Live broadcast room problem automatic reply system based on large language model

Abstract

The invention provides a live broadcast room problem automatic reply system based on a large language model, and relates to the field of live broadcast room automatic reply, and the system comprises a data input module which is used for inputting demand information by a user; the data acquisition module is used for acquiring demand information input by a user; the data identification module is used for analyzing and understanding the information processed by the data acquisition module; and the reply generation module is used for generating corresponding reply contents based on a large language model according to the contents identified by the data identification module. According to the live broadcast room problem automatic reply system based on the large language model, the interval setting module is arranged, a depreciable interval value is set by an owner of a live broadcast platform according to the cost of a product and the like, the depreciation can be performed according to the purchase condition of the user in the process of communicating with the user based on the large language model, and the user experience is improved. The system can autonomously adjust the price of the product, so that the sales volume of the commodity can be improved.

Description

Technical Field

[0001] The present invention relates to the field of automatic reply in live broadcast rooms, and in particular to an automatic reply system for live broadcast room questions based on a large language model. Background Art

[0002] The live broadcast room is an online space for real-time interaction between the host and the audience. Its core functions include content display, real-time communication and commercial conversion. During live broadcasts, due to the large number of viewers, the host may not be able to reply to each of them one by one. In addition, with the emergence of digital human live broadcasts, an automatic reply system is needed to automatically reply to the audience's questions and achieve efficient interaction through the automatic reply system.

[0003] Currently, the automatic reply system in the live broadcast room can usually only respond mechanically based on preset product information, such as product price information, and can only respond based on the set price. It lacks the function of flexibly adjusting prices based on user purchases to increase sales. Moreover, users of the live broadcast platform can only choose products pushed by the live broadcast platform. The live broadcast platform does not support customized products, and the sales form of products is relatively limited.

[0004] Therefore, it is necessary to provide an automatic reply system for live broadcast room questions based on a large language model to solve the above technical problems. Summary of the Invention

[0005] The present invention provides an automatic reply system for live broadcast room questions based on a large language model, which solves the problem that the existing live broadcast room reply system lacks the ability to flexibly adjust prices according to users' purchasing conditions, thereby increasing sales.

[0006] In order to solve the above technical problems, the present invention provides a live broadcast room question automatic reply system based on a large language model, comprising: a data input module, wherein the data input module is used for users to input demand information;

[0007] A data collection module, wherein the data collection module is used to collect demand information input by users;

[0008] A data recognition module, which analyzes and understands the information processed by the data acquisition module;

[0009] A reply generation module, which generates corresponding reply content based on the content identified by the data recognition module and the large language model;

[0010] A reply output module, configured to output the reply content generated by the reply generation module;

[0011] An information preset module, which is used to preset information about live broadcast products;

[0012] The information preset module includes an interval setting module, which is used to set the price interval of the product. The system allows custom prices within the set price interval;

[0013] Among them, the trigger conditions for the system to adjust product prices include passive user requests and active push price reduction strategies;

[0014] The data input module includes a reservation input module, which is used for users to customize products. The reply generation module also includes a reservation generation module, which is used to display the user's customized products;

[0015] Preferably, the user-defined product includes the following steps:

[0016] S11, the user clicks the reservation input module to enter the reservation interface;

[0017] S12. Select the product to be ordered. The ordering interface displays a plurality of corresponding product samples. Multiple characteristic parts of each product sample can be marked.

[0018] S13, the user selects features of different products and customizes their combinations to form new products;

[0019] S14. Each selectable feature of the product has multiple feature types preset, and the user selects the required corresponding feature by clicking.

[0020] Preferably, the reservation input module further includes a search box, through which the user searches for desired products or product features.

[0021] Preferably, the data acquisition module includes a preprocessing module and a feature classification module, the preprocessing module is used to clean the data collected by the data acquisition module, the feature classification module includes a product classification module and a data feature classification module, the product classification module is used to classify the data information of the same product;

[0022] Within a preset time interval, the data feature classification module is used to classify data information under the same product, wherein the reply generation module makes a unified reply to the same type of data information.

[0023] Preferably, the data recognition module includes a language recognition module, a graphic recognition module and a speech recognition module. The language recognition module recognizes text information input by the user based on a large language model, the graphic recognition module is used to recognize graphic information input by the user, and the speech recognition module is used to recognize speech information input by the user.

[0024] Preferably, the reply generation module includes an information generation module, and the information generation module includes a language generation module, a graphic generation module and a speech generation module. The language generation module generates text reply information based on a large language model, the graphic generation module is used to generate graphic reply information, and the speech generation module is used to generate voice reply information.

[0025] Preferably, the reply generation module further includes a history calling module. When the data identification module identifies that the question information input by the user is repeated information, the history calling module calls historical reply information for reply.

[0026] Preferably, the data input module includes a preview input module, which is used for the user to build a morphological model according to his or her body shape, and to preview the desired clothing product by clicking on the morphological model. The reply generation module also includes a preview generation module, which is used to display the morphological model built by the user.

[0027] The construction of the morphological model for wearing preview specifically includes the following steps:

[0028] S21, the user clicks the preview input module to enter the preview interface;

[0029] S22, the user selects a base model, and multiple local features of the base model can be adjusted separately;

[0030] S23, the user selects the overall size of the basic model according to his / her body shape, and then adjusts the local features of the body;

[0031] S24. After the morphological model is constructed, select the preview product and wear it on the morphological model for preview.

[0032] Preferably, the basic model includes a full-body model and a partial-body model.

[0033] Preferably, the reply generation module further includes a recommendation module. When the user selects a product, the recommendation module is used to recommend products that match the selected product or similar products to the user.

[0034] Compared with related technologies, the automatic reply system for live broadcast room questions based on a large language model provided by the present invention has the following beneficial effects:

[0035] The present invention provides an automatic reply system for live broadcast room questions based on a large language model. Passive user requests refer to users asking for price reductions, otherwise they will not buy. Active price reduction strategies include when users ask for prices but do not receive subsequent purchase messages. In this case, the system can push price reduction strategies. Alternatively, when product sales are poor, it can directly send price reduction information to users entering the live broadcast platform.

[0036] By setting up the interval setting module, merchants on the live broadcast platform can set the range of price reduction values ​​based on the cost of the product, etc. The automatic reply system can reduce prices based on the user's purchasing situation during communication with the user based on the large language model. The system can independently adjust the price of the product, thereby increasing the sales volume of the product. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a functional block diagram of the automatic reply system for live broadcast room questions based on a large language model provided by the present invention;

[0038] Figure 2 This is a functional block diagram of the reply generation module provided by the present invention;

[0039] Figure 3 A functional block diagram of the data input module provided by the present invention;

[0040] Figure 4 A block diagram of the steps for user-defined products provided by the present invention;

[0041] Figure 5 A flowchart of the steps for constructing a morphological model for wearing preview provided by the present invention;

[0042] Figure 6 The principle block diagram of the data acquisition module provided by the present invention;

[0043] Figure 7 This is a principle block diagram of the data identification module provided by the present invention. DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0045] The present invention provides a system for automatically replying to questions in a live broadcast room based on a large language model.

[0046] Please refer to Figures 1 to 3 In one embodiment of the present invention, the live broadcast room automatic reply system based on the large language model includes:

[0047] A data input module, wherein the data input module is used for users to input required information;

[0048] A data collection module, wherein the data collection module is used to collect demand information input by users;

[0049] A data recognition module, which analyzes and understands the information processed by the data acquisition module;

[0050] A reply generation module, which generates corresponding reply content based on the content identified by the data recognition module and the large language model;

[0051] A reply output module, configured to output the reply content generated by the reply generation module;

[0052] An information preset module, which is used to preset information about live broadcast products;

[0053] The information preset module includes an interval setting module, which is used to set the price interval of the product. The system allows custom prices within the set price interval;

[0054] Among them, the trigger conditions for the system to adjust product prices include passive user requests and active push price reduction strategies;

[0055] The data input module includes a reservation input module, which is used for users to customize products. The reply generation module also includes a reservation generation module, which is used to display users' customized products.

[0056] In this embodiment, a passive user request refers to a user requesting a price reduction, otherwise they will not make a purchase; an active price reduction strategy refers to a situation where a user inquires about a price but does not make a subsequent purchase. In this case, the system can push a price reduction strategy, or directly send a price reduction information to the user who enters the live broadcast platform when the product is not selling well.

[0057] By setting up the interval setting module, merchants on the live broadcast platform can set the range of price reduction values ​​based on the cost of the product, etc. The automatic reply system can reduce prices based on the user's purchasing situation during communication with the user based on the large language model. The system can independently adjust the price of the product, thereby increasing the sales of the product.

[0058] When lowering the price, you can set multiple selling prices within the price reduction range to implement a progressive bargaining strategy. For example, if the price reduction range is 10-30 yuan, you can set the price reduction amount to 10 yuan, 20 yuan, and 30 yuan. If the buyer demands a price reduction of more than 30 yuan, the system will not lower the price further.

[0059] The data input module is used by the user end facing the audience to input information. The user end enters the questions they want to ask through the input interface; the data acquisition module collects the information input by the user, and the data recognition module identifies the content of the information based on the large language model. The reply generation module generates the reply content based on the large language model and finally outputs the reply content;

[0060] The information preset module sets the product information in advance, including the product name, model, price, size, performance, and promotion information; the reply generation module indexes the product information in the information preset module and generates corresponding reply content.

[0061] By setting up a predetermined input module, users can customize the style of products, not limited to the product style in the live broadcast room, so as to better meet customers' demand for product styles and increase product sales;

[0062] The steps for building a large language model include:

[0063] a1: Collect and preprocess text data (including books, articles, web pages, and other text sources in various fields. This data will be used to train the model. During the preprocessing stage, operations such as cleaning, word segmentation, and part-of-speech tagging are required so that the model can better understand and process the text).

[0064] a2: Select the Transformer model architecture;

[0065] a3: Using the model architecture, training the model using preprocessed data, and choosing the backpropagation algorithm and optimizer involved in the training process to minimize the loss function and update the model parameters;

[0066] a4: During the training process, the validation set is used to evaluate the performance of the model, and the model is optimized and adjusted based on the evaluation results. The model can be optimized and adjusted by adjusting hyperparameters, adding regularization terms, etc.

[0067] a5: Use optimization techniques to optimize the performance and efficiency of the model. Optimization techniques such as pruning, quantization, and knowledge distillation can be used to optimize the model.

[0068] a6: Deploy the trained model to the live streaming platform's server through an API or other means. The model can be applied to various natural language processing tasks, such as text generation, question-answering systems, and machine translation.

[0069] Large language models can also use models such as the GPT series and BERT.

[0070] Please refer again Figure 1 As an optional method of this embodiment, the data acquisition module includes a preprocessing module and a feature classification module. The preprocessing module is used to clean the data collected by the data acquisition module. The feature classification module includes a product classification module and a data feature classification module. The product classification module is used to classify data information of the same product.

[0071] Within a preset time interval, the data feature classification module is used to classify data information under the same product, wherein the reply generation module makes a unified reply to the same type of data information.

[0072] The data collection module obtains user interaction data in real time through the API connection of the live broadcast platform;

[0073] The text cleaning steps include: denoising: removing emoticons, special characters, repeated characters, and garbled characters.

[0074] Standardization conversion: unify uppercase and lowercase letters, convert between full-width and half-width letters; replace synonyms to reduce semantic ambiguity;

[0075] Abnormal data filtering: Filter meaningless text, advertising spam, and remove invalid questions that are too short (e.g. <3 words) or too long (e.g. >200 words).

[0076] When there are multiple products in the live broadcast room, the product classification module groups the information of the same product together;

[0077] The data feature classification module further categorizes information about the same product, grouping information with the same semantics together. For example, it groups inquiries about product prices together, allowing multiple users to receive a unified response to the same question, eliminating the need for sequential responses.

[0078] Among them, statistics and responses to the same product information are carried out within a preset time. For example, within 1-5 seconds, users who ask about product prices are counted, and then a unified response is given to avoid long classification and statistics time, which leads to long response time and makes users unable to get answers to their questions for a long time.

[0079] In the input interface, users can choose to only display the content of the messages they sent and the content of the messages they replied to, so that users can more clearly see the replies they need, and users can click on the sent messages to be reminded to reply.

[0080] See also Figure 3 and Figure 4 As an optional method of this embodiment, the user-defined product includes the following steps:

[0081] S11, the user clicks the reservation input module to enter the reservation interface;

[0082] S12. Select the product to be ordered. The ordering interface displays a plurality of corresponding product samples. Multiple characteristic parts of each product sample can be marked.

[0083] S13, the user selects features of different products and customizes their combinations to form new products;

[0084] S14. Each selectable feature of the product has multiple feature types preset, and the user selects the required corresponding feature by clicking.

[0085] After entering the reservation interface, choose the type of product sold on the live broadcast platform, such as clothing, home furnishings, electrical appliances, etc.

[0086] For example, in the case of T-shirts, after selecting a T-shirt, multiple different types of T-shirts will appear on the upper side of the booking interface. Among them, one or more locations of each T-shirt, such as the collar, front logo, back logo, chest pocket, cuffs, etc., can be marked.

[0087] You can select a favorite T-shirt body and then choose to modify different features, such as replacing the neckline of the T-shirt body with the neckline shape of other T-shirts, or replacing the logo of other T-shirts with the logo of the T-shirt body, or choosing to set or not set pockets on the chest, etc.; thus, you can customize a piece of clothing that you like.

[0088] For graphic features such as logos, you can choose other T-shirt logos to replace the logo of the T-shirt body itself or add it to the T-shirt body, and adjust the positions of the two logos so that both logos are located on the customized T-shirt body.

[0089] For home furnishings and electrical appliances, changes are mainly made to patterns, designs, etc.

[0090] If the product is a bed, you can adjust the style of the bed backrest or the pattern on the backrest, etc., to customize the combination of the bed backrest and the bed body.

[0091] Among them, for S14, multiple feature types are preset at each selectable marked feature, that is, when the corresponding marked feature position is clicked, multiple different feature types will be displayed. For example, when the neckline position of the T-shirt is clicked, multiple different neckline types will be displayed. At this time, you can click on the corresponding favorite neckline type to change it, so that the neckline type can be adjusted more quickly; you can still choose features from other types of T-shirts or clothes to replace or add to the selected T-shirt.

[0092] The multiple feature types preset at each selectable marked feature are hidden under normal conditions and will only appear when the mouse is clicked or remains on the marked feature.

[0093] The replacement of image features can be achieved by using front-end graphics technologies (such as SVG, Canvas) and other technologies.

[0094] As an optional manner of this embodiment, the reservation input module further includes a search box, and the user searches for desired products or product features through the search box.

[0095] By setting up a search box, you can search for other styles of products or combine product features to complete customized products, thereby increasing product diversity.

[0096] As a preferred method of this embodiment, an upload box may be set to upload local product styles, such as logo patterns, etc.

[0097] See also Figure 7 The data recognition module includes a language recognition module, a graphic recognition module and a speech recognition module. The language recognition module recognizes the text information input by the user based on a large language model, the graphic recognition module is used to recognize the graphic information input by the user, and the speech recognition module is used to recognize the speech information input by the user.

[0098] By setting up language recognition modules, image recognition modules and voice recognition modules, users are allowed to input text sentences, image information and voice information at the input end of the live broadcast platform, increasing the diversity of the input end and facilitating different users to input information and communicate with the anchor, etc.

[0099] Among them, the language recognition module recognizes and understands sentences based on a large language model;

[0100] The graphic information recognition module mainly recognizes the image content and text content in the picture. By recognizing the image content, when the user enters the picture of the target product, the image recognition module can provide the user with the desired product or similar products.

[0101] Among them, convolutional neural networks (CNNs) can be used to identify graphic content: multiple convolutional layers are used to automatically extract hierarchical features of images (such as bottom-level edges → middle-level shapes → high-level semantics), and combined with fully connected layers to achieve classification or positioning.

[0102] To recognize text in images, you can use an end-to-end OCR model.

[0103] For speech recognition, speech recognition models can be used, such as Transformer-based end-to-end models (such as Fairseq-S2T, Listen, Attend and Spell (LAS), which use the self-attention mechanism to simultaneously capture audio timing and language context).

[0104] See also Figure 2As an optional method of this embodiment, the reply generation module includes an information generation module, and the information generation module includes a language generation module, a graphic generation module and a voice generation module. The language generation module generates text reply information based on a large language model, the graphic generation module is used to generate graphic reply information, and the voice generation module is used to generate voice reply information.

[0105] The language generation module, the image generation module, and the voice generation module enable the automatic reply system to generate text replies based on the large language model, as well as generate image messages and voice messages for reply;

[0106] Generate image information, such as sending pictures of products from different angles for users to view from multiple angles;

[0107] By setting up a voice generation module, it is convenient for a small number of illiterate people or people who are not convenient to type to reply through voice.

[0108] Please refer again Figure 2 As an optional method of this embodiment, the reply generation module also includes a history calling module. When the data identification module identifies that the question information input by the user is repeated information, the history calling module calls the historical reply information for reply.

[0109] By setting up a history call module, you can directly call historical reply information to reply to detected repeated questions without having to reorganize the reply information, thereby improving reply efficiency.

[0110] See also Figure 3 and Figure 5 As an optional method of this embodiment, the data input module includes a preview input module, which is used for the user to build a morphological model according to his or her body shape, and to preview the clothing product by clicking on the desired clothing product and wearing it on the morphological model. The reply generation module also includes a preview generation module, which is used to display the morphological model built by the user;

[0111] The construction of the morphological model for wearing preview specifically includes the following steps:

[0112] S21, the user clicks the preview input module to enter the preview interface;

[0113] S22, the user selects a base model, and multiple local features of the base model can be adjusted separately;

[0114] S23, the user selects the overall size of the basic model according to his / her body shape, and then adjusts the local features of the body;

[0115] S24. After the morphological model is constructed, select the preview product and wear it on the morphological model for preview.

[0116] By setting up a preview input module, the user can build his own morphological model and try on the selected clothes through the morphological model, so that he can more intuitively see the effect of the clothes after wearing them, reducing the risk of clothes not fitting well or not achieving the expected effect due to not being able to try them on.

[0117] Specifically, after entering the preview interface, the user selects a basic model, which includes a male model and a female model, and the corresponding gender models are further divided into adult models and child models;

[0118] After selecting the basic model, the overall size of the basic model is automatically selected according to the user's height and weight. For example, if the user is 175cm tall and 65kg, the model will be adjusted to the body shape of 175cm and 65kg. Then, the user can partially adjust the corresponding part of the morphological model according to their specific situation. For example, compared with the basic morphological model, the user's belly will be larger, so the belly can be adjusted separately. After the adjustment is completed, select the pre-ordered clothes and click on the clothes to wear them on the morphological model, so that you can intuitively see whether the clothes fit and the effect after wearing them.

[0119] As an optional method in this example, the adjustable parts of the morphological model include shoulders, chest, stomach, waist and hips, buttocks, thighs, etc.; the morphological model has a rotation function, and different positions can be viewed.

[0120] The adjustment method includes displaying a data input box when clicking the corresponding part, where you can enter the data of the corresponding part, such as waist circumference, leg circumference, chest circumference, etc., or setting a sliding event at the corresponding part, clicking the slider to slide left and right to achieve the range change of the corresponding part, etc.

[0121] The construction of the human body morphological model specifically includes the following steps:

[0122] b1. Divide the adjustable parts: determine the parts that need to be adjusted (such as chest, waist, hips), and assign independent parameters to each part (such as scaling factor, arc value);

[0123] b2. Deformation algorithm: Use the Blendshape modeling tool: Create the extreme shape of the part in the modeling tool and export it as a model containing Blendshape;

[0124] The code uses weight values ​​(0-1) to blend the base shape with the target shape to achieve smooth deformation. FFD (Free Form Deformation): Use the FFD3D plugin in Three.js to add a grid of control points for parts such as the waist. Drag the control points or move them by parameters to distort the local model.

[0125] b3. Interactive interface development:

[0126] For example, in control creation: add HTML<inputtype="range"> Slider, bound to the part parameter (e.g. id="waist-slider" controls waist circumference);

[0127] Establish event monitoring: javascript;

[0128] For example: update the model when the waist slider changes;

[0129]

[0130] b4. Real-time rendering and optimization:

[0131] Model update: Every time the parameters change, trigger mesh.needsUpdate = true to notify Three.js to redraw;

[0132] b5. Performance optimization: Reduce the deformation accuracy of non-critical parts and give priority to rendering core parts (such as the torso).

[0133] The establishment of the clothing model can be achieved through techniques such as bone skin binding, blend shape presets or cloth physical simulation, so that the parametric deformation of the clothing model and the human body model can be synchronized in real time, achieving the effect of dynamic changes of clothing as the human body parts are adjusted.

[0134] As an optional manner of this embodiment, the basic model includes a full-body model and a partial-body model.

[0135] By setting partial body models such as the upper body model and the lower body model, when choosing a top, only the upper body model can be selected. Correspondingly, when choosing pants, only the upper body model can be selected, thereby reducing the load on the system operation.

[0136] For different merchants, only models of corresponding parts can be set. For example, a merchant who only sells hats can choose to build a head model, where the head circumference of the head model can be adjusted; a merchant who sells shoes can only set a foot model, where the size of the foot and the height of the instep can be adjusted.

[0137] See also Figure 3 Among them, the data input module also includes a conventional input module, which is used to transmit text information, voice information, or picture information on the input interface for product communication with the automatic reply system or anchor.

[0138] See also Figure 2As an optional method of this embodiment, the reply generation module also includes a recommendation module. When the user selects a product, the recommendation module is used to recommend products that match the selected product or similar products to the user.

[0139] By setting up a recommendation module, supporting products or similar products can be recommended to users, thereby improving the user's shopping experience and increasing the sales volume of the live broadcast platform.

[0140] For example, when a user selects a top, we can recommend matching pants, a bottoming shirt, or shoes, etc., or recommend similar clothes to the user.

[0141] As an optional method of this embodiment, the live broadcast room question automatic reply system based on the large language model realizes the live broadcast automatic reply function based on the large language model and RPA (robotic process automation) technology.

[0142] Among them, the automatic reply system for live broadcast room questions based on the large language model can be applied to a digital human live broadcast platform or a real-person live broadcast platform.

[0143] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An automatic reply system for live broadcast room questions based on a large language model, characterized by: include: A data input module, wherein the data input module is used for users to input required information; A data collection module, wherein the data collection module is used to collect demand information input by users; A data recognition module, which analyzes and understands the information processed by the data acquisition module; A reply generation module, which generates corresponding reply content based on the content identified by the data recognition module and the large language model; A reply output module, configured to output the reply content generated by the reply generation module; An information preset module, which is used to preset information about live broadcast products; The information preset module includes an interval setting module, which is used to set the price interval of the product. The system allows custom prices within the set price interval; Among them, the trigger conditions for the system to adjust product prices include passive user requests and active push price reduction strategies; The data input module includes a reservation input module, which is used for users to customize products. The reply generation module also includes a reservation generation module, which is used to display users' customized products.

2. The automatic reply system for live broadcast room questions based on a large language model according to claim 1 is characterized in that: The user-defined product includes the following steps: S11, the user clicks the reservation input module to enter the reservation interface; S12. Select the product to be ordered. The ordering interface displays a plurality of corresponding product samples. Multiple characteristic parts of each product sample can be marked. S13, the user selects features of different products and customizes their combinations to form new products; S14. Each selectable feature of the product has multiple feature types preset, and the user selects the required corresponding feature by clicking.

3. The automatic reply system for live broadcast room questions based on a large language model according to claim 2 is characterized in that: The reservation input module further includes a search box, through which the user searches for desired products or product features.

4. The automatic reply system for live broadcast room questions based on a large language model according to claim 1 is characterized in that: The data acquisition module includes a preprocessing module and a feature classification module. The preprocessing module is used to clean the data collected by the data acquisition module. The feature classification module includes a product classification module and a data feature classification module. The product classification module is used to classify the data information of the same product. Within a preset time interval, the data feature classification module is used to classify data information under the same product, wherein the reply generation module makes a unified reply to the same type of data information.

5. The automatic reply system for live broadcast room questions based on a large language model according to claim 1 is characterized in that: The data recognition module includes a language recognition module, a graphic recognition module and a speech recognition module. The language recognition module recognizes text information input by the user based on a large language model, the graphic recognition module is used to recognize graphic information input by the user, and the speech recognition module is used to recognize speech information input by the user.

6. The automatic reply system for live broadcast room questions based on a large language model according to claim 1 is characterized in that: The reply generation module includes an information generation module, which includes a language generation module, a graphic generation module and a voice generation module. The language generation module generates text reply information based on a large language model, the graphic generation module is used to generate graphic reply information, and the voice generation module is used to generate voice reply information.

7. The automatic answering system for live broadcast room questions based on a large language model according to claim 6 is characterized in that: The reply generation module further includes a history calling module. When the data identification module identifies that the question information input by the user is repeated information, the history calling module calls historical reply information to make a reply.

8. The automatic reply system for live broadcast room questions based on a large language model according to claim 1 is characterized in that: The data input module includes a preview input module, which is used for the user to build a morphological model according to his or her body shape, and to preview the clothing product by clicking on the desired clothing product and wearing it on the morphological model. The reply generation module also includes a preview generation module, which is used to display the morphological model built by the user; The construction of the morphological model for wearing preview specifically includes the following steps: S21, the user clicks the preview input module to enter the preview interface; S22, the user selects a base model, and multiple local features of the base model can be adjusted separately; S23, the user selects the overall size of the basic model according to his / her body shape, and then adjusts the local features of the body; S24. After the morphological model is constructed, select the preview product and wear it on the morphological model for preview.

9. The automatic reply system for live broadcast room questions based on a large language model according to claim 8 is characterized in that: The basic model includes a full-body model and a partial-body model.

10. The automatic reply system for live broadcast room questions based on a large language model according to claim 1 is characterized in that: The reply generation module also includes a recommendation module. When the user selects a product, the recommendation module is used to recommend products that match the selected product or similar products to the user.

Images