Limited domain event reasoning method based on mixed atlas

Abstract

The invention relates to a defined domain event reasoning method based on a mixed graph, and belongs to the field of artificial intelligence, big data, knowledge graphs and natural language processing. In order to solve the problem that a traditional data processing method cannot effectively assist a decision maker to make a fast decision, event knowledge and logic relations contained in the multi-modal data characteristics of multiple sources are extracted, the multi-modal data are represented in a unified mode through a multi-modal networked synthesis representation method, and therefore the multi-modal data characteristics of multiple sources are obtained. Multi-modal data fusion is realized, and a mixed knowledge graph network capable of dynamically storing event nodes and relationships is constructed; and on the basis of the mixed knowledge graph, predicting different future development trends by adopting a trend-simulation reasoning prediction method. According to the method, the event development trend can be predicted and reasoned, and a decision maker is assisted to solve problems and formulate coping strategies through deduction and research.

Description

Technical Field

[0001] The present invention belongs to the fields of artificial intelligence, big data, knowledge graphs, and natural language processing, and specifically relates to a limited domain event reasoning method based on a hybrid graph. Background Art

[0002] In certain technical fields, data processing and analysis have become the key to decision-making and deployment. As the environment changes rapidly, the sources of information are wide and huge, and the update speed is extremely fast. Decision makers must be able to respond quickly to major issues. This requires decision makers to master a large amount of effective information in a short period of time to support timely and accurate decision-making. As an important source of information, intelligence data, traditional data processing methods usually rely on manual rapid sorting and integration in the brain, and construct a description of the situation based on time and causality. However, this large amount of unstructured text data that relies on manual processing often cannot guarantee the accuracy of processing, and cannot effectively assist decision makers in making quick decisions. With the continuous development of deep learning and natural language processing technology, data processing will gradually shift from the traditional information mode to the intelligent mode.

[0003] In order to meet the needs of fast and accurate analysis and processing, there is an urgent need for an intelligent technology based on event extraction that can automatically process and analyze intelligence data, mine key events and their correlations, and present the multi-dimensional characteristics of data through visualization, thereby assisting in improving action efficiency and decision-making quality. The present invention proposes a limited domain event reasoning method based on a hybrid graph, and uses advanced deep learning and natural language processing methods to build an efficient intelligent data processing framework. This technology can perform event reasoning in a specific field through a hybrid graph to achieve efficient analysis of complex intelligence data and intelligent decision support. Summary of the invention

[0004] 1. Technical issues to be resolved The technical problem to be solved by the present invention is how to provide a limited domain event reasoning method based on a hybrid graph to solve the problem that traditional data processing methods cannot effectively assist decision makers in making quick decisions.

[0005] (II) Technical solution In order to solve the above technical problems, the present invention proposes a limited domain event reasoning method based on a hybrid graph, which comprises the following steps: S1. Based on the characteristics of multimodal data from various sources, the event knowledge and logical relationships contained therein are extracted, and a multimodal network synthesis representation method is used to achieve a unified representation of multimodal data, realize multimodal data fusion, and construct a hybrid knowledge graph network that can dynamically store event nodes and relationships, including: event graph and time series feature graph; S2. On the basis of the time series feature map, a time series event map is constructed by combining event information and key element information; on the basis of the reason map and the time series event map, the trend-simulation reasoning prediction method is used to predict different future development trends, and finally the comprehensive monitoring and visual analysis of multimodal data from multiple sources is realized to create a deduction and simulation platform for auxiliary decision-making.

[0006] (III) Beneficial effects This paper proposes a limited domain event reasoning method based on hybrid graph, the main advantages of which are reflected in the following aspects: 1) A collection and management method of intelligence data centered on events is proposed: information extraction of multi-source heterogeneous data is carried out with "events" as the core, and text, images, relational data, situation and other knowledge are integrated with specific scenarios to accurately and vividly express dynamic knowledge in a visual way. This kind of systematic knowledge is more in line with human cognitive habits and helps relevant personnel quickly learn key knowledge and their interrelationships from massive data.

[0007] 2) A method for event deduction and prediction based on a hybrid knowledge graph is proposed. With the support of multi-source intelligence fusion results, a comprehensive analysis of the regional situation is conducted, implicit information is mined, the situation map is dynamically updated, and the development trend of events is predicted and inferred. This helps decision makers solve problems and formulate response strategies through deduction research. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 This is a flow chart of the limited domain event reasoning method based on the hybrid graph of the present invention; Figure 2 A schematic diagram of event logic extraction for the generative adversarial network model of the present invention; Figure 3 It is a schematic diagram of event reasoning that combines the event graph and the time sequence event graph. DETAILED DESCRIPTION

[0009] In order to make the purpose, content and advantages of the present invention more clear, the specific implementation methods of the present invention are further described in detail below in conjunction with the drawings and examples.

[0010] This paper proposes a limited domain event reasoning method based on hybrid graph. Considering the complexity of intelligence data in practical applications, this task mainly faces the following two challenges: First of all, faced with massive multi-source heterogeneous intelligence data, traditional data organization methods can no longer support users' requirements for simple and fast access to information services. There is an urgent need to build a complete event-oriented hybrid knowledge graph to uniformly model heterogeneous intelligence data.

[0011] Secondly, at present, the utilization of big data is still insufficient. Small-scale data information is not enough to assist decision-makers in comprehensively and quickly understanding the situation and making correct decisions on the situation.

[0012] It is possible to mine and organize the situation context using a hybrid knowledge graph, and combine event-driven artificial intelligence analysis technology to mine the event correlation relationships of specific topics, and thus assist in designing joint simulation deduction scenarios. The present invention can solve the above two technical problems.

[0013] Such as Figure 1 shown is a flowchart of a method for reasoning about events in a limited domain based on a hybrid graph. The method includes the following steps: S1. For the multi-modal data characteristics from multiple sources, the multiple sources include: specific domain databases, documents, intelligence texts, open-source data, streaming media, etc., extract the event knowledge and logical relationships contained therein, and use a multi-modal networked synthesis representation method to achieve unified representation of multi-modal data, realize multi-modal data fusion, and construct a hybrid knowledge graph network that can dynamically store event nodes and relationships, including: an event logic graph and a time-series feature graph; S2. Based on the time-series feature graph, combine event information and key element information to construct a time-series event graph; on the basis of the event logic graph and the time-series event graph, use the trend-simulation reasoning prediction method to predict different future development trends, and finally realize comprehensive monitoring and visualization analysis of multi-modal data from multiple sources, and build a deduction and simulation platform for assisting decision-making.

[0014] This process is mainly divided into two steps: automatic construction of a hybrid graph and evolution inference centered on events.

[0015] The process of constructing a hybrid graph centered on events is mainly divided into four parts: data collection and cleaning, event logic extraction, multi-modal data fusion, and event rule learning.

[0016] S11. Data collection and cleaning, collect multi-modal data from multiple sources, and perform cleaning and word segmentation on the text; Data collection is a basic link in the information processing chain. In addition to connecting data from multiple sources such as open-source data and closed-source data in the traditional way, more attention is paid to the collection of high-concurrency real-time data, and the data scale, timeliness, and authenticity are evaluated. Clearly define the overall goal, task scenario, and research object of information processing, list the requirements list, and formulate an evaluation plan and index system; the cleaning process of the data set includes: removing irrelevant content such as stop words and symbols through a customized data cleaning model. Because the original document text data contains punctuation, stop words, etc., before performing text processing tasks, it is necessary to clean and segment the text, and convert the text to be processed into the data input format required for further analysis.

[0017] S12. Event logic extraction. The Wasserstein generative adversarial network model is used to process multimodal data, automatically extracting thematic events, attributes, and logical relationships (including sequence, transition, causality, progression, association, explanation, etc.) to mine knowledge information about events in the real world.

[0018] This technology uses a bidirectional LSTM structure to build a generator and a discriminator. The generator includes a basic model and an enhanced model for event and logic extraction. In addition to inputting the collected labeled data, it introduces external resources for interpretation and supplementation to generate the label probability distribution of events and relationships. The discriminator scores the generation quality of the generator to provide feedback to the generator, so that the generator can update the gradient according to the feedback information to further improve the quality of the label probability. Figure 2 shown.

[0019] S13, multimodal data fusion. According to the event logic extraction results, construct the event nodes and their network representation, use the relationship between the event nodes to fuse the multimodal data, and obtain a unified representation network for multimodality. The present invention uses a multimodal network synthesis representation method to uniformly represent and organize data from different sources, different means, and different structures, thereby achieving multimodal data fusion.

[0020] First, based on the results of S12 event logic extraction, we construct event nodes, which refer to events, facts and other related relationships in the event graph. We use DeepWalk to treat event nodes as basic elements, and use truncated random walk sequences to represent the neighbor sequence of an event node. In this way, through multiple truncated random walks, we get a data set containing multiple sequences, each of which reflects the structural information of an event node and its neighbors. Then, we apply word embedding models such as Word2Vec to the obtained sequences to obtain the network representation of event nodes.

[0021] When faced with a logical node composed of multimodal data, a feature representation can be constructed for each modality first. For example, text data is output as a vector representation through the TF-IDF method, and image data is processed through several layers of convolutional networks and then obtained as a vector representation through a fully connected layer. The embedded representations of different modalities are then mapped to the embedding space of the same dimension. Each iteration maximizes the similarity between the embedded representations of linked nodes and minimizes the similarity of unlinked nodes, ultimately achieving the fusion of multimodal data and obtaining a unified representation network for multimodality.

[0022] S14, learning of rules of events. First, through the learning of rules of events, we face the multimodal unified representation network, and through certain cognitive reasoning methods, we fully explore the logical relationship between event nodes, so that the multimodal unified representation network can reason in a human cognitive way and complete the construction of the event graph. Then, we complete the semantic, temporal and spatial associations between event nodes, and construct a time series feature graph based on time information and event sequence. In the time series feature graph, nodes represent different states or stages of events, and edges represent the temporal relationship between events. The construction of the time series feature graph can improve the temporal development context and temporal dependency of events.

[0023] On the basis of the time series feature map, the time series event map is constructed by combining key elements such as event information and environmental information, deployment and capability information, and dynamic targets. On the basis of the reason map and the time series event map, the trend-simulation reasoning prediction method is used to predict different future development trends. Specifically, a model is established based on objective entities and events, and the fitting results of historical cases are compared with the actual results. The simulation deduction model is continuously revised and improved to predict different future development trends.

[0024] Traditional event reasoning methods mainly use script event prediction methods, which construct context information into multiple structured events and then select from a list of candidate events to achieve reasoning about subsequent possible events. However, the script event prediction process requires manual preparation of a complete event script, and cannot predict event types that do not appear in the script.

[0025] In order to overcome the limitations of traditional reasoning methods, the present invention combines event reasoning and prediction with event graphs and time series event graphs. Figure 3 shown.

[0026] During the application process, the causal graph is first used for reasoning: for new events that occur, the similarity between them and each node in the causal graph is calculated to find the most similar node to the new event; then, based on the generalization node in the causal graph of the most similar node to the new event, the subsequent evolution direction of the event based on the causal graph is deduced.

[0027] After completing the reasoning based on the event graph, the event prediction technology based on the cyclic event network is used according to the time series event graph to abstract the task of predicting future events into the reasoning task of the future state of the time series event graph. By cyclically encoding the information of all historical events in the time series event graph and aggregating the information of all temporally adjacent events in the time series event graph, the joint probability distribution of all events is inferred by aggregating various information such as historical events, temporally adjacent events, and simultaneous events, so as to predict the probability of future events.

[0028] Finally, for the event reasoning results based on the event graph and the time series event graph, the final event reasoning results are obtained by weighted average probability fusion.

[0029] Based on the above-mentioned event reasoning intelligent algorithm, through the comprehensive use of deployment and capability information, dynamic targets, environmental information and other factors, it is possible to realize dynamic interactive simulation and deduction of the events that have occurred and intuitive, multi-angle assessment of the situation, determine the enemy's personnel structure, deployment, direction of action and route, identify the enemy's style, infer the enemy's intentions, make a reasonable explanation of the current situation, and predict situation changes in the future. Ultimately, a comprehensive situation map formed by organizations such as activities, events, time, location and personnel elements is established, thereby assisting decision makers in early deployment and improving the level of confrontation.

[0030] The present invention discloses a limited domain event reasoning method based on a hybrid graph, the main advantages of which are reflected in the following aspects: 1) A collection and management method of intelligence data centered on events is proposed: information extraction of multi-source heterogeneous data is carried out with "events" as the core, and text, images, relational data, situation and other knowledge are integrated with specific scenarios to accurately and vividly express dynamic knowledge in a visual way. This kind of systematic knowledge is more in line with human cognitive habits and helps relevant personnel quickly learn key knowledge and their interrelationships from massive data.

[0031] 2) A method for event deduction and prediction based on a hybrid knowledge graph is proposed. With the support of multi-source intelligence fusion results, a comprehensive analysis of the regional situation is conducted, implicit information is mined, the situation map is dynamically updated, and the development trend of events is predicted and inferred. This helps decision makers solve problems and formulate response strategies through deduction research.

[0032] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for reasoning about limited-domain events based on a hybrid graph, characterized in that The method comprises the following steps: S1. Based on the characteristics of multimodal data from various sources, the event knowledge and logical relationships contained therein are extracted, and a multimodal network synthesis representation method is used to achieve a unified representation of multimodal data, realize multimodal data fusion, and construct a hybrid knowledge graph network that can dynamically store event nodes and relationships, including: event graph and time series feature graph; S2. On the basis of the time series feature map, a time series event map is constructed by combining event information and key element information; on the basis of the reason map and the time series event map, the trend-simulation reasoning prediction method is used to predict different future development trends, and finally the comprehensive monitoring and visual analysis of multimodal data from multiple sources is realized to create a deduction and simulation platform for auxiliary decision-making.

2. The method for reasoning about limited-domain events based on a hybrid map according to claim 1, wherein The S1 includes: S11. Data collection and cleaning: Collect multimodal data from various sources, clean the text, and perform word segmentation; S12, event logic extraction: Use the Wasserstein generative adversarial network model to process multimodal data, automatically extract thematic events, attributes and logical relationships, and mine the knowledge information of the real world; S13, multimodal data fusion: According to the event logic extraction results, construct the event nodes and their network representation, use the relationship between the event nodes to fuse the multimodal data, and obtain a unified representation network for multimodality; S14. Learning of rules of events: First, through learning of rules of events, we face the multimodal unified representation network and use cognitive reasoning methods to fully explore the logical relationship between event nodes, so that the multimodal unified representation network can reason in a human cognitive way and complete the construction of the event graph; then, we complete the semantic and spatiotemporal associations between event nodes and construct a temporal feature graph based on time information and event sequence. In the temporal feature graph, nodes represent different states or stages of events, and edges represent the temporal relationship between events. The construction of the temporal feature graph improves the temporal development context and temporal dependency of events.

3. The method for reasoning about limited-domain events based on a hybrid map according to claim 2, wherein The S11 includes: data collection is a basic link in the information processing chain. In addition to importing data from open source data and closed source data in the traditional way, it pays more attention to the collection of high-concurrency real-time data, and evaluates the data scale, timeliness, and authenticity; clarifies the overall goals, task scenarios, and research objects of information processing, lists a list of requirements, and formulates an evaluation plan and indicator system; before performing text processing tasks, cleans and segmentes the text, and converts the text to be processed into the required data input format.

4. The method for reasoning about limited-domain events based on a hybrid map according to claim 2, wherein The S12 includes: using a bidirectional LSTM structure to build a generator and a discriminator, the generator includes an enhanced model for event and logic extraction, in addition to inputting the collected labeled data, introducing external resources for interpretation and supplement, generating label probability distribution of events and relationships, and the discriminator scores the generation quality of the generator to provide feedback to the generator, so that the generator can update the gradient according to the feedback information to improve the quality of the label probability.

5. The method for reasoning about limited-domain events based on a hybrid graph as claimed in claim 4, wherein Logical relationships include: sequence, transition, cause and effect, progression, association and explanation.

6. The method for reasoning about limited-domain events based on a hybrid graph as claimed in claim 4, wherein, The S13 comprises: constructing a logic node according to the event logic extraction result of S12, where the logic node refers to an event, fact and other related relationships in the logic graph, using DeepWalk to regard the logic node as an element, and using a truncated random walk sequence to represent a neighbor sequence of a logic node, and obtaining a data set containing multiple sequences through multiple truncated random walks, each sequence reflects the structural information of a logic node and its neighbors; then applying the Word2Vec word embedding model to the obtained sequence, thereby obtaining a network representation of the logic node; When faced with a logical node composed of multimodal data, we first construct a feature representation for each modality, and then map the embedded representations of different modalities to the embedding space of the same dimension. Each iteration maximizes the similarity between the embedded representations of linked nodes and minimizes the similarity of unlinked nodes, ultimately achieving the fusion of multimodal data and obtaining a unified representation network for multimodality.

7. The method for reasoning about limited-domain events based on a hybrid map according to claim 6, wherein The text data is output as a vector representation using the TF-IDF method, and the image data is processed through several layers of convolutional networks and then through a fully connected layer to obtain a vector representation.

8. The method for reasoning about limited-domain events based on a hybrid map according to any one of claims 2-7, characterized in that The S2 includes: on the basis of the time series feature map, by combining event information and key elements, the key elements include: environmental information, deployment and capability information and dynamic goals; constructing a time series event map; on the basis of the reason map and the time series event map, using the trend-simulation reasoning prediction method to predict different future development trends. Specifically, a model is established according to objective entities and events, and the fitting results of historical cases are compared with the actual results, and the simulation deduction model is continuously revised and improved to predict different future development trends.

9. The method for reasoning about limited-domain events based on a hybrid graph as claimed in claim 8, wherein, Apply the event graph to reason: for a new event, calculate its similarity with each node in the event graph and find the most similar node to the new event; Then, according to the generalized node of the most similar node of the new event in the causal graph, the subsequent evolution direction of the event based on the causal graph is deduced.

10. The method for reasoning about limited-domain events based on a hybrid graph as claimed in claim 9, wherein, After completing the reasoning based on the event graph, according to the time series event graph, the event prediction technology based on the recurrent event network is used to abstract the task of predicting future events into the reasoning task of the future state of the time series event graph; By cyclically encoding the information of all historical events in the time series event graph and aggregating the information of all temporally adjacent events in the time series event graph, the information of historical events, temporally adjacent events and simultaneous events is aggregated to infer the joint probability distribution of all events, thereby predicting the probability of future events. Finally, for the event reasoning results based on the event graph and the time series event graph, the final event reasoning result is obtained by probability fusion using weighted average.

Images