An intelligence situation mining method based on event extraction

Abstract

The present invention relates to an intelligence situation mining method based on event extraction, and belongs to the field of information extraction / situation awareness. The present invention extracts intelligence-related vocabulary from existing intelligence situation text data, and classifies the intelligence situation text data into an event scenario type library based on the category and frequency of the domain words. Design event templates for different event scenario types, extract event elements from intelligence situation text data based on the event templates, and form an event list. Retrieve intelligence situation sequence data based on event elements, associate and match similar attributes, and form an intelligence situation library. The present invention solves the problem of mining intelligence situations in a large number of intelligence texts and situation sequences.

Description

Technical Field

[0001] The present invention belongs to the field of information extraction / situation awareness, and in particular relates to an intelligence situation mining method based on event extraction. Background Art

[0002] In view of the prominent problems faced by event analysis, such as low timeliness, more qualitative analysis, and less situation analysis, only by achieving deep cognition and intelligent processing of the situation can we quickly obtain important information from a large amount of situation data, firmly grasp the "intelligence information advantage", and promote the development of event analysis technology based on intelligence situation text and sequence data, improve the event handling model, and provide support services for command decision-making. Therefore, it is urgent to develop an intelligence situation mining method and analysis system based on event extraction. The mining of intelligence situation requires judging or selecting the text or situation sequence generated by various information sources, discovering new clues from existing information, combining the information of the activities, and using analytical thinking to perceive and characterize the intelligence theme behind the information. At present, these important intelligence links need to be realized with cumbersome manual work. In addition, the existing methods are generally complex and difficult to cope with the massive intelligence situation information on the Internet.

[0003] In view of the above-mentioned deficiencies, the present invention proposes an intelligence situation mining method and analysis system based on event extraction, establishes an analysis model for intelligence situation text data and sequence data, and constructs an intelligence situation mining and analysis system in combination with event extraction technology, which transforms the data organization mode from the surface features of information to the semantic level, realizes efficient and accurate correlation mining and fusion of intelligence text and situation sequences, improves the depth and efficiency of intelligence situation mining, and provides support for the review and prediction of future intelligence events. Summary of the invention

[0004] 1. Technical issues to be resolved

[0005] The technical problem to be solved by the present invention is how to provide an intelligence situation mining method based on event extraction to solve the problem that important intelligence links need to rely on cumbersome manual implementation, and the existing methods are generally highly complex and difficult to cope with the massive intelligence situation information on the Internet.

[0006] (II) Technical solution

[0007] In order to solve the above technical problems, the present invention proposes an intelligence situation mining method based on event extraction, which comprises the following steps:

[0008] S1. Event scenario classification: Extract intelligence-related words from the existing intelligence situation text data, and classify the intelligence situation text data into event scenario type libraries based on the categories and frequencies of domain words;

[0009] S2. Event element extraction: Design event templates for different event scenario types, extract event elements from intelligence situation text data based on the event templates, and form an event list;

[0010] S3. Intelligence situation correlation and fusion: Retrieve intelligence situation sequence data based on event elements, correlate and match similar attributes, and form an intelligence situation database.

[0011] (III) Beneficial effects

[0012] The present invention proposes an intelligence situation mining method based on event extraction, the main advantages of which are reflected in the following aspects:

[0013] (1) A method for event scene classification is designed. For intelligence situation text data, domain words are extracted from it. Based on the different contributions of words and domain words in sentences to event scene classification, a text classifier is constructed and the text classification results are output. This solves the problem that traditional event classification methods cannot be applied to text data without trigger words.

[0014] (2) A method for event element extraction is proposed. Event templates are constructed according to different event types. The basic information and specific information to be extracted are formulated. Then, an end-to-end encoder-decoder framework is constructed to solve the problem of difficulty in extracting location elements and time elements in document-level multi-event intelligence situation text.

[0015] (3) A method for intelligence situation association and fusion is proposed. Based on the common elements in the event list and the intelligence situation sequence attributes, a feature template is constructed to achieve similarity association at the semantic level. The dynamic information in the situation sequence is analyzed based on the association results and finally integrated into the structured event description. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is the overall framework of the present invention;

[0017] Figure 2 The event list and situation association matching framework of the present invention;

[0018] Figure 3 This is a schematic diagram of an example of the intelligence situation fusion process of the present invention. DETAILED DESCRIPTION

[0019] 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.

[0020] The present invention belongs to the technical field of information extraction / situation awareness, and in particular relates to an intelligence situation mining method and analysis system based on event extraction.

[0021] The present invention discloses an intelligence situation mining method and analysis system based on event extraction. Through the existing intelligence situation text data, relevant words in the intelligence field are extracted, and the intelligence situation text data is classified into an event scene type library according to the category and frequency of the field words. For different event scene types, event templates are designed, and event elements in the intelligence situation text data are extracted according to the event templates to form an event list. According to the event elements, the intelligence situation sequence data is retrieved, and similar attributes are associated and matched to form an intelligence situation library. It mainly solves the mining of intelligence situations in a large number of intelligence texts and situation sequences, and specifically includes:

[0022] (1) It is necessary to solve the problem of extracting domain event words from intelligence texts and building text classification models, and classify the intelligence texts to be processed according to different event scenarios;

[0023] (2) It is necessary to solve the problem of extracting event elements from intelligence texts, build event templates based on event scene classification, and extract event elements from the text.

[0024] (3) It is necessary to solve the problem that intelligence situation sequence data is difficult to mine and utilize, support rule matching and similarity analysis based on sequence attributes and event elements, and meet the comprehensive analysis of intelligence situation.

[0025] The intelligence situation mining method based on event extraction of the present invention comprises the following steps:

[0026] S1. Event scenario classification: Extract intelligence-related words from the existing intelligence situation text data, and classify the intelligence situation text data into event scenario type libraries based on the categories and frequencies of domain words.

[0027] The current popular event scene classification is mainly based on the method of trigger word detection. This method maps the trigger words and event scene types in advance. For example, the trigger words of the reconnaissance event are "reconnaissance", "probe", "interception", etc. The trigger words directly map the event scene category, ignoring the influence of other related field words in the text on the event type. In addition, the annotation of trigger words requires strong field knowledge, and there are problems such as difficulty in formulating annotation standards and high annotation costs. The event scene classification method of the present invention starts from event instances and uses the similarity of similar events in element composition and sentence structure to classify event scenes.

[0028] S11. Data preprocessing

[0029] For intelligence situation text data, we first use regular expressions to perform data cleaning operations such as character replacement, such as converting English characters to lowercase and replacing dates, longitude and latitude, etc. with a standard unified format, and then use word segmentation tools to perform word segmentation processing.

[0030] S12. Domain word extraction

[0031] Domain words include the subject and domain information of the event. For example, reconnaissance events include reconnaissance platforms, reconnaissance targets, and reconnaissance methods. Domain words can be used to infer the event type described in the original text. When extracting domain words, named entity recognition is used to extract domain words from the original text data, including weapon equipment models, mission types, combat units, etc.

[0032] S13. Construction of event scenario type system

[0033] Based on the characteristics of the subject, environment, mission objectives, etc. in the intelligence situation text, an event scenario type system is manually formulated.

[0034] S14. Text Classification

[0035] Text classification completes the event scene classification of intelligence situation by encoding words and domain words, encoding event scene types, and classifying scene types for intelligence situation texts labeled with domain words.

[0036] S141, word and domain word encoding

[0037] Using the word segmentation results and domain word extraction results of intelligence situation text data, the language model Skip-gram model is used to initialize word encoding d w and domain word encoding d e , learn word vectors from the intelligence situation corpus, and generate a vector for each word, which contains the word and the type of the word. For example, words and domain words are two different words. Domain words are words that contain the subject and domain information of the event. Words are words other than domain words. O represents non-domain words, and the word is encoded as [word / O]. For PER, which represents the type of person, the domain word is encoded as [name / PER].

[0038] S142, event scene type code

[0039] The event scenario type encoding of the intelligence situation text is represented by two randomly initialized encoding vectors t1 and t2, where t1 is used to capture local information and t2 is used to capture global information.

[0040] S143, scene type classification

[0041] The long short-term memory network LSTM is used to process the word and domain word encoding of the intelligence situation text, and h i It represents the result after each layer of LSTM, and uses the attention mechanism to evaluate the impact of each word in the intelligence situation text on the judgment of the event scene type.

[0042]

[0043] where α kis the attention score of the k-th word.

[0044] It can be obtained that the intelligence situation text is represented by S att =α T H, where α=[α1,…,α n ] is the attention vector of the word, which indicates the importance of the word in the sentence. Here, the word includes: single words and domain words, H = [h1,…,h n ] represents the output of each layer of LSTM.

[0045] S144, use Represents the local features of the text of the domain word, using Represents the global features of the text, including the features of words and domain words, h n It is the output of the nth layer LSTM. The weighted sum of the two is calculated by the Sigmoid function to obtain the event scene type.

[0046] o=σ(λ·v att +(1-λ)·v global )

[0047] Where σ represents the Sigmoid function, and λ is a hyperparameter used to balance the local and global features of the text;

[0048] S145. Through continuous training, adjust t1 and t2 so that the event scene type conforms to the event scene type system established in S13.

[0049] S2. Event element extraction: Design event templates for different event scenario types, extract event elements from intelligence situation text data based on event templates, and form an event list

[0050] The purpose of event element extraction is to extract the basic elements that can define the event scenario from the unstructured intelligence situation text, and include the attribute information in the intelligence situation sequence as much as possible, and finally present it in the form of a structured event list, where the event list includes the subject, time, location, event type, etc.

[0051] S21. Event template construction

[0052] Domain experts build event templates for S1 to obtain different event scenario types based on the needs of intelligence situation mining. Event templates include basic information and specific information. Basic information includes event name, event type, personnel, location, start time, end time, etc. Specific information includes attribute information (equipment type, equipment model, etc.) and dynamic information (speed, heading, etc.) of the specific event subject.

[0053] S22. Event Extraction

[0054] Event extraction uses an end-to-end encoder-decoder framework, which includes a document-level encoder, a role decoder, and an event decoder. The document-level encoder uses the Transformer framework to interactively model all sentences and domain words contained in the sentences. The role decoder and event decoder are used to extract the elements of multiple events in parallel in a document.

[0055] The location elements in intelligence situation texts are complex. The description of a geographical location usually contains one or more place name coordinates, and there is a nested relationship between place names. Extracting them individually will cause semantic loss, and a deep model is needed to learn the context for judgment.

[0056] The time element is also complex. The time element in intelligence situation data has a start and an end. The time indicator needs to determine the start time and the end time based on the context. It is difficult to express the exact meaning by only extracting the time indicator.

[0057] In addition, the domain words and event trigger words in the intelligence text are scattered in different sentences. In order to obtain more complete event information, it is necessary to use a multi-granularity event extraction model to perform multi-granularity document-level relationship extraction, and extract different descriptions of multiple places and times in the document, such as event start time, event end time, subject arrival time, subject departure time, etc., to achieve accurate description of the event.

[0058] To solve the above problems, the present invention uses document-level relation extraction technology to dynamically adjust the word semantic vector according to the context content, and adopts a strategy of parallel annotation of place names. When extracting place name arguments, the place names are annotated in parallel and then integrated in the order of appearance, that is, point = (point1, point2, pointn3, ..., point n ), solve the problem that multiple place names in intelligence situation text data represent a complete semantics and the number is not fixed; when extracting time arguments, the time indicator is extracted independently according to the semantics in the context to extract the start time and the completion time, and the time modifiers (for example, [start xx hours and xx minutes], [arrive at xx hours and xx minutes]) are extracted together. Finally, using global information (global information refers to multiple intelligence situation texts, which may contain a description of an event and are finally integrated into one event), based on single-document and cross-document multi-granularity event extraction models, event elements in text data are extracted according to event templates, and event element information between documents is sorted to form an event list, realizing the integration of descriptions of the same event in multiple intelligence situation text data.

[0059] S3. Intelligence situation correlation and fusion: Retrieve intelligence situation sequence data based on event elements, correlate and match similar attributes, and form an intelligence situation database.

[0060] Intelligence situation correlation fusion mines items related to events from a large amount of situation sequence data based on the attributes of the event list.

[0061] The formats of intelligence situation text data and real-time intelligence situation sequence data are quite different. We cannot simply use the overlap of keywords to indicate whether the data are related. Therefore, we use the event element extraction results to semantically fuse the subject, time, location and other elements contained therein to form an event representation. Finally, we calculate the matching degree between the intelligence situation text data and the intelligence situation sequence data. Figure 2 shown.

[0062] The model is mainly divided into three parts, namely, intelligence situation text event representation generation, intelligence situation sequence event representation generation and data association matching.

[0063] S31. Generate intelligence situation text event representation

[0064] According to the attributes of the intelligence situation text, the main information of the event and the time and place of the event are obtained from the event list to generate the event representation S1.

[0065] S1=(subject, time, ..., place)

[0066] S32. Generate intelligence situation sequence event representation

[0067] Perform data analysis and processing on structured intelligence situation sequence data, filter according to event type attributes in the event list, statistically analyze the time period T when event representation S1 usually occurs, cut the situation data in the time window [tT, t+T] when the S1 event occurs, represent the situation data within the time window, and generate multiple situation event representations S i .

[0068] S i =(subject, time, ..., place)

[0069] S33, data association matching connects S1 and S i Input the association matching model to obtain the semantic matching score C of the two data:

[0070] C = association matching (S1, S i )

[0071] Set the association matching threshold τ. If the matching score C>τ, it means that the situation S i The same event as event S1. Situation data S i The set θ of situations associated with event S1 in represents the situation information associated with and matched to event S1. By analyzing θ, we can obtain the start time and end time of event S1 and various dynamic information of the event subject.

[0072] Embodiment 1:

[0073] Intelligence Situation Fusion:

[0074] Taking the "** event" as an example, for the intelligence situation text data, it can be associated with the intelligence situation sequence data based on the results of the event elements, including event type, occurrence time, enemy equipment, our equipment, event location and other information. The intelligence situation sequence contains the situation information corresponding to the event, including equipment type, equipment model, time and space information, etc. The situation information, as data that describes the specific process of the event, is more detailed and intuitive than the document type.

[0075] There is a certain correlation between event elements and intelligence situation sequence data. By statistically analyzing the sequence attributes in the correlation result θ state, we can obtain features such as trajectory, maximum speed, and average speed. The obtained features can be integrated into the attribute description of the event.

[0076] The present invention discloses an intelligence situation mining method and analysis system based on event extraction, the main advantages of which are embodied in the following aspects:

[0077] (1) A method for event scene classification is designed. For intelligence situation text data, domain words are extracted from it. Based on the different contributions of words and domain words in sentences to event scene classification, a text classifier is constructed and the text classification results are output. This solves the problem that traditional event classification methods cannot be applied to text data without trigger words.

[0078] (2) A method for event element extraction is proposed. Event templates are constructed according to different event types. The basic information and specific information to be extracted are formulated. Then, an end-to-end encoder-decoder framework is constructed to solve the problem of difficulty in extracting location elements and time elements in document-level multi-event intelligence situation text.

[0079] (3) A method for intelligence situation association and fusion is proposed. Based on the common elements in the event list and the intelligence situation sequence attributes, a feature template is constructed to achieve similarity association at the semantic level. The dynamic information in the situation sequence is analyzed based on the association results and finally integrated into the structured event description.

[0080] 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. An intelligence situation mining method based on event extraction, characterized in that: The method comprises the following steps: S1. Event scenario classification: Extract intelligence-related words from the existing intelligence situation text data, and classify the intelligence situation text data into event scenario type libraries based on the categories and frequencies of domain words; S2. Event element extraction: Design event templates for different event scenario types, extract event elements from intelligence situation text data based on the event templates, and form an event list; S3, Intelligence situation association fusion: Search the intelligence situation sequence data based on event elements, associate and match similar attributes, and form an intelligence situation database; in, The step S1 specifically includes the following steps: S11. Data preprocessing For intelligence situation text data, firstly, regular expressions are used to clean the data, and then word segmentation tools are used to perform word segmentation. S12. Domain word extraction Domain words include the subject and domain information of the event. When extracting domain words, named entity recognition is used to extract domain words from the original text data, including weapon equipment models, mission types, and combat units. S13. Construction of event scenario type system According to the characteristics of the subject, environment, and mission objectives in the intelligence situation text, an event scenario type system is manually formulated; S14. Text Classification The intelligence situation text with domain word labels is classified into event scenarios through word and domain word coding, event scenario type coding and scenario type classification; The step S14 specifically includes the following steps: S141, word and domain word encoding Using the word segmentation results and domain word extraction results of intelligence situation text data, the language model Skip-gram model is used to initialize word encoding d w and domain word encoding d e , learn word vectors from the intelligence situation corpus, generate a vector for each word, which contains the word and the type of word; S142, event scene type code The event scenario type encoding of the intelligence situation text is represented by two randomly initialized encoding vectors t1 and t2, where t1 is used to capture local information and t2 is used to capture global information; S143, scene type classification The long short-term memory network LSTM is used to process the word and domain word encoding of the intelligence situation text, and h i It represents the result after each layer of LSTM, and uses the attention mechanism to evaluate the impact of each word in the intelligence situation text on the judgment of the event scene type. where α k is the attention score of the kth word; It can be obtained that the intelligence situation text is represented by S att =α T H, where α=[α1,…,α n ] is the attention vector of the word, indicating the importance of the word in the sentence, H = [h1,…,h n ] represents the output of each layer of LSTM; S144, use Represents the local features of the text of the domain word, using Represents the global characteristics of the intelligence situation text, h n is the output of the nth layer LSTM. The weighted sum of the two is calculated by the Sigmoid function to obtain the event scene type: o=σ(λ·v att +(1-λ)·v global ) Where σ represents the Sigmoid function, and λ is a hyperparameter used to balance the local and global features of the text; S145. Through continuous training, adjust t1 and t2 so that the event scene type conforms to the event scene type system established in S13.

2. The intelligence situation mining method based on event extraction as claimed in claim 1 is characterized in that: The data cleaning includes: converting English characters to lowercase and replacing dates and longitude and latitude into a standard unified format.

3. The intelligence situation mining method based on event extraction as claimed in claim 1, characterized in that: The step S2 specifically includes the following steps: S21. Event template construction Domain experts build event templates for S1 to obtain different event scenario types based on the needs of intelligence situation mining; S22. Event extraction Event extraction uses an end-to-end encoder-decoder framework, which includes a document-level encoder, a role decoder, and an event decoder. The document-level encoder uses the Transformer framework to interactively model all sentences and domain words contained in the sentences. The role decoder and event decoder are used to extract elements of multiple events in parallel in the document.

4. The intelligence situation mining method based on event extraction as claimed in claim 3 is characterized in that: The event template includes two parts: basic information and specific information. The basic information includes the event name, event type, personnel, location, start time and end time. The specific information includes the attribute information and dynamic information of the specific event subject.

5. The intelligence situation mining method based on event extraction as claimed in claim 4 is characterized in that: The attribute information of the event subject includes: equipment type and equipment model, and the dynamic information includes: speed and heading.

6. The intelligence situation mining method based on event extraction as claimed in claim 3 is characterized in that: The step S22 specifically includes: Use document-level relationship extraction technology to dynamically adjust word semantic vectors based on contextual content; The strategy of parallel labeling of place names is adopted. When extracting place name arguments, the place names are labeled in parallel and then integrated in the order of appearance, that is, point = (point1, point2, point3, ..., point n ); When extracting time arguments, the start time and completion time of the time indicator are extracted independently according to the semantics in the context, and the time-modifying components are extracted together; Finally, by utilizing global information, multi-granularity event extraction models based on single documents and cross-documents are used to extract event elements from text data according to event templates, and the event element information between documents is sorted out to form an event list, thereby realizing the integration of descriptions of the same event in multiple intelligence situation text data.

7. The intelligence situation mining method based on event extraction according to any one of claims 4 to 6, characterized in that: The step S3 specifically includes the following steps: S31. Generate intelligence situation text event representation According to the attributes of the intelligence situation text, the main information of the event and the time and place of the event are obtained from the event list to generate the event representation S1. S1=(subject, time, ..., place); S32. Generate intelligence situation sequence event representation Perform data analysis and processing on the structured real-time intelligence situation sequence data, filter according to the event type attribute in the event list, statistically analyze the time period T of the event representation S1, cut the situation data in the time window [tT, t+T] of the S1 event occurrence time, represent the situation data in the time window, and generate multiple situation event representations S i , S i =(subject, time, ..., place) S33, data association matching connects S1 and S i Input the association matching model to obtain the semantic matching score C of the two data: C = association matching (S1, S i ) Set the association matching threshold τ. If the matching score C>τ, it means that the situation S i This is the same event as event S1, and the situation data S i The set θ of situations associated with event S1 in represents the situation information associated with and matched to event S1.

8. The intelligence situation mining method based on event extraction as claimed in claim 7 is characterized in that: The step S33 also includes: analyzing θ to obtain the start time and end time of event S1 and various dynamic information of the event subject.

Images