An AI-assisted reading method and system
By creating an auxiliary knowledge base and linking multi-source data, and by optimizing the AI model in conjunction with human thinking, the problem of insufficient deep understanding and interpretation capabilities of AI-assisted reading technology has been solved, resulting in a reading experience that better meets human expectations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA DATANG TECH & ECONOMY RES INST CO LTD
- Filing Date
- 2025-12-11
- Publication Date
- 2026-07-31
AI Technical Summary
Existing AI-assisted reading technologies are unable to engage in deep thinking and interpretation; they can only analyze and extract information based on surface details, lacking the ability to gain a deep understanding.
Create an auxiliary knowledge base, establish connections between data and articles, gain in-depth understanding through multi-source data, and optimize the AI model by combining human thinking to make it closer to human thought patterns.
It achieves a deep understanding and profound thinking about the article, outputs interpretation results that meet the expectations of human readers, and provides interpretations that go beyond the surface.
Smart Images
Figure CN121724010B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to an AI-assisted reading method and system. Background Technology
[0002] In recent years, AI-assisted reading applications have become increasingly common. Currently, AI technology, through big data analysis and machine learning algorithms, can recommend personalized reading content to users based on their reading history, preferences, and behavioral data. During the reading process, it can also perform functions such as information extraction and text analysis, helping users quickly understand the main idea and key information of books. However, AI-assisted reading also has some significant shortcomings, namely, its limited comprehension ability. It analyzes and extracts information based more on surface information than on in-depth thinking and interpretation. Therefore, researchers in this field still need to further optimize the assisted reading capabilities of AI. Summary of the Invention
[0003] This invention provides an AI-assisted reading method, comprising: Step 1: Create an auxiliary knowledge base and establish the relationship between the data in the base and the articles; Step 2: Based on the relationship between the data in the auxiliary knowledge base and the articles, establish a training dataset; Step 3: Use the training dataset to incrementally train the AI-assisted reading model to improve the model's ability to deeply understand the sentences in the article; Step 4: Conduct a recognition test on the trained AI-assisted reading model, and optimize the model's understanding focus based on the test results to make it closer to human thinking patterns. Step 5: The optimized AI-assisted reading model provides services to users through the AI interactive window.
[0004] The AI-assisted reading method described above involves creating an auxiliary knowledge base and establishing relationships between the data in the base and the articles. This process is divided into the following sub-steps: Annotation data, creative background data, and author's life experience data of articles obtained from different channels are stored in an auxiliary knowledge base; The article is processed into lexical units, and the article's lexical units are associated with annotation data in the auxiliary knowledge base; Based on the background data of the creation and the author's life experience, annotation data that does not match the emotional tone of the article is filtered out for word units.
[0005] The AI-assisted reading method described above, which filters out annotation data that does not match the emotional tone of the text based on the author's creative background data and life experience data, is specifically divided into the following sub-steps: Construct an emotion dictionary; Extract keywords from the author's life experiences prior to the year the article was written, keywords from the background data of the article's creation, and keywords from each footnote; For each extracted keyword, or each word element in the article, the word with the highest similarity is retrieved from the sentiment dictionary, and a sentiment index is assigned to it based on the retrieved word. Then, the sentiment color deviation between each annotation data and the current article is calculated based on the assigned sentiment index. Remove annotation data for each morpheme that has a sentiment bias exceeding a preset threshold.
[0006] The AI-assisted reading method described above, which incrementally trains the AI-assisted reading model using a training dataset, consists of the following sub-steps: Design a loss function for an AI-assisted reading model; The AI-assisted reading model is trained using a training dataset, and the model parameters are fine-tuned during the training process based on the return value of the loss function. The output accuracy of the AI-assisted reading model is verified by sampling, and the training weights of the model are adjusted based on the verification results.
[0007] The AI-assisted reading method described above involves conducting a recognition test on the trained AI-assisted reading model and optimizing the model's comprehension focus based on the test results. This process is divided into the following sub-steps: Collect test results on human readers' level of agreement with the interpretation results of the AI-assisted reading model; The processing weights of various data features in the AI-assisted reading model are adjusted based on the collected recognition test results; Repeat the above testing and adjustment steps until the sense of identification score reaches the predetermined standard.
[0008] The present invention also provides an AI-assisted reading system, comprising: an auxiliary knowledge base creation module, a training dataset preparation module, an AI-assisted reading model training module, an AI-assisted reading model optimization module, and an AI interaction module; The auxiliary knowledge base creation module is used to create an auxiliary knowledge base and establish the relationship between the data in the base and the articles. The training dataset preparation module is used to create a training dataset based on the relationship between data in the auxiliary knowledge base and articles; The AI-assisted reading model training module is used to incrementally train the AI-assisted reading model using a training dataset, thereby improving the model's ability to deeply understand sentences in articles. The AI-assisted reading model optimization module is used to conduct a recognition test on the trained AI-assisted reading model and optimize the model's understanding focus based on the test results, making it closer to human thinking patterns. The AI interaction module is used to provide services to users through an AI interaction window using an optimized AI-assisted reading model.
[0009] The beneficial effects achieved by this invention are as follows: it enables a deeper understanding of articles by leveraging multi-source data, providing users with interpretations that go beyond the surface and inspiring them to think deeply about the articles; it also optimizes existing AI-assisted reading models by incorporating human thinking, enabling them to output interpretation results that better meet the expectations of human readers. Attached Figure Description
[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0011] Figure 1 This is a flowchart of an AI-assisted reading method provided in Embodiment 1 of this application; Figure 2 This is a schematic diagram of an AI-assisted reading system provided in Embodiment 2 of this application. Detailed Implementation
[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0013] Example 1 like Figure 1 As shown, Embodiment 1 of this application provides an AI-assisted reading method, including: Step S10: Create an auxiliary knowledge base and establish the relationship between the data in the base and the articles; The auxiliary knowledge base draws its data from various sources, including academic forums, journals, websites, and databases. This auxiliary knowledge helps AI-assisted reading models discover implicit information and deeper meanings within articles. Of course, the knowledge in the auxiliary knowledge base also needs to be updated according to the articles being read in the reading application. Specifically: Step S11: Obtain the article's annotation data, creative background data, and the author's life experience from different channels and store the data in the auxiliary knowledge base; The annotation data of an article refers to the further explanations, reflections, or expansions made by officials or readers regarding the words and sentences in the article, which belongs to the category of reading experience data; the creative background refers to the cultural and social context in which the article is situated, which can be combined with the author's life experience data to understand the emotional tone of the article; the author's life experience data refers to the experience data of the author before the writing period, such as changes in relatives and friends, emotional turning points, and life setbacks that directly affected the expression of emotions, which is obtained through the author's autobiography or biographical documents or official entries; these three types of data are the foundation of human interpretation of articles and are also the key to AI learning interpretation thinking.
[0014] Step S12: Perform lexicalization on the article and associate the lexical units in the article with the annotation data in the auxiliary knowledge base; Natural language processing techniques are used to process the article into a sequence of terms. Then, the annotation data for each term is retrieved from the auxiliary knowledge base. An association table is created between the terms and the retrieved annotation data, with a one-to-many relationship between the terms and the annotation data.
[0015] Step S13: Based on the creative background data and the author's life experience data, filter out annotation data that does not match the emotional tone of the article for the word units; To understand the emotional tone of an article, one must not only refer to the original text but also consider its historical context and the author's life experiences. This allows for a better understanding of the author's views and attitudes within a specific social environment. Specifically: Step S131: Construct an emotion dictionary; First, several common emotion words were assigned initial sentiment indices, such as happiness, fear, sadness, pain, and anger, with sentiment indices of 0.2, 0.4, 0.6, 0.7, and 0.9, respectively. Then, pre-trained language models (such as GPT-3, LLaMA 2 / 3, Doubao, etc.) are used to retrieve words that are semantically similar to these common words, and sentiment indices are assigned based on similarity (calculated based on cosine similarity). The higher the similarity, the closer the assigned sentiment index is to the common words. For example, the similarity between "happy" and "joyful" is 0.95, so the sentiment index of "joyful" is assigned as 0.2 multiplied by 0.95. Then, the words that have already been assigned sentiment indices are used to search and assign values to other words until the vocabulary of the sentiment dictionary reaches the required size.
[0016] Step S132: Extract keywords from the author's life experience data prior to the year the article was written, keywords from the background data of the article's creation, and keywords from each annotation data; Keyword extraction is still achieved through natural language processing (NLP) technology, which can remove prepositions, auxiliary words, and useless words from the text, retaining only keywords with actual meaning.
[0017] Step S133: For each extracted keyword or each word element in the article, query the word with the highest similarity from the sentiment dictionary, assign a sentiment index to the word based on the query, and then calculate the sentiment color deviation between each annotation data and the current article based on the assigned sentiment index. The formula for calculating emotional color deviation is expressed as follows:
[0018] in The result of calculating the sentiment bias between the i-th annotation and the current article is given. These are the creative background, the author's life experiences, and the emotional discrimination coefficient of the original text. The sentiment index of the k1th creative background keyword. This represents the sentiment index of keywords related to the k2th author's life experiences. This represents the sentiment index of the k3rd word in the article, where k1 takes values from 1 to... , The total number of background keywords for the current article, k2 takes a value of 1~ , k3 represents the number of keywords related to the author's life experiences in the current article, with a value of 1~ , For the term associated with the i-th annotation, the number of terms contained in the original text paragraph to which it belongs. This represents the sentiment index of the k0th annotated keyword, where k0 ranges from 1 to... , The number of keywords contained in the i-th annotation data.
[0019] Step S134: Remove annotation data for each word whose emotional color deviation exceeds a preset threshold; Remove annotation data from the association table that deviates from the emotional tone of the current article by more than a preset threshold. The preset threshold is user-defined data.
[0020] Step S20: Based on the relationship between the data in the auxiliary knowledge base and the articles, establish a training dataset; After the previous step, the annotation data in the auxiliary knowledge base has been associated with the words in the article. In this step, it is necessary to organize these associated data into a training dataset. The training dataset contains positive samples and negative samples. For an input word, the positive sample is the annotation data with an association relationship, and the negative sample is the annotation data without an association relationship.
[0021] Step S30: Use the training dataset to incrementally train the AI-assisted reading model to improve the model's ability to deeply understand the sentences in the article; Existing AI-assisted reading models only refer to the original text and can only query and summarize information within the original text. Now, the model needs to be able to combine multi-source text annotation data for in-depth interpretation. Specifically: Step S31: Design a loss function for the AI-assisted reading model; The loss function of an AI-assisted reading model must allow the output to converge towards the feature space of positive samples, enabling the model to output correct annotation data based on the input lexical units. A well-designed loss function L() is expressed as: Where x is the input word, and Y represents the positive sample set of word x. This represents the negative sample set of word x. This represents the embedding vector generated by the AI-assisted reading model for word x. Let i be the i-th positive sample in the positive sample set Y. This indicates that the AI-assisted reading model is The generated embedding vector, i takes values from 1 to n, where n is the length of the positive sample set Y. negative sample set The j-th negative sample in the dataset, This indicates that the AI-assisted reading model is The generated embedding vector, j takes values from 1 to m, where m is the negative sample set. Length, This is an adjustable parameter used to control the smoothness of the function; Step S32: Train the AI-assisted reading model using the training dataset, and fine-tune the model parameters based on the return value of the loss function during the training process; During training, observe the change curve of the loss function and fine-tune the model parameters until the change curve no longer shows a significant downward trend, then stop training to avoid overfitting of the model.
[0022] Step S33: Sample and verify the output accuracy of the AI-assisted reading model, and adjust the training weights of the model based on the verification results; Extract sentences from different articles for interpretation and verify whether the model output is all associated annotation data. If not, increase the weight of the training data of that article to improve the interpretation accuracy and generalization ability of the model. Repeat the verification and adjustment until the accuracy of the sampling verification results reaches 100%.
[0023] Step S40: Conduct an identification test on the trained AI-assisted reading model, and optimize the model's understanding focus based on the test results to make it closer to human thinking patterns; Currently, AI-assisted reading models interpret texts in a somewhat random way. This means that the annotations output by the model may not reflect the key information that human readers need. Therefore, this step aims to make the model's output more focused on aspects that humans deem important. Specifically: Step S41: Collect test results on human readers' acceptance of the interpretation results of the AI-assisted reading model; The identification test involves human readers providing feedback on the interpretation results generated by an AI-assisted reading model. The test questions are randomly selected words and phrases from the article. If the human reader agrees with the generated interpretation results, they click the "agree" option; otherwise, they click the "disagree" option. This feedback data is collected as the test results of the identification test.
[0024] Step S42: Adjust the processing weights of various data features by the AI-assisted reading model based on the collected recognition test results; The weighting of various data features by an AI-assisted reading model essentially reflects the model's emphasis in understanding the text. For example, if the model assigns a higher weight to data feature A than to B, the resulting interpretation will naturally be closer to the spatial vector of data feature B. Superficially, it will appear that the model understands the content with data feature B more deeply and values it more. Substituting the collected consensus test results into the weighting formula: In the process, the AI-assisted reading model is fine-tuned based on the calculation results, among which... This represents the new processing weights for the t-th data feature in the AI-assisted reading model. The weights for the AI-assisted reading model in processing the original features of the t-th data item are: To handle the adjustment factor for the weights, Let be the number of interpretation results with the t-th data feature. The number of interpretations with the t-th data feature that are accepted by human readers. This represents the total number of interpretation results generated by the model in this test.
[0025] Step S43: Repeat the above testing and adjustment steps until the sense of identification score reaches the predetermined standard; The identification score is also calculated based on the identification test results, and its calculation formula is as follows: Where P is the calculated score of identification. This represents the total number of interpretation results generated by the model in this test. The number of interpretations that resonate with human readers.
[0026] The predetermined standard is a custom setting; in this embodiment, the standard is set to 8 points.
[0027] Step S50: The optimized AI-assisted reading model provides services to users through the AI interactive window; The AI interaction window is an indispensable part of the interaction between readers and AI. Users input questions into the window, which performs necessary formatting of the question text before inputting it into the AI-assisted reading model. The output is then presented to the user in natural and fluent language. At the same time, the AI interaction window is also responsible for surveying readers' acceptance of the output results. The survey method is the same as that used in the acceptance test. Subsequently, the AI-assisted reading model is regularly optimized and adjusted based on these survey data.
[0028] Example 2 like Figure 2 As shown, Embodiment 2 of this application provides an AI-assisted reading system, including: an auxiliary knowledge base creation module 21, a training dataset preparation module 22, an AI-assisted reading model training module 23, an AI-assisted reading model optimization module 24, and an AI interaction module 25; The auxiliary knowledge base creation module 21 is used to create an auxiliary knowledge base and establish the relationship between the data in the base and the articles; specifically, it includes: an auxiliary knowledge import submodule, a relationship table creation submodule, and a relationship data filtering submodule. 1. The auxiliary knowledge import submodule is used to import annotation data, creative background data, and author's life experience data from different channels into the auxiliary knowledge base; The annotation data of an article refers to the further explanations, reflections, or expansions made by officials or readers regarding the words and sentences in the article, which belong to the data of reading experience; the creative background refers to the cultural and social background in which the article is situated, and can be combined with the author's life experience data to understand the emotional tone of the article; these three types of data are the foundation of human interpretation of articles and also the key to AI learning interpretation thinking.
[0029] 2. The association table creation submodule is used to associate the words in the article with the annotation data in the auxiliary knowledge base; Natural language processing techniques are used to process the article into a sequence of terms. Then, the annotation data for each term is retrieved from the auxiliary knowledge base. An association table is created between the terms and the retrieved annotation data, with a one-to-many relationship between the terms and the annotation data.
[0030] 3. The associated data filtering submodule is used to filter out annotation data that does not match the emotional tone of the article based on the creative background data and the author's life experience data; To understand the emotional tone of an article, one must not only refer to the original text but also consider its historical context and the author's life experiences. This allows for a better understanding of the author's views and attitudes within a specific social environment. Specifically: I. Construct an emotional dictionary; First, several common emotion words were assigned initial sentiment indices, such as happiness, fear, sadness, pain, and anger, with sentiment indices of 0.2, 0.4, 0.6, 0.7, and 0.9, respectively. Then, pre-trained language models (such as GPT-3, LLaMA 2 / 3, Doubao, etc.) are used to retrieve words that are semantically similar to these common words, and sentiment indices are assigned based on similarity (calculated based on cosine similarity). The higher the similarity, the closer the assigned sentiment index is to the common words. For example, the similarity between "happy" and "joyful" is 0.95, so the sentiment index of "joyful" is assigned as 0.2 multiplied by 0.95. Then, words with pre-assigned sentiment indices are used to match and assign values to other words until the vocabulary of the sentiment dictionary reaches the required size.
[0031] II. Extract keywords from the author's life experience data prior to the year the article was written, keywords from the background data of the article's creation, and keywords from each footnote; Keyword extraction is still achieved through natural language processing (NLP) technology, which can remove prepositions, auxiliary words, and useless words from the text, retaining only keywords with actual meaning.
[0032] III. For each extracted keyword, or each word element in the article, retrieve the word with the highest similarity from the sentiment dictionary, assign a sentiment index to the retrieved word, and then calculate the sentiment color deviation between each annotation data and the current article based on the assigned sentiment index. The formula for calculating emotional color deviation is expressed as follows: ,in The result of calculating the sentiment bias between the i-th annotation and the current article is given. These are the creative background, the author's life experiences, and the emotional discrimination coefficient of the original text. The sentiment index of the k1th creative background keyword. This represents the sentiment index of keywords related to the k2th author's life experiences. This represents the sentiment index of the k3rd word in the article, where k1 takes values from 1 to... , The total number of background keywords for the current article, k2 takes a value of 1~ , k3 represents the number of keywords related to the author's life experiences in the current article, with a value of 1~ , For the term associated with the i-th annotation, the number of terms contained in the original text paragraph to which it belongs. This represents the sentiment index of the k0th annotated keyword, where k0 ranges from 1 to... , The number of keywords contained in the i-th annotation data.
[0033] IV. Remove annotation data for each lexical unit that has a sentiment bias exceeding a preset threshold; Remove annotation data from the association table that deviates from the emotional tone of the current article by more than a preset threshold. The preset threshold is user-defined data.
[0034] Training dataset preparation module 22 is used to build a training dataset based on the relationship between data in the auxiliary knowledge base and articles; After the previous step, the annotation data in the auxiliary knowledge base has been associated with the words in the article. In this step, it is necessary to organize these associated data into a training dataset. The training dataset contains positive samples and negative samples. For an input word, the positive sample is the annotation data with an association relationship, and the negative sample is the annotation data without an association relationship.
[0035] The AI-assisted reading model training module 23 is used to incrementally train the AI-assisted reading model using the training dataset to improve the model's ability to deeply understand sentences in the text; specifically, it includes: a loss function design submodule, a model parameter fine-tuning submodule, and a training weight adjustment submodule; 1. The loss function design submodule is used to design a loss function for the AI-assisted reading model; The loss function of an AI-assisted reading model must allow the output to converge towards the feature space of positive samples, enabling the model to output correct annotation data based on the input lexical units. A well-designed loss function L() is expressed as: Where x is the input word, and Y represents the positive sample set of word x. This represents the negative sample set of word x. This represents the embedding vector generated by the AI-assisted reading model for word x. Let i be the i-th positive sample in the positive sample set Y. This indicates that the AI-assisted reading model is The generated embedding vector, i takes values from 1 to n, where n is the length of the positive sample set Y. negative sample set The j-th negative sample in the dataset, This indicates that the AI-assisted reading model is The generated embedding vector, j takes values from 1 to m, where m is the negative sample set. Length, This is an adjustable parameter used to control the smoothness of the function.
[0036] 2. The model parameter fine-tuning submodule is used to train the AI-assisted reading model using the training dataset and to fine-tune the model parameters based on the return value of the loss function during the training process. During training, observe the change curve of the loss function and fine-tune the model parameters until the change curve no longer shows a significant downward trend, then stop training to avoid overfitting of the model.
[0037] 3. The training weight adjustment submodule is used to sample and verify the output accuracy of the AI-assisted reading model, and adjust the training weights of the model based on the verification results. Extract sentences from different articles for interpretation and verify whether the model output is all associated annotation data. If not, increase the weight of the training data of that article to improve the interpretation accuracy and generalization ability of the model. Repeat the verification and adjustment until the accuracy of the sampling verification results reaches 100%.
[0038] The AI-assisted reading model optimization module 24 is used to conduct a recognition test on the trained AI-assisted reading model and optimize the model's understanding focus based on the test results, making it closer to human thinking patterns; specifically, it includes: a recognition test result collection submodule, a weight adjustment processing submodule, and an adjustment result verification submodule; 1. The submodule for collecting identification test results is used to collect identification test results of human readers on the interpretation results of the AI-assisted reading model; The identification test involves human readers providing feedback on the interpretation results generated by an AI-assisted reading model. The test questions are randomly selected words and phrases from the article. If the human reader agrees with the generated interpretation results, they click the "agree" option; otherwise, they click the "disagree" option. This feedback data is collected as the test results of the identification test.
[0039] 2. The weight adjustment submodule is used to adjust the processing weights of various data features by the AI-assisted reading model based on the collected recognition test results; The weighting of various data features in an AI-assisted reading model essentially reflects the model's emphasis in understanding the text. For example, if the model assigns a higher weight to data feature A than to B, the resulting interpretation will naturally be closer to the spatial vector of data feature B. Superficially, it will appear that the model understands the content with data feature B more deeply and values it more. The collected consensus test results are then substituted into the weighting adjustment formula. In the process, the AI-assisted reading model is fine-tuned based on the calculation results, among which... This represents the new processing weights for the t-th data feature in the AI-assisted reading model. The weights for the AI-assisted reading model in processing the original features of the t-th data item are: To handle the adjustment factor for the weights, Let be the number of interpretation results with the t-th data feature. The number of interpretations with the t-th data feature that are accepted by human readers. This represents the total number of interpretation results generated by the model in this test.
[0040] 3. Adjust the result verification submodule to verify whether the recognition score of the AI-assisted reading model meets the predetermined standard; The identification score is also calculated based on the identification test results, and its calculation formula is as follows: Where P is the calculated score of identification. This represents the total number of interpretation results generated by the model in this test. The number of interpretations that resonate with human readers.
[0041] The predetermined standard is a custom setting; in this embodiment, the standard is set to 8 points.
[0042] AI interaction module 25 is used to provide services to users through an AI interaction window using an optimized AI-assisted reading model; The AI interaction window is an indispensable part of the interaction between readers and AI. Users input questions into the window, which performs necessary formatting of the question text before inputting it into the AI-assisted reading model. The output is then presented to the user in natural and fluent language. At the same time, the AI interaction window is also responsible for surveying readers' acceptance of the output results. The survey method is the same as that used in the acceptance test. Subsequently, the AI-assisted reading model is regularly optimized and adjusted based on these survey data.
[0043] Corresponding to the above embodiments, the present invention provides a computer storage medium, including: at least one memory and at least one processor; The memory is used to store one or more program instructions; A processor is used to run one or more program instructions to execute an AI-assisted reading method.
[0044] Corresponding to the above embodiments, this embodiment of the invention provides a computer-readable storage medium containing one or more program instructions, which are executed by a processor to provide an AI-assisted reading method.
[0045] The embodiments disclosed in this invention provide a computer-readable storage medium storing computer program instructions, which, when executed on a computer, cause the computer to perform the aforementioned AI-assisted reading method.
[0046] In this embodiment of the invention, the processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0047] The various methods, steps, and logic diagrams disclosed in the embodiments of this invention can be implemented or executed. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The processor reads information from the storage medium and, in conjunction with its hardware, completes the steps of the above methods.
[0048] The storage medium can be memory, such as volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.
[0049] Among them, non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory.
[0050] Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (Synchlink DRAM, SLDRAM), and direct memory bus RAM (DRRAM).
[0051] The storage media described in the embodiments of the present invention are intended to include, but are not limited to, these and any other suitable types of memory.
[0052] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using a combination of hardware and software. When applied as software, the corresponding functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of computer programs from one place to another. Storage media can be any available medium that can be accessed by a general-purpose or special-purpose computer.
[0053] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.
Claims
1. An AI-assisted reading method, characterized in that, include: Step 1: Create an auxiliary knowledge base and establish the relationship between the data in the base and the articles; Specifically, it includes: Annotation data, creative background data, and author's life experience data of articles obtained from different channels are stored in an auxiliary knowledge base; The article is processed into lexical units, and the lexical units in the article are associated with the annotation data in the auxiliary knowledge base; Based on the background data of the creation and the author's life experience, annotation data that does not match the emotional tone of the article is filtered out for word elements. This is specifically divided into the following sub-steps: Construct an emotion dictionary; Extract keywords from the author's life experience data prior to the year the article was written, keywords from the background data of the article's creation, and keywords from each footnote; For each extracted keyword, or each word element in the article, the word with the highest similarity is retrieved from the sentiment dictionary, and a sentiment index is assigned to it based on the retrieved word. Then, the sentiment color deviation between each annotation data and the current article is calculated based on the assigned sentiment index. Remove annotation data with a sentiment bias exceeding a preset threshold from each lexical unit; Step 2: Based on the relationship between the data in the auxiliary knowledge base and the articles, establish a training dataset; Step 3: Use the training dataset to incrementally train the AI-assisted reading model to improve the model's ability to deeply understand the sentences in the article; Step 4: Conduct a recognition test on the trained AI-assisted reading model, and optimize the model's understanding focus based on the test results to make it closer to human thinking patterns. Step 5: The optimized AI-assisted reading model provides services to users through the AI interactive window.
2. The AI-assisted reading method according to claim 1, characterized in that, Incremental training of the AI-assisted reading model using the training dataset involves the following sub-steps: Design a loss function for an AI-assisted reading model; The AI-assisted reading model is trained using a training dataset, and the model parameters are fine-tuned during the training process based on the return value of the loss function. The output accuracy of the AI-assisted reading model is verified by sampling, and the training weights of the model are adjusted based on the verification results.
3. The AI-assisted reading method according to claim 1, characterized in that, The trained AI-assisted reading model undergoes a recognition test, and the model's comprehension focus is optimized based on the test results. This process involves the following sub-steps: Collect test results on human readers' level of agreement with the interpretation results of the AI-assisted reading model; The processing weights of various data features in the AI-assisted reading model are adjusted based on the collected recognition test results; Repeat the above testing and adjustment steps until the sense of identification score reaches the predetermined standard.
4. An AI-assisted reading system, characterized in that, include: The module includes: knowledge base creation module, training dataset preparation module, AI-assisted reading model training module, AI-assisted reading model optimization module, and AI interaction module. The auxiliary knowledge base creation module is used to create an auxiliary knowledge base and establish the relationship between the data in the base and the articles. The auxiliary knowledge base creation module specifically includes: The auxiliary knowledge import submodule is used to import annotation data, creative background data, and author's life experience data from different channels into the auxiliary knowledge base; The association table creation submodule is used to associate the lexical elements in the article with the annotation data in the auxiliary knowledge base; The related data filtering submodule is used to filter out annotation data that does not match the emotional tone of the article based on the author's background data and biographical data. Specifically, it is used for: Construct an emotion dictionary; Extract keywords from the author's life experience data prior to the year the article was written, keywords from the background data of the article's creation, and keywords from each footnote; For each extracted keyword, or each word element in the article, the word with the highest similarity is retrieved from the sentiment dictionary, and a sentiment index is assigned to it based on the retrieved word. Then, the sentiment color deviation between each annotation data and the current article is calculated based on the assigned sentiment index. Remove annotation data with a sentiment bias exceeding a preset threshold from each lexical unit; The training dataset preparation module is used to create a training dataset based on the relationship between data in the auxiliary knowledge base and articles; The AI-assisted reading model training module is used to incrementally train the AI-assisted reading model using a training dataset, thereby improving the model's ability to deeply understand sentences in articles. The AI-assisted reading model optimization module is used to conduct a recognition test on the trained AI-assisted reading model and optimize the model's understanding focus based on the test results, making it closer to human thinking patterns. The AI interaction module is used to provide services to users through an AI interaction window using an optimized AI-assisted reading model.
5. The AI-assisted reading system according to claim 4, characterized in that, The AI-assisted reading model training module specifically includes: The loss function design submodule is used to design a loss function for the AI-assisted reading model. The model parameter fine-tuning submodule is used to train the AI-assisted reading model using the training dataset and fine-tune the model parameters based on the return value of the loss function during the training process. The training weight adjustment submodule is used to sample and verify the output accuracy of the AI-assisted reading model, and adjust the model's training weights based on the verification results.
6. The AI-assisted reading system according to claim 4, characterized in that, The AI-assisted reading model optimization module specifically includes: The submodule for collecting identification test results is used to collect identification test results from human readers regarding the interpretation results of the AI-assisted reading model. The weight adjustment submodule is used to adjust the processing weights of various data features by the AI-assisted reading model based on the collected recognition test results. Adjust the result verification submodule to verify whether the recognition score of the AI-assisted reading model meets the predetermined standard.
7. A computer storage medium, characterized in that, include: At least one memory and at least one processor; Memory, used to store one or more program instructions; A processor for running one or more program instructions to perform an AI-assisted reading method as described in any one of claims 1-3.