CN109902283B - An information output method and device - Google Patents

Abstract

The application discloses an information output method and device, wherein the method comprises the following steps: acquiring a fault description text, wherein the fault description text is used for describing faults occurring in a network; generating a semantic vector of the fault description text through a semantic generation model; acquiring semantic vectors respectively corresponding to related texts of various types of target data, wherein the target data are used for assisting in analyzing reasons of fault generation; calculating the correlation between the semantic vector of the fault description text and the semantic vector of the related text of each target data; determining and outputting first data, wherein the first data is target data with the largest correlation between semantic vectors in various target data and semantic vectors of fault description texts, or the first data is target data with the correlation between semantic vectors in various target data and semantic vectors of fault description texts larger than a preset threshold value. By implementing the method, the data related to the fault description text and used for assisting in analyzing the fault cause can be accurately found out.

Description

An information output method and device

technical field

The present application relates to the field of communication technologies, and in particular to an information output method and device.

Background technique

When network equipment fails, it will affect normal communication and bring serious losses to people's work and life, so timely repair of network equipment failure is very important. At present, when a network device fails, front-line engineers will collect data from the fault site to assist in analyzing the cause of the fault, for example, collecting key performance indicators (KPIs), device alarms, and device logs within a period of time before and after the network device fault occurs and other parameter data. And the first-line engineers will describe the fault phenomenon and get the fault description text. Front-line engineers feed back the collected KPI and other data and fault description text to the operation and maintenance department in the form of fault work orders. According to the fault description text in the fault work order, the operation and maintenance engineer manually selects some KPIs, device alarms, device logs and other parameter data from the data collected by the front line with their own professional knowledge. Further, abnormality detection and mutual corroboration are performed on the selected data, so as to analyze the root cause of the fault and provide guidance for repairing faulty network equipment. This fault detection method, which manually selects parameter data related to the fault description text from KPI, device alarm, device log and other parameter data for viewing and analysis, has low efficiency and slow speed, and cannot meet the increasing network demand.

In the prior art, the text with the same keywords as the fault description text is searched, and the fault is checked and analyzed according to the relevant parameter data of the text. However, there may not be the same keywords in the relevant text that can be used to assist in analyzing the cause of the fault and the fault description text with high correlation. Therefore, the data associated with the fault description text and used to assist in analyzing the cause of the fault cannot be accurately found through the existing methods.

Contents of the invention

The present application provides an information output method and device, which can automatically and accurately find data related to the fault description text to assist in analyzing the cause of the fault.

In the first aspect, the present application provides an information output method, the method includes: obtaining a fault description text, which is used to describe a fault occurring in the network; generating a semantic vector of the fault description text through a semantic generation model; obtaining multiple Semantic vectors corresponding to the relevant texts of each type of target data, the target data is used to assist in analyzing the cause of the fault; calculate the correlation between the semantic vectors of the fault description text and the semantic vectors of the relevant texts of each target data; determine and Output the first data, the first data is the target data with the greatest correlation between the semantic vector and the semantic vector of the fault description text in each target data, or the first data is the correlation between the semantic vector and the fault description text in each target data The target data for which the relevance of the semantic vector is greater than a preset threshold.

The present application can accurately find the target data associated with the fault description text by comparing the correlation between the semantic vector of the fault description text and the semantic vector of the relevant text of the target data. For example, the fault is described as "slow Internet access by industry users", and the name of the key indicator for fault analysis related to it analyzed in this application is "downlink bandwidth control packet loss ratio". It can be seen that there is literally no matching and correlation between the two, but this application has learned the domain knowledge of "Internet speed and packet loss ratio are related" through semantic analysis and mining, and realized the correlation between the two. linked analysis. Therefore, by implementing the method described in the first aspect, data related to the fault description text for assisting in analyzing the cause of the fault can be automatically and accurately found.

In a possible implementation manner, before acquiring the fault description text, semantic vectors corresponding to relevant texts of multiple types of target data may also be generated through a semantic generation model.

And the semantic vectors corresponding to the relevant texts of various target data can also be saved respectively; correspondingly, the specific implementation manner of obtaining the semantic vectors respectively corresponding to the relevant texts of various target data is: obtaining the relevant texts of various target data stored respectively Corresponding semantic vector.

By implementing this embodiment, the semantic vectors corresponding to the relevant texts of various target data can be generated and saved in advance, and after receiving the fault description text, the semantic vectors corresponding to the relevant texts of various target data can be directly used to match the fault Correlation calculation is performed on the semantic vectors of the description text, so that it is not necessary to temporarily generate semantic vectors corresponding to the relevant texts of various target data after receiving the fault description text. It can be seen that by implementing this embodiment, it is beneficial to quickly calculate the correlation between the semantic vector of the fault description text and the semantic vector of the relevant text of each type of target data.

In a possible implementation, the above semantic generation model is generated according to the training of the word vector matrix corresponding to the historical fault description text, and the word vector matrix includes the word vector corresponding to each word in the historical fault description text, and the word vector is used for Indicates the semantics of words.

By implementing the semantic generation model trained in this embodiment, the semantics of the text can be expressed more accurately.

In a possible implementation manner, the above-mentioned multiple types of target data include at least two of key performance indicators, equipment alarms, and equipment logs; when the above-mentioned target data is a key performance indicator, the relevant text of the target data is a key The name of the performance indicator; when the above target data is a device alarm, the relevant text of the target data is the identification of the device alarm; when the above target data is a device log, the relevant text of the target data is the content fragment of the device log.

In a second aspect, the present application provides a training method for a semantic generation model, the method comprising: obtaining a word vector set corresponding to the training text, the word vectors included in the word vector set correspond to the words in the training text one by one, the The word vector is used to represent the semantics of the word; the historical fault description text is converted into a word vector matrix composed of at least one word vector according to the word vector set; the semantic generation model is obtained according to the word vector matrix training, and the semantic generation model is used to generate text semantic vector.

Optionally, after obtaining the word vector set corresponding to the training text, the word vector set corresponding to the training text may be saved, so that the word vectors in the word vector set can be used later.

It can be seen that the method described in the second aspect is to gradually model the semantics at the lexical level to the semantics at the sentence level to obtain a semantic generation model. This semantic generation model training method is in line with the basic principles of language generation. Therefore, by implementing the semantic generation model trained by the method described in the second aspect, the semantics of the text can be expressed more accurately.

In a possible implementation, the specific implementation of converting the historical fault description text into a word vector matrix composed of at least one word vector according to the word vector set is: performing word segmentation processing on the historical fault description text to obtain the historical fault description text A corresponding word sequence consisting of at least one word; obtaining word vectors corresponding to words included in the word sequence from the word vector set; forming a word vector matrix from word vectors corresponding to each word included in the word sequence.

By implementing this embodiment, the historical fault description text can be accurately converted into a word vector matrix composed of at least one word vector.

In a possible implementation manner, when there is no word vector corresponding to the word included in the word sequence in the word vector set, a random vector is generated as a word vector corresponding to the word included in the word sequence.

By implementing this embodiment, the historical fault description text can be accurately converted into a word vector matrix composed of at least one word vector.

In a possible implementation, the specific implementation of obtaining the semantic generation model according to the word vector matrix training is: obtaining the fault equipment type corresponding to the historical fault description text; training the classification model according to the word vector matrix and category labels, the category labels include The faulty device type; a semantic generation model is obtained according to the classification model.

By implementing the semantic generation model trained in this embodiment, the semantics of the text can be expressed more accurately.

In a possible implementation, the specific implementation of training the classification model according to the word vector matrix and category labels is as follows: the word vector matrix and category labels are input into the neural network for iterative training, and the input neural network The word vectors in the word vector matrix and the parameters of the neural network are adjusted to generate the classification model. The semantic generation model trained through this embodiment can express the semantics of the text more accurately.

Optionally, the word vectors corresponding to the corresponding words in the word vector set may also be updated using the word vectors in the word vector matrix input in the last iteration training. By implementing this embodiment, the word vectors in the word vector set can be corrected according to the historical fault description text corpus with domain knowledge, so that the word vectors in the word vector set can better express the semantic information of words in the domain knowledge.

In a third aspect, an information output device is provided, and the information output device can execute the method in the above-mentioned first aspect or a possible implementation manner of the first aspect. This function may be implemented by hardware, or may be implemented by executing corresponding software on the hardware. The hardware or software includes one or more units corresponding to the above functions. This unit can be software and/or hardware. Based on the same inventive concept, the problem-solving principles and beneficial effects of the information output device can be referred to the above-mentioned first aspect or the possible implementation manners and beneficial effects of the first aspect, and repeated descriptions will not be repeated.

In a fourth aspect, a model training device is provided, and the model training device can execute the method in the above-mentioned second aspect or a possible implementation manner of the second aspect. This function may be implemented by hardware, or may be implemented by executing corresponding software on the hardware. The hardware or software includes one or more units corresponding to the above functions. This unit can be software and/or hardware. Based on the same inventive concept, the problem-solving principle and beneficial effects of the model training device can be referred to the above-mentioned second aspect or the possible implementation manners and beneficial effects of the second aspect, and repeated descriptions will not be repeated.

According to a fifth aspect, an information output device is provided, and the information output device includes: a processor, a memory, and a communication interface; the processor, the communication interface, and the memory are connected; wherein, the communication interface may be a transceiver. The communication interface is used to implement communication with other network elements. Wherein, one or more programs are stored in the memory, and the processor invokes the programs stored in the memory to realize the above-mentioned first aspect or the solution in the possible implementation manners of the first aspect, and the information output device solves the implementation of the problem For the manner and beneficial effects, reference may be made to the above first aspect or the possible implementation manners and beneficial effects of the first aspect, and repeated descriptions will not be repeated.

In a sixth aspect, a model training device is provided, which includes: a processor, a memory, and a communication interface; the processor, the communication interface, and the memory are connected; wherein, the communication interface may be a transceiver. The communication interface is used to implement communication with other network elements. Wherein, one or more programs are stored in the memory, and the processor invokes the programs stored in the memory to realize the above-mentioned second aspect or the solution in the possible implementation manner of the second aspect, and the model training device solves the implementation of the problem For the manner and beneficial effects, reference may be made to the above-mentioned second aspect or possible implementation manners and beneficial effects of the second aspect, and repeated descriptions will not be repeated.

In a seventh aspect, a computer program product is provided, which, when running on a computer, causes the computer to execute the above-mentioned first aspect, the second aspect, the possible implementation manner of the first aspect, or the possible implementation manner of the second aspect Methods.

An eighth aspect provides a chip product of an information output device, which implements the method in the first aspect or any possible implementation manner of the first aspect.

A ninth aspect provides a chip product of a model training device, which implements the method in the second aspect or any possible implementation manner of the second aspect.

In a tenth aspect, a computer-readable storage medium is provided. Instructions are stored in the computer-readable storage medium. When the computer-readable storage medium is run on a computer, the computer executes the method of the above-mentioned first aspect or a possible implementation of the first aspect. methods in methods.

In the eleventh aspect, a computer-readable storage medium is provided. Instructions are stored in the computer-readable storage medium. When the computer-readable storage medium is run on a computer, the computer executes the method of the above-mentioned second aspect or the possible method of the second aspect. method in the implementation.

Description of drawings

FIG. 1 is a schematic flow diagram of an information output method provided in an embodiment of the present application;

Fig. 2 is a schematic flow chart of a training method for a semantic generation model provided by an embodiment of the present application;

Fig. 3 is the schematic diagram of the neural network adopted by a kind of CBOW algorithm that the embodiment of the application provides;

FIG. 4 is a schematic structural diagram of a neural network for training a classification model provided in an embodiment of the present application;

Fig. 5 is a schematic structural diagram of an information output device provided by an embodiment of the present application;

Fig. 6 is a schematic structural diagram of a model training device provided by an embodiment of the present application;

FIG. 7 is a schematic structural diagram of another information output device provided by an embodiment of the present application;

Fig. 8 is a schematic structural diagram of another model training device provided by an embodiment of the present application.

Detailed ways

The specific embodiments of the present application will be further described in detail below in conjunction with the accompanying drawings.

The embodiments of the present application provide an information output method and device, which can automatically determine and output data related to the fault description text to assist in analyzing the cause of the fault.

The information output method and device provided by this application will be introduced in detail below.

Please refer to FIG. 1 . FIG. 1 is a schematic flowchart of an information output method provided by an embodiment of the present application. As shown in Figure 1, the information output method includes the following parts 101-105, wherein:

101. The information output device acquires a fault description text.

Wherein, the fault description text is a text describing a fault phenomenon, that is, the fault description text is used to describe a fault occurring in the network. For example, the fault description text may be "slow Internet access of industry users", "communication interruption of online charging system (OCS)" and so on. The fault description text may be sent to the information output device by other devices. For example, the first-line engineer describes the fault phenomenon, obtains the fault description text, and sends the collected data (such as key performance indicators, etc.) and the fault description text to the operation and maintenance department in the form of a fault work order to assist in the analysis of the cause of the fault. Information output device.

102. The information output device generates the semantic vector of the fault description text through the semantic generation model.

In a possible implementation manner, the semantic generation model may be generated by training a word vector matrix corresponding to the historical fault description text, and the word vector matrix includes word vectors corresponding to each word in the historical fault description text.

Optionally, for the training method of the semantic generation model, please refer to the training method of the semantic generation model described in FIG. 2 below. That is to say, the semantic generation model used by the information output device may be the semantic generation model trained by the model training device in FIG. 2 below. The information output device in FIG. 1 and the model training device in FIG. 2 can be deployed on the same device or deployed on different devices. When the information output device in Figure 1 and the model training device in Figure 2 are deployed in different devices, after the model training device has trained the semantic generation model, it can send the semantic generation model to the information output device, so that the information output device can receive The semantic generative model generates semantic vectors of fault description text. When the information output device in Figure 1 and the model training device in Figure 2 are deployed in the same device, the information output device can obtain the semantic generation model from the model training device, so that the information output device can generate a fault description through the semantic generation model Semantic vectors of text.

Of course, the semantic generation model may not be trained in the manner described in FIG. 2 , but may also be trained in other ways, which is not limited in this embodiment of the present application.

In a possible implementation manner, the specific implementation manner in which the information output device generates the semantic vector of the fault description text through the semantic generation model is as follows:

The information output device converts the fault description text into a word vector matrix according to the word vector set, and then inputs the word vector matrix into the semantic generation model to generate a semantic vector of the fault description text, wherein the word vector set includes multiple word vectors. Optionally, the word vector set may be generated by the model training device in FIG. 2 below and sent to the information output device.

Optionally, the specific implementation manner in which the information output device converts the fault description text into a word vector matrix according to the word vector set is: the information output device performs word segmentation processing on the fault description text, and obtains a word consisting of at least one word corresponding to the fault description text Sequence; obtain the word vectors corresponding to the words included in the word sequence from the word vector set; form the word vector matrix of the fault description text with the word vectors corresponding to each word included in the word sequence. When the word vector corresponding to the word included in the word sequence does not exist in the word vector set, a random vector is generated as the word vector corresponding to the word included in the word sequence.

For example, the fault description text includes 4 words, and the word sequence obtained by performing word segmentation on the fault description text is "industry", "user", "Internet access", and "slow". The information output device finds word vector 1 corresponding to "industry", word vector 2 corresponding to "user", and word vector 3 corresponding to "surfing the Internet" from the word vector set. If the word vector corresponding to "slow" is not found, a random vector word is generated. Vector 4 is used as the word vector corresponding to "slow". The information output device forms word vectors 1 to 4 into a word vector matrix of the fault description text. Then input the word vector matrix into the semantic generation model to generate the semantic vector of the fault description text.

103. The information output device acquires semantic vectors respectively corresponding to relevant texts of multiple types of target data.

Wherein, the target data is used to assist in analyzing the cause of the fault. Wherein, the execution order of part 103 and part 102 can be in no particular order, part 102 can be executed first and then part 103 can be executed, or part 103 can be executed first and then part 102 can be executed.

In a possible implementation, the multiple types of target data include at least two of key performance indicators (KPIs), device alarms, and device logs; when the target data is a key performance indicator, the relevant text of the target data is the name of the key performance indicator; when the target data is a device alarm, the relevant text of the target data is the identification of the device alarm; when the target data is a device log, the relevant text of the target data is the content fragment of the device log. Wherein, each type of target data is multiple.

For example, the various types of target data include key performance indicators and equipment alarms. The various types of target data mentioned above are 100 different key performance indicators and 20 different equipment alarms. The 100 key performance indicators are respectively KPI 1 to KPI 100. The 20 equipment alarms are respectively equipment alarm 1 ~20. The semantic vectors corresponding to the relevant texts of various types of target data acquired by the information output device are the semantic vectors corresponding to the names of key performance indicators 1 to 100, and the semantic vectors corresponding to the signs of equipment alarms 1 to 20 . That is to say, the information output device will obtain 120 semantic vectors.

In a possible implementation manner, before receiving the fault description text, the information output device may generate semantic vectors respectively corresponding to relevant texts of various types of target data through a semantic generation model.

Optionally, after the information output device generates the semantic vectors respectively corresponding to the relevant texts of various types of target data, it may store the semantic vectors respectively corresponding to the relevant texts of the various types of target data. After receiving the fault description text, the semantic vectors corresponding to the related texts storing the various target data can be obtained, so as to perform correlation calculation with the semantic vectors of the fault description text. By implementing this embodiment, the semantic vectors corresponding to the relevant texts of various target data can be generated and saved in advance, and after receiving the fault description text, the semantic vectors corresponding to the relevant texts of various target data can be directly used to match the fault Correlation calculation is performed on the semantic vectors of the description text, so that it is not necessary to temporarily generate semantic vectors corresponding to the relevant texts of various target data after receiving the fault description text. It can be seen that by implementing this embodiment, it is beneficial to quickly calculate the correlation between the semantic vector of the fault description text and the semantic vector of the relevant text of each type of target data.

In a possible implementation, the principle of the information output device to generate the semantic vector corresponding to the relevant text of the target data through the semantic generation model is the same as the principle of the information output device to generate the semantic vector of the fault description text through the semantic generation model. repeat.

104. The information output device calculates the correlation between the semantic vector of the fault description text and the semantic vector of the relevant text of each type of target data.

For example, there are two types of target data, which are 100 different key performance indicators and 20 different equipment alarms, and the 100 key performance indicators are respectively KPI 1 to KPI 100. 20 equipment alarms They are device alarms 1 to 20 respectively. The information output device calculates the correlation between the semantic vectors of the fault description text and the semantic vectors of the relevant texts of 100 key performance indicators, and calculates the correlation between the semantic vectors of the fault description text and the semantic vectors of the relevant texts of 20 equipment alarms . Thus, 120 correlations would be obtained.

In a possible implementation, the angle between the vectors can be used as a measure of correlation, and the correlation between the semantic vector of the fault description text and the semantic vector of the relevant text of the target data can be expressed as:

Among them, cos(θ) is the correlation between the semantic vector of the fault description text and the semantic vector of the relevant text of the target data, n is the dimension number of the semantic vector of the fault description text and the relevant text of the target data, x _i is the fault description text The semantic vector of the i-th dimension, y _i The semantic vector of the i-th dimension of the relevant text of the target data.

105. The information output device determines and outputs the first data.

Wherein, after the information output device calculates the correlation between the semantic vector of the fault description text and the semantic vector of the relevant text of each type of target data in each type of target data, the information output device determines and outputs the first data, the first data is The target data with the greatest correlation between the semantic vector and the semantic vector of the fault description text in each type of target data, or the first data is the target data whose correlation between the semantic vector and the semantic vector of the fault description text in each type of target data is greater than a preset threshold target data.

For example, the two types of target data obtained are 100 different key performance indicators and 20 different equipment alarms, respectively, and the 100 key performance indicators are respectively KPI 1 to KPI 100. The 20 equipment alarms are respectively These are device alarm 1 to device alarm 20. The correlations between the semantic vectors of the fault description text and the semantic vectors of the relevant texts of key performance indicators 1 to 100 are correlations 1 to 100, respectively. Correlation 1 is the largest correlation, and the information output device outputs a key performance indicator 1 . The correlations between the semantic vectors of the fault description text and the semantic vectors of the relevant texts of equipment alarm 1-equipment alarm 20 are correlations 101-120, respectively. If the correlation 120 is the maximum correlation, the information output device outputs the equipment alarm 20 .

For another example, if the correlation 1 and the correlation 2 are correlations greater than a preset threshold, the information output device outputs the key performance indicator 1 and the key performance indicator 2 . If the correlation 101 and the correlation 102 are greater than the preset threshold, the information output device outputs the equipment alarm 1 and the equipment alarm 2 .

The target data with greater correlation between the semantic vector and the semantic vector of the fault description text indicates that the target data is more related to the fault description text, and the user may need to view the target data to analyze the cause of the fault. For example, the fault description text is "ocs communication interruption", and the name of the key indicator is "ocs communication interruption times", the semantic vector of the fault description text is highly correlated with the semantic vector of the key indicator name, and the user may need to check This key indicator is used to analyze the cause of the failure. It can be seen that by implementing the method described in FIG. 1 , data related to the fault description text to assist in analyzing the cause of the fault can be automatically found.

In the prior art, the text with the same keywords as the fault description text is searched, and the fault is checked and analyzed according to the relevant parameter data of the text. However, there may not be the same keywords in the relevant text that can be used to assist in analyzing the cause of the fault and the fault description text with high correlation. Therefore, the data associated with the fault description text and used to assist in analyzing the cause of the fault cannot be accurately found through the existing methods. The embodiment of the present application can accurately find the target data associated with the fault description text by comparing the correlation between the semantic vector of the fault description text and the semantic vector of the relevant text of the target data. For example, the fault is described as "slow Internet access by industry users", and the name of the key indicator for fault analysis related to it analyzed in the embodiment of the present application is "downlink bandwidth control packet loss ratio". It can be seen that there is literally no matching and correlation between the two, but this application has learned the domain knowledge of "Internet speed and packet loss ratio are related" through semantic analysis and mining, and realized the correlation between the two. linked analysis.

Therefore, by implementing the method described in FIG. 1 , data related to the fault description text for assisting in analyzing the cause of the fault can be automatically and accurately found.

Please refer to FIG. 2 . FIG. 2 is a schematic flowchart of a method for training a semantic generation model provided by an embodiment of the present application. As shown in Figure 2, the training method of the semantic generation model includes the following parts 201-203, wherein:

201. The model training device acquires a word vector set corresponding to the training text.

Wherein, the word vectors included in the word vector set are in one-to-one correspondence with the words in the training text. For example, if the training text includes 10,000 words, the word vector set also includes 10,000 word vectors. The word vector is used to represent the semantics of the word. Optionally, after obtaining the word vector set corresponding to the training text, the word vector set corresponding to the training text can be saved, so that the word vectors in the word vector set can be used later.

The training text is the corpus. In a possible implementation manner, the training text may be an encyclopedia type text. Word vectors learned from encyclopedic texts have good general semantics.

In a possible implementation, the model training device first preprocesses the training text, and then performs word segmentation processing on each sentence text to obtain the word-segmented training text, and obtains the word segmentation through word2vec tools or other tools The set of word vectors corresponding to the subsequent training text.

For example, the training text is "Mathematics is a subject that uses symbolic language to study the changes in the structure of quantities and concepts such as space. I like mathematics." The model training device splits the training text into two sentences, which are "Mathematics is a subject that uses symbolic language to study the changes in the structure of quantities and concepts such as space" and "I like mathematics". The two sentences are then segmented separately. The training text after word segmentation is "Mathematics is a subject that uses symbolic language to study the changes in the structure of quantities and concepts such as space. I like mathematics." The model training device uses the word2vec tool to traverse the word-segmented training text sentence by sentence, and the word vector corresponding to each word in the training text is obtained after the traversal. The model training device stores a word vector set composed of word vectors corresponding to each word in the training text.

The model training device can use the word2vec tool and the CBOW algorithm to obtain the word vector set corresponding to the training text after word segmentation. The idea of the CBOW algorithm is to predict the current word through a given context word. The goal of CBOW algorithm training is to maximize the probability of occurrence of a word when the context of a word is given. After training, each word gets a corresponding word vector in the output layer. Although the modeling idea of the CBOW algorithm is a classification process, it will generate word vectors as a by-product.

For example, FIG. 3 is a schematic diagram of a neural network used by the CBOW algorithm. As shown in Figure 3, the neural network consists of three layers, which are input layer, mapping layer and output layer. Wherein, the output layer includes the constructed Huffman tree. A leaf node of the Huffman tree represents the word vector of a word in the training text, and the word vector of the word corresponding to each leaf node is randomly initialized. Each non-leaf node has a built-in weight vector whose dimension is the same as the word vector of the input layer.

Among them, the input layer is a word vector of n-1 words around a certain word w(t). n is the window size. For example, if n is 5, the n-1 words around word w(t) are the first two words and the last two words of word w(t). The first two and last two words of word w(t) are w(t-2), w(t-1), w(t+1), w(t+2), respectively. Correspondingly, the word vectors of these n-1 words are recorded as v(w(t-2)), v(w(t-1)), v(w(t+1)), v(w(t +2)). The input layer transfers the n-1 word vectors to the mapping layer, and the mapping layer adds the n-1 word vectors, that is, correspondingly adds the dimensions of the n-1 word vectors. For example, the mapping layer input is pro(t)=v(w(t-2))+v(w(t-1))+v(w(t+1))+v(w(t+2)) .

The projection layer inputs the summed vector pro(t) to the root node of the Huffman tree. After inputting pro(t) into the root node, the probability from the root node to each leaf node will be calculated. The training process of the model is expected to have the highest probability of reaching the leaf node corresponding to w(t) from the root node. In the training text of , the same context will appear many times, so each weight vector will be continuously corrected in the process of traversing the training text to achieve this effect. After traversing all the words in the training text, the word vector corresponding to each leaf node of the Huffman tree is the word vector corresponding to each word in the training text. Here, "all words in the training text" includes repeated words in the training text.

Among them, each time the leaf node corresponding to the word w(t) passes through an intermediate node from the root node, it is equivalent to performing a binary classification, and the classifier can use a softmax regression classifier. Among them, the classification probability of each classification is:

Among them, _θi represents the ith weight vector. pro(t) is the sum of the word vectors of the context of w(t), and e is a natural constant.

Assume that the path traversed from the root node to the leaf node corresponding to the word w(t) contains L intermediate nodes, and the parameters on these nodes form a parameter vector as [θ ₁ , θ ₂ , θ ₃ ,..., θ _L ], then The probability of the root node to the leaf node corresponding to the word w(t) is the product of the probability of each binary classification, that is, the probability of the root node to the leaf node corresponding to the word w(t) is:

表示i从1到L逐一递增对P(context(w(t)),θ_i)连乘求积。从根节点到其他叶子节点的概率的计算方法同理，在此不赘述。Among them, P(w(t)|context(w(t))) is the probability from the root node to the leaf node corresponding to word w(t),

Indicates that i is incremented one by one from 1 to L to multiply and multiply P(context(w(t)), θ _i ). The calculation method of the probability from the root node to other leaf nodes is the same, and will not be repeated here.

202. The model training device converts the historical fault description text into a word vector matrix consisting of at least one word vector according to the word vector set.

Specifically, the model training device can convert a large amount of historical fault description text into a word vector matrix. The model training device obtains a semantic generation model according to a large amount of word vector matrix training. For example, if there are historical fault description text 1 to historical fault description text 100, the historical fault description text 1 to historical fault description text 100 can be respectively converted into word vector matrices, that is, 100 word vector matrices are obtained. The model training device trains the 100 word vector matrices to obtain a semantic generation model.

In a possible implementation, the model training device converts the historical fault description text into a word vector matrix consisting of at least one word vector according to the word vector set. The specific implementation may be: the model training device performs word segmentation on the historical fault description text Process to obtain a word sequence consisting of at least one word corresponding to the historical fault description text; obtain word vectors corresponding to words included in the word sequence from the word vector set; form the history with word vectors corresponding to each word included in the word sequence A matrix of word vectors for the fault description text. When the word vector corresponding to the word included in the word sequence does not exist in the word vector set, a random vector may be generated as the word vector corresponding to the word included in the word sequence. It can be seen that by implementing this embodiment, the historical fault description text can be accurately converted into a word vector matrix composed of at least one word vector.

For example, the historical fault description text 1 includes 4 words, and the word sequence obtained by performing word segmentation on the historical fault description text 1 is "industry", "user", "Internet access", and "slow". The model training device finds word vector 1 corresponding to "industry", word vector 2 corresponding to "user", and word vector 3 corresponding to "Internet" from the set of word vectors. If the word vector corresponding to "slow" is not found, a random vector word is generated Vector 4 is used as the word vector corresponding to "slow". The model training device forms the word vectors 1-4 into the word vector matrix 1 of the historical fault description text 1 . The principles of converting other historical fault description texts 2 to 100 into word vector matrices are the same as the conversion of historical fault description text 1 into word vector matrices, and will not be repeated here.

203. The model training device trains the word vector matrix to obtain a semantic generation model.

Specifically, after obtaining the word vector matrix, the model training device can input the word vector matrix into the neural network for training to obtain a semantic generation model. This semantic generative model is used to generate semantic vectors of text. This semantic vector is used to represent the semantics of the text.

It can be seen that the method described in Figure 2 is to gradually model the semantic generation model from the semantic level of the vocabulary to the semantic level of the sentence level. This training method of the semantic generation model is in line with the basic principles of language generation. Therefore, the semantic generation model trained by implementing the method described in FIG. 2 can express the semantics of the text more accurately.

In a possible implementation, the specific implementation of the model training device to obtain the semantic generation model according to the word vector matrix training is: the model training device obtains the fault equipment type corresponding to the historical fault description text; the model training device obtains the fault equipment type according to the word vector matrix and category The classification model is trained with the label, and the category label includes the type of the faulty equipment; the model training device obtains the semantic generation model according to the classification model. By implementing the semantic generation model trained in this embodiment, the semantics of the text can be expressed more accurately.

For example, the faulty device type corresponding to the historical fault description text may be a router, a wired device, or a wireless device. For example, the fault described in the historical fault description text is a fault generated by a router, and the faulty device type corresponding to the historical fault description text is a router. Front-line engineers can collect the faulty device type corresponding to each fault description text, and then add the fault description text, the faulty device type corresponding to the fault description text, and the data used to assist in analyzing the cause of the fault to the work order, and send the work order to The operation and maintenance terminal analyzes the cause of the failure. Therefore, the model training device can obtain the type of faulty equipment corresponding to the historical fault description text from the work order.

Wherein, the classification model obtained through training is a model used to generate the type of faulty equipment corresponding to the fault description text. For example, the word vector matrix corresponding to the fault description text 1 is input into the classification model, and the classification model can output the faulty equipment type corresponding to the fault description text 1.

In a possible implementation, the specific implementation of the model training device training the classification model according to the word vector matrix and category labels is as follows: the word vector matrix and category labels are input into the neural network for iterative training, and the input The word vectors in the word vector matrix of the neural network and the parameters of the neural network are adjusted to generate a classification model. By implementing this embodiment, the trained classification model can accurately classify the fault description text.

Optionally, the model training device may also use the word vectors in the adjusted word vector matrix to update the word vectors corresponding to the corresponding words in the word vector set. By implementing this optional method, the word vectors in the word vector set can be corrected according to the historical fault description text corpus with domain knowledge, so that the word vectors in the word vector set can better express the semantic information of words in the fault domain.

For example, FIG. 4 is a structural schematic diagram of a neural network used for training a classification model. As shown in Figure 4, the neural network includes convolutional layers, pooling layers, and fully connected layers. The word vector matrix 1 of the historical fault description text 1 includes word vectors {w1, w2, w3, w4, w5, w6}. Each word vector has 128 dimensions. After the model training device obtains the word vector matrix 1, it inputs the word vector matrix 1 into the neural network. As shown in Figure 4, there are two convolution kernels in the neural network. Of course, there may be more than two convolution kernels in practical applications, and the embodiment of the present application uses two convolution kernels as an example for illustration. The convolution kernel 1 on the left performs pairwise convolution on the word vector matrix 1 including the word vector. For example, w1 is convolved with w2 to get C1, w2 is convolved with w3 to get C2, w3 is convolved with w4 to get C3, w4 is convolved with w5 to get C4, and w5 is convolved with w6 to get C5. The convolution kernel 2 on the right performs three-three convolutions on the word vector matrix 1 including the word vector. For example, w1, w2, and w3 are convolved to get C6, w2, w3, and w4 are convolved to get C7, w3, w4, and w5 are convolved to get C8, and w4, w5, and w6 are convolved to get C9. In practical applications, other numbers of word vectors are also convolved, and the embodiment of the present application uses two-two convolution and three-three convolution as examples for illustration.

It can be seen that the convolution kernel 1 can generate a feature map (feature map) C=[C1, C2, . . . , C5], and the convolution kernel 2 can generate a feature map C=[C6, C7, C8, C9]. After the model training device obtains the feature map generated by each convolution kernel, for each feature map, the maximum value in each dimension is selected as the text feature vector generated by the current convolution kernel through a maximum pooling operation. The model training device splices all text feature vectors to obtain the final semantic vector of historical fault description text 1. That is, as shown in Figure 4, the model training device selects the largest value from the first dimension of C1~C5, selects the largest value from the second dimension of C1~C5, and selects the largest value from the third dimension of C1~C5. Select the largest value, and so on, until the largest value is selected from the 128th dimension of C1 to C5. The model training device forms the maximum value of the selected 128 dimensions into the text feature vector 1 corresponding to the convolution kernel 1 . Similarly, the model training device also obtains the text feature vector 2 corresponding to the convolution kernel 2 . The model training device splices the text feature vector 1 and the text feature vector 2 to obtain the final semantic vector of the historical fault description text 1.

The model training device inputs the semantic vector of the historical fault description text 1 into the fully connected layer, and uses the fault device type (such as a router) corresponding to the historical fault description text 1 as a category label, and inputs it into the fully connected layer. The model training device analyzes the semantic vector of the historical fault description text 1 at the fully connected layer, and the analysis shows that the maximum probability of the faulty device type is a switch. Since the most probable faulty device type (i.e. switch) obtained by analyzing the semantic vector of historical fault description text 1 is different from the category label (i.e. router) corresponding to historical fault description text 1, the model training device records the historical fault description The type of faulty equipment with the greatest probability obtained by analyzing the semantic vector of text 1 is incorrect. Similarly, the model training device inputs the word vector matrix of the historical fault description text 2 into the neural network for training according to the above process, obtains the semantic vector of the historical fault description text 2, and inputs the fault equipment corresponding to the historical fault description text 2 in the fully connected layer type (such as a switch) as a class label. The model training device analyzes the semantic vector of the historical fault description text 2, and the analysis shows that the maximum probability of the faulty device type is a firewall. Therefore, the type of faulty equipment with the highest probability obtained by analyzing the semantic vector of the historical fault description text 2 recorded by the model training device is incorrect. Assuming that there are 100 historical fault description texts, the remaining 98 historical fault description texts are similarly input to the neural network for training of the classification model according to the above-mentioned method of historical fault description text 1. After completing the first round of training for historical fault description texts 1 to 100, assuming that the maximum probability of the faulty device type obtained from the analysis of the semantic vectors corresponding to historical fault description texts 1 to 50 is incorrect, the model training device will be correct for the parameters of the neural network and The word vectors in the word vector matrix corresponding to historical fault description texts 1 to 50 are adjusted. After the adjustment is completed, the historical fault description texts 1-100 are retrained with the new word vector matrix and the parameters of the neural network until the most probable faulty equipment type and classification are obtained according to the semantic vector analysis corresponding to the historical fault description texts 1-100 If the tags match, a classification model is generated, that is, a classification model is generated by iteratively training a neural network.

Finally, the model training device uses the word vectors in the word vector matrix input in the last round of iteration training to update the word vectors corresponding to the corresponding words in the word vector set. For example, the historical fault description text 1 is "slow Internet access" before the last round of iterative training. After the iterative training is completed, use word vector 1 to replace the word vector corresponding to "go online" in the word vector set. Historical fault description text 2 is "OCS communication interruption". After the training is complete, use word vector 2 to replace the word vector corresponding to "interrupt" in the word vector set. The same is true for other historical fault description texts, so I won’t repeat them here.

In a possible implementation manner, the specific implementation manner in which the model training device obtains the semantic generation model according to the classification model is: the model training device uses the part above the fully connected layer in the classification model as the semantic generation model. By implementing the semantic generation model generated in this embodiment, the semantic vector of the text can be accurately generated.

In the embodiment of the present invention, the device can be divided into functional modules according to the above method examples. For example, each functional module can be divided corresponding to each function, or two or more functions can be integrated into one module. The above-mentioned integrated modules can be implemented in the form of hardware or in the form of software function modules. It should be noted that the division of modules in the embodiment of the present invention is schematic, and is only a logical function division, and there may be another division manner in actual implementation.

Please refer to FIG. 5, which is an information output device provided by the implementation of the present invention. The information output device includes: an acquisition module 501 , a generation module 502 , a calculation module 503 and an output module 504 . in:

The obtaining module 501 is used to obtain the fault description text; the generation module 502 is used to generate the semantic vector of the fault description text through the semantic generation model, and the fault description text is used to describe the fault occurring in the network; the obtaining module 501 is also used to obtain Semantic vectors corresponding to the relevant texts of multiple types of target data, the target data is used to assist in the analysis of the cause of the fault; the calculation module 503 is used to calculate the semantic vector of the fault description text and the semantics of the relevant text of each target data The correlation of the vector; the output module 504 is used to determine and output the first data, the first data is the target data with the greatest correlation between the semantic vector and the semantic vector of the fault description text in each target data, or the first data is the target data for each target data The target data whose correlation between the semantic vector and the semantic vector of the fault description text in the target data is greater than a preset threshold.

In a possible implementation manner, the generation module 502 is further configured to generate semantic vectors respectively corresponding to relevant texts of various target data through a semantic generation model before the acquisition module 501 acquires the fault description text.

In a possible implementation, the semantic generation model is trained and generated according to the word vector matrix corresponding to the historical fault description text, and the word vector matrix includes the word vector corresponding to each word in the historical fault description text, and the word vector is used to represent the word semantics.

In a possible implementation, the multiple types of target data include at least two of key performance indicators, device alarms, and device logs; when the target data is a key performance indicator, the relevant text of the target data is a key performance indicator name; when the target data is a device alarm, the relevant text of the target data is the identification of the device alarm; when the target data is a device log, the relevant text of the target data is the content fragment of the device log.

Please refer to Fig. 6, Fig. 6 is a model training device provided by the implementation of the present invention. The model training device includes an acquisition module 601, a conversion module 602 and a training module 603, wherein:

Acquisition module 601 is used to obtain the word vector set corresponding to the training text, and the word vector included in the word vector set corresponds to the words in the training text one by one; the conversion module 602 is used to convert the historical fault description text into A word vector matrix consisting of at least one word vector; the training module 603 is further configured to train a semantic generation model according to the word vector matrix, and the semantic generation model is used to generate a semantic vector of the text.

In a possible implementation, the conversion module 602 is specifically configured to: perform word segmentation processing on the historical fault description text to obtain a word sequence consisting of at least one word corresponding to the historical fault description text; obtaining the word sequence from the word vector set includes The word vectors corresponding to the words in the word sequence; the word vectors corresponding to each word included in the word sequence form a word vector matrix.

In a possible implementation manner, the conversion module 602 is further specifically configured to: generate a random vector as a word vector corresponding to the word included in the word sequence when there is no word vector corresponding to the word included in the word sequence in the word vector set.

In a possible implementation, the training module 603 trains the semantic generation model according to the word vector matrix by: obtaining the type of faulty equipment corresponding to the historical fault description text; training the classification model according to the word vector matrix and category labels, the category The label includes the type of the faulty device; the semantic generation model is obtained according to the classification model.

In a possible implementation, the training module 603 trains the classification model according to the word vector matrix and category labels as follows: input the word vector matrix and category labels into the neural network for iterative training, and input the neural network The word vectors in the word vector matrix of the network and the parameters of the neural network are adjusted to generate a classification model.

Please refer to FIG. 7 . FIG. 7 is a schematic structural diagram of an information output device disclosed in an embodiment of the present application. As shown in FIG. 7 , the information output device 700 includes a processor 701 , a memory 702 and a communication interface 703 . Wherein, the processor 701, the memory 702 and the communication interface 703 are connected.

Wherein, the processor 701 may be a central processing unit (central processing unit, CPU), a general purpose processor, a coprocessor, a digital signal processor (digital signal processor, DSP), an application-specific integrated circuit (application-specific integrated circuit, ASIC) , field programmable gate array (fieldprogrammable gate array, FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. The processor 701 may also be a combination that implements computing functions, for example, a combination of one or more microprocessors, a combination of a DSP and a microprocessor, and the like.

Wherein, the communication interface 703 is used to implement communication with other network elements.

Wherein, the processor 701 invokes the program code stored in the memory 702 to execute the steps performed by the information output device in the above method embodiments.

Please refer to FIG. 8 , which is a schematic structural diagram of a model training device disclosed in an embodiment of the present application. As shown in FIG. 8 , the model training device 800 includes a processor 801 , a memory 802 and a communication interface 803 . Wherein, the processor 801, the memory 802 and the communication interface 803 are connected.

Wherein, the processor 801 may be a central processing unit (central processing unit, CPU), a general processor, a coprocessor, a digital signal processor (digital signal processor, DSP), an application-specific integrated circuit (application-specific integrated circuit, ASIC) , field programmable gate array (fieldprogrammable gate array, FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. The processor 801 may also be a combination of computing functions, for example, a combination of one or more microprocessors, a combination of a DSP and a microprocessor, and the like.

Wherein, the communication interface 803 is used to implement communication with other network elements.

Wherein, the processor 801 invokes the program code stored in the memory 802 to execute the steps performed by the model training apparatus in the above method embodiments.

Based on the same inventive concept, the problem-solving principle of each device provided in the embodiment of the present application is similar to that of the method embodiment of the present application. Therefore, the implementation of each device can refer to the implementation of the method. For a concise description, it is not repeated here.

In the foregoing embodiments, the descriptions of each embodiment have their own emphases, and for parts not described in detail in a certain embodiment, reference may be made to relevant descriptions of other embodiments.

Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, rather than limiting them; although the application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: It is still possible to modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the technical solutions of the various embodiments of the present application. scope.

Claims (15)

1. An information output method, characterized in that the method comprises:

Acquiring a fault description text, wherein the fault description text is used for describing faults occurring in a network;

generating a semantic vector of the fault description text through a semantic generation model;

acquiring semantic vectors respectively corresponding to related texts of multiple types of target data, wherein the target data are used for assisting in analyzing reasons generated by the faults;

calculating the correlation between the semantic vector of the fault description text and the semantic vector of the related text of each target data;

determining and outputting first data, wherein the first data is target data with the largest correlation between semantic vectors in each target data and semantic vectors of the fault description text, or the first data is target data with the correlation between semantic vectors in each target data and semantic vectors of the fault description text being larger than a preset threshold;

wherein the multiple types of target data comprise at least two of key performance indexes, equipment alarms and equipment logs; when the target data is the key performance index, the related text of the target data is the name of the key performance index; when the target data is the equipment alarm, the related text of the target data is the identification of the equipment alarm; and when the target data is the equipment log, the related text of the target data is a content segment of the equipment log.

2. The method of claim 1, wherein prior to the obtaining the fault description text, the method further comprises:

and generating semantic vectors respectively corresponding to the related texts of the target data of multiple types through the semantic generation model.

3. The method according to claim 1 or 2, wherein the semantic generation model is generated according to word vector matrix training corresponding to the historical fault description text, the word vector matrix comprising word vectors corresponding to words in the historical fault description text, the word vectors being used for representing the semantics of the words.

4. A method of training a semantic generation model, the method comprising:

acquiring a word vector set corresponding to a training text, wherein word vectors included in the word vector set correspond to words in the training text one by one, and the word vectors are used for representing the semantics of the words;

converting the historical fault description text into a word vector matrix consisting of at least one word vector according to the word vector set;

acquiring a fault equipment type corresponding to the historical fault description text;

training a classification model according to the word vector matrix and a class label, wherein the class label comprises the fault equipment type;

And obtaining a semantic generation model according to the classification model, wherein the semantic generation model is used for generating semantic vectors of the text.

5. The method of claim 4, wherein said converting the historical fault description text from the set of word vectors into a word vector matrix of at least one word vector comprises:

word segmentation processing is carried out on the historical fault description text, and a word sequence which corresponds to the historical fault description text and consists of at least one word is obtained;

acquiring word vectors corresponding to words included in the word sequence from the word vector set;

and forming word vector matrixes by word vectors corresponding to the words included in the word sequences.

6. The method of claim 5, wherein the method further comprises:

when the word vector corresponding to the word included in the word sequence does not exist in the word vector set, generating a random vector as the word vector corresponding to the word included in the word sequence.

7. The method according to any one of claims 4-6, wherein training a classification model according to the word vector matrix and the class labels comprises:

and inputting the word vector matrix and the category labels into a neural network for iterative training, and adjusting the word vectors in the word vector matrix input into the neural network and parameters of the neural network during each iterative training to generate the classification model.

8. An information output apparatus, characterized in that the information output apparatus includes:

the system comprises an acquisition module, a fault description module and a fault analysis module, wherein the acquisition module is used for acquiring a fault description text, and the fault description text is used for describing faults in a network;

the generation module is used for generating semantic vectors of the fault description text through a semantic generation model;

the acquisition module is further used for acquiring semantic vectors corresponding to related texts of various types of target data respectively, wherein the target data are used for assisting in analyzing the cause of the fault;

the calculating module is used for calculating the correlation between the semantic vector of the fault description text and the semantic vector of the related text of each target data;

the output module is used for determining and outputting first data, wherein the first data is target data with the largest correlation between semantic vectors in each target data and semantic vectors of the fault description text, or the first data is target data with the correlation between semantic vectors in each target data and semantic vectors of the fault description text being larger than a preset threshold;

wherein the multiple types of target data comprise at least two of key performance indexes, equipment alarms and equipment logs; when the target data is the key performance index, the related text of the target data is the name of the key performance index; when the target data is the equipment alarm, the related text of the target data is the identification of the equipment alarm; and when the target data is the equipment log, the related text of the target data is a content segment of the equipment log.

9. The apparatus of claim 8, wherein the device comprises a plurality of sensors,

the generation module is further configured to generate semantic vectors corresponding to related texts of multiple target data respectively through the semantic generation model before the acquisition module acquires the fault description text.

10. The apparatus according to claim 8 or 9, wherein the semantic generation model is generated according to a word vector matrix training corresponding to a historical fault description text, the word vector matrix comprising word vectors corresponding to words in the historical fault description text, the word vectors being used to represent the semantics of the words.

11. A model training apparatus, characterized in that the model training apparatus comprises:

the training text obtaining module is used for obtaining a word vector set corresponding to the training text, wherein word vectors included in the word vector set correspond to words in the training text one by one;

the conversion module is used for converting the historical fault description text into a word vector matrix consisting of at least one word vector according to the word vector set;

the training module is used for acquiring the type of the fault equipment corresponding to the historical fault description text; and training a classification model according to the word vector matrix and a class label, wherein the class label comprises the fault equipment type; and the semantic generation model is used for generating a semantic vector of the text according to the classification model.

12. The apparatus of claim 11, wherein the conversion module is specifically configured to:

word segmentation processing is carried out on the historical fault description text, and a word sequence which corresponds to the historical fault description text and consists of at least one word is obtained;

acquiring word vectors corresponding to words included in the word sequence from the word vector set;

and forming word vector matrixes by word vectors corresponding to the words included in the word sequences.

13. The apparatus of claim 12, wherein the conversion module is further specifically configured to:

when the word vector corresponding to the word included in the word sequence does not exist in the word vector set, generating a random vector as the word vector corresponding to the word included in the word sequence.

14. The apparatus according to any one of claims 11 to 13, wherein the training module trains the classification model according to the word vector matrix and the class label by:

and inputting the word vector matrix and the category labels into a neural network for iterative training, and adjusting the word vectors in the word vector matrix input into the neural network and parameters of the neural network during each iterative training to generate the classification model.

15. A computer readable storage medium having instructions stored therein which, when run on a computer, cause the computer to perform the method of any of the preceding claims 1 to 7.

Images