CN112488316A - Event intention reasoning method, device, equipment and storage medium - Google Patents
Event intention reasoning method, device, equipment and storage medium Download PDFInfo
- Publication number
- CN112488316A CN112488316A CN202011460065.6A CN202011460065A CN112488316A CN 112488316 A CN112488316 A CN 112488316A CN 202011460065 A CN202011460065 A CN 202011460065A CN 112488316 A CN112488316 A CN 112488316A
- Authority
- CN
- China
- Prior art keywords
- event
- fuzzy
- intention
- inference
- current occurrence
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 66
- 238000012545 processing Methods 0.000 claims abstract description 65
- 238000003062 neural network model Methods 0.000 claims abstract description 47
- 239000013598 vector Substances 0.000 claims description 50
- 238000013528 artificial neural network Methods 0.000 claims description 37
- 238000012549 training Methods 0.000 claims description 28
- 230000008569 process Effects 0.000 claims description 27
- 239000011159 matrix material Substances 0.000 claims description 18
- 238000004590 computer program Methods 0.000 claims description 3
- 239000000203 mixture Substances 0.000 claims description 3
- 230000001537 neural effect Effects 0.000 abstract description 6
- 238000000605 extraction Methods 0.000 description 17
- 238000010586 diagram Methods 0.000 description 10
- 230000006870 function Effects 0.000 description 10
- 238000004891 communication Methods 0.000 description 7
- 230000009471 action Effects 0.000 description 5
- 210000002569 neuron Anatomy 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 230000002457 bidirectional effect Effects 0.000 description 2
- 238000005352 clarification Methods 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 238000003672 processing method Methods 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 238000007476 Maximum Likelihood Methods 0.000 description 1
- 230000001364 causal effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000008451 emotion Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 239000013604 expression vector Substances 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000008450 motivation Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 230000010355 oscillation Effects 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 230000001960 triggered effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/04—Inference or reasoning models
- G06N5/048—Fuzzy inferencing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/043—Architecture, e.g. interconnection topology based on fuzzy logic, fuzzy membership or fuzzy inference, e.g. adaptive neuro-fuzzy inference systems [ANFIS]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Fuzzy Systems (AREA)
- Automation & Control Theory (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Mathematical Analysis (AREA)
- Computational Mathematics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The application discloses an event intention inference method, an event intention inference device, event intention inference equipment and a storage medium. According to the method and the device, the event intention reasoning problem of the meaning fuzzy information is solved through the fuzzy theory, meanwhile, the fuzzy theory is combined with the neural network model, the classification neural network model is used for fuzzification processing, the strong feature expression capability of the deep neural model can be learned, the fuzzy reasoning capability of the fuzzy theory is inherited, and therefore the method and the device are more suitable for reasoning the fuzzy event intention, and the reasoning accuracy of the event intention is improved.
Description
Technical Field
The present application relates to the field of natural language processing technologies, and in particular, to an event intention inference method, apparatus, device, and storage medium.
Background
With a large amount of structured and unstructured text data such as research reports, dynamic messages, overview documents, parameter charts and the like in various field scenes, the data are increasing in an exponential level. The real-life intent of how to capture an event from large-scale knowledge data becomes difficult to guess.
In particular, for some field scenarios, the event is intended to be ambiguous, e.g., whether the military field inference event is intended to be reconnaissance, the uncertainty it represents is ambiguous, as the motivation or concept for reconnaissance is ambiguous itself. In addition, in other fields, the perception, emotion, judgment and the like of an object need to be inferred from a large amount of text data, and the intentions are fuzzy. For such ambiguous event intentions, the inference implementation process will be more difficult.
Disclosure of Invention
In view of the above problems, the present application is proposed to provide an event intention inference method, apparatus, device and storage medium, which can implement an inference acquisition process of ambiguous event intentions. The specific scheme is as follows:
an event intent inference method, comprising:
acquiring event information of a current occurrence event;
fuzzifying the event information by using a pre-trained classification neural network model to obtain a fuzzification result;
combining and reasoning the fuzzification result by adopting a fuzzy rule to obtain a fuzzy reasoning result;
and performing defuzzification processing on the fuzzy inference result to obtain the event intention of the current occurrence event.
Preferably, the process of performing fuzzification processing, combining and reasoning on the fuzzification result, and defuzzification processing on the event information includes:
inputting the event information into a pre-trained fuzzy neural network intention inference model to obtain the event intention of the current occurrence event predicted by the model; wherein,
and the fuzzy layer of the fuzzy neural network intention inference model performs fuzzification processing on the event information by using a pre-trained classification neural network model, the fuzzy inference layer combines and infers fuzzy results by adopting a fuzzy rule to obtain a fuzzy inference result, and the defuzzification layer performs defuzzification processing on the fuzzy inference result to obtain the predicted event intention.
Preferably, the fuzzifying the event information by using the pre-trained classification neural network model as a fuzzy membership function includes:
acquiring text features of the event information, wherein the text features are composed of n-dimensional vectors (x)1,x2,...xn) Composition is carried out;
corresponding the text features to n-dimensional vectors (x)1,x2,...xn) Expanding the data into n data;
weight matrix W by classifying neural network modelsm*nAnd performing matrix dot product operation on the n expanded data to obtain n output vectors as fuzzification results, wherein each output vector is an m-dimensional vector and corresponds to the attribution degree of the m class labels.
Preferably, the combining and reasoning of the fuzzification result by using the fuzzy rule includes:
combining and reasoning said n output vectors using fuzzy rules, wherein,
the number of the fuzzy rules is the same as the number of clustering centers after n dimensions are clustered in n-dimensional vectors corresponding to the text features.
Preferably, the method further comprises the following steps:
acquiring an event correlation map established based on historical event data, wherein points in the event correlation map represent events, directed edges between the points represent incidence relations of the events, and weights on the directed edges represent the occurrence frequency of the incidence relations in the historical event data;
determining inference events which are possibly generated after the current occurrence event based on the event correlation map, and determining an event intention of the current occurrence event based on the event type of the inference events;
and determining the final event intention of the current occurrence event by combining the event intention of the current occurrence event determined based on the inference event and the event intention of the current occurrence event obtained after the defuzzification processing.
Preferably, the determining, based on the event correlation map, inference events that may occur after the current occurrence event includes:
obtaining an event inference prediction model trained based on the event correlation map, the event inference prediction model configured to predict a state representation of a degree of correlation between event pairs based on input event pairs;
determining a candidate event corresponding to the current occurrence event based on the event correlation map, wherein the candidate event is an event which may occur after the current occurrence event;
respectively forming an event pair by each candidate event and the current occurrence event, and inputting the event pair into the event inference prediction model to obtain the correlation degree between each candidate event and the current occurrence event output by the model;
and determining inference events from each of the candidate events based on the correlation.
Preferably, the determining a candidate event corresponding to the current occurrence event based on the event correlation map includes:
determining a target point corresponding to the current occurrence event in the event correlation map;
and searching points directly or indirectly pointed by the target point according to a set searching distance by taking the target point as a center, and forming a candidate event by the events corresponding to the searched points.
Preferably, the determining inference events from each of the candidate events based on the correlation comprises:
determining the average value of each weight on a directed edge link between each candidate event and the current occurrence event in the event correlation map, and taking the average value as the average weight between each candidate event and the current occurrence event;
weighting the correlation degree between the candidate event and the current occurrence event by using the average weight between the candidate event and the current occurrence event to obtain the weighted correlation degree between each candidate event and the current occurrence event;
and selecting the candidate event with the highest weighted correlation as the reasoning event.
Preferably, the method further comprises the following steps:
generating a model by utilizing a pre-trained interpretable text, processing the event information of the current occurrence event, the event intention and the domain knowledge of the domain to which the current occurrence event belongs, and obtaining the interpretable text of the event intention output by the model;
the interpretable text generation model is obtained by taking event information and event intention of a training event and field knowledge in the field to which the training event belongs as training samples and taking interpretable text corresponding to the training event as sample labels for training.
Preferably, the processing event information of the current occurrence event, the event intention and domain knowledge in the field to which the current occurrence event belongs by using the pre-trained interpretable text generation model comprises:
respectively coding event information by using an interpretable text generation model to obtain an event hidden layer state, and coding domain knowledge to obtain a domain knowledge hidden layer state;
respectively carrying out attention processing on the event hidden layer state and the domain knowledge hidden layer state by utilizing an interpretable text generation model to obtain an event attention representation and a domain knowledge attention representation;
and processing the event attention representation and the domain knowledge attention representation by using an interpretable text generation model to obtain an interpretable text of the event intention.
Preferably, the domain knowledge includes domain keywords included in the event information and a preset event intention inference rule.
An event intention inference apparatus, comprising:
the event information acquisition unit is used for acquiring the event information of the current occurrence event;
the fuzzification processing unit is used for fuzzifying the event information by using a pre-trained classification neural network model to obtain a fuzzification result;
the fuzzy reasoning unit is used for combining and reasoning the fuzzification result by adopting a fuzzy rule to obtain a fuzzy reasoning result;
and the defuzzification processing unit is used for performing defuzzification processing on the fuzzy inference result to obtain the event intention of the current occurrence event.
An event intention inference device, comprising: a memory and a processor;
the memory is used for storing programs;
the processor is used for executing the program to realize the steps of the event intention reasoning method.
A storage medium having stored thereon a computer program which, when executed by a processor, carries out the steps of the event intent inference method as described above.
According to the technical scheme, the event intention inference method introduces a fuzzy theory and a neural network model, firstly obtains the event information of the current event, then fuzzifies the event information by the pre-trained classification neural network model to obtain a fuzzified result, fuzzifies the event information by using the classification neural network model to enable the fuzzified result to have strong characteristic expression capability of a deep neural model and description and expression capability of the fuzzy theory on the fuzzified information, further combines and infers the fuzzified result by adopting a fuzzy rule to obtain a fuzzy inference result, and finally defuzzifies the fuzzified inference result to obtain the event intention of the current event. Therefore, the method and the device solve the problem of event intention inference of the meaning fuzzy information through the fuzzy theory, combine the fuzzy theory and the neural network model, and use the classification neural network model for fuzzification processing, so that the strong feature expression capability of the deep neural model can be learned, and the fuzzy inference capability of the fuzzy theory is inherited, so that the method and the device are more suitable for inferring the fuzzy event intention, and the inference accuracy of the event intention is improved.
Drawings
Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are only for purposes of illustrating the preferred embodiments and are not to be construed as limiting the application. Also, like reference numerals are used to refer to like parts throughout the drawings. In the drawings:
FIG. 1 is a schematic diagram of a fuzzy intent inference process disclosed in an embodiment of the present application;
FIG. 2 is a schematic flow chart of an event intention inference method disclosed in an embodiment of the present application;
FIG. 3 illustrates a schematic diagram of an intention inference model structure of a fuzzy neural network;
FIG. 4 illustrates a BERT model structure diagram;
FIG. 5 illustrates an event correlation map diagram;
FIG. 6 illustrates a schematic diagram of an interpretable text-generating model;
fig. 7 is a schematic structural diagram of an event intention inference device according to an embodiment of the present application;
fig. 8 is a schematic structural diagram of an event intention inference device provided in an embodiment of the present application.
Detailed Description
The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only a part of the embodiments of the present application, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.
The application provides an event intention reasoning scheme which can carry out intention reasoning on a current occurrence event.
The fuzzy theory is introduced to solve the problem of uncertainty of the fuzzy concept, so that more accurate reasoning can be realized for the fuzzy event intention.
Ambiguity in fuzzy theory represents the conceptual uncertainty of a problem, has no definite meaning, and is essentially different from randomness in a probability problem. The randomness in the probability problem is due to the occurrence of insufficient conditions, such that there is no deterministic causal relationship between conditions and events, thereby presenting uncertainty as to whether an event has occurred. Probability randomness also exists for events in certain fields, for example, military event content may reflect enemy and may not react, and obvious determination cannot be achieved.
The basis of the fuzzy theory is a fuzzy set, and the concept of the fuzzy set is different from the classical set. In the classical set, x ∈ U for any element in the domain of discourse U, and the set contained in the domain of discourse U
x either belongs to the set A or does not, i.e., x ∈ A and x ■ A must be one of them, and above the functional representation there is:
this is a binary representation of whether it is or not. Compared with the binary (0 and 1) judgment mode, in the fuzzy set theory, the concepts of the membership degree and the fuzzy membership function are introduced, and do not belong to or do not belong to the classical set. The fuzzy membership function maps the membership of an element to a fuzzy set to [0, 1%]One value in the interval is called membership. For an element in the fuzzy set, which belongs to the fuzzy set and does not belong to the fuzzy set, the boundary is fuzzy, and the membership degree is used to measure the attribution range of the element to the fuzzy setAnd (4) degree. For an element x ∈ U in the universe of discourse U, and a fuzzy set
The fuzzy set theory is described using mathematical language as follows:
{(x,μA(x))|x∈U}
wherein, muA(. -) represents the membership function of the element under the fuzzy set A, μA(x) Has a value range of [0,1 ]],μA(x) Closer to 1, the higher the degree of membership of x to the fuzzy set a, i.e. the more likely it belongs to the set a.
The application provides a fuzzy intention reasoning process, as shown in fig. 1.
The whole reasoning process comprises three parts, namely fuzzification processing, fuzzy reasoning and defuzzification processing.
And in the fuzzification processing step, the input variable is fuzzified by using a pre-trained classification neural network.
And in the fuzzy reasoning step, the fuzzy reasoning result is combined and reasoned through the fuzzy rules in the fuzzy rule base to obtain the fuzzy reasoning result.
And (5) a defuzzification processing link, which is used for carrying out clarification processing on the fuzzy reasoning result to obtain the final event intention.
The scheme can be realized based on a terminal with data processing capacity, and the terminal can be a mobile phone, a computer, a server, a cloud terminal and the like.
Next, as described in conjunction with fig. 2, the event inference method of the present application may include the following steps:
and step S100, acquiring event information of the current occurrence event.
Specifically, the current occurrence event is an event that needs to be inferred by the intention of the event. In this step, event information of the current occurrence event is obtained, and the event information includes description of the current occurrence event. Taking an event in the military field as an example, an exemplary event message may be: the aircraft of the country A appears in the territorial sea of the country B, and the true intention of the aircraft of the country A, such as whether to execute a detection task and the like, needs to be inferred aiming at the military event.
And S110, fuzzifying the event information by using a pre-trained classification neural network model to obtain a fuzzification result.
Specifically, in the scheme provided by this embodiment, the event information is fuzzified by using the classification neural network model, so that the fuzzification result has strong feature expression capability of the deep neural model and fuzzy information description and expression capability of the fuzzy theory.
The pre-trained classification neural network model can adopt a stack type self-encoder SAE or a feedforward neural network and other classification neural network architectures, a deep layer model is trained according to sample data to serve as the classification neural network model, and the trained classification neural network model has good classification performance, so that the fuzzification effect of the classification neural network model on input vectors is guaranteed.
And step S120, combining and reasoning the fuzzification result by adopting a fuzzy rule to obtain a fuzzy reasoning result.
Specifically, the fuzzy rule base can be configured in advance, and then fuzzy rules in the fuzzy rule base are adopted to combine and reason the fuzzification result obtained in the last step, so that a fuzzy reasoning result is obtained.
And S130, performing defuzzification processing on the fuzzy inference result to obtain the event intention of the current occurrence event.
Specifically, defuzzification processing, namely clarification processing, can be performed on the fuzzy inference result, and the fuzzification result is converted into a determined inference result, namely the event intention of the current occurrence event is obtained.
The event intention reasoning method provided by the embodiment of the application introduces a fuzzy theory and a neural network model, firstly obtains event information of a current occurrence event, then fuzzifies the event information by a pre-trained classification neural network model to obtain a fuzzified result, fuzzifies the event information by using the classification neural network model to enable the fuzzified result to simultaneously have strong feature expression capability of a deep neural model and description and expression capability of the fuzzy information by the fuzzy theory, further combines and reasons the fuzzified result by adopting a fuzzy rule to obtain a fuzzy reasoning result, and finally defuzzifies the fuzzy reasoning result to obtain the event intention of the current occurrence event. Therefore, the method and the device solve the problem of event intention inference of the meaning fuzzy information through the fuzzy theory, combine the fuzzy theory and the neural network model, and use the classification neural network model for fuzzification processing, so that the strong feature expression capability of the deep neural model can be learned, and the fuzzy inference capability of the fuzzy theory is inherited, so that the method and the device are more suitable for inferring the fuzzy event intention, and the inference accuracy of the event intention is improved.
In some embodiments of the present application, the above-mentioned processes of steps S110-S130 may be implemented by a pre-trained fuzzy neural network intention inference model.
Specifically, the fuzzy neural network intention inference model can be trained in advance. In training, the fuzzy neural network intention inference model is trained by using historical event data, wherein the historical event data comprises event information of historical events and real event intentions of the historical events.
Referring to fig. 3, in the present embodiment, the fuzzy neural network intention inference model includes an input layer, a fuzzification layer, a fuzzy inference layer, a deblurring layer, and an output layer.
The input layer inputs the text representation features extracted from the event information.
The fuzzy layer can adopt a stack type self-encoder SAE or other various classification neural network architectures such as a feedforward neural network and the like, and carries out fuzzification processing on an input vector according to a classification neural network model obtained by sample data training. Taking the classification neural network model based on SAE training as an example, the fuzzy layer using the SAE model for fuzzy processing is herein defined as SAE fuzzy layer. And the SAE fuzzification layer fuzzifies the input event information to obtain a fuzzification result.
And the fuzzy reasoning layer combines and reasons the fuzzy results by adopting a fuzzy rule to obtain a fuzzy reasoning result.
Each neuron in the fuzzy inference layer represents a fuzzy rule. In the fuzzy inference layer, a mode of connecting with neuron weights of the fuzzification layer can be used, and the magnitude of the weights represents the contribution degree of the response fuzzy rule to the fuzzy inference result.
And the deblurring layer performs deblurring processing on the fuzzy inference result to obtain the predicted event intention.
Specifically, the deblurring layer may normalize the fuzzy inference result of the previous layer, and the sum of the output results of the fuzzy rules after normalization is 1. Through normalization processing, the corrected amplitude has a specified constraint range when the network carries out reverse parameter adjustment in the subsequent steps, so that large-amplitude oscillation of the network is avoided, and model convergence is facilitated.
Further, the defuzzification layer defuzzifies each normalized rule output result, that is, selects the element with the maximum membership degree in the fuzzy inference result set as an output value, so as to obtain the classification result of the input sample, that is, the intention of the event which occurs at present.
For the back propagation process of the fuzzy neural network intention inference model, each parameter adjustment in the model can use a back propagation algorithm, wherein the parameters to be adjusted mainly comprise weight parameters of a deblurring layer and weight parameters from the deblurring layer to a fuzzy inference layer.
After the fuzzy neural network intention inference model is obtained through training, the event information of the current occurrence event can be input into the model, and the event intention of the current occurrence event is predicted by the model.
In this embodiment, the input vector is fuzzified by using a stacked self-encoder or other classification neural network architecture according to a classification neural network model trained by sample data, so that the constructed fuzzy neural network intention inference model can have strong feature learning capability of a deep learning model and fuzzy inference capability of a fuzzy neural network at the same time.
In some embodiments of the present application, since the fuzzy neural network intends to have a fuzzy inference layer after the fuzzy layer of the inference model, the input vector of the classification neural network model for performing the fuzzy processing in the fuzzy layer may be processed in a different manner from that of the conventional classification neural network.
Conventional classification neural networks give an input vector x (x) for an m-classification problem1,x2...xn) There are n dimensions, and the processing mode is to directly use the weight matrix Wm*nPerforming matrix dot product operation W.x on xTFinally, a corresponding output vector y ═ y (y) is obtained1,y2...ym)TAnd each value in the output vector corresponds to the attribution degree of one category respectively.
In the present application, the process of fuzzifying the event information by the classification neural network model may refer to the following steps:
s1, obtaining the text characteristic of the input event information, wherein the text characteristic is formed by n-dimensional vector x ═ x1,x2,...xn) And (4) forming.
S2, corresponding the text features to n-dimensional vectors (x)1,x2,...xn) And expanded into n pieces of data.
In an alternative manner, each of the n expanded data may have only one-dimensional nonzero value, such as: { (x)1,0...0),(0,x2...0)…(0,0...xn)}. Or, in other implementation manners, one element in each data may be a certain element in the n-dimensional vector, and the other elements are preset values, such as 1,2, and the like. Weight matrix W of subsequent classification neural network modelm*nPerforming matrix dot product operation W.x on the n data after expansionTAnd then, subtracting a preset bias to obtain a fuzzification result.
S3 weight matrix W through classifying neural network modelsm*nPerforming matrix dot product operation W.x on the n data after expansionTAnd obtaining n output vectors as fuzzification results.
Where n output vectors can be represented as:
each of which
N is an m-dimensional vector corresponding to the degree of attribution of each of the m class labels, i.e., the degree of attribution of each of the m class labels
In this embodiment, the input vector is expanded into n data, so that the fuzzy rules in the subsequent fuzzy inference layer determine how each dimension of the input vector is combined to contribute to the final result.
A simple example is used for illustration, for the sake of clear description, it is assumed that a 2-class problem is involved, the dimension of the input vector is 3, a simplified classification neural network structure is used, and only an input layer and an output layer are provided, and then only one weight matrix W is provided, and the values are recorded as:
the traditional processing method of the neural network structure is to process input data x ═ x (x)1,x2,x3) Performing matrix dot product operation W.xTThe results obtained were:
y＝(w11x1+w12x2+w13x3,w21x1+w22x2+w23x3)T
the above y is a two-dimensional vector representing degrees of attribution to the 2 classification categories, respectively.
The present invention is different from the classification neural network model in the processing method, and the classification neural network model expands input data to obtain x { (x) as a result of the expansion1,0,0),(0,x2,0),(0,0,x3)}。
Further, matrix dot product operation W.x is carried out on the expanded result and the weight matrix WTTo obtain 3 output vectors, which are:
y1＝(w11x1,w21x1)T
y2＝(w12x2,w22x2)T
y3＝(w13x3,w23x3)T
the fuzzy inference layer will make combined inference on these output vectors according to fuzzy rules. Obviously, the traditional way of classifying neural networks processes input vectors is to restrict the combination rule to 3 output vectors for direct summation, since y is y1+y2+y3From this point of view, the conventional classification neural network is equivalent to the special case of directly adding all output vectors with the fuzzy rule fixed.
Further, for the fuzzy inference layer in the fuzzy neural network intention inference model, a fuzzy rule is adopted to combine and infer the fuzzified result, which may be that the fuzzy rule is adopted to combine and infer the n output vectors output by the classification neural network model.
Each neuron in the fuzzy inference layer represents a fuzzy rule, and for the selection of the quantity of the fuzzy rules, the traditional fuzzy neural network model has the problem of dimension disaster. Therefore, in this embodiment, n dimensions in the input data may be clustered, and the number of clustering centers is obtained as the number of fuzzy rules. That is, the number of fuzzy rules in the fuzzy inference layer is the same as the number of clustering centers after n dimensions are clustered in the n-dimensional vector corresponding to the text feature of the event information.
In some embodiments of the present application, there is further provided another implementation of inferring an intent of an event.
In practical situations, events are not independent but rather are a complex network of interconnected events, and the follow-up events of an event are often not only one. An event alone does not truly reflect the true intent, i.e., intent inference can also involve the evolution of the event, and if the rich association information between events is not sufficiently utilized, the true intent behind the actionable event cannot be discerned.
Therefore, in the embodiment, the subsequent possible events of the current event are judged, and the subsequent possible events are also used as a part of influence factors when the intention of the current event is inferred. On the basis, the relevance among the events is modeled, the dynamic change characteristics of the events are considered, and the true intention of the events is inferred and predicted.
In the embodiment, for dynamically changing events, event extraction is performed based on historical event data, an association relationship is established by using the extracted events, and event intention inference is performed by using event association.
The process of extracting events based on historical event data may refer to the following descriptions:
the Event extraction can be divided into two types, namely Meta Event (Meta Event) extraction and Topic Event (Topic Event) extraction. The embodiment of the present application specifically refers to meta-event extraction, that is, an event is composed of an event Trigger (Trigger) and a parameter (Argument) describing an event structure. Event-triggered words, i.e., core words that represent the occurrence of an event, are mostly verbs or nouns. The parameter (Argument) is also called an event Argument, and represents attribute information of a participant, time, place, and the like of the event. Thus, the event extraction becomes a task of event identification and event parameter identification, and the event parameter identification extracts corresponding parameters according to a preset event template, which is similar to a slot filling task and can be called argument identification. In terms of meta-event extraction, event extraction methods including pattern matching based, machine learning based, neural network based, and the like may be employed in embodiments of the present application.
Taking an event extraction method based on a neural network as an example for explanation, in the embodiment, event extraction is converted into a supervised multivariate classification task, and the method specifically comprises sentence-level event extraction, and a BERT + CRF neural network event extraction method based on a joint model and fusing external resources.
Specifically, in this embodiment, a word vector may be trained by using a BERT pre-training language model, so that text semantic information is completely stored, the context bidirectional feature extraction capability of the model is improved, and the problems of event trigger word extraction and boundary division of parameter content are well solved. Compared with the traditional BilSTM model, the semantic information is more fully utilized, and the recognition rate of the model to the element content is improved.
The present embodiment adopts a BERT pre-training language model, and the structure is shown in fig. 4.
Wherein E is1,E2,...,EnAs an input vector to the model, T1,T2,...,TnIs the output vector of the model.
Conditional Random Fields (CRF) are an algorithm in the task of sequence labeling, in which linear chain element random fields can be used, and belong to a discriminant model for predicting output sequences from input sequences. For a given sequence X ═ X1,x2,...xn) The corresponding label is Y ═ Y1,y2,...yn) If the following conditions are satisfied:
P(Yi|X,Y1,...,Yi-1,Yi+1,...,Yn)＝P(Yi|X,Yi-1,Yi+1)
let P (N, K) be the weight matrix of the decoded layer output, and then the evaluation score S (x, y), i.e.
Wherein A is a transfer matrix, k is the number of labels, and n is the sequence length. The maximum probability of a sequence tag y can be calculated using the SoftMax function, i.e.:
the maximum posterior probability of P (y | x) is solved by a maximum likelihood method during training:
the training and decoding of the linear conditional random field may use the Viterbi algorithm.
In the embodiment, a BERT + CRF neural network model is used for event extraction, each single character in historical event data is encoded by using a BERT pre-training language model to obtain a word vector corresponding to the single character, the word vector containing context information is input into a CRF layer to be decoded, the CRF layer can output a label sequence with the maximum probability to obtain the category of each character, and the category label can comprise various trigger words of different types and different parameters. Based on the above, the meta-event extraction process of the historical event data can be completed, and each extracted event can be obtained.
After the event extraction is completed on the historical event data, in order to better utilize the complex association information between the extracted events, an event association map is constructed in the embodiment, and then the inference event which is most likely to occur after the current occurrence event is inferred based on the event association map.
Points in the event association graph represent events, directed edges among the points represent association relations of the events, and weights on the directed edges represent the frequency of the association relations in historical event data.
Referring to fig. 5, an event correlation map diagram is illustrated.
The event correlation map illustrated in fig. 5 includes five events ABCDE, where three events BCD have strong correlation and form a ring structure, which is implicit connection information and provides more information more effectively when learning the event representation.
After extracting events based on historical event data and establishing an event correlation map, the present embodiment may determine inference events that may occur after the current occurrence event based on the event correlation map. On the basis, the event intention of the current occurrence event is determined based on the event type of the inference event.
It should be noted that, in this embodiment, an inference event that may occur after a current occurrence event can be determined based on the event correlation map, and then an event intention of the current occurrence event can be determined, and the event intention is defined as a second event intention. In addition, the event intention of the currently occurring event determined based on the fuzzy neural network intention inference model in the foregoing embodiment is defined as the first event intention. On the basis, the final event intention of the current event can be determined by combining the first event intention and the second event intention in the embodiment of the application.
In an optional implementation manner, the attention degree of the user to different event intentions may be preset, on this basis, the final weight of the first event intention may be determined based on the confidence degree of the first event intention and the attention degree thereof, and similarly, the final weight of the second event intention may be determined based on the confidence degree of the second event intention and the attention degree thereof. And selecting the event intention with the maximum final weight as the final event intention of the current event.
In some embodiments of the present application, in the process of determining inference events that may occur after a current occurrence event based on the event correlation map, reference may be made to the following implementation steps:
and S1, acquiring an event inference prediction model trained on the event correlation map.
Specifically, after the event correlation map is established, the event inference prediction model can be obtained by training and learning on the map by using a neural network model.
The event inference prediction model is configured to predict a state representation of a correlation between pairs of events based on the input pairs of events. That is, the event inference prediction model can learn the representation of each event in the event association graph and the association relationship between events. On the basis, for the input pair of events, the event inference prediction model can predict the correlation degree between the input pair of events.
And S2, determining a candidate event corresponding to the current occurrence event based on the event correlation map.
Specifically, the candidate event is an event that may occur after the current event occurs.
In the simplest way, each event in the event correlation map can be directly used as a candidate event which is possibly generated after the current event.
In addition, the embodiment of the present application further provides another way of determining a candidate event, specifically:
and S21, determining a target point corresponding to the current occurrence event in the event correlation map.
And S22, searching points directly or indirectly pointed by the target point according to the set searching distance by taking the target point as a center, and forming candidate events by the events corresponding to the searched points.
The set search distance may be a set maximum jump number, and if a maximum jump from the target point to the outside is set three times, a candidate event is formed based on events corresponding to the searched points.
And S3, respectively forming event pairs by each candidate event and the current occurrence event, and inputting the event pairs into the event inference prediction model to obtain the correlation degree between each candidate event and the current occurrence event output by the model.
Specifically, for each candidate event of the determined current occurrence event, each candidate event and the current occurrence event are combined into an event pair input event inference prediction model, and the correlation degree between each candidate event output by the model and the current occurrence event is obtained.
And S4, determining inference events from each candidate event based on the correlation.
In an alternative implementation manner, the candidate event with the largest correlation degree may be directly selected as the inference event most likely to occur after the current occurrence event.
In another optional implementation manner, a weighted value on a directed edge between events in the event correlation map may be further considered, the correlation between each candidate event and the current occurrence event is weighted, and an inference event is selected according to the final weighted correlation. Specific implementation steps can be referred to as follows:
and S41, determining the average value of the weights on the directional side link between each candidate event and the current occurrence event in the event correlation map, and taking the average value as the average weight between each candidate event and the current occurrence event.
Specifically, for each candidate event, a shortest path between a target point corresponding to the current occurrence event and a point corresponding to the candidate event is searched in the event correlation map, and weights on directed side links forming the shortest path are averaged, and the result is used as an average weight between the candidate event and the current occurrence event.
S42, weighting the correlation degree between the candidate event and the current occurrence event by using the average weight between the candidate event and the current occurrence event to obtain the weighted correlation degree between each candidate event and the current occurrence event.
And S43, selecting the candidate event with the highest weighted correlation as the inference event.
In the above embodiments of the present application, two different ways of determining the event intention of the currently occurring event are described. Although the inferred event intention of the currently occurring event can be obtained for the user, the event intention is not sufficiently interpretable. Examples are as follows:
currently, a piece of military event information is acquired: "10 Yue 13 late to 14 early morning, the A Navy 1 EP-3E scout with the number of 123456 takes off from the X base and carries out the action over the air in the sea area in the southwest of the Y island".
By adopting the event intention inference scheme of any one of the previous embodiments of the application, the intention of the military event can be inferred to be 'scout intelligence'.
For the user, the user may not understand why the "spy intelligence" event intention can be inferred through the military event information, that is, the event intention inference process is invisible to the user, so that the interpretability of the corresponding event intention may not be strong enough.
In order to solve the problem, in the embodiment of the application, the knowledge used in the intention reasoning process is described in a book-oriented mode, so that the reasoning process is visualized, and the interpretability of the reasoning event intention is increased.
Specifically, the interpretable text generation model is trained in advance, and the model is obtained by taking event information and event intention of a training event and field knowledge of the field to which the event intention belongs as training samples and taking interpretable text corresponding to the training event as sample labels.
And generating a model based on the trained interpretable text, and processing the event information of the current occurrence event, the event intention and the domain knowledge of the domain to which the current occurrence event belongs to obtain the interpretable text of the event intention output by the model.
Wherein, the event intention in the input interpretable text generation model can be the event intention determined in any one of the previous embodiments. The domain knowledge of the domain to which the current occurrence event belongs may include domain keywords included in the event information of the current occurrence event, and a preset event intention inference rule, where the event intention inference rule may be generated in advance by a model or set manually, and the application is not limited strictly.
For example, the event information of the current occurrence event is as follows: "10 Yue 13 late to 14 early morning, the A Navy 1 EP-3E scout with the number of 123456 takes off from the X base and carries out the action over the air in the sea area in the southwest of the Y island". The event determined for the current occurrence is intended to be "scout intelligence".
The "EP-3E scout", "above sea area", "early morning", and "Y island" in the event information may be used as the domain keywords.
In addition, the preset event intention inference rule may include: "the enemy plane moving from night to early morning stage is generally an outgoing scout, the task executed by the scout is probably a scout task, and the enemy plane going in and out of the sensitive area is often used for searching the intelligence", etc.
Based on this, the domain keywords and the artificial rules can be used as domain knowledge, and the interpretable text generating model processes the event information, the event intention and the domain knowledge to obtain the interpretable text of the event intention of the reconnaissance intelligence, which is output by the model, as follows:
since the EP-3E aircraft is a scout and visits the sensitive area Y island in the morning, the EP-3E aircraft is intended to be "scout intelligence".
Obviously, according to the scheme of the application, the interpretable text of the event intention can be generated by combining the domain knowledge, so that the interpretability of the event intention of the current occurrence event obtained by inference is higher.
In some embodiments of the application, the processing, by using an interpretable text generating model, the event information of the current occurrence event, the event intention, and domain knowledge in the field to which the current occurrence event belongs to obtain the interpretable text may specifically include:
and S1, coding the event information by using the interpretable text generation model to obtain an event hidden layer state, and coding the domain knowledge to obtain a domain knowledge hidden layer state.
And S2, performing attention processing on the event hidden layer state and the domain knowledge hidden layer state respectively by using an interpretable text generation model to obtain an event attention representation and a domain knowledge attention representation.
And S3, processing the event attention expression and the domain knowledge attention expression by utilizing an interpretable text generation model to obtain an interpretable text of the event intention.
In conjunction with the interpretable text generating model illustrated in fig. 6, the interpretable text generating model provided in the embodiment of the present application has two kinds of input, namely, a first encoder for encoding event information and a second encoder for encoding domain knowledge. The first encoder may employ bi-directional LSTM to encode semantic information of the event information, with hidden states of the LSTM to characterize the event information.
The second encoder may employ a bi-directional recurrent neural network and a threshold recurrent unit GMU. When the domain knowledge comprises domain keywords and artificial rules, the second encoder encodes the domain keywords and the artificial rules respectively, splices encoding results, and represents domain knowledge information by using a hidden layer state output by the last layer of the bidirectional cyclic neural network.
It will be appreciated that the resulting interpretable text is typically associated with only a small portion of the event information or only a small portion of the domain knowledge information. Thus, embodiments of the present application may use an attention mechanism to generate a high-level representation of event attention and domain knowledge attention. Finally, the overall expression vector c is obtainediAnd then generating interpretable text for the intent of the event.
The event intention inference device provided by the embodiment of the application is described below, and the event intention inference device described below and the event intention inference method described above may be referred to correspondingly.
Referring to fig. 7, fig. 7 is a schematic structural diagram of an event intention inference device disclosed in the embodiment of the present application.
As shown in fig. 7, the apparatus may include:
an event information acquiring unit 11 configured to acquire event information of a currently occurring event;
the fuzzification processing unit 12 is configured to perform fuzzification processing on the event information by using a pre-trained classification neural network model to obtain a fuzzification result;
the fuzzy reasoning unit 13 is used for combining and reasoning the fuzzification result by adopting a fuzzy rule to obtain a fuzzy reasoning result;
and the defuzzification processing unit 14 is configured to perform defuzzification processing on the fuzzy inference result to obtain an event intention of the current event.
Optionally, the functions of the fuzzification processing unit, the fuzzy inference unit and the defuzzification processing unit may be implemented by a fuzzy neural network intention inference model unit, specifically:
inputting the event information into a pre-trained fuzzy neural network intention inference model by a fuzzy neural network intention inference model unit to obtain the event intention of the current occurrence event predicted by the model; wherein,
and the fuzzy layer of the fuzzy neural network intention inference model performs fuzzification processing on the event information by using a pre-trained classification neural network model, the fuzzy inference layer combines and infers fuzzy results by adopting a fuzzy rule to obtain a fuzzy inference result, and the defuzzification layer performs defuzzification processing on the fuzzy inference result to obtain the predicted event intention.
Optionally, the process of performing the fuzzification processing on the event information by the fuzzification processing unit using the pre-trained classification neural network model may include:
acquiring text features of the event information, wherein the text features are composed of n-dimensional vectors (x)1,x2,...xn) Composition is carried out;
corresponding the text features to n-dimensional vectors (x)1,x2,...xn) Expanding the data into n data;
weight matrix W by classifying neural network modelsm*nAnd performing matrix dot product operation on the n expanded data to obtain n output vectors as fuzzification results, wherein each output vector is an m-dimensional vector and corresponds to the attribution degree of the m class labels.
Optionally, each of the n expanded data may have only one-dimensional non-zero, such as { (x)1,0...0),(0,x2...0)...(0,0...xn)}。
Optionally, the fuzzy inference unit adopts a fuzzy rule to combine and infer the fuzzification result, and the process may include:
combining and reasoning said n output vectors using fuzzy rules, wherein,
the number of the fuzzy rules is the same as the number of clustering centers after n dimensions are clustered in n-dimensional vectors corresponding to the text features.
Optionally, the apparatus of the present application may further include:
the event correlation map acquisition unit is used for acquiring an event correlation map established based on historical event data, the middle points of the event correlation map represent events, directed edges between the points represent the incidence relation of the events, and the weight on the directed edges represents the occurrence frequency of the incidence relation in the historical event data;
the event intention reasoning unit is used for determining a reasoning event which is possibly generated after the current occurrence event based on the event correlation map and determining the event intention of the current occurrence event based on the event type of the reasoning event;
and the event intention comprehensive processing unit is used for determining the final event intention of the current event based on the event intention of the current event determined based on the inference event and the event intention of the current event obtained after the defuzzification processing.
Optionally, the process of determining, by the event intention inference unit, an inference event that may occur after the current occurrence event based on the event association map may include:
obtaining an event inference prediction model trained based on the event correlation map, the event inference prediction model configured to predict a state representation of a degree of correlation between event pairs based on input event pairs;
determining a candidate event corresponding to the current occurrence event based on the event correlation map, wherein the candidate event is an event which may occur after the current occurrence event;
respectively forming an event pair by each candidate event and the current occurrence event, and inputting the event pair into the event inference prediction model to obtain the correlation degree between each candidate event and the current occurrence event output by the model;
and determining inference events from each of the candidate events based on the correlation.
Optionally, the process of determining the candidate event corresponding to the current occurrence event by the event intention inference unit based on the event association map may include:
determining a target point corresponding to the current occurrence event in the event correlation map;
and searching points directly or indirectly pointed by the target point according to a set searching distance by taking the target point as a center, and forming a candidate event by the events corresponding to the searched points.
Optionally, the process of determining an inference event from each candidate event by the event intention inference unit based on the correlation degree may include:
determining the average value of each weight on a directed edge link between each candidate event and the current occurrence event in the event correlation map, and taking the average value as the average weight between each candidate event and the current occurrence event;
weighting the correlation degree between the candidate event and the current occurrence event by using the average weight between the candidate event and the current occurrence event to obtain the weighted correlation degree between each candidate event and the current occurrence event;
and selecting the candidate event with the highest weighted correlation as the reasoning event.
Optionally, the apparatus of the present application may further include:
the interpretable text generating unit is used for processing the event information of the current occurrence event, the event intention and the domain knowledge of the domain to which the current occurrence event belongs by utilizing a pre-trained interpretable text generating model to obtain an interpretable text output by the model and aiming at the event intention;
the interpretable text generation model is obtained by taking event information and event intention of a training event and field knowledge in the field to which the training event belongs as training samples and taking interpretable text corresponding to the training event as sample labels for training.
Optionally, the process of processing the event information of the current occurrence event, the event intention, and the domain knowledge of the domain to which the current occurrence event belongs by using the pre-trained interpretable text generation model by the interpretable text generation unit may include:
respectively coding event information by using an interpretable text generation model to obtain an event hidden layer state, and coding domain knowledge to obtain a domain knowledge hidden layer state;
respectively carrying out attention processing on the event hidden layer state and the domain knowledge hidden layer state by utilizing an interpretable text generation model to obtain an event attention representation and a domain knowledge attention representation;
and processing the event attention representation and the domain knowledge attention representation by using an interpretable text generation model to obtain an interpretable text of the event intention.
The event intention reasoning device provided by the embodiment of the application can be applied to event intention reasoning equipment, such as a terminal: mobile phones, computers, etc. Alternatively, fig. 8 shows a block diagram of a hardware structure of the event intention inference device, and referring to fig. 8, the hardware structure of the event intention inference device may include: at least one processor 1, at least one communication interface 2, at least one memory 3 and at least one communication bus 4;
in the embodiment of the application, the number of the processor 1, the communication interface 2, the memory 3 and the communication bus 4 is at least one, and the processor 1, the communication interface 2 and the memory 3 complete mutual communication through the communication bus 4;
the processor 1 may be a central processing unit CPU, or an application Specific Integrated circuit asic, or one or more Integrated circuits configured to implement embodiments of the present invention, etc.;
the memory 3 may include a high-speed RAM memory, and may further include a non-volatile memory (non-volatile memory) or the like, such as at least one disk memory;
wherein the memory stores a program and the processor can call the program stored in the memory, the program for:
acquiring event information of a current occurrence event;
fuzzifying the event information by using a pre-trained classification neural network model to obtain a fuzzification result;
combining and reasoning the fuzzification result by adopting a fuzzy rule to obtain a fuzzy reasoning result;
and performing defuzzification processing on the fuzzy inference result to obtain the event intention of the current occurrence event.
Alternatively, the detailed function and the extended function of the program may be as described above.
Embodiments of the present application further provide a storage medium, where a program suitable for execution by a processor may be stored, where the program is configured to:
acquiring event information of a current occurrence event;
fuzzifying the event information by using a pre-trained classification neural network model to obtain a fuzzification result;
combining and reasoning the fuzzification result by adopting a fuzzy rule to obtain a fuzzy reasoning result;
and performing defuzzification processing on the fuzzy inference result to obtain the event intention of the current occurrence event.
Alternatively, the detailed function and the extended function of the program may be as described above.
Finally, it should also be noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.
The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, the embodiments may be combined as needed, and the same and similar parts may be referred to each other.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims (14)
1. An event intention inference method, comprising:
acquiring event information of a current occurrence event;
fuzzifying the event information by using a pre-trained classification neural network model to obtain a fuzzification result;
combining and reasoning the fuzzification result by adopting a fuzzy rule to obtain a fuzzy reasoning result;
and performing defuzzification processing on the fuzzy inference result to obtain the event intention of the current occurrence event.
2. The method of claim 1, wherein the process of fuzzifying the event information, combining and reasoning the fuzzification results, and defuzzifying comprises:
inputting the event information into a pre-trained fuzzy neural network intention inference model to obtain the event intention of the current occurrence event predicted by the model; wherein,
and the fuzzy layer of the fuzzy neural network intention inference model performs fuzzification processing on the event information by using a pre-trained classification neural network model, the fuzzy inference layer combines and infers fuzzy results by adopting a fuzzy rule to obtain a fuzzy inference result, and the defuzzification layer performs defuzzification processing on the fuzzy inference result to obtain the predicted event intention.
3. The method of claim 1, wherein the fuzzifying the event information with a pre-trained classification neural network model comprises:
acquiring text features of the event information, wherein the text features are composed of n-dimensional vectors (x)1,x2,...xn) Composition is carried out;
corresponding the text features to n-dimensional vectors (x)1,x2,...xn) Expanding the data into n data;
weight matrix W by classifying neural network modelsm*nAnd performing matrix dot product operation on the n expanded data to obtain n output vectors as fuzzification results, wherein each output vector is an m-dimensional vector and corresponds to the attribution degree of the m class labels.
4. The method of claim 3, wherein the combining and reasoning the fuzzification results using the fuzzy rule comprises:
combining and reasoning said n output vectors using fuzzy rules, wherein,
the number of the fuzzy rules is the same as the number of clustering centers after n dimensions are clustered in n-dimensional vectors corresponding to the text features.
5. The method of claim 1, further comprising:
acquiring an event correlation map established based on historical event data, wherein points in the event correlation map represent events, directed edges between the points represent incidence relations of the events, and weights on the directed edges represent the occurrence frequency of the incidence relations in the historical event data;
determining inference events which are possibly generated after the current occurrence event based on the event correlation map, and determining an event intention of the current occurrence event based on the event type of the inference events;
and determining the final event intention of the current occurrence event by combining the event intention of the current occurrence event determined based on the inference event and the event intention of the current occurrence event obtained after the defuzzification processing.
6. The method according to claim 5, wherein the determining inference events that are likely to occur after the currently occurring event based on the event correlation graph comprises:
obtaining an event inference prediction model trained based on the event correlation map, the event inference prediction model configured to predict a state representation of a degree of correlation between event pairs based on input event pairs;
determining a candidate event corresponding to the current occurrence event based on the event correlation map, wherein the candidate event is an event which may occur after the current occurrence event;
respectively forming an event pair by each candidate event and the current occurrence event, and inputting the event pair into the event inference prediction model to obtain the correlation degree between each candidate event and the current occurrence event output by the model;
and determining inference events from each of the candidate events based on the correlation.
7. The method according to claim 6, wherein the determining a candidate event corresponding to the currently occurring event based on the event correlation map comprises:
determining a target point corresponding to the current occurrence event in the event correlation map;
and searching points directly or indirectly pointed by the target point according to a set searching distance by taking the target point as a center, and forming a candidate event by the events corresponding to the searched points.
8. The method of claim 6, wherein determining inference events from each of the candidate events based on the correlation comprises:
determining the average value of each weight on a directed edge link between each candidate event and the current occurrence event in the event correlation map, and taking the average value as the average weight between each candidate event and the current occurrence event;
weighting the correlation degree between the candidate event and the current occurrence event by using the average weight between the candidate event and the current occurrence event to obtain the weighted correlation degree between each candidate event and the current occurrence event;
and selecting the candidate event with the highest weighted correlation as the reasoning event.
9. The method according to any one of claims 1-8, further comprising:
generating a model by utilizing a pre-trained interpretable text, processing the event information of the current occurrence event, the event intention and the domain knowledge of the domain to which the current occurrence event belongs, and obtaining the interpretable text of the event intention output by the model;
the interpretable text generation model is obtained by taking event information and event intention of a training event and field knowledge in the field to which the training event belongs as training samples and taking interpretable text corresponding to the training event as sample labels for training.
10. The method of claim 9, wherein processing event information of the current occurrence, the event intent, and domain knowledge of a domain to which the current occurrence belongs using a pre-trained interpretable text generation model comprises:
respectively coding event information by using an interpretable text generation model to obtain an event hidden layer state, and coding domain knowledge to obtain a domain knowledge hidden layer state;
respectively carrying out attention processing on the event hidden layer state and the domain knowledge hidden layer state by utilizing an interpretable text generation model to obtain an event attention representation and a domain knowledge attention representation;
and processing the event attention representation and the domain knowledge attention representation by using an interpretable text generation model to obtain an interpretable text of the event intention.
11. The method according to claim 9, wherein the domain knowledge includes domain keywords included in the event information and a preset event intention inference rule.
12. An event intention inference apparatus, comprising:
the event information acquisition unit is used for acquiring the event information of the current occurrence event;
the fuzzification processing unit is used for fuzzifying the event information by using a pre-trained classification neural network model to obtain a fuzzification result;
the fuzzy reasoning unit is used for combining and reasoning the fuzzification result by adopting a fuzzy rule to obtain a fuzzy reasoning result;
and the defuzzification processing unit is used for performing defuzzification processing on the fuzzy inference result to obtain the event intention of the current occurrence event.
13. An event intention inference device, comprising: a memory and a processor;
the memory is used for storing programs;
the processor, which is used for executing the program, realizes the steps of the event intention reasoning method as claimed in any one of claims 1-11.
14. A storage medium having stored thereon a computer program, characterized in that the computer program, when being executed by a processor, carries out the steps of the event intention inference method according to any one of claims 1-11.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011460065.6A CN112488316B (en) | 2020-12-11 | 2020-12-11 | Event intention reasoning method, device, equipment and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011460065.6A CN112488316B (en) | 2020-12-11 | 2020-12-11 | Event intention reasoning method, device, equipment and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112488316A true CN112488316A (en) | 2021-03-12 |
| CN112488316B CN112488316B (en) | 2022-12-02 |
Family
ID=74916338
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011460065.6A Active CN112488316B (en) | 2020-12-11 | 2020-12-11 | Event intention reasoning method, device, equipment and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112488316B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113590879A (en) * | 2021-08-05 | 2021-11-02 | 哈尔滨理工大学 | System, method, computer and storage medium for shortening timestamp and solving multi-event video question-answering through network |
| CN115018472A (en) * | 2022-08-03 | 2022-09-06 | 中国电子科技集团公司第五十四研究所 | Interactive incremental information analysis system based on interpretable mechanism |
| CN117909507A (en) * | 2024-03-19 | 2024-04-19 | 金盾检测技术股份有限公司 | AI-based data classification system |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5524176A (en) * | 1993-10-19 | 1996-06-04 | Daido Steel Co., Ltd. | Fuzzy expert system learning network |
| US6078911A (en) * | 1991-10-15 | 2000-06-20 | General Electric Company | Compilation of rule bases for fuzzy logic control |
| CN102323751A (en) * | 2011-06-28 | 2012-01-18 | 浙江大学 | Preparatory grinding system control method based on fuzzy intelligence control and optimization method |
| CN102866321A (en) * | 2012-08-13 | 2013-01-09 | 广东电网公司电力科学研究院 | Self-adaptive stealing-leakage prevention diagnosis method |
| US20130103630A1 (en) * | 2009-08-19 | 2013-04-25 | Bae Systems Plc | Fuzzy inference methods, and apparatuses, systems and apparatus using such inference apparatus |
| US20160055427A1 (en) * | 2014-10-15 | 2016-02-25 | Brighterion, Inc. | Method for providing data science, artificial intelligence and machine learning as-a-service |
| CN108665070A (en) * | 2018-05-16 | 2018-10-16 | 深圳大学 | Limit TS fuzzy reasoning methods based on extreme learning machine and system |
| CN110705812A (en) * | 2019-04-15 | 2020-01-17 | 中国石油大学(华东) | Industrial fault analysis system based on fuzzy neural network |
-
2020
- 2020-12-11 CN CN202011460065.6A patent/CN112488316B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6078911A (en) * | 1991-10-15 | 2000-06-20 | General Electric Company | Compilation of rule bases for fuzzy logic control |
| US5524176A (en) * | 1993-10-19 | 1996-06-04 | Daido Steel Co., Ltd. | Fuzzy expert system learning network |
| US20130103630A1 (en) * | 2009-08-19 | 2013-04-25 | Bae Systems Plc | Fuzzy inference methods, and apparatuses, systems and apparatus using such inference apparatus |
| CN102323751A (en) * | 2011-06-28 | 2012-01-18 | 浙江大学 | Preparatory grinding system control method based on fuzzy intelligence control and optimization method |
| CN102866321A (en) * | 2012-08-13 | 2013-01-09 | 广东电网公司电力科学研究院 | Self-adaptive stealing-leakage prevention diagnosis method |
| US20160055427A1 (en) * | 2014-10-15 | 2016-02-25 | Brighterion, Inc. | Method for providing data science, artificial intelligence and machine learning as-a-service |
| CN108665070A (en) * | 2018-05-16 | 2018-10-16 | 深圳大学 | Limit TS fuzzy reasoning methods based on extreme learning machine and system |
| CN110705812A (en) * | 2019-04-15 | 2020-01-17 | 中国石油大学(华东) | Industrial fault analysis system based on fuzzy neural network |
Non-Patent Citations (3)
| Title |
|---|
| CHANG GYOON LIM ET AL: "A Performance Analysis of User’s Intention Classification from EEG Signal by a Computational Intelligence in BCI", 《ICMLSC 2018》 * |
| 吴旭等: "基于模糊聚类和模糊推理的电网连锁故障预警方法", 《电网技术》 * |
| 张颖等: "基于规则的模糊离散事件系统建模与控制研究", 《控制与决策》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113590879A (en) * | 2021-08-05 | 2021-11-02 | 哈尔滨理工大学 | System, method, computer and storage medium for shortening timestamp and solving multi-event video question-answering through network |
| CN113590879B (en) * | 2021-08-05 | 2022-05-31 | 哈尔滨理工大学 | System, method, computer and storage medium for shortening timestamp and solving multi-event video question-answering through network |
| CN115018472A (en) * | 2022-08-03 | 2022-09-06 | 中国电子科技集团公司第五十四研究所 | Interactive incremental information analysis system based on interpretable mechanism |
| CN115018472B (en) * | 2022-08-03 | 2022-11-11 | 中国电子科技集团公司第五十四研究所 | Interactive incremental information analysis system based on interpretable mechanism |
| CN117909507A (en) * | 2024-03-19 | 2024-04-19 | 金盾检测技术股份有限公司 | AI-based data classification system |
| CN117909507B (en) * | 2024-03-19 | 2024-05-17 | 金盾检测技术股份有限公司 | AI-based data classification system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112488316B (en) | 2022-12-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112488316B (en) | Event intention reasoning method, device, equipment and storage medium | |
| CN112084331B (en) | Text processing and model training method and device, computer equipment and storage medium | |
| CN100504856C (en) | Method and apparatus for characterizing documents based on clusters of related words | |
| CN111382868B (en) | Neural network structure searching method and device | |
| CN113158554B (en) | Model optimization method and device, computer equipment and storage medium | |
| Ra et al. | DeepAnti-PhishNet: Applying deep neural networks for phishing email detection | |
| Al Iqbal et al. | Knowledge based decision tree construction with feature importance domain knowledge | |
| CN112418525A (en) | Method and device for predicting social topic group behaviors and computer storage medium | |
| CN113011471A (en) | Social group dividing method, social group dividing system and related devices | |
| CN116781346A (en) | Convolution two-way long-term and short-term memory network intrusion detection method based on data enhancement | |
| CN113627550A (en) | Image-text emotion analysis method based on multi-mode fusion | |
| Rogova et al. | Considerations of context and quality in information fusion | |
| CN116227467A (en) | Model training method, text processing method and device | |
| Souha et al. | Pre-trained models for intent classification in chatbot: Comparative study and critical analysis | |
| US11983633B2 (en) | Machine learning predictions by generating condition data and determining correct answers | |
| Song et al. | Spammer detection using graph-level classification model of graph neural network | |
| CN112486467B (en) | Interactive service recommendation method based on dual interaction relation and attention mechanism | |
| Alatas et al. | Mining fuzzy classification rules using an artificial immune system with boosting | |
| CN113010772B (en) | Data processing method, related equipment and computer readable storage medium | |
| CN114093447A (en) | Data asset recommendation method and device, computer equipment and storage medium | |
| Cepero-Pérez et al. | Proactive forest for supervised classification | |
| Makarova et al. | A case-based reasoning approach with fuzzy linguistic rules: Accuracy validation and application in interface design-support intelligent system | |
| Shao et al. | Feature-based inductive transfer learning through minimum encoding | |
| Nalamothu | Abusive and hate speech tweets detection with text generation | |
| Mazumder et al. | Performance Evaluation of Different Word Embedding Techniques Across Machine Learning and Deep Learning Models |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |