User experience design method based on natural language processing technology

Abstract

The invention discloses a user experience design method based on a natural language processing technology, relates to the technical field of natural language processing technologies and user experience design, and adopts the natural language processing technology to intelligently collect and analyze user feedback information. The method is constructed on the basis of advanced natural language understanding and sentiment analysis technologies, and an intelligent feedback management system realizes efficient mining and deep insight of mass text data and covers deep information extraction, sentiment recognition and quantification and automatic report generation; according to the application of the natural language processing technology in the field of user feedback collection and analysis, not only is the crossing from passive receiving and manual processing to active detection and intelligent analysis realized, but also the user experience is improved through seamless voice interaction and accurate path guidance. The user experience and the response speed of the enterprise to the user demand are greatly improved, and the intelligent upgrading of the feedback system is really realized.

Description

Technical Field

[0001] The present invention relates to the fields of natural language processing technology and user experience design technology, and in particular to a user experience design method based on natural language processing technology. Background Art

[0002] Traditional user experience design methods usually follow five stages: research, definition, development, evaluation, and release. Various methods and tools can be used to complete the work at each stage. Common methods include user interviews, focus groups, questionnaires, usability testing, and user testing.

[0003] Problems with traditional user experience design methods: 1. Difficult to adapt to rapidly changing needs. 2. Not user-centric enough. 3. Inefficient. 4. Difficult to quantify.

[0004] The reasons for these problems are: 1. The limitations of the method itself. Too much focus on user needs, while ignoring the user's goals and environment; it is difficult to quantify, making it difficult to evaluate the design effect. 2. Challenges brought by the development of the times. Market demand changes rapidly, user needs are diversified and personalized; product updates and iterations are fast.

[0005] The rapid development of natural language processing (NLP) technology has brought an indispensable driving force to user experience design. 1. It can be used to analyze user feedback and comments and automatically identify changes in user needs. For example, sentiment analysis technology can be used to identify user satisfaction with product features and discover potential problems. This information can help designers quickly understand changes in user needs and adjust design plans in a timely manner. 2. It can be used to build user portraits and understand user behavior and preferences. For example, text mining technology can be used to extract keywords and topics from user comments and analyze user behavior data. This information can help designers understand users more deeply and design products that better meet user needs. 3. It can be used to automate some user experience design tasks. Extract key information from text, such as user needs, function descriptions, etc. Generate user documents, product descriptions, etc. Analyze user test results and identify problems encountered by users. These automated tasks can help improve the efficiency of user experience work. 4. It can be used to quantify the effect of user experience design. Sentiment analysis technology can be used to evaluate user satisfaction with product design and generate user experience indicators.

[0006] With the popularity of online education, users have higher requirements for the interactivity and personalized experience of the platform. In order to meet these needs, the platform needs to be able to collect and analyze user feedback to continuously optimize the user experience. Summary of the invention

[0007] The technical problem to be solved by the present invention is to provide a user experience design method based on natural language processing technology in response to the deficiencies of the background technology.

[0008] The present invention adopts the following technical solutions to solve the above technical problems:

[0009] A user experience design method based on natural language processing technology, specifically comprising the following steps;

[0010] Step 1, data collection: using natural language processing technology to intelligently collect and analyze user feedback information;

[0011] Step 2, Natural Language Processing: Built on advanced natural language understanding and sentiment analysis technologies, the intelligent feedback management system achieves efficient mining and deep insights into massive text data, covering deep information extraction, sentiment recognition and quantification, and automated report generation;

[0012] Among them, deep information extraction: using natural language understanding technology to deeply analyze the massive amount of text data collected; through precise analysis of vocabulary, grammar, semantics and contextual relationships, accurately extract key issues, specific opinions, demand suggestions and even potential pain points in user feedback to form a structured information database;

[0013] Emotion recognition and quantification: The system is embedded with an emotion analysis module, which can accurately identify and quantify the emotional color in user feedback. By analyzing language features such as word selection, sentence structure, and tone intensity, the system can not only identify the user's emotional state, including satisfaction, dissatisfaction, confusion, and excitement, but also quantify its intensity, giving the feedback an emotional temperature.

[0014] Automatic report generation: Based on the in-depth analysis results, the system can automatically generate detailed user feedback reports; the reports not only include statistical overviews of various feedback topics and sentiment distribution trends, but may also provide targeted improvement suggestions and strategic inspirations;

[0015] Step 3, user experience design, specifically includes analyzing user feedback, user interviews and analysis, user behavior testing, and drawing user journey maps:

[0016] Among them, analyzing user feedback: conducting in-depth analysis of user feedback in an all-round and multi-dimensional manner, covering various communication channels including but not limited to comments, messages, phone calls, etc., to reveal users' praise for the product and their demands for improvement; using advanced text analysis technology to conduct semantic mining and sentiment analysis on massive comments and messages, and extracting detailed insights into users' preferences for product features, functional evaluation, and user experience;

[0017] User interview and analysis: Through high-precision transcription and structured organization of audio and video materials of the interview process, we ensure that every user's words are accurately captured and converted into text materials for analysis; we use natural language processing and text analysis tools to conduct in-depth mining of interview records to identify the core needs, expected improvements, and pain points revealed by users in the conversation;

[0018] User behavior testing: When users actually use products or services, pay close attention to key indicators such as operation paths, response time, interaction frequency, and error rates to reveal possible confusion points, blocking points, or efficiency bottlenecks in user interface design. Through sophisticated behavioral data analysis, identify interface design defects that directly affect user experience, such as difficult-to-understand icons, unreasonable layouts, and insufficient feedback. It can also reveal subtle problems that are not easy to detect but seriously affect user satisfaction, such as loading delays and complex operation processes.

[0019] Draw a user journey map: Through in-depth mining of user behavior data, identify key touch points, decision nodes, and emotional turning points in the user's interaction with the product or service, and thus build a detailed user journey map.

[0020] As a further preferred solution of the user experience design method based on natural language processing technology of the present invention, in step 1, data collection includes two modes: automatic detection and collection and active invitation and seamless guidance, as follows:

[0021] Automatically detect and collect all text corners containing customer feedback; use advanced text mining algorithms to capture all text data posted by users on the entire site in real time;

[0022] Active invitation and seamless guidance: For users who have not left feedback or who lack explicit feedback channels in specific scenarios, the system further innovatively introduces a natural voice input portal; through a friendly voice interaction interface, users are invited to express their needs or ideas in the most natural and convenient way.

[0023] As a further preferred embodiment of the user experience design method based on natural language processing technology of the present invention, in step 2, at the stage of natural language processing, a deep learning method is used to process the text to realize the functions of deep information extraction and emotion recognition quantification;

[0024] Among them, the core functions of emotion recognition and quantification adopt the basic BERT-CLS-ATT model. Through a CLS-ATT structure that can enhance the BERT model's semantic expression of sentences, the weighted summation of different attention weights of words in different positions is used to enhance the semantic expression of sentences; among them, the BERT model uses the Transformer encoding structure to build the entire model.

[0025] As a further preferred solution of the user experience design method based on natural language processing technology of the present invention, in step 2, the encoding link is specifically as follows:

[0026] The input of the BERT model is first converted into a vector form through an embedding operation, and then enters the encoding layer; the encoding layer uses a self-attention mechanism to capture semantic information and inputs the result into a feedforward neural network for parallel calculation; the encoding layer can be stacked in multiple layers, and each layer performs the same processing on the input until the last layer of encoding is completed and the final word embedding vector is output;

[0027] Input layer: The input sentence will be marked with start [CLS] and end [SEP] symbols and embedded into a vector, called token embedding; token embedding will be combined with position encoding so that the model can distinguish words at different positions in the sentence.

[0028] As a further preferred solution of the user experience design method based on natural language processing technology of the present invention, position coding uses sine and cosine functions to encode each position in the sentence, and the calculation formula is as follows:

[0029]

[0030] Among them, pos represents the position of each word in the sentence, i represents the dimension in the encoding, and model represents the dimension of the model; character embedding is similar to word embedding in other models, which is used to represent the semantic information of each word; sentence embedding is used to represent the overall information of the two sentences contained in each input sample, and each sentence has a unified embedding information; these three embedding vectors are superimposed together to form the input of the BERT model;

[0031] Encoder: In the encoder layer, the input and output of each Transformer block are a series of vectors, each of which contains multiple features. Through multiple layers of Transformer blocks, the BERT model can perform multiple nonlinear transformations on the input to obtain richer semantic representations. In the first layer of the encoder, each Transformer block consists of two sublayers: a multi-head self-attention sublayer and a feedforward neural network sublayer. The multi-head self-attention sublayer is responsible for interacting with the input, while the feedforward neural network sublayer maps the interaction results to a higher-dimensional space.

[0032] The calculation formula of the multi-head self-attention sub-layer is as follows:

[0033]

[0034] Where Q, K, and V represent query, key, and value, respectively.k is the dimension of the key and value; calculated by taking the dot product of the query and the key and dividing by The similarity between each query and the key can be calculated and converted into weights using the softmax function; these weights are then multiplied and added with the value to get the final self-attention representation;

[0035] The feedforward neural network sublayer consists of two linear transformations and one nonlinear transformation, and its calculation formula is as follows:

[0036] FFN(x)=max(0,xW1+b1)W2+b2

[0037] x is the input vector, W1 and W2 are trainable weight matrices, b1 and b2 are trainable bias vectors; the ReLU activation function is used here to introduce nonlinear functions;

[0038] The BERT model processes input by stacking multiple encoder layers, each of which contains a multi-head self-attention sublayer and a feed-forward neural network sublayer. In order to fully mine the information of the input sequence, the BERT model performs a residual connection layer normalization operation between each encoder layer.

[0039] The BERT model obtains a representation of the original input sequence by concatenating the outputs of all encoder layers.

[0040] As a further preferred solution of the user experience design method based on natural language processing technology of the present invention, in step 2, the training process is as follows:

[0041] In the pre-training stage, the BERT model uses two unsupervised pre-training tasks, namely the masked language model MLM and the next sentence prediction NSP;

[0042] In the masked language model task, the BERT model randomly masks some words in the input text and then predicts the masked words;

[0043] In the next sentence prediction task, the BERT model receives the input of two sentences and predicts whether the two sentences are adjacent or continuous. Fine-tuning: After pre-training, the BERT model needs to be fine-tuned in a specific NLP task. The fine-tuning process can be regarded as adding a new classifier on top of the BERT model and training the classifier in a specific NLP task. Specifically, the training process of the fine-tuning stage can be expressed by the following formula:

[0044] θ f =argminL task (D train ,θf ,θ bert )

[0045] Among them, θ f Represents the parameters of the newly added classifier, θ bert represents the parameters of the pre-trained BERT model, L task represents the loss function of a specific task, D train Represents the training dataset for the task; in the fine-tuning phase, a small batch stochastic gradient descent algorithm is usually used for training;

[0046] During fine-tuning, the top classifier of the BERT model is initialized with the parameters of the pre-training phase and then trained based on the specific task data. During training, the parameters of all layers can be fine-tuned, but the parameters of the pre-training phase are usually frozen, and only the parameters of the newly added classifier are updated.

[0047] The pre-training and fine-tuning stages of the BERT model can be regarded as training the model in different tasks, so that the BERT model can better understand natural language. Its model architecture is actually the Encoder in the Transformer.

[0048] As a further preferred embodiment of the user experience design method based on natural language processing technology of the present invention, in step 2, different places are usually paid attention to when judging the emotion of the sentence, that is, different attention is paid to the words in different positions of the sentence; a CLS-ATT structure that can enhance the model's semantic expression of the sentence is adopted, and in the CLS-ATT structure, the last layer of tag vectors of the BERT model other than the [CLS] tag is added to obtain a new feature vector by weighted summation, assuming that the feature vector of the last layer of other tag texts other than [CLS] is b i ,i=1,2,…,n weighted sum''then the new eigenvector is:

[0049]

[0050] And for a i is the global feature weight vector G and the text feature vector b i Mapping transformation obtains the new feature vector c i Do the dot product between them, and then perform the Softmax operation, that is:

[0051] c i =f(b i ),i=1,2,…,n

[0052]

[0053] Then, we combine this additional feature vector with the [CLS] tag in proportion to obtain the [CLS] vector with enhanced text features, namely Cls new , the calculation formula is as follows:

[0054] Cls new =λ*Cls+(1-λ)*E

[0055] Finally use this Cls new The feature vector is connected to the subsequent fully connected layer for sentiment classification task.

[0056] As a further preferred solution of the user experience design method based on natural language processing technology of the present invention, the overall structure of the model after combining the CLS-ATT structure is compared with the original BERT model. The new model structure adds a CLS-ATT text enhancement structure between the last layer output and the fully connected layer of BERT;

[0057] The core function of deep information extraction is mainly fine-tuned based on the open source neural network head model UIE of the ERINE structure. This model can be fine-tuned and trained with a small amount of manually annotated data to achieve the best results;

[0058] UIE selected the T5 model structure for parameter initialization. In order to enable the model to have information extraction capabilities, the pre-training model of the model is divided into the following parts;

[0059] 1. Text-to-Structure part:

[0060] For each pair (x, y), where x is the original text and y is the information structure to be extracted; construct the corresponding correct schema by extracting the spot and associating type in y, and automatically construct the wrong schema, so that the encoder and decoder of the model can generate the corresponding information structure y to be extracted based on the original text and schema; that is:

[0061]

[0062] 4. Structure generation part:

[0063] This part allows the model to generate a valid structure y that conforms to SEL and schema, thereby ensuring that the model can output a unified structure, and thus the model's decoder is trained accordingly;

[0064]

[0065] 5. Retrofitting semantic representation part:

[0066] This part allows the model to learn the basic capabilities of the language model, so the training objectives commonly used in masked language models are used;

[0067]

[0068] 4. Finetune part:

[0069] After completing the pre-training of UIE, in order to adapt to different information extraction tasks, the model needs to be further fine-tuned. For a given annotated corpus {(s,x,y)}, where s is the schema corresponding to the task, x is the original text, and y is the information structure to be extracted, the model is optimized through cross entropy loss; that is, re-training for a specific extraction task:

[0070]

[0071] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0072] 1. The application of natural language processing technology in the field of user feedback collection and analysis not only realizes the transition from passive reception and manual processing to active detection and intelligent analysis, but also greatly improves the user experience and the enterprise's response speed to user needs through seamless voice interaction and precise path guidance, and truly realizes the intelligent upgrade of the feedback system;

[0073] 2. This invention is built on advanced natural language understanding and sentiment analysis technologies. Our intelligent feedback management system achieves efficient mining and deep insight into massive text data, covering three core functions: deep information extraction, sentiment recognition and quantification, and automated report generation, aiming to comprehensively improve the value conversion of user feedback and the efficiency of corporate decision-making;

[0074] 3. In-depth information extraction: The system uses natural language understanding technology to conduct in-depth analysis of the massive amount of text data collected. Through precise analysis of vocabulary, grammar, semantics and contextual relationships, it accurately extracts key issues, specific opinions, demand suggestions and even potential pain points in user feedback to form a structured information database. This refined processing greatly improves the readability and analyzability of the feedback content, facilitating subsequent in-depth research and decision support.

[0075] 4. Emotion recognition and quantification: The system embeds an emotion analysis module, which can accurately identify and quantify the emotional color in user feedback. By analyzing language features such as word selection, sentence structure, and tone intensity, the system can not only identify the user's emotional state (such as satisfaction, dissatisfaction, confusion, excitement, etc.), but also quantify its intensity and give the feedback an emotional temperature. This emotional insight provides companies with a real-time barometer of user satisfaction, helping them to quickly respond to emotional fluctuations and adjust strategies in a timely manner.

[0076] 5. Automatic report generation: Based on the above in-depth analysis results, the system can automatically generate detailed user feedback reports. The report not only contains a statistical overview of various feedback topics and sentiment distribution trends, but also may provide targeted improvement suggestions and strategic inspirations; corporate decision makers no longer need to spend precious time collating data and writing reports. They only need to directly refer to the professional documents generated by the system to quickly grasp the overall situation and make wise decisions based on data-driven;

[0077] 6. Through comprehensive user feedback analysis, in-depth user interviews, rigorous behavioral testing, and refined user journey mapping, we build a complete user insight system that aims to fully understand user needs, accurately identify pain points, and timely optimize product experience, ultimately maximizing user satisfaction and product value. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without paying any creative work.

[0079] Figure 1 It is a structural diagram of a multi-terminal online system of the present invention;

[0080] Figure 2 It is the natural language permanent bottom suspension entrance of the present invention;

[0081] Figure 3 It is a structural diagram of the BERT model of the present invention;

[0082] Figure 4 It is the input structure of the BERT model of the present invention;

[0083] Figure 5 It is the Encoder structure diagram of the Transformer of the present invention;

[0084] Figure 6 It is the structure diagram of CLS-ATT of the present invention;

[0085] Figure 7 This is a diagram of the BERT-CLS-ATT model of the present invention;

[0086] Figure 8 It is the principle diagram of information extraction of UIE model of the present invention;

[0087] Fig. 9 It is a principle diagram of a user experience design method based on natural language processing technology of the present invention. DETAILED DESCRIPTION

[0088] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings:

[0089] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. The present invention is described in detail below based on the drawings and preferred embodiments, and the purpose and effect of the present invention will become clearer. It should be understood that the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0090] The present invention provides a user experience design method based on natural language processing technology, such as Fig. 9 As shown, the specific steps include the following:

[0091] 1. Data collection:

[0092] (1) Traditional method:

[0093] The collection of user feedback information usually relies on interactive platforms such as message areas and comment areas. In these areas, users are free to express their opinions, ask questions or share their experiences, forming an original data source for companies to review. However, the collection and analysis process of such information is highly dependent on human intervention, and there is a risk of inefficiency and subjective bias. Specifically, staff need to spend a lot of time browsing massive messages one by one, manually extracting key information, and then classifying, summarizing and interpreting them through subjective judgment. This cumbersome process is not only time-consuming and labor-intensive, and it is difficult to cope with a large amount of user feedback that is updated in real time, but also due to individual cognitive differences, it may lead to deviations in the understanding of the feedback content, affecting the accuracy and timeliness of decision-making.

[0094] (2) Collection by natural language:

[0095] In contrast, the use of advanced natural language processing technology for intelligent collection and analysis of user feedback information has significantly improved efficiency and accuracy.

[0096] 1) Automatically detect and collect all text corners containing customer feedback. No matter whether it is public comments, private messages or embedded feedback forms, nothing is missed. Intelligent identification and collection: The system uses advanced text mining algorithms to capture all text data posted by users on the entire site in real time. No matter what kind of interactive interface it hides in, it can be accurately located and integrated into the database. This automated process completely eliminates the inefficiency and omission risks of manual screening, ensuring that companies can fully and instantly understand the voice of users. Figure 1 shown. Figure 1 This is a structural diagram of a multi-terminal online system.

[0097] 2) Active invitation and seamless guidance. For users who have not yet left feedback or who lack explicit feedback channels in certain scenarios, the system further innovatively introduces a natural voice input portal. Through a friendly voice interaction interface, users are invited to express their needs or ideas in the most natural and convenient way. Once the user's voice input is received, the system immediately transcribes and understands it to ensure that no matter what medium the user chooses, their voice can be effectively captured. Figure 2 . Figure 2 It is a permanent floating entrance at the bottom for natural language.

[0098] To sum up, the application of natural language processing technology in the field of user feedback collection and analysis has not only achieved the transition from passive reception and manual processing to active detection and intelligent analysis, but also greatly improved the user experience and the company's response speed to user needs through seamless voice interaction and precise path guidance, truly realizing the intelligent upgrade of the feedback system.

[0099] 2 Natural Language Processing:

[0100] Built on advanced natural language understanding and sentiment analysis technologies, our intelligent feedback management system achieves efficient mining and deep insights into massive text data, covering three core functions: deep information extraction, sentiment recognition and quantification, and automated report generation, aiming to comprehensively enhance the value conversion of user feedback and the efficiency of corporate decision-making.

[0101] (1) Deep information extraction: The system uses natural language understanding technology to conduct in-depth analysis of the massive amount of text data collected. Through precise analysis of vocabulary, grammar, semantics and contextual relationships, it accurately extracts key issues, specific opinions, demand suggestions and even potential pain points in user feedback to form a structured information database. This refined processing greatly improves the readability and analyzability of the feedback content, facilitating subsequent in-depth research and decision support.

[0102] (2) Emotional recognition and quantification: The system embeds a sentiment analysis module that can accurately identify and quantify the emotional tone in user feedback. By analyzing language features such as word choice, sentence structure, and tone intensity, the system can not only identify the user's emotional state (such as satisfaction, dissatisfaction, confusion, excitement, etc.), but also quantify its intensity and give the feedback an emotional temperature. This kind of emotional insight provides companies with a real-time barometer of user satisfaction, helping them to quickly respond to emotional fluctuations and adjust strategies in a timely manner.

[0103] (3) Automated report generation: Based on the above in-depth analysis results, the system can automatically generate detailed user feedback reports. The report not only contains a statistical overview of various feedback topics and sentiment distribution trends, but also may provide targeted improvement suggestions and strategic inspiration. Corporate decision makers no longer need to spend precious time collating data and writing reports. They can directly refer to the professional documents generated by the system to quickly grasp the overall situation and make wise decisions based on data.

[0104] 3. User Experience Design

[0105] (1) Analyze user feedback

[0106] We conduct an all-round, multi-dimensional in-depth analysis of user feedback, covering various communication channels including but not limited to comments, messages, phone calls, etc., in order to accurately reveal users' praise for the product and their demands for improvement. Through the systematic sorting and interpretation of these diverse information, we strive to capture users' detailed feelings about all aspects of the product, so as to accurately portray their cognitive profile of the product value.

[0107] In this process, we use advanced text analysis technology to conduct semantic mining and sentiment analysis on massive comments and messages, and extract detailed insights into users' preferences for product features, functional evaluation, and user experience. At the same time, we convert telephone feedback into text and also include it in the above text analysis process to ensure that no matter what form the feedback takes, the user's voice can be fully listened to and objectively interpreted. By quantifying the user's likes and dislikes of various aspects of the product, we are able to build a three-dimensional and dynamic product satisfaction portrait, providing highly targeted data support for the company's product optimization and iteration.

[0108] (2) User interviews and analysis

[0109] In order to obtain more intuitive and in-depth user insights, we carefully organize user interviews and use advanced natural language technology equipment to record and post-process the entire process. By accurately transcribing and structuring the audio and video materials of the interview process, we ensure that every user's words are accurately captured and converted into text materials for analysis.

[0110] We then used natural language processing and text analysis tools to conduct in-depth mining of the interview records to identify the core needs, improvements, and pain points that users revealed in the conversations. Through fine-grained analysis of the interview content, we not only reveal users' opinions on specific aspects such as product functions, design, and services, but also reveal the underlying behavioral motivations, psychological expectations, and usage scenarios, providing a rich and detailed user story background for product design and user experience optimization.

[0111] (3) User behavior testing

[0112] To further explore potential problems in the process of user-product interaction, we design and implement rigorous user behavior testing. When users actually use products or services, we pay close attention to key indicators such as their operation path, response time, interaction frequency, error rate, etc., aiming to reveal possible confusion points, blocking points or efficiency bottlenecks in user interface design.

[0113] Through sophisticated behavioral data analysis, we can not only identify interface design defects that directly affect user experience, such as difficult-to-understand icons, unreasonable layouts, insufficient feedback, etc., but also reveal subtle problems that are not easy to detect but seriously affect user satisfaction, such as loading delays, complex operation processes, etc. These in-depth test results provide product teams with valuable optimization clues, helping them to accurately locate and fix problems in interface design and improve the overall user experience quality.

[0114] (4) Draw a user journey map

[0115] We are well aware of the importance of user journey and are committed to depicting the diverse characteristics and differentiated needs of user groups through detailed data analysis. Through in-depth mining of user behavior data, we identify key touchpoints, decision nodes, and emotional turning points in the user's interaction with products or services, thereby building a detailed user journey map.

[0116] In this map, we not only clearly mark the various stages of users from their first contact with the product to becoming loyal users, but also carefully describe the problems they may encounter, the solutions they expect, and the potential value-added opportunities at each stage. What is particularly critical is that we emphasize that at every touchpoint of the user journey, we must provide an experience design that is highly consistent with their current situation and does not interrupt their operational fluency. This is like when the user is concentrating on handling office affairs, there is a caring virtual assistant by his side, always ready to respond to his needs, provide seamless support for his task execution, and ensure that the user feels a high degree of comfort and efficiency throughout the process of using the product.

[0117] In summary, we have built a complete user insight system through comprehensive user feedback analysis, in-depth user interviews, rigorous behavioral testing, and refined user journey mapping, aiming to fully understand user needs, accurately identify pain points, and timely optimize product experience, ultimately maximizing user satisfaction and product value.

[0118] 4. Effect Evaluation

[0119] When evaluating the application effect of natural language processing technology in user experience design, it is necessary to comprehensively consider multiple dimensions to ensure that the technology not only effectively enhances the product's language interaction capabilities, but also effectively improves user satisfaction, ease of use and overall experience quality.

[0120] (1) User satisfaction

[0121] 1) Questionnaire survey: By designing standardized questionnaires or scales, collect users’ subjective evaluations of NLP functions, such as satisfaction scores, ease of use, accuracy of information understanding, and problem-solving efficiency.

[0122] 2) User feedback: Actively collect opinions and suggestions from users through in-app feedback mechanisms, social media, and customer service channels to understand their actual feelings about the NLP features and their desired improvements.

[0123] 3) Net Promoter Score (NPS): Measures the user's willingness to recommend a product to others, which can reflect the impact of NLP functions on user loyalty and word of mouth.

[0124] (2) Ease of use

[0125] 1) User interface adaptability: evaluate the degree of integration of NLP functions with the overall UI / UX design, such as the clarity of voice input prompts, text feedback layout, and error prompts.

[0126] 2) Learning curve: Observe the learning speed of new users when they first come into contact with NLP functions, and whether they can operate proficiently after a short period of use.

[0127] 3) Accessible design: Ensure that NLP technology is inclusive of different user groups (including users with special needs), such as supporting multiple languages ​​and dialects, and being compatible with assistive technologies.

[0128] (3) Task completion process and results

[0129] 1) Task success rate: This is the ratio of the number of successful attempts by users to complete specific tasks (such as querying information, executing instructions, generating text, etc.) using NLP functions to the total number of attempts.

[0130] 2) Task completion time: Record the time required for users to complete tasks related to NLP interaction and evaluate whether NLP improves operational efficiency.

[0131] 3) Error Rate and Correction Behavior: Tracks the misunderstandings, misidentifications, or incorrect responses that users encounter when using NLP features, and the number of actions they take to correct these issues.

[0132] 4) Emotional understanding and appropriate response: If sentiment analysis is involved, check whether the system can accurately identify user emotions and give appropriate emotional responses

[0133] In the natural language processing stage, deep learning methods are mainly used to process text in order to achieve the functions of deep information extraction and sentiment recognition quantification.

[0134] The core functions of emotion recognition and quantification mainly use the basic BERT-CLS-ATT model. Through a CLS-ATT structure that can enhance the semantic expression of sentences by the BERT model, the semantic expression of sentences is enhanced by using the weighted summation of different attention weights of words in different positions. Among them, BERT uses the Transformer encoding structure to build the entire model. Its model structure is as follows Figure 3 As shown: Figure 3 This is the BERT model structure diagram.

[0135] (1) Coding

[0136] The input of the model is first converted into a vector form through an embedding operation, and then enters the encoding layer. The encoding layer uses a self-attention mechanism to capture semantic information and inputs the results into a feedforward neural network for parallel calculation. The encoding layer can be stacked in multiple layers, and each layer performs the same processing on the input until the last layer of encoding is completed and the final word embedding vector is output.

[0137] Input layer: The input sentence is marked with start [CLS] and end [SEP] symbols and embedded into a vector, called a token embedding. These token embeddings are combined with position encodings so that the model can distinguish between words at different positions in the sentence. The input structure of the BERT model is as follows: Figure 4 As shown: Figure 4 Input structure for the BERT model.

[0138] Position encoding uses sine and cosine functions to encode each position in the sentence. The calculation formula is as follows:

[0139]

[0140] Among them, pos represents the position of each word in the sentence, i represents the dimension in the encoding, and model represents the dimension of the model. Character embedding is similar to word embedding in other models, which is used to represent the semantic information of each word. Sentence embedding is used to represent the overall information of the two sentences contained in each input sample, and each sentence has a unified embedding information. Finally, these three embedding vectors are superimposed together to form the input of the BERT model.

[0141] Encoder: In the encoder layer, the input and output of each Transformer block are a series of vectors, each of which contains multiple features. Through multiple layers of Transformer blocks, the BERT model can perform multiple nonlinear transformations on the input to obtain richer semantic representations. In the l-th layer encoder, each Transformer block consists of two sublayers: a multi-head self-attention sublayer and a feedforward neural network sublayer. Among them, the multi-head self-attention sublayer is responsible for interacting with the input, and the feedforward neural network sublayer maps the interaction results to a higher-dimensional space.

[0142] The calculation formula of the multi-head self-attention sub-layer is as follows:

[0143]

[0144] Where Q, K, and V represent query, key, and value, respectively. k is the dimension of the key and value. By taking the dot product of the query and the key and dividing by The similarity between each query and the key can be calculated and converted into weights using a softmax function. These weights are then multiplied with the value and added together to get the final self-attention representation.

[0145] The feedforward neural network sublayer consists of two linear transformations and one nonlinear transformation, and its calculation formula is as follows:

[0146] FFN(x)=max(0,xW1+b1)W2+b2

[0147] x is the input vector, W1 and W2 are trainable weight matrices, b1 and b2 are trainable bias vectors. The ReLU activation function is used here to introduce nonlinear functions.

[0148] The BERT model processes input by stacking multiple encoder layers, each of which contains a multi-head self-attention sublayer and a feedforward neural network sublayer. In order to fully mine the information of the input sequence, the BERT model performs a residual connection layer normalization operation between each encoder layer.

[0149] Finally, the BERT model obtains a representation of the original input sequence by concatenating the outputs of all encoder layers.

[0150] (2) Training process

[0151] During the pre-training stage, the BERT model used two unsupervised pre-training tasks, namely Masked Language Model (MLM) and Next Sentence Prediction (NSP).

[0152] In the masked language model task, the BERT model randomly masks some words in the input text and then predicts the masked words. This task can help the BERT model learn the semantic connection between words, and also prevent the model from over-relying on words in the context.

[0153] In the next sentence prediction task, the BERT model receives the input of two sentences and predicts whether the two sentences are adjacent or continuous. This task can help the BERT model learn the relationship between sentences, thereby improving the generalization ability of the model in tasks such as text classification. Fine-tuning, after pre-training is completed, the BERT model needs to be fine-tuned in specific NLP tasks. The fine-tuning process can be regarded as adding a new classifier on top of the BERT model and training the classifier in a specific NLP task. Specifically, the training process of the fine-tuning stage can be expressed by the following formula:

[0154] θ f =argminL task (D train ,θ f ,θ bert )

[0155] Among them, θ f Represents the parameters of the newly added classifier, θ bert represents the parameters of the pre-trained BERT model, L task represents the loss function of a specific task, D train Represents the training dataset for the task. In the fine-tuning phase, a small batch stochastic gradient descent algorithm is usually used for training.

[0156] During fine-tuning, the top classifier of the BERT model is initialized with the parameters of the pre-training phase and then trained based on specific task data. During training, the parameters of all layers can be fine-tuned, but the parameters of the pre-training phase are usually frozen, and only the parameters of the newly added classifier are updated.

[0157] In general, the pre-training and fine-tuning stages of the BERT model can be regarded as training the model in different tasks, so that the BERT model can better understand natural language. Its model architecture is actually the Encoder in the Transformer, such as Figure 5 shown. Figure 5 This is the Encoder structure diagram of Transformer.

[0158] The original BERT model uses the first [CLS] tag of the text to represent the text features of the entire text, and directly feeds it into the subsequent fully connected layer for fine-tuning. However, in order to more accurately express the semantics of complex text, some additional structures may be needed to enhance the expression. Intuitively, when judging the sentiment of a sentence, we usually pay attention to different places, that is, we pay different attention to words at different positions in the sentence. Therefore, we use a CLS-ATT structure that can enhance the model's expression of sentence semantics, such as Figure 6 As shown: Figure 6 This is the CLS-ATT structure diagram.

[0159] In the CLS-ATT structure, the last layer of BERT model tag vectors other than [CLS] are added to obtain a new feature vector by weighted summation. Assume that the feature vector of the last layer of tag text other than [CLS] is b i ,i=1,2,…,n. Then the new eigenvector is:

[0160]

[0161] And for a i , is the global feature weight vector G and the text feature vector b i Mapping transformation obtains the new feature vector c i Do the dot product between them, and then perform the Softmax operation, that is:

[0162] c i =f(b i ),i=1,2,…,n

[0163]

[0164] Then, we combine this additional feature vector with the [CLS] tag in proportion to obtain the [CLS] vector with enhanced text features, namely Cls new , the calculation formula is as follows:

[0165] Cls new =λ*Cls+(1-λ)*E

[0166] Finally use this Clsnew The feature vector is connected to the subsequent fully connected layer for sentiment classification task.

[0167] The overall structure of the model after combining the CLS-ATT structure can be seen Figure 7 Compared with the original BERT model, the new model structure adds a CLS-ATT text enhancement structure between the last layer output and the fully connected layer of BERT. Figure 7 This is the BERT-CLS-ATT model diagram.

[0168] The core function of deep information extraction is mainly fine-tuned based on the open source neural network head model UIE of the ERINE structure. This model can be fine-tuned and trained with a small amount of manually annotated data to achieve the best effect. The principle of information extraction of the UIE model is as follows Figure 8 As shown: Figure 8 This is the schematic diagram of UIE model information extraction.

[0169] UIE selected the T5 model structure for parameter initialization. In order to enable the model to have information extraction capabilities, the pre-training model of the model is divided into the following parts.

[0170] 1. Text-to-Structure

[0171] For each pair (x, y), where x is the original text and y is the information structure to be extracted, the correct schema is constructed by extracting the spot and associating type in y, and the wrong schema is automatically constructed, so that the encoder and decoder of the model can generate the corresponding information structure y to be extracted based on the original text and schema. That is:

[0172]

[0173] 6. Structure generation

[0174] This part allows the model to generate a valid structure y that conforms to SEL and schema, thereby ensuring that the model can output a unified structure, and thus the decoder of the model is trained accordingly.

[0175]

[0176] 7. Retrofitting semantic representation

[0177] This part allows the model to learn the basic capabilities of the language model, so the training objectives commonly used by masked language models are used.

[0178]

[0179] 4. Finetune

[0180] After completing UIE pre-training, in order to adapt to different information extraction tasks, the model needs to be further fine-tuned. For a given annotated corpus {(s,x,y)}, where s is the schema corresponding to the task, x is the original text, and y is the information structure to be extracted, the model is optimized through cross entropy loss. That is, retraining for a specific extraction task:

[0181]

[0182] Implementation steps:

[0183] 1. Data collection and feedback mechanism

[0184] Automatic data collection: The platform collects user interaction data in real time through a built-in log system, including but not limited to page browsing records, video viewing time, exercise completion status, etc.

[0185] User feedback invitation: After the course ends, the system automatically triggers a customized feedback interface, inviting users to evaluate the course content, teaching methods, interface design and other aspects.

[0186] 2. Natural Language Processing and Sentiment Analysis

[0187] Text preprocessing: Clean the text data of user feedback, including removing stop words and punctuation marks, and performing stem extraction.

[0188] Deep Information Extraction: Leverage deep learning models, such as recurrent neural networks (RNNs) or long short-term memory networks (LSTMs), to extract key information and user intent from text.

[0189] Sentiment recognition and quantification: Through sentiment analysis algorithms such as support vector machine (SVM) or deep belief network (DBN), users’ feedback emotions are classified into positive and negative polarities and quantitatively scored.

[0190] 3. Automated report generation

[0191] Multi-dimensional analysis: Analyze the collected data according to multiple dimensions such as user groups, course types, and time series.

[0192] Visual presentation: Use charts and graphs to present the analysis results in an intuitive way, such as a bar chart of sentiment distribution, a line chart of user activity, etc.

[0193] Automatic report generation: The system automatically generates a detailed user experience analysis report based on the preset template.

[0194] 4. User experience design iteration

[0195] User feedback analysis: Identify pain points and room for improvement in user experience based on data analysis results in automated reports.

[0196] User solution: arrange one-on-one interviews to gain an in-depth understanding of users’ usage experience and unmet needs, use natural language analysis to summarize interview questions, and further conduct cause analysis. Provide preliminary solutions.

[0197] User behavior testing: Design and implement user behavior testing, such as eye tracking experiments, to observe user behavior patterns in actual operations.

[0198] User journey mapping: Based on user feedback and behavioral testing results, map the user’s entire journey of using the platform to identify key touchpoints and potential areas for improvement.

[0199] 5. System iteration and optimization

[0200] Interface redesign: Iterate the user interface based on the user journey map, such as optimizing the navigation structure, adjusting the button layout, etc.

[0201] Functional optimization: Based on user feedback, optimize the platform's functions, such as adding personalized recommendation algorithms, improving the interactivity of exercises, etc.

[0202] Technical implementation: Apply design changes and technical optimizations to the online education platform and conduct A / B testing to verify the improvement effect.

[0203] 6. Continuous monitoring and feedback

[0204] Real-time monitoring: Continuously monitor user feedback and behavior data to ensure that new designs and technical optimizations can bring the expected results.

[0205] Periodic report updates: Regularly update the user experience analysis report to reflect the latest user feedback and behavior trends.

[0206] Expected Results

[0207] Improve user satisfaction: Improve users' overall satisfaction with the platform by continuously optimizing the user experience.

[0208] Enhance user stickiness: Improved user interface and functional design can increase user usage frequency and duration.

[0209] Improve learning efficiency: Personalized recommendations and interactive exercises can improve users' learning efficiency and results.

[0210] Optimize decision-making process: All design and functional optimization are based on data-driven analysis to ensure scientific and effective decision-making.

[0211] This case study shows how an intelligent online education platform systematically collects and analyzes user data, applies natural language processing and sentiment analysis techniques, and combines user experience design to continuously iterate and optimize platform functions to improve user satisfaction and learning efficiency. Through this continuous optimization process, the platform can better adapt to changes in user needs and maintain its competitiveness.

[0212] Those skilled in the art can understand that the above are only preferred examples of the invention and are not intended to limit the invention. Although the invention is described in detail with reference to the above examples, those skilled in the art can still modify the technical solutions recorded in the above examples or replace some of the technical features with equivalents. Any modification, equivalent replacement, etc. made within the spirit and principle of the invention should be included in the protection scope of the invention. All technical features in this embodiment can be freely combined according to actual needs.

[0213] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A user experience design method based on natural language processing technology, characterized by: The specific steps include: Step 1, data collection: using natural language processing technology to intelligently collect and analyze user feedback information; Step 2, Natural Language Processing: Built on advanced natural language understanding and sentiment analysis technologies, the intelligent feedback management system achieves efficient mining and deep insights into massive text data, covering deep information extraction, sentiment recognition and quantification, and automated report generation; Among them, deep information extraction: using natural language understanding technology to deeply analyze the massive amount of text data collected; through precise analysis of vocabulary, grammar, semantics and contextual relationships, accurately extract key issues, specific opinions, demand suggestions and even potential pain points in user feedback to form a structured information database; Emotion recognition and quantification: The system is embedded with an emotion analysis module, which can accurately identify and quantify the emotional color in user feedback. By analyzing language features such as word selection, sentence structure, and tone intensity, the system can not only identify the user's emotional state, including satisfaction, dissatisfaction, confusion, and excitement, but also quantify its intensity, giving the feedback an emotional temperature. Automatic report generation: Based on the in-depth analysis results, the system can automatically generate detailed user feedback reports; the reports not only include statistical overviews of various feedback topics and sentiment distribution trends, but may also provide targeted improvement suggestions and strategic inspirations; Step 3, user experience design, specifically includes collecting user feedback, user interviews and analysis, user behavior testing and drawing user journey maps: Collecting user feedback: Collecting user feedback in an all-round and multi-dimensional manner, including but not limited to comments, messages, phone calls and other communication channels, to reveal users' praise for the product and their demands for improvement; using advanced text analysis technology to conduct semantic mining and sentiment analysis on massive comments and messages, and extracting detailed insights into users' preferences for product features, functional evaluations, and user experience; User interview and analysis: Through high-precision transcription and structured organization of audio and video materials of the interview process, we ensure that every user's words are accurately captured and converted into text materials for analysis; we use natural language processing and text analysis tools to conduct in-depth mining of interview records to identify the core needs, expected improvements, and pain points revealed by users in the conversation; User behavior testing: When users actually use products or services, pay close attention to key indicators such as operation paths, response time, interaction frequency, and error rates to reveal possible confusion points, blocking points, or efficiency bottlenecks in user interface design. Through sophisticated behavioral data analysis, identify interface design defects that directly affect user experience, such as difficult-to-understand icons, unreasonable layouts, and insufficient feedback. It can also reveal subtle problems that are not easy to detect but seriously affect user satisfaction, such as loading delays and complex operation processes. Draw a user journey map: Through in-depth mining of user behavior data, identify key touch points, decision nodes, and emotional turning points in the user's interaction with the product or service, and thus build a detailed user journey map.

2. The user experience design method based on natural language processing technology according to claim 1, characterized in that: In step 1, data collection includes two methods: automatic detection and collection and active invitation and seamless guidance, as follows: Automatically detect and collect all text corners containing customer feedback; use advanced text mining algorithms to capture all text data posted by users on the entire site in real time; Active invitation and seamless guidance: For users who have not left feedback or who lack explicit feedback channels in specific scenarios, the system further innovatively introduces a natural voice input portal; through a friendly voice interaction interface, users are invited to express their needs or ideas in the most natural and convenient way.

3. The user experience design method based on natural language processing technology according to claim 1, characterized in that: In step 2, at the stage of natural language processing, a deep learning method is used to process the text to achieve the functions of deep information extraction and sentiment recognition quantification; Among them, the core functions of emotion recognition and quantification adopt the basic BERT-CLS-ATT model. Through a CLS-ATT structure that can enhance the BERT model's semantic expression of sentences, the weighted summation of different attention weights of words in different positions is used to enhance the semantic expression of sentences; among them, the BERT model uses the Transformer encoding structure to build the entire model.

4. The user experience design method based on natural language processing technology according to claim 3 is characterized by: In step 2, the encoding process is as follows: The input of the BERT model is first converted into a vector form through an embedding operation, and then enters the encoding layer; the encoding layer uses a self-attention mechanism to capture semantic information and inputs the result into a feedforward neural network for parallel calculation; the encoding layer can be stacked in multiple layers, and each layer performs the same processing on the input until the last layer of encoding is completed and the final word embedding vector is output; Input layer: The input sentence will be marked with start [CLS] and end [SEP] symbols and embedded into a vector, called token embedding; token embedding will be combined with position encoding so that the model can distinguish words at different positions in the sentence.

5. The user experience design method based on natural language processing technology according to claim 4, characterized in that: Position encoding uses sine and cosine functions to encode each position in the sentence. The calculation formula is as follows: Among them, pos represents the position of each word in the sentence, i represents the dimension in the encoding, and model represents the dimension of the model; character embedding is similar to word embedding in other models, which is used to represent the semantic information of each word; sentence embedding is used to represent the overall information of the two sentences contained in each input sample, and each sentence has a unified embedding information; these three embedding vectors are superimposed together to form the input of the BERT model; Encoder: In the encoder layer, the input and output of each Transformer block are a series of vectors, each of which contains multiple features. Through multiple layers of Transformer blocks, the BERT model can perform multiple nonlinear transformations on the input to obtain richer semantic representations. In the first layer of the encoder, each Transformer block consists of two sublayers: a multi-head self-attention sublayer and a feedforward neural network sublayer. The multi-head self-attention sublayer is responsible for interacting with the input, while the feedforward neural network sublayer maps the interaction results to a higher-dimensional space. The calculation formula of the multi-head self-attention sub-layer is as follows: Where Q, K, and V represent query, key, and value, respectively. k is the dimension of the key and value; calculated by taking the dot product of the query and the key and dividing by The similarity between each query and the key can be calculated and converted into weights using the softmax function; these weights are then multiplied and added with the value to get the final self-attention representation; The feedforward neural network sublayer consists of two linear transformations and one nonlinear transformation, and its calculation formula is as follows: FFN(x)=max(0,xW1+b1)W2+b2 x is the input vector, W1 and W2 are trainable weight matrices, b1 and b2 are trainable bias vectors; the ReLU activation function is used here to introduce nonlinear functions; The BERT model processes input by stacking multiple encoder layers, each of which contains a multi-head self-attention sublayer and a feed-forward neural network sublayer. In order to fully mine the information of the input sequence, the BERT model performs a residual connection layer normalization operation between each encoder layer. The BERT model obtains a representation of the original input sequence by concatenating the outputs of all encoder layers.

6. The user experience design method based on natural language processing technology according to claim 5, characterized in that: In step 2, the training process is as follows: In the pre-training stage, the BERT model uses two unsupervised pre-training tasks, namely the masked language model MLM and the next sentence prediction NSP; In the masked language model task, the BERT model randomly masks some words in the input text and then predicts the masked words; In the next sentence prediction task, the BERT model receives the input of two sentences and predicts whether the two sentences are adjacent or continuous. Fine-tuning: After pre-training, the BERT model needs to be fine-tuned in a specific NLP task. The fine-tuning process can be regarded as adding a new classifier on top of the BERT model and training the classifier in a specific NLP task. Specifically, the training process of the fine-tuning stage can be expressed by the following formula: i f =argminL task (D train ,i f ,i bert ) Among them, θ f Represents the parameters of the newly added classifier, θ bert represents the parameters of the pre-trained BERT model, L task represents the loss function of a specific task, D train Represents the training dataset for the task; in the fine-tuning phase, a small batch stochastic gradient descent algorithm is usually used for training; During fine-tuning, the top classifier of the BERT model is initialized with the parameters of the pre-training phase and then trained based on the specific task data. During training, the parameters of all layers can be fine-tuned, but the parameters of the pre-training phase are usually frozen, and only the parameters of the newly added classifier are updated. The pre-training and fine-tuning stages of the BERT model can be regarded as training the model in different tasks, so that the BERT model can better understand natural language. Its model architecture is actually the Encoder in the Transformer.

7. The user experience design method based on natural language processing technology according to claim 3 is characterized by: In step 2, when judging the sentiment of a sentence, different places are usually paid attention to, that is, different attention is paid to words in different positions of the sentence; a CLS-ATT structure is adopted to enhance the semantic expression of the sentence. In the CLS-ATT structure, the last layer of tag vectors of the BERT model other than the [CLS] tag is added to obtain a new feature vector by weighted summation. Assume that the feature vector of the last layer of tag text other than [CLS] is b i ,i=1,2,…,n weighted sum''then the new eigenvector is: And for a i is the global feature weight vector G and the text feature vector b i Mapping transformation obtains the new feature vector c i Do the dot product between them, and then perform the Softmax operation, that is: c i =f(b i ),i=1,2,…,n Then, we combine this additional feature vector with the [CLS] tag in proportion to obtain the [CLS] vector with enhanced text features, namely Cls new , the calculation formula is as follows: Cls new =λ*Cls+(1-λ)*E Finally use this Cls new The feature vector is connected to the subsequent fully connected layer for sentiment classification task.

8. The user experience design method based on natural language processing technology according to claim 7, characterized in that: The overall structure of the model after combining the CLS-ATT structure. Compared with the original BERT model, the new model structure adds the CLS-ATT text enhancement structure between the last layer output and the fully connected layer of BERT. The core function of deep information extraction is mainly fine-tuned based on the open source neural network head model UIE of the ERINE structure. This model can be fine-tuned and trained with a small amount of manually annotated data to achieve the best results; UIE selected the T5 model structure for parameter initialization. In order to enable the model to have information extraction capabilities, the pre-training model of the model is divided into the following parts; 1. Text-to-Structure part: For each pair (x, y), where x is the original text and y is the information structure to be extracted; construct the corresponding correct schema by extracting the spot and associating type in y, and automatically construct the wrong schema, so that the encoder and decoder of the model can generate the corresponding information structure y to be extracted based on the original text and schema; that is:

2. Structure generation part: This part allows the model to generate a valid structure y that conforms to SEL and schema, thereby ensuring that the model can output a unified structure, and thus the model's decoder is trained accordingly; 3. Retrofitting semantic representation part: This part allows the model to learn the basic capabilities of the language model, so the training objectives commonly used in masked language models are used; 4. Finetune part: After completing the pre-training of UIE, in order to adapt to different information extraction tasks, the model is further fine-tuned; for a given annotated corpus {(s,x,y)}, where s is the schema corresponding to the task, x is the original text, and y is the information structure to be extracted, the model is optimized through cross entropy loss; that is, re-training for a specific extraction task:

Images