Multi-round dialogue scene context completion system based on large model driving

Abstract

The invention discloses a multi-round dialogue scene context completion system based on large model driving, and relates to the field of artificial intelligence. A missing identification module; the entity same module obtains a correction statement; the intention recognition module is used for generating the weight of the character intention; the habit recognition module recognizes at least one habit word of the character tag; and the completion module adjusts the pre-filling part to obtain a target filling part, and fills the target filling part to the corresponding position of the multiple rounds of dialogues. By arranging a data processing module, an entity same module, an intention recognition module, a habit recognition module and a complementation module, missing parts in multiple rounds of dialogues can be accurately recognized, it is ensured that the anaphora relation in the multiple rounds of dialogues is clear, the reply content of the missing parts is limited, and the reply content of the missing parts is processed according to habits.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, specifically to a multi-turn dialogue scenario context completion system based on a large model. Background Technology

[0002] With the development of artificial intelligence technology, human-computer dialogue systems have been widely used in customer service, intelligent assistants and other fields. However, existing technologies have the following problems when handling multi-turn dialogues: there are often unclear references or missing information in the dialogue, which affects the completeness of dialogue understanding; it is difficult to accurately identify the true intentions of each participant in the dialogue; there is a lack of modeling of the personalized language habits of individuals; and it is impossible to automatically fill in the missing information in the dialogue, which affects the processing of subsequent dialogues. Summary of the Invention

[0003] To address the aforementioned technical issues, a multi-turn dialogue scenario context completion system based on a large model is provided. This technical solution resolves the problems mentioned in the background section.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] A multi-turn dialogue scene context completion system based on a large model includes:

[0006] The data processing module divides the multi-turn dialogue into at least one basic statement according to the time sequence and matches character tags to the basic statement.

[0007] The missing part identification module, based on semantic recognition, identifies at least one missing part in a multi-turn dialogue;

[0008] The entity is the same module, which identifies at least one object referent in the basic statement, and replaces the object referent in the basic statement according to the entity corresponding to the object referent to obtain the corrected statement;

[0009] An intent recognition module identifies at least one character's intent in the corrected statement and generates weights for the character's intent based on multi-turn dialogue.

[0010] A habit recognition module, which identifies at least one habitual word of a person's tag based on the corrected statement;

[0011] The completion module predicts the missing content based on the character's intention, obtains the pre-filled part, adjusts the pre-filled part according to common words, obtains the target filled part, and fills the target filled part into the corresponding position in the multi-turn dialogue.

[0012] Preferably, the segmentation into at least one basic statement and matching person tags to the basic statement includes the following steps:

[0013] Obtain at least one character who participates in multiple rounds of dialogue, divide the multiple rounds of dialogue according to the character's affiliation, obtain at least one basic statement, and arrange the basic statements in chronological order.

[0014] The name of the person who sends the basic statement is used as a character tag to match the corresponding basic statement.

[0015] Preferably, the step of identifying at least one missing part of a multi-turn dialogue based on semantic recognition includes the following steps:

[0016] The semantics of basic statements are obtained by using a large model, and the basic statements are classified according to the semantics to obtain question basic statements and non-question basic statements.

[0017] When the character tags of the basic statement of the question and the basic statement of the non-question are different, the response of the basic statement of the non-question to the basic statement of the question is a complete answer, a partial answer, or an explanation of why it cannot be answered. When the character tags of the basic statement of the question and the basic statement of the non-question are the same, there is no response relationship between the basic statement of the non-question and the basic statement of the question.

[0018] When no non-problem base statement responds to the problem base statement, the problem base statement is used as the target base statement.

[0019] The response to the target's basic statement will be treated as the missing part.

[0020] Preferably, identifying at least one object pronoun in the basic statement includes the following steps:

[0021] A commonly used dialogue lexicon is pre-acquired. The lexicon stores at least one part of speech and the meaning of the commonly used words under the part of speech. The part of speech is adjective, adverb, noun, verb and personal pronoun.

[0022] The basic statement is divided into at least one basic word group, which is a common word in a common dialogue vocabulary;

[0023] At least one part-of-speech (POS) allocation scheme is formed to create a basic word group, wherein the POS allocation scheme is composed of a random combination of one of the at least one POS of the basic word group;

[0024] The part-of-speech tags assigned to basic word groups in the part-of-speech tagging scheme are used as the characteristic part-of-speech tags of the basic word groups.

[0025] Obtain the meaning of basic phrases under the restriction of feature parts of speech, and use it as feature meaning. Summarize the feature meanings of basic phrases in basic sentences to form a semantic interpretation of the part of speech allocation scheme.

[0026] The target part-of-speech (POS) allocation scheme is selected, and the semantic interpretation of the target POS allocation scheme is consistent with the semantics of the basic sentences identified by the large model. The part-of-speech assigned to the basic word group in the target POS allocation scheme is taken as the true part-of-speech of the basic word group.

[0027] Basic phrases whose real parts of speech are nouns or personal pronouns are all used as object referents.

[0028] Preferably, the process of replacing object pronouns in the basic statement to obtain the revised statement includes the following steps:

[0029] The object pronoun that functions as a noun is denoted as a noun pronoun, and the object pronoun that functions as a personal pronoun is denoted as a pronoun pronoun.

[0030] The noun pronoun that appears before the pronoun referent and is closest to it is designated as the first noun pronoun referent. The noun pronoun that appears after the pronoun referent and is closest to it is designated as the second noun pronoun referent. The distance between the pronoun referent and the noun pronoun referent is the number of bytes between them. The pronoun referent corresponds to its first and second noun pronoun referents.

[0031] In the basic statement, replace the corresponding pronoun with the first noun pronoun to obtain the first preparatory statement;

[0032] In the basic statement, replace the corresponding pronoun with the second noun pronoun to obtain the second preparatory statement;

[0033] The first or second preparatory statement, which does not contradict the semantics of the multi-turn dialogue, is used as the corrective statement.

[0034] Preferably, identifying at least one person's intent in the revised statement includes the following steps:

[0035] The correction statements that contain the same object pronouns and have the same character tags are compiled into a set of correction statements, and the set of correction statements corresponds to the object pronouns.

[0036] Extract statements describing object pronouns from the set of corrected statements, remove duplicate content from the statements describing object pronouns to form user requirements, and match the user tags of the corrected statements in the set of corrected statements to the user requirements.

[0037] Summarize the needs of characters with the same character tags to form character intentions.

[0038] Preferably, the generation of character intent weights based on multi-turn dialogue includes the following steps:

[0039] Record the basic phrases that function as adjectives or adverbs in a character's intentions as emotional phrases;

[0040] Based on big data, at least one sample event of the occurrence of sentiment phrases is obtained in advance, and the importance of the sample event is obtained in advance.

[0041] The sentiment coefficient of a sentiment phrase is obtained by averaging the importance of at least one sample event in which the sentiment phrase appears.

[0042] The average emotional coefficient of the emotional phrases in the character's intention is taken to obtain the index of the character's intention. The indices of all the characters' intentions are summed to obtain the comprehensive index.

[0043] The weight of a character's intention is obtained by dividing the index of the character's intention by the overall index.

[0044] Preferably, the step of identifying at least one common word for a person's tag based on the modified statement includes the following steps:

[0045] The basic statements corresponding to the same character tags are compiled into a basic statement set, and the basic statement set corresponds to the character tag.

[0046] The basic phrases appearing in the basic statements in the basic statement set are compiled into a basic phrase set;

[0047] Classify the basic phrases in the basic phrase set to obtain at least one basic phrase category, and the basic phrases in the basic phrase category have the same meaning;

[0048] The frequency of basic phrases is statistically analyzed within the basic statements in the basic statement set.

[0049] The most frequent basic phrases in the basic phrase categories are taken as idiomatic words. The character tags corresponding to the basic sentence sets of the basic phrase categories are matched to the idiomatic words. The idiomatic words corresponding to the same character tags are summarized to obtain at least one idiomatic word for each character tag.

[0050] Preferably, the step of predicting the content of the missing part based on the character's intention to obtain the pre-filled part includes the following steps:

[0051] The intentions of the characters that intersect with the basic content of the target's statements are taken as the target character's intentions.

[0052] Divide the total number of words in a multi-turn dialogue by the number of basic questions to get the maximum number of words in the content.

[0053] The weights of at least one target person's intentions are summed to obtain a comprehensive weight. The weight of the target person's intention is divided by the comprehensive weight to obtain the proportion of the target person's intention. The maximum number of words in the content is multiplied by the proportion of the target person's intention to obtain the maximum response limit for the target person's intention.

[0054] The target question is obtained by identifying the question content of the target basic statement through a large model and intersecting the question content of the target basic statement with the intention of the target person.

[0055] The big data model is used to identify the semantic tendency of the target person's intention towards the target question. When the semantic tendency is that it cannot be answered, the content related to the target question within the target person's intention is summarized to form the reason for the inability to answer, and this is used as the answer to the target question. When the semantic tendency is that it can be answered, the content related to the target question within the target person's intention is summarized to obtain a preliminary answer. The content related to the preliminary answer is then retrieved from the big data and added to the preliminary answer to obtain the answer to the target question.

[0056] Reduce the answer to the target question to no more than the upper limit of the answer intended by the target person. When reducing, delete phrases with adverbs and adjectives, and delete sentences that are not fluent after deletion. Combine the reduced answer to the target question into the prepared filling part.

[0057] Preferably, adjusting the pre-filled portion according to common usage to obtain the target filled portion includes the following steps:

[0058] Use the character tags of the adjacent basic statements following the target basic statement as the target character tags.

[0059] The target person tag's idiomatic words are used to replace words with the same meaning in the pre-filled part, resulting in the target filled part. The target filled part is placed at the end of the target basic sentence corresponding to the target filled part.

[0060] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0061] By setting up data processing, entity identification, intent recognition, habit recognition, and completion modules, the system can accurately identify missing parts in multi-turn dialogues and uniformly process the referential relationships in multi-turn dialogues to ensure clarity. This guarantees accurate identification of participants' true intentions, and consequently, generates corresponding missing parts based on these intentions. Furthermore, based on the length settings and the weight of the participants' intentions, the system limits the content of responses to missing parts to prevent excessive supplementary content from affecting comprehension of the original multi-turn dialogue. Additionally, based on participants' habits, the system processes responses to missing parts to prevent significant differences in language between supplementary content and the original multi-turn dialogue, thus avoiding readability issues. Attached Figure Description

[0062] Figure 1 This is a flowchart illustrating the multi-turn dialogue scenario context completion system based on a large model driven by the present invention.

[0063] Figure 2 This invention is divided into at least one basic statement, and a flowchart is used to match the basic statement with the person's tag.

[0064] Figure 3 This is a schematic diagram of the process of identifying at least one missing part of a multi-turn dialogue based on semantic recognition according to the present invention.

[0065] Figure 4 This is a schematic diagram illustrating the process of identifying at least one object pronoun in a basic statement according to the present invention.

[0066] Figure 5 This is a flowchart illustrating the process of replacing object pronouns in basic statements to obtain modified statements, as described in this invention.

[0067] Figure 6 This is a schematic diagram illustrating the process of identifying the intent of at least one person in a modified statement according to the present invention.

[0068] Figure 7 This is a flowchart illustrating the process of generating character intent weights based on multi-turn dialogue in this invention.

[0069] Figure 8 This is a schematic diagram of the process of identifying at least one common word of a person's tag based on the modified statement according to the present invention;

[0070] Figure 9 This is a flowchart illustrating the process of predicting the content of missing parts based on the character's intention to obtain the pre-filled part according to the present invention.

[0071] Figure 10 This is a schematic diagram illustrating the process of adjusting the pre-filled portion according to conventional terminology to obtain the target filled portion in this invention. Detailed Implementation

[0072] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0073] Reference Figure 1 As shown, a multi-turn dialogue scene context completion system based on a large model includes:

[0074] The data processing module divides the multi-turn dialogue into at least one basic statement according to the time sequence and matches character tags to the basic statement.

[0075] The missing part identification module, based on semantic recognition, identifies at least one missing part in a multi-turn dialogue;

[0076] The entity is the same module, which identifies at least one object referent in the basic statement, and replaces the object referent in the basic statement according to the entity corresponding to the object referent to obtain the corrected statement;

[0077] An intent recognition module identifies at least one character's intent in the corrected statement and generates weights for the character's intent based on multi-turn dialogue.

[0078] A habit recognition module, which identifies at least one habitual word of a person's tag based on the corrected statement;

[0079] The completion module predicts the missing content based on the character's intention, obtains the pre-filled part, adjusts the pre-filled part according to common words, obtains the target filled part, and fills the target filled part into the corresponding position in the multi-turn dialogue.

[0080] Reference Figure 2 As shown, segmenting into at least one basic statement and matching person tags to the basic statement includes the following steps:

[0081] Obtain at least one character who participates in multiple rounds of dialogue, divide the multiple rounds of dialogue according to the character's affiliation, obtain at least one basic statement, and arrange the basic statements in chronological order.

[0082] The name of the person who sends the basic statement is used as a character tag to match the corresponding basic statement.

[0083] In this solution, to complete the context, at least one missing part of the multi-turn dialogue needs to be identified, and the missing part needs to be supplemented based on the content of the multi-turn dialogue. The supplementation cannot exceed the scope of the multi-turn dialogue. However, due to the unclear reference of personal pronouns and the lack of clear identification of characters' intentions in the multi-turn dialogue, this will cause some interference to the subsequent content supplementation. The supplemented content may be very long, which may lead to an excessive proportion of the text and affect readability. In addition, its language habits are different from those of the characters in the multi-turn dialogue, which may make the supplemented content seem abrupt in the multi-turn dialogue. Therefore, a series of steps are set up to deal with this problem.

[0084] Reference Figure 3 As shown, identifying at least one missing part of a multi-turn dialogue based on semantic recognition includes the following steps:

[0085] The semantics of basic statements are obtained by using a large model, and the basic statements are classified according to the semantics to obtain question basic statements and non-question basic statements.

[0086] When the character tags of the basic statement of the question and the basic statement of the non-question are different, the response of the basic statement of the non-question to the basic statement of the question is a complete answer, a partial answer, or an explanation of why it cannot be answered. When the character tags of the basic statement of the question and the basic statement of the non-question are the same, there is no response relationship between the basic statement of the non-question and the basic statement of the question.

[0087] When no non-problem base statement responds to the problem base statement, the problem base statement is used as the target base statement.

[0088] The response to the target's basic statement will be treated as the missing part.

[0089] During a dialogue, according to common sense, if a character's dialogue needs are not met, regardless of whether the question is asked, the character will ask again. Therefore, it will eventually appear as a question. Thus, we can identify the basic statement of the question. In multiple rounds of dialogue, we should respond to the basic statement of the question. If there is no response, it is the missing part.

[0090] Reference Figure 4 As shown, identifying at least one object pronoun in a basic statement includes the following steps:

[0091] A commonly used dialogue lexicon is pre-acquired. The lexicon stores at least one part of speech and the meaning of the commonly used words under the part of speech. The part of speech is adjective, adverb, noun, verb and personal pronoun.

[0092] The basic statement is divided into at least one basic word group, which is a common word in a common dialogue vocabulary;

[0093] At least one part-of-speech (POS) allocation scheme is formed to create a basic word group, wherein the POS allocation scheme is composed of a random combination of one of the at least one POS of the basic word group;

[0094] The part-of-speech tags assigned to basic word groups in the part-of-speech tagging scheme are used as the characteristic part-of-speech tags of the basic word groups.

[0095] Obtain the meaning of basic phrases under the restriction of feature parts of speech, and use it as feature meaning. Summarize the feature meanings of basic phrases in basic sentences to form a semantic interpretation of the part of speech allocation scheme.

[0096] The target part-of-speech (POS) allocation scheme is selected, and the semantic interpretation of the target POS allocation scheme is consistent with the semantics of the basic sentences identified by the large model. The part-of-speech assigned to the basic word group in the target POS allocation scheme is taken as the true part-of-speech of the basic word group.

[0097] Basic phrases whose real parts of speech are nouns or personal pronouns are all used as object referents.

[0098] In multi-turn dialogues, it is necessary to clarify the entities referred to by personal pronouns. Personal pronouns include you, I, he, it, etc. If the entities they refer to are unclear, it can easily interfere with the understanding of multi-turn dialogues. Therefore, it is necessary to clarify the entities referred to by personal pronouns. Since entities are usually nouns, it is necessary to first identify at least one object pronoun in the basic sentences, and then replace the object pronoun with the noun corresponding to the entity referred to by the personal pronoun.

[0099] Reference Figure 5 As shown, replacing object pronouns in the basic statement to obtain the corrected statement involves the following steps:

[0100] The object pronoun that functions as a noun is denoted as a noun pronoun, and the object pronoun that functions as a personal pronoun is denoted as a pronoun pronoun.

[0101] The noun pronoun that appears before the pronoun referent and is closest to it is designated as the first noun pronoun referent. The noun pronoun that appears after the pronoun referent and is closest to it is designated as the second noun pronoun referent. The distance between the pronoun referent and the noun pronoun referent is the number of bytes between them. The pronoun referent corresponds to its first and second noun pronoun referents.

[0102] In the basic statement, replace the corresponding pronoun with the first noun pronoun to obtain the first preparatory statement;

[0103] In the basic statement, replace the corresponding pronoun with the second noun pronoun to obtain the second preparatory statement;

[0104] The first or second preparatory statement, which does not contradict the semantics of the multi-turn dialogue, is used as the corrective statement.

[0105] During substitution, according to dialogue conventions, noun referents usually appear near pronoun referents. That is, if a noun referent appears in the previous sentence, a pronoun referent can replace it in the next sentence. However, since the same noun referent appears multiple times, in some positions, a pronoun referent referring to the same entity can appear before the noun referent. Because the noun referent referring to the same entity has already appeared before, a noun referent referring to the same entity can appear before or after the pronoun referent. During substitution, it is necessary to check the noun referents before or after the pronoun referent. The basis for the check is whether a contradiction will occur after the substitution. This is because when pronoun referents and noun referents referring to different entities are substituted for each other, a contradiction will occur because the entities corresponding to the descriptive statements are different.

[0106] Reference Figure 6 As shown, identifying at least one person's intent in a corrected statement includes the following steps:

[0107] The correction statements that contain the same object pronouns and have the same character tags are compiled into a set of correction statements, and the set of correction statements corresponds to the object pronouns.

[0108] Extract statements describing object pronouns from the set of corrected statements, remove duplicate content from the statements describing object pronouns to form user requirements, and match the user tags of the corrected statements in the set of corrected statements to the user requirements.

[0109] Summarize the needs of characters with the same character tags to form character intentions.

[0110] The identification of a person's intent mainly relies on person tags. During the identification process, the intents of the same person need to be merged to obtain the person's intent. That is, a person's intent cannot contain the intents of more than one person.

[0111] Reference Figure 7 As shown, the weighting of character intent based on multi-turn dialogue includes the following steps:

[0112] Record the basic phrases that function as adjectives or adverbs in a character's intentions as emotional phrases;

[0113] Based on big data, at least one sample event of the occurrence of sentiment phrases is obtained in advance, and the importance of the sample event is obtained in advance.

[0114] The sentiment coefficient of a sentiment phrase is obtained by averaging the importance of at least one sample event in which the sentiment phrase appears.

[0115] The average emotional coefficient of the emotional phrases in the character's intention is taken to obtain the index of the character's intention. The indices of all the characters' intentions are summed to obtain the comprehensive index.

[0116] The weight of a character's intention is obtained by dividing the index of the character's intention by the overall index.

[0117] The weight of the character's intention is mainly used to allocate content proportions when responding. This is because the supplementary content cannot be too long and its length needs to be limited. However, the response depends on the character's intention. Therefore, when supplementing, the length of the supplementary content corresponding to the character's intention needs to be limited according to the weight of the character's intention.

[0118] Reference Figure 8 As shown, identifying at least one common word for a person's tag based on the corrected statement includes the following steps:

[0119] The basic statements corresponding to the same character tags are compiled into a basic statement set, and the basic statement set corresponds to the character tag.

[0120] The basic phrases appearing in the basic statements in the basic statement set are compiled into a basic phrase set;

[0121] Classify the basic phrases in the basic phrase set to obtain at least one basic phrase category, and the basic phrases in the basic phrase category have the same meaning;

[0122] The frequency of basic phrases is statistically analyzed within the basic statements in the basic statement set.

[0123] The most frequent basic phrases in the basic phrase categories are taken as idiomatic words. The character tags corresponding to the basic sentence sets of the basic phrase categories are matched to the idiomatic words. The idiomatic words corresponding to the same character tags are summarized to obtain at least one idiomatic word for each character tag.

[0124] Each character has their own preferred language style, and supplementary content will be placed in each character's dialogue. When the supplementary content differs from the character's preferred language style, then...

[0125] Reference Figure 9 As shown, the process of predicting the missing content based on the character's intention to obtain the pre-filled portion includes the following steps:

[0126] The intentions of the characters that intersect with the basic content of the target's statements are taken as the target character's intentions.

[0127] Divide the total number of words in a multi-turn dialogue by the number of basic questions to get the maximum number of words in the content.

[0128] The weights of at least one target person's intentions are summed to obtain a comprehensive weight. The weight of the target person's intention is divided by the comprehensive weight to obtain the proportion of the target person's intention. The maximum number of words in the content is multiplied by the proportion of the target person's intention to obtain the maximum response limit for the target person's intention.

[0129] The target question is obtained by identifying the question content of the target basic statement through a large model and intersecting the question content of the target basic statement with the intention of the target person.

[0130] The big data model is used to identify the semantic tendency of the target person's intention towards the target question. When the semantic tendency is that it cannot be answered, the content related to the target question within the target person's intention is summarized to form the reason for the inability to answer, and this is used as the answer to the target question. When the semantic tendency is that it can be answered, the content related to the target question within the target person's intention is summarized to obtain a preliminary answer. The content related to the preliminary answer is then retrieved from the big data and added to the preliminary answer to obtain the answer to the target question.

[0131] Reduce the answer to the target question to no more than the upper limit of the answer intended by the target person. When reducing, delete phrases with adverbs and adjectives, and delete sentences that are not fluent after deletion. Combine the reduced answer to the target question into the prepared filling part.

[0132] Because of the large model, it has the function of content generation, which is a commonly used existing technology. However, the generated content may have different habits and be too long, so corresponding processing is required.

[0133] Reference Figure 10 As shown, adjusting the pre-filled portion according to common usage to obtain the target filled portion involves the following steps:

[0134] Use the character tags of the adjacent basic statements following the target basic statement as the target character tags.

[0135] The target person tag's idiomatic words are used to replace words with the same meaning in the pre-filled part, resulting in the target filled part. The target filled part is placed at the end of the target basic sentence corresponding to the target filled part.

[0136] The supplementary content is a response to the target basic statement. Therefore, it is necessary to determine the person tags of the adjacent basic statements after the target basic statement. The supplementary content will be incorporated into the statements generated by the target person tags. Therefore, idioms need to be determined based on the target person tags.

[0137] Furthermore, this solution also proposes a storage medium on which a computer-readable program is stored. When the computer-readable program is invoked, the aforementioned multi-turn dialogue scenario context completion system based on a large model is executed.

[0138] It is understandable that the storage medium can be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; an optical medium, such as a DVD; or a semiconductor medium, such as a solid-state drive (SSD).

[0139] In summary, the advantages of this invention are as follows: by setting up a data processing module, an entity identification module, an intent recognition module, a habit recognition module, and a completion module, it can accurately identify missing parts in multi-turn dialogues and process the referential relationships in multi-turn dialogues in a consistent manner, ensuring that the referential relationships in multi-turn dialogues are clear. This guarantees the accurate identification of the participants' true intent, and consequently, it can generate corresponding missing parts based on the participants' true intent. Simultaneously, based on the length settings and the weight of the participants' intents, the content of the responses to the missing parts is limited to prevent excessive supplementary content from affecting the understanding of the original multi-turn dialogue. Furthermore, based on the participants' habits, the content of the responses to the missing parts is processed to prevent the supplementary content from differing too much from the original multi-turn dialogue's terminology, thus affecting readability.

[0140] The foregoing has shown and described 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 embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A multi-turn dialogue scene context completion system based on a large model, characterized in that, include: The data processing module divides the multi-turn dialogue into at least one basic statement according to the time sequence and matches character tags to the basic statement. The missing part identification module, based on semantic recognition, identifies at least one missing part in a multi-turn dialogue; The entity is the same module, which identifies at least one object referent in the basic statement, and replaces the object referent in the basic statement according to the entity corresponding to the object referent to obtain the corrected statement; An intent recognition module identifies at least one character's intent in the corrected statement and generates weights for the character's intent based on multi-turn dialogue. A habit recognition module, which identifies at least one habitual word of a person's tag based on the corrected statement; The completion module predicts the missing content based on the character's intention, obtains the pre-filled part, adjusts the pre-filled part according to common words, obtains the target filled part, and fills the target filled part into the corresponding position in the multi-turn dialogue.

2. The multi-turn dialogue scene context completion system based on a large model driven by claim 1, characterized in that, The process of segmenting the data into at least one basic statement and matching the basic statement with person tags includes the following steps: Obtain at least one character who participates in multiple rounds of dialogue, divide the multiple rounds of dialogue according to the character's affiliation, obtain at least one basic statement, and arrange the basic statements in chronological order. The name of the person who sends the basic statement is used as a character tag to match the corresponding basic statement.

3. The multi-turn dialogue scene context completion system based on a large model driven by claim 2, characterized in that, The method of identifying at least one missing part of a multi-turn dialogue based on semantic recognition includes the following steps: The semantics of basic statements are obtained by using a large model, and the basic statements are classified according to the semantics to obtain question basic statements and non-question basic statements. When the character tags of the basic statement of the question and the basic statement of the non-question are different, the response of the basic statement of the non-question to the basic statement of the question is a complete answer, a partial answer, or an explanation of why it cannot be answered. When the character tags of the basic statement of the question and the basic statement of the non-question are the same, there is no response relationship between the basic statement of the non-question and the basic statement of the question. When no non-problem base statement responds to the problem base statement, the problem base statement is used as the target base statement. The response to the target's basic statement will be treated as the missing part.

4. The multi-turn dialogue scene context completion system based on a large model driven by claim 3, characterized in that, The process of identifying at least one object pronoun in a basic statement includes the following steps: A commonly used dialogue lexicon is pre-acquired. The lexicon stores at least one part of speech and the meaning of the commonly used words under the part of speech. The part of speech is adjective, adverb, noun, verb and personal pronoun. The basic statement is divided into at least one basic word group, which is a common word in a common dialogue vocabulary; At least one part-of-speech (POS) allocation scheme is formed to create a basic word group, wherein the POS allocation scheme is composed of a random combination of one of the at least one POS of the basic word group; The part-of-speech tags assigned to basic word groups in the part-of-speech tagging scheme are used as the characteristic part-of-speech tags of the basic word groups. Obtain the meaning of basic phrases under the restriction of feature parts of speech, and use it as feature meaning. Summarize the feature meanings of basic phrases in basic sentences to form a semantic interpretation of the part of speech allocation scheme. The target part-of-speech (POS) allocation scheme is selected, and the semantic interpretation of the target POS allocation scheme is consistent with the semantics of the basic sentences identified by the large model. The part-of-speech assigned to the basic word group in the target POS allocation scheme is taken as the true part-of-speech of the basic word group. Basic phrases whose real parts of speech are nouns or personal pronouns are all used as object referents.

5. A multi-turn dialogue scene context completion system based on a large model as described in claim 4, characterized in that, The process of replacing object pronouns in the basic statement to obtain the revised statement includes the following steps: The object pronoun that functions as a noun is denoted as a noun pronoun, and the object pronoun that functions as a personal pronoun is denoted as a pronoun pronoun. The noun pronoun that appears before the pronoun referent and is closest to it is designated as the first noun pronoun referent. The noun pronoun that appears after the pronoun referent and is closest to it is designated as the second noun pronoun referent. The distance between the pronoun referent and the noun pronoun referent is the number of bytes between them. The pronoun referent corresponds to its first and second noun pronoun referents. In the basic statement, replace the corresponding pronoun with the first noun pronoun to obtain the first preparatory statement; In the basic statement, replace the corresponding pronoun with the second noun pronoun to obtain the second preparatory statement; The first or second preparatory statement, which does not contradict the semantics of the multi-turn dialogue, is used as the corrective statement.

6. The multi-turn dialogue scene context completion system based on a large model as described in claim 5, characterized in that, The process of identifying at least one person's intent in the revised statement includes the following steps: The correction statements that contain the same object pronouns and have the same character tags are compiled into a set of correction statements, and the set of correction statements corresponds to the object pronouns. Extract statements describing object pronouns from the set of corrected statements, remove duplicate content from the statements describing object pronouns to form user requirements, and match the user tags of the corrected statements in the set of corrected statements to the user requirements. Summarize the needs of characters with the same character tags to form character intentions.

7. The multi-turn dialogue scene context completion system based on a large model driven by claim 6, characterized in that, The weighting of character intent based on multi-turn dialogue includes the following steps: Record the basic phrases that function as adjectives or adverbs in a character's intentions as emotional phrases; Based on big data, at least one sample event of the occurrence of sentiment phrases is obtained in advance, and the importance of the sample event is obtained in advance. The sentiment coefficient of a sentiment phrase is obtained by averaging the importance of at least one sample event in which the sentiment phrase appears. The average emotional coefficient of the emotional phrases in the character's intention is taken to obtain the index of the character's intention. The indices of all the characters' intentions are summed to obtain the comprehensive index. The weight of a character's intention is obtained by dividing the index of the character's intention by the overall index.

8. A multi-turn dialogue scene context completion system based on a large model as described in claim 7, characterized in that, The process of identifying at least one common word for a person's tag based on the corrected statement includes the following steps: The basic statements corresponding to the same character tags are compiled into a basic statement set, and the basic statement set corresponds to the character tag. The basic phrases appearing in the basic statements in the basic statement set are compiled into a basic phrase set; Classify the basic phrases in the basic phrase set to obtain at least one basic phrase category, and the basic phrases in the basic phrase category have the same meaning; The frequency of basic phrases is statistically analyzed within the basic statements in the basic statement set. The most frequent basic phrases in the basic phrase categories are taken as idiomatic words. The character tags corresponding to the basic sentence sets of the basic phrase categories are matched to the idiomatic words. The idiomatic words corresponding to the same character tags are summarized to obtain at least one idiomatic word for each character tag.

9. A multi-turn dialogue scene context completion system based on a large model as described in claim 8, characterized in that, The process of predicting the missing content based on the character's intention to obtain the pre-filled portion includes the following steps: The intentions of the characters that intersect with the basic content of the target's statements are taken as the target character's intentions. Divide the total number of words in a multi-turn dialogue by the number of basic questions to get the maximum number of words in the content. The weights of at least one target person's intentions are summed to obtain a comprehensive weight. The weight of the target person's intention is divided by the comprehensive weight to obtain the proportion of the target person's intention. The maximum number of words in the content is multiplied by the proportion of the target person's intention to obtain the maximum response limit for the target person's intention. The target question is obtained by identifying the question content of the target basic statement through a large model and intersecting the question content of the target basic statement with the intention of the target person. The big data model is used to identify the semantic tendency of the target person's intention towards the target question. When the semantic tendency is that it cannot be answered, the content related to the target question within the target person's intention is summarized to form the reason for the inability to answer, and this is used as the answer to the target question. When the semantic tendency is that it can be answered, the content related to the target question within the target person's intention is summarized to obtain a preliminary answer. The content related to the preliminary answer is then retrieved from the big data and added to the preliminary answer to obtain the answer to the target question. Reduce the answer to the target question to no more than the upper limit of the answer intended by the target person. When reducing, delete phrases with adverbs and adjectives, and delete sentences that are not fluent after deletion. Combine the reduced answer to the target question into the prepared filling part.

10. A multi-turn dialogue scene context completion system based on a large model driven by claim 9, characterized in that, The process of adjusting the pre-filled portion according to common usage to obtain the target filled portion includes the following steps: Use the character tags of the adjacent basic statements following the target basic statement as the target character tags. The target person tag's idiomatic words are used to replace words with the same meaning in the pre-filled part, resulting in the target filled part. The target filled part is placed at the end of the target basic sentence corresponding to the target filled part.

Images