CN113435179B

CN113435179B - Composition review method, device, equipment and storage medium

Info

Publication number: CN113435179B
Application number: CN202110705457.2A
Authority: CN
Inventors: 巩捷甫; 章继东; 呼啸; 宋巍; 王士进; 胡国平; 秦兵; 刘挺
Original assignee: Hebei Xunfei Institute Of Artificial Intelligence; Iflytek Beijing Co ltd; iFlytek Co Ltd
Current assignee: Hebei Xunfei Institute Of Artificial Intelligence; Iflytek Beijing Co ltd; iFlytek Co Ltd
Priority date: 2021-06-24
Filing date: 2021-06-24
Publication date: 2024-04-30
Anticipated expiration: 2041-06-24
Also published as: CN113435179A

Abstract

The application provides a composition review method, a device, equipment and a storage medium, wherein the method comprises the following steps: detecting whether a target composition to be reviewed is an abnormal composition; if not, correcting the target composition from the word level, the sentence level and the chapter level respectively to obtain correction results corresponding to the target composition at the word level, the sentence level and the chapter level respectively; determining a scoring grade of the target composition from the plurality of review dimensions to obtain a scoring grade of the target composition in the plurality of review dimensions; and generating comments of the target composition according to the grading of the target composition in a plurality of comment dimensions. The composition review method provided by the application can automatically review the composition to be reviewed, and avoids the problems caused by manual participation because manual participation is not needed, and the composition review method provided by the application can obtain the review results with rich contents, and the review results with rich contents can play a good guiding role for writers, so that the user experience is good.

Description

Composition review method, device, equipment and storage medium

Technical Field

The application relates to the technical field of text data processing, in particular to a method, a device, equipment and a storage medium for reviewing composition.

Background

Writing is an indispensable practical skill in daily life and study of people, and is also a necessary capability for students to master in important points in school education.

At present, most of the modes of reviewing the composition are manual, namely, the contents of the composition are reviewed by a reviewer, and the review of the composition is subjectively given according to the understanding of the contents.

However, in some cases, the number of the compositions to be reviewed is often large, the compositions require the reviewer to review the content one by one, which is time-consuming and labor-consuming, and the reviewer often presents a brief review due to the large workload of review, which has limited guidance effect on writers.

Disclosure of Invention

In view of the above, the application provides a method, a device and a storage medium for reviewing composition, which are used for solving the problems that the existing method for reviewing composition is time-consuming and labor-consuming, the given comment is very brief, and the guiding effect on writers is limited, and the technical scheme is as follows:

a composition review method comprising:

Detecting whether a target composition to be reviewed is an abnormal composition;

If not, correcting the target composition from the word level, the sentence level and the chapter level respectively to obtain correction results corresponding to the target composition at the word level, the sentence level and the chapter level respectively;

determining a scoring grade of the target composition from a plurality of review dimensions to obtain a scoring grade of the target composition over the plurality of review dimensions;

and generating comments of the target composition according to the grading of the target composition in the plurality of comment dimensions.

Optionally, the plurality of review dimensions are a plurality of review dimensions corresponding to the learning segment to which the target composition belongs in the set plurality of dimensions;

Determining a scoring grade of the target composition from a plurality of review dimensions to obtain a scoring grade of the target composition over the plurality of review dimensions, comprising:

And determining the grading of the target composition from a plurality of grading dimensions corresponding to the subject paragraph to which the target composition belongs, so as to obtain the grading of the target composition on the plurality of grading dimensions corresponding to the subject paragraph to which the target composition belongs.

Optionally, the generating the comment of the target composition according to the grading of the target composition in the multiple comment dimensions includes:

determining comments corresponding to the scoring grades of the target composition on each scoring dimension based on the pre-established scoring dimension and the corresponding relation between the scoring grades and the comments;

and generating comments of the target composition according to comments corresponding to the grading of the target composition on each comment dimension.

Optionally, the composition review method further includes:

determining a scoring grade of the target composition;

The generating the comment of the target composition according to the grading of the target composition in the plurality of comment dimensions comprises the following steps:

And generating the comments of the target composition according to the grading of the target composition in the plurality of comment dimensions and the grading of the whole target composition.

Optionally, the determining the grading of the target composition includes:

Predicting the overall score of the target composition based on a pre-established score prediction model, and taking the overall score as a first score of the target composition;

And/or, based on a pre-established differential prediction model, predicting the differential between the target composition and the normative texts in the normative text set, and based on the differential between the target composition and the normative texts in the normative text set and the scoring of the normative texts in the normative text set, predicting the overall scoring of the target composition as a second scoring of the target composition, wherein the normative texts in the normative text set and the target composition belong to the same subject;

and determining the grading of the whole target composition based on the first grading of the target composition and/or the second grading of the target composition.

Optionally, the detecting whether the target composition to be reviewed is an abnormal composition includes:

if the target composition meets at least one of the following conditions, determining that the target composition is an abnormal composition:

Condition one: the similarity between the target composition and a text in a pre-constructed celebrity famous chapter material library is larger than a preset similarity threshold;

condition II: the ratio of sentences with the confusion degree larger than a preset confusion degree threshold value in the target composition is larger than a preset ratio threshold value;

and (3) a third condition: sensitive information appears in the target composition.

Optionally, the modifying the target composition from the word level includes:

One or more of the following treatments are performed on the target composition: language accuracy analysis, network term retrieval and error punctuation recognition, and generating a correction result corresponding to the target composition at the word level according to the processing result; wherein the language accuracy analysis comprises an individual word error correction, and/or a grammar error detection, and/or idiom class error detection, and/or an ancient poetry error detection;

Correcting the target composition from sentence level, comprising:

One or more of the following treatments are performed on the target composition: graceful sentence recognition, advanced vocabulary statistics, sentence repair method recognition and descriptive sentence recognition, and generating a modification result corresponding to the target composition at a sentence level according to a processing result;

Correcting the target composition from the chapter level, comprising:

And performing chapter structure recognition and/or theme recognition on the target composition, and generating a modification result corresponding to the target composition at a chapter level according to the recognition result.

Optionally, performing error punctuation recognition on the target composition includes:

For each paragraph in the target document:

Removing punctuation marks in the paragraphs, and taking the paragraphs after the punctuation marks are removed as target texts;

predicting a label category corresponding to each word in the target text based on a pre-established punctuation prediction model; the label category corresponding to one word is used for indicating whether punctuation marks exist behind the word and why the punctuation marks exist;

determining punctuation marks predicted for the target text according to the label category corresponding to each word in the target text;

From the punctuation in the paragraph and the punctuation predicted for the target text, determining that the erroneous punctuation is used in the punctuation of the paragraph.

Optionally, the process of performing the word-writing error correction on the target composition includes:

For each sentence in the target work:

detecting wrongly written words in the sentence, and obtaining a candidate word set corresponding to the wrongly written words;

masking the error word in the sentence based on a pre-established mask language model, and predicting the probability that the word at the masking position in the masked sentence is each candidate word in the candidate word set corresponding to the wrongly written word;

based on the predicted probability, determining the correct word corresponding to the wrongly written word from the candidate word set corresponding to the wrongly written word;

correcting the wrongly written word in the sentence into the correct word corresponding to the wrongly written word.

Optionally, the detecting the wrongly written word in the sentence includes:

For each word in the sentence:

The word in the sentence is replaced by each word in the confusion word set corresponding to the word, and each sentence after replacement forms a candidate sentence subset;

calculating the confusion degree of the sentence and the confusion degree of the sentences in the candidate sentence set based on a pre-established statistical language model;

Based on the confusion of the sentence and the confusion of the sentences in the candidate sentence set, it is determined whether the word is a wrongly written word.

Optionally, the process of performing syntax error detection on the target composition includes:

For each sentence in the target work:

The method comprises the steps of obtaining syntactic dependency characteristics of a sentence, and characteristics of each word, word segmentation characteristics of each word, mutual information characteristics of each word and part-of-speech characteristics of each word in the sentence;

Determining a context vector for each word in the sentence based on the obtained features;

Determining whether the sentence has a grammar error according to the context vector of each word in the sentence, and determining the specific grammar error when the grammar error exists.

Optionally, the plurality of comment dimensions are a plurality of comment dimensions of one or more of the following four aspects: content, expression, structure, development, wherein:

the review dimensions for the content aspect include one or more of the following review dimensions: ideological health, emotion type, compliance with questions and content enrichment;

The review dimensions of the expression aspect include one or more of the following review dimensions: basic expression, line language specification, language fluency and text conforming;

the review dimension of the structural aspect includes: the structure is strict;

the review dimension of the development aspect includes: and (5) culture and collection.

Optionally, determining the scoring grade of the target composition in the review dimension of the thought health includes:

Judging whether each sentence in the target text contains a low-custom language or not;

Determining the grading of the target composition in the evaluation dimension of the thought health according to the judging result of each sentence in the target composition;

Determining a scoring grade of the target composition in the review dimension of the emotion type comprises:

Identifying the emotion type expressed by the target composition;

Determining grading of the target composition in the evaluation dimension of the emotion type according to the emotion type identification result of the target composition;

Determining a scoring grade of the target composition in the review dimension of the fitting topic comprises:

Acquiring a representation vector of each word in the title of the target composition and a representation vector of each word in the text of the target composition, and determining a target vector corresponding to each word in the title of the target composition based on the acquired vectors, wherein the target vector corresponding to one word can represent the related condition of the word and each word in the text of the target composition;

Determining the matching degree of the title and the text of the target composition according to the target vector corresponding to each word in the title of the target composition, and determining the grading of the target composition in the evaluation dimension conforming to the title according to the matching degree of the title and the text of the target composition;

Determining a scoring grade of the target composition in the review dimension of the content enrichment, comprising:

Determining a chapter representation vector of the target composition based on the target composition, basic information of the target composition and correction results corresponding to the target composition on sentence level and chapter level respectively;

and determining the grading of the target composition in the review dimension of the content enrichment by classifying the chapter representation vectors.

Optionally, determining a scoring grade of the target composition in the review dimension of the base expression includes:

based on the word making list and/or the common idiom library respectively corresponding to different school segments, carrying out word making and/or common idiom recognition on the target composition so as to obtain word use conditions of the target composition on different school segments;

performing sentence pattern recognition on the target composition based on the appointed sentence pattern to obtain the sentence pattern service condition of the target composition;

Determining scoring classification of the target composition in a review dimension of basic expression based on word use conditions of the target composition in different school segments and sentence pattern use conditions of the target composition;

Determining a scoring grade of the target composition in the review dimension of the line document specification comprises:

detecting whether the title, and/or paragraph, and/or network term, and/or punctuation mark, and/or genre format of the target composition meets the specification;

determining the grading of the target composition in the review dimension of the line text specification according to the detection result of the target composition in the line text specification;

determining a scoring grade of the target composition in the language fluency review dimension comprises:

extracting feature vectors from the basic information of the target composition, the statistical information of the word error correction and grammar error detection results and the collocation and combination of the occurrence of the target composition;

Determining a scoring grade of the target composition in the language fluency scoring dimension by classifying the extracted feature vectors;

determining a scoring grade of the target composition in the document meeting the review dimension comprises:

identifying the genre of the target composition based on a pre-established genre identification model, wherein the genre identification model is obtained by training a training composition marked with the genre;

And determining the grading of the target composition in the dimension of the document meeting the review according to whether the genre of the target composition is consistent with the appointed genre.

Optionally, the identifying the genre of the target composition based on the pre-established genre identification model includes:

Encoding each word in the target text based on the genre recognition model to obtain an encoding vector of each word in the target text;

Performing attention calculation on the coding vector of each word in the target text based on the genre recognition model to obtain an attention vector of each word in the target text, and obtaining a representation vector of each sentence in the target text based on the attention vector of each word in the target text;

Encoding the representation vector of each sentence in the target text based on the genre identification model to obtain an encoded vector of each sentence in the target text;

Performing attention calculation on the coding vector of each sentence in the target composition based on the genre identification model to obtain an attention vector of each sentence in the target composition, and determining a chapter representation vector of the target composition based on the attention vector of each sentence in the target composition;

classifying the chapter representative vector of the target composition based on the genre identification model to obtain the genre of the target composition.

Optionally, the correction result of the target composition at the chapter level includes the correction result of the target composition on the chapter structure;

determining a scoring grade of the target composition in a review dimension with strict structure comprises:

and determining the grading of the target composition in the evaluation dimension with strict structure according to the correction result of the target composition in the chapter structure.

Optionally, determining the scoring grade of the target composition in the review dimension of the literature comprises:

Determining a chapter representation vector of the target composition based on the target composition, the basic information of the target text and the correction result of the target composition at the sentence level;

and determining the grading of the target composition in the review dimension of the literature by classifying the chapter representation vectors.

A composition review device, comprising: the system comprises a detection module, a correction module, a grading determination module and a comment generation module;

The detection module is used for detecting whether the target composition to be reviewed is an abnormal composition;

The correcting module is used for correcting the target composition from the word level, the sentence level and the chapter level respectively when the target composition is not an abnormal composition, so as to obtain correction results of the target composition corresponding to the word level, the sentence level and the chapter level respectively;

The scoring grade determining module is used for determining the scoring grade of the target composition from a plurality of scoring dimensions so as to obtain the scoring grade of the target composition in the plurality of scoring dimensions;

and the comment generation module is used for generating comments of the target composition according to the grading of the target composition in the plurality of comment dimensions.

A composition review device comprising: a memory and a processor;

the memory is used for storing programs;

The processor is configured to execute the program to implement each step of the composition review method described in any one of the above.

A readable storage medium having stored thereon a computer program which, when executed by a processor, performs the steps of the composition review method of any of the preceding claims.

According to the method, the device, the equipment and the storage medium for reviewing the composition, whether the target composition to be reviewed is an abnormal composition is detected, if the target composition is not the abnormal composition, the target composition is modified from the word level, the sentence level and the chapter level respectively, so that the modification results of the target composition corresponding to the word level, the sentence level and the chapter level are obtained. The method for reviewing the composition provided by the application can automatically review the composition to be reviewed, and avoids the problems caused by manual participation because manual participation is not needed, and through the method for reviewing the composition provided by the application, not only can the corresponding correction results of the target composition on the word level, the sentence level and the chapter level respectively be obtained, but also the comments of the target composition on a plurality of review dimensions can be obtained, namely, the review results are rich, the rich review results can play a good guiding role on writers, and the user experience is good.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are required to be used in the embodiments or the description of the prior art will be briefly described below, and it is obvious that the drawings in the following description are only embodiments of the present invention, and that other drawings can be obtained according to the provided drawings without inventive effort for a person skilled in the art.

FIG. 1 is a schematic flow chart of a method for reviewing composition according to an embodiment of the present application;

FIG. 2 is a schematic diagram of an example of shallow features to be extracted for treatises and other body-building compositions provided by embodiments of the present application;

FIG. 3 is a schematic diagram of an example of word level, sentence level, and chapter level modification content and dimension of generated comments provided by an embodiment of the present application;

Fig. 4 is a topological structure diagram of a genre identification model according to an embodiment of the present application;

FIG. 5 is a schematic structural diagram of a GRU model according to an embodiment of the present application;

FIG. 6 is a schematic flow chart of generating comments of a target composition according to scoring classification of the target composition in multiple comment dimensions provided by an embodiment of the present application;

Fig. 7 is a schematic structural diagram of a composition review device according to an embodiment of the present application;

fig. 8 is a schematic structural diagram of a composition review device according to an embodiment of the present application.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

In view of the problems of the existing manual review mode, the inventor tries to propose an automatic review method, and for this reason, the inventor performs research, and the initial thought is as follows: and (3) adopting a scoring model-based scoring method, namely training a scoring model by using a scoring-marked training composition in advance, and scoring the scoring-to-be-scored composition by using the scoring model obtained by training.

The inventor finds that the scoring model-based scoring method only gives a score for the composition to be scored, and the writer only knows about the level of the composition by the score, but does not know about the problem of the composition, so that the scoring result obtained by the scoring model-based scoring method has limited guiding effect on writers.

In view of the problems of the scoring-model-based review method, the inventor further researches, and finally provides a composition review method with good effect through continuous research, wherein the composition review method is applicable to any application scene requiring the review of the composition, the automatic review of the composition can be realized through the composition review method, and the review result with strong guidance to the writer can be obtained through the composition review method, and the basic concept of the composition review method is as follows: and correcting the target composition from the word level, the sentence level and the chapter level to obtain correction results corresponding to the target composition at the word level, the sentence level and the chapter level respectively, determining the grading of the target composition from a plurality of grading dimensions, and generating the comment of the target composition according to the grading of the target composition at the plurality of grading dimensions.

The composition review method provided by the application can be applied to the electronic equipment with the data processing capability, and the electronic equipment can be a server at a network side or a terminal used at a user side, such as a PC, a notebook, a PAD, a smart phone and the like. The method for reviewing composition provided by the present application will be described by way of the following examples.

First embodiment

Referring to fig. 1, a flow chart illustrating a method for reviewing composition provided in an embodiment of the present application may include:

Step S101: and detecting whether the target composition to be reviewed is an abnormal composition, if not, executing the step S102a, and if so, executing the step S102b.

Specifically, the process of detecting whether the target composition to be reviewed is an abnormal composition may include: if the target composition meets at least one of the following conditions, determining that the target composition is an abnormal composition:

The celebrity famous chapter material library in the condition one can be constructed by integrating and de-duplicating common texts such as works of common celebrity famous chapter and part-compiled textbooks, and the similarity between a target composition and a text in the celebrity famous chapter material library is larger than a preset similarity threshold, so that most of contents of the target composition are the contents of the text in the celebrity famous chapter library, and in the case, the target composition is judged to be an abnormal composition.

The degree of confusion of the sentences in the second condition can be determined based on a pre-established sentence-level language model, the degree of confusion of the sentences is characterized by the degree of "meaning" of the sentences, and it is required to be explained that the ratio of the sentences with the degree of confusion larger than a preset degree of confusion threshold in the target composition is larger than a preset ratio threshold, and the ratio of the sentences with the degree of confusion not meaning "in the target composition is higher, and in this case, the target composition is judged to be an abnormal composition.

For the third condition, the method for detecting whether the sensitive information appears in the target composition may be: searching whether the sensitive words in the pre-collected and arranged sensitive word list appear in the target composition, and if the sensitive words in the sensitive word list appear in the target composition, determining that the sensitive information appears in the target composition, and in this case, judging that the target composition is an abnormal composition.

Step S102a: and correcting the target composition from the word level, the sentence level and the chapter level respectively to obtain correction results corresponding to the target composition at the word level, the sentence level and the chapter level respectively.

S102b: and generating abnormal composition comments.

Specifically, if the target composition meets the first condition, a celebrity famous chapter type comment may be generated, optionally, the celebrity famous chapter type comment may include information (such as a name) of a text whose similarity with the target composition is greater than a preset similarity threshold in a celebrity famous chapter material library, if the target composition meets the second condition, a comment with the target composition not meeting the meaning/chapter confusion may be generated, optionally, a sentence with the confusion greater than the preset confusion threshold (i.e., a sentence with the confusion degree not meeting the meaning) may be marked, if the target composition meets the third condition, a sensitive information type comment may be generated, optionally, a sensitive word in the target composition may be included in the sensitive information type comment, and optionally, a sensitive word in the target composition may also be marked.

Step S103: and determining the grading of the target composition from the plurality of review dimensions to obtain the grading of the target composition in the plurality of review dimensions.

In this step, "a plurality of review dimensions" is a set part or all of the plurality of review dimensions.

Alternatively, the plurality of comment dimensions in step S103 may be a plurality of comment dimensions in one or more of the following four aspects: content, expression, structure, development, wherein:

The review dimensions for the content aspect may include one or more of the following review dimensions: ideological health, emotion type, compliance with questions and content enrichment; the review dimensions of the expression aspects may include one or more of the following review dimensions: basic expression, line language specification, language fluency and text conforming; the review dimension of the structural aspect may include: the structure is strict; review dimensions of the development aspect may include: and (5) culture and collection.

It should be noted that, the plurality of comment dimensions in step S103 may be the 10 comment dimensions, or may be some of the 10 comment dimensions.

In the education field, considering that the review dimensions focused by different school segments are different, in order to obtain more targeted review results and better guide writers, the application provides the method for determining the scoring grade of the target composition from a plurality of review dimensions corresponding to the school segments to which the target composition belongs, so as to obtain the scoring grade of the target composition in a plurality of review dimensions corresponding to the school segments to which the target composition belongs. The following table shows review dimensions for each of the academic interests:

Table 1 review dimension for each school focus

As shown in the table, in the primary school stage 1-2, the ideological health and emotion types in the content aspect are emphasized, the basic expression application and the line text specification in the expression aspect are emphasized, in the primary school stage 3-4, the basic requirements of the primary school stage 1-2 are met, the two review dimensions of content enrichment and meeting the subjects are added in the content aspect, and in the primary school stage 5-6, the two review dimensions of language smoothness and the line text are added. In the junior middle school, the application of basic expression is not focused any more, and two review dimensions of the consistence of the cultural relics and the strict structure are increased.

Step S104: and generating comments of the target composition according to the grading of the target composition in the plurality of comment dimensions.

Optionally, the method for reviewing composition provided in this embodiment may further include: after determining that the target composition is not an abnormal composition, determining a scoring grade of the target composition as a whole before generating a comment of the target composition according to the scoring grade of the target composition in a plurality of comment dimensions. On the basis, the comments of the target composition can be generated according to the scoring grades of the target composition in a plurality of comment dimensions and the scoring grades of the whole target composition.

The implementation modes of grading the score of the target composition whole are various:

in one possible implementation: the overall score of the target composition can be predicted based on a pre-established score prediction model, and the score rank of the overall target composition can be determined according to the overall score of the target composition. The scoring prediction model is obtained by training a training composition marked with the integral score.

In another possible implementation: the score of the whole target composition can be predicted based on a pre-established score prediction model, the score of the target composition and the score of the template in the template set (the score may be positive or negative), and the whole score of the target composition is predicted based on the score of the target composition and the score of the template in the template set, and the whole score of the target composition is determined according to the whole score of the target composition. In the implementation mode, the normative texts in the normative text set and the target composition belong to the same theme, and the differential prediction model is obtained by training the training composition and the normative texts which belong to the same theme as the training composition and are marked with integral scores.

Wherein, based on the difference between the target composition and the normative in the normative set and the scoring of the normative in the normative set, the process of predicting the overall scoring of the target composition comprises: determining the score of the target composition as the score of the target composition on each template in the template set according to the score of the template and the difference between the target composition and the template; and calculating the average value of the scores of the target composition on each scope in the scope set, and calculating the average value as the overall score of the target composition.

In yet another possible implementation: the overall score of the target composition can be predicted based on a pre-established score prediction model to serve as a first score of the target composition; predicting the difference between the target composition and the template in the template set based on a pre-established difference prediction model, and predicting the overall score of the target composition as a second score of the target composition based on the difference between the target composition and the template in the template set and the score of the template in the template set; a scoring grade for the target composition as a whole is determined based on the first score for the target composition and the second score for the target composition.

It should be noted that, whether the scoring prediction model or the difference prediction model is the prediction thinking, features of the target composition in two aspects are obtained, wherein the features include shallow features, word use features, sentence level evaluation features, chapter structure analysis features, deep semantic features, and vector representations of the target composition are obtained according to the features of the target composition in two aspects, and then overall scoring or difference prediction is performed according to the vector representations of the target composition. In addition, it should be noted that, considering that the discussion paper has its uniqueness compared with other genres, the uniqueness of the discussion paper is mainly represented by that it has a certain requirement on chapter structure, and in view of this, when scoring or predicting the difference between the discussion paper, chapter structure features need to be acquired, while for other genres, such as the description, there is no harder requirement on chapter structure, but there is a certain requirement on expression, therefore, there is a need to acquire features related to expression, and of course, both have some common features. Fig. 2 shows a schematic diagram of an example of shallow features to be extracted for composition of treatises and other genres.

Given the differences in the characteristics used for scoring or scoring the composition of different genres, the genre of the target composition may be determined first, before scoring or scoring the target composition. In addition, in view of different study contents of different study sections, a scoring prediction model and/or a grading prediction model of different study sections can be established, and when the scoring prediction or the grading prediction is carried out on the target study, the scoring prediction model and/or the grading prediction model of the study section to which the target study belongs are adopted for prediction.

According to the composition review method provided by the embodiment of the application, whether the target composition to be reviewed is an abnormal composition is detected, if the target composition is not the abnormal composition, the target composition is modified from the word level, the sentence level and the chapter level respectively, so that modification results of the target composition corresponding to the word level, the sentence level and the chapter level are obtained. The method for reviewing the composition provided by the embodiment of the application can automatically review the composition to be reviewed, avoids the problems caused by manual participation because manual participation is not needed, can obtain the corresponding correction results of the target composition on the word level, the sentence level and the chapter level respectively, and can obtain the comments of the target composition on a plurality of review dimensions, namely, the review results are rich, the rich review results can play a good guiding role for writers, and the user experience is good.

Second embodiment

The present embodiment is mainly directed to "step S102 a" in the above embodiment: and correcting the target composition from the word level, the sentence level and the chapter level respectively to obtain the corresponding correction results of the target composition at the word level, the sentence level and the chapter level respectively for introduction.

First, a process of modifying a target composition from a word level will be described.

The process of altering the target composition from word level includes: one or more (preferably more) of the following treatments are performed on the target composition: language accuracy analysis, network term retrieval and error punctuation recognition, and generating a correction result corresponding to the target composition at the word level according to the processing result.

The language accuracy analysis comprises the steps of word correction, grammar error detection, idiom retrieval and ancient poetry error detection. Preferably, the language accuracy analysis includes both individual word error correction, grammar error detection, idiom class error detection, and ancient poetry error detection.

The process for correcting the written characters of the target composition comprises the following steps: for each sentence in the target work, performing:

step a1, detecting wrongly written words in the sentence, and obtaining a candidate word set corresponding to the wrongly written words.

Specifically, the process of detecting wrongly written words in the sentence includes: for each word in the sentence, respectively replacing the word in the sentence with each word in the confusion word set corresponding to the word, forming a candidate sentence subset by the replaced sentences, and calculating the confusion degree of the sentence and the confusion degree of the sentences in the candidate sentence set based on a pre-established statistical language model; based on the confusion of the sentence and the confusion of the sentences in the candidate sentence set, it is determined whether the word is a wrongly written word. Wherein, the confusion word set corresponding to a word can contain the shape near word, the sound near word and homonym of the word.

Alternatively, the statistical language model in this embodiment may be a language model based on a deep neural network, for example, an ELMO-based language model, a BERT-based language model, or the like, and these models may learn syntax semantic information of a deeper level of language through a large number of parameters.

Since the manner of determining whether each word in a sentence is a wrongly written word is the same, the present embodiment describes a process of determining whether a "store" word is a wrongly written word by taking a "store" word in the sentence "i am looking at a store" as an example:

assuming that the confusion word set corresponding to "store" is { electricity, point, occupied }, firstly, the "store" in "I'm looking at" is replaced by "electricity", "point", "occupied", respectively, to obtain a candidate sentence subset { I'm looking at TV, I'm looking at "point looking at" and I'm looking at "occupied }, after obtaining the candidate sentence subset, calculating the confusion degree of" I'm looking at "in" store "and the confusion degree of" I'm looking at "in" television "," I'm looking at "and" I'm occupied at "in candidate sentence set based on a pre-established statistical language model, calculating, the confusion degree of 'I'm looking at a store 'is 55, the confusion degree of' I'm looking at a television' is 23, the confusion degree of 'I'm looking at a point of view 'is 61, the confusion degree of' I'm looking at a occupation' is 42, and the confusion degree of 'I'm looking at a television 'and' I'm looking at a occupation' is higher than that of 'I'm looking at a store ', so that' store 'in' I'm looking at a store' can be determined to be wrongly written, candidate word sets corresponding to wrongly written words 'store' can be formed by 'electricity' and 'occupation', and correct words corresponding to 'store' can be determined from 'electricity' and 'occupation' subsequently.

And a2, masking the error word in the sentence based on a pre-established mask language model, and predicting the probability that the word at the masking position in the masked sentence is each candidate word in the candidate word set corresponding to the wrongly written word.

For the example "I look at store" above, the probability of masking the word at the location of the mask as "electric" and the probability of being "occupied" are predicted based on a pre-established masking language model to mask the "store".

Alternatively, the mask language model in this embodiment may be a Bert-based mask language model.

And a step a3 of determining correct words corresponding to the wrongly written words from the candidate word sets corresponding to the wrongly written words based on the predicted probability.

For the above example "i look at store", assuming that the probability of predicting the word at the mask position as "electric" is 0.99 via step a2, and assuming that the probability of the word at the mask position as "occupied" is 0.13, it is determined that the correct word corresponding to the misprinted word "store" is "electric".

And a4, correcting the wrongly written word in the sentence into a correct word corresponding to the wrongly written word.

For the example "i'm looking at store", after determining that the correct word corresponding to the wrong written word "store" is "electricity", the "i'm looking at store" is corrected to "i'm looking at television".

The grammar error detection process for the target composition comprises the following steps: for each sentence in the target work, performing:

Step b1, obtaining the syntactic dependency characteristics of the sentence, and the characteristics of each word, the word segmentation characteristics of each word, the mutual information characteristics of each word and the part-of-speech characteristics of each word in the sentence.

The word segmentation feature of a word is used for representing the word and which word is divided into a word, the mutual information feature of the word is used for representing the co-occurrence condition of the word and the front and rear contents of the word, and the part-of-speech feature of the word is used for representing the part-of-speech of the word where the word is located.

And b2, determining the context vector of each word in the sentence according to the acquired characteristics.

Alternatively, the Bert model and long and short term memory networks LSTM (preferably bidirectional LSTM) may be used to obtain a context vector for each word in the sentence.

The method comprises the steps of training a Bert model, namely a pretrained neural network model, pretraining the Bert model on a large-scale corpus by using a Transformer, so that the model learns general sentence representation, and because the pretrained Bert model learns general sentence representation, the parameters of the Bert model need to be finely adjusted in the training process of a grammar error detection task, so that the model learns context representation related to the grammar error detection task, LSTM is a cyclic neural network widely applied to sequence tasks, the sequence can be better modeled, long-distance dependency information is captured, and the bidirectional LSTM network represents bidirectional context information through splicing vectors of forward LSTM vector representation and reverse LSTM vector representation, so that the representation capability of the LSTM model is further enhanced.

And b3, determining whether the sentence has grammar errors or not according to the context vector of each word in the sentence, and determining the specific grammar errors when the grammar errors exist.

Alternatively, the context vector for each word in the sentence may be input into the CRF model for sequence labeling for syntax error detection. The CRF model can calculate the sentence level normalization probability, so that a globally optimal labeling sequence is selected, and for a sequence labeling task, the output labeling labels also have a certain dependency relationship, and the CRF algorithm can model the dependency relationship among the labeling labels by learning the transition probability among the labeling labels, so that a more reasonable labeling result is ensured to be selected when the model is decoded.

For idiom-class error detection, a fuzzy matching algorithm can be adopted to carry out fuzzy matching on the target composition and idioms in a pre-constructed idiom library, and the idioms with wrong writing in the target composition can be determined through a fuzzy matching result. Similar to idiotype error detection, fuzzy matching algorithm can be adopted to perform fuzzy matching on the target composition and the ancient poems in the pre-constructed ancient poems library, and the ancient poems with wrong writing in the target composition can be determined through the fuzzy matching result.

For network phrase retrieval, network non-canonical phrases occurring in the target composition may be retrieved based on a pre-built network non-canonical phrase resource table.

The punctuation errors generally include two kinds of punctuation errors, namely, punctuation errors of a format class, such as continuous use of punctuation, mismatching of pairs of punctuation, missing of whole punctuation, common use of a stop with a conjunction, misuse of a colon before a conjunction, and the like, and punctuation errors of a semantic class.

For punctuation errors of format classes, a rule-based error detection mode can be acquired, specifically, a corresponding regular expression is written for each punctuation error in advance, so that a rule base for punctuation error detection is constructed, when punctuation error detection is carried out on a target composition, punctuation marks in paragraphs are matched with rules in the rule base by taking the paragraphs as units, and if the rules are matched, the corresponding punctuation marks are determined to be wrong punctuation marks.

For punctuation errors of semantic classes, mainly solving the use errors of periods and commas related to semantics in texts, wherein the use errors of periods and commas related to semantics cannot be solved by rules. For punctuation errors of semantic classes, the present embodiment provides the following recognition means:

For each paragraph in the target work, perform:

And c1, removing punctuation marks in the paragraphs, and taking the paragraphs after the punctuation marks are removed as target texts.

And c2, predicting the label category corresponding to each word in the target text based on a pre-established punctuation prediction model.

The label category corresponding to a word is used for indicating whether punctuation marks exist behind the word and why the punctuation marks exist.

Specifically, the punctuation prediction model may include an encoding module (e.g., macBERT), a mapping module, and a classification module, where:

The encoding module encodes each word c _i in the target text and outputs a representation vector h _i of each word:

h_i＝MacBERT(c_i) (1)

The representation vector h _i of each word c _i in the target text is input to a mapping module, the mapping module maps the representation vector h _i of each word to the space of the label class, the mapping result is output, the mapping result is input to a classification module, and the classification module calculates the probability distribution p _i of each word in the target text on the space of the label class based on the softmax function and the input mapping result:

p_i＝Softmax(Linear(h_i)) (2)

After obtaining the probability distribution p _i of each word in the target text on the label class space, the label class corresponding to each word in the target text, namely the label class predicted for each word in the target text, can be determined according to the probability distribution p _i of each word in the target text on the label class space.

And c3, determining punctuation marks predicted for the target text according to the label category corresponding to each word in the target text.

And c4, determining that the wrong punctuation mark is used in the punctuation mark of the paragraph according to the punctuation mark in the paragraph and the punctuation mark predicted for the target text.

By comparing the punctuation in the paragraph with the punctuation predicted for the target text, it can be determined that the wrong punctuation was used in the paragraph.

The result of the target composition on the word level can indicate that the target composition has problems on the word level, such as wrongly written characters, wrong grammar, use of network nonstandard expression, wrongly written idioms, wrongly written ancient poems, wrongly used punctuations, and the like.

Next, a process of correcting the target composition from the sentence level will be described.

The process of correcting the target composition from sentence level comprises the following steps: and performing graceful sentence recognition, and/or repair method recognition and/or descriptive sentence recognition on the target composition, and generating a correction result corresponding to the target composition at a sentence level according to the recognition result. Preferably, graceful sentence recognition, convoy skill recognition and descriptive sentence recognition can be performed on the target composition at the same time, and further, correction results corresponding to the target composition at the sentence level are generated according to the three recognition results.

The process for carrying out graceful sentence recognition on the target composition comprises the following steps: each sentence in the target composition is input into a pre-established graceful sentence discrimination model to obtain a discrimination result of each sentence in the target composition. It should be noted that the graceful sentence discrimination model is a classification model that classifies an input sentence into graceful sentences/non-graceful sentences, specifically, each sentence in the target work is input into a classification model of graceful sentences/non-graceful sentences, which extracts feature vectors for the input sentence first, and then determines whether the input sentence is a graceful sentence or a non-graceful sentence based on the extracted feature vectors.

The method of the convoy has a lot, and is commonly known as ranking, quoting, metaphing and personification, and the processes of ranking convoy recognition, quoting convoy recognition, metaphing convoy recognition and personification convoy recognition are respectively described below.

The process for identifying the ranking and the congratulation of the target composition comprises the following steps: firstly, constructing an inverted storage structure for a target composition according to paragraphs, respectively extracting candidate ranking sentences, filtering the candidate ranking sentences, judging the candidate ranking sentences and the like on the basis, and then using mechanisms such as candidate ranking sentence segmentation, candidate ranking sentence complementation and the like to identify complete ranking sentences.

The process of citation and congratulation identification of the target composition comprises the following steps: for each sentence in the target work, searching in a pre-constructed reference resource library, if the sentence is searched in the reference resource library, determining the sentence as a reference sentence, and if the sentence is not searched in the reference resource library, determining the sentence as a reference sentence. The quoted resource library comprises common ancient poetry, cultural relics, modern celebrity, hangzhou, poem, classical sentences in common songs and the like.

The process of metaphorically identifying the target composition includes: for each sentence in the target composition, the metaphors (such as "like", "look like", etc.) are first screened from the sentence, then a first vector capable of characterizing the sentence component is obtained based on the metaphor recognition model and the screened metaphors, and a second vector characterizing the sentence (especially the front part of the metaphor and the rear part of the metaphor) is determined based on the metaphor recognition model and the obtained vector.

The process for anthropomorphic fix-up identification of the target composition comprises the following steps: and inputting each sentence into a pre-established anthropomorphic sentence recognition model aiming at each sentence in the target work, and obtaining a recognition result which is output by the anthropomorphic sentence recognition model and is used for indicating whether the sentence is an anthropomorphic sentence. After a sentence to be identified is input into the anthropomorphic sentence identification model, the anthropomorphic sentence identification model firstly performs feature extraction, and then classifies the extracted features, so that a classification result indicating whether the input sentence is an anthropomorphic sentence is obtained.

The process of identifying the descriptive sentence of the target composition comprises the following steps: for each sentence in the target work, the description type of the sentence and whether the sentence is in the text are identified based on a pre-established description sentence identification model, that is, whether each sentence in the target work is a description sentence in the text or not can be identified based on the description sentence identification model.

The descriptive sentence recognition model is obtained by training a training text with two labels, wherein one label is used for indicating the descriptive type of the training text, and the other label is used for indicating whether the training text has a text. Alternatively, the description type of one sentence may be one of the following description types: language, action, appearance, mind, scene. Alternatively, the descriptive sentence recognition model may include a semantic representation vector determination module (such as a bidirectional LSTM), a text semantic vector determination module for determining a semantic representation vector of the input sentence, a text recognition module for determining whether the input sentence is text mined based on the semantic representation vector of the input sentence, and a descriptive type recognition module for determining a descriptive type of the input sentence based on the semantic representation vector of the input sentence.

The correction result of the target composition at the sentence level can indicate whether the target composition uses a method of making a repair, and/or a graceful sentence, and/or a descriptive sentence, and when the method of making a repair is used, can also indicate which method of making a repair is used and which part of the target composition uses the method of making a repair, when the graceful sentence is used, can indicate which part of the target composition is a graceful sentence, and when the descriptive sentence is used, can indicate which type of text-in/text-out descriptive sentence is used, and can also indicate which part of the target composition is the type of text-in/text-out descriptive sentence.

Finally, the process of modifying the target text from the chapter level is introduced.

The process of modifying the target text from the chapter level includes: and performing chapter structure analysis and/or theme analysis on the target composition, and generating a modification result corresponding to the target composition at a chapter level according to the analysis result. Preferably, chapter structure analysis and theme analysis can be performed on the target composition at the same time, and then a modification result corresponding to the target composition at a chapter level is generated according to the two analysis results.

For the chapter structure recognition, the chapter structure of the target composition may be analyzed based on a set analysis rule, for example, whether the target composition is beginning to be the first part bring out the theme, whether the target composition is the first part and the last part, whether the structure is clear, whether the viewpoint is clear, and the like. For topic analysis, topic identification can be performed on the target composition based on a pre-established topic identification model, and on the basis of the topic identification, topic departure detection can be further performed, namely, whether the topic identified for the target composition deviates from a specified topic or not is detected.

The topic identification model is essentially a classification model, which determines the topic to which the target composition belongs from topics contained in a pre-constructed topic system.

The middle part of fig. 3 shows a schematic diagram of correcting the target composition from the word level, sentence level and chapter level, through the correcting process described above, the correcting results corresponding to the target composition at the word level, sentence level and chapter level can be obtained, and the correcting results corresponding to the target composition at the word level, sentence level and chapter level can enable the writer to learn information about wrongly written words, unhappy choice of words, irregular expressions, punctuation errors, graceful sentences used, used congratulation techniques, used descriptive expression modes, chapter structures, composition topics and the like in the composition.

Third embodiment

The present embodiment mainly corresponds to "step S103" in the above embodiment: and determining the grading of the target composition from the plurality of review dimensions to obtain the grading of the target composition in the plurality of review dimensions for introduction.

The application provides that the high-class composition scoring specification (the high-class composition requires judging the quality of the student composition from the aspects of content, expression and development level 3) can be used for commencing the composition from the four aspects of content, expression, structure and development, and the commentary dimension of each aspect is provided for the content examined in each aspect, preferably, as shown in fig. 3, the commentary dimension of the content aspect can comprise ideological health, emotion type, compliance with questions and content enrichment, the commentary dimension of the expression aspect can comprise basic expression, line specification, language fluency and document compliance, the commentary dimension of the structure aspect can comprise structure rigor, and the commentary dimension of the development aspect can comprise literature. It should be noted that, in practical application, the score of the target composition may be determined from all the above-mentioned dimensions, or a part of the dimensions may be selected from the above-mentioned dimensions, the score of the target composition may be determined from the selected dimensions, and the score of the target composition may be determined from which dimensions according to specific requirements.

Next, a description will be given of a process of determining a score rank of a target document in each of the above-mentioned review dimensions.

(1) Determining scoring rankings of target composition in four review dimensions of content

(1-1) Determining a scoring grade for the target composition in the review dimension of mental health

The process of determining the scoring grade of the target composition in the review dimension of the thought health comprises the following steps: judging whether each sentence in the target work contains a low-custom language or not; and determining the grading of the target composition in the evaluation dimension of the thought health according to the judging result of each sentence in the target composition.

Wherein the process of determining whether each sentence in the target work contains a low-custom language may include: and inputting each sentence in the target text to a pre-established sentence classification model to obtain a classification result which is output by the sentence classification model and used for indicating whether each sentence in the target text contains a low-custom language.

The evaluation dimension of the health of the idea comprises two grading grades, namely a first grade and a second grade, if the objective composition comprises a low-custom language, the grading of the objective composition in the evaluation dimension of the health of the idea is determined to be the second grade, and if the objective composition does not comprise the low-custom language, the grading of the objective composition in the evaluation dimension of the health of the idea is determined to be the first grade. It should be noted that, the review dimension of the health of the mind includes two grading steps, which is only an example, and the number of grading steps can be determined according to specific requirements in practical application.

(1-2) Determining a scoring grade for the target composition in the review dimension of emotion type

The process of determining the scoring grade of the target composition in the review dimension of emotion type comprises the following steps: identifying emotion types expressed by the target composition; and determining the grading of the target composition on the evaluation dimension of the emotion type according to the emotion type identification result of the target composition.

In one possible implementation manner, the emotion type expressed by the target composition can be determined based on an emotion classification method of an emotion dictionary, in another possible implementation manner, the emotion type expressed by the target composition can be determined based on an emotion classification model obtained through training in advance, and optionally, the emotion type expressed by the target composition can be one of positive emotion and negative emotion. Optionally, when determining the emotion type expressed by the target composition, the intensity of the emotion type expressed by the target composition can be determined, and then the grading of the target composition in the evaluation dimension of the emotion type is determined according to the emotion type expressed by the target composition and the intensity of the emotion type expressed by the target composition.

(1-3) Determining a scoring grade for the target composition in the review dimension of fitting the topic

Determining a scoring grade of the target composition in the review dimension conforming to the topic comprises: acquiring a representation vector of each word in the title of the target composition and a representation vector of each word in the text of the target composition, and determining a target vector corresponding to each word in the title of the target composition based on the acquired vectors, wherein the target vector corresponding to one word can represent the related condition of the word and each word in the text of the target composition; and determining the matching degree of the topics and the texts of the target composition according to the target vector corresponding to each word in the topics of the target composition, and determining the grading of the target composition on the review dimension conforming to the topics according to the matching degree of the topics and the texts of the target composition.

It should be noted that, the matching degree of the title and the text of the target composition can represent the condition that the target composition accords with the title, and the higher the matching degree of the title and the text of the target composition, the more accords with the title.

(1-4) Determining scoring for the target composition in the review dimension of content enrichment

Whether the content is full or not is mainly determined to be complete and detailed, and the content is full and comprehensive determination needs to be performed on the composition, because the correction result of the target composition includes some useful information, such as which sentences are graceful sentences, which sentences include the repair, which sentences include the description, and so on, in view of this, the embodiment determines the scoring of the target composition in the review dimension of the content is full in combination with the correction result.

The process of determining scoring for the target composition in the review dimension of content enrichment includes: determining a chapter representation vector of the target composition based on the target composition, basic information of the target composition, a correction result corresponding to the target composition at a sentence level and a correction result corresponding to the target composition at a chapter level (particularly a departure detection result of the target composition); and classifying the chapter representation vectors through a classifier to obtain scoring grades in the review dimension of the content enrichment of the target composition. The basic information of the target composition may include, but is not limited to, one or more of the following: the word part of the word used in the target composition, the sentence pattern used in the target composition, the length of the target composition, the number of words in the target composition, the number of single words in the target composition, the number of paragraphs in the target composition, the number of sentences in the target composition, and the like.

Optionally, after obtaining the scoring grade of the target composition in the review dimension of content enrichment based on the classifier, whether the scoring grade obtained based on the classifier is reasonable or not may be determined based on the correction result (such as the number of written sentences in the target composition) corresponding to the target composition at the sentence level, and if not, the scoring grade is corrected.

(2) Determining scoring rankings on multiple reviews of a target composition in terms of expression

(2-1) Determining a scoring grade of the target composition in the review dimension of the underlying expression

The basic expression aspect is mainly to judge whether the use condition of basic capability points (particularly words and sentence patterns) of composition is reasonable. The process of determining a scoring grade for the target composition in the review dimension of the underlying expression may include: based on the word making list and/or the common idiom library respectively corresponding to different learning segments, carrying out word making and/or common idiom recognition on the target composition so as to obtain word use conditions of the target composition on different learning segments; performing sentence pattern recognition (such as exclamation sentences and question sentences) on the target composition based on the appointed sentence pattern so as to obtain the sentence pattern service condition of the target composition; and determining the grading of the target composition in the evaluation dimension of the basic expression according to the word use condition of the target composition in different school segments and the sentence pattern use condition of the target composition.

Optionally, when determining the scoring grade of the target composition in the evaluation dimension of the basic expression according to the word use condition of the target composition in different school segments and the sentence pattern use condition of the target composition, the scoring grade of the target composition in the word use condition can be determined firstly based on the word use condition of the target composition in different school segments, the scoring grade of the target composition in the sentence pattern use condition can be determined according to the sentence pattern use condition of the target composition, and then the scoring grade of the target composition in the evaluation dimension of the basic expression can be determined according to the scoring grade of the target composition in the word use condition and the scoring grade of the target composition in the sentence pattern use condition.

Optionally, three grading steps can be set for the word use condition, namely a first step, a second step and a third step, wherein the third step represents that the word use condition is lower than the current school, the second step represents that the word use condition is synchronous with the current school, and the first step represents that the word use condition exceeds the current school. The embodiment counts the number of new words and/or idioms used in each school paragraph in the target composition, and determines the grading of the target composition on the word use condition based on the counted number and the set number threshold. It should be noted that, in order to obtain a reasonable grading, the number threshold needs to be set appropriately.

(2-2) Determining a scoring grade for the target composition in the review dimension of the line specification

The process of determining the scoring grade of the target composition in the review dimension of the line document specification comprises the following steps: detecting whether the title, and/or paragraph, and/or network term, and/or punctuation mark, and/or document format of the target composition meets the specification or not to obtain a detection result of the target composition on the line specification, and determining the grading of the target composition on the review dimension of the line specification according to the detection result of the target composition on the line specification.

For the detection of the questions and the paragraphs, the questions and the segmentation results in the target composition can be obtained based on a text processing mode, and whether the questions and the paragraphs of the target composition meet the specifications or not is determined according to the obtained questions and the segmentation results; for the detection of the network expression, the network expression in the target composition can be detected based on a pre-constructed network expression table, and whether the network expression in the target composition meets the specification is determined according to the detection result; for the detection of punctuation marks, whether the punctuation marks of the target composition are wrong or not can be detected based on rules or models (if the punctuation error detection is carried out when the target composition is corrected from word level, the punctuation error detection result can be directly utilized here); for the detection of the genre format, a target genre (i.e., the genre of the target text) may be obtained, and whether the format book of the target text meets the format requirement of the target genre is detected.

(2-3) Determining scoring for the target composition in the language fluency review dimension

Language fluency is a dimension evaluated from the viewpoints of grammar specification, line fluency and the like in composition. The process of determining the scoring grade of the target composition in the language fluency scoring dimension comprises the following steps: extracting feature vectors from basic information, statistics information of the word error correction and grammar error detection results and combined information of the target composition; and classifying the extracted feature vectors through a classifier to obtain grading of the target composition in the review dimension of language fluency.

The basic information of the target composition may include, but is not limited to, one or more of the following: the word part of the word used in the target composition, the sentence pattern used in the target composition, the length of the target composition, the number of words in the target composition, the number of single words in the target composition, the number of paragraphs in the target composition, the number of sentences in the target composition, and the like. The statistics information of the error correction and grammar error detection results of the target composition is information obtained by counting the wrongly written word and grammar error of the target composition, and the statistics information can comprise: the combination information of the target composition comprises collocation of the target composition, punctuation mark combination among sentences and the like.

It should be noted that, the association between language fluency and grammar error is relatively tight, and generally, under the condition of wrong words or more unhappy choice of words, the grading of language fluency can be greatly reduced, so that the application uses the statistical information of the word error correction and grammar error detection results as the important basis for determining the grading of the target composition in the evaluation dimension of language fluency.

Optionally, after obtaining the grading of the target composition in the language fluency evaluation dimension based on the classifier, determining whether the grading obtained based on the classifier is reasonable based on the statistics information of the word error correction and grammar error detection results of the target composition, and if not, correcting the grading of the target composition in the language fluency evaluation dimension.

(2-4) Determining a scoring grade for the target composition in terms of the document meeting the review dimension

The genre matching is mainly to determine whether the genre of the target composition is consistent with the specified genre (for example, the author is required to write a paper for the proposal, and the specified genre is the proposal).

The process of determining the scoring grade of the target composition in the document meeting the review dimension comprises the following steps: and identifying the genre of the target composition based on a pre-established genre identification model, and determining the grading of the target composition in the review dimension according to whether the genre of the target composition is consistent with the appointed genre. The genre recognition model is obtained by training a training composition marked with a genre.

The process of identifying the genre of the target composition based on the pre-established genre identification model comprises the following steps: the method comprises the steps of determining the probability that the genre of the target composition is the set genre based on the genre identification model, and determining the genre of the target composition according to the probability that the genre of the target composition is the set genre. Wherein the set genres may, but are not limited to, include one or more of the following genres: diary, letter, lecture, post-reading feel, poem, narrative, writer, writing, discussion, description, and the like.

More specifically, the process of determining the probability that the genre of the target composition is the set genres based on the genre identification model includes:

And d1, encoding each word in the target text based on the genre recognition model to obtain an encoding vector of each word in the target text.

Referring to fig. 4, a topology of a genre recognition model is shown, and as shown in fig. 4, the genre recognition model includes a word encoding module, a word level attention module, a sentence encoding module, a sentence level attention module, and a classification module.

When the genre of the target composition is identified, sentence segmentation is firstly carried out on the target composition, then word segmentation is carried out on each sentence obtained through the sentence segmentation, then the word coding module of each word obtained through the word segmentation is input into the genre identification model for coding, and the word coding module outputs the coding vector of each word in the target composition. The purpose of encoding words is to map the words to a high-dimensional semantic vector space.

And d2, performing attention calculation on the coding vector of each word in the target text based on the genre recognition model to obtain the attention vector of each word in the target text, and obtaining the representation vector of each sentence in the target text based on the attention vector of each word in the target text.

The method comprises the steps of inputting the coding vector of each word in the target text into a word level attention module, wherein the word level attention module can determine the attention weight of each word in the target text, and further can obtain the attention vector of each word in the target text based on the coding vector and the attention weight of each word in the target text, and further can obtain the representation vector of each sentence in the target text.

W _it in fig. 4 represents the t-th word of the i-th sentence in the object, α _it represents the attention weight of the word w _it, the attention vector of w _it can be determined according to the encoding result of w _it and α _it, and the expression vector s _i of the i-th sentence in the object can be determined according to the attention vector of each word in the i-th sentence. Note that u _w in fig. 2 represents a query vector when calculating the word-level attention weight.

Note that the purpose of performing attention computation on the code vector of each word in the target text is to highlight important word information, so that the important word information can be focused on by subsequent operations.

And d3, encoding the representation vector of each sentence in the target text based on the genre recognition model to obtain an encoded vector of each sentence in the target text.

The sentence coding module in this embodiment may be a bidirectional GRU model, and fig. 5 shows the structure of the GRU model, and the operation of the GRU model is as follows:

z_t＝σ(W_zx_t+U_zh_t-1+b_z) (4)

r_t＝σ(W_rx_t+U_rh_t-1+b_r) (6)

Where x _t denotes the input of the current time step, h _t-1 denotes the hidden vector of the last time step, z _t and r _t denote the update gate and reset gate, respectively, W _z,U_z,b_z and W _r,U_r,b_r are parameters of the update gate and reset gate, respectively, And h _t respectively memorizing the vector of the current time step and finally outputting the hidden vector of the current time step.

It should be noted that, the GRU model has two gates, namely an update gate and a reset gate, and the structural arrangement overcomes the problem that RNNs cannot solve remote dependence well, and has stronger characterization capability for longer sentences. To this end, the present embodiment uses a bi-directional GRU model to characterize sentences (i.e., encodes sentence representation vectors obtained in the previous step):

Where e _i represents a vector representation of the i-th word in the sentence.

Merging the results of the bidirectional GRU models to obtain the coding result of sentences:

And d4, performing attention calculation on the coding vector of each sentence in the target composition based on the genre recognition model to obtain an attention vector of each sentence in the target composition, and determining the chapter representation vector of the target composition based on the attention vector of each sentence in the target composition.

The encoded vector of each sentence in the target composition is input into a sentence-level attention module, the sentence-level attention module can determine the attention weight of each sentence in the target composition, and further, the attention vector of each sentence in the target composition can be obtained based on the encoded vector and the attention weight of each sentence in the target composition, and the chapter representation vector of the target composition can be further obtained based on the attention vector and the attention weight.

Note that the purpose of performing attention computation on the code vector of each sentence in the target work is to further highlight important word information, so that important word information can be focused on when classification is performed subsequently.

And d5, classifying the chapter representation vectors of the target composition based on the genre identification model to obtain the probability that the genres of the target composition are set for each genre.

Specifically, the chapter representation vector of the target composition is input into the classification module of the genre recognition model for classification. Optionally, keywords (such as titles) can be extracted from the target composition, the representing vectors of the keywords and the chapter representing vectors of the target composition are input into the classification module of the genre recognition model together for classification, and the classification effect of the model can be improved by adding the keyword information.

(3) Determining scoring rankings of target composition in structurally review dimensions

The review dimension of the target composition in terms of structure mainly comprises the review dimension of strict structure. Determining a scoring grade of the target composition in a structurally rigorous review dimension comprises: and determining the grading of the target composition in the evaluation dimension of strict structure according to the result of the modification of the target composition in the chapter structure.

(4) Determining scoring rankings of target composition in review dimension of development

The review dimension of the target composition in development mainly comprises the review dimension of the literature. If the composition has bright spots on any aspect of language expression, vocabulary, and congratulation, the composition can be regarded as having the literature.

The process of determining the scoring grade of the target composition in the review dimension of the literature comprises the following steps: determining a chapter representation vector of the target composition based on the target composition, basic information of the target text and an correcting result of the target composition at a sentence level; and classifying the chapter representation vectors to determine the grading of the target composition in the review dimension of the literature. It should be noted that, when determining the chapter representation vector of the target composition based on the target composition, the basic information of the target text, and the correction result of the target composition at the sentence level, the chapter representation vector of the target composition may be determined based on information (e.g., the number of good words, the word frequency of words, the number of words used, the abundance of languages, the organization arrangement of paragraphs, etc.) related to the extraction in the three information of the target composition, the basic information of the target text, and the correction result of the target composition at the sentence level.

The review dimension of the document may include three steps, namely a first step, a second step and a third step, wherein the first step represents the document, indicates that the whole vocabulary and the congratulation of the article are excellent, has multiple graceful expressions, has very prominent bright spots, the second step represents the document, indicates that the local vocabulary and the congratulation of the article are better, the bright spots are general, and the third step represents the document is insufficient, and indicates that the article lacks rich vocabulary and congratulation and has insufficient bright spots.

Fourth embodiment

The present embodiment is mainly described in "step S104" in the above embodiments: and generating the comment of the target composition according to the grading of the target composition in a plurality of comment dimensions, and introducing the realization process of the comment of the target composition.

Referring to fig. 6, a flow diagram illustrating generation of comments of a target composition according to scoring classification of the target composition in multiple comment dimensions may include:

Step S601: and determining the comment corresponding to the scoring grade of the target composition on each scoring dimension based on the pre-constructed scoring dimension and the corresponding relation between the scoring grade and the comment.

The following table shows an example of the correspondence of review dimensions, score rankings and reviews:

table 2 correspondence between review dimension, score rating and comment

For example, in the step S104, the "multiple review dimensions" include "meeting the topic" and "content enrichment", the scoring grade of the target composition in the review dimension "meeting the topic" is a first grade, the scoring grade of the target composition in the review dimension "content enrichment" is also a first grade, and then, based on the correspondence between the review dimension and the scoring grade indicated above and the scoring grade, the scoring grade of the target composition in the review dimension "meeting the topic" is obtained as "the content of the composition defines the meaning of the title", the scoring grade of the target composition in the review dimension "content enrichment" is "the content of the target composition is very rich, and the character is vividly highlighted.

It should be noted that, in the correspondence between the review dimension, the score rank and the review, one score rank may correspond to one review, or may correspond to two reviews, for example, two reviews corresponding to one rank in the review dimension of the content enrichment in the table above, in this case, if there are multiple reviews in the correspondence between the score ranks of the target composition in a certain dimension, one rank may be randomly selected from the multiple reviews, for example, the score rank of the target composition in the review dimension of the content enrichment is one rank, in the correspondence of the table above, one rank in the review dimension of the content enrichment corresponds to two reviews, and one rank may be randomly selected from the two reviews as the corresponding review of the score rank of the target composition in the review dimension of the content enrichment.

Step S602: and generating the comments of the target composition according to the comments corresponding to the grading of the target composition on each comment dimension.

In one possible implementation manner, the comment corresponding to the grading of the target composition in each review dimension can be directly used as the comment of the target composition; in another possible implementation manner, the comments corresponding to the scoring grades of the target composition in each scoring dimension can be spliced and combined by using the connecting word and/or the turning word based on the comment template so as to obtain a complete comprehensive comment as the comment of the target composition. The following is an example of a comprehensive comment:

"this work is written so well, you really expect that you can also do so with the next work-!

In terms of content, the ideas of the text are positive, so that not only are the meaning of titles of the composition content clear, but also the content is rich, and the personality of the person is vivid. In terms of expression, a relatively large number of advanced words are used, exclamation sentence pattern representation emphasis is used, sentence expression is relatively rich, but part of punctuation marks are used improperly.

Some promotion suggestions are given to you, the applicable object of mastering the words is very critical, in addition, the questionable sentence pattern is remembered when the questioning is expressed, and most importantly, attention is paid to the compliance of the punctuation with the use standardization.

Children's book of ancient people to know what things we want to know about, we want to learn to draw nutrition from the book. "

In the above embodiment, it is mentioned that, in addition to determining the score of the target composition in the plurality of review dimensions, the score of the target composition may be determined, in which case, in addition to the score of the target composition in each of the plurality of review dimensions, the score of the target composition may be determined, and in determining the score of the target composition, the score of the target composition may be determined based on the pre-established correspondence between the score of the target composition and the score, and finally the comprehensive score of the target composition may be generated based on the score of the target composition in each of the plurality of review dimensions and the score of the target composition.

Through the mode, comments with rich contents and strong guidance can be obtained.

Fifth embodiment

The embodiment of the application also provides a composition review device, which is described below, and the composition review device described below and the composition review method described above can be correspondingly referred to each other.

Referring to fig. 7, a schematic structural diagram of a composition review device provided in an embodiment of the present application may include: a detection module 701, a correction module 702, a score ranking determination module 703 and a comment generation module 704.

The detection module 701 is configured to detect whether a target composition to be reviewed is an abnormal composition;

The correcting module 702 is configured to correct the target composition from the word level, the sentence level, and the chapter level, respectively, when the target composition is not an abnormal composition, so as to obtain correction results corresponding to the target composition at the word level, the sentence level, and the chapter level, respectively;

a scoring grade determining module 703, configured to determine a scoring grade of the target composition from a plurality of review dimensions, so as to obtain the scoring grade of the target composition in the plurality of review dimensions;

And the comment generation module 704 is used for generating comments of the target composition according to the grading of the target composition in the plurality of comment dimensions.

In one possible implementation manner, the plurality of review dimensions are a plurality of review dimensions corresponding to the learning segment to which the target composition belongs in the set plurality of dimensions;

The scoring grade determining module 703 is specifically configured to determine a scoring grade of the target composition from a plurality of scoring dimensions corresponding to the learning segment to which the target composition belongs, so as to obtain the scoring grade of the target composition on the plurality of scoring dimensions corresponding to the learning segment to which the target composition belongs.

In one possible implementation, the comment generation module 704 includes: the comment determination sub-module and the comment generation sub-module.

The comment determination submodule is used for determining comments corresponding to the scoring grades of the target composition in each scoring dimension based on the corresponding relation between the pre-constructed scoring dimension and the scoring grade and the comments;

And the comment generation sub-module is used for generating comments of the target composition according to comments corresponding to the grading of the target composition on each comment dimension.

In one possible implementation, the composition review device may further include: and the overall grading determination module.

And the overall grading determination module is used for determining the grading of the target composition overall.

The comment generation module 704 is specifically configured to generate a comment of the target composition according to the score ranks of the target composition in the multiple review dimensions and the score ranks of the target composition as a whole.

In one possible implementation, the overall score ranking determination module may include: a first score determination sub-module, and/or a second score determination sub-module, and an overall score ranking determination sub-module.

The first score determining submodule is used for predicting the overall score of the target composition based on a pre-established score prediction model and taking the overall score as the first score of the target composition;

the second score determining submodule is used for predicting the difference between the target composition and the template in the template set based on a pre-established difference prediction model, and predicting the overall score of the target composition as the second score of the target composition based on the difference between the target composition and the template in the template set and the score of the template in the template set, wherein the template in the template set and the target composition belong to the same theme;

The overall score grading determination submodule is used for determining the score grading of the overall target composition based on the first score of the target composition and/or the second score of the target composition.

In a possible implementation manner, the detection module 701 is specifically configured to determine that the target composition is an abnormal composition if the target composition meets at least one of the following conditions:

In one possible implementation, the modification module 702 includes: a word level altering module, a sentence level altering module and a chapter level altering module.

And the word level correcting module is used for carrying out one or more of the following treatments on the target composition: language accuracy analysis, network term retrieval and error punctuation recognition, and generating a correction result corresponding to the target composition at the word level according to the processing result; wherein the language accuracy analysis comprises an individual word error correction, and/or a grammar error detection, and/or idiom class error detection, and/or an ancient poetry error detection;

A sentence-level correction module, configured to perform one or more of the following treatments on the target composition: graceful sentence recognition, advanced vocabulary statistics, sentence repair method recognition and descriptive sentence recognition, and generating a modification result corresponding to the target composition at a sentence level according to a processing result;

And the correction module of the chapter level is used for carrying out chapter structure recognition and/or theme recognition on the target composition, and generating a correction result corresponding to the target composition at the chapter level according to the recognition result.

In one possible implementation, the word-level altering module is specifically configured to, for each paragraph in the target composition, when performing error punctuation recognition on the target composition:

In one possible implementation, the word-level correction module is specifically configured to, for each sentence in the target composition, perform an individual word correction on the target composition:

In one possible implementation, the word-level altering module is specifically configured to, for each word in the sentence, when detecting a misplaced word in the sentence:

In one possible implementation, the word-level correction module is specifically configured to, for each sentence in the target composition, when performing syntax error detection on the target composition:

In one possible implementation, the plurality of comment dimensions are a plurality of comment dimensions of one or more of the following four aspects: content, expression, structure, development, wherein:

In one possible implementation, the score ranking determining module 703 is specifically configured to, when determining the score ranking of the target composition in the review dimension of the thought health:

Judging whether each sentence in the target composition contains a low-custom language or not, and determining the grading of the target composition in the evaluation dimension of the thought health according to the judging result of each sentence in the target composition;

the scoring grade determining module 703 is specifically configured to, when determining the scoring grade of the target composition in the review dimension of the emotion type:

identifying emotion types expressed by the target composition, and identifying results according to the emotion types of the target composition; determining a grading of the target composition in the evaluation dimension of the emotion type;

the scoring grade determining module 703 is specifically configured to, when determining the scoring grade of the target composition in the review dimension of meeting the topic:

Acquiring a representation vector of each word in the title of the target composition and a representation vector of each word in the text of the target composition, and determining a target vector corresponding to each word in the title of the target composition based on the acquired vectors, wherein the target vector corresponding to one word can represent the related condition of the word and each word in the text of the target composition; determining the matching degree of the title and the text of the target composition according to the target vector corresponding to each word in the title of the target composition, and determining the grading of the target composition in the evaluation dimension conforming to the title according to the matching degree of the title and the text of the target composition; determining a scoring grade of the target composition in the review dimension of the content enrichment, comprising: determining a chapter representation vector of the target composition based on the target composition, basic information of the target composition and correction results corresponding to the target composition on sentence level and chapter level respectively; and determining the grading of the target composition in the review dimension of the content enrichment by classifying the chapter representation vectors.

In one possible implementation, the score ranking determining module 703 is specifically configured to, when determining the score ranking of the target composition in the review dimension of the base expression:

Based on the word making list and/or the common idiom library respectively corresponding to different school segments, carrying out word making and/or common idiom recognition on the target composition so as to obtain word use conditions of the target composition on different school segments; performing sentence pattern recognition on the target composition based on the appointed sentence pattern to obtain the sentence pattern service condition of the target composition; determining scoring classification of the target composition in a review dimension of basic expression based on word use conditions of the target composition in different school segments and sentence pattern use conditions of the target composition;

the scoring grade determining module 703 is specifically configured to, when determining the scoring grade of the target composition in the review dimension of the line document specification:

detecting whether the title, and/or paragraph, and/or network term, and/or punctuation mark, and/or genre format of the target composition meets the specification; determining the grading of the target composition in the review dimension of the line text specification according to the detection result of the target composition in the line text specification;

the scoring grade determining module 703 is specifically configured to, when determining the scoring grade of the target composition in the language fluency review dimension:

Extracting feature vectors from the basic statistical information of the target composition, the statistical information of the word error correction and grammar error detection results and the collocation and combination of the target composition; determining a scoring grade of the target composition in the language fluency scoring dimension by classifying the extracted feature vectors;

the scoring grade determining module 703 is specifically configured to, when determining that the target composition meets the scoring grade in the review dimension:

Identifying the genre of the target composition based on a pre-established genre identification model, wherein the genre identification model is obtained by training a training composition marked with the genre; and determining the grading of the target composition in the dimension of the document meeting the review according to whether the genre of the target composition is consistent with the appointed genre.

The scoring ranking determination module 703 is specifically configured to, when identifying the genre of the target composition based on a pre-established genre identification model:

encoding each word in the target text based on the genre recognition model to obtain an encoding vector of each word in the target text; performing attention calculation on the coding vector of each word in the target text based on the genre recognition model to obtain an attention vector of each word in the target text, and obtaining a representation vector of each sentence in the target text based on the attention vector of each word in the target text; encoding the representation vector of each sentence in the target text based on the genre identification model to obtain an encoded vector of each sentence in the target text; performing attention calculation on the coding vector of each sentence in the target composition based on the genre identification model to obtain an attention vector of each sentence in the target composition, and determining a chapter representation vector of the target composition based on the attention vector of each sentence in the target composition; classifying the chapter representative vector of the target composition based on the genre identification model to obtain the genre of the target composition.

In one possible implementation manner, the modification result of the target composition at the chapter level includes modification result of the target composition at the chapter structure;

The scoring grade determining module 703 is specifically configured to, when determining the scoring grade of the target composition in the review dimension of strict structure:

In one possible implementation, the score ranking determining module 703 is specifically configured to, when determining the score ranking of the target composition in the review dimension of the document:

determining a chapter representation vector of the target composition based on the target composition, the basic information of the target text and the correction result of the target composition at the sentence level; and determining the grading of the target composition in the review dimension of the literature by classifying the chapter representation vectors.

The composition review device provided by the embodiment of the application can automatically review the composition to be reviewed, and avoids the problems caused by manual participation because manual participation is not needed, and the composition review device provided by the embodiment of the application not only can obtain the corresponding correction results of the target composition on the word level, the sentence level and the chapter level respectively, but also can obtain the comments of the target composition on a plurality of review dimensions, namely, the review results are rich, the rich review results can play a good guiding role for writers, and the user experience is good.

Sixth embodiment

The embodiment of the application also provides a composition review device, referring to fig. 8, which shows a schematic structural diagram of the composition review device, the composition review device may include: at least one processor 801, at least one communication interface 802, at least one memory 803, and at least one communication bus 804;

In the embodiment of the present application, the number of the processor 801, the communication interface 802, the memory 803 and the communication bus 804 is at least one, and the processor 801, the communication interface 802 and the memory 803 complete communication with each other through the communication bus 804;

The processor 801 may be a central processing unit CPU, or an Application-specific integrated Circuit ASIC (Application SPECIFIC INTEGRATED Circuit), or one or more integrated circuits configured to implement embodiments of the present invention, etc.;

the memory 803 may include a high-speed RAM memory, and may further include a non-volatile memory (non-volatile memory), etc., such as at least one magnetic disk memory;

wherein the memory stores a program, the processor is operable to invoke the program stored in the memory, the program operable to:

Alternatively, the refinement function and the extension function of the program may be described with reference to the above.

Seventh embodiment

The embodiment of the present application also provides a computer-readable storage medium storing a program adapted to be executed by a processor, the program being configured to:

Finally, it is further noted that relational terms such as first and second, and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.

In the present specification, each embodiment is described in a progressive manner, and each embodiment is mainly described in a different point from other embodiments, and identical and similar parts between the embodiments are all enough to refer to each other.

The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for reviewing a composition, comprising:

generating comments of the target composition according to grading of the target composition on the plurality of comment dimensions;

The modifying the target composition from word level includes:

Performing error punctuation recognition on the target composition;

performing error punctuation recognition on the target composition, including:

For each paragraph in the target document:

2. The composition review method according to claim 1, wherein the plurality of review dimensions are a plurality of review dimensions corresponding to a school to which the target composition belongs among the set plurality of dimensions;

3. The composition review method of claim 1 wherein the generating a review of the target composition according to the scoring rankings of the target composition in the plurality of review dimensions comprises:

4. The composition review method of claim 1 further comprising:

determining a scoring grade of the target composition;

5. The composition review method of claim 4 wherein the determining a scoring grade for the target composition whole comprises:

6. The method for reviewing a composition according to claim 1, wherein said detecting whether a target composition to be reviewed is an abnormal composition comprises:

7. The method for reviewing a composition according to claim 1, wherein,

The modifying the target composition from word level further comprises: one or more of the following treatments are performed on the target composition: language accuracy analysis and network term retrieval, and generating a correction result corresponding to the target composition at the word level according to the processing result; wherein the language accuracy analysis comprises an individual word error correction, and/or a grammar error detection, and/or idiom class error detection, and/or an ancient poetry error detection;

Correcting the target composition from sentence level, comprising:

Correcting the target composition from the chapter level, comprising:

8. The composition review method of claim 7 wherein the process of correcting the target composition for the written word comprises:

For each sentence in the target work:

9. The method of claim 8, wherein detecting mispronounced words in the sentence comprises:

For each word in the sentence:

10. The composition review method of claim 7 wherein the act of grammatically error detecting the target composition comprises:

For each sentence in the target work:

11. The composition review method of claim 1 wherein the plurality of review dimensions are of one or more of the following four aspects: content, expression, structure, development, wherein:

12. The composition review method of claim 11 wherein determining a scoring grade of the target composition in the thought health review dimension comprises:

Identifying the emotion type expressed by the target composition;

13. The composition review method of claim 11 wherein determining a scoring grade of the target composition in the review dimension of the base expression comprises:

14. The composition review method of claim 13 wherein the identifying the genre of the target composition based on the pre-established genre identification model comprises:

15. The composition review method of claim 11 wherein the target composition includes a change in chapter structure of the target composition in change results at chapter level;

16. The composition review method of claim 11 wherein determining a scoring grade of the target composition in the review dimension comprises:

17. A composition review device, comprising: the system comprises a detection module, a correction module, a grading determination module and a comment generation module;

the comment generation module is used for generating comments of the target composition according to the grading of the target composition in the plurality of comment dimensions;

The altering module alters the target composition from word level, comprising:

the correction module carries out error punctuation recognition on the target composition;

the correcting module carries out error punctuation recognition on the target composition, and the correcting module comprises the following steps:

The correction module, for each paragraph in the target document: removing punctuation marks in the paragraphs, and taking the paragraphs after the punctuation marks are removed as target texts; predicting a label category corresponding to each word in the target text based on a pre-established punctuation prediction model; the label category corresponding to one word is used for indicating whether punctuation marks exist behind the word and why the punctuation marks exist; determining punctuation marks predicted for the target text according to the label category corresponding to each word in the target text; from the punctuation in the paragraph and the punctuation predicted for the target text, determining that the erroneous punctuation is used in the punctuation of the paragraph.

18. A composition review device, comprising: a memory and a processor;

the memory is used for storing programs;

the processor is configured to execute the program to implement the steps of the composition review method according to any one of claims 1 to 16.

19. A readable storage medium having stored thereon a computer program, which when executed by a processor, implements the steps of the composition review method of any one of claims 1 to 16.