CN112949300A - Typhoon early warning planning model automatic generation method and system based on deep learning - Google Patents

Abstract

The invention discloses a method and a system for automatically generating a typhoon early warning planning model based on deep learning, which comprises the following steps: A. acquiring a natural language text related to typhoon early warning; B. deep learning is carried out based on NLP, and a word vector set related to typhoon early warning planning is established; C. and constructing an action expression aiming at the typhoon early warning plan, and generating the output of the Bayesian network of the current domain. The method and the system for automatically generating the typhoon early warning planning model based on the deep learning automatically create the typhoon emergency plan planning field model by using the planning field definition language and introduce the natural language processing method to convert the action sequence problem required by generating the Bayesian network into the text sequence marking problem, thereby providing a typhoon emergency planning frame for generating the typhoon early warning planning model aiming at the text characteristics of the typhoon advance plan and forming the automatic production technology of the typhoon early warning planning model.

Description

Typhoon early warning planning model automatic generation method and system based on deep learning

Technical Field

The invention relates to a software implementation method and a software implementation system for a typhoon early warning plan model, in particular to an improvement on a deep learning and automatic generation method and a system for a planning model in the typhoon response field.

Background

In the prior art, natural disasters are one of the major problems facing human beings in the world, and the sustainable development of economy and society is seriously influenced and the survival of human beings is threatened. Although natural disasters are unavoidable, if the overall story line of the weather development and corresponding countermeasures can be extracted from the past disaster events, reference can be provided for the decision on prevention and control of newly occurring disasters. Based on the improvement of the computing power of the prior art, in order to obtain the story line and the disaster relief measures of the disaster events in the past year, the key information of the disaster occurrence is correctly and effectively analyzed and extracted, and the disaster prevention and relief measures are summarized, so that it is very important to automatically form a new measure scheme for dealing with the possible disaster events.

The intelligent planning by an arithmetic system is an important branch of artificial intelligence, and the main idea is to recognize and analyze the surrounding environment, and to reason a plurality of selectable executable rescue measures and actions under the condition of providing limited resources and relevant constraints according to the preset target and the past countermeasures, so as to comprehensively make an action sequence for realizing the set target, and the action sequence is called a plan in the prior art.

The intelligent planning focuses on realizing strategies or action sequences, and can be applied to multiple fields of intelligent driving, action planning of intelligent robots, aerospace technology and the like. One of the challenging research contents is based on intelligent programming narration, and how to obtain useful information from narrative text to construct a field model becomes a large focus of the narrative programming field.

For intelligent planning modeling, some methods have been studied and improved at present. Framer is a classical method of learning planning domain models from natural language descriptions, which starts from natural sentence input, uses the CoreNLP tool of stanford to extract action templates, classifies the actions, extracts consistent expression forms from original sentences, and finally uses the LOCM tool to obtain planning domain models. storyFramer is improved and expanded on the basis of Framer, the Framer method is applied to narration planning, actions and attributes are extracted by using a CoreNLP tool, and then a user input link is introduced, so that the robustness and the practicability of the whole framework are improved.

However, experiments show that the NLP tool used in the method cannot well distinguish complex situations such as ambiguous words, which may cause part-of-speech errors in the labeling, and thus the extracted action template is redundant and redundant, which may result in an increase in subsequent user interaction. Meanwhile, there is no planning and early warning scheme for typhoon weather disasters in the prior art, so the prior art is still to be improved and developed.

Disclosure of Invention

The invention aims to provide a method and a system for automatically generating a typhoon early warning planning model based on deep learning, provides a brand-new method and a brand-new system for generating a typhoon text-oriented emergency planning field model, and improves the operability and the practicability of the typhoon text as much as possible.

The technical scheme of the invention is as follows:

a typhoon early warning planning model automatic generation method based on deep learning comprises the following steps:

A. acquiring a natural language text related to typhoon early warning;

B. deep learning is carried out based on NLP, and a word vector set related to typhoon early warning planning is established;

C. and constructing an action expression aiming at the typhoon early warning plan, and generating the output of the Bayesian network of the current domain.

The method for automatically generating the typhoon early warning planning model based on the deep learning comprises the following steps:

b1, converting the natural language text input in the step A into word vectors through a BERT model, and labeling parts of speech to words by using a BilSTM-CRF model;

and B2, extracting action triples of the subject, the predicate and the object according to the part-of-speech tagging result, and generating an action set and an object set.

The method for automatically generating the typhoon early warning planning model based on the deep learning, wherein the step B1 further comprises:

b11, segmenting natural language texts related to typhoon early warning based on a BERT model, wherein each word is represented by a hidden feature space vector with given dimensionality;

b12, inputting the sentence represented by the vector into the BilSTM model, and obtaining the part-of-speech predicted labels of all words in the sentence according to the output.

The method for automatically generating the typhoon early warning planning model based on the deep learning comprises the following steps of: and carrying out de-duplication processing on the repeated word vectors.

The method for automatically generating the typhoon early warning planning model based on the deep learning further comprises the following steps after the step B12:

and B13, adding a constraint to the final prediction label when the label is input into the CRF model, and ensuring the output result to be reasonable and effective, wherein the constraint is formed by automatic learning in the training process through deep learning.

The method for automatically generating the typhoon early warning planning model based on the deep learning, wherein the step B2 further comprises:

and B21, marking each associated object as a candidate parameter of the output action for each action ternary.

The method for automatically generating the typhoon early warning planning model based on the deep learning comprises the following steps: and setting a user interaction link for carrying out merging action, homogeneous item setting of the object, and a precondition and a postcondition for classifying the object and/or adding the action.

The typhoon early warning planning model automatic generation method based on deep learning is characterized in that three general categories are set in the user interaction link: and the department, the typhoon attribute and the direction are used for selecting from the candidate parameters output by the user before and corresponding to the corresponding category.

The typhoon early warning planning model automatic generation method based on deep learning judges possible causal relationship among each action by using a Glange causal relationship test method, and adds a precondition and a postcondition for each action according to the causal relationship.

Any one of the systems of the method for automatically generating the typhoon early warning planning model based on the deep learning comprises the following steps:

the natural language text acquisition module is used for acquiring a natural language text related to typhoon early warning;

an NLP deep learning module, which is used for deep learning based on NLP and establishing a word vector set related to typhoon early warning planning;

and the typhoon early warning plan construction module is used for constructing action expression aiming at the typhoon early warning plan and generating Bayesian network output of the current domain.

The method and the system for automatically generating the typhoon early warning planning model based on the deep learning have the advantages that natural language processing (NLP,_{Natural Language Processing}) The method comprises the steps of converting an action sequence problem required by the Bayesian network generation into a text sequence labeling problem, thereby providing a typhoon emergency planning network generation framework aiming at the text characteristics of a typhoon pre-plan, and constructing a structural representation of a domain model file by adopting an NLP (non line segment) technology to form an automatic production technology of the Bayesian network early warning planning model for the typhoon.

Furthermore, a final output model for user interaction perfection and optimization is introduced, so that the typhoon prediction planning model disclosed by the invention is better adapted to the actual situation, and an information inlet facilitating interaction with a user is provided.

Drawings

Fig. 1 is a schematic diagram of a model framework generation flow of the method and system for automatically generating a typhoon early warning planning model based on deep learning according to the preferred embodiment of the invention.

Fig. 2 is an exemplary diagram of triplet information in the preferred embodiment of the method and system for automatically generating a typhoon early warning planning model based on deep learning according to the present invention.

Fig. 3 is another exemplary diagram of triple information extraction in the preferred embodiment of the method and system for automatically generating a typhoon early warning planning model based on deep learning according to the present invention.

Fig. 4 is a system block diagram of a method and a system for automatically generating a typhoon early warning planning model based on deep learning according to a preferred embodiment of the present invention.

Detailed Description

The following describes in detail preferred embodiments of the present invention.

The preferred embodiment of the method and the system for automatically generating the typhoon early warning planning model based on deep learning, as shown in fig. 1, is an application implementation for typhoon early warning planning based on the algorithm in the existing intelligent planning field.

Firstly, natural language texts related to typhoon early warning are required to be acquired, the natural language texts comprise instructions and descriptions of historical common typhoon early warning and handling measures and the like, and mainly comprise all natural language texts. The input natural language text converts the input natural language sentence into a word vector by a BERT model (bidirectional encoderpressationfront transformations), and labels parts of speech to words using a BiLSTM-CRF model.

And then, extracting action triples according to the part-of-speech tagging result, and eliminating repeated items to generate an action set and an object set. And then, through user interaction links, the method comprises the steps of combining actions, setting homogeneous items of the objects, classifying the objects and/or adding preconditions and postconditions of the actions and the like.

Finally, a Bayesian network of the current domain is generated.

In the BilSTM-CRF model, the ratio of LSTM: the full name Long Short-Term Memory is one of RNN (Current Neural network). LSTM is well suited for modeling time series data, such as text data, due to its design features. BilSTM: the abbreviation of Bi-directional Long Short-Term Memory is composed of forward LSTM and backward LSTM. It can be seen that it is well suited to do the sequence labeling task with upper and lower relationships, and can form the relationship with the time sequence, and prevent the simple static word vector from generating the correlation which is not practical, so it is often used to model the context information in NLP, and it can be simply understood that the bidirectional LSTM is the improved version of LSTM, and the LSTM is the improved version of RNN.

It is briefly stated that RNN, in order to predict the final result, first uses the first word to predict, and certainly, only uses the prediction result of the first prediction as inaccurate, but uses this result as a feature, together with the second word, to continue predicting the result; then, combining the new prediction result with the third word to make new prediction; this process is then repeated; up to the last word. Thus, if there are n words entered, the result is in fact predicted n times, giving n predicted sequences. Throughout the process, the model shares a set of parameters.

Thus, the RNN reduces the number of parameters of the model, prevents overfitting, and is inherently designed to handle sequence problems, and can reduce static logical association errors in time order, and is therefore particularly suited to handle sequence problems. LSTM improves RNN to capture longer distance information, but both LSTM and RNN have a problem in that it is pushed from left to right so that later words are more important than earlier words. Thus a bi-directional BilSTM appears that does an LSTM once from left to right, then does an LSTM once from right to left, and then combines the two results to make the result more logical to reality.

The CRF (Conditional Random Fields) is an algorithm model that is difficult in the field of machine learning, and the difficulty lies in its definition, which involves probabilities such as probability map models and clusters; the method is approximate to perfect in mathematics, and relates to probability, expectation calculation and knowledge in optimization, but the method is better in application effect in natural language processing.

In the preferred embodiment of the method and the system for automatically generating the typhoon early warning planning model based on the deep learning, the finally obtained action expression is a simplified expression of a natural language input statement, namely a triple in design: subject, predicate, object. These triplets may represent all actions and properties described in the original sentence, the mentioned objects and their role in the original sentence.

In order to extract triples from a sentence, in the preferred embodiment of the present invention, syntactic analysis is performed by means of the best currently successful BERT-BiLSTM-CRF model to complete named entity recognition and part-of-speech tagging. The part of speech tagging function, which is closest to the actual requirement of the present invention, can be divided into three parts.

First, word embedding is performed based on the BERT model, and compared with english, the first difficulty in processing chinese text is to segment sentences as accurately as possible. The important reason for selecting this model in the preferred embodiment of the present invention is that it has a segmentation mechanism and the segmentation accuracy is close to 100%. Through word embedding, a potentially low-dimensional semantic representation of each word and the entire sentence is obtained, that is, each word can be represented by a hidden feature space vector of a given dimension.

Secondly, the sentences represented by the vectors are input into a BilSTM model, and the part-of-speech predicted labels of all words in the sentences are obtained according to the output of the BilSTM model. In order to greatly reduce invalid sequences in the prediction result, the output of the second part is further input into the CRF model in the preferred embodiment of the invention, constraints can be added to the final prediction label to ensure that the output result is reasonable and effective, and more importantly, the constraints can be formed through deep learning automatic learning in the training process, so that no more labor cost is brought.

Then, in the preferred embodiment of the present invention, based on the obtained part-of-speech tag, an action triple can be extracted, where the action represents using a verb as an action name and includes all related objects. An example of a part-of-speech result of a natural language input is illustrated in fig. 2. In the figures, the word labeled "NN" represents a noun, usually as a candidate subject or object (and modifier) in the current domain; the word labeled "VV" represents a verb, i.e., a predicate, typically representing one possible candidate action; while the "JJ" label represents an adjective and the "PU" label represents a punctuation mark, all without concern.

It can be seen that the subject of the sentence is "central weather station", the verb of the predicate is "issue", the object is "typhoon warning", and the modifier is "orange". Therefore, in the preferred embodiment of the method of the present invention, a motion triple can be extracted from the sentence according to the result, and particularly, modifiers are also added to the motion triple to keep the detailed information of the subject or object. The output of the sentence will be ("central weather station", "release", "typhoon orange warning") and it can be seen that the triplet contains the key information of the sentence.

In a preferred embodiment of the method and system of the present invention, in addition to creating triples, for each action triplet, each associated object is marked as a candidate parameter for the output action. In the stage of user interaction, the method and the system set three general categories, namely departments, typhoon attributes and directions, and a user can select from the previously output candidate parameters and correspond to the corresponding categories to determine the clear action direction of the typhoon.

The typhoon is particularly common disaster weather which is easy to occur, and the forming and developing time and space of the typhoon are very short and fast moving, so the typhoon early warning planning has practical significance.

For the precondition and postcondition of each action, there are few methods to identify them, and in the preferred embodiment of the method and system of the present invention, the gram jeopardy test method in the field of economics is used, and the correlation between two actions is analyzed based on statistics to determine whether they are causality. Then, the preconditions and postconditions associated with each action are screened out.

In the preferred embodiment of the method and the system of the invention, a user interaction interface is also provided for correcting and constraining the typhoon early warning plan. In natural language texts, the same action or object can often have a plurality of different expression forms, and the merging technology of similar expressions is still a difficult problem in the field of NLP. Therefore, in order to eliminate the same kind of terms, the method and the system of the present invention require the user to select and delete the same meaning expression by himself, so that the same action or object can have a fixed expression, in other words, no two words in a single domain represent the same action or object.

Where the user is required to classify the objects into different categories. In the setting of the preferred embodiment of the method and the system of the invention, three types of objects can be selected, namely departments, typhoon attributes and directions, which are information having great significance with typhoon early warning planning.

In the setting of the preferred embodiment of the method and system of the present invention, two predicates are assigned to each action by using the grand causal relationship test method, but because the statistical-based method may not be comprehensive in practical situations, the user is required to modify and add the predicates, so that the action preconditions are more reasonable.

The preferred embodiment of the invention takes the following input sentences as an example to verify the technical effect, and specifically takes the typhoon early warning plan of the mansion as an implementation ground:

s1 "typhoon news is forwarded in time in city flood prevention office"

S2 report of city flood prevention to market committee "

S3 "city weather station timely issuing typhoon notice"

As shown in fig. 3, the part-of-speech tagging result, the method and system of the present invention extract actions and objects from the above statements according to the steps, and the obtained actions and the objects and parameters thereof are shown in table 1 below.

TABLE 1

Specifically, for the sentence S1, words labeled "NN" and "VV" in series are first selected, here, "city flood prevention office forwards typhoon news" in time, and then a word of "NN" (i.e., a word labeled "NN" before "VV") appears before the word labeled "VV") as the first item of the triplet; subsequently, the word labeled "VV" naturally becomes the second term, while the last remaining words labeled "NN" (i.e., all words labeled "NN" after "VV") are the third term. At this time, the action triplets are ("city flood office", "forward", "typhoon news"), which can almost express the meaning of the original S1.

However, in order to preserve as much information as possible in the sentence, the preferred embodiment of the present invention provides that all modifiers (if any) in the sentence are preserved at their respective modification positions to avoid semantic loss. The modifier in S1 is the word labeled "AD". Therefore, the final triplet output in the sentence S1 is ("city flood office", "forward in time", "typhoon news").

For extracted triples, the present method and system will identify the major action elements and parameters and add their preconditions and postconditions as much as possible for each action using the granger causal relationship test, as such information is essential for the construction of bayesian nets.

The set of actions and their objects obtained after automatic processing of the building typhoon response data set organized as a preferred embodiment are shown in table 2 below.

TABLE 2

In the preferred embodiment of the present invention, a user interaction step follows. Firstly, the user needs to screen out the same kind of objects expressing different meanings and combining the same kind of objects. As shown in table 2, in the generated action predicate object table, there are some homogeneous items in the object column, such as wharf and city wharf, which requires the user to select one and delete the rest of the repeated items.

After the problem of duplicate items is solved, the user can classify the objects. In combination with practical requirements, the preferred embodiment of the present invention limits categories to three categories, department, typhoon attributes and location. For the current domain, the main task of the user interaction is object modification, i.e. to classify different objects into given categories, the classification results of which are also shown in table 2, where the parts requiring user interaction have been specifically indicated. It can be seen that relatively few parts need to be input by the user, which also highlights the automation and intelligence of the present technique.

In the preferred embodiment of the method and system for automatically generating a typhoon early warning planning model based on deep learning of the present invention, a system for implementing the method is provided, as shown in fig. 4, which specifically includes: the natural language text acquisition module is used for acquiring a natural language text related to typhoon early warning; an NLP deep learning module, which is used for deep learning based on NLP and establishing a word vector set related to typhoon early warning planning; and the typhoon early warning plan construction module is used for constructing action expression aiming at the typhoon early warning plan and generating Bayesian network output of the current domain. The modules are realized in a software system, are used for realizing typhoon early warning planning, particularly emergency plan preparation for typhoon, and have the processing capability which is faster and more intelligent than people.

The invention provides a typhoon early warning planning model automatic generation method and system based on deep learning, and provides a planning field model generation framework for typhoon texts, namely, according to the characteristics of the typhoon texts, a BERT-BilSTM-CRF model is used for part-of-speech tagging, action triples are extracted, an action set and an object set are generated, and a Bayesian network model in the typhoon field is generated after a user interaction link.

The method and the system focus on the intelligent planning direction, and the research direction of the method and the system is put on the automatic generation of the planning because the current typhoon early warning planning scheme is mainly evaluated and formulated by human experience. Because of the existence of the Bayesian network model in the prior art, the problems of the method and the system of the invention in the preferred embodiment are converted into the automatic generation of the Bayesian network model in the planning field aiming at typhoon early warning. The method can be applied to typhoon emergency response generation, and can generate corresponding emergency response measures aiming at new typhoon information and according to the time sequence process of typhoon development as long as the prior typhoon information of a local domain and the instructions and effects of the corresponding measures are collected to generate a field Bayesian network model.

It is worth mentioning that the method and the system have higher automation degree, can be operated without professional planning field experts, and have high practicability.

It will be understood that modifications and variations can be made by persons skilled in the art in light of the above teachings and all such modifications and variations are intended to be included within the scope of the invention as defined in the appended claims.

Claims (10)

1. A typhoon early warning planning model automatic generation method based on deep learning comprises the following steps:

A. acquiring a natural language text related to typhoon early warning;

B. deep learning is carried out based on NLP, and a word vector set related to typhoon early warning planning is established;

C. and constructing an action expression aiming at the typhoon early warning plan, and generating the output of the Bayesian network of the current domain.

2. The method for automatically generating the typhoon early warning planning model based on the deep learning of claim 1, wherein the step B further comprises:

b1, converting the natural language text input in the step A into word vectors through a BERT model, and labeling parts of speech to words by using a BilSTM-CRF model;

and B2, extracting action triples of the subject, the predicate and the object according to the part-of-speech tagging result, and generating an action set and an object set.

3. The method for automatically generating the typhoon early warning planning model based on the deep learning of claim 2, wherein the step B1 further comprises:

b11, segmenting natural language texts related to typhoon early warning based on a BERT model, wherein each word is represented by a hidden feature space vector with given dimensionality;

b12, inputting the sentence represented by the vector into the BilSTM model, and obtaining the part-of-speech predicted labels of all words in the sentence according to the output.

4. The method for automatically generating the typhoon early warning planning model based on the deep learning of claim 3, wherein the step B12 further comprises: and carrying out de-duplication processing on the repeated word vectors.

5. The method for automatically generating the typhoon early warning planning model based on the deep learning of claim 3 or 4, wherein after the step B12, the following steps are further executed:

and B13, adding a constraint to the final prediction label when the label is input into the CRF model, and ensuring the output result to be reasonable and effective, wherein the constraint is formed by automatic learning in the training process through deep learning.

6. The method for automatically generating the typhoon early warning planning model based on the deep learning of claim 5, wherein the step B2 further comprises:

and B21, marking each associated object as a candidate parameter of the output action for each action ternary.

7. The method for automatically generating the typhoon early warning planning model based on the deep learning of claim 6, wherein the step C further comprises: and setting a user interaction link for carrying out merging action, homogeneous item setting of the object, and a precondition and a postcondition for classifying the object and/or adding the action.

8. The method for automatically generating the typhoon early warning planning model based on the deep learning of claim 7 is characterized in that in the user interaction link, three general categories are set: and the department, the typhoon attribute and the direction are used for selecting from the candidate parameters output by the user before and corresponding to the corresponding category.

9. The method for automatically generating a typhoon early warning planning model based on deep learning of claim 8, wherein a precondition and a postcondition are added for each action based on the granger causal relationship test method.

10. The system for the automatic generation method of the deep learning-based typhoon early warning planning model according to any one of claims 1 to 9, is characterized by comprising the following steps:

the natural language text acquisition module is used for acquiring a natural language text related to typhoon early warning;

an NLP deep learning module, which is used for deep learning based on NLP and establishing a word vector set related to typhoon early warning planning;

and the typhoon early warning plan construction module is used for constructing action expression aiming at the typhoon early warning plan and generating Bayesian network output of the current domain.

Images