Prompt word generation method, device and equipment for generating read-along teaching content

Abstract

The embodiment of the specification provides a prompt word generation method, device and equipment for generating reading teaching content, wherein the prompt word generation method for generating reading teaching content comprises the following steps: obtaining reading dialogue data comprising teacher reply text and student answer text; converting the reading dialogue data into structured data; merging the structured data with a historical memory graph to obtain an updated memory graph; calculating comprehensive scores of each structured data in the updated memory graph; selecting target structured data in which the word element occupancy meets the upper limit constraint of the word element quantity of a large language model from each structured data according to the comprehensive scores; filling the target structured data into a prompt word template to obtain prompt word information; the prompt word information is used for inputting a large language model to generate next round reading teaching content; and the next round reading teaching content comprises at least one type of data in question, teacher reply, blackboard writing and reading dialogue task report.

Description

Technical Field

[0001] This specification relates to the field of computer technology, and in particular to a method, apparatus, computing device, medium, and program product for generating prompt words for generating accompanying reading teaching content. Background Technology

[0002] With the continuous development of artificial intelligence, combining AI with teaching tasks can better assist students in learning. However, as dialogue between teachers and students increases during the teaching process, the number of lexical units required to be input into the large language model also increases, and these units are prone to containing useless information, leading to wasted lexical units and inaccurate processing results from the large language model in handling teaching tasks. Summary of the Invention

[0003] In view of this, embodiments of this specification provide a method for generating prompt words for generating accompanying reading teaching content. One or more embodiments of this specification also relate to a prompt word generation device for generating accompanying reading teaching content, a computing device, a computer-readable storage medium, and a computer program product, to address the technical deficiencies existing in the prior art.

[0004] According to a first aspect of the embodiments of this specification, a method for generating prompt words for generating accompanying reading teaching content is provided, including: Acquire the accompanying reading dialogue data; the accompanying reading dialogue data includes the teacher's reply text and the student's answer text; The accompanying dialogue data is converted into structured data; the structured data includes data type labels, semantic units, and the original evidence text corresponding to the semantic units; The structured data is merged with the historical memory graph to obtain an updated memory graph; the updated memory graph contains multiple nodes; each node contains one piece of structured data. Calculate the overall score of each structured data in the updated memory map; the overall score is positively correlated with the task relevance of the structured data to the reading-aloud dialogue task; Based on the comprehensive score, target structured data whose word count conforms to the upper limit constraint of the word count of the large language model is selected from each of the structured data. The target structured data is filled into the prompt word template to obtain prompt word information; the prompt word information is used to input the large language model to generate the next round of accompanying reading teaching content; the next round of accompanying reading teaching content includes at least one type of data, such as follow-up questions, teacher replies, blackboard writing, and accompanying reading dialogue task reports.

[0005] According to a second aspect of the embodiments of this specification, a prompt word generation device for generating accompanying reading teaching content is provided, characterized in that it includes: The data acquisition module is used to acquire the accompanying reading dialogue data; the accompanying reading dialogue data includes the teacher's reply text and the student's answer text; The data conversion module is used to convert the accompanying dialogue data into structured data; the structured data includes data type labels, semantic units, and the original evidence text corresponding to the semantic units; The memory graph update module is used to merge the structured data with the historical memory graph to obtain an updated memory graph; the updated memory graph contains multiple nodes; each node contains one piece of structured data; The scoring calculation module is used to calculate the comprehensive score of each structured data in the updated memory map; the comprehensive score is positively correlated with the task relevance of the structured data to the reading companion dialogue task; The data selection module is used to select target structured data from the various structured data based on the comprehensive score, whose word occupancy meets the upper limit constraint of the word number of the large language model. The prompt word generation module is used to fill the target structured data into the prompt word template to obtain prompt word information; the prompt word information is used to input the large language model to generate the next round of accompanying reading teaching content; the next round of accompanying reading teaching content includes at least one type of data, such as follow-up questions, teacher replies, blackboard writing, and accompanying reading dialogue task reports.

[0006] According to a third aspect of the embodiments of this specification, a computing device is provided, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the above-described method for generating prompt words for generating accompanying reading teaching content.

[0007] According to a fourth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores computer-executable instructions, which, when executed by a processor, implement the steps of the above-described method for generating prompt words for generating accompanying reading teaching content.

[0008] According to a fifth aspect of the embodiments of this specification, a computer program product is provided, wherein when the computer program product is executed in a computer, the computer is instructed to perform the steps of the above-described method for generating prompt words for generating accompanying reading teaching content.

[0009] At least one embodiment in this specification achieves the following beneficial effects: By acquiring accompanying dialogue data containing teacher response text and student answer text, the accompanying dialogue data is converted into structured data and merged with a historical memory graph to obtain an updated memory graph containing multiple structured data. A comprehensive score is calculated for each structured data in the updated memory graph. Based on the comprehensive score and the upper limit of the number of lexical units in the large language model, target structured data is selected from the structured data. This target structured data is then used to fill the prompt word template to obtain prompt word information, enabling the large language model to generate the next round of accompanying reading teaching content based on the prompt word information. On the one hand, by converting the accompanying dialogue data into structured data, the meaning of the accompanying dialogue data can be preserved while reducing redundant or useless information in the prompt word information, thus improving the practicality of the prompt word information. On the other hand, by selecting appropriate target structured data based on the upper limit of the number of lexical units in the large language model and the comprehensive score, the problem of the large language model being unable to read the complete prompt word information, missing some key information, and resulting in poor-quality generated accompanying reading teaching content can be avoided. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating a method for generating prompt words for generating accompanying reading teaching content, provided in one embodiment of this specification. Figure 2 This is a schematic diagram of a prompt word generation device for generating accompanying reading teaching content, provided in one embodiment of this specification. Figure 3 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0011] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0012] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0013] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0014] Furthermore, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0015] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0016] Sliding window summary: This can represent the segmentation and semantic compression of earlier historical dialogues while retaining the original dialogues from recent rounds.

[0017] Semantic units: can represent structured information extracted from reading dialogues that is related to reading comprehension, student profile, preferences, errors, and emotional state.

[0018] Long-term memory: can represent information such as student profiles, reading abilities, interaction preferences, and mastered / unmastered knowledge points that are saved across different stages and lessons.

[0019] Lexical budget: The context length limit that can be used for system prompts, book content, historical dialogues, and task instructions each time the large language model is invoked.

[0020] Large Language Models (LLMs) are deep learning-based natural language processing models with massive numbers of parameters (e.g., billions to trillions). Through large-scale pre-training, they can achieve multi-task generalization capabilities. They do not require training for specific tasks and can perform tasks such as text generation, translation, and question answering.

[0021] A token is the smallest unit of data used to process text; it is the smallest piece of information that a large model can understand and generate. It can be a continuous string of characters, a single character, half a word, a full word, or a punctuation mark.

[0022] In a guided reading course, a single student goes through multiple stages, including greetings, listening and reading aloud, silent reading, comprehension and question-and-answer sessions, blackboard explanations, phased comprehensive challenge tasks, and report generation. These stages involve a considerable amount of content, making it easy for prompts to contain irrelevant information. In general technologies, servers can retain dialogue history based on the most recent rounds, or simply concatenate the history into plain text like "Teacher:xxx / Student:xxx". However, even plain text contains a significant amount of irrelevant content, which can interfere with the accuracy of the large language model in processing teaching tasks.

[0023] In general technology, during the generation of phased comprehensive challenge tasks, the preceding questions and student answers from multiple rounds of follow-up questions can be carried to help the large language model determine whether to continue following up or whether to summarize and conclude. However, there is still a lot of irrelevant content in the information, which interferes with the accuracy of the large language model in processing teaching tasks.

[0024] This specification provides a method for generating prompt words for generating accompanying reading teaching content. This specification also relates to an apparatus for generating prompt words for generating accompanying reading teaching content, a computing device, and a computer-readable storage medium, which will be described in detail in the following embodiments.

[0025] See Figure 1 , Figure 1 A flowchart illustrating a method for generating prompt words for accompanying reading instruction content according to an embodiment of this specification is shown. From a programming perspective, the entity executing the process can be a program mounted on a server or terminal. From a hardware perspective, the entity executing the process can be a server or platform deployed with a large language model for generating accompanying reading instruction content. Figure 1 As shown, the specific steps include:

[0026] Step 102: Obtain the accompanying reading dialogue data; the accompanying reading dialogue data includes the teacher's reply text and the student's answer text.

[0027] In one embodiment of this specification, the accompanying reading dialogue data can represent the data generated by the dialogue interaction between the teacher and the student in an accompanying reading teaching scenario. The teacher's response text can represent the text content output by the AI ​​(Artificial Intelligence, AI) accompanying reading teacher to the student during the accompanying reading dialogue. The student's response text can represent the student's response text to the teacher's questions, guidance, or teaching activities during the accompanying reading dialogue. The student's response text can be obtained by converting the student's voice response into text through speech recognition technology; alternatively, the student's response text can also be the text directly received from the student's input.

[0028] In practical applications, the accompanying dialogue data can also include image information or handwritten content. The server can use optical character recognition technology to convert the image information or handwritten content into text.

[0029] In one embodiment of this specification, in a guided reading teaching scenario, the AI-powered guided reading teacher can engage in multiple rounds of guided reading dialogue with the student, which may include explaining the book content, asking questions, and providing feedback. Each round of guided reading dialogue can generate both the teacher's response text and the student's answer text. The server can collect the text data generated from each round of guided reading dialogue in real time, using it as input for subsequent structured transformation, memory map updates, and prompt word generation.

[0030] In one embodiment of this specification, the server can extract the teacher's response text and the student's answer text generated in each round of peer-to-peer reading dialogue from the dialogue log after each round of interaction. The teacher's response text can be generated and output using a large language model; the student's answer text can be obtained by converting the student's voice response into text using speech recognition technology, or by directly receiving text input by the student through input devices such as a keyboard or touchscreen. In practical applications, the peer-to-peer reading dialogue data can also include dialogue-related metadata, such as book identifiers, dialogue round identifiers, timestamps, etc., for traceability in subsequent processing. The peer-to-peer reading dialogue data can be obtained by the server after a round of peer-to-peer reading dialogue interaction, or the peer-to-peer reading dialogue data can be obtained by the server in batches at preset time intervals, and so on.

[0031] Step 104: Convert the accompanying dialogue data into structured data.

[0032] The structured data includes data type labels, semantic units, and the original evidence text corresponding to the semantic units.

[0033] In one embodiment of this specification, structured data can represent data organized according to a predefined field format, such as JOSN format, key-value pair format, and XML format. Data type tags can be used to identify the category to which the structured data belongs. Data type tags can indicate what type of information the structured data describes, such as teaching knowledge points, question content, student comprehension deviations, student interests and preferences, and student emotional states. Semantic units represent information units with independent semantic meaning extracted from the accompanying dialogue data. Semantic units can be the result of semantic understanding and information compression of the original dialogue text. A piece of structured data contains at least one semantic unit. Evidence text can represent the original text fragment that generated the corresponding semantic unit. Evidence text can be used to trace semantic units, clarifying from which segment or sentence of dialogue text the semantic unit was extracted. In practical applications, the server can use a large model to convert the accompanying dialogue data into structured data; or the server can convert the accompanying dialogue data into structured data according to preset rules.

[0034] Step 106: Merge the structured data with the historical memory map to obtain the updated memory map.

[0035] The updated memory graph contains multiple nodes; each node contains one piece of structured data.

[0036] In one embodiment of this specification, the historical memory graph can represent a memory graph that has been established and accumulated prior to the current dialogue round. The historical memory graph contains historical structured data, with each piece of historical structured data stored in a node. The historical memory graph can reflect the student's accumulated performance in previous reading dialogues. The memory graph can represent a data structure that organizes memory data in a graph structure. Nodes in the memory graph are used to store structured data, and nodes can be connected by edges to represent the relationships between structured data, such as coming from the same book or belonging to the same knowledge point. An updated memory graph can contain historical structured data and newly incorporated structured data.

[0037] Step 108: Calculate the overall score of each structured data in the updated memory map.

[0038] The overall score is positively correlated with the task relevance of the structured data to the reading dialogue task.

[0039] In one embodiment of this specification, the comprehensive score represents the reference value of structured data in the current guided reading dialogue task; it is a quantitative score calculated for the structured data. The higher the comprehensive score, the greater the reference value of the structured data for the current task. The guided reading dialogue task can represent the teaching task to be performed in the current round of guided reading dialogue, or it can represent the teaching task to be performed in multiple rounds of guided reading dialogue, including the current round. The task types of guided reading dialogue tasks can include follow-up questions, teacher responses, blackboard explanations, and the generation of guided reading dialogue task reports. Task relevance represents the degree of correlation between the structured data and the current guided reading dialogue task. The higher the task relevance, the more valuable the structured data is to the current guided reading dialogue task, and the higher the comprehensive score can be.

[0040] Step 110: Based on the comprehensive score, select target structured data from the various structured data whose word occupancy meets the upper limit constraint of the word number of the large language model.

[0041] In one embodiment of this specification, a token can represent the basic unit when a large language model processes text. A Chinese word can be segmented into one or more tokens; an English word can also be segmented into one or more tokens. Token occupancy can represent the number of tokens occupied by structured data. The token occupancy of different structured data can be the same or different. The longer the text length of the structured data, the larger the token occupancy can be. The upper limit of the number of tokens can represent the maximum number of tokens that the large language model can accept in a single processing. The upper limit of the number of tokens of the large language model is determined by the model's own architecture. The target structured data can be structured data selected from structured data and used to populate the prompt word template.

[0042] Step 112: Fill the target structured data into the prompt word template to obtain prompt word information.

[0043] The prompt word information is used to input the large language model to generate the next round of accompanying reading teaching content; the next round of accompanying reading teaching content includes at least one type of data, such as follow-up questions, teacher replies, blackboard writing, and accompanying reading dialogue task reports.

[0044] In one embodiment of this specification, the prompt word template can represent a predefined text frame containing placeholders. The prompt word template can include a fixed text portion and a dynamically filled portion. The fixed text portion includes system role settings, teaching objective descriptions, etc., while the dynamically filled portion is used to fill in target structured data, recent dialogue text, book content, etc. The prompt word information can represent the complete text obtained after filling the prompt word template with target structured data, which can be used as input to a large language model to guide the large language model in generating specific types of teaching content. The next round of accompanying reading teaching content can represent the teaching output generated by the large language model based on the prompt word information, used for the next round of accompanying reading dialogue. Follow-up questions can represent a teaching interaction where the teacher raises further questions based on the student's answer, guiding the student to think more deeply. Teacher responses can represent the teacher's response to the student's answer, including explanation of knowledge points, encouraging feedback, and corrective guidance. Blackboard writing can represent the teaching points presented by the teacher in written form to assist students in reading and understanding. The accompanying reading dialogue task report can represent a summary text generated after the accompanying reading course or after the completion of a phase task, including student learning performance, ability changes, and recommendations.

[0045] It should be understood that the order of some steps in the methods described in one or more embodiments of this specification may be interchanged according to actual needs, or some steps may be omitted or deleted.

[0046] Figure 1 The method described here acquires accompanying dialogue data containing both teacher and student responses, converts this data into structured data, and merges it with a historical memory graph to obtain an updated memory graph containing multiple structured data entries. A comprehensive score is calculated for each structured data entry in the updated memory graph. Based on the comprehensive score and the upper limit of the word count in the large language model, target structured data is selected from the structured data. This target structured data is then used to fill the prompt word template, generating prompt word information. This prompt word information allows the large language model to generate the next round of accompanying reading content. On one hand, converting the accompanying dialogue data into structured data ensures the meaning of the dialogue while reducing redundant or useless information in the prompt word information, improving its practicality. On the other hand, selecting appropriate target structured data based on the upper limit of the word count in the large language model and the comprehensive score avoids the problem of excessive word counts in the prompt word information, which could prevent the large language model from reading the complete prompt word information, missing key information, and resulting in poor-quality accompanying reading content.

[0047] based on Figure 1 In addition to the method described herein, this specification also provides some specific implementation methods of the method, which will be described below.

[0048] To further extract semantics from the teacher's response text and the student's answer text, obtaining their respective key-value pair data, and generating structured data based on the key-value pair data, in one or more embodiments of this specification, optionally, the conversion of the accompanying dialogue data into structured data may specifically include: Semantic extraction is performed on the teacher's response text to obtain the first key-value pair data; the first key-value pair data includes a first data type label, a first semantic unit, and the first evidence original text. Semantic extraction is performed on the student's response text to obtain second key-value pair data; the second key-value pair data includes a second data type label, a second semantic unit, and the original evidence text. The structured data is generated based on the first key-value pair data and the second key-value pair data.

[0049] In one embodiment of this specification, semantic extraction can refer to the process of identifying or extracting information fragments with specific meanings from natural language text. The result of semantic extraction can be semantic information represented in a structured form; or it can be words from the original text. Key-value pair data can represent data stored in the form of key-value pairs. Each key-value pair can contain a field name and a field content corresponding to the field name. Key-value pair data can be in JSON format, a dictionary structure, or a similar data representation. There can be one or more first key-value pairs. There can be one or more second key-value pairs.

[0050] In one embodiment of this specification, the teacher's response text carries information about teaching behavior, such as the knowledge points explained, the questions asked, and the content displayed on the blackboard. The student's response text carries information about the student's learning status, such as the student's level of understanding, emotional state, ability performance, and interests. The server can also set a confidence level for each piece of structured data or for each key-value pair to facilitate the subsequent filtering of target structured data. The initial confidence level can be a preset value, which can be adjusted based on the accumulation of multiple rounds of dialogue.

[0051] In one embodiment of this specification, semantic extraction is performed separately on the teacher's response text and the student's answer text. Different extraction strategies can be used to extract the text independently, avoiding confusion regarding information sources and data types caused by mixing the two types of text. This supports targeted retrieval and filtering based on information source and category in subsequent memory management. Furthermore, the teacher's response text and the student's answer text can be integrated into a memory graph as structured data. This allows the memory graph to record the complete reading-along teaching process, providing a memory foundation of teaching behaviors and learning feedback for the subsequent generation of targeted reading-along teaching content.

[0052] In one embodiment of this specification, the server can also obtain structured data by matching information of specific patterns from teacher response text and student answer text using preset regular expressions or keyword templates; or, it can also obtain structured data by recognizing information from teacher response text and student answer text using a large model.

[0053] In practical applications, the server can merge original evidence as edges into the memory graph, and can merge other data in the structured data besides the original evidence as nodes into the memory graph.

[0054] In one embodiment of this specification, before converting the accompanying dialogue data into structured data, the accompanying dialogue data needs to be preprocessed to obtain a preprocessed result, which is then converted into structured data. Specifically, the preprocessing of the accompanying dialogue data involves determining the tags in the teacher's response text; these tags may include spoken tags, whiteboard tags, and model inference tags; retaining the text content marked by the spoken tags and whiteboard tags; extracting the content marked by the model inference tags to obtain preprocessed teacher text containing spoken text and whiteboard text; determining the type of student's response content; if the student's response content is speech-based, speech recognition technology is used to convert the student's response content into student response text; if the student's response content is image-based, optical character recognition technology is used to convert the student's response content into student response text; if the student's response content is text-based, the student response text is determined. The server can then convert the preprocessed teacher text and student response text into structured data.

[0055] To extract multi-dimensional teaching behavior information from teacher response text, first semantic units can be extracted from different types of teaching information. In one or more embodiments of this specification, optionally, the teacher response text includes spoken text and blackboard text; the semantic extraction of the teacher response text may specifically include: Extract the teaching knowledge points and questions from the spoken text; Extract key knowledge points from the blackboard text; The teaching knowledge point information, the question content, and the key knowledge point information are used as the first semantic unit; Determine the first data type tag corresponding to the first semantic unit; The original text fragments from the spoken text and the written text that generate the first semantic unit are used as the original text of the first evidence.

[0056] In one embodiment of this specification, the spoken text can represent teaching content verbally expressed to students by the AI-assisted reading teacher in natural language. The spoken text can be real-time, student-oriented content presented in a conversational tone, such as explaining knowledge points, asking questions, or offering encouragement. It can use spoken text tags such as... <reply> and< / reply> Mark the spoken text.

[0057] The blackboard text can represent the key teaching points presented in written form by the AI-assisted reading teacher. The blackboard text can be presented in a descriptive tone to aid student reading and comprehension, such as textual summaries of key knowledge points, formula explanations, etc., and can use tags such as... <board> and< / board> Mark the whiteboard text.

[0058] Teaching knowledge information refers to the knowledge-based information extracted from the spoken text that is related to the content of the book or the teaching objectives. Teaching knowledge point information can be a semantic summary or the original content of the knowledge explanation in the spoken text. Question content refers to the information of questions posed by the teacher to students, extracted from the spoken text. Question content can be a semantic summary or the original content of the question section in the spoken text. Key knowledge point information refers to the key teaching points extracted from the blackboard text and presented in written form. Key knowledge point information can be a semantic summary or the original content of the content displayed in the blackboard text. Original text fragments can refer to unprocessed, continuous segments of original text extracted from the original text. Original text fragments can be a sentence, a paragraph, or a phrase.

[0059] In one embodiment of this specification, the server can utilize a large model to extract teaching knowledge points and questions from the spoken text. Specifically, the spoken text is input into the large model, and preset extraction prompts guide the model to identify the key knowledge points explained and the questions posed in the spoken text. The extraction prompts can specify the types of information to be extracted and the output format. Correspondingly, the server can also utilize the large model to extract key knowledge points from whiteboard text.

[0060] In one embodiment of this specification, the server may also refer to the book content and teaching objectives related to the reading dialogue task to determine which content in the spoken text belongs to knowledge points, which content belongs to questions, and whether the knowledge points are aligned with the current teaching objectives, thereby enabling accurate acquisition of knowledge point information and question content.

[0061] The first data type label can be obtained by matching the content source and information type of the first semantic unit with a preset data type label mapping table. For example, if the first semantic unit comes from teaching knowledge point information in the spoken text, the corresponding first data type label can be determined as "teaching knowledge point"; if the first semantic unit comes from the question content in the spoken text, the corresponding first data type label can be determined as "question"; if the first semantic unit comes from the key knowledge point information in the blackboard text, the corresponding first data type label can be determined as "blackboard key points". Alternatively, the first data type label can also be determined using a large model.

[0062] In one embodiment of this specification, the server can locate the original text of the first evidence by sentence boundaries; alternatively, the server can locate the original text of the first evidence by semantic boundaries; or alternatively, the server can locate the original text of the first evidence by markers. Sentence boundary location can mean that if the first semantic unit is extracted from a sentence in the teacher's reply text, then that complete sentence is taken as the original text of the first evidence. Semantic boundary location can mean that if the first semantic unit is extracted from a clause in the teacher's reply text, then that clause is taken as the original text of the first evidence. Marker location can mean that if the start and end position markers of the first semantic unit in the source text are recorded during the extraction process, then the corresponding text segment is extracted as the original text of the first evidence based on the markers. If the first semantic unit is extracted from multiple unconnected original text segments, the multiple original text segments can be merged as the original text of the first evidence, and the boundaries of each original text segment can be marked by delimiters.

[0063] In one embodiment of this specification, the server can also extract teaching strategy information from the spoken text, such as descriptions of teaching strategies like "ask questions before explaining" or "analogy-based guidance"; it can also extract encouraging feedback information from the spoken text, such as positive feedback from teachers to students like "You answered very well" or "Your understanding is very insightful"; and it can also extract difficulty adjustment information from the spoken text, such as teachers' adjustments to the teaching difficulty like "This question is a bit difficult, let's think about it from a different angle".

[0064] In one embodiment of this specification, the server can also extract structural relationship information from the whiteboard text, such as structured information like character relationship diagrams and event timelines displayed on the whiteboard; it can also extract comparison information from the whiteboard text, such as comparison table content displayed on the whiteboard.

[0065] In one embodiment of this specification, semantic extraction is performed on the spoken text and blackboard text in the teacher's restored text, and then different first data type labels are assigned to the first semantic units such as teaching knowledge point information, question content, and key knowledge point information. This enables targeted retrieval and filtering of target structured data according to the teaching behavior type, providing multi-dimensional reference data.

[0066] Optionally, in one or more embodiments of this specification, the semantic extraction of the student's answer text may specifically include: Extract key answer information from the student's response text; Based on the student's response text, identify the student's emotional state; The key answer information and the emotional state are used as the second semantic unit; Determine the second data type tag corresponding to the second semantic unit; The original text fragment from the student's answer text that produces the second semantic unit is used as the original text of the second evidence.

[0067] In one embodiment of this specification, key answer information can represent information extracted from student response text that reflects the student's cognitive state regarding the book content or teaching questions. Key answer information can be a semantic summary of content with teaching evaluation value in the student response text, and can include data from at least one category: comprehension and expression, misconceptions, and ability performance. Specifically, comprehension and expression can represent the student's degree of understanding and expression of the book content or teaching questions; misconceptions can represent the student's misunderstanding or biased understanding of the book content or knowledge points, reflecting aspects of the student's cognition that need correction; and ability performance can represent the student's comprehensive reading level in at least one dimension, such as reading comprehension, expression, or depth of thinking.

[0068] Emotional state can represent the emotional tendency or psychological state exhibited by a student when answering questions. Emotional state is identified by analyzing features such as tone words, emotional vocabulary, and expression style in the student's answer text. Emotional state can include at least one of the following: positive, pleasant, negative, frustrated, confused, hesitant, engaged, empathetic, indifferent, or perfunctory. Emotional state can be determined by matching the student's answer text with a pre-defined emotional vocabulary list; alternatively, it can be obtained by inputting the student's answer text into an emotion recognition model; or, if the student's answer text is converted from speech, the server can obtain it by fusing acoustic features such as intonation, speed, and pauses with text features. The second data type label can be obtained by matching the content source and information type of the sub-semantic unit with a pre-defined data type label mapping table; alternatively, the second data type label can be obtained by inputting the second semantic unit into the larger model and outputting it. The method of obtaining the second piece of evidence can be the same as the method of obtaining the first piece of evidence.

[0069] In one embodiment of this specification, the server can use a large model to identify student response text and obtain key response information and emotional state; alternatively, it can identify key response information and emotional state in student response text based on preset recognition rules. In practical applications, the server can also record the student's thinking time and expression time, using these as auxiliary references to identify student emotions.

[0070] For example, if a student's answer text contains phrases such as "I like foxes the most", "Foxes want to be tamed because then they will be different", and "I think the fox was very sad when the Little Prince left", then the second semantic unit can be extracted: hobby = likes foxes; causal understanding = taming establishes a unique relationship between the two; emotional understanding = the fox was sad because of the separation.

[0071] In one embodiment of this specification, the server can also extract information from student responses indicating students' interest and preferences for specific book content, teaching methods, or topics; it can also extract information from student responses indicating students' cooperation with teaching activities. In practical applications, before converting the accompanying dialogue data into structured data, the accompanying dialogue data can be deduplicated to reduce interference from irrelevant information and improve the efficiency of semantic extraction from teacher responses and student answers.

[0072] In one embodiment of this specification, by identifying key answer information in the student's answer text, the key answer information can be used as a reference so that the subsequently generated accompanying reading teaching content is more targeted at the knowledge level. At the same time, by identifying the student's emotional state, corresponding emotional feedback can be provided to the student, which improves the quality of the generated accompanying reading teaching content and also enhances the student's learning enthusiasm.

[0073] To avoid duplicate storage of semantically consistent structured data, semantically consistent or highly similar structured data can be merged. In one or more embodiments of this specification, optionally, the historical memory map includes historical structured data with confidence levels; merging the structured data with the historical memory map to obtain an updated memory map may specifically include: Calculate the similarity between the structured data and the historical structured data; If the similarity is greater than a preset similarity threshold, the confidence of the historical structured data is enhanced to obtain an updated memory map; the updated memory map records the time information for updating the confidence.

[0074] In one embodiment of this specification, confidence can represent a quantitative measure of the credibility of the information expressed by structured data. Similarity can represent the degree of semantic similarity between structured data and historical structured data. Similarity can be calculated using a preset similarity calculation formula, such as cosine similarity or Euclidean distance. The preset similarity threshold can be determined based on expert experience or actual user needs; for example, the preset similarity threshold can be a value such as 0.9, 0.85, or 0.95. Time information can represent the timestamp for updating the confidence.

[0075] In one embodiment of this specification, the confidence level of historical structured data is enhanced to obtain an updated memory map. Specifically, the server can use a formula... Calculate the new confidence level for historical structured data; where This represents the enhanced new confidence level; η represents the preset learning rate, a value determined based on expert experience, such as 0.1, 0.05, etc., or it can be determined based on similarity. The value of η is positively correlated with similarity. For example, if the preset similarity threshold is 0.8, then if the similarity is 0.8, then η = 0.03; if the similarity is 0.85, then η can be 0.07; pos represents the positive verification signal, which can be 1; clip() can indicate that the new confidence level is within the range of [0,1]; then the update time field of the node representing the historical structured data is updated to the current time to record the time of this confidence enhancement. In practical applications, the server can also supplement the original evidence text of the new structured data into the node corresponding to the historical structured data; and record the dialogue turn identifier corresponding to the new structured data into the node.

[0076] In one embodiment of this specification, the server can configure different preset similarity thresholds for different data type tags. For example, the threshold for structured data labeled with the "error concept" tag can be 0.9; the threshold for structured data labeled with the "interest preference" tag can be 0.8. Alternatively, the server can configure the same preset similarity threshold for different data type tags. The server can also set an upper limit on the number of original evidence texts contained in each node of the memory graph. If the upper limit is exceeded, the original evidence texts can be sorted from earliest to latest according to the time when they were added to the node, and the excess original evidence texts at the top can be removed; or, the semantic similarity between each original evidence text and the semantic unit can be determined, and the original evidence texts within the upper limit number that have the highest semantic similarity can be retained.

[0077] In practical applications, before calculating the similarity between structured data and historical structured data, it is possible to determine whether the data type labels of the structured data and historical structured data are the same. If the data type labels are the same, then the similarity between the structured data and historical structured data is calculated. Alternatively, before calculating the similarity, it is possible to identify target historical structured data in the historical memory map that has the same data type label as the structured data, and then calculate the similarity between the structured data and the target historical structured data.

[0078] In one embodiment of this specification, the similarity between new structured data and historical structured data can objectively determine the degree of semantic consistency between new information and existing memories, providing a reliable data foundation for the merging strategy. This improves the accuracy of merging data into the historical memory map, avoids wasting storage space by including duplicate and redundant data in the updated memory map, and also provides complete evidentiary support for subsequent memory conflict tracing by enhancing confidence and adding new original evidence.

[0079] In one or more embodiments of this specification, the method may optionally further include: If the similarity is less than the preset similarity threshold, then it is determined whether there is a conflict between the structured data and the historical structured data; If there is a conflict between the structured data and the historical structured data, the structured data and the historical structured data are merged based on the number of original evidence texts and the confidence level to obtain a merged result. The merge result is updated in the historical memory map to obtain the updated memory map; the updated memory map displays the conflict status at the merge result and records the conflict source information; the conflict source information includes the structured data and the historical structured data.

[0080] In one embodiment of this specification, a conflict can indicate a situation where the data type labels of structured data and historical structured data are the same, but the meanings of the semantic units are opposite or contradictory. A conflict can reflect a change or contradiction in a student's understanding of the same type of information across different rounds of dialogue. For example, if a student expresses "I think the Little Prince doesn't like his rose" in the third round of dialogue and "I think the Little Prince misses his rose" in the eighth round, both sets of structured data have data type labels indicating understanding, but the meanings of the semantic units are opposite, then a conflict is determined to exist.

[0081] The number of original evidence texts can represent the number of original evidence texts contained in historical structured data, or it can represent the number of original evidence texts contained in structured data. The merged result can represent the result data obtained after resolving conflicts between conflicting structured data and historical structured data. The merged result can replace the original historical structured data stored in the corresponding node of the memory graph.

[0082] A conflict status can be represented by a marker field in a memory graph node that indicates the presence of an information conflict at that node. If the node does not have an information conflict, the conflict status can be hidden. Conflict source information can represent detailed information about the original data that caused the conflict, allowing for tracing the source and content of the conflict.

[0083] In one embodiment of this specification, it is determined whether there is a conflict between structured data and historical structured data. Specifically, the structured data and historical structured data can be input into a large model to obtain the judgment result output by the large model; or, if the similarity is less than a preset similarity threshold, it is determined whether the data type label of the structured data and the data type label of the historical structured data are consistent. If the data type labels are consistent, it indicates that there is a conflict.

[0084] In one embodiment of this specification, structured data and historical structured data are merged. Specifically, the existing original evidence text in the node corresponding to the historical structured data is retained, and the original evidence text of the structured data is added to the node to obtain an updated node. The node may contain the original evidence text that generated the historical semantic unit and the original evidence text that generated the new semantic unit, and the original evidence text is marked with a dialogue round identifier to distinguish it. The confidence level is weighted and summed based on the number of original evidence texts to obtain a new confidence level. The new confidence level is assigned to the structured data represented by the node.

[0085] In practical applications, the node containing historical structured data can retain the historical semantic units contained in the historical structured data, and record the new semantic units contained in the structured data that conflict with the historical semantic units as conflict source information; or, if the original text of the structured data is more sufficient as evidence, the historical semantic units in the node containing the historical structured data can be replaced with the new semantic units in the structured data, and the historical semantic units can be recorded as conflict source information; or, the historical semantic units and new semantic units can be recorded in the semantic unit field of the node containing the historical structured data.

[0086] In practical applications, servers can also merge data based on time relevance. For example, structured data is generated closer to the current time than historical structured data, and is better able to reflect students' recent cognitive state. When merging, structured data can be given a higher time weight, making the merged result more inclined to adopt new semantic units.

[0087] In practical applications, conflicts can encompass various scenarios. For example, if the evidentiary strength of historical structured data and other structured data is relatively close, the merging result is uncertain and can be marked as pending review for manual review to obtain the final merging result. Alternatively, if the evidence in the historical structured data is stronger, the merging result will retain the historical semantic units and keep the new semantic units in the conflict source information, marking the conflict state. Or, if the evidence in the structured data is stronger, the merging result will be new semantic units, while the historical semantic units will be kept in the conflict source information, marking the conflict state. In practical applications, if subsequently generated structured data strengthens the confidence of one side in the conflict, the high-confidence structured data will be retained, the low-confidence structured data will be deleted, and the conflict state will no longer be displayed.

[0088] In one embodiment of this specification, by identifying historical structured data that conflict with the newly generated structured data, the newly generated structured data and the historical structured data are merged, and the conflict status and conflict source information are retained in the merging result so that subsequent processing can identify conflict nodes and take a weight reduction strategy.

[0089] In one or more embodiments of this specification, the method may optionally further include: If the structured data does not conflict with the historical structured data, the structured data is merged into the historical memory graph as a new node to obtain the updated memory graph.

[0090] In one embodiment of this specification, the server can store structured data into a new node. The new node can contain data type labels, semantic units, evidence text, book identifiers, dialogue round identifiers, confidence levels, and update times. It can be represented using a JSON structure {type_i, value_i, evidence_i, bookId, dialogId, conf_i, t_i}, where type_i can represent the data type label of the i-th node, value_i can represent the semantic unit of the i-th node, evidence_i can represent the evidence text of the i-th node, dialogId can represent the dialogue round identifier, bookId can represent the book identifier, conf_i can represent the confidence information of the i-th node, and t_i can represent the update time of the i-th node.

[0091] In one embodiment of this specification, the structured data or memory map can also store the student's rhythmic characteristics. Rhythmic characteristics can represent information reflecting changes in the student's learning rhythm and state during the guided reading dialogue, and may include at least one of the following: consecutive correct answers, consecutive incorrect answers, historical instances of non-cooperation, and slow-start waiting time. Consecutive correct answers can represent the number of times a student answers correctly in multiple consecutive rounds of dialogue. When the number of consecutive correct answers is high, the system can prompt to increase the difficulty of the questions or reduce superficial explanations. Consecutive incorrect answers can represent the number of times a student answers incorrectly in multiple consecutive rounds of dialogue. When the number of consecutive incorrect answers reaches a preset number, the system can reduce the difficulty of the questions, switch to an explanation mode, or use selection-based questions. Slow-start waiting time can represent the time required for a student to go from starting guided reading to entering an actively participating state, used to reserve more guidance time for slow-starting students in subsequent lessons. Rhythmic characteristics can be stored in the memory map as a data type of structured data, or they can be attached as an independent field to the node information.

[0092] In one embodiment of this specification, the confidence level of a new node can be set to a preset global default value; alternatively, a corresponding preset initial confidence level can be determined based on the data type label and set as the confidence level of the new node; or, the confidence level can be generated simultaneously with the generation of structured data using a large model, and this confidence level can be used as the confidence level of the new node. In practical applications, when creating a new node, an association edge can also be established between the new node and existing nodes. The type of association edge can include at least one of the following: association with the same book, association with the same dialogue round, and association with the same knowledge point. The new node can also establish a time-series association with the nodes generated from the previous round of accompanying reading dialogue data, so as to track the time evolution trajectory of the student profile based on the time-series association.

[0093] In practical applications, as the number of dialogue rounds increases, the number of nodes in the memory graph will also increase. To avoid the problem of a large number of nodes in the memory graph making management and data tracking difficult, multiple nodes of the same data type can be semantically similar and merged into a summary node. The semantic unit of the summary node can be a general description of the semantic units of multiple semantically similar nodes. Alternatively, structured data that has not been used as target structured data to fill prompt words for a long time can be identified, removed from the memory graph, and saved to a preset database. This avoids excessive occupation of the memory graph by long-term inactive structured data and also avoids situations where data cannot be found.

[0094] By comparing structured data with historical structured data, appropriate methods are selected to merge the structured data into the memory graph, resulting in an updated memory graph. This allows the memory graph to retain more complete information while reducing the storage of redundant information, thus improving the quality of the memory graph.

[0095] To quantify the reference value of each structured data point in the memory graph for the current reading-aloud dialogue task from multiple dimensions, in one or more embodiments of this specification, optionally, the calculation of the comprehensive score of each structured data point in the updated memory graph may specifically include: Obtain the confidence level and time recentity of each of the structured data; the time recentity represents the inverse index of the time interval between the generation time of the structured data and the current time. Calculate the task relevance of each of the structured data sets to the accompanying dialogue task; Calculate the cost value of each structured data based on the number of lexical units in each structured data; The confidence level, the time relevance, the task relevance, and the cost value are weighted to obtain a comprehensive score for each of the structured data.

[0096] In one embodiment of this specification, time recency can be used to reflect the freshness of data. An inverse indicator can represent a negative correlation between time interval and time recency; the larger the time interval, the lower the time recency; the smaller the time interval, the higher the time recency. Generation time can represent the time when the node corresponding to the structured data was created or last updated; the generation time is stored in the node's update time field. Cost value can represent a quantitative measure of the token resources used to populate a prompt word template with a piece of structured data.

[0097] In one embodiment of this specification, the server can be based on the formula score_k=α rel_k+β conf_k+γ recency_k-△ cost_k calculates the overall score of the structured data; where score_k represents the overall score of the k-th structured data; rel_k represents the task relevance of the k-th structured data; conf_k represents the confidence level of the k-th structured data; recency_k represents the time recentity of the k-th structured data; cost_k represents the cost value of the k-th structured data; α, β, γ, and Δ can be the policy weights corresponding to the aforementioned dimensions.

[0098] In one embodiment of this specification, the values ​​of the strategy weights can be adjusted according to the task type of the accompanying dialogue task. For example, if the task type is a follow-up question task, the value of α is larger than other values; if the task type is an accompanying dialogue task report, the values ​​of β and γ are larger than other values.

[0099] In one embodiment of this specification, a comprehensive score is obtained by weighted summation of multiple dimensions, which provides a quantitative basis for the selection of target structured data under subsequent lexical budget constraints, so as to select target structured data that maximizes information value within a limited lexical budget.

[0100] Optionally, in one or more embodiments of this specification, calculating the task relevance of each of the structured data to the accompanying dialogue task may specifically include: Determine the task type of the accompanying dialogue task corresponding to the structured data; Based on the task type, determine the weight value of each semantic unit in the structured data; Calculate the relevance of each semantic unit to the subtask of the accompanying dialogue task; The task relevance is obtained by weighting and summing the subtask relevance of each semantic unit using the weight values ​​of each semantic unit.

[0101] In one embodiment of this specification, the task type can represent a specific category label for the accompanying dialogue task. Task types include follow-up questions, small talk, blackboard writing, reports, and challenge tasks. Structured data can represent multiple semantic units belonging to the same dialogue round. Based on the data type label of the semantic unit, a corresponding weight value is determined. Semantic similarity is calculated between the semantic unit and the task text representing the accompanying dialogue task, and the similarity calculation result is determined as the sub-task relevance. Then, the sub-task relevance of each semantic unit is multiplied by its corresponding weight value, resulting in multiple products. These products are then summed to obtain the task relevance. Thus, the task relevance of structured data can be determined through multiple semantic units belonging to the same dialogue round, enabling subsequent selection of data to be filled into the prompt word template on a dialogue round basis, improving the accuracy of prompt words and the accuracy of the model's prompt word-based generation results.

[0102] Optionally, in one or more embodiments of this specification, selecting target structured data from the various structured data whose lexical occupancy meets the upper limit constraint of the number of lexical units in a large language model based on the comprehensive score may specifically include: The structured data are combined and processed to obtain multiple structured data sets; the number of lexical units contained in each structured data set is less than or equal to the upper limit of the number of lexical units in the large language model; Calculate the sum of the comprehensive scores of the structured data contained in the structured data set to obtain the total score of each structured data set; The structured data set with the highest total score is determined as the target structured data set; the target structured data set contains the target structured data.

[0103] In one embodiment of this specification, the structured data set may contain one or more pieces of structured data. The number of lexical units contained in the structured data set can represent the sum of the lexical units occupied by each piece of structured data in the structured data set. The total score can represent the sum of the comprehensive scores of each piece of structured data belonging to the same data set, with the sum being the total score of the structured data set.

[0104] In one embodiment of this specification, the various structured data are combined. Specifically, the knapsack optimization formula can be used to combine the various structured data; or, a combination enumeration method can be used, with the upper limit of the number of lexical units in the large language model as a constraint, to combine the various structured data.

[0105] Sort the structured datasets from highest to lowest total score and select the structured data items in the top-ranked dataset as the target structured data; or sort the structured datasets from lowest to highest total score and select the structured data items in the bottom-ranked dataset as the target structured data.

[0106] In practical applications, prompt word information can include system prompt words, pre-set N-round dialogue texts, accompanying dialogue tasks, book excerpts targeted by the accompanying dialogue tasks, and target structured data. System prompt words can be some prompt words carried in the prompt word template; the pre-set N-round dialogue texts can represent the most recently generated N-round accompanying dialogue data; the accompanying dialogue task can represent the task corresponding to this round of accompanying dialogue. Target structured data can be selected from structured data generated before the most recently generated N-round accompanying dialogue data in the memory graph, used as a reference for long-term memory to improve the accuracy of model processing. By selecting target structured data and filling it into the prompt word template, the large language model can reason based on memory combinations that maximize information value within a limited word budget when generating accompanying teaching content, improving the personalization and targeting of the generated content.

[0107] In one or more embodiments of this specification, optionally, the structured data has a confidence level; the method may further include: Obtain feedback signals from the structured data; the feedback signals include at least one of positive feedback signals and negative feedback signals; The time decay value is determined based on the time interval between the generation time of the structured data and the current time. The confidence level is updated based on the feedback signal and the time decay value to obtain the target confidence level of the structured data.

[0108] In one embodiment of this specification, the feedback signal can represent a signal that verifies the correctness or validity of the information expressed by the structured data. A positive feedback signal can represent a feedback signal that verifies the information expressed by the structured data as correct. A negative feedback signal can represent a feedback signal that verifies the information expressed by the structured data as incorrect or outdated. The time decay value can represent a value determined based on the time interval for decaying the confidence level during confidence updates, and can be obtained through a lookup table.

[0109] In one embodiment of this specification, the feedback signal can be student correction feedback obtained by monitoring whether the student's statements in subsequent dialogues contradict existing structured data; alternatively, the feedback signal can be feedback obtained by teachers or review users annotating the structured data for correctness through an annotation interface, using the annotation results as the feedback signal; or behavioral feedback can be obtained by monitoring the student's behavior during the guided reading process. For example, if a student exhibits non-cooperation or skips activities under the guidance of teaching content generated based on a certain piece of structured data, a negative feedback signal is generated for that piece of data. The time decay value can be linearly decayed, for example, the time decay value = a preset decay coefficient. Time interval; the time decay value can also be stepped decay, for example, dividing the time interval into several intervals, with different intervals corresponding to different decay values. Time decay can be configured by number of days or class sessions, and different decay periods can be used for structured data with different data type tags.

[0110] The server can be based on the formula conf_i'=clip(conf_i+eta). pos_feedback-mu The function `neg_feedback-decay(t),0,1)` calculates the target confidence level. Here, `conf_i'` represents the target confidence level; `conf_i` represents the confidence level before the update; `eta` represents the positive feedback coefficient, which is a first preset positive number; `pos` represents the positive feedback signal value; `mu` represents the negative feedback coefficient, which is a second preset positive number; `neg_feedback` is the negative feedback signal value; `decay(t)` is the time decay value; and `clip(x,0,1)` restricts `x` to the range of 0 to 1, taking 0 if less than 0 and 1 if greater than 1.

[0111] After determining the target confidence level, the value of the confidence field of the corresponding structured data is modified to the target confidence level. This allows the memory map with the updated confidence level to participate in the scoring and selection in the next round of dialogue. This enables the erroneous profiles to be gradually corrected through the feedback mechanism, reducing the risk of long-term contamination of the accompanying reading teaching content by erroneous information.

[0112] In one or more embodiments of this specification, the server may also manage the companion dialogue context using a combination of short-term original text window, segmented summarization, and structured long-term memory compression. Specifically, the server may maintain an original message queue according to the dialogue time and dynamically select the context content based on the upper limit of the number of lexical units in the large language model before each invocation.

[0113] The server can retain the original text of the most recent N rounds of dialogue through a short-term original text window to maintain the continuity of the dialogue. N can be a preset window size parameter, which can be adaptively adjusted according to the student's grade, question type, student activity level, or the context length of the large language model. The original text of the most recent N rounds of dialogue can be directly input into the large language model as part of the prompt word information without compression.

[0114] Segmented summarization involves extracting semantic units corresponding to earlier historical dialogues from an updated memory graph, and then merging these semantic units to generate a segmented summary, for earlier historical dialogues that exceed the short original text window. The length of the segmented summary is constrained by the maximum number of lexical units in the summary, allowing it to summarize the main content of earlier historical dialogues with fewer lexical units. Segmented summarization can be generated by a large language model, or by a small model, rule templates, or embedding clustering. Segmented summarization can be used as part of the cue word information along with the short original text window and the target structured data. For an example of segmented summarization, semantic units could include: interest preference = liking foxes; causal understanding = taming establishes a unique relationship between the two; emotion understanding = the fox is sad because of separation, etc. Segmented summarization can compress these multiple semantic units into summary information such as "students can understand the relational meaning of 'taming,' recognize the fox's emotions during separation, and have a high interest in the fox character," thereby reducing the number of lexical units.

[0115] Long-term memory can represent the server storing information accumulated across stages and lessons, such as student profiles, reading abilities, interaction preferences, and mastered / unmastered knowledge points, in a structured data format within a memory graph. Target structured data is selected from this graph through comprehensive scoring and lexical budget constraints. The memory graph can be stored in a relational database, vector database, key-value store, or local client-side cache. In practical applications, the target structured data for writing prompt words does not include the original evidence text to avoid the evidence text occupying too many lexical units and failing to provide a complete long-term user profile.

[0116] By combining the three layers of compression mentioned above, the server can control the total number of lexical units while maintaining the continuity of recent conversations, retaining key information from earlier history, and providing personalized references across stages through structured long-term memory.

[0117] In one or more embodiments of this specification, the architecture adopted by the server may include at least one of the following components: a semantic unit extraction component for extracting semantic units from the companion dialogue data; a memory graph writing component for merging structured data into the memory graph, preserving the original evidence text, dialogue turn identifiers, and conflict states during the writing process without directly overwriting old memories; a task-related selection component for calculating relevance based on the task type of the current companion dialogue task, applying different weights to different types of information; a lexical budget optimization component for selecting context fragment combinations with higher returns under lexical budget constraints; and a feedback update component for updating the memory confidence based on feedback signals and time decay values. The functions of the above components may be implemented using large models or classifiers, etc.

[0118] The above one or more steps have the same or similar technical features as the aforementioned method for generating prompt words for generating accompanying reading teaching content. Please refer to the detailed description of the foregoing embodiments, and it will not be repeated here.

[0119] The various technical features in the above embodiments can be combined arbitrarily, as long as there is no conflict or contradiction between the combinations of features. However, due to space limitations, they have not been described one by one. Therefore, the arbitrary combination of various technical features in the above embodiments is also within the scope of this specification.

[0120] Through the above method, in this embodiment of the specification, by converting the accompanying dialogue data into structured data carrying original text evidence, it is possible to support evidence tracing and facilitate manual verification. At the same time, it is possible to select high-value structured datasets as target structured data to fill the prompt word template through comprehensive scoring, thereby improving the density of effective contextual information.

[0121] On the other hand, by including structured data such as short-term original text windows, segmented summaries, and evidence-based long-term memory maps in the prompt word information, the amount of word usage is reduced while maintaining continuity across stages. Furthermore, the confidence level in the memory map can be updated based on feedback signals and time intervals, thereby improving the accuracy of the confidence level in the memory map and preventing occasional answers from easily becoming erroneous long-term memories.

[0122] Corresponding to the above method embodiments, this specification also provides an embodiment of a prompt word generation device for generating accompanying reading teaching content. Figure 2 This specification illustrates a schematic diagram of a prompt word generation device for generating accompanying reading teaching content, according to one embodiment of the present specification. Figure 2 As shown, the device includes: Data acquisition module 202 is used to acquire accompanying reading dialogue data; the accompanying reading dialogue data includes teacher reply text and student answer text; The data conversion module 204 is used to convert the accompanying dialogue data into structured data; the structured data includes data type labels, semantic units, and the original evidence text corresponding to the semantic units; The memory graph update module 206 is used to merge the structured data with the historical memory graph to obtain an updated memory graph; the updated memory graph contains multiple nodes; each node contains one piece of structured data; The scoring calculation module 208 is used to calculate the comprehensive score of each structured data in the updated memory map; the comprehensive score is positively correlated with the task relevance of the structured data to the reading companion dialogue task; Data selection module 210 is used to select target structured data from each of the structured data whose word occupancy meets the upper limit constraint of the word number of the large language model based on the comprehensive score. The prompt word generation module 212 is used to fill the target structured data into the prompt word template to obtain prompt word information; the prompt word information is used to input the large language model to generate the next round of accompanying reading teaching content; the next round of accompanying reading teaching content includes at least one type of data, such as follow-up questions, teacher replies, blackboard writing, and accompanying reading dialogue task reports.

[0123] based on Figure 2 The embodiments of this specification also provide some specific implementations of the method, which will be described below.

[0124] Optionally, the data conversion module may be specifically used for: Semantic extraction is performed on the teacher's response text to obtain the first key-value pair data; the first key-value pair data includes a first data type label, a first semantic unit, and the first evidence original text. Semantic extraction is performed on the student's response text to obtain second key-value pair data; the second key-value pair data includes a second data type label, a second semantic unit, and the original evidence text. The structured data is generated based on the first key-value pair data and the second key-value pair data.

[0125] Optionally, the teacher's response text includes spoken text and whiteboard text; the data conversion module can specifically be used for: Extract the teaching knowledge points and questions from the spoken text; Extract key knowledge points from the blackboard text; The teaching knowledge point information, the question content, and the key knowledge point information are used as the first semantic unit; Determine the first data type tag corresponding to the first semantic unit; The original text fragments from the spoken text and the written text that generate the first semantic unit are used as the original text of the first evidence.

[0126] Optionally, the data conversion module may be specifically used for: Extract key answer information from the student's response text; Based on the student's response text, identify the student's emotional state; The key answer information and the emotional state are used as the second semantic unit; Determine the second data type tag corresponding to the second semantic unit; The original text fragment from the student's answer text that produces the second semantic unit is used as the original text of the second evidence.

[0127] Optionally, the historical memory map includes historical structured data with confidence levels; the memory map update module can specifically be used for: Calculate the similarity between the structured data and the historical structured data; If the similarity is greater than a preset similarity threshold, the confidence of the historical structured data is enhanced to obtain an updated memory map; the updated memory map records the time information for updating the confidence.

[0128] Optionally, the memory map update module can be specifically used for: If the similarity is less than the preset similarity threshold, then it is determined whether there is a conflict between the structured data and the historical structured data; If there is a conflict between the structured data and the historical structured data, the structured data and the historical structured data are merged based on the number of original evidence texts and the confidence level to obtain a merged result. The merge result is updated in the historical memory map to obtain the updated memory map; the updated memory map displays the conflict status at the merge result and records the conflict source information; the conflict source information includes the structured data and the historical structured data.

[0129] Optionally, the memory map update module can be specifically used for: If the structured data does not conflict with the historical structured data, the structured data is merged into the historical memory graph as a new node to obtain the updated memory graph.

[0130] Optionally, the scoring calculation module may be specifically used for: Obtain the confidence level and time recentity of each of the structured data; the time recentity represents the inverse index of the time interval between the generation time of the structured data and the current time. Calculate the task relevance of each of the structured data sets to the accompanying dialogue task; Calculate the cost value of each structured data based on the number of lexical units in each structured data; The confidence level, the time relevance, the task relevance, and the cost value are weighted to obtain a comprehensive score for each of the structured data.

[0131] Optionally, the scoring calculation module may be specifically used for: Determine the task type of the accompanying dialogue task corresponding to the structured data; Based on the task type, determine the weight value of each semantic unit in the structured data; Calculate the relevance of each semantic unit to the subtask of the accompanying dialogue task; The task relevance is obtained by weighting and summing the subtask relevance of each semantic unit using the weight values ​​of each semantic unit.

[0132] Optionally, the data selection module may be specifically used for: The structured data are combined and processed to obtain multiple structured data sets; the number of lexical units contained in each structured data set is less than or equal to the upper limit of the number of lexical units in the large language model; Calculate the sum of the comprehensive scores of the structured data contained in the structured data set to obtain the total score of each structured data set; The structured data set with the highest total score is determined as the target structured data set; the target structured data set contains the target structured data.

[0133] Optionally, the structured data has a confidence level; the device can also be used for: Obtain feedback signals from the structured data; the feedback signals include at least one of positive feedback signals and negative feedback signals; The time decay value is determined based on the time interval between the generation time of the structured data and the current time. The confidence level is updated based on the feedback signal and the time decay value to obtain the target confidence level of the structured data.

[0134] The above is an illustrative scheme of a prompt word generation device for generating accompanying reading teaching content according to this embodiment. It should be noted that the technical solution of this prompt word generation device for generating accompanying reading teaching content belongs to the same concept as the technical solution of the prompt word generation method for generating accompanying reading teaching content described above. Details not described in detail in the technical solution of the prompt word generation device for generating accompanying reading teaching content can be found in the description of the technical solution of the prompt word generation method for generating accompanying reading teaching content described above.

[0135] Figure 3 A structural block diagram of a computing device according to one embodiment of this specification is shown. The components of the computing device 300 include, but are not limited to, a memory 310 and a processor 320. The processor 320 is connected to the memory 310 via a bus 330, and a database 350 is used to store data.

[0136] The computing device 300 also includes an access device 340, which enables the computing device 300 to communicate via one or more networks 360. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 340 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0137] In one embodiment of this specification, the aforementioned components of the computing device 300 and Figure 3 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 3 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0138] The computing device 300 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 300 can also be a mobile or stationary server.

[0139] The processor 320 is used to execute the following computer-executable instructions, which, when executed by the processor, implement the steps of the above-described method for generating prompt words for generating accompanying reading teaching content.

[0140] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the above-described method for generating prompt words for generating accompanying reading teaching content belong to the same concept. For details not described in detail in the technical solution of the computing device, please refer to the description of the technical solution of the above-described method for generating prompt words for generating accompanying reading teaching content.

[0141] An embodiment of this specification also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the above-described method for generating prompt words for generating accompanying reading teaching content.

[0142] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium belongs to the same concept as the technical solution of the above-described method for generating prompt words for generating accompanying reading teaching content. For details not described in detail in the technical solution of the storage medium, please refer to the description of the technical solution of the above-described method for generating prompt words for generating accompanying reading teaching content.

[0143] An embodiment of this specification also provides a computer program product, wherein when the computer program product is executed in a computer, the computer is instructed to perform the steps of the above-described method for generating prompt words for generating accompanying reading teaching content.

[0144] The above is an illustrative scheme of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product and the technical solution of the above-described method for generating prompt words for generating accompanying reading teaching content belong to the same concept. For details not described in detail in the technical solution of the computer program product, please refer to the description of the technical solution of the above-described method for generating prompt words for generating accompanying reading teaching content.

[0145] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0146] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0147] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0148] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0149] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A method for generating prompt words for generating accompanying reading teaching content, characterized in that, include: Obtain the accompanying reading dialogue data; The accompanying reading dialogue data includes teacher's reply text and student's answer text; The accompanying dialogue data is converted into structured data; the structured data includes data type labels, semantic units, and the original evidence text corresponding to the semantic units; The structured data is merged with the historical memory graph to obtain an updated memory graph; the updated memory graph contains multiple nodes. Each of the nodes contains one piece of the structured data; Calculate the overall score of each structured data in the updated memory map; the overall score is positively correlated with the task relevance of the structured data to the reading-aloud dialogue task; Based on the comprehensive score, target structured data whose word count conforms to the upper limit constraint of the word count of the large language model is selected from each of the structured data. The target structured data is populated into the prompt word template to obtain prompt word information; the prompt word information is used to input the large language model to generate the next round of accompanying reading teaching content; The next round of guided reading instruction includes at least one type of data: follow-up questions, teacher responses, blackboard writing, and guided reading dialogue task reports.

2. The method according to claim 1, characterized in that, The process of converting the accompanying dialogue data into structured data specifically includes: Semantic extraction is performed on the teacher's response text to obtain the first key-value pair data; the first key-value pair data includes a first data type label, a first semantic unit, and the first evidence original text. Semantic extraction is performed on the student's response text to obtain second key-value pair data; the second key-value pair data includes a second data type label, a second semantic unit, and the original evidence text. The structured data is generated based on the first key-value pair data and the second key-value pair data.

3. The method according to claim 2, characterized in that, The teacher's response text includes spoken text and written text; the semantic extraction of the teacher's response text specifically includes: Extract the teaching knowledge points and questions from the spoken text; Extract key knowledge points from the blackboard text; The teaching knowledge point information, the question content, and the key knowledge point information are used as the first semantic unit; Determine the first data type tag corresponding to the first semantic unit; The original text fragments from the spoken text and the written text that generate the first semantic unit are used as the original text of the first evidence.

4. The method according to claim 2, characterized in that, The semantic extraction of the student's answer text specifically includes: Extract key answer information from the student's response text; Based on the student's response text, identify the student's emotional state; The key answer information and the emotional state are used as the second semantic unit; Determine the second data type tag corresponding to the second semantic unit; The original text fragment from the student's answer text that produces the second semantic unit is used as the original text of the second evidence.

5. The method according to claim 1, characterized in that, The historical memory map includes structured historical data with confidence levels; merging the structured data with the historical memory map to obtain an updated memory map specifically includes: Calculate the similarity between the structured data and the historical structured data; If the similarity is greater than a preset similarity threshold, the confidence of the historical structured data is enhanced to obtain an updated memory map; the updated memory map records the time information for updating the confidence.

6. The method according to claim 5, characterized in that, The method further includes: If the similarity is less than the preset similarity threshold, then it is determined whether there is a conflict between the structured data and the historical structured data; If there is a conflict between the structured data and the historical structured data, the structured data and the historical structured data are merged based on the number of original evidence texts and the confidence level to obtain a merged result. The merge result is updated in the historical memory map to obtain the updated memory map; the updated memory map displays the conflict status at the merge result and records the conflict source information; the conflict source information includes the structured data and the historical structured data.

7. The method according to claim 6, characterized in that, The method further includes: If the structured data does not conflict with the historical structured data, the structured data is merged into the historical memory graph as a new node to obtain the updated memory graph.

8. The method according to claim 1, characterized in that, The calculation of the comprehensive score for each of the structured data in the updated memory map specifically includes: Obtain the confidence level and time recentity of each of the structured data; the time recentity represents the inverse index of the time interval between the generation time of the structured data and the current time. Calculate the task relevance of each of the structured data sets to the accompanying dialogue task; Calculate the cost value of each structured data based on the number of lexical units in each structured data; The confidence level, the time relevance, the task relevance, and the cost value are weighted to obtain a comprehensive score for each of the structured data.

9. The method according to claim 8, characterized in that, The calculation of the task relevance between each of the structured data and the accompanying dialogue task specifically includes: Determine the task type of the accompanying dialogue task corresponding to the structured data; Based on the task type, determine the weight value of each semantic unit in the structured data; Calculate the relevance of each semantic unit to the subtask of the accompanying dialogue task; The task relevance is obtained by weighting and summing the subtask relevance of each semantic unit using the weight values ​​of each semantic unit.

10. The method according to claim 1, characterized in that, The step of selecting target structured data from the various structured data sets based on the comprehensive score, where the word count conforms to the upper limit constraint of the word count in the large language model, specifically includes: The structured data are combined and processed to obtain multiple structured data sets; the number of lexical units contained in each structured data set is less than or equal to the upper limit of the number of lexical units in the large language model; Calculate the sum of the comprehensive scores of the structured data contained in the structured data set to obtain the total score of each structured data set; The structured data set with the highest total score is determined as the target structured data set; the target structured data set contains the target structured data.

11. The method according to claim 1, characterized in that, The structured data has a confidence level; the method further includes: Obtain feedback signals from the structured data; the feedback signals include at least one of positive feedback signals and negative feedback signals; The time decay value is determined based on the time interval between the generation time of the structured data and the current time. The confidence level is updated based on the feedback signal and the time decay value to obtain the target confidence level of the structured data.

12. A device for generating prompt words for generating accompanying reading teaching content, characterized in that, include: The data acquisition module is used to acquire the accompanying dialogue data; The accompanying reading dialogue data includes teacher's reply text and student's answer text; The data conversion module is used to convert the accompanying dialogue data into structured data; the structured data includes data type labels, semantic units, and the original evidence text corresponding to the semantic units; The memory graph update module is used to merge the structured data with the historical memory graph to obtain an updated memory graph; the updated memory graph contains multiple nodes. Each of the nodes contains one piece of the structured data; The scoring calculation module is used to calculate the comprehensive score of each structured data in the updated memory map; the comprehensive score is positively correlated with the task relevance of the structured data to the reading companion dialogue task; The data selection module is used to select target structured data from the various structured data based on the comprehensive score, whose word occupancy meets the upper limit constraint of the word number of the large language model. The prompt word generation module is used to populate the target structured data into the prompt word template to obtain prompt word information; The prompt word information is used to input the large language model to generate the next round of accompanying reading teaching content; The next round of guided reading instruction includes at least one type of data: follow-up questions, teacher responses, blackboard writing, and guided reading dialogue task reports.

13. A computing device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the prompt word generation method for generating accompanying reading teaching content as described in any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that, It stores computer-executable instructions, which, when executed by a processor, implement the steps of the prompt word generation method for generating accompanying reading teaching content as described in any one of claims 1 to 11.

15. A computer program product, characterized in that, It includes a computer program / instruction that, when executed by a processor, implements the steps of the prompt word generation method for generating accompanying reading teaching content as described in any one of claims 1 to 11.

Images