English pronunciation real-time error correction method based on deep neural network driving

Abstract

The invention discloses an English pronunciation real-time error correction method based on deep neural network driving, and relates to the technical field of English pronunciation, and the method comprises the steps: obtaining a past sample manuscript of a user and a sample pronunciation of reading the sample manuscript; collecting to-be-classified sample manuscripts, preprocessing the sample manuscripts, performing feature extraction on text resources, and converting text information into a vocabulary for representation; obtaining a mapping relation between the voice and the vocabulary; obtaining a semantic use habit of a user; the actual intention of the user is obtained; according to the actual intention of the user, a correct voice text is generated, and error correction is performed on the real-time dialogue voice based on the correct voice text. Through feature extraction of the sample manuscript, analysis of the sample pronunciation, obtaining of the semantic use habit of the user and obtaining of the actual intention of the user, the problem that pronunciation errors of each student for the same word are different can be solved, and the accuracy of error correction is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of English pronunciation, and in particular to a real-time error correction method for English pronunciation driven by a deep neural network. Background Art

[0002] For non-native English learners, the standard and correct pronunciation of English words is the basis for learning English well, because English is a phonetic language, and the biggest feature of phonetic languages is phonetics. How an English word should be written almost entirely depends on the pronunciation of the word, and correct English pronunciation is the guarantee for normal English communication with others. Therefore, the pronunciation teaching of English words is crucial. In current English teaching, students' English pronunciation is taught by teachers in class. Due to limited teaching staff, some universities even have only 3 to 4 English teachers in a department or college. One teacher usually has to teach dozens of students and cannot correct students' English word pronunciation one by one.

[0003] Pronunciation correction for word or text reading can be identified and corrected by comparing with standard pronunciation. However, for independent English conversation, it is difficult to identify and correct errors because each student has different pronunciation errors for the same word. At the same time, students may have errors in grammar and word usage, which will further increase the difficulty of identification. Summary of the Invention

[0004] In order to solve the above technical problems, a real-time English pronunciation error correction method driven by a deep neural network is provided. This technical solution solves the problems raised in the above background technology.

[0005] In order to achieve the above objects, the technical solution adopted by the present invention is:

[0006] A real-time English pronunciation correction method driven by a deep neural network, including:

[0007] Obtain the user's past sample manuscripts and sample pronunciations of the sample manuscripts, where the sample manuscripts contain all the vocabulary the user has mastered;

[0008] Collect sample documents to be classified, preprocess them, extract features from text resources, and convert text information into a vocabulary for representation;

[0009] Analyze the sample pronunciation and obtain the mapping relationship between speech and vocabulary based on the analysis results;

[0010] Acquire the user's semantic usage habits, including the user's incorrect usage habits;

[0011] Obtain the user's real-time conversation voice and obtain the user's actual intention based on semantic usage habits;

[0012] According to the actual intention of the user, the correct speech text is generated, and based on the correct speech text, the real-time conversation speech is corrected.

[0013] Preferably, the steps of collecting sample documents to be classified, preprocessing them, extracting features from text resources, and converting text information into a vocabulary for representation include the following steps:

[0014] Removing irrelevant characters from the sample document, wherein the irrelevant characters include special symbols and spaces;

[0015] Based on the dictionary-based word segmentation method, the English text of the sample manuscript is segmented into at least one independent word;

[0016] Traverse all sample documents, collect all independent words that appear, build a vocabulary, and use the independent words in the vocabulary as feature words;

[0017] The frequency of occurrence of characteristic words in all sample manuscripts is counted as characteristic frequency, and the characteristic frequency is paired with the characteristic words.

[0018] Preferably, the step of analyzing the sample pronunciation and obtaining a mapping relationship between speech and vocabulary based on the analysis result comprises the following steps:

[0019] Based on historical data, the correctly pronounced speech is compared with the standard speech to obtain at least one sample similarity, and the minimum value of the sample similarity is used as the preset value;

[0020] The sample pronunciation is recognized using the standard speech of the characteristic vocabulary, and the portion of the sample pronunciation with a similarity higher than a preset value is intercepted as the characteristic pronunciation, and the characteristic vocabulary is paired with the characteristic pronunciation;

[0021] At least one unrecognized part of the sample pronunciation is respectively regarded as an unrecognized speech segment;

[0022] Establishing a mapping relationship between the characteristic vocabulary of the sample document and the characteristic pronunciation of the corresponding sample pronunciation, where the characteristic vocabulary and the characteristic pronunciation established in the mapping relationship are paired, and the relative position of the characteristic vocabulary in the sample document is consistent with the relative position of the characteristic pronunciation in the sample pronunciation;

[0023] In the sample document, at least one portion for which no mapping relationship is established is used as a characteristic portion;

[0024] Acquire characteristic words at both ends of the characteristic part in the sample document as target characteristic words;

[0025] Acquire a feature pronunciation that has a mapping relationship with the target feature vocabulary from the sample pronunciation as the target feature pronunciation;

[0026] The part of the sample pronunciation sandwiched between the two target feature pronunciations is regarded as an unrecognized speech segment and corresponds to the feature part;

[0027] When the feature part contains only one feature word, the corresponding unrecognized speech segment is used as the feature speech to establish a mapping relationship with the feature word contained in the feature part;

[0028] When the characteristic part contains multiple characteristic words, the lower limit of the user's pronunciation interval is obtained, and at least one discontinuity point of the unrecognized speech segment is identified based on the lower limit of the pronunciation interval. Based on the discontinuity point, the unrecognized speech segment is divided into at least one characteristic speech, and a mapping relationship is established between the characteristic speech in the unrecognized speech segment and the characteristic words in the characteristic part in sequence.

[0029] Preferably, the obtaining of the user's semantic usage habits comprises the following steps:

[0030] Obtain the user's previously revised English composition text;

[0031] Using a dependency syntax analysis tool to perform dependency syntax analysis on the preprocessed English composition text to reveal dependency relationships between words, including subject-verb relationships, verb-object relationships, and attributive-predicate relationships;

[0032] The output of dependency syntax analysis is a tree structure representation, where each node represents a word and each edge represents the dependency relationship between words;

[0033] In the dependency parsing tree, identify noun phrases and nouns as entity candidates, and confirm the accuracy of the entities based on part-of-speech tags and context information;

[0034] Store the extracted entities in the entity library;

[0035] In the dependency parsing tree, verb phrases and verbs are identified as relation candidates, and the subject and object of the verb are analyzed to determine the semantic relationship between entities;

[0036] Verify and correct the extracted relationships based on contextual information and domain semantics;

[0037] Attribute extraction is achieved by analyzing the attributives and modifiers in the dependency syntax tree;

[0038] Create an empty semantic network, define the network structure and the attributes of nodes and edges, set the attribute fields of entities and relationships, the attribute fields of entities include name, type and description, and the attribute fields of relationships include relationship type and confidence;

[0039] Add the extracted entities and relationships to the semantic network, create nodes for each entity, set the node attributes, create edges for each pair of entities with relationships, and set the edge attributes;

[0040] The semantic network is presented as a graphical interface, with nodes represented as circles and edges as curves, and different types of entities and relationships are distinguished by different colors, shapes and sizes;

[0041] Take the semantic network as the user's semantic usage habits.

[0042] Preferably, obtaining the user's actual intention based on semantic usage habits includes the following steps:

[0043] Using the characteristic voice, the real-time conversation voice is recognized, and the real-time conversation voice is recognized in chronological order to obtain at least one target characteristic voice;

[0044] Based on the mapping relationship between the characteristic speech and the characteristic vocabulary, the target characteristic speech is replaced with the characteristic vocabulary that has a mapping relationship with the target characteristic speech to obtain the user's real-time pronunciation text;

[0045] Use semantic networks to infer the intention of real-time pronunciation text and obtain the user's actual intention.

[0046] Preferably, the method of using a semantic network to infer the intention of the real-time pronunciation text to obtain the actual intention of the user comprises the following steps:

[0047] The real-time pronunciation text is segmented according to vocabulary to obtain at least one actual word, actual edges are connected between adjacent actual words, two entities consistent with the adjacent actual words are obtained in the semantic network as feature entities, and attributes of the edge between the two feature entities are assigned to the actual edge;

[0048] The actual edge with conflicting attributes is used as the target actual edge, and the actual word connected by the target actual edge is used as the target actual word;

[0049] Form at least one replacement for the target actual word using entities in the semantic network;

[0050] Segmenting the real-time pronunciation text after applying the replacement scheme to obtain at least one corrected word, connecting correction edges between adjacent corrected words, and obtaining two entities consistent with the adjacent corrected words in the semantic network as corrected entities. If no edge exists between the two corrected entities, the replacement scheme is deleted; otherwise, the attributes of the edge between the two corrected entities are assigned to the corrected edge.

[0051] The number of correction edges with conflicting attributes is counted as the check value of the replacement plan, and the replacement plan with the smallest check value is regarded as the actual intention of the user.

[0052] Preferably, the generation of the correct voice text according to the actual intention of the user comprises the following steps:

[0053] Use the replacement solution with the smallest checksum value to replace the real-time pronunciation text to obtain the correct voice text.

[0054] Preferably, the error correction of the real-time conversation speech based on the correct speech text comprises the following steps:

[0055] Obtain the standard pronunciation of each word in the correct pronunciation text as the reference pronunciation, and number the reference pronunciations in sequence;

[0056] Perform interval recognition on the user's real-time conversation voice. When the pronunciation interval is greater than the lower limit of the pronunciation interval, the pronunciation interval is used as the cutting point.

[0057] Using a cutting point to segment the real-time conversation speech into at least one actual word speech, and numbering the actual word speech in sequence;

[0058] The actual word voice with the same number is compared with the reference voice. When the similarity between the actual word voice and the reference voice is less than a preset value, the reference voice is used to correct the actual word voice.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] By extracting features from sample documents, analyzing sample pronunciations, obtaining the user's semantic usage habits and the user's actual intentions, the system can identify the text corresponding to the pronunciation based on the pre-set user's usage habits and pronunciation habits, and infer the user's true intentions based on the context of the conversation, thereby forming a correct text and removing the user's writing errors. By comparing the correct text with the standard voice, real-time error correction of English pronunciation can be completed. This solution can solve the problem that each student has different pronunciation errors for the same word. At the same time, it can also correct possible errors in students' grammar and word usage to ensure the accuracy of error correction. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 Schematic diagram of the flow of the real-time English pronunciation error correction method driven by a deep neural network of the present invention;

[0062] Figure 2 A flowchart of the present invention for collecting sample documents to be classified, preprocessing them, extracting features from text resources, and converting text information into a vocabulary for representation;

[0063] Figure 3A flowchart of analyzing a sample pronunciation and obtaining a mapping relationship between speech and vocabulary based on the analysis results of the present invention;

[0064] Figure 4 A schematic diagram of the process of obtaining the semantic usage habits of a user according to the present invention;

[0065] Figure 5 A schematic diagram of a process for obtaining the user's actual intention based on semantic usage habits of the present invention;

[0066] Figure 6 A flow chart of the present invention for using a semantic network to infer the intention of a real-time pronunciation text to obtain the actual intention of the user;

[0067] Figure 7 The present invention is a flowchart of correcting errors in real-time conversation speech based on correct speech text. DETAILED DESCRIPTION

[0068] The following description is intended to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are merely examples, and those skilled in the art may conceive of other obvious variations.

[0069] Reference Figure 1 As shown in FIG, the real-time English pronunciation error correction method driven by a deep neural network includes:

[0070] Obtain the user's past sample manuscripts and sample pronunciations of the sample manuscripts, where the sample manuscripts contain all the vocabulary the user has mastered;

[0071] Collect sample documents to be classified, preprocess them, extract features from text resources, and convert text information into a vocabulary for representation;

[0072] Analyze the sample pronunciation and obtain the mapping relationship between speech and vocabulary based on the analysis results;

[0073] Acquire the user's semantic usage habits, including the user's incorrect usage habits;

[0074] Obtain the user's real-time conversation voice and obtain the user's actual intention based on semantic usage habits;

[0075] According to the actual intention of the user, the correct speech text is generated, and based on the correct speech text, the real-time conversation speech is corrected.

[0076] When users are having a conversation in English, there may be errors in their writing. At the same time, when there are errors, pronunciation errors may also occur. Since everyone's pronunciation errors are different, it is impossible to accurately identify the problems by directly comparing them with the standard voice. Therefore, the correction is not targeted. It may be that they want to express a certain word, but when correcting the error, they use the correct pronunciation of another word to correct it. Therefore, the correction is not targeted and the effect needs to be improved. In order to accurately correct the error, it is necessary to obtain the actual meaning expressed by the user, and then correct the error according to the standard voice of the actual expression. In this way, it can be more targeted. In this solution, corresponding steps are set to achieve this effect.

[0077] Reference Figure 2 As shown in FIG, collecting sample documents to be classified, preprocessing them, extracting features from text resources, and converting text information into a vocabulary for representation include the following steps:

[0078] Removing irrelevant characters from the sample document, wherein the irrelevant characters include special symbols and spaces;

[0079] Based on the dictionary-based word segmentation method, the English text of the sample manuscript is segmented into at least one independent word;

[0080] Traverse all sample documents, collect all independent words that appear, build a vocabulary, and use the independent words in the vocabulary as feature words;

[0081] The frequency of occurrence of characteristic words in all sample manuscripts is counted as characteristic frequency, and the characteristic frequency is paired with the characteristic words.

[0082] The usage of characteristic frequency is used as a verification in the mapping relationship between speech and vocabulary. Because there may be a certain deviation in the correspondence between characteristic speech and characteristic vocabulary, further verification is required to ensure the accuracy of the mapping. Since the sample pronunciation of the sample manuscript and the sample reading manuscript are in a corresponding relationship, the characteristic speech in the sample pronunciation and the characteristic vocabulary in the sample manuscript are in a corresponding relationship, and the frequency of occurrence in the sample pronunciation and the sample manuscript is consistent. Therefore, verification can be performed based on this.

[0083] There may be incorrect pronunciations in the sample pronunciation, but users are accustomed to the incorrect pronunciation of the same words. Therefore, the pronunciation errors are the same each time. Therefore, by establishing a mapping relationship between feature speech and feature vocabulary, the actual intention of the real-time conversation speech can be inferred. Because, even if there are incorrect pronunciations in the real-time conversation speech, according to the mapping relationship, it can be determined which word it wants to express, and then the pronunciation can be corrected according to the determined text.

[0084] Reference Figure 3 As shown, analyzing the sample pronunciation and obtaining the mapping relationship between speech and vocabulary based on the analysis results includes the following steps:

[0085] Based on historical data, the correctly pronounced speech is compared with the standard speech to obtain at least one sample similarity, and the minimum value of the sample similarity is used as the preset value;

[0086] The sample pronunciation is recognized using the standard speech of the characteristic vocabulary, and the portion of the sample pronunciation with a similarity higher than a preset value is intercepted as the characteristic pronunciation, and the characteristic vocabulary is paired with the characteristic pronunciation;

[0087] At least one unrecognized part of the sample pronunciation is respectively regarded as an unrecognized speech segment;

[0088] Establishing a mapping relationship between the characteristic words of the sample document and the characteristic pronunciations of the corresponding sample pronunciations, where the characteristic words and the characteristic pronunciations established in the mapping relationship are paired, and the relative positions of the characteristic words in the sample document are consistent with the relative positions of the characteristic pronunciations in the sample pronunciations;

[0089] In the sample document, at least one portion for which no mapping relationship is established is used as a characteristic portion;

[0090] Acquire characteristic words at both ends of the characteristic part in the sample document as target characteristic words;

[0091] Acquire a feature pronunciation that has a mapping relationship with the target feature vocabulary from the sample pronunciation as the target feature pronunciation;

[0092] The part of the sample pronunciation sandwiched between the two target feature pronunciations is regarded as an unrecognized speech segment and corresponds to the feature part;

[0093] When the feature part contains only one feature word, the corresponding unrecognized speech segment is used as the feature speech to establish a mapping relationship with the feature word contained in the feature part;

[0094] When the characteristic part contains multiple characteristic words, the lower limit of the user's pronunciation interval is obtained, and at least one discontinuity point of the unrecognized speech segment is identified based on the lower limit of the pronunciation interval. Based on the discontinuity point, the unrecognized speech segment is divided into at least one characteristic speech, and a mapping relationship is established between the characteristic speech in the unrecognized speech segment and the characteristic words in the characteristic part in sequence.

[0095] Here, there are multiple unrecognized speech segments, each of which is a continuous pronunciation of the user, that is, the pronunciation of several words spoken continuously, and the unrecognized speech segment does not contain a characteristic pronunciation;

[0096] There are multiple feature parts, each of which is a continuous text without a mapping relationship, that is, each word in the feature part has no mapping relationship;

[0097] Since the unrecognized speech segment and the characteristic part are formed in the same way, there is a corresponding relationship between them. That is, the unrecognized speech segment is the speech of one of the characteristic parts, which is of course the speech with mispronunciation. In the future, we only need to establish a corresponding relationship between the two.

[0098] When determining the mapping relationship, it is necessary to determine which word corresponds to each pronunciation of the user. Since the user may have pronunciation errors, it is impossible to determine it directly at one time. First, by determining the words with correct pronunciation, the vocabulary and correct pronunciation can be marked in the sample document and sample pronunciation. Then, according to the position of the correct mapping relationship in the sample document and sample pronunciation, for the remaining unrecognizable speech, the correspondence between the unrecognized speech segment and the characteristic part is determined based on the relative position relationship of the recognized speech. The unrecognized speech segment and the characteristic part may be multiple words, so it is necessary to segment the vocabulary according to the interval of pronunciation, and then determine the pronunciation of the user for each vocabulary.

[0099] It should be noted that users can only pronounce the words they know and will not pronounce unknown words. Therefore, when sampling their pronunciation, it is only necessary to ensure that the sample document contains all the words the user knows.

[0100] Reference Figure 4 As shown, obtaining the user's semantic usage habits includes the following steps:

[0101] Obtain the user's previously revised English composition text;

[0102] Using a dependency syntax analysis tool to perform dependency syntax analysis on the preprocessed English composition text to reveal dependency relationships between words, including subject-verb relationships, verb-object relationships, and attributive-predicate relationships;

[0103] The output of dependency syntax analysis is a tree structure representation, where each node represents a word and each edge represents the dependency relationship between words;

[0104] In the dependency parsing tree, identify noun phrases and nouns as entity candidates, and confirm the accuracy of the entities based on part-of-speech tags and context information;

[0105] Store the extracted entities in the entity library;

[0106] In the dependency parsing tree, verb phrases and verbs are identified as relation candidates, and the subject and object of the verb are analyzed to determine the semantic relationship between entities;

[0107] Verify and correct the extracted relationships based on contextual information and domain semantics;

[0108] Attribute extraction is achieved by analyzing the attributives and modifiers in the dependency syntax tree;

[0109] Create an empty semantic network, define the network structure and the attributes of nodes and edges, set the attribute fields of entities and relationships, the attribute fields of entities include name, type and description, and the attribute fields of relationships include relationship type and confidence;

[0110] Add the extracted entities and relationships to the semantic network, create nodes for each entity, set the node attributes, create edges for each pair of entities with relationships, and set the edge attributes;

[0111] The semantic network is presented as a graphical interface, with nodes represented as circles and edges as curves, and different types of entities and relationships are distinguished by different colors, shapes and sizes;

[0112] Take the semantic network as the user's semantic usage habits.

[0113] Usage habits are mainly used to set the user's habitual expressions, so that the text corresponding to the real-time conversation voice can be modified to obtain the user's actual intention.

[0114] Reference Figure 5 As shown, based on semantic usage habits, obtaining the user's actual intention includes the following steps:

[0115] Using the characteristic voice, the real-time conversation voice is recognized, and the real-time conversation voice is recognized in chronological order to obtain at least one target characteristic voice;

[0116] Based on the mapping relationship between the characteristic speech and the characteristic vocabulary, the target characteristic speech is replaced with the characteristic vocabulary that has a mapping relationship with the target characteristic speech to obtain the user's real-time pronunciation text;

[0117] Use semantic networks to infer the intention of real-time pronunciation text and obtain the user's actual intention.

[0118] Reference Figure 6 As shown, using a semantic network to infer the intention of the real-time pronunciation text to obtain the user's actual intention includes the following steps:

[0119] The real-time pronunciation text is segmented according to vocabulary to obtain at least one actual word, actual edges are connected between adjacent actual words, two entities consistent with the adjacent actual words are obtained in the semantic network as feature entities, and attributes of the edge between the two feature entities are assigned to the actual edge;

[0120] The actual edge with conflicting attributes is used as the target actual edge, and the actual word connected by the target actual edge is used as the target actual word;

[0121] Form at least one replacement for the target actual word using entities in the semantic network;

[0122] Segmenting the real-time pronunciation text after applying the replacement scheme to obtain at least one corrected word, connecting correction edges between adjacent corrected words, and obtaining two entities consistent with the adjacent corrected words in the semantic network as corrected entities. If no edge exists between the two corrected entities, the replacement scheme is deleted; otherwise, the attributes of the edge between the two corrected entities are assigned to the corrected edge.

[0123] The number of correction edges with conflicting attributes is counted as the check value of the replacement plan, and the replacement plan with the smallest check value is regarded as the actual intention of the user.

[0124] Since adjacent words must be connected, the modified edge must be assigned attributes. If it is not assigned attributes, it means that there is a problem with the replacement plan and it needs to be abandoned.

[0125] When making a replacement, the replacement scheme is mainly screened through conflicts, and the conflict expressing the correct replacement scheme must be the least.

[0126] Generating the correct speech text according to the user's actual intention includes the following steps:

[0127] Use the replacement solution with the smallest checksum value to replace the real-time pronunciation text to obtain the correct voice text.

[0128] Reference Figure 7 As shown, based on the correct speech text, the error correction of the real-time conversation speech includes the following steps:

[0129] Obtain the standard pronunciation of each word in the correct pronunciation text as the reference pronunciation, and number the reference pronunciations in sequence;

[0130] Perform interval recognition on the user's real-time conversation voice. When the pronunciation interval is greater than the lower limit of the pronunciation interval, the pronunciation interval is used as the cutting point.

[0131] Using a cutting point to segment the real-time conversation speech into at least one actual word speech, and numbering the actual word speech in sequence;

[0132] The actual word voice with the same number is compared with the reference voice. When the similarity between the actual word voice and the reference voice is less than a preset value, the reference voice is used to correct the actual word voice.

[0133] Furthermore, the present solution also proposes a storage medium on which a computer-readable program is stored. When the computer-readable program is called, it executes the above-mentioned real-time English pronunciation error correction method driven by a deep neural network.

[0134] It is understandable that the storage medium may be a magnetic medium, such as a floppy disk, a hard disk, or a magnetic tape; an optical medium, such as a DVD; or a semiconductor medium, such as a solid state disk (SSD).

[0135] In summary, the advantages of the present invention are: feature extraction through sample documents, analysis of sample pronunciation, acquisition of the user's semantic usage habits and the user's actual intentions, so that the text corresponding to the pronunciation can be identified based on the pre-set user's usage habits and pronunciation habits, and the user's true intentions can be guessed based on the context of the conversation, thereby forming a correct text and removing the user's writing errors, so that real-time error correction of English pronunciation can be completed by comparing the correct text with the standard voice. Through this solution, the problem that each student has different pronunciation errors for the same word can be solved. At the same time, possible errors in students' grammar and word usage can be corrected to ensure the accuracy of error correction.

[0136] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A real-time English pronunciation error correction method driven by a deep neural network, characterized in that: include: Obtain the user's past sample manuscripts and sample pronunciations of the sample manuscripts, where the sample manuscripts contain all the vocabulary the user has mastered; Collect sample documents to be classified, preprocess them, extract features from text resources, and convert text information into a vocabulary for representation; Analyze the sample pronunciation and obtain the mapping relationship between speech and vocabulary based on the analysis results; Acquire the user's semantic usage habits, including the user's incorrect usage habits; Obtain the user's real-time conversation voice and obtain the user's actual intention based on semantic usage habits; According to the actual intention of the user, the correct speech text is generated, and based on the correct speech text, the real-time conversation speech is corrected.

2. A learning resource optimization and updating method for a learning machine according to claim 1, characterized in that: The steps of collecting sample documents to be classified, preprocessing them, extracting features from text resources, and converting text information into a vocabulary for representation include the following steps: Removing irrelevant characters from the sample document, wherein the irrelevant characters include special symbols and spaces; Based on the dictionary-based word segmentation method, the English text of the sample manuscript is segmented into at least one independent word; Traverse all sample documents, collect all independent words that appear, build a vocabulary, and use the independent words in the vocabulary as feature words; The frequency of occurrence of characteristic words in all sample manuscripts is counted as characteristic frequency, and the characteristic frequency is paired with the characteristic words.

3. The learning resource optimization and updating method for a learning machine according to claim 2, characterized in that: Analyzing the sample pronunciation and obtaining a mapping relationship between speech and vocabulary based on the analysis result includes the following steps: Based on historical data, the correctly pronounced speech is compared with the standard speech to obtain at least one sample similarity, and the minimum value of the sample similarity is used as the preset value; The sample pronunciation is recognized using the standard speech of the characteristic vocabulary, and the portion of the sample pronunciation with a similarity higher than a preset value is intercepted as the characteristic pronunciation, and the characteristic vocabulary is paired with the characteristic pronunciation; At least one unrecognized part of the sample pronunciation is respectively regarded as an unrecognized speech segment; Establishing a mapping relationship between the characteristic vocabulary of the sample document and the characteristic pronunciation of the corresponding sample pronunciation, where the characteristic vocabulary and the characteristic pronunciation established in the mapping relationship are paired, and the relative position of the characteristic vocabulary in the sample document is consistent with the relative position of the characteristic pronunciation in the sample pronunciation; In the sample document, at least one portion for which no mapping relationship is established is used as a characteristic portion; Acquire characteristic words at both ends of the characteristic part in the sample document as target characteristic words; Acquire a feature pronunciation that has a mapping relationship with the target feature vocabulary from the sample pronunciation as the target feature pronunciation; The part of the sample pronunciation sandwiched between the two target feature pronunciations is regarded as an unrecognized speech segment and corresponds to the feature part; When the feature part contains only one feature word, the corresponding unrecognized speech segment is used as the feature speech to establish a mapping relationship with the feature word contained in the feature part; When the characteristic part contains multiple characteristic words, the lower limit of the user's pronunciation interval is obtained, and at least one discontinuity point of the unrecognized speech segment is identified based on the lower limit of the pronunciation interval. Based on the discontinuity point, the unrecognized speech segment is divided into at least one characteristic speech, and a mapping relationship is established between the characteristic speech in the unrecognized speech segment and the characteristic words in the characteristic part in sequence.

4. The learning resource optimization and updating method for a learning machine according to claim 3, characterized in that: The acquisition of the user's semantic usage habits comprises the following steps: Obtain the user's previously revised English composition text; Using a dependency syntax analysis tool to perform dependency syntax analysis on the preprocessed English composition text to reveal dependency relationships between words, including subject-verb relationships, verb-object relationships, and attributive-predicate relationships; The output of dependency syntax analysis is a tree structure representation, where each node represents a word and each edge represents the dependency relationship between words; In the dependency parsing tree, identify noun phrases and nouns as entity candidates, and confirm the accuracy of the entities based on part-of-speech tags and context information; Store the extracted entities in the entity library; In the dependency parsing tree, verb phrases and verbs are identified as relation candidates, and the subject and object of the verb are analyzed to determine the semantic relationship between entities; Verify and correct the extracted relationships based on contextual information and domain semantics; Attribute extraction is achieved by analyzing the attributives and modifiers in the dependency syntax tree; Create an empty semantic network, define the network structure and the attributes of nodes and edges, set the attribute fields of entities and relationships, the attribute fields of entities include name, type and description, and the attribute fields of relationships include relationship type and confidence; Add the extracted entities and relationships to the semantic network, create nodes for each entity, set the node attributes, create edges for each pair of entities with relationships, and set the edge attributes; The semantic network is presented as a graphical interface, with nodes represented as circles and edges as curves, and different types of entities and relationships are distinguished by different colors, shapes and sizes; Take the semantic network as the user's semantic usage habits.

5. The method for real-time English pronunciation error correction based on deep neural network driving according to claim 4 is characterized in that: The method of obtaining the user's actual intention based on semantic usage habits includes the following steps: Using the characteristic voice, the real-time conversation voice is recognized, and the real-time conversation voice is recognized in chronological order to obtain at least one target characteristic voice; Based on the mapping relationship between the characteristic speech and the characteristic vocabulary, the target characteristic speech is replaced with the characteristic vocabulary that has a mapping relationship with the target characteristic speech to obtain the user's real-time pronunciation text; Use semantic networks to infer the intention of real-time pronunciation text and obtain the user's actual intention.

6. The method for real-time English pronunciation error correction based on deep neural network drive according to claim 5, characterized in that: The method of using the semantic network to infer the intention of the real-time pronunciation text to obtain the actual intention of the user includes the following steps: The real-time pronunciation text is segmented according to vocabulary to obtain at least one actual word, actual edges are connected between adjacent actual words, two entities consistent with the adjacent actual words are obtained in the semantic network as feature entities, and attributes of the edge between the two feature entities are assigned to the actual edge; The actual edge with conflicting attributes is used as the target actual edge, and the actual word connected by the target actual edge is used as the target actual word; Form at least one replacement for the target actual word using entities in the semantic network; Segmenting the real-time pronunciation text after applying the replacement scheme to obtain at least one corrected word, connecting correction edges between adjacent corrected words, and obtaining two entities consistent with the adjacent corrected words in the semantic network as corrected entities. If no edge exists between the two corrected entities, the replacement scheme is deleted; otherwise, the attributes of the edge between the two corrected entities are assigned to the corrected edge. The number of correction edges with conflicting attributes is counted as the check value of the replacement plan, and the replacement plan with the smallest check value is regarded as the actual intention of the user.

7. The method for real-time English pronunciation error correction based on deep neural network drive according to claim 6 is characterized in that: The generation of the correct speech text according to the actual intention of the user comprises the following steps: Use the replacement solution with the smallest checksum value to replace the real-time pronunciation text to obtain the correct voice text.

8. The method for real-time English pronunciation error correction based on deep neural network drive according to claim 7, characterized in that: The method of correcting the real-time conversation speech based on the correct speech text comprises the following steps: Obtain the standard pronunciation of each word in the correct pronunciation text as the reference pronunciation, and number the reference pronunciations in sequence; Perform interval recognition on the user's real-time conversation voice. When the pronunciation interval is greater than the lower limit of the pronunciation interval, the pronunciation interval is used as the cutting point. Using a cutting point to segment the real-time conversation speech into at least one actual word speech, and numbering the actual word speech in sequence; The actual word voice with the same number is compared with the reference voice. When the similarity between the actual word voice and the reference voice is less than a preset value, the reference voice is used to correct the actual word voice.

Images