WO2021243828A1 - Text processing method and apparatus based on machine learning, and computer device and medium - Google Patents

Abstract

Disclosed are a text processing method and apparatus based on machine learning, and a computer device and a medium, which relate to the field of intelligent decision making in artificial intelligence. The method comprises: acquiring question answering data to be processed, and preprocessing said question answering data to obtain standard question answering data, the standard question answering data comprising standard material information and standard question information (S10); inputting the standard question information in the standard question answering data into a preset question answering classification model to obtain a question type of the standard question information (S20); inputting the standard material information, the standard question information and the corresponding question type into a preset target machine reading comprehension model for prediction to obtain initial answer information, the initial answer information comprising a plurality of pieces of evaluation data information and problem-solving thought information corresponding to the standard question information, wherein the target machine reading comprehension model is obtained by means of training using a convolutional neural network-pretraining language model (S30); and determining final evaluation data from among the plurality of pieces of evaluation data information according to the problem-solving thought information, and recording the final estimation data and the problem-solving thought information as target answer information in a preset integration mode (S40). The method improves the accuracy of an answer obtained by means of machine reading.

Description

Text processing method, device, computer equipment and medium based on machine learning

This application claims the priority of a Chinese patent application filed with the Chinese Patent Office on June 5, 2020, the application number is 202010502599.4, and the invention title is "machine learning-based text processing methods, devices, computer equipment and media", and its entire contents Incorporated in this application by reference.

Technical field

This application relates to the field of intelligent decision-making in artificial intelligence, and in particular to a text processing method, device, computer equipment and medium based on machine learning.

Background technique

At present, deep learning has achieved fruitful results in image recognition, speech recognition and other fields. Machine Reading Comprehension (MRC) has become a new hot spot in the field of artificial intelligence research and application. Its main function is to read and understand a given article or Context, automatically give answers to related questions. At present, the traditional method of machine reading comprehension mainly adopts the method of determining the correct answer based on similarity or correlation. This type of method determines the correct answer by calculating the most similarity or correlation between the sentence of the option and the background material. However, Sentences that are semantically equivalent are often expressed in different forms of syntactic structure. Methods based on similarity and relevance can only find sentences in the background material that have a higher degree of similarity to the syntax structure or semantic expression of the options, and cannot understand the semantics. The nuances, and the nuances between sentences are the first priority in language processing. The inventor realizes that such methods are based on background materials to make correct answers and cannot output the corresponding problem-solving process; therefore, the accuracy of the answers obtained by current machine reading is low, and they cannot be used to assist teaching/learning in a real sense. effect.

Summary of the invention

The embodiments of the present application provide a text processing method, device, computer equipment, and medium based on machine learning to solve the problem of low accuracy of answers obtained by machine reading.

A text processing method based on machine learning, including:

Obtain the answer data to be processed, and preprocess the answer data to be processed to obtain standard answer data, where the standard answer data includes standard material information and standard question information;

Input the standard question information in the standard answer data into a preset answer classification model to obtain the question type of the standard question information;

The standard material information, the standard question information, and the corresponding question type are input into a preset target machine reading comprehension model for prediction, and initial answer information is obtained. The initial answer information includes multiple evaluation data information and Problem-solving idea information corresponding to the standard topic information, wherein the target machine reading comprehension model is obtained by training using a convolutional neural network-pre-training language model;

The final evaluation data is determined from a plurality of the evaluation data information according to the problem-solving idea information, and the final evaluation data and the problem-solving idea information are recorded as target answer information in a preset integration manner.

A text processing device based on machine learning includes:

The preprocessing module is used to obtain the answer data to be processed, and preprocess the answer data to be processed to obtain standard answer data, where the standard answer data includes standard material information and standard question information;

The first input module is configured to input the standard question information in the standard answer data into a preset answer classification model to obtain the question type of the standard question information;

The prediction module is used to input the standard material information, the standard question information, and the corresponding question type into a preset target machine reading comprehension model for prediction to obtain initial answer information, where the initial answer information includes multiple Pieces of evaluation data information and problem-solving ideas information corresponding to the standard topic information, wherein the target machine reading comprehension model is obtained by training using a convolutional neural network-pre-training language model;

The determining module is configured to determine final evaluation data from a plurality of the evaluation data information according to the problem-solving idea information, and record the final evaluation data and the problem-solving idea information as a target answer in a preset integration manner information.

A computer device includes a memory and a processor, the processor and the memory are connected to each other, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is used to execute the The program instructions of the memory, wherein:

Obtain the answer data to be processed, and preprocess the answer data to be processed to obtain standard answer data, where the standard answer data includes standard material information and standard question information;

Input the standard question information in the standard answer data into a preset answer classification model to obtain the question type of the standard question information;

The standard material information, the standard question information, and the corresponding question type are input into a preset target machine reading comprehension model for prediction, and initial answer information is obtained. The initial answer information includes multiple evaluation data information and Problem-solving idea information corresponding to the standard topic information, wherein the target machine reading comprehension model is obtained by training using a convolutional neural network-pre-training language model;

The final evaluation data is determined from a plurality of the evaluation data information according to the problem-solving idea information, and the final evaluation data and the problem-solving idea information are recorded as target answer information in a preset integration manner.

A computer-readable storage medium, the computer-readable storage medium stores a computer program, the computer program includes program instructions, and when the program instructions are executed by a processor, they are used to implement the following steps:

Obtain the answer data to be processed, and preprocess the answer data to be processed to obtain standard answer data, where the standard answer data includes standard material information and standard question information;

Input the standard question information in the standard answer data into a preset answer classification model to obtain the question type of the standard question information;

The standard material information, the standard question information, and the corresponding question type are input into a preset target machine reading comprehension model for prediction, and initial answer information is obtained. The initial answer information includes multiple evaluation data information and Problem-solving idea information corresponding to the standard topic information, wherein the target machine reading comprehension model is obtained by training using a convolutional neural network-pre-training language model;

The final evaluation data is determined from a plurality of the evaluation data information according to the problem-solving idea information, and the final evaluation data and the problem-solving idea information are recorded as target answer information in a preset integration manner.

The above-mentioned text processing methods, devices, computer equipment and media based on machine learning obtain standard answer data by taking the answer data to be processed and preprocessing the answer data to be processed. The standard answer data includes standard material information and standard question information; The standard question information in the answer data is input into the preset answer classification model to obtain the question type of the standard question information; the standard material information, standard question information and the corresponding question type are input into the preset target machine reading comprehension model Predict and obtain initial answer information. The initial answer information includes multiple evaluation data information and problem-solving ideas information corresponding to the standard problem information. The target machine reading comprehension model is obtained by training with a convolutional neural network-pre-training language model的; According to the problem-solving idea information, the final evaluation data is determined from a plurality of the evaluation data information, and the final evaluation data and the problem-solving idea information are recorded as the target answer information in a preset integration mode; by using a convolutional neural network- The target machine reading comprehension model trained by the pre-training language model predicts the answer to the answer data to be processed, and obtains the target answer information that contains both the evaluation data information and the corresponding problem-solving idea information; thereby further improving the accuracy of the answer obtained by machine reading And the real meaning plays a role in assisting teaching/learning.

Description of the drawings

FIG. 1 is a schematic diagram of an application environment of a text processing method based on machine learning in an embodiment of the present application;

2 is a flowchart of a text processing method based on machine learning in an embodiment of the present application;

FIG. 3 is another flowchart of a text processing method based on machine learning in an embodiment of the present application;

4 is another flowchart of a text processing method based on machine learning in an embodiment of the present application;

FIG. 5 is another flowchart of a text processing method based on machine learning in an embodiment of the present application;

Fig. 6 is another flowchart of a text processing method based on machine learning in an embodiment of the present application;

FIG. 7 is a functional block diagram of a text processing device based on machine learning in an embodiment of the present application;

FIG. 8 is another functional block diagram of a text processing device based on machine learning in an embodiment of the present application;

FIG. 9 is another functional block diagram of a text processing device based on machine learning in an embodiment of the present application

Fig. 10 is a schematic diagram of a computer device in an embodiment of the present application.

detailed description

The technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of them. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of this application.

The text processing method based on machine learning provided by the embodiment of the present application can be applied to the application environment shown in FIG. 1. Specifically, the text processing method based on machine learning is applied in a text processing system based on machine learning. The text processing system based on machine learning includes a client and a server as shown in FIG. Communication is used to solve the problem of low accuracy of answers obtained by machine reading. Among them, the client is also called the client, which refers to the program that corresponds to the server and provides local services to the client. The client can be installed on, but not limited to, various personal computers, notebook computers, smart phones, tablet computers, and portable wearable devices. The server can be implemented with a standalone server or a server cluster composed of multiple servers.

In an embodiment, as shown in FIG. 2, a text processing method based on machine learning is provided. The application of the method to the server in FIG. 1 is taken as an example for description, including the following steps:

S10: Obtain the answer data to be processed, and preprocess the answer data to be processed to obtain standard answer data. The standard answer data includes standard material information and standard question information.

Among them, the answer data to be processed refers to the reading comprehension data to be processed. Each piece of reading comprehension data is regarded as a pending answer data. The language of the answer data to be processed can be Chinese or English. Specifically, the answer data to be processed mainly includes reading materials and question information. Among them, the topic information is mainly composed of questions and corresponding candidate answers. The reading material can be single-paragraph text or multi-paragraph text. A piece of reading material in the answer data to be processed may correspond to one or more item information. Optionally, to obtain the answer data to be processed, any piece of reading comprehension data can be obtained directly from the test system, or any piece of reading comprehension data on the paper answer sheet can be scanned and recognized.

Specifically, the preprocessing of the answer data to be processed mainly includes format judgment and processing of the answer data to be processed, to determine whether the format of the answer data to be processed meets preset conditions. In this embodiment, only the answer data to be processed in English format can Input to the machine reading comprehension model for answer prediction. Therefore, if the text format of the answer data to be processed is Chinese, the answer data to be processed in Chinese format needs to be converted into the answer data to be processed in English format.

Further, after the standard format text of the answer to be processed is determined, the answer data to be processed is assembled into the answer data in json format, and the json string in the answer data to be processed meets the requirements, such as judging the answer to be processed Whether the key in the data is vacant, whether the value type meets the requirements, whether the value length is within the range, etc. If the json string in the answer data to be processed does not meet the requirements, the answer data to be processed is returned to the client interface and Perform an abnormal display, prompting the user that the pending answer data is illegal data, and the pending answer data needs to be re-acquired.

Preferably, in order to avoid the reduction of the efficiency of answer prediction due to the excessively large number of characters in the acquired answer data to be processed, in this embodiment, if the number of characters in the answer data to be processed exceeds the preset character threshold, then It is necessary to perform character segmentation processing on the answer data to be processed according to the real-time situation. For example, one answer data to be processed that originally contains a piece of reading material and multiple item information can be divided into multiple answer data to be processed, and each answer data to be processed includes One reading material and one topic information.

Specifically, after preprocessing the answer data to be processed, qualified standard answer data is obtained. The standard answer data includes standard material information and standard question information. Among them, the standard material information is material information that meets the requirements after preprocessing the material information in the answer data to be processed. The standard item information is the item information that meets the requirements after preprocessing the item information in the answer data to be processed.

S20: Input the standard question information in the standard answer data into the preset answer classification model to obtain the question type of the standard question information.

Specifically, one standard answer data may include one or more standard question information, and the question types corresponding to different standard question information may be different. For example, the standard question information included in a standard answer data may be a full-text inference question, a paragraph reasoning question, or a summary multiple-choice question. In this embodiment, in order to improve the accuracy of model prediction, before the standard answer data is input into the machine reading comprehension model for prediction, the type of each standard question information in the standard answer data is determined.

Specifically, by inputting each standard question information in the standard answer data into a preset answer classification model, the question type of each standard question information can be obtained. Among them, the answer classification model is a pre-trained model that can identify the standard question information, thereby determining the question type of the standard question information. In this embodiment, the question type of the classified standard question information may be a vocabulary question, a highlight question, a full text inference question, an insertion question, a paragraph reasoning question, a summary multiple choice question, or a connection question.

Among them, the answer classification model is preferably a machine learning Bayesian model. Specifically, the machine learning Bayes model is obtained by training a large amount of topic information that has been classified and labeled in advance. Among them, Bayesian decision theory (Bayesian decision theory) is the basic method to implement decision-making under the framework of probability. It is a combination of Decision theory + Probability theory. It discusses how to make optimal decisions in an environment containing uncertainty. For classification tasks, in an ideal situation where all relevant probabilities are known, Bayesian Decision theory considers how to select the optimal category label based on these probabilities and misjudgment losses (probability knowledge + knowledge of the loss caused by the decision → optimal decision).

S30: Input the standard material information, standard question information and the corresponding question type into the preset target machine reading comprehension model for prediction, and obtain initial answer information. The initial answer information includes multiple evaluation data information and information corresponding to the standard question information. Problem-solving ideas information, where the target machine reading comprehension model is obtained by training using a convolutional neural network-pre-training language model.

Among them, the target machine reading comprehension model refers to a pre-trained model that can predict answers and analyze problem-solving ideas. The target machine reading comprehension model is obtained by training with a convolutional neural network-pretrained language model. Convolutional neural network-pre-training language model is a model obtained by combining convolutional neural network model and pre-training language model. Understandably, the convolutional neural network-pre-training language model is equivalent to the model formed by connecting the convolutional neural network and the pre-training language network model.

Specifically, the target machine reading comprehension model mainly includes a prediction layer, a reasoning layer, an encoding layer, and a data layer. In this embodiment, the prediction layer includes several prediction units, and each prediction unit corresponds to one type of standard title information. For example, the prediction layer can include vocabulary item unit, highlight item unit, full-text inference item unit, insertion item unit, paragraph inference item unit, summary multiple-choice item unit, and connection item unit. Specifically, when the standard topic information is input into the prediction layer of the target machine reading comprehension model, it is necessary to input the standard topic information into the corresponding prediction unit for prediction according to the topic type of the standard topic information; thus, the standard topic is obtained. At least one standard candidate text of the information. The reasoning layer mainly includes the RoBerta unit and the XLNet unit. The RoBerta unit mainly obtains the selection probability value of each standard candidate text by combining the standard candidate text and the standard material information. The XLNet unit mainly processes the standard candidate text and standard material information, and obtains the key information of the standard material information. Among them, the selection probability value is the probability value used to evaluate the standard candidate text as the correct answer. The key information of the standard material information is the information after labeling and parsing each sentence in the standard material information. For example: mark which sentence in the standard material information is the central opinion sentence, the sub-thesis sentence, and the non-opinion sentence.

Further, after obtaining the selection probability value of each standard candidate text and the key information of the standard material information, the coding layer is used to perform feature encoding on the selection probability value of each standard candidate text and the key information of the standard material information. And input the selection probability value of each standard candidate text for feature encoding and the key information of the standard material information into the data layer, so as to obtain the initial answer information. The initial answer information includes multiple evaluation data information and problem-solving ideas information corresponding to the standard problem information. Among them, the evaluation data information is the selection probability value corresponding to each candidate answer in the standard topic information. Since one standard question information includes at least two candidate answers, the initial answer information obtained includes multiple evaluation data information. Each candidate answer corresponds to an evaluation data message. Problem-solving thinking information is the process of analyzing the normal answer derived from the standard topic information, that is, the reason and understanding process of why this answer was chosen.

S40: Determine final evaluation data from multiple evaluation data information according to the problem-solving idea information, and record the final evaluation data and the problem-solving idea information as target answer information in a preset integration manner.

Specifically, since the initial answer information includes multiple evaluation data information, and each evaluation data information is a probability value corresponding to each candidate answer in the standard topic information. Therefore, after the probability value corresponding to each candidate answer in the standard question information is determined, the probability value corresponding to each candidate answer is screened according to the problem-solving idea information and the question requirements in the standard question information. The final evaluation data is determined in the evaluation data information, that is, the correct answer corresponding to the standard question is determined, and then the final evaluation data corresponding to the standard question and the corresponding problem-solving idea information are recorded as the target answer in a preset integration method information. Among them, the preset integration method can be to directly combine the final evaluation data and the corresponding problem-solving idea information.

Exemplarily, if the obtained initial answer information includes 4 evaluation data information, which are candidate answer A: 0.81, candidate answer B: 0.92, candidate answer C: 0.95 and candidate answer D: 0.01, the question requirements in the standard question information are Which is a conclusion that is impossible to infer from the material. Therefore, the final evaluation data is determined as the candidate answer D from the four evaluation data information combined with the problem-solving idea information. Understandably, the probability value corresponding to the candidate answer D is the smallest probability value, that is, the candidate answer D is unlikely to be inferred from the material, so the final evaluation data is the candidate answer D. Finally, the final evaluation data and problem-solving ideas information are recorded as target answer information in a preset integration manner. Understandably, the target answer information includes the correct answer to the question and the reason and understanding process of why the answer was chosen.

In this embodiment, the answer data to be processed is obtained, and the answer data to be processed is preprocessed to obtain standard answer data. The standard answer data includes standard material information and standard question information; the standard question information in the standard answer data is input to the preset In the answer classification model, the question type of the standard question information is obtained; the standard material information, the standard question information and the corresponding question type are input into the preset target machine reading comprehension model for prediction, and the initial answer information is obtained. The initial answer information includes Multiple evaluation data information and problem-solving ideas information corresponding to the standard topic information. Among them, the target machine reading comprehension model is obtained by training using a convolutional neural network-pre-training language model; based on the problem-solving idea information from multiple evaluation data information Determine the final evaluation data in the process, and record the final evaluation data and the problem-solving idea information as the target answer information in a preset integration method; the target machine reading comprehension model obtained by using the convolutional neural network-pre-training language model training to deal with the answer The data is used for answer prediction, and the target answer information that contains both the evaluation data information and the corresponding problem-solving idea information is obtained; thus, the accuracy and true meaning of the answers obtained by machine reading are further improved, which plays a role in assisting teaching/learning.

In one embodiment, as shown in FIG. 3, preprocessing the answer data to be processed includes the following steps:

S101: Standardize the text form of the answer data to be processed to obtain the initial answer data.

Specifically, since the language of the acquired answer data to be processed may be in Chinese format or English format, and in this embodiment, only the answer data to be processed in English format can be input into the machine reading comprehension model for answer prediction, therefore, In this step, the text format of the answer data to be processed is standardized, that is, the answer data to be processed is converted into a unified English format to obtain the initial answer data.

S102: Convert the initial answer data into a json data format to obtain candidate answer data.

Specifically, after the initial answer data is determined, the initial answer data is assembled into candidate answer data in json format. Among them, the json data format is a lightweight data exchange format that uses a text format completely independent of programming languages to store and represent data. The concise and clear hierarchical structure of the json data format is not only easy for humans to read and write, but also easy for machine analysis and generation, and can effectively improve network transmission efficiency. Therefore, by converting the initial answer data into the json data format, it is beneficial to the subsequent rapid and accurate data processing.

Specifically, classes or functions that convert various data formats (map, xml or yaml, etc.) into json data format can be pre-written and packaged into a conversion script to convert the initial answer data into candidate answer data in json data format. . When performing data format conversion, first obtain the corresponding conversion scripts according to the data format of the initial answer data, and then execute the corresponding conversion scripts to convert the initial answer data into a json data format to obtain candidate answer data.

S103: Determine whether the json character string in the candidate answer data meets the preset requirement, and if the json character string in the candidate answer data meets the preset requirement, determine the candidate answer data as the standard answer data.

Specifically, judging whether the json string in the candidate answer data meets the preset requirements is mainly to judge whether the key in the json string is vacant, whether the value type meets the requirements, and whether the value length is within the range, etc. In a specific embodiment, the preset type range and the preset length range of the value in the json string that meet the requirements have been preset. If the key in the json string in the candidate answer data is not empty, the value type is within the preset type range, and the length of the value is within the preset length range, it is determined that the json string in the candidate answer data meets the preset requirements , Determine the candidate answer data as the standard answer data.

In another specific embodiment, if it is determined that the json string in the candidate answer data does not meet the preset requirements, that is, the key in the json string in the candidate answer data is vacant, or the value type is not within the preset type range, Or the length of the value is not within the preset length range, the answer data to be processed is returned to the client interface and an abnormal display is performed, prompting the user that the answer data to be processed is illegal data, and the answer data to be processed needs to be retrieved.

In this embodiment, the text format of the answer data to be processed is standardized to obtain the initial answer data; the initial answer data is converted into a json data format to obtain candidate answer data; it is judged whether the json string in the candidate answer data meets the preset requirements , If the json string in the candidate answer data meets the preset requirements, the candidate answer data is determined as the standard answer data; thereby improving the accuracy and uniformity of the obtained standard answer data, and ensuring that the subsequent data is input to the target machine The accuracy of the predictions made in the reading comprehension model.

In one embodiment, as shown in FIG. 4, inputting standard material information, standard question information and corresponding question types into a preset target machine reading comprehension model for prediction to obtain initial answer information specifically includes the following steps:

S301: Input standard material information, standard topic information, and corresponding topic types into the prediction layer of the target machine's reading comprehension model to obtain a standard candidate text set of standard topic information. The standard candidate text set includes at least one standard preparation Select the text.

Among them, the standard candidate text set refers to the text set obtained by separately concatenating the question in the standard topic information and each candidate answer. Among them, the standard candidate text set contains at least one standard candidate text.

Specifically, after the topic type of the standard topic information is determined, the standard material information, the standard topic information, and the corresponding topic type are input into the prediction layer of the target machine reading comprehension model. In this embodiment, the processing logic of the prediction layer corresponding to different types of standard title information is different. That is, multiple types of processing units are included in the prediction layer of the target machine reading comprehension model. Specifically, in this embodiment, the prediction layer of the target machine reading comprehension model includes a vocabulary item unit, a highlight item unit, a full-text inference item unit, an insertion item unit, a paragraph inference item unit, a summary multiple-choice item unit, and a connection item. unit. When inputting standard topic information into the prediction layer of the target machine's reading comprehension model, according to the topic type input of the standard topic information, input the standard topic information into the corresponding prediction unit for prediction; thereby obtaining at least one criterion of the standard topic information Alternative text. For example: if the topic type of the standard topic information is a vocabulary question, when the standard topic information is input into the prediction layer of the target machine's reading comprehension model, it will be based on the topic type associated with the standard topic information: vocabulary question. The standard topic information is automatically input into the vocabulary item unit of the prediction layer of the target machine's reading comprehension model, so as to obtain the standard candidate text set of the standard topic information.

S302: Input each standard candidate text and standard material information in the standard candidate text set into the inference layer of the target machine reading comprehension model to obtain the selection probability value of each standard candidate text and key information of the standard material information.

Among them, the reasoning layer is used to judge whether each standard candidate text can be inferred from the standard material information. The reasoning layer includes RoBerta unit and XLNet unit. Among them, RoBERTa is the enhancement and tuning of BERT. RoBERTa mainly made improvements to the previously proposed BERT in three aspects. One is the specific details of the model and improved the optimization function; the second is the training strategy level, which uses a dynamic mask to train the model, which proves the NSP (Next Sentence Prediction) The lack of training strategy uses a larger batch size; the third is the data level, on the one hand, a larger data set is used, on the other hand, BPE (Byte-Pair Encoding) is used to process text data . XLNet is a general autoregressive pre-training method that learns bidirectional contextual information by maximizing the log likelihood of all possible factorization orders.

Specifically, each standard candidate text and standard material information in the standard candidate text set output by the prediction layer is input into the inference layer of the target machine reading comprehension model; the RoBerta unit is used to process the standard candidate text and standard material information , So as to obtain the selection probability value of each standard candidate text, and use the XLNet unit to process the standard candidate text and the standard material information to obtain the key information of the standard material information. Among them, the selection probability value is the probability value used to evaluate the standard candidate text as the correct answer. The range of the selection probability value is 0-1. The higher the selection probability value, the greater the probability that the corresponding standard candidate text is the correct answer. The key information of the standard material information is the information after labeling and parsing each sentence in the standard material information. For example: which of the standard material information are the central opinion sentence, the sub-thesis sentence and the non-opinion sentence, etc.

Further, the target machine reading comprehension model also includes an encoding layer and a data layer. Among them, the encoding layer is mainly responsible for feature encoding of the standard candidate text and standard material information input to the inference layer. In this embodiment, the encoding layer mainly uses the BERT encoder method and the XLNet encoder method standard candidate text and standard material information for feature encoding. Feature coding. The problem that the data layer solves is the dependence of the Base model, because our reasoning model is not from 0 to 1, but is based on the industry's large-scale training model to do some migration, so the data we are based on include RACE, SQuAD, etc.

S303: Combine the selection probability value of each standard candidate text and the key information of the standard material information to obtain initial answer information.

Specifically, after obtaining the selection probability value of each standard candidate text and the key information of the standard material information, the selection probability value of each standard candidate text and the key information of the standard material information are combined, namely The initial answer information can be obtained.

In this embodiment, standard material information, standard topic information, and corresponding topic types are input into the prediction layer of the target machine's reading comprehension model to obtain a standard candidate text set of standard topic information. The standard candidate text set includes At least one standard candidate text; input each standard candidate text and standard material information in the standard candidate text set into the inference layer of the target machine reading comprehension model to obtain the selection probability value and standard material of each standard candidate text The key information of the information; the selection probability value of each standard candidate text and the key information of the standard material information are combined to obtain the initial answer information; thereby improving the accuracy of the generated initial answer information.

In one embodiment, as shown in FIG. 5, before the standard material information, standard topic information, and corresponding topic types are input into a preset target machine reading comprehension model for prediction, the machine learning-based text processing method also Specifically include the following steps:

S11: Obtain a preset number of sample answer data, and each sample answer data includes key paragraph information, sample questions, and corresponding candidate answer sets.

Among them, the sample answer data refers to the reading comprehension data used for model training. Optionally, the sample answer data can be obtained by directly acquiring several pieces of reading comprehension data from the test system, or by scanning and identifying the reading comprehension data on the paper answer sheet. Each of the sample answer data includes key paragraph information, sample questions, and corresponding candidate answer sets. Among them, the key paragraph information refers to the material information corresponding to the sample question. The sample question refers to the question of the question in the sample answer data. The sample question and the corresponding candidate answer set are the candidate answer items corresponding to the sample question. For example: According to paragraph 2, Athens had all of the following before being a city-state EXCEPT is a sample problem; Aa council made up of aristocrats; B an assembly made up of men; Ca; Democratic that was fully who were elected yearly is the candidate answer set corresponding to the sample question.

It should be noted that, to obtain a preset number of sample answer data, the preset number can be M, where M is a positive integer. The specific value of M can be set according to actual needs. The higher the value of M, the higher the accuracy of subsequent model training, but the extraction efficiency will decrease, and the selection of M can be comprehensively considered in terms of accuracy and efficiency.

S12: Combine the sample question of each sample answer data with each candidate answer in the corresponding candidate answer set to obtain a sample candidate text set of each sample answer data. The sample candidate text set includes at least one sample Alternative text.

Specifically, the sample questions of each sample answer data and each candidate answer in the corresponding candidate answer set are respectively spliced to obtain at least one sample candidate text of each sample answer data.

For example: If the sample question is According to paragraph 2, Athens had all of the following before becoming a city-state EXCEPT; the alternative answer set is Aa council up of aristocrats; B an assembly that was made up of men; Ca constitution Fully democratic; D.officials who were elected yearly; Then, after stitching the sample questions of each sample answer data with each candidate answer in the corresponding candidate answer set, 4 sample candidate texts can be obtained respectively It is: "Athens had a council made up of aristocrats before being a city-state"; Athens had an assembly made up of men before becoming a city-state; "Athens had a before a city-state"; "Athens had a before a city-state"; "; "Athens had officials who were elected annually before becoming a city-state".

S13: Mark the key paragraph information of each sample answer data to obtain the marked data of the key paragraph information.

Specifically, the key paragraph information of each sample answer data is annotated to obtain annotation data of the key paragraph information, where the annotation data is data used to annotate the key information of each sentence in the key paragraph information. For example, the labeling data can be used to label which sentences in the key paragraph information are central point sentences, which sentences are sub-thesis sentences, and which sentences are non-point sentences.

S14: Input the sample candidate text set, key paragraph information, and corresponding annotation data in each sample answer data as training samples into the convolutional neural network-pre-training language model for training, to obtain the target machine reading comprehension model.

Specifically, the sample candidate text set, key paragraph information, and corresponding annotation data in each sample answer data are input as training samples into the convolutional neural network-pre-training language model for training, and the target machine reading comprehension can be obtained Model. Among them, the convolutional neural network-pre-training language model is a model obtained by combining the convolutional neural network model and the pre-training language model. Understandably, the convolutional neural network-pre-training language model is equivalent to the model formed by connecting the convolutional neural network and the pre-training language network model.

In this embodiment, a preset number of sample answer data is obtained, and each sample answer data includes key paragraph information, sample questions, and corresponding candidate answer sets; the sample questions of each sample answer data and the corresponding candidate Each candidate answer in the answer set is spliced to obtain a sample candidate text set of each sample answer data. The sample candidate text set includes at least one sample candidate text; the key paragraph information of each sample answer data is marked, Obtain the annotation data of key paragraph information; input the sample candidate text set, key paragraph information and corresponding annotation data in each sample answer data as training samples into the convolutional neural network-pre-training language model for training, and obtain the target Machine reading comprehension model; thereby improving the accuracy of the generated target machine reading comprehension model.

In one embodiment, as shown in FIG. 6, the sample candidate text set, key paragraph information, and corresponding annotation data in each sample answer data are input as training samples into the convolutional neural network-pre-training language model After training and obtaining the target machine reading comprehension model, the text comprehension processing method based on machine learning also specifically includes the following steps:

S15: Receive an update instruction, and detect whether the minimum risk training loss function in the target machine's reading comprehension model is minimized.

S16: When the minimum risk training loss function is not minimized, after optimizing and adjusting the parameters of the target machine reading comprehension model for a preset number of times, use the preset evaluation function and selected verification answer data to understand the adjusted target machine reading comprehension The accuracy of the model output answer is evaluated, and the evaluation result is obtained; among them, the parameters of the target machine reading comprehension model are optimized and adjusted, including a minimization process for the minimum risk training loss function.

S17: If the evaluation result meets the preset evaluation requirements, record the adjusted target machine reading comprehension model as the new target machine reading comprehension model, so that the standard material information, standard topic information and corresponding topic types can be re-input to the new target machine reading comprehension model. Make predictions in the target machine reading comprehension model to get the initial answer information.

Among them, the update instruction refers to an instruction used to trigger the optimization of the target machine's reading comprehension model. Optionally, the update instruction may be generated when the target machine's reading comprehension model is required to have a more accurate predictive ability, or a trigger cycle may be preset for periodic generation, etc. Specifically, an update instruction is received, and it is detected whether the minimum risk training loss function in the reading comprehension model of the target machine is minimized. If the minimum risk training loss function in the target machine reading comprehension model is not minimization, then the goal is to minimize the minimum risk training loss function, and the parameters of the target machine reading comprehension model are optimized for a preset number of times, and then the target is executed The training of the machine reading comprehension model continuously optimizes the probability distribution of the output answers of the target machine reading comprehension model, so that the answers to the sample questions in the predicted sample answer data are getting closer and closer to the standard answers. Therefore, through a preset number of iterative optimization adjustments, an adjusted target machine reading comprehension model can be obtained. Among them, the minimum risk training refers to the use of the loss function Δ(y,y ⁽ⁿ⁾ ) to describe the degree of difference between the answer y predicted by the model and the standard answer y ⁽ⁿ⁾ , and to try to find a set of parameters to make the model in the training set The expected value of the loss.

Specifically, the calculation formula of the minimum risk training loss function R(θ) is:

Among them, x ⁽ⁿ⁾ is the sample question in the sample answer data; y is the answer output by the target machine reading comprehension model, P(y|x ⁽ⁿ⁾ ；θ) is the target machine reading comprehension model when the model parameter is θ The output probability value of the answer, Y(x ⁽ⁿ⁾ ) is the set of all possible output answers of the target machine reading comprehension model corresponding to x ^{(n) , Δ(y,y (n)} ) is the answer output by the target machine reading comprehension model The degree of difference (ie loss) from the standard answer y ^(n). In this example, rouge evaluation is used to calculate the loss between the answer output by the target machine's reading comprehension model and the standard answer y ⁽ⁿ⁾ , and define Δ(y, y ⁽ⁿ⁾ ) = 1-rouge(y, y ⁽ⁿ⁾ ). Based on rouge-L, the longest subsequence can be automatically matched. The rouge evaluation in this embodiment adopts rouge-L, and the corresponding calculation formula is: in the above formula, x and y are the text sequence of the standard answer and the model output answer; N is The length of the standard answer; n is the length of the model output answer; β is a hyperparameter, which can be set as required, and the value is 1.2 in this embodiment; LCS is the longest common subsequence. Of course, in specific applications, personalized settings can be made according to specific tasks and needs. Further, after optimizing and adjusting the parameters of the target machine reading comprehension model for a preset number of times, the preset evaluation function and selected verification answer data are used to evaluate the accuracy of the adjusted target machine reading comprehension model output answer. Obtain the evaluation result; among them, an optimization adjustment is performed on the parameters of the target machine reading comprehension model, including a minimization process for the minimum risk training loss function.

Among them, the evaluation result refers to the result obtained after the effect evaluation of the target machine reading comprehension model after parameter adjustment. The verification answer data refers to the data set used to verify the effect of the target machine's reading comprehension model after parameter adjustment. Each verification answer data includes key paragraph information, sample questions and corresponding candidate answer sets. Specifically, after the target machine reading comprehension model is optimized and adjusted for a preset number of times, the selected verification answer data is input into the adjusted target machine reading comprehension model, and then the preset evaluation function is used, such as ROUGE (Recall-Oriented Understudy ForGisting Evaluation, evaluation of the understanding of improvement evaluation, BLEU (Bilingual Evaluation Understudy, bilingual evaluation) evaluates the accuracy of the answers output by the adjusted target machine reading comprehension model, and obtains the evaluation result.

Further, after the evaluation result is obtained, it is determined whether the evaluation result meets the preset evaluation requirements. If the evaluation result meets the preset evaluation requirements, the optimization adjustment of the target machine reading comprehension model is stopped, and the adjusted target machine reading comprehension The model is recorded as a new target machine reading comprehension model. Among them, the preset evaluation requirement is when the loss function in the reading comprehension model of the target machine reaches the minimum until it converges. That is, when the evaluation result indicates that the loss function in the target machine reading comprehension model converges during the iterative optimization and adjustment process, and the minimum optimized loss function is obtained, it means that the evaluation result meets the preset evaluation requirements, and the optimization of the target machine reading comprehension model is stopped. Adjust and record the adjusted target machine reading comprehension model as a new target machine reading comprehension model, so that the standard material information, standard topic information and corresponding topic types can be re-input into the new target machine reading comprehension model for prediction , To obtain the initial answer information, thereby further improving the accuracy of the obtained initial answer information.

In another specific embodiment, if the obtained evaluation result does not meet the preset evaluation requirements, continue to optimize and adjust the target machine reading comprehension model to minimize the loss function until it converges, until the evaluation result meets the preset Assess the requirements, and finally record the adjusted target machine reading comprehension model as the new target machine reading comprehension model. Understandably, in this embodiment, each time the target machine reading comprehension model performs an iterative optimization adjustment, an evaluation result will be output accordingly, so that after a preset number of iterative optimization adjustments and evaluations, multiple evaluations will be correspondingly obtained. As a result, until the evaluation results meet the preset evaluation requirements, stop the iterative optimization and adjustment of the target machine reading comprehension model

In this embodiment, an update instruction is received to detect whether the minimum risk training loss function in the target machine reading comprehension model is minimized; when the minimum risk training loss function is not minimized, the parameters of the target machine reading comprehension model are preset After the optimization and adjustment of the number of times, the preset evaluation function and the selected verification answer data are used to evaluate the accuracy of the adjusted target machine reading comprehension model output answer, and the evaluation result is obtained; among them, the parameters of the target machine reading comprehension model are evaluated. An optimization adjustment, including a minimization process for the minimum risk training loss function; if the evaluation result meets the preset evaluation requirements, the adjusted target machine reading comprehension model is recorded as a new target machine reading comprehension model, so as to facilitate Standard material information, standard question information and corresponding question types are re-input into the new target machine reading comprehension model for prediction, and initial answer information is obtained, thereby further improving the accuracy and accuracy of the obtained initial answer information.

It should be understood that the size of the sequence number of each step in the foregoing embodiment does not mean the order of execution. The execution sequence of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present application.

In one embodiment, a text processing device based on machine learning is provided, and the text processing device based on machine learning has a one-to-one correspondence with the text processing method based on machine learning in the foregoing embodiment. As shown in FIG. 7, the machine learning-based text processing device includes a preprocessing module, a first input module 20, a prediction module 30, and an integration module 40. The detailed description of each functional module is as follows:

The preprocessing module 10 is configured to obtain answer data to be processed, and preprocess the answer data to be processed to obtain standard answer data, where the standard answer data includes standard material information and standard question information;

The first input module 20 is configured to input the standard question information in the standard answer data into a preset answer classification model to obtain the question type of the standard question information;

The prediction module 30 is configured to input the standard material information, the standard question information, and the corresponding question type into a preset target machine reading comprehension model for prediction to obtain initial answer information, where the initial answer information includes A plurality of evaluation data information and problem-solving ideas information corresponding to the standard topic information, wherein the target machine reading comprehension model is obtained by training using a convolutional neural network-pre-training language model;

The determining module 40 is configured to determine final evaluation data from a plurality of the evaluation data information according to the problem-solving idea information, and record the final evaluation data and the problem-solving idea information in a preset integration manner as a target Answer information.

Preferably, as shown in FIG. 8, the preprocessing module 10 includes:

The standardization unit 101 is used to standardize the text form of the answer data to be processed to obtain initial answer data;

The conversion unit 102 is configured to convert the initial answer data into a json data format to obtain candidate answer data;

The judging unit 103 is configured to judge whether the json character string in the candidate answer data meets preset requirements, and if the json character string in the candidate answer data meets the preset requirements, determine the candidate answer data as a standard answer data.

Preferably, as shown in FIG. 9, the prediction module 30 includes:

The first input unit 301 is configured to input the standard material information, the standard topic information, and the corresponding topic type into the prediction layer of the target machine reading comprehension model to obtain the standard topic information A selected text set, where the standard candidate text set includes at least one standard candidate text;

The second input unit 302 is configured to input each standard candidate text and the standard material information in the standard candidate text set into the inference layer of the target machine reading comprehension model to obtain each The selection probability value of the standard candidate text and the key information of the standard material information;

The combining unit 303 is configured to combine the selection probability value of each standard candidate text and the key information of the standard material information to obtain initial answer information.

Preferably, the text processing device based on machine learning further includes:

The obtaining module is used to obtain a preset number of sample answer data, each of the sample answer data includes key paragraph information, sample questions, and corresponding candidate answer sets;

The splicing module is used to splice the sample question of each sample answer data with each candidate answer in the corresponding candidate answer set to obtain the sample candidate text of each sample answer data Set, the sample candidate text set includes at least one sample candidate text;

An annotation module, configured to annotate the key paragraph information of each of the sample answer data to obtain the annotation data of the key paragraph information;

The second input module is used to input the sample candidate text set, the key paragraph information and the corresponding annotation data in each of the sample answer data as training samples into the convolutional neural network-pre-training language model Perform training to obtain the target machine reading comprehension model.

Preferably, the text processing device based on machine learning further includes:

The detection module is configured to receive update instructions and detect whether the minimum risk training loss function in the target machine reading comprehension model is minimized;

The optimization adjustment module is used to optimize and adjust the parameters of the target machine reading comprehension model for a preset number of times when the minimum risk training loss function is not minimized, and then use the preset evaluation function and the selected verification answer data, Evaluate the accuracy of the output answers of the adjusted target machine reading comprehension model to obtain the evaluation result; wherein, performing an optimization adjustment on the parameters of the target machine reading comprehension model includes performing the minimum risk training loss function Minimize the processing flow at one time;

The recording module is used to record the adjusted target machine reading comprehension model as a new target machine reading comprehension model when the evaluation result meets the preset evaluation requirements, so as to facilitate the standard material information, standard topic information and corresponding The question type of is re-input into the new target machine reading comprehension model for prediction, and the initial answer information is obtained.

For the specific definition of the text processing device based on machine learning, please refer to the above definition of the text processing method based on machine learning, which will not be repeated here. The various modules in the above-mentioned machine learning-based text processing device can be implemented in whole or in part by software, hardware, and a combination thereof. The foregoing modules may be embedded in the form of hardware or independent of the processor in the computer device, or may be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the foregoing modules.

In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. 10. The computer equipment includes a processor, a memory, a network interface, and a database connected through a system bus. Among them, the processor of the computer device is used to provide calculation and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device is used to store the data used in the text processing method based on machine learning in the foregoing embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. The computer program is executed by the processor to realize a text processing method based on machine learning.

In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, it implements the machine learning-based Text processing method.

In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the machine learning-based text processing method in the foregoing embodiment is implemented. Wherein, the computer-readable storage medium may be non-volatile or volatile.

Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage. In the medium, when the computer program is executed, it may include the procedures of the above-mentioned method embodiments. Wherein, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and/or volatile memory. Non-volatile memory may include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not a limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain Channel (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

The above are only specific implementations of this application, but the protection scope of this application is not limited to this. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed in this application. Should be covered within the scope of protection of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

Claims (20)

1. A text processing method based on machine learning, which includes:

   Obtain the answer data to be processed, and preprocess the answer data to be processed to obtain standard answer data, where the standard answer data includes standard material information and standard question information;

   Input the standard question information in the standard answer data into a preset answer classification model to obtain the question type of the standard question information;

   The standard material information, the standard question information, and the corresponding question type are input into a preset target machine reading comprehension model for prediction, and initial answer information is obtained. The initial answer information includes multiple evaluation data information and Problem-solving idea information corresponding to the standard topic information, wherein the target machine reading comprehension model is obtained by training using a convolutional neural network-pre-training language model;

   The final evaluation data is determined from a plurality of the evaluation data information according to the problem-solving idea information, and the final evaluation data and the problem-solving idea information are recorded as target answer information in a preset integration manner.
2. The text processing method based on machine learning of claim 1, wherein the preprocessing of the answer data to be processed comprises:

   Standardize the text form of the answer data to be processed to obtain initial answer data;

   Convert the initial answer data into a json data format to obtain candidate answer data;

   It is determined whether the json character string in the candidate answer data meets a preset requirement, and if the json character string in the candidate answer data meets the preset requirement, the candidate answer data is determined as standard answer data.
3. The text processing method based on machine learning according to claim 1, wherein said inputting said standard material information, said standard topic information and corresponding said topic type into a preset target machine reading comprehension model is performed Predict, get the initial answer information, including:

   The standard material information, the standard topic information, and the corresponding topic type are input into the prediction layer of the target machine reading comprehension model to obtain a standard candidate text set of the standard topic information, and the standard is prepared The selected text set includes at least one standard candidate text;

   Input each standard candidate text and the standard material information in the standard candidate text set into the reasoning layer of the target machine reading comprehension model to obtain the selection probability value of each standard candidate text And key information of the standard material information;

   Combine the selection probability value of each standard candidate text and the key information of the standard material information to obtain initial answer information.
4. The text processing method based on machine learning according to claim 1, wherein said inputting said standard material information, said standard topic information and corresponding said topic type into a preset target machine reading comprehension model Before making a prediction, it also includes:

   Obtain a preset number of sample answer data, each of the sample answer data includes key paragraph information, sample questions, and corresponding candidate answer sets;

   Respectively splicing the sample question of each sample answer data with each candidate answer in the corresponding candidate answer set to obtain a sample candidate text set of each sample answer data, the sample The candidate text set includes at least one sample candidate text;

   Labeling the key paragraph information of each of the sample answer data to obtain the labeling data of the key paragraph information;

   Input the sample candidate text set, the key paragraph information and the corresponding annotation data in each of the sample answer data as training samples into the convolutional neural network-pre-training language model for training, and obtain the target machine reading Understand the model.
5. The machine learning-based text processing method according to claim 4, wherein the sample candidate text set, the key paragraph information and the corresponding annotation data in each of the sample answer data are used as training The samples are input into the convolutional neural network-pre-training language model for training, and after the target machine reading comprehension model is obtained, the machine learning-based text processing method further includes:

   Receiving an update instruction, and detecting whether the minimum risk training loss function in the target machine reading comprehension model is minimized;

   When the minimum risk training loss function is not minimized, after the parameters of the target machine reading comprehension model are optimized and adjusted a preset number of times, the preset evaluation function and the selected verification answer data are used to compare the adjusted The accuracy of the output answers of the target machine reading comprehension model is evaluated to obtain the evaluation result; wherein, an optimization adjustment of the parameters of the target machine reading comprehension model includes performing a minimization process on the minimum risk training loss function;

   If the evaluation result meets the preset evaluation requirements, the adjusted target machine reading comprehension model is recorded as a new target machine reading comprehension model, so that the standard material information, the standard topic information and the corresponding The question type is input into the new target machine reading comprehension model for prediction, and initial answer information is obtained.
6. 8. The machine learning-based text processing method according to claim 1, wherein, before preprocessing the to-be-processed answer data, the method further comprises:

   Acquiring the number of characters included in the answer data to be processed;

   When the number of characters is greater than a preset character threshold, character segmentation processing is performed on the to-be-processed answer data to obtain a plurality of to-be-processed answer data.
7. The text processing method based on machine learning according to claim 1, wherein the answer classification model is a machine learning Bayes model, and the machine learning Bayes model is obtained by training based on pre-labeled topic information and topic types.
8. A text processing device based on machine learning, which includes:

   The preprocessing module is used to obtain the answer data to be processed, and preprocess the answer data to be processed to obtain standard answer data, where the standard answer data includes standard material information and standard question information;

   The first input module is configured to input the standard question information in the standard answer data into a preset answer classification model to obtain the question type of the standard question information;

   The prediction module is used to input the standard material information, the standard question information, and the corresponding question type into a preset target machine reading comprehension model for prediction to obtain initial answer information, where the initial answer information includes multiple Pieces of evaluation data information and problem-solving ideas information corresponding to the standard topic information, wherein the target machine reading comprehension model is obtained by training using a convolutional neural network-pre-training language model;

   The determining module is used to determine final evaluation data from a plurality of the evaluation data information according to the problem-solving idea information, and record the final evaluation data and the problem-solving idea information as a target answer in a preset integration manner information.
9. A computer device includes a memory and a processor, the processor and the memory are connected to each other, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is used to execute the The program instructions of the memory, wherein:

   Obtain the answer data to be processed, and preprocess the answer data to be processed to obtain standard answer data, where the standard answer data includes standard material information and standard question information;

   Input the standard question information in the standard answer data into a preset answer classification model to obtain the question type of the standard question information;

   The standard material information, the standard question information, and the corresponding question type are input into a preset target machine reading comprehension model for prediction, and initial answer information is obtained. The initial answer information includes multiple evaluation data information and Problem-solving idea information corresponding to the standard topic information, wherein the target machine reading comprehension model is obtained by training using a convolutional neural network-pre-training language model;

   The final evaluation data is determined from a plurality of the evaluation data information according to the problem-solving idea information, and the final evaluation data and the problem-solving idea information are recorded as target answer information in a preset integration manner.
10. The computer device of claim 11, wherein the processor is configured to:

    Standardize the text form of the answer data to be processed to obtain initial answer data;

    Convert the initial answer data into a json data format to obtain candidate answer data;

    It is determined whether the json character string in the candidate answer data meets a preset requirement, and if the json character string in the candidate answer data meets the preset requirement, the candidate answer data is determined as standard answer data.
11. The computer device of claim 11, wherein the processor is configured to:

    The standard material information, the standard topic information, and the corresponding topic type are input into the prediction layer of the target machine reading comprehension model to obtain a standard candidate text set of the standard topic information, and the standard is prepared The selected text set includes at least one standard candidate text;

    Input each standard candidate text and the standard material information in the standard candidate text set into the reasoning layer of the target machine reading comprehension model to obtain the selection probability value of each standard candidate text And key information of the standard material information;

    Combine the selection probability value of each standard candidate text and the key information of the standard material information to obtain initial answer information.
12. The computer device of claim 11, wherein the processor is configured to:

    Obtain a preset number of sample answer data, each of the sample answer data includes key paragraph information, sample questions, and corresponding candidate answer sets;

    Respectively splicing the sample question of each sample answer data with each candidate answer in the corresponding candidate answer set to obtain a sample candidate text set of each sample answer data, the sample The candidate text set includes at least one sample candidate text;

    Labeling the key paragraph information of each of the sample answer data to obtain the labeling data of the key paragraph information;

    Input the sample candidate text set, the key paragraph information and the corresponding annotation data in each of the sample answer data as training samples into the convolutional neural network-pre-training language model for training, and obtain the target machine reading Understand the model.
13. The computer device of claim 14, wherein the processor is configured to:

    Receiving an update instruction, and detecting whether the minimum risk training loss function in the target machine reading comprehension model is minimized;

    When the minimum risk training loss function is not minimized, after the parameters of the target machine reading comprehension model are optimized and adjusted a preset number of times, the preset evaluation function and the selected verification answer data are used to compare the adjusted The accuracy of the output answers of the target machine reading comprehension model is evaluated to obtain the evaluation result; wherein, an optimization adjustment of the parameters of the target machine reading comprehension model includes performing a minimization process on the minimum risk training loss function;

    If the evaluation result meets the preset evaluation requirements, the adjusted target machine reading comprehension model is recorded as a new target machine reading comprehension model, so that the standard material information, the standard topic information and the corresponding The question type is input into the new target machine reading comprehension model for prediction, and initial answer information is obtained.
14. The computer device of claim 9, wherein the processor is configured to:

    Acquiring the number of characters included in the answer data to be processed;

    When the number of characters is greater than a preset character threshold, character segmentation processing is performed on the to-be-processed answer data to obtain a plurality of to-be-processed answer data.
15. 9. The computer device according to claim 9, wherein the answer classification model is a machine learning Bayes model, and the machine learning Bayes model is obtained by training based on pre-labeled topic information and topic types.
16. A computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, the computer program includes program instructions, and when the program instructions are executed by a processor, they are used to implement the following steps:

    Obtain the answer data to be processed, and preprocess the answer data to be processed to obtain standard answer data, where the standard answer data includes standard material information and standard question information;

    Input the standard question information in the standard answer data into a preset answer classification model to obtain the question type of the standard question information;

    The standard material information, the standard question information, and the corresponding question type are input into a preset target machine reading comprehension model for prediction, and initial answer information is obtained. The initial answer information includes multiple evaluation data information and Problem-solving idea information corresponding to the standard topic information, wherein the target machine reading comprehension model is obtained by training using a convolutional neural network-pre-training language model;

    The final evaluation data is determined from a plurality of the evaluation data information according to the problem-solving idea information, and the final evaluation data and the problem-solving idea information are recorded as target answer information in a preset integration manner.
17. 15. The computer-readable storage medium of claim 16, wherein when the program instructions are executed by the processor, they are further used to implement the following steps:

    Standardize the text form of the answer data to be processed to obtain initial answer data;

    Convert the initial answer data into a json data format to obtain candidate answer data;

    It is determined whether the json character string in the candidate answer data meets a preset requirement, and if the json character string in the candidate answer data meets the preset requirement, the candidate answer data is determined as standard answer data.
18. 15. The computer-readable storage medium of claim 16, wherein when the program instructions are executed by the processor, they are further used to implement the following steps:

    The standard material information, the standard topic information, and the corresponding topic type are input into the prediction layer of the target machine reading comprehension model to obtain a standard candidate text set of the standard topic information, and the standard is prepared The selected text set includes at least one standard candidate text;

    Input each standard candidate text and the standard material information in the standard candidate text set into the reasoning layer of the target machine reading comprehension model to obtain the selection probability value of each standard candidate text And key information of the standard material information;

    Combine the selection probability value of each standard candidate text and the key information of the standard material information to obtain initial answer information.
19. 15. The computer-readable storage medium of claim 16, wherein when the program instructions are executed by the processor, they are further used to implement the following steps:

    Obtain a preset number of sample answer data, each of the sample answer data includes key paragraph information, sample questions, and corresponding candidate answer sets;

    Respectively splicing the sample question of each sample answer data with each candidate answer in the corresponding candidate answer set to obtain a sample candidate text set of each sample answer data, the sample The candidate text set includes at least one sample candidate text;

    Labeling the key paragraph information of each of the sample answer data to obtain the labeling data of the key paragraph information;

    Input the sample candidate text set, the key paragraph information and the corresponding annotation data in each of the sample answer data as training samples into the convolutional neural network-pre-training language model for training, and obtain the target machine reading Understand the model.
20. The computer-readable storage medium of claim 19, wherein when the program instructions are executed by the processor, they are further used to implement the following steps:

    Receiving an update instruction, and detecting whether the minimum risk training loss function in the target machine reading comprehension model is minimized;

    When the minimum risk training loss function is not minimized, after the parameters of the target machine reading comprehension model are optimized and adjusted a preset number of times, the preset evaluation function and the selected verification answer data are used to compare the adjusted The accuracy of the output answers of the target machine reading comprehension model is evaluated to obtain the evaluation result; wherein, an optimization adjustment of the parameters of the target machine reading comprehension model includes performing a minimization process on the minimum risk training loss function;

    If the evaluation result meets the preset evaluation requirements, the adjusted target machine reading comprehension model is recorded as a new target machine reading comprehension model, so that the standard material information, the standard topic information and the corresponding The question type is input into the new target machine reading comprehension model for prediction, and initial answer information is obtained.

Images