TWI745878B - Chat robot system and chat robot model training method - Google Patents

Abstract

A chat robot model training method includes: recording an association between an intent and an entity and an answer by means of a knowledge base; analyzing the intent and entity in a plurality of training corpora; using a generalization ability of a machine learning model to separate a plurality of synonym words in the training corpus, and removing the synonym words from the training corpus to obtain a corpus set; inputting the corpus set into a neural network model to obtain a training model, testing the training model to obtain a test result; analyzing the test results according to a predefined rule to generate an analysis result, and generating a new version of the training corpus according to the analysis result.

Description

Chat robot system and chat robot model training method

The invention relates to an interactive robot system, in particular to a chat robot system and a chat robot model training method.

When a user browses a webpage, some webpages can open an interactive window for the user to input a message, and the interactive window can display reply information based on the user's input message. In addition, the user can also input questions into a robot with a user interface in text or voice, and the robot can use the user interface to respond to the user's question.

Traditional web servers or robots require a large amount of training corpus for machine learning. When the user input information matches the training corpus, the corresponding reply information can be found.

However, the same user's intention (for example, to turn on the light) can be expressed in multiple ways (for example, I want to turn on the light, feel dark, turn on the light for me... etc.), which traditional web servers or robots cannot exhaust All the words that users may use are used as training corpus. Even if a large amount of training corpus is used, it may lead to an increase in the amount of calculations of the system, requiring hardware upgrades or costly time. Therefore, how to efficiently generate a training model of a chat robot has become one of the problems to be solved in this field.

In order to solve the above-mentioned problems, one aspect of the present disclosure provides a chat robot system. The chat robot system includes a storage device and a processor. The storage device is used for storing a knowledge base for recording the association between a combination of an intention and an entity and an answer. The processor is used to analyze the intentions and entities in a plurality of training corpus, and use a generalization ability of a machine learning model to separate the plural synonyms in these training corpus, and to separate these synonyms After the training corpus is removed, a corpus is obtained. After the corpus is input to a neural network model, a training model is obtained, and the training model is tested to obtain a test result. Wherein, when the processor analyzes the test result from the test result using a predefined rule to generate an analysis result, and generates a new version of the training corpus according to the analysis result.

Another aspect of the present invention is to provide a chat robot model training method. The chatbot model training method includes: using a knowledge base to record the association between the combination of an intention and an entity and an answer; analyzing the intentions and entities in a plurality of training corpus; using a generalization of a machine learning model ( generalization) Ability to separate plural synonymous characters in these training corpus, and remove these synonymous characters from these training corpora to obtain a corpus; input the corpus into a type of neural network model, Obtain a training model, test the training model to obtain a test result; analyze the test result according to a predefined rule to generate an analysis result, and generate a new version of the training corpus according to the analysis result.

The chat robot system and chat robot model training method shown in the present invention remove synonymous words in the training corpus through the generalization ability of the machine learning model to generate the size of the corpus used for the input neural network model , Which can reduce the amount of calculation and save time. By analyzing the training model, a new version of the training corpus can be generated based on the analysis result, and the new version of the training corpus can be input into the neural network model to achieve the effect of optimizing the training model.

The following description is a preferred implementation of the invention, and its purpose is to describe the basic spirit of the invention, but not to limit the invention. The actual content of the invention must refer to the scope of the claims that follow.

It must be understood that the words "including", "including" and other words used in this specification are used to indicate the existence of specific technical features, values, method steps, operations, elements, and/or components, but they do not exclude Add more technical features, values, method steps, job processing, components, components, or any combination of the above.

Words such as "first", "second", and "third" used in the claims are used to modify the elements in the claims, and are not used to indicate that there is an order of priority, antecedent relationship, or an element Prior to another element, or the chronological order of execution of method steps, is only used to distinguish elements with the same name.

Please refer to FIGS. 1 and 2. FIG. 1 is a block diagram of the integrated chat robot system 100 according to an embodiment of the present invention. FIG. 2 is a flowchart of a chat robot model training method 200 according to an embodiment of the present invention.

In one embodiment, as shown in FIG. 1, the integrated multi-platform login system 100 includes a storage device 10 and a processor 20. In one embodiment, the storage device 10 is coupled to the processor 20.

In an embodiment, the chat robot system 100 may be a notebook computer, a desktop computer, a tablet, a mobile phone, or other electronic devices. In one embodiment, the chat robot system 100 includes a user interface and a screen through which the user can input messages, and the screen can display user input messages and reply messages output by the processor 20. For example, after a user opens a web page on a laptop computer, and the web page is linked to the chatbot system 100, a message window is opened to receive the user's input message. For example, the user enters "Tokyo Repair Center" in the message window Where?" The chat robot system 100 finds the answer after executing the chat robot model training method 200, and responds to the address of the repair center in Tokyo in the message window.

In one embodiment, the storage device 10 can be implemented as a read-only memory, flash memory, floppy disk, hard disk, optical disk, flash drive, tape, a database accessible by the network, or those familiar with the art Easily think about storage media with the same function.

In one embodiment, the processor 20 may be composed of a volume circuit such as a micro controller, a microprocessor, a digital signal processor, and an application specific integrated circuit (application specific integrated circuit, ASIC) or a logic circuit.

Please refer to Figure 2, the chat robot model training method 200 is described below.

In step 210, the storage device 10 stores a knowledge base for recording the association between a combination of a question type and an entity and an answer.

In one embodiment, the sentence input by the user may correspond to the same or different intentions, and the intention refers to the operation that the user wants to perform or the type of question, which can be input by the user through the chatbot system 100 or manually Sentences are classified into intentions. Examples of sentences and intentions are shown in Table 1 below. sentence intention hi Greeting hello Greeting hi there Greeting where is the repair center in Tokyo Find Repair Center I want to find a repair center in Taipei city Find Repair Center i want to go to the acer repair center Find Repair Center why my wifi is not working Troubleshooting I am not able to use my wifi Troubleshooting my wifi doesn't work Troubleshooting Launch the system backup Use application Use the system backup Use application Please open the Acer Recovery Use application In Table 1, the chat robot system 100 can map multiple sentences to a high-dimensional vector space through neural algorithms, such as word embedding algorithms or intent classification algorithms, and then go Calculate the distance between each other, and finally classify the question. Those with similar distances are regarded as the same type of intent; however, the present invention is not limited to this, as long as the method can be used to classify the sentence intent. By marking the intention of each sentence, the chatbot system 100 can learn the meaning behind each sentence input by the user.

In one embodiment, the chat robot system 100 also needs important words in sentences, called entities (or keywords), for example, in the sentences "where is the repair center in Tokyo" and "where is the repair center in Taipei" In this sentence, "Tokyo" and "Taipei" are called entities. Because the same intention may occur on different entities, the chatbot system 100 will learn to recognize the entities in the sentence.

In one embodiment, the chatbot system 100 can apply a named entity recognition (Named Entity Recognition, NER) technology to identify entities, where the algorithm used in this technology is a conditional random field (CRF), which is A discriminative probability model, often used to mark or analyze sequence data, such as natural language text, to predict tags in sentences. In one embodiment, the chatbot system 100 can determine the tag of the entity by the context of the sentence. For example, the place after the preposition "in" is usually the place, so the "Tokyo" following the "in" is marked as Location, and treat the location "Tokyo" as an entity.

With the above method, when the user enters "where is the repair center in Tokyo", the chatbot system 100 can understand the semantic meaning and simplify the sentence information to the intention "find repair center", and the entity is "Tokyo". Then, the knowledge base can record the association between an intent (for example, "find repair center") and an entity (for example, "Tokyo") and an answer (for example, "shinjuku" (Shinjuku)). In other words, the knowledge base regards the intent and the entity as a set of indexes, through which the answers can be associated (or searched).

In step 220, the processor 20 analyzes the intentions and entities in the plurality of training corpora.

In one embodiment, the sentences in Table 1 can be regarded as training corpus, and the processor 20 analyzes the intention of the sentence through a word vector algorithm or an intention classification algorithm.

In one embodiment, the processor 20 uses named entity recognition technology to identify entities.

In step 230, the processor 20 uses a generalization capability of a machine learning model to separate the plural synonyms in the training corpus, and remove these synonyms from the training corpus, To obtain a corpus.

In one embodiment, the synonymous characters can be defined by the user in advance.

FIG. 3 is a schematic diagram of a chat robot model training method according to an embodiment of the present invention. Please also refer to Figure 3 below. In the pre-process stage, the processor 20 removes the synonymous characters 32 from the training corpus 30 to generate a corpus 34 as data for subsequent training.

In one embodiment, the machine learning model can be implemented by logistic regression, linear regression, nearest neighbor algorithm, K-Means clustering, neural network, etc.

In one embodiment, generalization ability refers to the adaptability of the machine learning model to fresh samples (for example, training corpus 30). The purpose of learning is to learn the rules behind the pairing. For data other than the training corpus 30 with the same rules, the trained network can also give appropriate output. This ability is called generalization ability.

For example, in the sentences "where is the repair center in Tokyo" and "where is the repair center in Taipei", the sentence patterns "Tokyo" and "Taipei" are entities, which are regarded as important keywords. And mark it as a location. When a new piece of data is entered, such as "where is the repair center in Beijing", even if the data has not appeared before, the generalization ability of the machine learning model can still be based on rules (such as sentence patterns, prepositions, or Other rules) It is determined that "Beijing" is an entity and marked as a location.

In this way, even if a large amount of training corpus 30 is not used, the generalization ability of the machine learning model can be used to determine the entity of the new data and its label.

Since a good machine learning model can generalize to a certain degree, in other words, even if certain words do not require special training, the machine learning model can recognize these words, so the generalization ability of machine learning can be used to Synonymous characters 32 are separated from the training corpus 30.

Among them, the synonyms 32 are, for example, "wifi", "wi-fi", and "wireless". These three words appear in a sentence to a high degree and represent the same meaning. Therefore, these three words can be taken as "wifi" "Is retained as keywords, while "wi-fi" and "wireless" are removed from the training corpus 30. The training corpus 30 with the synonym 32 (only one keyword retained) is removed is called the corpus 34. In one embodiment, "wifi", "wi-fi", and "wireless" may be referred to as a set of synonyms.

By removing the synonymous words 32 with repeated meanings, the amount of data in the corpus 34 is reduced, which can greatly reduce the amount of computation required for the subsequent training corpus 34.

In step 240, the processor 20 inputs the corpus 34 into a type of neural network model 36, obtains a training model 38, tests the training model 38, and obtains a test result 44.

In one embodiment, as shown in Figure 3, in the training phase 320, the corpus is input to the neural network model 36, and the neural network model 36 will sequentially perform convolution layers, ReLU layers, Convolutional layer, ReLU layer, pooling layer, ReLU layer, convolutional layer, ReLU layer, pooling layer and fully connected. However, the calculation of each layer in the neural network model 36 can be adjusted according to the implementation, and it is not limited to this. After a series of calculations, the neural network model 36 outputs a training model 38.

In one embodiment, the intent and entity of the test data 40 are known, and the intent and entity and their corresponding results are also known. In the test phase 330, after the test data 40 is input to the training model 38, the training model 38 outputs the test result 44.

In one embodiment, the processor 20 inputs the test data 40 and the synonymous characters 32 into the training model 38 to test whether the training model 38 can correctly determine the synonymous characters 32 and obtain the answer corresponding to the synonymous characters 32 to train the model 38 Output test result 44.

Then, the processor 20 analyzes the test result 44 to verify the accuracy of the training model 38.

In step 250, the processor 20 analyzes the test result 44 according to a predefined rule 46 to generate an analysis result 48, and generates a new version of the training corpus 50 according to the analysis result 48.

In one embodiment, as shown in FIG. 3, in the analysis stage 340, the processor 20 inputs the test result 44 into the analysis module 45. The analysis module 45 can be implemented by a circuit, firmware or software. The analysis module Pre-defined rules 46 are stored in 45. After the test result 44 is subjected to a series of judgments through the pre-defined rules 46, the analysis module 45 outputs the analysis result 48, and generates a new version of the training corpus 50 according to the analysis result 48.

In one embodiment, please refer to FIG. 4, which is a flowchart of a method 400 for applying a predefined rule 46 to analyze a test result 44 according to an embodiment of the present invention.

In step 410, the processor 20 reads the test result 44.

In one embodiment, the test result 44 includes test one result and test two result. The predefined rule 46 defines test one to compare whether the predicted result 44 is equal to the expected label, and the predefined rule 46 defines test two to be output by the training model 38. Intentions and entities find answers in the knowledge base. In one embodiment, the test result 44 is shown in Table 2. sentence Results of test one Results of test two where is the repair center in Tokyo PASS PASS I want to find a repair center in Taipei city PASS FAIL why my wifi is not working FAIL FAIL In Table 2, “PASS” is used for passing the test, and “FALL” is used for testing errors.

In one embodiment, if both the result of test 1 and the result of test 2 pass, it means that the test result 44 has no errors and no subsequent steps are required. If one of the results of test one and the result of test two is wrong, step 420 is entered.

In step 420, the processor 20 analyzes the error pattern by using a predefined rule 46.

In one embodiment, the error type includes a lack of synonymous word type, a lack of answer type, or another error type.

In one embodiment, if the predefined rule 46 judges that the result of test one is wrong, and the result of test two is also wrong, it means that the training model 38 cannot recognize the predefined word, and this word is not in the synonymous character set, nor can it The intent and entity marked with the training model 38 find the associated answer in the knowledge base. The pre-defined rule 46 judges that this type of error is a type of lack of synonymous characters. For example, in the sentence "why my wifi is not working" in Table 2, the predefined rule 46 judges that the result of test one is wrong, and the result of test two is also wrong, it means that the training model 38 may not be able to recognize "wifi", and This word is also not in the synonymous character set, so the associated answer cannot be found in the knowledge base.

In one embodiment, if the predefined rule 46 judges that the result of test one is passed and the result of test two is wrong, it means that the training model 38 can correctly identify the intention and entity of the sentence, but cannot use the intention and entity as an index to find knowledge The answer to which the library is associated. The pre-defined rule 46 judges this type of error as a lack of answer type. For example, in the sentence "I want to find a repair center in Taipei city" in Table 2, the pre-defined rule 46 determines that the result of test one is passed and the result of test two is wrong, which means that the training model 38 may not be able to use the intention and entity as The index finds the answers associated with the knowledge base.

In one embodiment, other error types refer to unknown errors, such as insufficient neurons in a neural network, insufficient number of layers, or an algorithm error, which requires the assistance of a data engineer to adjust.

The error pattern analyzed by the processor 20 according to the above method can be regarded as the analysis result 48, and different operations are performed corresponding to different error patterns in the subsequent steps, for example, a new version of the training corpus 50 is generated according to the analysis result 48 and the corpus 34.

In step 430, the processor 20 determines that the error type is the lack of synonyms or the lack of answers. When the processor 20 determines that the error type is the type lacking synonymous characters, step 440 is entered. When the processor 20 determines that the error type is a lack of answer type, step 450 is entered. In addition, when the processor 20 determines that the error type is another error type, the process ends.

In step 440, when the processor 20 analyzes that at least one error pattern in the test result 44 is the lack of a synonymous character pattern, it means that a specific synonymous character (for example, "wifi") cannot be analyzed by the training model 38, then process The device 20 adds a specific synonym (for example, "wifi") to the corpus 34 to update the corpus 34, and inputs the updated corpus 34 (ie, the new version of the training corpus 50) into the neural network model 36, To generate a new version of the training model.

In this way, the chat robot system 100 can add unrecognized synonyms to the corpus 34 to avoid recurring errors. After repeating the process 320~340 many times, the training model can be continuously optimized.

In step 450, when the processor 20 analyzes the intention and the entity in the test result 44 to be correct, but at least one error type is a lack of answer type, representing the intention (for example, to solve the "wifi" damage problem) and the entity (for example, The association between the combination of "wifi") and the answer (for example, restarting) is not established, and the processor 20 searches for the answer associated with the knowledge base according to the entity and intent as an index (for example, the entities "wifi" and "wireless" are Synonymous words, you can use "wireless" and intent as an index to find out the answer associated with the knowledge base), and associate the entity and intent with the answer to update the knowledge base.

The chat robot system and chat robot model training method shown in the present invention remove synonymous words in the training corpus through the generalization ability of the machine learning model to generate the size of the corpus used for the input neural network model , Which can reduce the amount of calculation and save time. By analyzing the training model, a new version of the training corpus can be generated based on the analysis result, and the new version of the training corpus can be input into the neural network model to achieve the effect of optimizing the training model.

The method of the present invention, or a specific type or part thereof, can exist in the form of code. The program code can be included in physical media, such as floppy disks, CDs, hard disks, or any other machine-readable (such as computer-readable) storage media, or not limited to external forms of computer program products. Among them, When the program code is loaded and executed by a machine, such as a computer, the machine becomes a device for participating in the present invention. The code can also be transmitted through some transmission media, such as wire or cable, optical fiber, or any transmission type. When the code is received, loaded and executed by a machine, such as a computer, the machine becomes used to participate in this Invented device. When implemented in a general-purpose processing unit, the program code combined with the processing unit provides a unique device that operates similar to the application of a specific logic circuit.

Although the present invention has been disclosed in the above embodiments, it is not intended to limit the present invention. Anyone who is familiar with the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection of the present invention The scope shall be subject to the definition of the attached patent application scope.

100: Chatbot system 10: Storage device 20: Processor 200: Chatbot model training method 210~250, 410~450: steps 310: Pre-processing stage 30: Training corpus 32: Synonyms 34: Corpus 320: Training phase 330: Test phase 340: Analysis phase 36: Neural Network Model 38: Training model 40: Test data 44: Test results 45: Analysis Module 46: Pre-defined rules 48: Analysis results 50: New version of training corpus 400: The method of applying pre-defined rules to analyze test results

FIG. 1 is a block diagram of an integrated chat robot system according to an embodiment of the present invention. Figure 2 is a flowchart of a chat robot model training method according to an embodiment of the present invention. FIG. 3 is a schematic diagram of a chat robot model training method according to an embodiment of the present invention. FIG. 4 is a flowchart of a method for analyzing test results by applying predefined rules according to an embodiment of the present invention.

200: Chatbot model training method 210~250: Steps

Claims (6)

A chat robot system includes: a storage device for storing a knowledge base for recording the association between a combination of an intention and an entity and an answer; and a processor for analyzing a plurality of trainings The intention and the entity in the corpus will be used to separate the plural synonymous characters in the training corpus using a generalization ability of a machine learning model, and the synonymous characters will be separated from the training words. After the data is removed, a corpus is obtained. After the corpus is input into a neural network model, a training model is obtained, and the training model is tested to obtain a test result; wherein, when the processor receives the test result from the In the result, a predefined rule is used to analyze the test result to generate an analysis result, and a new version of the training corpus is generated according to the analysis result; wherein, the processor applies the predefined rule to analyze at least one of the test results Error type, the at least one error type can be regarded as the analysis result, and the at least one error type includes a lack of synonyms, a lack of answers, or other error types; wherein, when the processor analyzes the When the at least one error pattern in the test result is the lack of synonymous character pattern, it means that a specific synonymous character cannot be analyzed by the training model. Then the processor adds the specific synonymous character to the corpus to update the corpus The corpus, the updated corpus is input into this type of neural network model to generate the new version of the training model. For example, the chat robot system described in claim 1, wherein, when the processor analyzes the intention in the test result and the entity judges correct, but the at least one error type is the lack of answer type, which represents the intention And the association between the entity’s combination and the answer has not been established, the processor is based on the entity and the intention Use as an index to find the answer associated with the knowledge base, and associate the entity and the intention with the answer to update the knowledge base. For example, the chat robot system described in item 1 of the scope of patent application, wherein the processor is further used to input a plurality of test data and the synonyms into the training model to obtain the test result, and analyze the test result to verify the Train the model. A chatbot model training method includes: recording the association between a combination of an intent and an entity and an answer through a knowledge base; analyzing the intent and the entity in a plurality of training corpora; using a machine learning model A generalization ability separates plural synonymous characters in the training corpus, and removes the synonymous characters from the training corpus to obtain a corpus; input the corpus A type of neural network model to obtain a training model, test the training model to obtain a test result; and analyze the test result according to a predefined rule to generate an analysis result, and generate a new version of the training corpus according to the analysis result ; Apply the predefined rules to analyze at least one error type in the test result, the at least one error type, the at least one error type can be regarded as the analysis result, and the at least one error type includes a lack of synonyms Pattern, a lack of answer pattern, or another error pattern; when the at least one error pattern in the test result is the lack of synonym pattern, it means that a specific synonym cannot be analyzed by the training model, then The same as The semantic characters are added to the corpus to update the corpus, and the updated corpus is input into the neural network model to generate the new version of the training corpus. For example, the chat robot model training method described in item 4 of the scope of patent application further includes: when the intention in the verification result and the entity judgment are correct, but the at least one error type is the lack of answer type, which represents the intention And the association between the combination of the entity and the answer has not been established, then the answer to the knowledge base is searched based on the entity and the intent as an index, and the entity and the intent are associated with the answer to update the answer knowledge base. For example, the chat robot model training method described in item 4 of the scope of patent application further includes: inputting a plurality of test data and the synonyms into the training model to obtain the test result, and analyzing the test result to verify the training model.

Images