CN-114328874-B - Open test question answer scoring method, device, equipment and storage medium

Abstract

The invention discloses a method, a device, equipment and a storage medium for scoring answers of open test questions. The method comprises the steps of obtaining target test questions and target answers, respectively extracting word and sentence characteristics of the target test questions and the target answers to obtain test question feature sets and answer feature sets, carrying out similarity measurement on the test question feature sets and the answer feature sets to obtain semantic similarity and character similarity, determining test question intention vectors of the test question feature sets and answer intention vectors of the answer feature sets by adopting a pre-trained target intention analysis model, and determining answer scores of the target answers according to the semantic similarity, the character similarity, the test question intention vectors and the answer intention vectors and the pre-trained target scoring model. The open test question answer scoring method provided by the invention can be integrated in the intelligent equipment, so that a user can use the intelligent equipment to score answers of open test questions rapidly and accurately, the paper reading efficiency is improved, and the open test question answer scoring method has breakthrough improvement in the field of intelligent education.

Inventors

• LIU XIAOJUN
• CAO XIAOYE

Assignees

• 上海米学人工智能信息科技有限公司

Dates

Publication Date

20260512

Application Date

20211229

Claims (10)

1. 1. The open test question answer scoring method is characterized by comprising the following steps: Acquiring a target test question and a target answer, and respectively extracting word and sentence characteristics of the target test question and the target answer to obtain a test question characteristic set and an answer characteristic set; Performing similarity measurement on the test question feature set and the answer feature set to obtain semantic similarity and character similarity; Determining a test question intention vector of the test question feature set and an answer intention vector of the answer feature set by adopting a pre-trained target intention analysis model; Determining an answer score of the target answer according to the semantic similarity, the character similarity, the test question intention vector and the answer intention vector by combining a pre-trained target scoring model; the training process of the target intention analysis model comprises the following steps: Constructing an intention analysis model to be trained comprising a test question intention analysis network, an answer intention analysis network and an intention classifier, and initializing the number of intention categories of the intention classifier into a preset number of categories; Training the intention analysis model to be trained based on training word embedding data, and updating model parameters of the intention analysis model to be trained; Testing the updated intention analysis model to be trained based on the test word embedding data, and adjusting the number of intention categories of the intention classifier; the intention analysis model to be trained is trained and tested again, the adjustment trend of the number of the intention categories is determined, and when the adjustment trend meets the adjustment stability condition, the intention analysis model to be trained updated for the last time is determined to be a target intention analysis model; the method for testing the intention analysis model to be trained based on the test word embedding data comprises the steps of: acquiring test word embedding data of a preset test number group, and inputting each test word embedding data into an updated intention analysis model to be trained, wherein the test word embedding data comprises a test question word embedding sequence and a test answer word embedding sequence; Aiming at each group of test word embedded data, determining a test question intention vector of the test question word embedded sequence by utilizing the test question intention analysis network, and determining a test answer intention vector of the test answer word embedded sequence by utilizing the answer intention analysis network; determining a vector distance distribution condition according to Euclidean distance between each test question intention vector and the corresponding test answer intention vector; And when the vector distance distribution condition meets a preset distribution condition, subtracting one from the number of the intention categories of the intention classifier, otherwise, adding one to the number of the intention categories of the intention classifier.
2. 2. The method for scoring answers to open test questions according to claim 1, wherein the extracting the word and sentence feature of the target test questions and the target answers to obtain a test question feature set and an answer feature set respectively comprises: sentence embedding operation is carried out on the target test questions and the target answers respectively, and test question sentence embedding vectors and answer sentence embedding vectors are obtained; performing word segmentation operation on the target test questions to obtain a test question Chinese word segmentation set and a test question English word segmentation set, and performing word embedding operation on the test question Chinese word segmentation in the test question Chinese word segmentation set to obtain a test question word embedding sequence; performing word segmentation operation on the target answers to obtain an answer Chinese word segmentation set and an answer English word segmentation set, and performing word embedding operation on answer Chinese word segmentation in the answer Chinese word segmentation set to obtain an answer word embedding sequence; and forming a test question feature set by the test question sentence embedding vector, the test question word embedding sequence, the test question Chinese word segmentation set and the test question English word segmentation set, and forming an answer feature set by the answer sentence embedding vector, the answer word embedding sequence, the answer Chinese word segmentation set and the answer English word segmentation set.
3. 3. The method for scoring answers to open test questions according to claim 2, wherein the step of measuring similarity between the test question feature set and the answer feature set to obtain semantic similarity and character similarity comprises: Determining semantic similarity according to the test question embedding vector and the answer sentence embedding vector; Performing Chinese character similarity measurement on the test question Chinese word segmentation set and the answer Chinese word segmentation set to obtain Chinese character similarity, and performing English character similarity measurement on the test question English word segmentation set and the answer English word segmentation set to obtain English character similarity; and determining the character similarity according to the Chinese character similarity and the English character similarity.
4. 4. The method for scoring answers to open questions of claim 2, wherein determining the answer intention vector of the set of questions features and the answer intention vector of the set of answer features using a pre-trained target intention parsing model comprises: Performing intention analysis on the test question word embedding sequence by using a target test question intention analysis network in a target intention analysis model to obtain a test question intention vector; And carrying out intention analysis on the answer word embedding sequence by using a target answer intention analysis network in a target intention analysis model to obtain an answer intention vector.
5. 5. The method for scoring answers to open questions according to claim 1, wherein the training the intent-to-train analytical model based on training word embedding data, updating model parameters of the intent-to-train analytical model, comprises: acquiring training word embedding data of a preset training quantity group, inputting each training word embedding data into the intention analysis model to be trained for training, and updating model parameters of the intention analysis model to be trained, wherein the training word embedding data comprises a training test word embedding sequence and a training answer word embedding sequence; and returning to the operation of re-acquiring the embedded data of the training words of the preset training quantity group and updating the model parameters until the updating times reach the preset updating times.
6. 6. The method for scoring answers to open questions according to claim 5, wherein the step of inputting each of the training word embedded data into the intention-to-be-trained analytical model to train, and updating model parameters of the intention-to-be-trained analytical model comprises: aiming at each group of training word embedded data, determining training test question intention vectors of the training test question word embedded sequence by utilizing the test question intention analysis network, determining training answer intention vectors of the training answer word embedded sequence by utilizing the answer intention analysis network, and carrying out intention classification on the training test question intention vectors and the training answer intention vectors according to an intention classifier to obtain training test question intention types and training answer intention types; Obtaining an intention fitting loss function according to the intention type of each training test question and the corresponding intention type of the training answer; And back-propagating the intention analysis model to be trained through the intention fitting loss function, and updating model parameters of the intention analysis model to be trained.
7. 7. The open question answer scoring method of claim 1, wherein the training process of the objective scoring model comprises: scoring and labeling the training feature combination data to obtain standard scores corresponding to the training feature combination data, wherein the training feature combination data comprises standard semantic similarity, standard character similarity, standard test question intention vectors and standard answer intention vectors; inputting the training feature combination data into a scoring model to be trained to obtain an output actual score; Obtaining a scoring fitting loss function according to the standard score and the actual score; and back-propagating the scoring model to be trained through the scoring fitting loss function to obtain the scoring model.
8. 8. An open test question answer scoring device, comprising: the language processing module is used for acquiring a target test question and a target answer, and respectively extracting word and sentence characteristics of the target test question and the target answer to obtain a test question characteristic set and an answer characteristic set; the similarity determining module is used for carrying out similarity measurement on the test question feature set and the answer feature set to obtain semantic similarity and character similarity; The intention recognition module is used for determining a test question intention vector of the test question feature set and an answer intention vector of the answer feature set by adopting a pre-trained target intention analysis model; The score determining module is used for determining the answer score of the target answer according to the semantic similarity, the character similarity, the test question intention vector and the answer intention vector and by combining a pre-trained target scoring model; the training process of the target intention analysis model comprises the following steps: Constructing an intention analysis model to be trained comprising a test question intention analysis network, an answer intention analysis network and an intention classifier, and initializing the number of intention categories of the intention classifier into a preset number of categories; Training the intention analysis model to be trained based on training word embedding data, and updating model parameters of the intention analysis model to be trained; Testing the updated intention analysis model to be trained based on the test word embedding data, and adjusting the number of intention categories of the intention classifier; the intention analysis model to be trained is trained and tested again, the adjustment trend of the number of the intention categories is determined, and when the adjustment trend meets the adjustment stability condition, the intention analysis model to be trained updated for the last time is determined to be a target intention analysis model; the method for testing the intention analysis model to be trained based on the test word embedding data comprises the steps of: acquiring test word embedding data of a preset test number group, and inputting each test word embedding data into an updated intention analysis model to be trained, wherein the test word embedding data comprises a test question word embedding sequence and a test answer word embedding sequence; Aiming at each group of test word embedded data, determining a test question intention vector of the test question word embedded sequence by utilizing the test question intention analysis network, and determining a test answer intention vector of the test answer word embedded sequence by utilizing the answer intention analysis network; determining a vector distance distribution condition according to Euclidean distance between each test question intention vector and the corresponding test answer intention vector; And when the vector distance distribution condition meets a preset distribution condition, subtracting one from the number of the intention categories of the intention classifier, otherwise, adding one to the number of the intention categories of the intention classifier.
9. 9. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the open question answer scoring method of any one of claims 1-7 when the program is executed by the processor.
10. 10. A storage medium containing computer executable instructions which, when executed by a computer processor, are for performing the open question answer scoring method of any one of claims 1-7.

Description

Open test question answer scoring method, device, equipment and storage medium Technical Field The embodiment of the invention relates to the technical field of intelligent education, in particular to an open test question answer scoring method, device, equipment and storage medium. Background In daily teaching and examination, there are a large number of open questions such as paragraph admissions, general article ideas, etc. At present, the open answer is modified by manual work of teaching experts, the efficiency is extremely low, and subjective factors of different scoring operators can cause the phenomenon of unfair scoring. The existing open test question scoring model can only carry out simple character recognition and logic judgment on answers, and cannot accurately judge whether the answers accord with questions or not, and the scoring accuracy is not high. Disclosure of Invention The invention provides a method, a device, equipment and a storage medium for scoring answers of open test questions, which are used for rapidly and accurately scoring the answers of the open test questions. In a first aspect, an embodiment of the present invention provides a method for scoring answers of open test questions, including: Acquiring a target test question and a target answer, and respectively extracting word and sentence characteristics of the target test question and the target answer to obtain a test question characteristic set and an answer characteristic set; Performing similarity measurement on the test question feature set and the answer feature set to obtain semantic similarity and character similarity; Determining a test question intention vector of the test question feature set and an answer intention vector of the answer feature set by adopting a pre-trained target intention analysis model; and determining the answer score of the target answer according to the semantic similarity, the character similarity, the test question intention vector and the answer intention vector by combining a pre-trained target scoring model. Optionally, the extracting the word and sentence feature of the target test question and the target answer to obtain a test question feature set and an answer feature set includes: sentence embedding operation is carried out on the target test questions and the target answers respectively, and test question sentence embedding vectors and answer sentence embedding vectors are obtained; performing word segmentation operation on the target test questions to obtain a test question Chinese word segmentation set and a test question English word segmentation set, and performing word embedding operation on the test question Chinese word segmentation in the test question Chinese word segmentation set to obtain a test question word embedding sequence; performing word segmentation operation on the target answers to obtain an answer Chinese word segmentation set and an answer English word segmentation set, and performing word embedding operation on answer Chinese word segmentation in the answer Chinese word segmentation set to obtain an answer word embedding sequence; and forming a test question feature set by the test question sentence embedding vector, the test question word embedding sequence, the test question Chinese word segmentation set and the test question English word segmentation set, and forming an answer feature set by the answer sentence embedding vector, the answer word embedding sequence, the answer Chinese word segmentation set and the answer English word segmentation set. Optionally, the performing similarity measurement on the test question feature set and the answer feature set to obtain semantic similarity and character similarity includes: Determining semantic similarity according to the test question embedding vector and the answer sentence embedding vector; Performing Chinese character similarity measurement on the test question Chinese word segmentation set and the answer Chinese word segmentation set to obtain Chinese character similarity, and performing English character similarity measurement on the test question English word segmentation set and the answer English word segmentation set to obtain English character similarity; and determining the character similarity according to the Chinese character similarity and the English character similarity. Optionally, the determining the test question intention vector of the test question feature set and the answer intention vector of the answer feature set by using a pre-trained target intention analysis model includes: Performing intention analysis on the test question word embedding sequence by using a target test question intention analysis network in a target intention analysis model to obtain a test question intention vector; And carrying out intention analysis on the answer word embedding sequence by using a target answer intention analysis network in a target intention analysis model to obtain an answer intention vec