Content tag identification method, system and device

Abstract

The present invention provides a content label recognition method for analyzing the practicality and timeliness of content on an OTA platform, comprising the following steps: obtaining content data on the OTA platform, the content data including text information and image information; performing preliminary label recognition on the content data using at least one graphic labeling model to obtain at least one preliminary label; performing keyword matching on the content data using a keyword model to obtain at least one keyword matching result; inputting the content data, preliminary labels, and keyword matching results into a large text model; the large text model outputs a practicality label or timeliness label for the content data based on a preset prompt word. The present invention can significantly reduce manual intervention, allowing operators to quickly filter out practical and timeliness content that meets the requirements from massive content, thereby improving content operation efficiency.

Description

Background Art

[0002] With the popularization of the Internet and mobile Internet, online travel agency (OTA) platforms have become an important channel for users to obtain travel information and make reservations. OTA platforms have accumulated a large amount of user-generated content (UGC), which contains rich travel information and has important reference value for users' travel decisions. However, how to efficiently filter out valuable information for users from massive content has become an important challenge facing OTA platforms. At present, OTA platform content screening and analysis mainly rely on two methods: manual screening and keyword-based matching. Manual screening relies on the subjective judgment of operators. Through manual browsing and screening of content, content that meets operational requirements is selected for recommendation or display. This method is still applicable when the amount of content is small, but when faced with massive content, it is inefficient, costly, and difficult to ensure the timeliness of information. In addition, the subjectivity of manual screening may also lead to information bias. The keyword-based matching method retrieves and matches content by presetting a keyword library. Although this method is relatively efficient, it can only match based on literal meaning and cannot understand the deep semantics of the content, resulting in limited accuracy and recall. The keyword matching method is also difficult to effectively capture the subtle differences and timeliness changes of the content. Therefore, existing technologies have significant deficiencies in processing the massive content screening and practicality and timeliness analysis of OTA platforms, and cannot effectively meet the platform's growing content operation needs. Existing technologies are difficult to accurately identify whether content has practical value, such as whether it contains useful travel advice, price information, etc.; it is also difficult to effectively judge the timeliness of information, such as whether the information is recently released or whether it describes a recent event. These limitations make it difficult for users to quickly obtain the information they need, which reduces the user experience and also affects the content operation efficiency of the platform.

[0003] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute the prior art known to ordinary technicians in the field. Summary of the invention

[0004] In view of the problems in the prior art, the purpose of the present invention is to provide a content tag identification method, system and device for analyzing the practicality and timeliness of OTA platform content, so as to overcome the defects of low efficiency and insufficient accuracy of existing OTA platform content screening and analysis methods.

[0005] The embodiment of the present invention provides a content tag identification method for analyzing the practicability and timeliness of OTA platform content, comprising the following steps:

[0006] Obtain content data on the OTA platform, the content data includes text information and image information;

[0007] Using at least one image-text labeling model to perform preliminary label recognition on the content data to obtain at least one preliminary label;

[0008] Using the keyword model to perform keyword matching on the content data to obtain at least one keyword matching result;

[0009] Input content data, preliminary tags, and keyword matching results into the text big model;

[0010] The text model outputs practicality labels or timeliness labels for content data based on preset prompt words.

[0011] In some embodiments, the image and text tagging model includes a text model and an image model, wherein:

[0012] The text model is used to process the text information in the content data and output the text score;

[0013] The image model is used to process the image information in the content data and output the image score;

[0014] The text score and image score are combined to obtain the final score, and a preset threshold is used to determine whether to assign a corresponding preliminary label to the content data.

[0015] In some embodiments, the text model is a roberta-base model and the image model is a CLIP model.

[0016] In some embodiments, the text model and the image model are trained separately by adapter fine-tuning.

[0017] In some embodiments, the keyword model performs keyword matching based on a preset keyword library, where the keyword library includes keywords for representing practicality and keywords for representing timeliness.

[0018] In some embodiments, the keywords used to characterize practicality include "traffic", "ticket price", "transportation", "accommodation", "strategy", and "guide"; the keywords used to characterize timeliness include "yesterday", "today", "recently", "upcoming", "exhibition", "event", and "match".

[0019] In some embodiments, the text large model is the qwen2-7b model.

[0020] In some embodiments, the prompt words include the title, text, preliminary tags, and keyword matching results of the content data.

[0021] In some embodiments, the method further includes: filtering the practicality tags and timeliness tags output by the large text model according to preset rules.

[0022] In some embodiments, the preset rules include word limit and time limit.

[0023] In some embodiments, the vllm reasoning framework is used to accelerate the reasoning process of large text models.

[0024] The embodiment of the present invention further provides a content tag identification system, which is used to implement the above-mentioned content tag identification method.

[0025] An embodiment of the present invention further provides a content tag identification device, including:

[0026] processor;

[0027] a memory in which executable instructions for the processor are stored;

[0028] The processor is configured to execute the steps of the above-mentioned content tag identification method by executing executable instructions.

[0029] An embodiment of the present invention further provides a computer-readable storage medium for storing a program, which implements the steps of the content tag identification method when the program is executed by a processor.

[0030] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure.

[0031] The content tag recognition method provided by the present invention, by combining the image and text tagging model, the keyword model and the large text model, has the following beneficial effects: the present invention can greatly reduce manual intervention, so that operators can quickly screen out practical and timely content that meets the requirements from massive content, thereby improving content operation efficiency. The method combining multimodal information fusion and deep semantic understanding can more accurately identify the practicality and timeliness characteristics of content, reduce the misjudgment rate caused by traditional keyword matching methods, and improve the accuracy of tag recognition. By quickly and accurately screening out high-quality and valuable content, the user experience on the OTA platform can be effectively improved and user stickiness can be enhanced. The method provided by the present invention can provide more accurate data support for the content operation of the OTA platform, thereby realizing more intelligent content operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Other features, objectives and advantages of the present invention will become more apparent from a reading of the detailed description of non-limiting embodiments made with reference to the following accompanying drawings.

[0033] Figure 1 is a flow chart of a content tag identification method according to an embodiment of the present invention;

[0034] Figure 2 is a structural diagram of a content tag identification device according to an embodiment of the present invention;

[0035] Figure 3 It is a schematic diagram of the structure of a computer-readable storage medium according to an embodiment of the present invention. DETAILED DESCRIPTION

[0036] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that the disclosure will be more comprehensive and complete and to fully convey the concepts of the example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0037] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the steps. For example, some steps can be decomposed, while some steps can be combined or partially combined. Therefore, the actual execution order may change according to the actual situation. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0038] In the present invention, the basic mechanism of label recognition and feature analysis of multimodal content is implemented through information processing technology, the key information in the content is extracted by artificial intelligence model, and the screening results are optimized through multi-stage processing, aiming to solve the technical problems of low efficiency and insufficient accuracy of the existing OTA platform content screening and analysis methods, thereby realizing efficient and accurate content analysis. First of all, multimodal information fusion is a technology that integrates information (such as text, image, audio, video, etc.) from different sources or modalities. Its basic principle is to use the complementarity and redundancy between different modal information to improve the accuracy and robustness of information understanding and analysis. In the present invention, multimodal information fusion specifically refers to the fusion of text information and image information. Text information usually contains descriptions and comments on things, while image information can intuitively show the form and characteristics of things. By fusing text information and image information, the meaning of the content can be understood more comprehensively and accurately. For example, if a travel note only contains text descriptions, it may be difficult to accurately judge whether the described attractions are attractive; but if it also contains photos of the attractions, you can more intuitively understand its scenery, thereby more accurately judging the practicality of the travel note. Secondly, keyword matching is a technology for finding and matching keywords in text based on a predefined keyword library. Its basic principle is to determine whether the text contains relevant keywords by comparing the text with the words in the keyword library. Keyword matching technology is simple and efficient, and can quickly and preliminarily screen out potentially relevant texts. In the present invention, keyword matching is used to preliminarily screen content that may contain practical and timeliness information. For example, if a travel note contains keywords such as "strategy" and "guide", it can be preliminarily judged that it may be practical; if it contains keywords such as "recently" and "upcoming", it can be preliminarily judged that it may be timeliness. However, keyword matching technology can only match based on literal meaning, and cannot understand the deep semantics of the text, which is prone to false recall and missed recall. Finally, text understanding based on large language models is a technology that uses pre-trained large language models (LLM) to perform deep semantic understanding of text. Large language models learn rich language knowledge and contextual relationships by training on large-scale text corpora, and can better understand the meaning, emotion and intention of text. In the present invention, large language models are used to perform refined analysis and label recognition on the content after preliminary screening. For example, even if a travel note contains the keyword "yesterday", the large language model can determine whether its main content is related to the events that occurred "yesterday" based on the context, thereby more accurately judging its timeliness. In addition, the large language model can also combine the results of multimodal information fusion to understand the content more comprehensively and improve the accuracy of tag recognition. The present invention can effectively improve the efficiency and accuracy of content screening and analysis on the OTA platform by combining technical principles such as multimodal information fusion, keyword matching, and text understanding based on large language models.Specifically, multimodal information fusion can provide a more comprehensive understanding of content, keyword matching can quickly and preliminarily screen content, and text understanding based on large language models can perform refined analysis and tag recognition, thereby overcoming the shortcomings of existing technologies and achieving more efficient and accurate content tag recognition. This combination of multiple technical principles can effectively improve the final recognition effect.

[0039] like Figure 1 As shown, an embodiment of the present invention provides a content tag identification method for analyzing the practicality and timeliness of OTA platform content, including the following steps:

[0040] S100, obtaining content data on an OTA platform, where the content data includes text information and image information;

[0041] In an embodiment of the present invention, content data is user-generated content (UGC) collected from an online travel agent (OTA) platform, and these contents exist in two modes: text and image. Text information usually includes travel notes, strategies, comments, descriptions, etc., and image information includes photos, screenshots, illustrations, etc. It can be understood by those skilled in the art that the content data of the OTA platform comes from a wide range of sources, for example, it can be obtained through API interfaces, web crawling, etc. Text information can be structured data (such as title, text, tags) or unstructured data (such as user comments), and image information can be picture files in various formats. This step is intended to provide a data basis for subsequent tag recognition and is a prerequisite for subsequent steps. By simultaneously acquiring text information and image information, the foundation is laid for subsequent multimodal information fusion, thereby more comprehensively understanding the content.

[0042] S200, using at least one image-text labeling model to perform preliminary label recognition on content data to obtain at least one preliminary label;

[0043] In an embodiment of the present invention, the image and text labeling model refers to a model that can process text and image information at the same time and output preliminary labels. The model can be a single model or a combination of multiple models. Preliminary labels refer to labels that preliminarily classify or mark content, such as preliminary judgments on content themes, emotions, styles, etc. Those skilled in the art will appreciate that the image and text labeling model can adopt a variety of deep learning models, such as BERT, RoBERTa, etc. for text models, and ResNet, EfficientNet, etc. for image models. By performing preliminary label recognition on the content through the image and text labeling model, content features can be quickly extracted to provide a reference for subsequent refined label recognition. This step introduces multimodal information fusion and utilizes the complementary information of text and images to preliminarily identify content features, thereby providing more accurate input for the subsequent refined recognition of large models, thereby improving the accuracy of overall recognition.

[0044] S300, performing keyword matching on content data using a keyword model to obtain at least one keyword matching result;

[0045] In an embodiment of the present invention, a keyword model refers to a model for keyword matching based on a preset keyword library. The keyword library contains keywords used to characterize practicality and / or timeliness. The keyword matching result refers to whether the text contains the keywords in the keyword library, and which keywords are contained. Those skilled in the art will appreciate that the keyword model can use a simple string matching algorithm or a more complex text matching algorithm. The purpose of keyword matching is to quickly and preliminarily determine whether the content contains keywords related to practicality and / or timeliness, and provide clues for subsequent refined tag identification. This step uses keyword matching to quickly locate potentially related words in the content, providing important reference information for the subsequent large model judgment, thereby improving recognition efficiency and recall rate.

[0046] S400, inputting content data, preliminary tags and keyword matching results into the text macro model;

[0047] In the embodiment of the present invention, the text big model refers to a pre-trained large language model, such as BERT, RoBERTa, GPT series models, etc. The information input to the text big model includes the original content data (text and image information), as well as the preliminary label and keyword matching results obtained by the image and text labeling model and the keyword model. This step aims to use the powerful semantic understanding ability of the big model to perform refined analysis and label recognition on the content.

[0048] S500: The text large model outputs a practicality label or a timeliness label of the content data based on a preset prompt word.

[0049] In an embodiment of the present invention, the prompt word prompt refers to input information that guides the text large model to perform a specific task. By designing appropriate prompt words, the large model can be guided to better understand the task objectives and output more accurate results. Practicality labels and timeliness labels are labels for the final classification or marking of content, indicating whether the content has practical value and whether it is timely, respectively. This step is the core step of the present invention. By combining the information obtained in the previous steps, the large model is used to perform deep semantic understanding and comprehensive judgment, and finally output the practicality and / or timeliness labels of the content.

[0050] Through the implementation of the above steps, the embodiment of the present invention effectively combines multimodal information fusion, keyword matching and text understanding based on a large language model, and utilizes the comprehensiveness of multimodal information fusion, the rapidity of keyword matching and the deep semantic understanding ability of a large language model to more accurately and efficiently identify the practicality and timeliness information in the content of the OTA platform, thereby improving the efficiency and accuracy of content screening.

[0051] In some optional embodiments, the image and text marking model includes a text model and an image model, wherein:

[0052] The text model is used to process the text information in the content data and output the text score;

[0053] In the embodiment of the present invention, the text model refers to a model for processing text information and outputting a corresponding score. The score can characterize the characteristics or attributes of the text information in a certain aspect, such as the sentiment tendency (positive, negative, neutral), topic relevance, quality score, etc. of the text. The text model can adopt various natural language processing models, such as but not limited to: models based on recurrent neural networks (RNN), such as LSTM, GRU, etc.; models based on convolutional neural networks (CNN), such as TextCNN, etc.; models based on Transformer, such as BERT, RoBERTa, ALBERT, etc. In the specific implementation process, the text model receives text information in the content data as input, such as the text of the travel notes, the comment content, etc., and outputs one or more text scores. For example, the text model can output a score representing the sentiment tendency of the text, and the higher the score, the more positive the sentiment. This step is intended to use the text model to extract key features in the text information and provide a basis for subsequent fusion and label recognition. For example, if the words "recommended" and "worth going" appear multiple times in the text content, the text model may output a higher positive sentiment score.

[0054] The image model is used to process the image information in the content data and output the image score;

[0055] In the embodiment of the present invention, the image model refers to a model for processing image information and outputting a corresponding score. The score can characterize the characteristics or attributes of the image information in a certain aspect, such as the clarity, beauty, subject relevance, etc. of the image. The image model can adopt various computer vision models, such as: models based on convolutional neural networks (CNN), such as ResNet, EfficientNet, VGGNet, etc.; models based on Transformers, such as Vision Transformer (ViT), etc. In the specific implementation process, the image model receives image information in the content data as input, such as photos of scenic spots, pictures uploaded by users, etc., and outputs one or more image scores. For example, the image model can output a score representing the clarity of the image, and the higher the score, the clearer the image. This step is intended to use the image model to extract key features in the image information, such as the beauty of the image, whether it contains specific objects, etc., to provide a basis for subsequent fusion and label recognition. For example, if the image content is clear and the composition is beautiful, the image model may output a higher beauty score.

[0056] The text score and image score are combined to obtain the final score, and a preset threshold is used to determine whether to assign a corresponding preliminary label to the content data.

[0057] In an embodiment of the present invention, fusion refers to the process of combining or integrating text scores and image scores to obtain a comprehensive final score. The fusion method can be adopted in a variety of ways, such as: weighted average, weighted average of text scores and image scores according to preset weights. Splicing, splicing the text score and image score into a vector. Based on neural network fusion, the relationship between text scores and image scores is learned using a neural network model, and the final score is output. The preset threshold refers to the score value used to determine whether to assign a corresponding preliminary label to the content data. If the final score exceeds the preset threshold, the corresponding preliminary label is assigned to the content data; otherwise, it is not assigned. For example, if the final score represents the relevance of the content, the preset threshold is 0.8, and if the final score is 0.9, the preliminary label of "relevant" is assigned to the content data. This step is intended to effectively fuse multimodal information, comprehensively consider text and image information, so as to more accurately judge the relevance or other attributes of the content, and provide preliminary labels for subsequent refined label recognition. For example, the text content of a travel note is average, but the accompanying pictures are very beautiful. By integrating the text score and image score, we can more accurately judge that the travel note has a high reference value and assign a corresponding preliminary label.

[0058] By using a text model and an image model to process text and image information respectively and perform effective fusion, the features of the content data can be extracted more comprehensively, thereby improving the accuracy of preliminary label recognition, providing more reliable input for subsequent refined label recognition, and ultimately improving the overall effect of content label recognition. The text model and image model of the embodiment of the present invention can select different model architectures and parameters according to the specific application scenario. The fusion method can also select different fusion strategies according to the specific situation. The preset threshold can be adjusted according to actual needs. The models and fusion methods listed above are only examples and are not limitations of the present invention.

[0059] In some optional embodiments, the text model is a roberta-base model, and the image model is a CLIP model. The roberta-base model has excellent performance in natural language processing tasks. The roberta-base model is an improved version of the BERT model. Through a larger training data set and longer training time, as well as improvements in training strategies such as dynamic masking, it has achieved significant performance improvements in multiple natural language processing tasks. Those skilled in the art will understand that the roberta-base model can better understand the semantic information of the text and extract key features in the text, thereby providing a more accurate basis for subsequent fusion and label recognition. For example, the roberta-base model can better understand the emotional tendency, topic relevance and other information in the text, thereby more accurately judging whether the text is practical or timely. The CLIP model has shown powerful capabilities in image and text cross-modal tasks. The CLIP model maps images and texts to the same vector space by contrastive learning, thereby effectively measuring the similarity between images and texts. Those skilled in the art will understand that the CLIP model can effectively extract visual features of images and associate them with text information, thereby providing richer information for subsequent fusion and label recognition. For example, the CLIP model can identify information such as objects and scenes contained in an image and match it with the text description, thereby more accurately determining whether the image is related to the text content. Combining the roberta-base model with the CLIP model can fully utilize the advantages of both models to achieve more accurate multimodal information fusion and label recognition. For example, when the text content of a travel note describes a scenic spot and the accompanying picture also contains a photo of the scenic spot, the roberta-base model can extract the descriptive information about the scenic spot in the text, and the CLIP model can extract the visual features of the scenic spot in the photo. By fusing these two types of information, the authenticity and relevance of the travel note can be more accurately determined, thereby more accurately assigning it a corresponding preliminary label.

[0060] By using the roberta-base model and the CLIP model to process text and image information respectively and perform effective fusion, the advantages of the two models in their respective fields can be fully utilized, the features of content data can be extracted more accurately, the accuracy of preliminary label recognition can be improved, and more reliable input can be provided for subsequent refined label recognition, and finally the overall effect of content label recognition can be improved. Especially when processing OTA platform content containing rich text information and image information, this embodiment can more effectively improve the efficiency and accuracy of content screening and analysis. Although the present embodiment preferably uses the roberta-base model as the text model and the CLIP model as the image model, this is not a limitation of the present invention. Those skilled in the art can select other suitable text models and image models according to specific application scenarios. For example, the text model can use models such as bert-base and albert, and the image model can use models such as resnet and efficientnet. As long as effective processing and feature extraction of text and image information can be achieved and used for subsequent fusion and label recognition, it should be regarded as falling within the scope of protection of the present invention. For example, when computing resources are limited, a model with a smaller parameter amount can be used to improve processing efficiency. When there are higher requirements for specific features of an image (such as face recognition), a model optimized for the feature can be selected.

[0061] In some optional embodiments, the text model and the image model are trained respectively by means of adapter fine-tuning. Adapter fine-tuning is an efficient parameter fine-tuning method, which is particularly suitable for large pre-trained models. The core idea is to insert a small amount of adapter layers in each layer of the pre-trained model, and only update the parameters of these adapter layers during the training process of the downstream task, while freezing the main parameters of the pre-trained model. This can greatly reduce the amount of parameters that need to be trained, reduce the demand for computing resources, and also better retain the general knowledge that has been learned in the pre-trained model. Specifically, the text model and the image model are trained respectively by means of adapter fine-tuning, which means inserting adapter layers in the roberta-base model and the CLIP model respectively, and fine-tuning for specific downstream tasks (i.e., the text and image feature extraction tasks in the present invention). For the text model (roberta-base model), the adapter layer can be inserted after each self-attention module and feedforward neural network module of its Transformer layer. These adapter layers usually include a bottleneck structure that first reduces the dimension and then increases the dimension to reduce the amount of parameters. During the training process, only the parameters of these adapter layers are updated, while the main parameters of the roberta-base model remain unchanged. This allows the roberta-base model to better adapt to text feature extraction tasks, such as sentiment analysis, topic classification, etc., without significantly changing its original language understanding ability. For the image model (CLIP model), an adapter layer can also be inserted after each self-attention module and feedforward neural network module in its Transformer layer. Similar to the text model, these adapter layers also contain a bottleneck structure. During the training process, only the parameters of these adapter layers are updated, while the main parameters of the CLIP model remain unchanged. This allows the CLIP model to better adapt to image feature extraction tasks, such as image classification, object detection, etc., without significantly changing its original image and text cross-modal matching capabilities. Training text models and image models through adapter fine-tuning has the following advantages: high efficiency. Since only a small number of adapter parameters are trained, the demand for computing resources and training time is greatly reduced, especially when using large pre-trained models. The advantage is more obvious. Effectiveness. Adapter fine-tuning can usually achieve performance similar to that of full model fine-tuning, and even better in some cases. Scalability. By inserting different adapter layers, the pre-trained model can be adapted to different downstream tasks without retraining the entire model.

[0062] By training the text model and the image model by using the adapter fine-tuning method, the requirements for computing resources and training time can be significantly reduced while ensuring the performance of the model, and the training efficiency and deployment speed of the model can be improved. This is particularly important for resource-constrained application scenarios. At the same time, adapter fine-tuning can also better retain the general knowledge that has been learned in the pre-trained model, avoid overfitting, and improve the generalization ability of the model. Although the present embodiment preferably uses the adapter fine-tuning method to train the text model and the image model, this is not a limitation of the present invention. Those skilled in the art can select other suitable fine-tuning methods according to specific application scenarios, such as full model fine-tuning, prompt tuning, prefix tuning, etc. As long as the text model and the image model can be effectively trained and the required text and image features can be extracted, they should be regarded as falling within the scope of protection of the present invention. For example, when computing resources are sufficient, full model fine-tuning can be used to obtain higher performance. When the model needs to be deployed quickly, a lighter-weight fine-tuning method such as prompt tuning can be used.

[0063] In some optional embodiments, the keyword model performs keyword matching based on a preset keyword library, where the keyword library includes keywords for representing practicality and keywords for representing timeliness.

[0064] In the embodiment of the present invention, the function of the keyword model is to determine whether the content contains keywords related to practicality and / or timeliness by comparing the content data (mainly text information) with the preset keyword library. Keyword matching can adopt a variety of methods, such as: exact matching, comparing the text with the words in the keyword library word by word to determine whether they are completely consistent; fuzzy matching, allowing certain differences between the text and the words in the keyword library, such as spelling errors, synonyms, etc. It can be measured using methods such as edit distance and cosine similarity; regular expression matching, using regular expressions to define more complex matching rules, such as matching specific word combinations, word patterns, etc. The keyword library is the basis of keyword matching, and its quality directly affects the accuracy of the matching results. The construction of the keyword library can adopt a variety of methods, such as: manual construction, where the keyword library is manually constructed by domain experts based on experience and knowledge. Automatic construction, by analyzing a large amount of text data, extracting words with high frequency and strong discrimination as keywords. Methods such as word frequency statistics and TF-IDF can be used; semi-automatic construction, combining manual construction and automatic construction methods, first extracting candidate keywords by automatic methods, and then manually screening and correcting them.

[0065] In an embodiment of the present invention, the keyword library is divided into two sub-libraries: a practical keyword library and a timeliness keyword library. Practical keywords are used to characterize whether the content has practical use or reference value. For example, for travel strategy content, practical keywords may include "strategy", "guide", "skills", "suggestions", "experience", "route", "itinerary", "list", "essential", "recommended", "notes", etc. Timeliness keywords are used to characterize whether the content has time validity or freshness. For example, timeliness keywords may include "latest", "recent", "recently", "upcoming", "just", "new", "update", "release", "online", "first release", etc., as well as specific dates, times, seasons and other words. By dividing the keyword library into two sub-libraries of practicality and timeliness, keyword matching can be performed more accurately, and more abundant information can be provided for subsequent tag identification. For example, if a travel note contains the two keywords "strategy" and "latest", it can be preliminarily judged that the travel note is both practical and timeliness.

[0066] By using a preset keyword library containing practical and timeliness keywords for keyword matching, it is possible to quickly and effectively make a preliminary judgment on whether the content contains information related to practicality and / or timeliness, and provide important reference information for subsequent refined tag identification, thereby improving the efficiency and recall rate of content screening. At the same time, by dividing the keyword library into two sub-libraries, matching can be performed more accurately, and the accuracy of the matching results can be improved. The specific method of keyword matching and the method of constructing the keyword library can be selected and adjusted according to the specific application scenario. The content of the keyword library can also be updated and expanded according to actual needs. The keywords listed above are only examples and are not limitations of the present invention. For example, different keyword libraries can be constructed according to different OTA platforms and content types. More complex matching algorithms, such as matching algorithms based on semantic similarity, can be used to improve the accuracy of matching.

[0067] In some embodiments, the keywords used to characterize practicality include "traffic", "ticket price", "transportation", "accommodation", "strategy", and "guide"; the keywords used to characterize timeliness include "yesterday", "today", "recently", "upcoming", "exhibition", "event", and "match".

[0068] The keywords used to characterize practicality are closely related to the actual needs of travel and can effectively characterize the practical value of the content. Specifically:

[0069] “Traffic”: refers to the number of tourists at a certain place or attraction. For travelers, understanding the traffic can help them avoid crowded times and places and optimize their travel experience.

[0070] "Ticket price": refers to the ticket price to enter a certain attraction or participate in a certain event. For travelers, knowing the ticket price can help them make budgets and plans.

[0071] "Transportation": refers to the means of transportation to a certain place or attraction, including public transportation, taxis, self-driving, etc. For travelers, understanding transportation information can help them choose the appropriate mode of travel.

[0072] "Accommodation": refers to accommodation arrangements during the trip, including hotels, homestays, youth hostels, etc. For travelers, understanding accommodation information can help them choose the appropriate accommodation location and type.

[0073] "Strategies" and "Guides" refer to documents or materials that provide travel advice, tips, information, etc. For travelers, strategies and guides are important reference materials that can help them better plan their trips and solve problems.

[0074] These keywords cover all aspects that travelers may need to pay attention to before, during and after their trip, and can effectively characterize the practicality of the content.

[0075] Keywords used to represent timeliness are closely related to time and can effectively represent the timeliness of content. Specifically:

[0076] "Yesterday" and "Today": refer to the day before and today of the current date, and are used to indicate that the content has just happened or is happening.

[0077] "Recently": refers to a period of time in the past. The specific time range can be defined based on actual conditions, such as the past week, the past month, etc.

[0078] "Soon": refers to a period of time in the future. The specific time range can also be defined according to actual conditions, such as the next week, the next month, etc.

[0079] "Exhibition", "activity", "match": These words are usually associated with a specific time period, such as the date of an exhibition, the time of an activity, the time of a match, etc.

[0080] These keywords can effectively help identify whether the content is timely, such as the latest policies of a certain attraction, the latest arrangements of an event, etc.

[0081] By applying these practical and timeliness keywords to the keyword model, it is possible to quickly and preliminarily determine whether the content contains information related to practicality and / or timeliness, and provide an important reference for subsequent refined tag identification. By using these specific practical and timeliness keywords, keyword matching can be performed more effectively, and the accuracy and efficiency of preliminary judgments can be improved, thereby providing more reliable input for subsequent refined tag identification, and ultimately improving the overall effect of content tag identification. For example, content containing the latest travel information or practical strategies can be screened out more quickly to better meet the needs of users. The keywords listed above are only examples and are not limitations of the present invention. Those skilled in the art can add, delete or modify keywords according to the specific application scenarios and content characteristics of the OTA platform. For example, more targeted keywords can be added according to different tourist destinations or travel themes. More complex matching rules can be used, such as matching synonyms, near synonyms or related words of keywords to improve the recall rate of matching.

[0082] In some optional embodiments, the large text model is a qwen2-7b model. The qwen2-7b model shows strong performance and multi-language support capabilities in natural language processing tasks. qwen2-7b is a large language model with 7 billion parameters based on Transformer, which is trained on ultra-large-scale pre-training data. The pre-training data types are diverse and cover a wide range, including a large number of online texts, professional books, codes, etc. Compared with previous versions, the qwen2 series models have significantly improved in performance, especially in multi-language processing, long text understanding and complex problem solving. In an embodiment of the present invention, the content data, preliminary tags and keyword matching results are input into the qwen2-7b model, and combined with preset prompt words, the model can be guided to better understand the task objectives and output more accurate practical labels or timeliness labels. For example, by designing appropriate prompt words, the model can be guided to focus on key information in the content, such as "Does this guide provide practical travel advice?" or "Does the event described in this travel note happen recently?".

[0083] By adopting the qwen2-7b model as the large text model, its powerful language understanding ability and multilingual support ability can be fully utilized to more accurately identify the practicality and timeliness information in the content data, thereby improving the efficiency and accuracy of content screening and analysis. Especially when processing OTA platform content containing complex semantics and multiple languages, this embodiment can more effectively improve the quality and efficiency of content analysis. At the same time, the relatively moderate model scale of the qwen2-7b model also makes the present invention easier to deploy and use in practical applications. Although the present embodiment preferably adopts the qwen2-7b model as the large text model, this is not a limitation of the present invention. Those skilled in the art can select other suitable large text models according to specific application scenarios and requirements, such as other versions of the qwen model, the llama series model, the glm series model, etc. As long as the effective understanding and processing of text information can be achieved, and the practicality or timeliness label is output in combination with the preliminary label and keyword matching results, it should be regarded as falling within the scope of protection of the present invention. For example, when computing resources are limited, a model with a smaller parameter amount can be selected to improve processing efficiency; when higher performance is required, a model with a larger parameter amount can be selected.

[0084] In some optional embodiments, the prompt words include the title, text, preliminary labels and keyword matching results of the content data. In an embodiment of the present invention, the prompt words can guide the text big model to better understand the task objectives and output more accurate practicality or timeliness labels. The prompt word composition method adopted in this embodiment comprehensively considers the various aspects of the content data, and can provide more comprehensive context information for the text big model. Specifically: the title of the content data, the title is usually a high-level summary of the content, contains the most important subject information, and can help the text big model quickly understand the core content of the content. The text of the content data, the text is the specific development of the content, contains detailed descriptions and explanations, and can provide richer details for the text big model. Preliminary labels, the preliminary labels are the labels output by the previous steps (such as the image and text labeling model), which can provide preliminary judgment results for the text big model as reference information. Keyword matching results, the keyword matching results are the results output by the keyword model, indicating whether the content contains keywords related to practicality and / or timeliness, and can provide direct evidence for the text big model. By combining the above information into prompt words, more comprehensive context information can be provided for the text big model, so as to more accurately judge the practicality and timeliness of the content. The specific form of the prompt word can be adjusted according to the text model used and the specific application scenario. For example, task instructions described in natural language can be used, or structured data formats can be used. The key is to clearly express the task goal and provide sufficient context information.

[0085] By using prompt words containing various information, more comprehensive contextual information can be provided for the text large model, so as to more accurately judge the practicality and timeliness of the content. This helps to improve the accuracy and efficiency of content tag recognition, and ultimately improves the user experience. For example, it is possible to more accurately screen out content that is of practical help to users and has the latest information, so as to better meet the needs of users. Although the present embodiment preferably uses prompt words containing title, text, preliminary tags and keyword matching results, this is not a limitation of the present invention. Those skilled in the art can adjust the composition of prompt words according to specific application scenarios and requirements. For example, other information such as the source, author, release time, etc. of the content can be added as needed. Different prompt word templates or strategies can be used to optimize the performance of the text large model.

[0086] In some optional embodiments, the method further includes: filtering the practicality tags and timeliness tags output by the large text model according to preset rules. Although the large text model has a strong language understanding ability, in some cases, the tags it outputs may have certain deviations or uncertainties. Therefore, it is necessary to filter by preset rules to improve the quality of the tags. These preset rules can be designed based on a variety of factors, such as: confidence threshold, the large text model usually assigns a confidence score to each output tag, indicating the degree of certainty of the model for the tag. A confidence threshold can be set to retain only tags with a confidence higher than the threshold. For example, the confidence threshold can be set to 0.8, and only tags that the model believes are more than 80% sure are retained. Tag conflict detection, some tags may be mutually exclusive, such as "high practicality" and "low practicality" are a pair of mutually exclusive tags. If the model outputs these two tags at the same time, they need to be selected or eliminated according to certain rules. For example, the confidence of the two tags can be compared, and the tag with higher confidence can be retained. Keyword verification, the tags output by the large text model can be compared with the keyword matching results to verify the rationality of the tags. For example, if the label output by the model is "high timeliness", but the keyword matching results do not contain any timeliness keywords, it can be considered that the label may be wrong and needs to be removed or re-evaluated. Context consistency check, you can check whether the label is consistent with the context of the content data. For example, if the label output by the model is "high practicality", but the content data does not provide any actual suggestions or information, it can be considered that the label may be wrong. Manual review, for some labels that are difficult to judge through automatic rules, manual review can be performed to ensure the accuracy of the label. This is usually used to deal with some boundary cases or complex situations. The specific form of the preset rules can be adjusted according to the specific application scenarios and needs. For example, different confidence thresholds can be set according to different content types or label types. More complex rule combinations can be used to improve the accuracy of filtering.

[0087] By filtering the labels output by the large text model according to preset rules, the accuracy and reliability of the labels can be effectively improved, and the impact of incorrect labels on subsequent processing can be reduced. This helps to improve the overall effect of content label recognition and ultimately improve the user experience. For example, it is possible to more accurately recommend high-quality content that meets the needs of users. Although this embodiment describes some commonly used filtering rules, this is not a limitation of the present invention. Those skilled in the art can design other suitable filtering rules based on specific application scenarios and requirements. For example, machine learning methods can be used to automatically learn filtering rules.

[0088] In some optional embodiments, the preset rules include word count limit and time limit. These two limits are mainly used to filter the content data itself, rather than the output tags of the text large model. They are usually used to filter content in the earlier preprocessing stage to reduce the amount of data for subsequent processing and improve efficiency. The word count limit refers to limiting the text length of the content data. Some content may be too short or too long and not suitable for subsequent tag recognition. For example, some comments may have only a few words and cannot provide enough information; some articles may be too long and exceed the processing capacity of the text large model. Therefore, a word count range can be set to retain only the content within the range. For example, the minimum word count can be set to 10 words and the maximum word count can be set to 10,000 words. The word count limit needs to be adjusted according to the specific application scenario and content type. For example, for short text platforms such as Weibo, a smaller maximum word count limit can be set; for long text platforms such as news, a larger maximum word count limit can be set. Time limit refers to limiting the release time of content data. Some content may be too old, not timely, and does not meet user needs. For example, some travel guides from a few years ago may be outdated and no longer applicable. Therefore, a time range can be set to retain only content published within that time range. For example, only content published within the last year can be retained. The setting of time limits also needs to be adjusted according to specific application scenarios and content types. For example, for content with high timeliness requirements such as news, a shorter time range can be set; for content with low timeliness requirements such as historical materials, a longer time range can be set. Word count limits and time limits are usually applied in the content preprocessing stage to filter content data before inputting it into subsequent models (such as graphic and text tagging models, keyword models, and large text models). This can effectively reduce the amount of data to be processed subsequently, improve processing efficiency, and avoid the model processing inappropriate content.

[0089] By filtering content data using preset rules such as word limit and time limit, the amount of data for subsequent processing can be effectively reduced, processing efficiency can be improved, and the model can be prevented from processing inappropriate content, thereby improving the overall efficiency and quality of content analysis. For example, it is possible to screen out high-quality content that meets the requirements more quickly and reduce unnecessary consumption of computing resources. Although this embodiment describes two commonly used filtering rules, word limit and time limit, this is not a limitation of the present invention. Those skilled in the art can design other suitable filtering rules based on specific application scenarios and requirements. For example, filtering can be performed based on factors such as the quality, source, and author of the content.

[0090] In some optional embodiments, the vllm reasoning framework is used to accelerate the reasoning process of large text models. vllm is an efficient large language model (LLM) reasoning and service library based on Python. It uses technologies such as PagedAttention to significantly improve the reasoning speed and throughput of LLM, while reducing video memory usage and making more efficient use of hardware resources. This is crucial for deploying and using large language models in practical applications, especially when processing large amounts of content data. The main advantages of using vllm to accelerate the reasoning process of large text models include: PagedAttention, which is inspired by virtual memory and paging management in the operating system, blocks the KV cache (key-value cache) of each sequence, dynamically allocates KV cache video memory to requests, greatly improves video memory utilization, and can support larger models and longer sequences. Continuous batch processing, vllm supports continuous batch processing requests, and can merge multiple requests into one batch for processing, making full use of the parallel computing capabilities of the GPU and improving reasoning throughput. Efficient memory management, in addition to PagedAttention, vllm also uses other memory optimization technologies, such as zero copy, to further reduce memory usage and data transmission overhead. Ease of use: vllm is seamlessly integrated with the Hugging Face Transformers library, making it easy to load and use various pre-trained models, including qwen2-7b. Compared to using the Hugging Face Transformers library directly for inference, using vllm can significantly increase the inference speed, especially when processing multiple requests or long texts. This is because vllm manages GPU memory more effectively through technologies such as PagedAttention, avoiding problems such as memory overflow, thereby better utilizing the performance of the GPU.

[0091] By using the vllm reasoning framework to accelerate the reasoning process of the large text model, the efficiency of content analysis can be significantly improved, the consumption of computing resources can be reduced, and the needs of practical applications can be better met. Especially when processing massive OTA platform content data, this embodiment can more effectively improve the speed and efficiency of content analysis, thereby providing services to users faster. Although the vllm reasoning framework is preferably used in this embodiment, this is not a limitation of the present invention. Those skilled in the art can select other suitable reasoning acceleration frameworks or technologies, such as DeepSpeed, FasterTransformer, etc., according to specific application scenarios and requirements. As long as the reasoning efficiency of the large text model can be effectively improved, it should be deemed to fall within the scope of protection of the present invention.

[0092] The embodiment of the present invention further provides a content tag identification system, which is used to implement the steps of the content tag identification method.

[0093] In the content tag identification method of this embodiment, the serial number of each step is only used to distinguish the steps, but not to limit the specific execution order of the steps. The execution order of the above steps can be adjusted and changed as needed.

[0094] In the content tag identification system of the present invention, the functions of each module can be implemented by using the specific implementation of the content tag identification method as described above, which will not be described in detail here.

[0095] An embodiment of the present invention further provides a content tag identification device, comprising a processor; a memory, in which executable instructions of the processor are stored; wherein the processor is configured to execute the steps of the content tag identification method by executing the executable instructions.

[0096] It will be appreciated by those skilled in the art that various aspects of the present invention may be implemented as systems, methods or program products. Therefore, various aspects of the present invention may be specifically implemented in the following forms, namely: complete hardware implementation, complete software implementation (including firmware, microcode, etc.), or a combination of hardware and software implementations, which may be collectively referred to herein as "circuits", "modules" or "platforms".

[0097] Refer to the following Figure 2 The electronic device 600 according to this embodiment of the present invention is described. Figure 2 The electronic device 600 shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0098] like Figure 2As shown, the electronic device 600 is in the form of a general computing device. The components of the electronic device 600 may include, but are not limited to: at least one processing unit 610, at least one storage unit 620, a bus 630 connecting different system components (including the storage unit 620 and the processing unit 610), a display unit 640, etc.

[0099] The storage unit stores program codes, which can be executed by the processing unit 610, so that the processing unit 610 performs the steps according to various exemplary embodiments of the present invention described in the content tag identification method section of this specification. For example, the processing unit 610 can perform the following steps: Figure 1 Follow the steps shown in .

[0100] The storage unit 620 may include a readable medium in the form of a volatile storage unit, such as a random access memory unit (RAM) 6201 and / or a cache memory unit 6202 , and may further include a read-only memory unit (ROM) 6203 .

[0101] The storage unit 620 may also include a program / utility 6204 having a set (at least one) of program modules 6205, such program modules 6205 including but not limited to: an operating system, one or more application programs, other program modules and program data, each of which or some combination may include the implementation of a network environment.

[0102] Bus 630 may represent one or more of several types of bus structures, including a memory unit bus or memory unit controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of a variety of bus architectures.

[0103] The electronic device 600 may also communicate with one or more external devices 700 (e.g., keyboards, pointing devices, Bluetooth devices, etc.), may also communicate with one or more devices that enable a user to interact with the electronic device 600, and / or communicate with any device that enables the electronic device 600 to communicate with one or more other computing devices (e.g., routers, modems, etc.). Such communication may be performed via an input / output (I / O) interface 650. Furthermore, the electronic device 600 may also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) via a network adapter 660. The network adapter 660 may communicate with other modules of the electronic device 600 via a bus 630. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with the electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0104] In the content tag identification device, the program in the memory implements the steps of the content tag identification method when being executed by the processor, and therefore, the device can also obtain the technical effect of the above-mentioned content tag identification method.

[0105] The embodiment of the present invention also provides a computer-readable storage medium for storing a program, which implements the steps of the content tag identification method when executed by a processor. In some possible implementations, various aspects of the present invention can also be implemented in the form of a program product, which includes a program code. When the program product is executed on a terminal device, the program code is used to enable the terminal device to perform the steps of various exemplary implementations of the present invention described in the above content tag identification method section of this specification.

[0106] refer to Figure 3 As shown, a program product 800 for implementing the above method according to an embodiment of the present invention is described, which can adopt a portable compact disk read-only memory (CD-ROM) and include program code, and can be executed on a terminal device, such as a personal computer. However, the program product of the present invention is not limited thereto. In this document, a readable storage medium can be any tangible medium containing or storing a program, which can be used by or in combination with an instruction execution system, an apparatus or a device.

[0107] The program product may use any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0108] The computer readable storage medium may include a data signal propagated in a baseband or as part of a carrier wave, wherein a readable program code is carried. This propagated data signal may take a variety of forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination of the above. The readable storage medium may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by an instruction execution system, an apparatus, or a device or used in combination with it. The program code contained on the readable storage medium may be transmitted with any appropriate medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination of the above.

[0109] Program code for performing the operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a separate software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., through the Internet using an Internet service provider).

[0110] When the program in the computer storage medium is executed by the processor, the steps of the content tag identification method are implemented. Therefore, the computer storage medium can also obtain the technical effect of the content tag identification method.

[0111] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A content tag identification method for analyzing the practicability and timeliness of OTA platform content, characterized in that: The steps include: Acquire content data on an OTA platform, the content data including text information and image information; Using at least one graphic labeling model to perform preliminary label recognition on the content data to obtain at least one preliminary label; Perform keyword matching on the content data using a keyword model to obtain at least one keyword matching result; Inputting the content data, the preliminary tags and the keyword matching results into a text macro model; The text macro model outputs a practicality label or a timeliness label of the content data based on a preset prompt word.

2. The content tag identification method according to claim 1, characterized in that: The image and text marking model includes a text model and an image model, wherein: The text model is used to process the text information in the content data and output a text score; The image model is used to process the image information in the content data and output an image score; The text score and the image score are combined to obtain a final score, and a determination is made based on a preset threshold whether to assign a corresponding preliminary label to the content data.

3. The content tag identification method according to claim 2, characterized in that: The text model and the image model are trained separately by adapter fine-tuning.

4. The content tag identification method according to claim 1, characterized in that: The keyword model performs keyword matching based on a preset keyword library, wherein the keyword library includes keywords for representing practicality and keywords for representing timeliness.

5. The content tag identification method according to claim 4, characterized in that: The keywords used to characterize practicality include "traffic flow", "ticket price", "transportation", "accommodation", "strategy", and "guide"; the keywords used to characterize timeliness include "yesterday", "today", "recently", "upcoming", "exhibition", "activity", and "match".

6. The content tag identification method according to claim 1, characterized in that: The prompt words include the title, text, preliminary tags and keyword matching results of the content data.

7. The content tag identification method according to claim 1, characterized in that: The method further includes: filtering the practicality labels and timeliness labels output by the large text model according to preset rules.

8. The content tag identification method according to claim 1, characterized in that: A parallel reasoning framework is used to accelerate the reasoning process of the large text model.

9. A content tag identification system, characterized in that: Used to implement the content tag identification method described in any one of claims 1 to 8.

10. A content tag recognition device, characterized in that: include: processor; a memory storing executable instructions of the processor; The processor is configured to execute the steps of the content tag identification method according to any one of claims 1 to 8 by executing the executable instructions.

Images