Work automation processing platform based on artificial intelligence and processing method thereof

Abstract

The invention relates to the field of work automation processing, and provides an artificial intelligence-based work automation processing platform and a processing method thereof, and the platform comprises a work task module, a task issue module, an issue template module, a template agenda module, an agenda vectorization module and a management and control automation module. By adopting an end-to-end and knowledge-driven design thought and through a plurality of advanced machine learning technologies, each link of an enterprise work task is automatically optimized, and finally fine management, automatic control and efficient closed-loop management of an enterprise work process are realized.

Description

Technical Field

[0001] The present invention relates to the field of work automation processing, and provides a work automation processing platform based on artificial intelligence and its processing method. Background Art

[0002] With the continuous expansion of the scale of enterprises, the various work processes of enterprises are complex, which will reduce the work efficiency of enterprises, resulting in problems such as task omission or irregular work development. To a certain extent, the complex work processes hinder the healthy development of enterprises. Therefore, using modern computer technology to achieve automated processing of work tasks has become a trend of the times.

[0003] After retrieval, the invention patent with the Chinese patent number CN114169778A discloses a distribution task workload evaluation system and method based on a work platform, which relates to the field of computer information processing technology. In the prior art, such work automation systems generally upload information to the cloud, resulting in low work efficiency, insufficiently fine control, and lack of intelligence due to a large amount of data maintenance. Summary of the Invention

[0004] In view of the problems of insufficient refinement and intelligence in the existing work automation systems, the present invention provides a work automation processing platform based on artificial intelligence and its processing method.

[0005] The technical solution of the present invention is as follows:

[0006] The present invention proposes a work automation processing platform based on artificial intelligence, which includes: a work tasking module, a task topic module, a topic templating module, a template agendizing module, an agenda vectorizing module, and a control automation module. Among them,

[0007] The work tasking module, through training a deep learning model combined with an attention mechanism, completes the automatic mapping of the work process to task definition, and converts the work process description into specific task splitting and definition;

[0008] The task topic module, through pre-training the task topic field with a natural language model, learns the optimal matching relationship between tasks and topics, and defines and divides topics for the task descriptions read from the work tasking module;

[0009] The topic templating module, through a retrieval method of word vector matching, identifies and matches topic templates for the topics of the task topic module;

[0010] The template agendizing module, through a matching learning method based on a graph neural network, identifies and matches the agenda of the topic template, and performs time allocation and priority identification for the agenda;

[0011] The agenda vectorization module converts the agenda of the template agenda module into a vector form through word embedding technology;

[0012] The control automation module defines the workflow nodes and transfer logic of other modules through a workflow engine, and drives the automatic execution of the processes of other modules.

[0013] In some embodiments, the work tasking module uses a large number of real workflow descriptions and corresponding task definitions as training data, and is trained using a seq2seq model. The seq2seq model consists of an encoder and a decoder. The encoder uses a bidirectional LSTM to encode the input workflow description. The specific algorithm of the seq2seq model encoder is formula (1);

[0014] ht = f(xt, ht-1) (1)

[0015] Where, ht is the hidden state vector of the encoder at the current time step t, xt is the input of the encoder at the current time step t, and f is the recurrent function of the encoder.

[0016] In some embodiments, the decoder of the seq2seq model integrates an attention mechanism and can automatically generate corresponding task splitting and definitions. The specific algorithm of the seq2seq model decoder is formula (2), and the specific algorithms of the attention mechanism are formula (3), formula (4) and formula (5);

[0017] p(yt|y1:t-1, x) = g(yt-1, st, ct) (2)

[0018] Where, yt is the target word that the decoder needs to predict and output at time step t, y1:t-1 is the output sequence generated by the decoder from time step 1 to t-1, x is the source language sequence input by the encoder, p(yt|y1:t-1, x) is the predicted probability distribution of the decoder for the output word yt at time step t, yt-1 is the word generated by the decoder at the previous time step t-1, st is the hidden state at time step t, ct is the context vector, and g is the recurrent model of the decoder;

[0019] etj = a(st-1, hj) (3)

[0020] αtj = exp(etj) / Σk exp(etk) (4)

[0021] ct = Σj αtjhj (5)

[0022] Among them, a is the correlation function, hj is the state of the encoder at each moment, etj is the attention weight scoring value of the encoder state hj at time t, st-1 is the state of the decoder at the current step, hj is the state of the encoder at time step j, a is the scoring function of the attention mechanism, αtj is the normalized attention weight, Σkexp(etk) is the normalization factor of etj, and ct is the dynamic context vector.

[0023] In some embodiments, the task topicization module uses a fine-tuned BERT model, performs large-scale pre-training for the task topic domain and uses real task data and corresponding topic annotations for fine-tuning to obtain a fine-tuned model specific to task topic generation. The pre-training function of the fine-tuned BERT model is formula (6), the fine-tuning function of the fine-tuned BERT model is formula (7), and the inference function of the fine-tuned BERT model is formula (8);

[0024] L = Σ E(Ti, T^i) + Σ E(Si, S^i) (6)

[0025] Among them, T is the text sequence, S is the text sequence pair, E is the loss function, L is the overall training loss of the BERT model, E is the loss function, Ti is the i-th word of the text sequence paragraph, T^i is the prediction of the i-th word of the text sequence paragraph by the model, ΣE(Ti, T^i) is the total loss value of the text sequence paragraph, Si is the i-th word of the text sequence pair, S^i is the prediction of the i-th word of the text sequence pair by the model, and ΣE(Si, S^i) is the total loss value of the text sequence pair;

[0026] L = Σ E(yi, y^i) (7)

[0027] Among them, y is the topic annotation, y^ is the topic predicted by the BERT model, L is the loss function during BERT fine-tuning, E is the loss function, yi is the true target value annotated in the sample, y^i is the target value predicted by the model, and Σ is the sum of the loss function over all samples;

[0028] topic = argmax P(topic|t) (8)

[0029] Among them, topic is the most likely topic predicted by the model, argmax is to find the value that makes P(topic|t) the largest, P(topic|t) is the prediction probability of each topic topic under the condition of the input statement t calculated by the model, and t is the input sentence statement.

[0030] In some embodiments, the topic templatization module represents topics and topic templates as word vectors by constructing a topic word vector space, and performs similarity matching between the vectors of the topics and the vectors of the topic templates to achieve topic recognition and matching of topic templates.

[0031] In some embodiments, the template agenda module constructs a knowledge graph based on the matching learning method of a graph neural network from the agenda of the topic template and its time and priority data, uses the graph neural network to perform vector representation learning on the knowledge graph and the template agenda of the template agenda module, and performs similarity matching through the knowledge graph and the template agenda to generate the matching time allocation and priority arrangement of the agenda; the encoding function of the graph neural network is formula (9), the similarity function of the graph neural network is formula (10), and the inference function of the graph neural network is formula (11);

[0032] hv = σ(Σu∈N(v) Wu,v hu) (9)

[0033] where hv is the feature representation of node v, σ is a non-linear activation function, N(v) is the set of neighbor nodes of node v, u is a neighbor node of node v, Wu,v is the weight matrix from node u to node v, hu is the feature representation of node u, and Σ is the summation operation over all neighbors of node v;

[0034] S(g1,g2) = f(h1, h2) (10)

[0035] where S(g1, g2) is the similarity score between graph g1 and graph g2, g1 is the input source graph, g2 is the template graph to be matched, h1 is the graph neural network encoding vector of source graph g1, h2 is the graph neural network encoding vector of template graph g2, f is the similarity calculation function, and h1 and h2 are the encoding representations of the two agenda graphs respectively;

[0036] gt = argmax S(g, gi) (11)

[0037] gt is the result of the finally matched template graph, argmax is to find the gi that makes S(g, gi) maximum, S(g, gi) is the matching similarity score between the source graph g and the i-th template graph gi, g is the input source graph, such as a new agenda graph, and gi is the i-th template graph.

[0038] In some embodiments, the agenda vectorization module trains a Word2Vec word vector model using a large amount of meeting agenda data, obtains the word vectors of the agenda text of the template agenda module, performs weighted averaging on the word vectors to obtain the vector representation of the agenda statement, calculates the similarity of different agenda statements through the cosine similarity of the vector representation, and analyzes the content features of the agenda by using clustering and topic modeling methods; the word vector training function of the Word2Vec model is formula (12), the agenda statement vector function of the Word2Vec model is formula (13), and the similarity calculation function of the Word2Vec model is formula (14);

[0039] L = Σ log p(wt|wt-k, ..., wt+k) (12)

[0040] Where L is the loss function of word vector training, log p(wt|wt-k,...,wt+k) is the logarithmic prediction probability of the current word wt based on its context words wt-k to wt+k, wt is the current word, that is, the central word to be predicted, wt-k is the kth word before the central word wt, wt+k is the kth word after the central word wt, and Σ is the sum of all words in the training corpus;

[0041] vs = Σ wivi / |S| (13)

[0042] Where vs is the vector representation of the agenda statement, wi is the weight coefficient of the ith word in the statement, vi is the word vector of the ith word in the statement, Σ is the sum of the word vectors of all words in the statement, and |S| is the length of the statement;

[0043] similarity(vs1,vs2) = cos(vs1, vs2) (14)

[0044] Where similarity(vs1,vs2) is the cosine similarity of the vectors of statements vs1 and vs2, vs1 is the statement vector of agenda statement 1, vs2 is the statement vector of agenda statement 2, and cos is the cosine similarity calculation function.

[0045] In some embodiments, the task execution dependency of the workflow engine of the control automation module is formula (15), and the task status update formula of the workflow engine is formula (16);

[0046] Tn+1 ≠ Tn (15)

[0047] Where Tn+1 depends on Tn, Tn is the nth task in the workflow, Tn+1 is the (n + 1)th task in the workflow, and ≠ means that the two tasks are different tasks;

[0048] St+1 = F(St, At) (16)

[0049] Where, St is the task status at the current time step t, St+1 is the task status at the next time step t+1, At is the action performed on the task at the current time step t, and F is the state transition function.

[0050] In some embodiments, the control automation module is also provided with a rule engine, a monitoring and warning system, a knowledge graph function, and a reinforcement learning function. The rule engine calculates and analyzes data in real time according to business rules to achieve automated decision-making for other modules; the monitoring and warning system tracks the work progress and anomalies of other modules, generates warnings and gives feedback; the reinforcement learning function trains an agent through the process data of other modules to provide decision-making suggestions and process optimization solutions for other modules; the knowledge graph function stores the knowledge of the work process, constructs a work knowledge system, and supports the process optimization and decision-making of other modules; the rule matching of the rule engine is formula (17), and the rule condition check is formula (18);

[0051] IF C1 and C2 THEN A1(17)

[0052] Where, IF is the start symbol of the rule condition, C1 is the first condition for the rule to be triggered, and is the logical "AND" relationship of multiple conditions, C2 is the second condition for the rule to be triggered, THEN is the separator between the condition and the action, and A1 is the action to be executed when the condition is met;

[0053] check(C1, data) -> true / false (18)

[0054] Where, check is a function to detect whether the rule condition is triggered, C1 is the rule condition to be detected, data is the input data used to detect whether the condition is met, and true / false is the boolean value output by the function, where true means the condition is triggered and false means the condition is not triggered;

[0055] The policy gradient of the reinforcement learning function is formula (19); the state value of the reinforcement learning function is formula (20); the reward expectation of the reinforcement learning function is formula (21);

[0056]

[0057] Where, is the gradient of the policy network parameter θ, J(θ) is the expected return corresponding to the policy network parameter θ, and Eπθ is the expectation of taking random samples according to the policy πθ, is the gradient of the log probability of the policy πθ with respect to θ, πθ(at|st) is the probability that the policy πθ takes the action at in the state st, and Qπ(st, at) is the value of taking the action at in the state st;

[0058] V(s) = E[Rt|st = s] (20)

[0059] where V(s) is the state value, E[] is the expected value, Rt| is the immediate reward obtained at time t, and st is the state at time t;

[0060] Q(s, a) = E[Rt| st = s, at = a] (21)

[0061] where Q(s, a) is the reward expectation, E[] is the expected value, Rt| is the immediate reward obtained in the state at time t, st is the state at time t, and at is the action taken at time t.

[0062] The present invention proposes a processing method for a work automation processing platform based on artificial intelligence. The method includes:

[0063] Step 1: Define the workflow nodes and dependencies in the workflow engine of the control automation module, construct the workflow model of the workflow, and set the state machine rules for each node in the rule engine of the control automation module;

[0064] Step 2: The workflow engine drives the workflow according to the workflow model to implement the taskification, topicization, templatization, agendaization, and vectorization of the workflow in the work taskification module, task topicization module, topic templatization module, template agendaization module, and agenda vectorization module;

[0065] Step 3: The reinforcement learning function and knowledge graph function of the control automation module collect the data and knowledge of the processing workflow in Step 2 and continuously improve and optimize the decision-making scheme of the platform.

[0066] Implementing the present invention has the following beneficial effects:

[0067] The present invention proposes a work automation processing platform based on artificial intelligence and its processing method. Implementing the present invention has the following beneficial effects:

[0068] 1. The taskification module of the present invention uses a seq2seq model and an attention mechanism to automatically split and define tasks for enterprise work processes. This module can handle more complex work processes and learn task definition patterns in different industries and scenarios. The attention mechanism can capture key information in work descriptions and generate matching task definitions. As the model iteratively learns, the taskification effect can be continuously optimized, the quality of task definitions can be improved, and automated and refined management of work processes can be achieved.

[0069] 2. The task-topicization module of the present invention automatically divides and defines topics for work tasks by fine-tuning the BERT model. This model can handle more complex task descriptions and learn task-topic correspondence patterns in different scenarios. The model can continuously improve its effect by incremental learning using new data, achieving automated topicization of work tasks.

[0070] 3. The topic-templatization module of the present invention automatically identifies work topics and matches and defines template topics through a retrieval method based on word vector matching. This model can handle more complex topics and consider semantic similarity. As the word vector space continues to expand, the matching effect will continue to improve, enabling efficient and automatic templatization of work topics.

[0071] 4. The template-agendization module of the present invention uses a matching learning method based on a graph neural network to automatically identify and match and define template agendas for meeting agendas or work processes, and perform time allocation and priority identification. The graph neural network can learn the structural knowledge of agendas to achieve better agenda understanding and matching. As the knowledge graph expands, the model can improve the matching effect through incremental learning, thereby achieving automated standardization of work processes.

[0072] 5. The agenda-vectorization module of the present invention automatically converts meeting agendas into vector forms through word embedding technology. The converted word vectors can efficiently and automatically obtain semantic features, enabling intelligent agenda management and analysis.

[0073] 6. The control automation module of the present invention uses technologies such as a workflow engine, a rule engine, a monitoring and warning system, and machine learning to achieve closed-loop automation of enterprise work. Compared with traditional methods, this model has powerful orchestration capabilities, high integration, supports variable instances, operates efficiently, is flexibly configured, has real-time monitoring, and is user-friendly. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 It is a schematic diagram of an enterprise work process of a work automation processing platform based on artificial intelligence proposed in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0075] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0076] As Figure 1 shown, the present invention proposes a work automation processing platform based on artificial intelligence. The platform includes: a work tasking module, a task topic module, a topic templating module, a template agenda module, an agenda vectorization module, and a control automation module.

[0077] 1. Work Tasking Module

[0078] The work tasking module completes the automatic mapping from the work process to the task definition by training a deep learning model combined with the attention mechanism, and converts the work process description into specific task splitting and definition. Through the use of the seq2seq model and the attention mechanism, the platform realizes the automatic mapping from the work flow to the task definition, and provides real-time task allocation, task monitoring, and task evaluation.

[0079] By subdividing the work process and tasks into independent task units, each task is assigned a clear goal, executor, and time requirement, and the relevant tasks are interconnected according to the actual business to form a new and complete work process, realizing the refined management and automatic control of the work process.

[0080] The seq2seq model consists of an encoder and a decoder. The encoder uses a bidirectional LSTM to encode the input work process description, and the decoder integrates the attention mechanism and can automatically generate the corresponding task splitting and definition. The training data is a large number of real work process descriptions and corresponding task definitions. By maximizing the probability of the real task definition, the model parameters are adjusted to realize the end-to-end mapping learning from the work process to the task definition. In the deployment, when a new work process description is input, the seq2seq model can automatically output the split and defined tasks, realizing work tasking.

[0081] The specific algorithm formulas of the seq2seq model and the attention mechanism are as follows:

[0082] The encoder of the seq2seq model is as formula (1) below:

[0083] ht = f(xt, ht-1) (1)

[0084] Where:

[0085] ht is the hidden state at time t, and f is a recursive function such as LSTM.

[0086] $h_t$ is the hidden state vector of the encoder at the current time step $t$, which represents the semantic encoding information of the sentence after the input at time step $t$.

[0087] $x_t$ is the input of the encoder at the current time step $t$, that is, the $t$-th element in the input sequence $X$, representing the $t$-th word of the input sentence.

[0088] $h_{t - 1}$ is the hidden state vector of the encoder at the previous time step $t - 1$, which contains the encoding information of the input sentence up to time step $t - 1$.

[0089] $f$ is the recurrent function of the encoder, usually an LSTM or GRU, which is a function that maps the current input $x_t$ and the previous state $h_{t - 1}$ to the current state $h_t$.

[0090] The decoder of the seq2seq model is as follows in Equation (2):

[0091] $p(y_t|y_{1:t - 1}, x) = g(y_{t - 1}, s_t, c_t)\ (2)$

[0092] Where:

[0093] $p(y_t|y_{1:t - 1}, x)$ is the predicted probability distribution of the output word $y_t$ by the decoder at time $t$.

[0094] $y_t$ is the target word that the decoder needs to predict and output at time $t$.

[0095] $y_{1:t - 1}$ is the output sequence generated by the decoder from time 1 to $t - 1$.

[0096] $x$ is the source language sequence input to the encoder.

[0097] $s_t$ is the hidden state of the decoder at the current time $t$, which encodes the decoding context information up to $t$.

[0098] $c_t$ is the dynamic context vector generated by the attention mechanism, which focuses on the information of the source language sequence $x$.

[0099] $y_{t - 1}$ is the word generated by the decoder at the previous time step $t - 1$, representing the previous context semantics.

[0100] $g$ is the recurrent model of the decoder, which maps the previous context information to the prediction distribution of the current word.

[0101] The attention mechanism is given by Equations (3), (4), and (5)

[0102] $e_{tj} = a(s_{t - 1}, h_j)\ (3)$

[0103] $\alpha_{tj} = \frac{\exp(e_{tj})}{\sum_{k}\exp(e_{tk})}\ (4)$

[0104] $c_t=\sum_{j}\alpha_{tj}h_j\ (5)$

[0105] Where:

[0106] $a$ is the correlation function, and $h_j$ is the state of the encoder at each moment.

[0107] $e_{tj}$ is the attention weight score value of the encoder state $h_j$ at time $t$.

[0108] $s_{t - 1}$ is the state of the decoder at the current step, which is used to calculate the attention weight.

[0109] $h_j$ is the state of the encoder at time step $j$, which contains the encoding information of the source language side.

[0110] $a$ is the scoring function of the attention mechanism, which calculates the correlation between $s_{t - 1}$ and $h_j$.

[0111] $\alpha_{tj}$ is the attention weight after softmax normalization, which represents the attention degree of time $t$ to the encoder state $h_j$.

[0112] $\sum_{k}\exp(e_{tk})$ is the normalization factor of $e_{tj}$, which sums the scoring values of all encoder states.

[0113] $c_t$ is the dynamic context vector, which is the weighted sum of the attention weight and the encoder state.

[0114] Compared with the methods in the prior art, this seq2seq model can handle more complex work processes and learn the task definition patterns in different industries and scenarios. The attention mechanism can capture the key information of the job description and generate matching task definitions. As the model iteratively learns, the tasking effect can be continuously optimized, the quality of task definition can be improved, and the automated and refined management of the work process can be realized.

[0115] 2. Task topicization

[0116] The task topicization module pre-trains the task topic field through a natural language model, learns the optimal matching relationship between tasks and topics, and defines and divides the topics of the read work tasking module or the input task description. The task topicization module fine-tunes the BERT model, and the platform automatically divides and defines the topics of the work tasks.

[0117] First, conduct large-scale pre-training for the task topic field to obtain a fine-tuned BERT model with general semantic representations. Then, use real task data and corresponding topic annotations for fine-tuning to obtain a fine-tuned model specific to task topic generation. During fine-tuning, learn the optimal matching relationship between tasks and topics by maximizing the probability of real topics. When in use, input the work task description, and the BERT model can output the most matching task topic.

[0118] The objective function of BERT pre-training is formula (6):

[0119] L = Σ E(Ti, T^i) + Σ E(Si, S^i) (6)

[0120] Where:

[0121] T is the text sequence, S is the text sequence pair, and E is the loss function.

[0122] L is the overall training loss of the BERT model.

[0123] E is the loss function, which calculates the error between the predicted value and the true value.

[0124] Ti is the i-th word in the text sequence paragraph.

[0125] T^i is the prediction of the i-th word in the text sequence paragraph by the model.

[0126] ΣE(Ti, T^i) is the sum of the loss values of the text sequence paragraph, which realizes Masked LM training.

[0127] Si is the i-th word in the text sequence pair.

[0128] S^i is the prediction of the i-th word in the text sequence pair by the model.

[0129] ΣE(Si, S^i) is the sum of the loss values of the text sequence pair, which realizes Next Sentence Prediction training.

[0130] Pre-training is carried out through Masked LM and Next Sentence Prediction tasks.

[0131] The fine-tuning objective function is formula (7):

[0132] L = Σ E(yi, y^i) (7)

[0133] Where y is the topic annotation and y^ is the topic predicted by the BERT model.

[0134] L is the loss function during BERT fine-tuning.

[0135] E is the loss function, which calculates the error between the predicted value and the true value.

[0136] yi is the true target value annotated in the sample, which is the true topic of the sentence here.

[0137] y^i is the target value predicted by the model, that is, the predicted topic.

[0138] Σ is the sum of the loss function over all samples.

[0139] Fine-tuning training is performed by maximizing the probability of the true topic.

[0140] The inference function is given by Equation (8):

[0141] topic = argmax P(topic|t) (8)

[0142] For the new task t, the BERT model generates the most likely topic:

[0143] topic=argmax P(topic|t)

[0144] topic is the most likely topic predicted by the model.

[0145] argmax is to find the value that maximizes P(topic|t).

[0146] P(topic|t) is the predicted probability of each topic topic given the input statement t calculated by the model.

[0147] t is the input sentence statement.

[0148] When building the model, a pre-trained BERT model is used as the feature extractor. A multi-layer fully-connected network is connected as the task topic classifier.

[0149] For the pre-trained model, a large-scale general corpus is used to pre-train the BERT model to learn general semantic representations and encode semantic information.

[0150] For the fine-tuning model, the BERT model is fine-tuned using the real task description and the corresponding topic annotation data, and the mapping relationship from the task features to the topic is learned through the output layer classification.

[0151] When the model is applied to input new task description text, the BERT encoder extracts semantic features and the classifier outputs the most likely topic.

[0152] Compared with the methods of the prior art, this model can handle more complex task descriptions and learn the task topic correspondence patterns in different scenarios. The model effect can be continuously improved by incremental learning using new data, realizing the automatic topicization of work tasks. The platform automatically divides and defines topics for work tasks and provides real-time task allocation, progress monitoring, and result evaluation.

[0153] By decomposing work tasks into specific topics, it is possible to gain a clearer understanding of the content, objectives, and implementation steps of each topic, facilitating enterprises to better organize and coordinate the work of the team and avoid task overlap and conflicts. At the same time, task topicalization can also improve the transparency and traceability of work, making the work process clearer and more controllable.

[0154] 3. Topic templatization

[0155] The topic templatization module identifies and matches the topics of the task topicalization module with topic templates through a retrieval method based on word vector matching. The specific implementation method of the retrieval method of word vector matching in this module is as follows:

[0156] 1. Construct a topic word vector space, where each topic is represented as a word vector.

[0157] 2. Also perform vector representation on the defined topic templates.

[0158] 3. For new work topics, calculate the similarity between its vector and each topic template vector.

[0159] 4. Select the topic template with the highest similarity to the new topic vector as the matching output.

[0160] Compared with the methods of the existing technology, this model can handle more complex topics and consider semantic similarity. As the word vector space continues to expand, the matching effect will continue to improve. It can achieve efficient automatic templatization of work topics, provide real-time template customization, modification, guidance, and prompts, automatically summarize the update records of topic templates, and generate improvement suggestions.

[0161] Through topic templatization, repetitive work can be significantly reduced, errors and omissions can be minimized, work efficiency can be improved, consistency and standardization can be provided, ensuring that the execution of tasks meets the requirements. It can also improve the traceability and manageability of work, making the work process more controllable and transparent.

[0162] 4. Template agendaization

[0163] The template agendaization module identifies, matches the templates of the agendas of topic templates, and performs time allocation and priority identification on the agendas through a matching learning method based on graph neural networks. The specific implementation method of this template agendaization module is as follows:

[0164] 1. Construct a knowledge graph with the agenda and its time and priority data.

[0165] 2. Use graph neural networks to learn the vector representation of the agenda knowledge graph.

[0166] 3. Perform graph neural network learning vector representation learning on the template agendas.

[0167] 4. For the new agenda knowledge graph, calculate the similarity between its vector and the template agenda vector.

[0168] 5. Select the template agenda with the highest similarity and generate a matching time allocation and priority arrangement.

[0169] The encoding function of the graph neural network is formula (9):

[0170] hv = σ(Σu∈N(v) Wu,v hu) (9)

[0171] hv is the feature representation of node v, that is, the vector after encoding node v.

[0172] σ is a non-linear activation function, such as ReLU, etc.

[0173] N(v) is the set of neighbor nodes of node v.

[0174] u is a neighbor node of node v.

[0175] Wu,v is the weight matrix from node u to node v.

[0176] hu is the feature representation of node u, that is, the encoding vector of node u.

[0177] Σ is the summation operation over all neighbor nodes of node v.

[0178] N(v) are the neighbors of node v, W is the weight matrix, and hu are the features of the neighbor nodes.

[0179] The similarity function of the graph neural network is formula (10):

[0180] S(g1, g2) = f(h1, h2) (10)

[0181] S(g1, g2) is the similarity score between graph g1 and graph g2.

[0182] g1 is the input source graph, such as the new agenda knowledge graph.

[0183] g2 is the template graph to be matched, such as the template agenda graph.

[0184] h1 is the graph neural network encoding vector of source graph g1.

[0185] h2 is the graph neural network encoding vector of template graph g2.

[0186] f is the similarity calculation function, such as cosine similarity.

[0187] h1 and h2 are the encoding representations of the two agenda graphs respectively, and f is the similarity function.

[0188] The inference function of the graph neural network is formula (11):

[0189] gt = argmax S(g, gi) (11)

[0190] gt is the result of the finally matched template graph.

[0191] argmax is to find the gi that makes the maximum value of S(g, gi).

[0192] S(g, gi) is the matching similarity score between the source graph g and the i-th template graph gi.

[0193] g is the input source graph, such as a new agenda knowledge graph.

[0194] gi is the i-th template graph, and there are n template agenda graphs in total.

[0195] Compared with the methods of the prior art, the graph neural network can learn the structural knowledge of the agenda, achieve better agenda understanding and matching. The graph neural network can expand with the knowledge graph and improve the matching effect through incremental learning, so as to realize the automation and standardization of the work process.

[0196] Through template agenda setting, the time and energy for organizing meetings can be saved, the workload of meeting preparation can be reduced, the efficiency of meetings can be significantly improved, consistency and standardization can be provided, ensuring that the topics of meetings are fully discussed and decisions are made. It can also improve the efficiency and participation of meetings, making the meetings more targeted and productive.

[0197] 5. Agenda Vectorization

[0198] The agenda vectorization module converts the agenda of the template agenda setting module into a vector form through word embedding technology; the agenda vectorization module automatically converts the meeting agenda into a vector form through the Word2Vec word embedding technology. The specific implementation method of the template agenda setting module is as follows:

[0199] 1. Collect a large amount of meeting agenda data and train a Word2Vec word vector model.

[0200] 2. Segment the agenda text and obtain word vectors using a lookup word list.

[0201] 3. Perform weighted averaging on all word vectors in the agenda to obtain the vector representation of the agenda statement.

[0202] 4. Calculate the similarity of different agenda statements through the cosine similarity of vectors.

[0203] 5. Use methods such as clustering and topic modeling to analyze the content characteristics of the agenda vector.

[0204] 6. Continuously repeat the above steps to expand the agenda corpus and improve the accuracy of agenda vector training.

[0205] The word vector training function of the Word2Vec model is formula (12):

[0206] L = Σ log p(wt|wt-k, ..., wt+k) (12)

[0207] L is the loss function of word vector training.

[0208] log p(wt|wt-k,...,wt+k) is the logarithmic prediction probability of the current word wt based on its context words wt-k to wt+k.

[0209] wt is the current word, that is, the central word to be predicted.

[0210] wt-k is the kth word before the central word wt, the left context of wt.

[0211] wt+k is the kth word after the central word wt, the right context of wt.

[0212] Σ is to sum all words in the training corpus to calculate the global loss function.

[0213] The Word2Vec model learns word vectors by maximizing the prediction probability of words through context.

[0214] The agenda statement vector function of the Word2Vec model is formula (13):

[0215] vs = Σ wivi / |S| (13)

[0216] vs is the vector representation of the agenda statement.

[0217] wi is the weight coefficient of the ith word in the statement.

[0218] vi is the word vector of the ith word in the statement.

[0219] Σ is to sum the word vectors of all words in the statement.

[0220] |S| is the length of the statement, that is, the number of words.

[0221] The similarity calculation function of the Word2Vec model is formula (14):

[0222] similarity(vs1, vs2) = cos(vs1, vs2) (14)

[0223] similarity(vs1, vs2) is the vector cosine similarity between statements vs1 and vs2.

[0224] vs1 is the statement vector of agenda statement 1.

[0225] vs2 is the statement vector of agenda statement 2.

[0226] cos is the cosine similarity calculation function.

[0227] The Word2Vec model calculates the similarity between statements through the vector cosine value.

[0228] Compared with the methods of the prior art, word vectors can efficiently and automatically obtain semantic features, enabling intelligent agenda management and analysis.

[0229] Through agenda vectorization, it can help enterprises better understand and analyze the content and structure of meetings, identify and compare the similarities and differences between different meetings, so as to better understand the development and change trends of meetings. It can provide objective data support to help make decisions and improve the effectiveness of meetings. It can also promote the knowledge management and information sharing of meetings, making the results of meetings more easily spread and applied.

[0230] 6. Control Automation

[0231] The control automation module realizes the closed-loop automation of enterprise work processes through technologies such as workflow engines, rule engines, monitoring and warning systems, reinforcement learning functions, and knowledge graph functions.

[0232] Among them, the workflow engine defines the workflow nodes and transfer logics of other modules to drive the automatic execution of processes. The rule engine sets state machine rules for each node, calculates and analyzes data in real time according to business rules, and realizes automated business decisions. The monitoring and warning system tracks the work progress and anomalies of other modules, generates warnings and gives feedback. The reinforcement learning function trains agents to adjust the work plans of other modules and give feedback, providing decision-making suggestions and process optimization plans. The knowledge graph function constructs a work knowledge system, stores the knowledge generated during the work processes of other modules, and supports process optimization and decision-making.

[0233] Among them, the task execution dependency relationship of the workflow engine is formula (15):

[0234] Tn+1 ≠ Tn (15)

[0235] Tn+1 depends on Tn.

[0236] Task status update: St+1 = F(St, At).

[0237] Tn is the nth task in the workflow.

[0238] Tn+1 is the (n + 1)-th task in the workflow, that is, the subsequent task.

[0239] ≠ indicates that the two tasks are different tasks.

[0240] The task status update formula of the workflow engine is formula (16):

[0241] St+1 = F(St, At) (16)

[0242] St is the task status at the current time step t.

[0243] St+1 is the task status at the next time step t + 1.

[0244] At is the action performed on the task at the current time step t.

[0245] F is the state transition function, which determines the new state based on the current state and action.

[0246] The rule matching of the rule engine is formula (17):

[0247] IF C1 and C2 THEN A1(17)

[0248] IF is the start symbol of the rule condition.

[0249] C1 is the first condition for rule triggering.

[0250] and is the logical "AND" relationship for multiple conditions.

[0251] C2 is the second condition for rule triggering.

[0252] THEN is the separator symbol between the condition and the action.

[0253] A1 is the action to be performed when the condition is met.

[0254] The rule condition check is formula (18):

[0255] check(C1, data) -> true / false (18)

[0256] check is the function to detect whether the rule condition is triggered.

[0257] C1 is the rule condition to be detected.

[0258] data is the input data used to detect whether the condition is met.

[0259] True / false is the boolean value output by the function. True indicates that the condition is triggered, and false indicates that the condition is not triggered. -> is the function mapping relationship symbol.

[0260] The policy gradient of the reinforcement learning function is given by Equation (19):

[0261] is the gradient of the policy network parameter θ.

[0262] J(θ) is the expected return corresponding to the policy network parameter θ.

[0263] Eπθ is the expectation of taking random samples according to the policy πθ.

[0264] is the gradient of the log probability of the policy πθ with respect to θ.

[0265] πθ(at|st) is the probability that the policy πθ takes the action at in the state st.

[0266] Qπ(st, at) is the value (long-term return) of taking the action at in the state st.

[0267] The state value of the reinforcement learning function is given by Equation (20):

[0268] V(s) = E[Rt|st = s] (20)

[0269] where V(s) is the state value.

[0270] E[] is the expected value.

[0271] Rt| is the immediate reward obtained at time t.

[0272] st is the state at time t.

[0273] The reward expectation of the reinforcement learning function is given by Equation (21):

[0274] Q(s, a) = E[Rt| st = s, at = a] (21)

[0275] where Q(s, a) is the reward expectation,

[0276] E[] is the expected value.

[0277] Rt| is the immediate reward obtained in the state at time t.

[0278] st is the state at time t.

[0279] at is the action taken at time t.

[0280] These above-mentioned technologies of this module operate collaboratively to collect data, analyze and make decisions, optimize processes, and achieve end-to-end closed-loop automation of enterprise work management. It can achieve high efficiency, accuracy, and reliability in the management and control process, improve work efficiency, reduce the risk of human errors, and provide better decision-making basis and suggestions for managers.

[0281] Compared with the methods of the prior art, this module has powerful orchestration capabilities, high integration, supports variable instances, operates efficiently, is flexible in configuration, monitors in real time, and is user-friendly.

[0282] The present invention provides a processing method for an artificial intelligence-based work automation processing platform. The method includes:

[0283] Step 1: Define the workflow nodes and dependencies in the workflow engine of the control automation module to construct a workflow model of the workflow, and set state machine rules for each node in the rule engine of the control automation module;

[0284] Step 2: The workflow engine drives the workflow according to the workflow model to realize the taskification, topicization, templatization, agendaization, and vectorization of the workflow in the work tasking module, task topicization module, topic templatization module, template agendaization module, and agenda vectorization module;

[0285] Step 3: The reinforcement learning function and knowledge graph function of the control automation module collect the data and knowledge of the processing workflow in Step 2 and continuously improve and optimize the decision-making scheme of the platform.

[0286] The present invention proposes an artificial intelligence-based work automation processing platform and its processing method. Taking the entire process of the weekly meeting process automation processing as an example, the platform will be implemented step by step as follows:

[0287] Step 1: The workflow engine of the control automation module defines the weekly meeting process nodes and dependencies, such as "formulating the agenda" depends on "collecting topics", etc., and constructs a weekly meeting workflow model.

[0288] Step 2: The rule engine of the control automation module sets state machine rules for each node, such as triggering "formulating the agenda" when the topic collection is completed, etc.

[0289] Step 3: Input the weekly meeting workflow into the work tasking module, and the workflow engine drives the execution of the weekly meeting tasks according to the workflow model, such as sending notifications, updating node status, etc.

[0290] Step 4: The work tasking module, task topicization module, topic templatization module, template agendaization module, and agenda vectorization module sequentially realize the taskification, topicization, templatization, agendaization, and vectorization of the weekly meeting, generate a draft agenda of the weekly meeting, and vectorize the agenda.

[0291] Step 5: The reinforcement learning function of the control automation module performs reinforcement learning on the data of each weekly meeting, proposes adjustment plans, and improves the meeting effect; the knowledge graph function of the control automation module stores the knowledge of the weekly meeting, continuously improves the knowledge system, and supports the decision-making of the platform.

[0292] In the above process, the control automation module monitors the task progress in real time, sets rule reminders, and generates execution decisions. The above steps, combined with the algorithm formula, realize the closed-loop standardized management of the weekly meeting process and achieve the purpose of automatic control.

[0293] The above embodiments only represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the present invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention patent shall be subject to the appended claims.

Claims

1. An artificial intelligence-based work automation processing platform, characterized in that: The platform includes: a work task module, a task topic module, a topic template module, a template agenda module, an agenda vectorization module and a management and control automation module, wherein: The work taskization module automatically maps the workflow to the task definition by training a deep learning model combined with an attention mechanism, and converts the workflow description into specific task splitting and definition; The task topic module pre-trains the task topic domain through a natural language model, learns the optimal matching relationship between tasks and topics, and reads the task description of the work task module to define and divide topics; The topic template module identifies and matches the topic template of the task topic module through a search method of word vector matching; The template agenda module identifies the agenda of the topic template, matches the template agenda, and performs time allocation and priority identification on the agenda through a matching learning method based on a graph neural network; The agenda vectorization module converts the agenda of the template agenda module into a vector form through word embedding technology; The management and control automation module defines the workflow nodes and flow logic of other modules through the workflow engine, and drives the automatic execution of the processes of other modules.

2. The work automation processing platform based on artificial intelligence according to claim 1 is characterized in that: The work taskization module uses a large number of real workflow descriptions and corresponding task definitions as training data and adopts a seq2seq model for training. The seq2seq model consists of an encoder and a decoder. The encoder uses a bidirectional LSTM to encode the input workflow description. The specific algorithm of the seq2seq model encoder is formula (1); ht = f(xt, ht-1) (1) Among them, ht is the hidden state vector of the encoder at the current time step t, xt is the input of the encoder at the current time step t, and f is the loop function of the encoder.

3. The work automation processing platform based on artificial intelligence according to claim 2 is characterized in that: The decoder of the seq2seq model integrates the attention mechanism and automatically generates the corresponding task splitting and definition. The specific algorithm of the seq2seq model decoder is formula (2), and the specific algorithm of the attention mechanism is formula (3), formula (4) and formula (5); p(yt|y1:t-1,x) = g(yt-1,st,ct) (2) Where yt is the target word that the decoder needs to predict the output at time t, y1:t-1 is the output sequence that the decoder has generated from time 1 to t-1, x is the source language sequence input by the encoder, p(yt|y1:t-1, x) is the predicted probability distribution of the output word yt by the decoder at time t, yt-1 is the word that the decoder has generated at the previous time t-1, st is the hidden state at time t, ct is the context vector, and g is the recursive model of the decoder; etj = a(st-1,hj) (3) αtj = exp(etj) / Σk exp(etk) (4) ct = Σj αtjhj (5) Among them, a is the correlation function, hj is the state of the encoder at each moment, etj is the attention weight score of the encoder state hj at time t, st-1 is the state of the decoder at the current step, hj is the state of the encoder at time step j, a is the scoring function of the attention mechanism, αtj is the normalized attention weight, Σkexp(etk) is the normalization factor of etj, and ct is the dynamic context vector.

4. The work automation processing platform based on artificial intelligence according to claim 1 is characterized in that: The task topicalization module adopts a fine-tuned BERT model, performs large-scale pre-training on the task topic field, and uses real task data and corresponding topic annotations for fine-tuning, thereby obtaining a fine-tuned BERT model specific to task topic generation. The pre-training function of the fine-tuned BERT model is formula (6), the fine-tuning function of the fine-tuned BERT model is formula (7), and the inference function of the fine-tuned BERT model is formula (8); L = Σ E(Ti, T^i) + Σ E(Si, S^i) (6) Where T is a text sequence, S is a text sequence pair, E is a loss function, L is the overall training loss of the BERT model, E is a loss function, Ti is the i-th word in the text sequence paragraph, T^i is the model's prediction of the i-th word in the text sequence paragraph, ΣE(Ti, T^i) is the sum of the loss values ​​of the text sequence paragraph, Si is the i-th word in the text sequence pair, S^i is the model's prediction of the i-th word in the text sequence pair, and ΣE(Si, S^i) is the sum of the loss values ​​of the text sequence pair; L = Σ E(yi, y^i) (7) Where y is the topic label, y^ is the topic predicted by the BERT model, L is the loss function when BERT is fine-tuned, E is the loss function, yi is the true target value labeled in the sample, y^i is the target value predicted by the model, and Σ is the sum of the loss function over all samples; topic = argmax P(topic|t) (8) Among them, topic is the most likely topic predicted by the model, argmax is to find the value that maximizes P(topic|t), P(topic|t) is the predicted probability of each topic under the condition of input sentence t calculated by the model, and t is the input sentence.

5. The work automation processing platform based on artificial intelligence according to claim 1 is characterized in that: The topic template module constructs a topic word vector space, represents the topic and the topic template as word vectors respectively, and performs similarity matching between the topic vector and the topic template vector to achieve topic identification and matching of the topic template.

6. The artificial intelligence-based work automation processing platform according to claim 1, characterized in that: The template agenda module is based on a matching learning method of a graph neural network, constructs the agenda of the topic template and its time and priority data into a knowledge graph, uses a graph neural network to perform vector representation learning on the knowledge graph and the template agenda of the template agenda module, and performs similarity matching between the knowledge graph and the template agenda to generate the matching time allocation and priority arrangement of the agenda; the encoding function of the graph neural network is formula (9), the similarity function of the graph neural network is formula (10), and the inference function of the graph neural network is formula (11); hv = σ(Σu∈N(v) Wu, v hu) (9) Where hv is the feature representation of node v, σ is a nonlinear activation function, N(v) is the set of neighbor nodes of node v, u is a neighbor node of node v, Wu,v is the weight matrix from node u to node v, hu is the feature representation of node u, and Σ is the summation operation of all neighbor nodes of node v; S(g1, g2) = f(h1, h2) (10) Among them, S(g1, g2) is the similarity score between graph g1 and graph g2, g1 is the input source graph, g2 is the template graph to be matched, h1 is the graph neural network encoding vector of the source graph g1, h2 is the graph neural network encoding vector of the template graph g2, f is the similarity calculation function, h1 and h2 are the encoding representations of the two agenda graphs respectively; gt = argmax S(g, gi) (11) gt is the final matching template graph result, argmax is gi that maximizes S(g, gi), S(g, gi) is the matching similarity score between the source graph g and the i-th template graph gi, g is the input source graph, such as the new agenda graph, and gi is the i-th template graph.

7. The work automation processing platform based on artificial intelligence according to claim 1, characterized in that: The agenda vectorization module uses a large amount of conference agenda data to train the Word2Vec model, and obtains the word vectors of the agenda text of the template agenda module, performs weighted averaging on the word vectors to obtain the vector representation of the agenda sentence, calculates the similarity of different agenda sentences through the cosine similarity of the vector representation, and uses clustering and topic modeling methods to analyze the content characteristics of the agenda; The word vector training function of the Word2Vec model is formula (12), the agenda sentence vector function of the Word2Vec model is formula (13), and the similarity calculation function of the Word2Vec model is formula (14); L = Σ log p(wt|wt-k, ..., wt+k) (12) Where L is the loss function of word vector training, log p(wt|wt-k, ..., wt+k) is the logarithmic predicted probability of the current word wt based on its context word wt-k to wt+k, wt is the current word, that is, the central word to be predicted, wt-k is the kth word before the central word wt, wt+k is the kth word after the central word wt, and Σ is the sum of all words in the training corpus; vs = Σ wivi / |S| (13) Where vs is the vector representation of the agenda sentence, wi is the weight coefficient of the i-th word in the sentence, vi is the word vector of the i-th word in the sentence, Σ is the sum of the word vectors of all words in the sentence, and |S| is the length of the sentence; similarity(vs1, vs2) = cos(vs1, vs2) (14) Among them, similarity(vs1, vs2) is the vector cosine similarity of sentences vs1 and vs2, vs1 is the sentence vector of agenda sentence 1, vs2 is the sentence vector of agenda sentence 2, and cos is the cosine similarity calculation function.

8. The artificial intelligence-based work automation processing platform according to claim 1, characterized in that: The task execution dependency of the workflow engine of the management and control automation module is formula (15), and the task status update formula of the workflow engine is formula (16); Tn+1 ≠ Tn (15) Among them, Tn+1 depends on Tn, Tn is the nth task in the workflow, Tn+1 is the n+1th task in the workflow, and ≠ means that the two tasks are different tasks; St+1 = F(St, At) (16) Among them, St is the task state at the current time step t, St+1 is the task state at the next time step t+1, At is the action performed on the task at the current time step t, and F is the state transition function.

9. The artificial intelligence-based work automation processing platform according to claim 1, characterized in that: The management and control automation module is also equipped with a rule engine, a monitoring and early warning system, a knowledge graph function and a reinforcement learning function. The rule engine calculates and analyzes data in real time according to business rules to achieve automated decision-making for other modules; the monitoring and early warning system tracks the work progress and anomalies of other modules, generates early warnings and provides feedback; The reinforcement learning function trains the intelligent agent through the process data of other modules, and provides decision suggestions and process optimization solutions for other modules; The knowledge graph function stores the knowledge of the workflow, builds a work knowledge system, and supports process optimization and decision-making of other modules; The rule matching of the rule engine is formula (17), the rule condition check is formula (18); the policy gradient of the reinforcement learning function is formula (19); the state value of the reinforcement learning function is formula (20); and the reward expectation of the reinforcement learning function is formula (21); IF C1 and C2 THEN A1(17) Among them, IF is the start symbol of the rule condition, C1 is the first condition for the rule to be triggered, and is the logical "and" relationship of multiple conditions, C2 is the second condition for the rule to be triggered, THEN is the separator symbol between the condition and the action, and A1 is the action to be executed when the condition is met; check(C1, data) -> true / false (18) Among them, check is a function that detects whether the rule condition is triggered, C1 is the rule condition to be detected, data is the input data used to detect whether the condition is met, true / false is the Boolean value output by the function, true means the condition is triggered, false means the condition is not triggered; in, is the gradient of the policy network parameter θ, J(θ) is the expected return corresponding to the policy network parameter θ, Eπθ is the expectation of taking a random sample according to the policy πθ, is the gradient of the log probability of policy πθ with respect to θ, πθ(at|st) is the probability of policy πθ taking action at in state st, and Qπ(st, at) is the value of taking action at in state st; V(s) = E[Rt|st = s] (20) Where V(s) is the state value, E[] is the expected value, Rt| is the immediate reward obtained at time t, and st is the state at time t; Q(s, a) = E[Rt| st = s, at = a] (21) Among them, Q(s, a) is the reward expectation, E[] is the expected value, Rt| is the immediate reward obtained at the state at time t, st is the state at time t, and at is the action taken at time t.

10. A processing method for a work automation processing platform based on artificial intelligence, characterized in that: The method comprises: Step 1: define workflow nodes and dependencies in the workflow engine of the control automation module, build a workflow model of the workflow, and set state machine rules for each node in the rule engine of the control automation module; Step 2: The workflow engine drives the workflow according to the workflow model to sequentially implement the taskization, topicization, templateization, agendaization and vectorization of the workflow in the work taskization module, task topicization module, topic templateization module, template agendaization module and agenda vectorization module; Step 3: The management and control automation module strengthens the learning function and knowledge graph function by collecting data and knowledge of the processing workflow in step 2, and continuously improves and optimizes the decision-making plan of the platform.

Images