Document fluency detection method and device, electronic equipment and medium

[0039] The following describes exemplary embodiments of the present application with reference to the accompanying drawings, which include various details of the embodiments of the present application to facilitate understanding, and should be considered as merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. Likewise, for clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0040] The method, device, electronic equipment, and medium for detecting document fluency of the present application are described below with reference to the accompanying drawings.

[0041] The language expression of the composition is the main testing aspect in the composition review. The language expression testing includes two aspects: whether the language expression is beautiful and whether the language expression is smooth. Among them, the graceful language expression is mainly reflected in emotion and language beauty, while the fluent language is mainly reflected in the smoothness of sentences and whether there are grammar and typos.

[0042] At present, for the automatic review of composition, only the text in the composition can be corrected, and only the typos or lower-level grammatical errors of the fixed template can be corrected. The types of errors that can be resolved are single. For the composition Frequent errors such as ill-sentence errors, logical errors, or improper use of words, lack of effective and mature technology, and low detection accuracy and recall rates.

[0043] In response to the above problems, this application discloses a method for detecting document fluency, by calculating the gracefulness value, smoothness value and error value of the document to be detected, and then according to the calculated gracefulness value, smoothness value and error degree Value, the fluency value of the document to be detected is generated, thereby realizing the automatic detection of document grace, smoothness and error degree, thereby realizing the detection of document fluency. When the solution of this application is applied to composition review, automatic review of the composition is realized, and the composition can be scored from multiple aspects of grace, smoothness and error, and the review standard is unified, which can improve the accuracy of composition scoring.

[0044] figure 1 It is a schematic flow chart of the method for detecting document fluency according to the first embodiment of the present application. The method can be executed by the document fluency detecting apparatus provided by the present application, or by the electronic equipment provided by the present application, where the electronic equipment may be The server can also be a terminal device such as a desktop computer, a notebook computer, a smart phone, and a wearable device. Hereinafter, the document fluency detection device provided by the application executes the document fluency detection method provided by the application as an example to explain the application.

[0045] Such as figure 1 As shown, the method for detecting document fluency may include the following steps:

[0046] Step 101: Obtain a document to be detected.

[0047] Among them, the document to be detected may be, for example, a student's composition, an article created by a user, and so on.

[0048] In this embodiment, the document to be detected can be obtained in different ways.

[0049] As an example, when the document to be detected is a student’s handwritten composition, the teacher can take a picture of the student’s handwritten composition to obtain a picture containing the user’s composition, and upload the picture to the document fluency detection device. The detection device performs optical character recognition (Optical Character Recognition, OCR) on the picture uploaded by the teacher, obtains the composition of the student, and uses the obtained composition as the document to be detected.

[0050] As an example, when students need to self-check the composition they have made, or other users need to check the created articles, they can directly enter the text in the input interface provided by the document fluency detection device. After the entry is completed, the document The fluency detection device obtains the document to be detected.

[0051] Step 102: Calculate the grace value of the document to be detected.

[0052] The language of beauty acts on the creator’s sense of language, which includes sense of commensurability, sense of proportion, and sense of rhythm. Specifically, the sense of language is to regulate the use of words and follow the order of words. Languages ​​that can arouse aesthetic feelings are usually sentences with clear expressions, coherent and decent expressions, vividness, texture and distinctive personality.

[0053] In the embodiment of the present application, in order to evaluate whether the document to be detected has a graceful expression, the grace value of the document to be detected can be calculated, and whether the expression of the document to be detected is graceful or not can be evaluated based on the grace value of the document to be detected.

[0054] As a possible implementation, in this embodiment, a graceful sentence two classifier can be pre-trained, and the graceful sentence probability of each sentence in the document to be detected is calculated by the graceful sentence two classifier, and then according to the graceful sentence of each sentence Probability, get the gracefulness value of the document to be detected. For example, the average value of the graceful sentence probability of each sentence can be calculated, and the obtained average value can be used as the gracefulness value of the document to be detected.

[0055] Among them, the training data of the elegant sentence two classifier can be obtained from Internet data and labeled. For example, a large number of primary and middle school students’ essays and Internet writers’ articles can be obtained from Internet data, and the obtained articles and essays can be labeled and labeled. Whether the sentences in the composition and the article are graceful sentences, generate a large number of effectively labeled sentence samples as training data to train the two-classification network, and obtain the trained graceful sentence two-classifier for the graceful sentence detection of the document to be detected.

[0056] Step 103: Calculate the smoothness value and the error degree value of the document to be detected.

[0057] In this embodiment, after the document to be detected is obtained, the smoothness value of the document to be detected can also be calculated, and the error degree value of the document to be detected can be calculated.

[0058] Linguistic fluent refers to the standard, accurate, coherent and appropriate language. The fluent language does not lie in the gorgeous rhetoric, but in the fluent manner and accurate expression. Therefore, in this embodiment, the consistency value of the document to be detected can be determined by calculating the confusion degree of each paragraph in the document to be detected and the abnormality of the collocation of each sentence.

[0059] Among them, the confusion degree of the paragraph can be calculated through the pre-trained Chinese Deep Neural Network (DNN) language model. The Chinese DNN language model can calculate the confusion degree of each sentence. The confusion degree can reflect the smoothness of the sentence. The smaller the degree, the more smooth the sentence. According to the confusion of each sentence, the confusion value of the paragraph can be obtained. For example, the average value of the perplexity of each sentence can be calculated as the confusion degree of the paragraph, or the perplexity of each sentence can be weighted and summed to obtain the confusion degree of the paragraph, which is not limited in this application.

[0060] The collocation of each sentence in the paragraph is abnormal. The perplexity of each sentence can be compared with the preset perplexity threshold. The sentence with the perplexity greater than the perplexity threshold is determined as an abnormal sentence, and the abnormal sentence contained in the paragraph is calculated in the entire paragraph. The proportion of each sentence in the paragraph is obtained.

[0061] Then, according to the collocation abnormality of the paragraph and the confusion degree of the paragraph, the smoothness value of the paragraph can be determined. According to the smoothness value of each paragraph in the document to be detected, the smoothness value of the detected document can be determined.

[0062] As an example, the collocation abnormalities of the same paragraph and the confusion degree of the paragraph may be weighted and summed to obtain the smoothness value of the paragraph.

[0063] As an example, a calculation formula for the smoothness value of a paragraph can be predefined, and the smoothness value of the paragraph can be calculated according to the predefined calculation formula. In this example, the specific calculation method of the compliance value of the paragraph will be given in the subsequent embodiments, and in order to avoid repetition, a detailed description is not given here.

[0064] Further, after the smoothness value of each paragraph in the document to be detected is determined, the smoothness value of the document to be inspected can be determined according to the smoothness value of each paragraph. For example, the smoothness value of each paragraph can be weighted and summed to obtain the smoothness value of the document to be detected. For another example, the average value of the smoothness value of each paragraph can be calculated, and the obtained average value can be determined as the smoothness value of the document to be detected.

[0065] In this embodiment, when calculating the error degree value of the document to be detected, it is possible to detect errors such as linguistic and typos in the document to be detected. Among them, the detectable linguistic errors involve common errors in writing, including incomplete composition, improper collocation, and repetition. In addition, the typos detection focuses on the detection of typos and confusing characters and similar characters. By detecting grammatical errors and typos in the document to be detected, the error degree value of the document to be detected can be determined.

[0066] As an example, an error detection model used for error detection of documents to be detected can be pre-trained. The error detection model can detect the type of error contained in the document to be detected and the score of the corresponding error type, based on the score of the detected error type , You can determine the error degree value of the document to be detected.

[0067] As an example, an error detection model for error detection of documents to be detected can be pre-trained. The error detection model can detect the types of errors contained in the documents to be detected, and set corresponding score values ​​for each error type in advance, and then After the error detection model is used to detect the error type in the document to be detected, the error degree value of the document to be detected can be determined by summing according to the preset corresponding score values ​​of each error type.

[0068] It should be noted that the order of execution of step 102 and step 103 is in no particular order. The two can be executed simultaneously or sequentially. This embodiment only uses step 103 to be executed after step 102 as an example to explain the present application, but not As a limitation of this application.

[0069] Step 104: Generate a fluency value of the document to be detected according to the gracefulness value, the smoothness value and the error degree value of the document to be detected.

[0070] In this embodiment, after obtaining the gracefulness value, smoothness value and error value of the document to be detected, the fluency value of the document to be detected can be determined according to the obtained gracefulness value, smoothness value and error value.

[0071] As an example, according to the degree of influence of the grace, smoothness, and error of language expression on the fluency of language expression, preset weights can be assigned to grace, smoothness, and error in advance. Among them, grace, smoothness and The sum of the weights of the error degree is 1. Furthermore, after obtaining the gracefulness value, smoothness value and error degree value of the document to be detected, weighted sum calculation is performed on the gracefulness value, smoothness value and error degree value of the document to be detected according to the preset weight to obtain the Check the fluency value of the document.

[0072] The method for detecting document fluency in this embodiment obtains the document to be detected and calculates the gracefulness value, smoothness value and error value of the document to be detected, and then according to the gracefulness value, smoothness value and error of the document to be detected The degree value generates the fluency value of the document to be detected, thereby realizing automatic detection of document grace, smoothness, and error degree, thereby realizing the detection of document fluency. Moreover, when applied to composition review, the automatic review of the composition is realized, and the composition can be scored from multiple aspects of grace, smoothness, and error. Compared with the method of scoring the composition based on the teacher manually, it can save manpower And time, improve the efficiency of composition review, and the review standards are unified, which can improve the accuracy of composition scoring.

[0073] The beauty of the sentence can be reflected in terms of wording, sentence segmentation, emotional color, etc. In order to accurately detect the gracefulness value of the document to be recognized, in a possible implementation of the embodiment of the present application, the to-be-recognized can be obtained from different angles. The characteristics of the document are used to calculate the gracefulness value of the document to be recognized. Attached below figure 2 Give details.

[0074] figure 2 It is a schematic flowchart of a method for detecting document fluency according to the second embodiment of the present application. Such as figure 2 As shown in figure 1 Based on the illustrated embodiment, step 102 may include the following steps:

[0075] In step 201, the document to be detected is segmented to form multiple clauses.

[0076] As an example, for the obtained document to be inspected, the document to be inspected can be divided into multiple clauses according to the comma ",", semicolon ";" and other sentence symbols.

[0077] Step 202: Obtain the main structure features of multiple clauses.

[0078] Among them, the main structure feature can be composed of the correlation between the subject, predicate and object in the sentence and part of speech collocation.

[0079] In this embodiment, for each clause obtained by segmentation, the dependency analysis tool in natural language processing can be used to perform dependency syntactic analysis on the clause, and the dependency analysis tool uses the dependency relationship between words in the sentence to express the meaning of the word. Syntactic structure information, such as subject-predicate-object, definite adverbial complement and other structural information, is used to obtain sentence structure information.

[0080] For example, image 3 It is an example diagram of sentence dependency analysis. Such as image 3 As shown, for the sentence "the meeting announced the first batch of senior academicians", after analysis by the dependency analysis tool, it can be determined that the core of the sentence is "announcement", "meeting" and "announcement" constitute a subject-verb relationship, "announcement" It constitutes a verbal-object relationship with the "list", and the relationship between the "first batch" and "academicians", "senior" and "academicians", and "academicians" and "list" are all defined as China.

[0081] For the main structure of the clause, the correlation between subject, predicate and object and part of speech collocation can be calculated to form the main structure feature. Among them, the correlation between the subject, predicate and object can be calculated using the word granular language model n-gram model. Specifically, the n-gram language model is used to calculate the preplexity (ppl) of the subject, predicate and object collocation. Among them, the training corpus of the n-gram language model can be a large amount of collected article corpus. For example, the training corpus can be about 20 million article corpus. Only the subject, predicate, object, and main components are extracted from a large amount of article corpus as the training data pair. n-gram language model for training. Calculating the subject-predicate-object correlation of the jth clause in each clause can be calculated using the following formula (1).

[0082]

[0083] Among them, sim j Represents the subject-predicate-object correlation of the jth clause, ppl max Represents the maximum value of language confusion of all clauses, ppl j Represents the value of language confusion in the jth clause.

[0084] Furthermore, according to the subject-predicate-object correlation and part-of-speech collocation in the j-th clause, the main structure feature of the j-th sentence can be formed.

[0085] Still with image 3 Take the example shown as an example. If the extracted subject, predicate, and object are conference, announcement, and list respectively, the language confusion degree of the "meeting announcement list" is calculated, and the subject, predicate and object correlation between the meeting announcement lists is calculated according to the language confusion degree. The calculated subject-predicate-object correlation, the position of the subject-predicate-object component in the sentence "The conference announced the first batch of senior academicians" and the confidence of each component obtained by the dependency analysis model are determined as the main structure of the sentence feature.

[0086] Using the above process to analyze each clause, you can determine the main structure characteristics of each clause.

[0087] Step 203: Acquire basic sentence features of multiple clauses.

[0088] In this embodiment, for each clause, the part of speech and named entity appearing in the clause can be acquired as the basic sentence features of the clause.

[0089] Step 204: Generate a gracefulness value of the document to be detected according to the main structure features and basic sentence features of the multiple clauses.

[0090] In this embodiment, after acquiring the main structure features and basic sentence features of the multiple clauses, the grace value of the document to be detected can be generated according to the main structure features and basic sentence features of the multiple clauses.

[0091] As a possible implementation, after acquiring the main structure features and basic sentence features of multiple clauses, the main structure features and basic sentence features of the same clause can be spliced ​​together to obtain the spliced ​​features. Then input the spliced ​​features into a pre-trained graceful sentence two classifier to obtain the graceful sentence probability of each clause, and then determine the gracefulness value of the document to be detected according to the graceful sentence probability of each clause. For example, the average probability of graceful sentences of each clause can be calculated as the gracefulness value of the document to be detected.

[0092] In the method for detecting document fluency in this embodiment, multiple clauses are formed by segmenting the document to be detected, and the main structure features and basic sentence features of the multiple clauses are obtained, and then based on the main structure features and basic sentences of the multiple clauses The feature generates the gracefulness value of the document to be detected. Therefore, when determining the gracefulness of the document to be detected, not only the objective indicator of the appropriateness of subject-verb-object collocation is considered, but also the underlying feature of the basic sentence feature is considered. It helps to improve the accuracy of gracefulness detection.

[0093] For a sentence, in addition to the subject, predicate, and object components, it also includes some modifiers, such as attributives, adjectives and other modifiers. The rationality of language modification also has a certain impact on the grace of the sentence. Therefore, in a possible implementation of the embodiment of the present application, the modification dependent characteristics of multiple clauses can also be obtained, and when determining the gracefulness value of the document to be detected, according to the main structure characteristics and basic sentence characteristics of the multiple clauses And modification depends on the feature to generate the gracefulness value of the document to be detected.

[0094] Among them, the modification dependent characteristics of the clause can be determined according to the correlation between the modifiers such as adjectives and adverbs and the modified components. The specific implementation process will be given in the subsequent embodiments and will not be described in detail here.

[0095] When generating the gracefulness value of the document to be tested based on the main structure features, basic sentence features, and modification-dependent features of multiple clauses, the main structure features, basic sentence features, and modification-dependent features of the same clause can be combined first. Splicing, and then input the spliced ​​features into the pre-trained graceful sentence two classifier to obtain the graceful sentence probability of each clause, and then determine the gracefulness value of the document to be detected according to the graceful sentence probability of each clause . For example, the average probability of graceful sentences of each clause can be calculated as the gracefulness value of the document to be detected.

[0096] By obtaining the modification-dependent features of multiple clauses, the gracefulness value of the document to be tested is generated according to the main structure features of multiple clauses, basic sentence features, and modification-dependent features, thus, based on the main structure features of the sentence and the basic sentence features In the above, the objective indicator of the rationality of sentence modification is further considered, which improves the comprehensiveness of language grace detection, and realizes the gracefulness evaluation of the document to be tested from multiple angles.

[0097] The relevance between sentences also affects the gracefulness of sentences. In a possible implementation of the embodiment of this application, on the basis of acquiring the main structure features, basic sentence features, and modification dependency features of multiple clauses, Obtain the clause relevance features of multiple clauses, and then generate the gracefulness value of the document to be detected based on the main structure features, basic sentence features, modification dependency features, and clause relevance features of multiple clauses. Figure 4 Give details.

[0098] Figure 4 Is a schematic flow chart of a method for detecting document fluency according to the third embodiment of the present application, such as Figure 4 As shown in figure 1 Based on the illustrated embodiment, step 102 may include the following steps:

[0099] In step 301, the document to be detected is segmented to form multiple clauses.

[0100] Step 302: Obtain the main structure features of multiple clauses.

[0101] Step 303: Obtain basic sentence features of multiple clauses.

[0102] In this embodiment, the description of step 301 to step 303 can refer to the description of step 201 to step 203 in the foregoing embodiment, which is not repeated here.

[0103] Step 304: Obtain modification dependent characteristics of multiple clauses.

[0104] For the adjectives, adverbs, attributives and other modifiers contained in the clauses, for example, in the example sentence "the conference announced the first batch of senior academicians", "the first batch of senior academicians" is the modifier, and the difference between the modifier and the modifier can be calculated For example, calculating the correlation between the modified component "first batch of senior academicians" in the example sentence and the modified component "list", the calculation method can refer to the calculation method of calculating subject-verb-object correlation in the foregoing embodiment. Specifically, the n-gram language model is used to calculate the language confusion degree ppl between the modifiers in each clause and the modifiers, and then the correlation of the modifiers in each clause is calculated according to the above formula (1), and then according to each modifier Relevance is modified to depend on characteristics.

[0105] In the embodiments of this application, when the clause contains multiple modifiers, for example, both adjectives and adverbs are included, the correlation between each modifier and the modified components can be calculated separately to obtain multiple modifiers related Then calculate the average of the correlations of multiple modifiers as the modification-dependent feature of the clause, or choose the largest modifier-related correlation as the modification-dependent feature of the clause.

[0106] Step 305: Obtain clause characteristics of multiple clauses.

[0107] Step 306: Generate clause correlation features between the multiple clauses according to the clause features of the multiple clauses.

[0108] In this embodiment, for each of the divided clauses, the clause characteristics of each clause can be obtained, where the clause characteristics can be expressed by the correlation between the clauses.

[0109] As a possible implementation, a text similarity model based on Convolutional Neural Networks (CNN) can be used to calculate the correlation between each clause. Among them, the text similarity model based on CNN is based on text CNN (Text CNN) structure, trained with encyclopedia-level corpus data. Use the text similarity model to calculate the correlation between each clause, that is, obtain the clause characteristics of each clause.

[0110] Then, according to the clause features of multiple clauses, clause correlation features between multiple clauses can be generated.

[0111] As an example, the average value of the clause features of all clauses can be calculated, that is, the average value of the relevance between the clauses of all the clauses can be calculated as the clause relevance feature between multiple clauses.

[0112] It should be noted that in this embodiment, the order in which the features of the multiple clauses are acquired is in no particular order, and each feature can be acquired at the same time, or each feature can be acquired one after the other. This embodiment only uses the successive acquisition of each feature as an example for explanation. This application shall not be regarded as a limitation to this application.

[0113] By obtaining the clause features of multiple clauses and then obtaining the clause correlation features between multiple clauses, it provides conditions for considering the relevance between sentences during gracefulness detection, which is beneficial to improve the accuracy of gracefulness detection.

[0114] Step 307: Combine the main structure features, basic sentence features, modification dependent features, and clause correlation features of multiple clauses to generate input features.

[0115] Whether the sentence is beautiful includes not only whether the objective sentence subject-predicate-object collocation is appropriate, whether the sentence modification is reasonable, but also includes subjective feelings such as the beauty of the word language. Therefore, in this embodiment, when acquiring the characteristics of the clause, both the main structural characteristics are acquired The two features describing objective indicators, the characteristics of the basic sentence and the clause correlation feature of the sentence are also obtained. Then, the main structure characteristics, basic sentence characteristics, and modification dependence of the same clause are obtained by splicing. The feature and clause correlation feature are combined to obtain the input feature corresponding to each clause.

[0116] Step 308: Input the input features into the random forest model to generate the gracefulness value of the document to be detected.

[0117] In this embodiment, after the input features are acquired, the input features can be input into the random forest model to generate the gracefulness value of the document to be detected.

[0118] Among them, the training samples of the random forest model can be elementary and middle school composition annotation data, network article annotation data, etc., to indicate whether the sentences in the training samples are graceful sentences, and the standards and labeling results of graceful sentences can be labeled and calibrated by the language teacher. Generate a large number of effectively labeled sentence samples to train the random forest model, and then input the input features corresponding to each clause contained in the document to be detected into the trained random forest model to obtain the gracefulness value of the document to be detected.

[0119] In a possible implementation of the embodiment of this application, the random forest model can also output the probability that the clause corresponding to each input feature is a graceful sentence or a normal sentence, and then count the number of graceful sentences contained in the document to be detected in the total sentence The ratio of the number to get the gracefulness value of the document to be detected.

[0120] Figure 5 It is an example diagram of the gracefulness detection process based on multiple features. Such as Figure 5 As shown, according to the subject, predicate, and object in the sentence, the main structure characteristics are generated, the modification dependent characteristics of the sentence are generated according to the modifiers such as adjectives and adverbs, and the basic sentence characteristics are generated according to the part of speech and named entities appearing in the sentence. Relationship between ( Figure 5 Take Clause 1 and Clause 2 as examples) to generate clause correlation features, and then combine the features and input them into the random forest model to obtain the detection results of graceful sentences or ordinary sentences.

[0121] The method for detecting document fluency in this embodiment obtains the main structure characteristics, basic sentence characteristics, modification dependent characteristics, and clause correlation characteristics of multiple clauses in the document to be detected, and combines the main structure characteristics and basic sentence characteristics of multiple clauses. Features, modification dependent features, and clause correlation features are spliced ​​to generate input features, and then the input features are input to the random forest model to generate the gracefulness value of the document to be detected. Therefore, when evaluating the gracefulness of the document, consider Subject-verb collocation, modifier collocation, relevance between sentences and basic sentence features realize multi-angle evaluation of gracefulness, which helps to improve the accuracy of gracefulness evaluation.

[0122] Image 6 Is a schematic flow chart of a method for detecting document fluency according to the fourth embodiment of the present application, such as Image 6 As shown in figure 1 On the basis of the illustrated embodiment, in step 103, calculating the smoothness value of the document to be detected can be implemented through the following steps:

[0123] Step 401: Obtain each paragraph of the document to be detected.

[0124] Step 402: Calculate the paragraph confusion value of each paragraph.

[0125] In this embodiment, for each paragraph included in the document to be detected, the paragraph confusion value of each paragraph can be calculated.

[0126] As a possible implementation method, when calculating the paragraph confusion value of each paragraph, you can first obtain multiple sentences in each paragraph, and then calculate multiple language confusion values ​​of multiple sentences separately, and then according to the paragraph Multiple language confusion values ​​corresponding to multiple sentences among them generate paragraph confusion values ​​corresponding to paragraphs.

[0127] In specific implementation, a deep neural network (Deep Neural Networks, DNN) language model can be trained using the long sentence corpus of large data of hundreds of millions of documents, and the DNN language model language can calculate the smoothness of long sentences. When the DNN language model language is used to calculate the fluency of long sentences, the DNN language model language is specifically used to calculate the language confusion value of the long sentence. The smaller the language confusion, the smoother the sentence.

[0128] In this embodiment, after the DNN language model is used to obtain the language confusion value of each sentence, the paragraph confusion value of the paragraph can be further obtained.

[0129] As an example, the average value of language confusion values ​​corresponding to each sentence in a paragraph can be calculated, and the obtained average value can be used as the paragraph confusion value of the paragraph.

[0130] Or, as an example, the magnitude of the language confusion value corresponding to each sentence in the paragraph can be compared, and the maximum value of the language confusion value is used as the paragraph confusion value of the paragraph.

[0131] By acquiring multiple sentences in each paragraph, and then calculating multiple language confusion values ​​of multiple sentences separately, and then generating paragraph confusion values ​​corresponding to the paragraph according to multiple language confusion values ​​corresponding to multiple sentences in the paragraph Thus, the detection of the paragraph smoothness corresponding to each paragraph in the document to be detected is realized, which lays the foundation for the detection of the smoothness of the document to be detected.

[0132] Step 403: Calculate the short sentence collocation abnormal value of each paragraph.

[0133] In this embodiment, for each paragraph included in the document to be detected, the short sentence collocation abnormal value in each paragraph can also be calculated.

[0134] In a possible implementation of the embodiment of the present application, when calculating the short sentence collocation abnormal value of each paragraph, the sentence in each paragraph can be obtained first, and then the multiple short sentences in the sentence can be obtained. The collocation value between the sentences, and compare the collocation value between multiple sentences with the preset perplexity threshold. If the collocation value between multiple phrases is greater than the preset perplexity threshold, Then mark the sentence as an abnormal sentence, and then obtain the number of abnormal sentences in each paragraph, and generate short sentences with abnormal values ​​according to the number of abnormal sentences.

[0135] Among them, the preset perplexity threshold can be set according to experience or actual needs, or can be determined through samples. When determining based on samples, a certain number of sentences with poor fluency can be used as samples and labeled by experts to determine the preset perplexity threshold.

[0136] In this embodiment, the collocation degree value between multiple short sentences can be determined by the n-gram language model, and the collocation degree value can be represented by the perplexity value output by the n-gram language model.

[0137] In specific implementation, you can use crawler technology to obtain tens of thousands of composition phrase data as training data, train the n-gram language model, use the n-gram language model to detect each phrase and collocation, and calculate the perplexity value of each phrase and collocation. Indicates the degree of collocation between the phrase and the collocation.

[0138] In this embodiment, after the n-gram language model is used to obtain the collocation value between multiple short sentences, the collocation value between multiple short sentences can be compared with the preset perplexity threshold. When there are more than one sentence When the collocation value between short sentences is greater than the preset perplexity threshold, the sentence is marked as an abnormal sentence, the number of abnormal sentences contained in the paragraph is counted, and the number of abnormal sentences contained in a paragraph is calculated as the number of abnormal sentences in the paragraph The ratio of the total number of sentences to get the short sentence collocation abnormal value of the paragraph.

[0139] By obtaining multiple short sentences contained in the sentence in a paragraph, and calculating the collocation value between multiple short sentences, when the collocation value between multiple short sentences is greater than the preset perplexity threshold, the sentence is marked as an abnormal sentence , And then generate short sentence collocation abnormal values ​​according to the number of abnormal sentences in each paragraph, thereby realizing the detection of abnormal collocation short sentences in the paragraph, by determining the short sentence collocation abnormal value of the paragraph, to detect the document to be detected The smoothness laid the foundation.

[0140] It should be noted that, in this embodiment, the order of execution of step 402 and step 403 is in no particular order. This embodiment only uses step 403 to be executed after step 402 as an example to explain the application, and cannot be used as a limitation to the application.

[0141] Step 404: Generate the fluency value of each paragraph according to the paragraph confusion value and short sentence collocation abnormal value of each paragraph.

[0142] In this embodiment, after the paragraph confusion value and short sentence collocation abnormal value of each paragraph in the document to be detected are determined, the smoothness of each paragraph can be generated according to the paragraph confusion value and short sentence collocation abnormal value of each paragraph value.

[0143] As a possible implementation, the fluent value of the paragraph can be generated by the following formula (2).

[0144]

[0145] Where P i Indicates the smoothness value of the i-th paragraph, ep i Match outliers for the short sentence of the i-th paragraph, ppl max Is the maximum value among multiple language confusion values ​​in the i-th paragraph, ppl i Is the paragraph confusion value of the i-th paragraph.

[0146] By using the preset formula to calculate the pliability value of each paragraph, a unified calculation standard is provided for the calculation of the pliability value, which helps to ensure the accuracy of the pliability value calculation.

[0147] Step 405: Generate a smoothness value of the document to be detected according to the smoothness value of each paragraph.

[0148] In this embodiment, after the smoothness value of each paragraph in the document to be detected is determined, the smoothness value of the document to be detected can be further obtained. For example, the average value of the smoothness value of each paragraph can be calculated, and the obtained average value can be determined as the smoothness value of the document to be detected.

[0149] In the detection of document fluency in this embodiment, the paragraph confusion value of each paragraph in the document to be detected is calculated, and the short sentence collocation abnormal value of each paragraph is calculated, and then the paragraph confusion value and short sentence of each paragraph are calculated. Match the abnormal value to generate the smoothness value of each paragraph, and then generate the smoothness value of the document to be inspected according to the smoothness value of each paragraph, thereby realizing the smoothness inspection of the document to be inspected. When detecting the smoothness, both Considering the fluency of the sentence, and considering the degree of collocation between short sentences in the sentence, so as to improve the accuracy of fluency detection.

[0150] Word errors and language errors in the document also affect the score of the document. Therefore, in a possible implementation of the embodiment of the present application, the error degree value of the document to be detected can also be determined according to the type of error contained in the document to be detected , In order to improve the accuracy of scoring the documents to be detected. Attached below Figure 7 Give details.

[0151] Figure 7 Is a schematic flow chart of a method for detecting document fluency according to the fifth embodiment of the present application, such as Figure 7 As shown in figure 1 On the basis of the illustrated embodiment, in step 103, calculating the error degree value of the document to be detected can be implemented through the following steps:

[0152] Step 501: Obtain the error type of each sentence in the document to be detected, where the error type includes typos and collocation errors.

[0153] Among them, the detection of typos and mistakes can focus on detecting typos, miscellaneous characters, and similar characters in the document to be detected; collocation errors can refer to the linguistic defects contained in the document to be detected, including but not limited to missing components and improper collocation , Repetition, etc.

[0154] As a possible implementation manner, in this embodiment, when obtaining the error type of each sentence in the document to be detected, the error detection model can be used for detection through a pre-trained error detection model. Figure 8 It is an example diagram of the network structure of the error detection model. Such as Figure 8 As shown, the error detection model includes a feature layer, a bidirectional Long Short-Term Memory (LSTM) layer, and a Conditional Random Fields (CRF) layer, where the feature layer is a knowledge-enhanced semantic representation model ( Enhanced Representation from kNowledge IntEgration, ERNIE) and dependency analysis model.

[0155] use Figure 8 The error detection model shown, taking the example sentence "We are all a family" as an example, the example sentence is input into the error detection model, and the embedding features are obtained through the feature layer. Specifically, the ERNIE model is used for word embedding, and the ERNIE model is added I learned the word segmentation entity recognition, learned the expression of words and entities, and used the weight vectors of the last four layers of the ERNIE model for splicing as dynamic semantic features. Use the dependency analysis model to label the example sentences to obtain the characteristics of the grammatical structure of the example sentences. Among them, the dependency analysis model uses the dependency relationship between words in the sentence to express the syntactic structure information of the words, such as the subject-predicate-object, definite adverbial complement, etc. Relations, used to obtain structural information of sentences. Furthermore, the discrete features of the grammatical structure are continuous to obtain the embedding of the word granularity, which is spliced ​​with the word granularity embedding of the semantic characteristic to obtain the embedding characteristic of the sentence. Then, the embedded features of the sentence are input to the bidirectional LSTM layer, and the bidirectional LSTM is used to learn the context. Then, the output of the bidirectional LSTM layer is input to the CRF layer, and the conditional random field is used for error type labeling, and the error type of the sentence is output.

[0156] In this embodiment, use Figure 8 The error detection model shown can obtain the error type of each sentence in the document to be detected.

[0157] Step 502: Generate an error degree value of the document to be detected according to the error type of each sentence.

[0158] In this embodiment, after the error type of each sentence in the document to be detected is obtained, the error degree value of the document to be detected can be generated according to the error type of each sentence.

[0159] As an example, the corresponding score value can be set for each error type in advance, and then, after the error type of each sentence in the document to be detected is obtained, the score value corresponding to each error type set in advance is calculated by summing The method can determine the error degree value of the document to be detected.

[0160] The document fluency detection method of this embodiment obtains the error type of each sentence in the document to be detected, and generates the error degree value of the document to be detected according to the error type of each sentence, thereby achieving the document to be detected The error detection can detect typos and collocation errors in the documents to be detected, which is beneficial to improve the accuracy of the scoring of the documents to be detected.

[0161] Further, in a possible implementation manner of the embodiment of the present application, such as Figure 7 As shown, after obtaining the error type of each sentence in the document to be detected, the following steps may be further included:

[0162] Step 503: Determine whether it belongs to a correctable error type according to the error type.

[0163] Among them, the types of correctable errors can be preset. For example, the types of correctable errors can include typos and related word collocation errors.

[0164] Step 504: If it belongs to the correctable error type, obtain the corresponding correction result and give a prompt.

[0165] In this embodiment, after obtaining the error type of each sentence in the document to be detected, the obtained error type can be compared with the correctable error type to determine whether there is a correctable error type in the document to be detected. When the error type of a sentence belongs to the correctable error type, the corresponding correction result is obtained and prompted to remind the user of the correct word or collocation, thereby helping the user correct the error and improve the writing level.

[0166] In practical applications, you can prepare common error-prone word comparison dictionaries, similar word dictionaries, and fixed collocation dictionaries based on expert knowledge and corpus data obtained by crawlers, and use these dictionary data to generate confusion dictionaries to confuse each word to be corrected in the dictionary The word corresponds to a confusion list, which is a candidate set of the error correction model. For the detected error type and the location of the error, extract the context as a short sentence, extract the whole sentence as a long sentence, replace the words in the wrong position with the candidate result of the confusion dictionary, and use the replacement results of the short sentence and the long sentence respectively The language model calculates the degree of confusion, and determines the candidate error correction result with the smallest degree of confusion as the final error correction result and prompts the user.

[0167] The document fluency detection method of this embodiment obtains and prompts the corresponding correction result when the error type belongs to the correctable error type, thereby helping the user correct the error and improve the user's writing level.

[0168] According to an embodiment of the present application, the present application also provides a device for detecting document fluency.

[0169] It should be noted that in practical applications, the method for detecting document fluency provided in this application can be implemented through a cloud computing platform. The cloud computing platform can deploy beautiful sentence two classifiers, DNN language models, n-gram models and other models, and process the documents to be detected through the cloud computing platform to obtain the fluency value of the documents to be detected.

[0170] Picture 9 It is a schematic structural diagram of a document fluency detection device according to the sixth embodiment of the present application. Such as Picture 9 As shown, the device 60 for detecting document fluency includes: an acquisition module 610, a first calculation module 620, a second calculation module 630, a third calculation module 640, and a generation module 650.

[0171] Wherein, the obtaining module 610 is used to obtain the document to be detected.

[0172] The first calculation module 620 is configured to calculate the gracefulness value of the document to be detected.

[0173] The second calculation module 630 is configured to calculate the smoothness value of the document to be detected.

[0174] The third calculation module 640 is configured to calculate the error degree value of the document to be detected.

[0175] The generating module 650 is configured to generate the fluency value of the document to be detected according to the grace value of the document to be detected, the smoothness value and the error degree value.

[0176] In a possible implementation manner of the embodiment of this application, such as Picture 10 As shown in Picture 9 On the basis of the illustrated embodiment, the first calculation module 620 includes:

[0177] The segmentation unit 621 is configured to segment the document to be detected to form multiple clauses;

[0178] The first acquiring unit 622 is configured to acquire the main structural features of the multiple clauses;

[0179] The second obtaining unit 623 is configured to obtain basic sentence features of the multiple clauses; and

[0180] The first generating unit 624 is configured to generate the gracefulness value of the document to be detected according to the main structure features and basic sentence features of the multiple clauses.

[0181] Further, in a possible implementation manner of the embodiment of the present application, such as Picture 11 As shown in Picture 10 On the basis of the illustrated embodiment, the first calculation module 620 further includes:

[0182] The third acquiring unit 625 is configured to acquire the modification dependent characteristics of the multiple clauses;

[0183] In this embodiment, the first generating unit 624 is further configured to generate the gracefulness value of the document to be detected according to the main structure feature of the multiple clauses, the basic sentence feature, and the modification dependent feature.

[0184] In a possible implementation manner of the embodiment of this application, such as Picture 11 As shown, the first calculation module 620 further includes:

[0185] The fourth obtaining unit 626 is configured to obtain the clause features of the multiple clauses;

[0186] The determining unit 627 is configured to generate clause correlation features between the multiple clauses according to the clause features of the multiple clauses;

[0187] In this embodiment, the first generating unit 624 is further configured to generate the document to be detected according to the main structure features of the multiple clauses, the basic sentence features, the modification dependent features, and the clause correlation features. Gracefulness value.

[0188] As a possible implementation manner, the first generating unit 624 combines the main structure features of the multiple clauses, the basic sentence features, the modification dependent features, and the clause correlation features to generate input features. And input the input feature to the random forest model to generate the gracefulness value of the document to be detected.

[0189] In a possible implementation manner of the embodiment of this application, such as Picture 12 As shown in Picture 9 Based on the illustrated embodiment, the second calculation module 630 includes:

[0190] The fifth acquiring unit 631 is configured to acquire each paragraph of the document to be detected.

[0191] The first calculation unit 632 is configured to calculate the paragraph confusion value of each paragraph.

[0192] As a possible implementation manner, the first calculation unit 632 obtains multiple sentences in each paragraph, calculates multiple language confusion values ​​of the multiple sentences, and calculates multiple language confusion values ​​according to the Multiple language confusion values ​​corresponding to multiple sentences generate paragraph confusion values ​​corresponding to the paragraphs.

[0193] The second calculation unit 633 is used to calculate the short sentence collocation abnormal value of each paragraph.

[0194] As a possible implementation manner, the second calculation unit 633 obtains the sentence in each paragraph, obtains multiple short sentences in the sentence, and calculates the collocation degree value between the multiple short sentences , If the collocation value between the multiple short sentences is greater than the preset perplexity threshold, mark the sentence as an abnormal sentence, and obtain the number of abnormal sentences in each paragraph, according to the abnormal sentence The number of generated short sentence collocation abnormal value.

[0195] The second generating unit 634 is configured to generate the smoothness value of each paragraph according to the paragraph confusion value of each paragraph and the short sentence collocation abnormal value.

[0196] As a possible implementation manner, the second generating unit generates the smoothness value of the paragraph through the following formula:

[0197]

[0198] Among them, ep i Matching outliers for the phrase in the i-th paragraph, ppl max Is the maximum value among the plurality of language confusion values ​​in the i-th paragraph, ppl i Is the paragraph confusion value of the i-th paragraph.

[0199] The third generating unit 635 is configured to generate the smoothness value of the document to be detected according to the smoothness value of each paragraph.

[0200] In a possible implementation manner of the embodiment of this application, such as Figure 13 As shown in Picture 9 On the basis of the illustrated embodiment, the third calculation module 640 includes:

[0201] The sixth acquiring unit 641 is configured to acquire the error type of each sentence in the document to be detected, where the error type includes typos and collocation errors.

[0202] The fourth generating unit 642 is configured to generate an error degree value of the document to be detected according to the error type of each sentence.

[0203] Further, as Figure 13 As shown, the device 60 for detecting document fluency further includes:

[0204] The judging module 660 is configured to judge whether it belongs to a correctable error type according to the error type.

[0205] The correction module 670 is configured to obtain a corresponding correction result and prompt when the error type belongs to the correctable error type.

[0206] It should be noted that the foregoing explanation of the embodiment of the document fluency detection method is also applicable to the document fluency detection device of this embodiment, and the implementation principle is similar, and will not be repeated here.

[0207] The document fluency detection device of the embodiment of the present application obtains the document to be detected and calculates the gracefulness value, smoothness value and error value of the document to be detected, and then according to the gracefulness value, smoothness value and The error degree value generates the fluency value of the document to be detected, thereby realizing automatic detection of document grace, smoothness and error degree, thereby realizing the detection of document fluency. Moreover, when applied to composition review, the automatic review of the composition is realized, and the composition can be scored from multiple aspects of grace, smoothness, and error. Compared with the method of scoring the composition based on the teacher manually, it can save manpower And time, improve the efficiency of composition review, and the review standards are unified, which can improve the accuracy of composition scoring.

[0208] According to the embodiments of the present application, the present application also provides an electronic device and a readable storage medium.

[0209] Such as Figure 14 What is shown is a block diagram of an electronic device used to implement the method for detecting document fluency in an embodiment of the present application. Electronic devices are intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples, and are not intended to limit the implementation of the application described and/or required herein.

[0210] Such as Figure 14 As shown, the electronic device includes one or more processors 701, a memory 702, and interfaces for connecting various components, including a high-speed interface and a low-speed interface. The various components are connected to each other by using different buses, and can be installed on a common motherboard or installed in other ways as required. The processor may process instructions executed in the electronic device, including instructions stored in or on the memory to display graphical information of the GUI on an external input/output device (such as a display device coupled to an interface). In other embodiments, if necessary, multiple processors and/or multiple buses can be used with multiple memories and multiple memories. Similarly, multiple electronic devices can be connected, and each device provides part of the necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Figure 14 Take a processor 701 as an example.

[0211] The memory 702 is a non-transitory computer-readable storage medium provided by this application. Wherein, the memory stores instructions executable by at least one processor, so that the at least one processor executes the method for detecting document fluency provided in this application. The non-transitory computer-readable storage medium of the present application stores computer instructions, and the computer instructions are used to make a computer execute the method for detecting document fluency provided in the present application.

[0212] As a non-transitory computer-readable storage medium, the memory 702 can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions/modules corresponding to the document fluency detection method in the embodiment of the present application ( For example, attach Picture 9 The illustrated acquisition module 610, first calculation module 620, second calculation module 630, third calculation module 640, and generation module 650). The processor 701 executes various functional applications and data processing of the server by running non-transitory software programs, instructions, and modules stored in the memory 702, that is, realizing the document fluency detection method in the foregoing method embodiment.

[0213] The memory 702 may include a storage program area and a storage data area, where the storage program area can store an operating system and an application program required by at least one function; the storage data area can store usage information of the electronic device according to the method for detecting document fluency. Created data, etc. In addition, the memory 702 may include a high-speed random access memory, and may also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state storage devices. In some embodiments, the memory 702 may optionally include a memory remotely provided with respect to the processor 701, and these remote memories may be connected to an electronic device that executes the method for detecting document fluency through a network. Examples of the aforementioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0214] The electronic device that executes the method for detecting document fluency may further include: an input device 703 and an output device 704. The processor 701, the memory 702, the input device 703, and the output device 704 may be connected by a bus or other means, Figure 14 Take the bus connection as an example.

[0215] The input device 703 can receive input digital or character information, and generate key signal input related to the user settings and function control of the electronic device that executes the document fluency detection method, such as touch screen, keypad, mouse, track pad, touch pad , Pointing stick, one or more mouse buttons, trackball, joystick and other input devices. The output device 704 may include a display device, an auxiliary lighting device (for example, LED), a tactile feedback device (for example, a vibration motor), and the like. The display device may include, but is not limited to, a liquid crystal display (LCD), a light emitting diode (LED) display, and a plasma display. In some embodiments, the display device may be a touch screen.

[0216] Various implementations of the systems and technologies described herein can be implemented in digital electronic circuit systems, integrated circuit systems, application specific ASICs (application specific integrated circuits), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include: being implemented in one or more computer programs, the one or more computer programs may be executed and/or interpreted on a programmable system including at least one programmable processor, the programmable processor It can be a dedicated or general-purpose programmable processor, which can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit the data and instructions to the storage system, the at least one input device, and the at least one output device. An output device.

[0217] These calculation programs (also called programs, software, software applications, or codes) include machine instructions of a programmable processor, and can be implemented using high-level procedures and/or object-oriented programming languages, and/or assembly/machine language Calculation program. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and/or device used to provide machine instructions and/or data to a programmable processor ( For example, magnetic disks, optical disks, memory, programmable logic devices (PLD)), including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal used to provide machine instructions and/or data to a programmable processor.

[0218] In order to provide interaction with the user, the systems and techniques described here can be implemented on a computer that has: a display device for displaying information to the user (for example, a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) ); and a keyboard and a pointing device (for example, a mouse or a trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (for example, visual feedback, auditory feedback, or tactile feedback); and can be in any form (including Acoustic input, voice input, or tactile input) to receive input from the user.

[0219] The systems and technologies described herein can be implemented in a computing system that includes back-end components (for example, as a data server), or a computing system that includes middleware components (for example, an application server), or a computing system that includes front-end components (for example, A user computer with a graphical user interface or a web browser, through which the user can interact with the implementation of the system and technology described herein), or includes such background components, middleware components, Or any combination of front-end components in a computing system. The components of the system can be connected to each other through any form or medium of digital data communication (for example, a communication network). Examples of communication networks include: local area network (LAN), wide area network (WAN), and the Internet.

[0220] The computer system can include clients and servers. The client and server are generally far away from each other and usually interact through a communication network. The relationship between the client and the server is generated through computer programs running on the corresponding computers and having a client-server relationship with each other.

[0221] According to the technical solution of the embodiment of the present application, by calculating the gracefulness value, smoothness value and error degree value of the document to be detected, and then according to the calculated gracefulness value, smoothness value and error degree value, the smoothness of the document to be detected is generated Therefore, the automatic detection of document grace, smoothness, and error is realized, thereby realizing the detection of document fluency; and, when applied to composition review, automatic review of composition is realized, which can be based on grace, The scoring of essays on multiple aspects of smoothness and error can save manpower and time, improve the efficiency of composition review, and improve the accuracy of composition scoring compared with the way that teachers manually grade the composition.

[0222] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the present application can be performed in parallel, sequentially, or in a different order, as long as the desired result of the technical solution disclosed in the present application can be achieved, this is not limited herein.

[0223] The foregoing specific implementations do not constitute a limitation on the protection scope of the present application. Those skilled in the art should understand that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any amendments, equivalent substitutions and improvements made within the spirit and principles of this application shall be included in the protection scope of this application.