obtain a unit vector. ; Simultaneously, the context semantic memory vector E generated in step 1 needs to be mapped to the same dimensional space as pn for direct comparison. This mapping can be accomplished using a linear projection matrix P, and the projected vector is also L2 normalized to obtain... .

[0016] Within this unified semantic space, the system calculates the semantic deviation D between the new input and the context, which is defined as the complement of their cosine similarity, i.e.: .

[0017] Furthermore, step 4 includes: Step 4-1: Based on the text content newly input by the target user within a unit of time, extract the lexical feature vector, sentence structure feature vector, and conflict judgment vector corresponding to the text content. After concatenating these three vectors, perform dimensionality reduction to obtain the real-time behavior representation vector under the current input state. Calculate the difference between the real-time behavior representation vector and the fingerprint vector. By constructing a comprehensive risk assessment function, when the value obtained by the comprehensive risk assessment function after a certain input is completed is greater than the threshold, trigger an early warning prompt. Step 4-2: Concatenate the vocabulary feature vector, sentence structure feature vector, and conflict judgment vector corresponding to each unit of time, and perform dimensionality reduction through the same fully connected neural network layer as in Step 1-3 to obtain the current input state vector of the current target user. Compare the difference between the current input state vector and the behavior fingerprint vector and record it as P. Step 4-3: Establish a comprehensive evaluation function R, R=α×D+β×P, α+β=1. When R is greater than the threshold after a certain text input by the target user, a reminder is triggered and the submission of the result of that text input is paused.

[0018] For the target user u's latest input s within the current time window i Repeat the feature extraction process from step one to obtain the lexical feature vector, sentence structure feature vector, and conflict judgment vector for the current window. Meanwhile, if the conflict judgment vector cannot be statistically analyzed within a single window due to the absence of conflict events, a historical moving average or a default vector is used to fill the vector, ensuring that the vector dimension remains unchanged.

[0019] Concatenate the three current feature vectors mentioned above to form x. u,curr Then, input the same fully connected neural network layer as in the first step, that is, share the weight matrix W and bias vector b, to obtain the current input state vector f. u,curr : , Calculate the cosine distance between the current input state vector and the historical behavior fingerprint vector as the behavior risk deviation. : ; The larger the value, the more the current behavior deviates from the user's historical pattern; Finally, a comprehensive evaluation function is established. : .

[0020] To better implement the above methods, a data collaboration and decision support system based on deep learning is also proposed. The system includes: a user behavior modeling module, a context memory network management module, a semantic risk assessment module, and an assessment and decision module. The user behavior modeling module is used to identify the behavioral characteristics of target users and establish user behavior fingerprint vectors. It takes all historical operation logs of target users in online collaborative office scenarios as input data, extracts keywords and their corresponding word frequency information, grammatical structure type distribution and user's handling behavior records for conflict events, and constructs lexical feature vectors, sentence structure feature vectors and conflict judgment vectors respectively. These are then concatenated to form the original high-dimensional feature vector. After performing dimensionality reduction mapping, the fingerprint vector is obtained. The context memory network management module is used to obtain the text input content and corresponding timestamp information of all participating users within a unit time interval, extract key semantic words from the text content, convert each feature word into a word vector representation using a word embedding model, and combine it with the timestamp of the current input and the user's identity encoding vector to form a document operation vector. Temporal features are extracted from continuous document operation vectors, and the context relationship vector is calculated using a weighted summation method. The semantic risk assessment module is used to assess the semantic deviation of the target user's current input content relative to the global context. It collects the new text content input by the target user per unit time, uses a word embedding model to transform it into a new document operation vector, and calculates the difference between the new document operation vector and the context relation vector in the normalized dimensional space. The evaluation and decision-making module is used to extract the lexical feature vector, sentence structure feature vector, and conflict judgment vector corresponding to the target user's current input content. After concatenating these three vectors, dimensionality reduction is performed to obtain the real-time behavior representation vector under the current input state. The difference between the real-time behavior representation vector and the fingerprint vector is calculated. By constructing a comprehensive risk assessment function, when the value obtained by the comprehensive risk assessment function after a certain input is completed is greater than the threshold, an early warning prompt is triggered.

[0021] Furthermore, the user behavior modeling module includes an input feature management unit, a conflict judgment management unit, and a behavior fingerprint vector management unit; The input feature management unit manages the input features recorded by the user, including lexical features and sentence structure features. The conflict judgment management unit manages the user's historical records of conflict handling operations. The behavior fingerprint vector management unit is used to form the user's original feature vector based on the collected user features, and inputs it into a fully connected neural network layer for dimensionality reduction to obtain the user's behavior fingerprint vector. Furthermore, the context memory network management module includes: an operation sequence management unit, a document operation vector management unit, and a relation vector management unit; The operation sequence management unit is used to collect the input content and identity identifier of each operator in the current shared document operation; the document operation vector management unit is used to vectorize the input content and identity identifier and sort them to obtain the document operation sequence; and the relation vector management unit is used to calculate the context relation vector through a bidirectional long short-term memory network encoder and attention mechanism. Furthermore, the semantic risk assessment module includes: a current input semantic management unit and a semantic deviation calculation unit; The current input semantic management unit is used to calculate the document operation vector corresponding to the text content input by the target user in the current unit of time, and the semantic deviation calculation unit is used to calculate the difference between the document operation vector and the context relation vector. Furthermore, the evaluation and decision-making module includes: a current input state vector management unit, a behavior anomaly calculation unit, and a comprehensive evaluation unit; The current input state vector management unit manages the lexical feature vector, sentence structure feature vector, and conflict judgment vector corresponding to the user's current unit of time. The behavior anomaly calculation unit compares the difference between the current input state vector and the behavior fingerprint vector. The comprehensive evaluation unit determines the corresponding risk response strategy based on the evaluation structure of the comprehensive evaluation function.

[0022] Compared with the prior art, the beneficial effects of the present invention are: 1. By performing structured processing of user behavior to achieve fingerprint modeling, dual-dimensional risk assessment, and dynamic intervention mechanism, a technically complete, data-driven, and deployable collaborative auxiliary decision-making method was constructed. 2. By combining the matching degree analysis of real-time behavior state vectors and historical behavior fingerprints, a consistency verification mechanism for individual behavior patterns was established, which effectively identifies style abrupt changes, misoperations, or abnormal intentions; the comprehensive risk assessment function integrates the dual dimensions of semantic deviation and behavioral abnormality, supports dynamic threshold control and personalized parameter adjustment, and reduces the probability of semantic noise injection in collaborative documents. 3. By deeply integrating user behavior fingerprints and contextual semantic memory, a collaborative auxiliary decision-making mechanism with forward-looking risk warning capabilities has been constructed. This mechanism addresses the shortcomings of traditional methods, which can only provide pre-emptive prevention based on post-event arbitration, thereby improving the content credibility, semantic coherence, and decision reliability of the collaborative office system. Attached Figure Description

[0023] Fig. 1 This is a schematic diagram of the data collaboration and decision support system based on deep learning according to the present invention; Fig. 2 This is a flowchart illustrating the data collaboration and decision support method based on deep learning of this invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example: Figs. 1-2 As shown, this invention provides a technical solution: a data collaboration and decision support system and method based on deep learning, the method comprising: Step 1: Identify the behavioral characteristics of the target user and establish the user's behavioral fingerprint vector. Take all the historical operation logs of the target user in the online collaborative office scenario as input data, extract the keywords and their corresponding word frequency information, grammatical structure type distribution and user's handling behavior records for conflict events, and construct the lexical feature vector, sentence structure feature vector and conflict judgment vector respectively. Concatenate them in sequence to form the original high-dimensional feature vector, and obtain the fingerprint vector after dimensionality reduction mapping. Step 1 includes: Step 1-1: Take a user in the remote collaborative office scenario as the target user, collect all historical operation logs of the target user in the online collaborative office scenario, collect the keywords in the operation logs and the word frequency corresponding to the keywords, establish a vocabulary feature vector, collect the grammatical structure features of the target user's input content records, and establish a sentence structure feature vector. Step 1-2: Collect the processing records of the target user when there are conflicting operations in the network collaborative office scenario. Classify the target user's processing records by three tags: accept processing, do not accept processing, and abandon processing. Collect the proportion of each tag and establish a conflict judgment vector. Steps 1-3: Concatenate the lexical feature vector, sentence structure feature vector, and conflict judgment vector to obtain the original feature vector of the target user. Then, perform dimensionality reduction through a fully connected neural network layer to output the behavioral fingerprint vector of the target user. Step 2: Obtain the text input content and corresponding timestamp information of all participating users within a unit time interval, extract key semantic words from their text content, use a word embedding model to convert feature words into word vector representations, and combine them with the timestamp of this input and the user's identity encoding vector to form a document operation vector. Perform temporal feature extraction on continuous document operation vectors, and calculate the context relationship vector using a weighted summation method. Step 2 includes: Step 2-1: In the operation record of the current shared document, obtain the text content entered by all users within a unit of time, the timestamp of each input, obtain the identity information of all users who participated in the operation of the current shared document within a unit of time, and set the identity identifier for each user; Step 2-2: Obtain the text content input by any user each time, extract the feature words in the text content, convert the feature words into word vectors, concatenate the word vectors, timestamps and identity identifiers in a single input process to obtain document operation vectors, sort the document operation vectors according to the time sequence of input to obtain the document operation sequence; Steps 2-3: Input the document operation sequence into the bidirectional long short-term memory network encoder to obtain the hidden state sequence, where the i-th hidden state in the hidden state sequence is denoted as hi, and the attention weight αi corresponding to the i-th hidden state is calculated. ,in, W represents the weight matrix, and Q represents the total number of hidden states in the hidden state sequence. Steps 2-4: Calculate the context relation vector E. ; Step 3: Evaluate the semantic deviation of the target user's current input content relative to the global context, collect the new text content input by the target user per unit time, and convert it into a new document operation vector. In the normalized dimensional space, calculate the difference between the new document operation vector and the context relation vector. Step 3 includes: Step 3-1: Collect the text content input by the target user within a unit of time. Using the same word embedding model as in Step 2, obtain the document operation vector of the text content. Normalize the document operation vector and the context relation vector respectively, and then map them to the same dimension. Step 3-2: Calculate the difference D between the document operation vector and the context relation vector in the same dimension; Step 4: Extract the lexical feature vector, sentence structure feature vector, and conflict judgment vector corresponding to the target user's current input content. After concatenating these three vectors, perform dimensionality reduction to obtain the real-time behavior representation vector under the current input state. Calculate the difference between the real-time behavior representation vector and the fingerprint vector. By constructing a comprehensive risk assessment function, when the value obtained by the comprehensive risk assessment function after a certain input is completed is greater than the threshold, a warning prompt is triggered. Step 4 includes: Step 4-1: Based on the text content newly input by the target user within a unit of time, extract the lexical feature vector, sentence structure feature vector, and conflict judgment vector corresponding to the text content. After concatenating these three vectors, perform dimensionality reduction to obtain the real-time behavior representation vector under the current input state. Calculate the difference between the real-time behavior representation vector and the fingerprint vector. By constructing a comprehensive risk assessment function, when the value obtained by the comprehensive risk assessment function after a certain input is completed is greater than the threshold, trigger an early warning prompt. Step 4-2: Concatenate the vocabulary feature vector, sentence structure feature vector, and conflict judgment vector corresponding to each unit of time, and perform dimensionality reduction through the same fully connected neural network layer as in Step 1-3 to obtain the current input state vector of the current target user. Compare the difference between the current input state vector and the behavior fingerprint vector and record it as P. Step 4-3: Establish a comprehensive evaluation function R, R=α×D+β×P, α+β=1. When R is greater than the threshold after a certain text input by the target user, a reminder is triggered and the submission of the result of that text input is paused.

[0026] The following provides a specific implementation example to demonstrate the complete computation process of the present invention in a real collaborative scenario.

[0027] In this embodiment, the shared document is a draft of a market analysis report.

[0028] Time window Second.

[0029] Current time Window Coverage .

[0030] User u (ID=101) is a senior market analyst with a history of stable behavior.

[0031] A total of Q=4 operations occurred within the window: User 101 entered: "Q3 sales increased by 12% year-on-year"; User 102 entered: "Mainly benefited from channel expansion in East China"; User 101 entered: "But the profit margin dropped to 8%"; User 101 tried typing: "Because the cost of raw materials has skyrocketed by 300%".

[0032] Step 1: Construct the behavioral fingerprint vector of user 101 ; Historical vocabulary characteristics: High-frequency words include "sales revenue", "year-on-year growth", "profit margin", "channels", "costs", etc., which are normalized after TF-IDF weighting.

[0033] Sentence structure characteristics: 70% are subject-verb-object structures, 20% are adversative complex sentences, and 10% are data declarative sentences. .

[0034] Conflict resolution record: 60% accepted, 30% rejected, 10% abandoned. .

[0035] After concatenation, input to a fully connected layer, output .

[0036] Step 2: Construct the context relation vector E; Semantic vectors are extracted from the four inputs, concatenated with timestamps, and normalized to obtain (0.04, 0.07, 0.125, 0.15) respectively. The user ID is then embedded, resulting in... After BiLSTM encoding, .

[0037] The weights calculated by the attention mechanism are: ; Weighted summation .

[0038] Step 3: Calculate the semantic deviation D; Latest Input "Because raw material costs have skyrocketed by 300%."

[0039] Its semantic vector After projection .

[0040] Calculate the Euclidean distance with E after normalization. ,but: .

[0041] Step 4: Calculate the behavioral risk deviation P; Input characteristics of user 101 in the current window: Vocabulary: The addition of extreme expressions such as "surge" and "300%" is inconsistent with the historical style of "moderate growth" and "minor fluctuations"; Sentence structure: Uses a cause-and-effect sentence, which is in accordance with convention (no abnormalities); Conflict vector: using historical values ​​[0.6, 0.3, 0.1].

[0042] After splicing, the same fully connected layer is obtained .

[0043] If the cosine similarity is calculated to be 0.75, then: ; Comprehensive risk assessment: set up , ,but: ; Set threshold .

[0044] because The system triggered an intervention: it paused the submission and prompted, "The current input 'because the cost of raw materials has soared by 300%' deviates significantly from the context and has an abnormal style. It is recommended to verify the accuracy of the data."

[0045] The system includes: The module includes a user behavior modeling module, a context memory network management module, a semantic risk assessment module, and an assessment and decision-making module. The user behavior modeling module is used to identify the behavioral characteristics of target users and establish user behavior fingerprint vectors. It takes all historical operation logs of target users in online collaborative office scenarios as input data, extracts keywords and their corresponding word frequency information, grammatical structure type distribution and user's handling behavior records for conflict events, and constructs lexical feature vectors, sentence structure feature vectors and conflict judgment vectors respectively. These are then concatenated to form the original high-dimensional feature vector. After performing dimensionality reduction mapping, the fingerprint vector is obtained. The user behavior modeling module includes an input feature management unit, a conflict judgment management unit, and a behavior fingerprint vector management unit. The input feature management unit manages the input features recorded by the user, including lexical features and sentence structure features. The conflict judgment management unit manages the user's historical records of conflict handling operations. The behavior fingerprint vector management unit is used to form the user's original feature vector based on the collected user features, input it into a fully connected neural network layer for dimensionality reduction, and obtain the user's behavior fingerprint vector. The context memory network management module is used to obtain the text input content and corresponding timestamp information of all participating users within a unit time interval, extract key semantic words from the text content, convert each feature word into a word vector representation using a word embedding model, and combine it with the timestamp of the current input and the user's identity encoding vector to form a document operation vector. Temporal features are extracted from continuous document operation vectors, and the context relationship vector is calculated using a weighted summation method. The context memory network management module includes: an operation sequence management unit, a document operation vector management unit, and a relation vector management unit. The operation sequence management unit is used to collect the input content and identity identifier of each operator in the current shared document operation. The document operation vector management unit is used to vectorize the input content and identity identifier and sort them to obtain the document operation sequence. The relation vector management unit is used to calculate the context relation vector through a bidirectional long short-term memory network encoder and an attention mechanism. The semantic risk assessment module is used to assess the semantic deviation of the target user's current input content relative to the global context. It collects the new text content input by the target user per unit time, uses a word embedding model to transform it into a new document operation vector, and calculates the difference between the new document operation vector and the context relation vector in the normalized dimensional space. The semantic risk assessment module includes a current input semantic management unit and a semantic deviation calculation unit. The current input semantic management unit is used to calculate the document operation vector corresponding to the text content input by the target user in the current unit of time, and the semantic deviation calculation unit is used to calculate the difference between the document operation vector and the context relationship vector. The evaluation and decision-making module is used to extract the lexical feature vector, sentence structure feature vector and conflict judgment vector corresponding to the current input content of the target user. After concatenating these three vectors, dimensionality reduction is performed to obtain the real-time behavior representation vector under the current input state. The difference between the real-time behavior representation vector and the fingerprint vector is calculated. By constructing a comprehensive risk assessment function, when the value obtained by the comprehensive risk assessment function after a certain input is completed is greater than the threshold, an early warning prompt is triggered. The evaluation and decision-making module includes: a current input state vector management unit, a behavior anomaly calculation unit, and a comprehensive evaluation unit. The current input state vector management unit manages the lexical feature vector, sentence structure feature vector, and conflict judgment vector corresponding to the user's current unit of time. The behavior anomaly calculation unit compares the difference between the current input state vector and the behavior fingerprint vector. The comprehensive evaluation unit determines the corresponding risk response strategy based on the evaluation structure of the comprehensive evaluation function.

[0046] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A data collaboration and decision support method based on deep learning, characterized by: The methods include: Step 1: Identify the behavioral characteristics of the target user and establish the user's behavioral fingerprint vector. Take all the historical operation logs of the target user in the online collaborative office scenario as input data, extract the keywords and their corresponding word frequency information, grammatical structure type distribution and user's handling behavior records for conflict events, and construct the lexical feature vector, sentence structure feature vector and conflict judgment vector respectively. Concatenate them in sequence to form the original high-dimensional feature vector, and obtain the fingerprint vector after dimensionality reduction mapping. Step 2: Obtain the text input content and corresponding timestamp information of all participating users within a unit time interval, extract key semantic words from their text content, use a word embedding model to convert feature words into word vector representations, and combine them with the timestamp of this input and the user's identity encoding vector to form a document operation vector. Perform temporal feature extraction on continuous document operation vectors, and calculate the context relationship vector using a weighted summation method. Step 3: Evaluate the semantic deviation of the target user's current input content relative to the global context, collect the new text content input by the target user per unit time, and transform it into a new document operation vector. In the normalized dimensional space, calculate the difference between the new document operation vector and the context relation vector. Step 4: Extract the lexical feature vector, sentence structure feature vector, and conflict judgment vector corresponding to the target user's current input content. After concatenating these three vectors, perform dimensionality reduction processing to obtain the real-time behavior representation vector under the current input state. Calculate the difference between the real-time behavior representation vector and the fingerprint vector. By constructing a comprehensive risk assessment function, when the value obtained by the comprehensive risk assessment function after a certain input is completed is greater than the threshold, a warning prompt is triggered.

2. The data collaboration and decision support method based on deep learning according to claim 1, characterized in that: Step 1 includes: Step 1-1: Take a user in the remote collaborative office scenario as the target user, collect all historical operation logs of the target user in the online collaborative office scenario, collect the keywords in the operation logs and the word frequency corresponding to the keywords, establish a vocabulary feature vector, collect the grammatical structure features of the target user's input content records, and establish a sentence structure feature vector. Step 1-2: Collect the processing records of the target user when there are conflicting operations in the network collaborative office scenario. Classify the target user's processing records by three tags: accept processing, do not accept processing, and abandon processing. Collect the proportion of each tag and establish a conflict judgment vector. Steps 1-3: Concatenate the lexical feature vector, sentence structure feature vector, and conflict judgment vector to obtain the original feature vector of the target user. Then, perform dimensionality reduction through a fully connected neural network layer to output the behavioral fingerprint vector of the target user.

3. The data collaboration and decision support method based on deep learning according to claim 2, characterized in that: Step 2 includes: Step 2-1: In the operation record of the current shared document, obtain the text content entered by all users within a unit of time, the timestamp of each input, obtain the identity information of all users who participated in the operation of the current shared document within a unit of time, and set the identity identifier for each user; Step 2-2: Obtain the text content input by any user each time, extract the feature words in the text content, convert the feature words into word vectors, concatenate the word vectors, timestamps and identity identifiers in a single input process to obtain document operation vectors, sort the document operation vectors according to the time sequence of input to obtain the document operation sequence; Steps 2-3: Input the document operation sequence into the bidirectional long short-term memory network encoder to obtain the hidden state sequence, where the i-th hidden state in the hidden state sequence is denoted as hi, and the attention weight αi corresponding to the i-th hidden state is calculated. ,in, W represents the weight matrix, and Q represents the total number of hidden states in the hidden state sequence. Steps 2-4: Calculate the context relation vector E. .

4. The data collaboration and decision support method based on deep learning according to claim 3, characterized in that: Step 3 includes: Step 3-1: Collect the text content input by the target user within the unit time. Using the same word embedding model as in Step 2, obtain the document operation vector of the text content. Normalize the document operation vector and the context relationship vector respectively, and then map them to the same dimension. Step 3-2: Calculate the difference D between the document operation vector and the context relation vector in the same dimension.

5. The data collaboration and decision support method based on deep learning according to claim 4, characterized in that: Step 4 includes: Step 4-1: Based on the text content newly input by the target user within a unit of time, extract the lexical feature vector, sentence structure feature vector, and conflict judgment vector corresponding to the text content. After concatenating these three vectors, perform dimensionality reduction processing to obtain the real-time behavior representation vector under the current input state. Calculate the difference between the real-time behavior representation vector and the fingerprint vector. By constructing a comprehensive risk assessment function, when the value obtained by the comprehensive risk assessment function after a certain input is completed is greater than the threshold, trigger an early warning prompt. Step 4-2: Concatenate the vocabulary feature vector, sentence structure feature vector, and conflict judgment vector corresponding to each unit of time, and perform dimensionality reduction through the same fully connected neural network layer as in Step 1-3 to obtain the current input state vector of the current target user. Compare the difference between the current input state vector and the behavior fingerprint vector and record it as P. Step 4-3: Establish a comprehensive evaluation function R, R=α×D+β×P, α+β=1. When R is greater than the threshold after a certain text input by the target user, a reminder is triggered and the submission of the result of the certain text input is paused.

6. A deep learning-based data collaboration and decision support system, used to execute the deep learning-based data collaboration and decision support method according to any one of claims 1-5, characterized in that: The system includes: a user behavior modeling module, a context memory network management module, a semantic risk assessment module, and an assessment and decision-making module; The user behavior modeling module is used to identify the behavioral characteristics of target users and establish user behavior fingerprint vectors. It takes all historical operation logs of target users in online collaborative office scenarios as input data, extracts keywords and their corresponding word frequency information, grammatical structure type distribution and user's handling behavior records for conflict events, and constructs lexical feature vectors, sentence structure feature vectors and conflict judgment vectors respectively. These are then concatenated to form the original high-dimensional feature vector. After performing dimensionality reduction mapping, the fingerprint vector is obtained. The context memory network management module is used to obtain the text input content and corresponding timestamp information of all participating users within a unit time interval, extract key semantic words from the text content, convert each feature word into a word vector representation using a word embedding model, and combine it with the timestamp of the current input and the user's identity encoding vector to form a document operation vector. Temporal features are extracted from continuous document operation vectors, and the context relationship vector is calculated using a weighted summation method. The semantic risk assessment module is used to assess the semantic deviation of the target user's current input content relative to the global context. It collects the new text content input by the target user per unit time, uses a word embedding model to transform it into a new document operation vector, and calculates the difference between the new document operation vector and the context relation vector in the normalized dimensional space. The evaluation and decision-making module is used to extract the lexical feature vector, sentence structure feature vector, and conflict judgment vector corresponding to the current input content of the target user. After concatenating these three vectors, dimensionality reduction processing is performed to obtain the real-time behavior representation vector under the current input state. The difference between the real-time behavior representation vector and the fingerprint vector is calculated. By constructing a comprehensive risk assessment function, when the value obtained by the comprehensive risk assessment function after a certain input is completed is greater than the threshold, an early warning prompt is triggered.

7. The data collaboration and decision support system based on deep learning according to claim 6, characterized in that: The user behavior modeling module includes an input feature management unit, a conflict judgment management unit, and a behavior fingerprint vector management unit. The input feature management unit manages the input features recorded by the user, including lexical features and sentence structure features. The conflict judgment management unit manages the user's historical records of conflict handling operations. The behavior fingerprint vector management unit is used to form the user's original feature vector based on the collected user features, and inputs it into a fully connected neural network layer for dimensionality reduction to obtain the user's behavior fingerprint vector.

8. The data collaboration and decision support system based on deep learning according to claim 6, characterized in that: The context memory network management module includes: an operation sequence management unit, a document operation vector management unit, and a relation vector management unit; The operation sequence management unit is used to collect the input content and identity identifier of each operator in the current shared document operation. The document operation vector management unit is used to vectorize the input content and identity identifier and sort them to obtain the document operation sequence. The relation vector management unit is used to calculate the context relation vector through a bidirectional long short-term memory network encoder and attention mechanism.

9. The data collaboration and decision support system based on deep learning according to claim 6, characterized in that: The semantic risk assessment module includes: a current input semantic management unit and a semantic deviation calculation unit; The current input semantic management unit is used to calculate the document operation vector corresponding to the text content input by the target user in the current unit of time, and the semantic deviation calculation unit is used to calculate the difference between the document operation vector and the context relation vector.

10. The data collaboration and decision support system based on deep learning according to claim 6, characterized in that: The evaluation and decision-making module includes: a current input state vector management unit, a behavior anomaly calculation unit, and a comprehensive evaluation unit; The current input state vector management unit manages the lexical feature vector, sentence structure feature vector, and conflict judgment vector corresponding to the user's current unit of time. The behavior anomaly calculation unit compares the difference between the current input state vector and the behavior fingerprint vector. The comprehensive evaluation unit determines the corresponding risk response strategy based on the evaluation structure of the comprehensive evaluation function.

Images