Information generation model training method and information generation method

Abstract

The present disclosure provides a training method of an information generation model and an information generation method, relates to the technical field of artificial intelligence, in particular to the technical field of natural language processing, deep learning, large language model and the like, and can be applied to the intelligent recommendation scene. The method comprises the following steps: inputting collected multi-source signal data into a large language model to obtain at least one candidate topic; determining a positive example topic and a negative example topic from the at least one candidate topic; generating a positive training sample pair and a negative training sample pair according to the multi-source signal data, the positive example topic and the negative example topic; and training the large language model by using the positive training sample pair and the negative training sample pair to obtain an information generation model. The present disclosure improves the efficiency and accuracy of the information generation model in generating topic information, and reduces the same error generation rate of the information generation model.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, particularly to the fields of natural language processing, deep learning, and large language models, and can be applied to intelligent recommendation scenarios. Specifically, it relates to a training method for an information generation model and an information generation method. Background Technology

[0002] In today's rapidly evolving internet content ecosystem, recommendation systems, as the core bridge connecting users with massive amounts of content, directly determine user experience and product commercial value through their ability to perceive real-time trending topics and the accuracy of content distribution. Topic generation, a crucial link in the content operation and trending topic distribution of recommendation systems, can assign precise trending tags and dissemination directions to recommended content, helping the system quickly capture user interests and improve content exposure efficiency. Therefore, it is widely used in recommendation systems for various internet products such as news recommendations, social platforms, and short video distribution. Summary of the Invention

[0003] This disclosure proposes a training method for an information generation model and an information generation method.

[0004] According to a first aspect of this disclosure, a training method for an information generation model is provided, comprising: inputting collected multi-source signal data into a large language model to obtain at least one candidate topic; determining positive example topics and negative example topics from the at least one candidate topic; generating positive training sample pairs and negative training sample pairs based on the multi-source signal data, positive example topics, and negative example topics; and training the large language model using the positive training sample pairs and negative training sample pairs to obtain an information generation model.

[0005] According to a second aspect of this disclosure, an information generation method is provided, comprising: acquiring target data and target topic style; inputting the target data and target topic style into an information generation model, and outputting at least one target topic, wherein the information generation model is trained using the method described in the first aspect.

[0006] According to a third aspect of this disclosure, a training apparatus for an information generation model is provided, comprising: a topic generation module configured to input collected multi-source signal data into a large language model to obtain at least one candidate topic; a topic determination module configured to determine positive example topics and negative example topics from the at least one candidate topic; a training sample generation module configured to generate positive training sample pairs and negative training sample pairs based on the multi-source signal data, positive example topics, and negative example topics; and a training module configured to train the large language model using the positive training sample pairs and negative training sample pairs to obtain an information generation model.

[0007] According to a fourth aspect of this disclosure, an information generation apparatus is provided, comprising: an acquisition module configured to acquire target data and a target theme style; and an output module configured to input the target data and the target theme style into an information generation model and output at least one target theme, wherein the information generation model is trained using the method described in the first aspect.

[0008] According to a fifth aspect of this disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform a method as described in any implementation of the first or second aspect.

[0009] According to a sixth aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions for causing a computer to perform a method as described in any implementation of the first or second aspect.

[0010] According to a seventh aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the method as described in any of the implementations of the first or second aspect.

[0011] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0012] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 This is an exemplary system architecture diagram to which this disclosure can be applied; Figure 2 This is a flowchart of the first embodiment of a training method for generating a model based on the information disclosed herein; Figure 3 This is a flowchart of a second embodiment of a training method for generating a model based on the information disclosed herein; Figure 4 This is a flowchart of a third embodiment of a training method for generating a model based on the information disclosed herein; Figure 5 yes Figure 4 A flowchart of an embodiment of step 403; Figure 6 This is a flowchart of the fourth embodiment of the training method for generating a model based on the information disclosed herein; Figure 7This is a flowchart of an embodiment of the information generation method according to the present disclosure; Figure 8 This is a flowchart of another embodiment of the information generation method according to this disclosure; Figure 9 This is a schematic diagram of the structure of an embodiment of a training apparatus for generating a model based on the information in this disclosure; Figure 10 This is a schematic diagram of the structure of an embodiment of the information generation apparatus according to the present disclosure; Figure 11 This is a block diagram of an electronic device used to implement the training method or information generation method of the information generation model in the embodiments of this disclosure. Detailed Implementation

[0013] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0014] It should be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other. This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0015] Figure 1 An exemplary system frame 100 is shown, illustrating embodiments of the training method, training apparatus, or information generation apparatus of the information generation model of this disclosure.

[0016] like Figure 1 As shown, system architecture 100 may include terminal devices 101, 102, and 103, a network 104, and a server 105. Network 104 serves as the medium for providing communication links between terminal devices 101, 102, and 103 and server 105. Network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0017] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various applications for enabling information communication between the terminal devices 101, 102, and 103 and server 105 can be installed. These applications include cloud storage applications and instant messaging applications.

[0018] Terminal devices 101, 102, and 103 and server 105 can be either hardware or software. When terminal devices 101, 102, and 103 are hardware, they can be various electronic devices with displays, including but not limited to smartphones, tablets, laptops, and desktop computers. When terminal devices 101, 102, and 103 are software, they can be installed in the aforementioned electronic devices, and can be implemented as multiple software programs or software modules, or as a single software program or software module; no specific limitation is made here. When server 105 is hardware, it can be implemented as a distributed server cluster composed of multiple servers, or as a single server. When server 105 is software, it can be implemented as multiple software programs or software modules, or as a single software program or software module; no specific limitation is made here.

[0019] Server 105 can provide various services through its built-in applications. For example, users can operate the server through applications on terminal devices 101, 102, and 103 and send model training requests to server 105. Server 105 can receive and process these model training requests, performing the following processes: inputting the collected multi-source signal data into a large language model to obtain at least one candidate topic; determining positive and negative examples from the at least one candidate topic; generating positive and negative training sample pairs based on the multi-source signal data, positive and negative examples; and training the large language model using the positive and negative training sample pairs to obtain an information generation model.

[0020] It should be noted that the training method or information generation method of the information generation model provided in this embodiment is generally executed by the server 105, and correspondingly, the training device or information generation device of the information generation model is generally set in the server 105.

[0021] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0022] Continue to refer to Figure 2 The diagram illustrates a flow 200 of a first embodiment of a training method for an information generation model according to the present disclosure. The training method for the information generation model includes the following steps: Step 201: Input the collected multi-source signal data into the large language model to obtain at least one candidate topic.

[0023] In this embodiment, the execution entity of the training method for the information generation model (e.g.) Figure 1The server 105 shown will crawl data from the internet and other channels in real time or collect multi-source data through API (Application Programming Interface) interfaces. Multi-source data refers to data collected from multiple data sources, including but not limited to: social media data sources, news data sources, search data sources, and multimodal data sources. Social media data sources include commonly used social media apps (Applications), news data sources include news apps and news published and reported by news media, search data sources include search terms from search engines, and multimodal data sources include titles, subtitles, and other information from popular short videos. In other words, the aforementioned execution entity will collect multi-source signal data in real time from news, social media, user history behavior, and trending events.

[0024] Since the collected multi-source signal data is generally unstructured data, the aforementioned execution entity will preprocess the multi-source signal data using NLP (Natural Language Processing) technology, thereby transforming the multi-source signal data into structured signal vectors that can be understood by large models. The NLP technology here can include methods such as keyword extraction and topic clustering.

[0025] Subsequently, the aforementioned execution entity will input the structured signal vectors corresponding to the multi-source signal data into the basic large language model, so as to utilize the semantic understanding and topic generation capabilities of the basic large model to generate a batch of topic information, which will be recorded as candidate topics.

[0026] It's important to note that in the field of artificial intelligence, large models refer to deep neural networks with over 1 billion parameters. These networks are capable of processing massive amounts of data and performing various complex tasks, such as natural language processing, computer vision, and speech recognition. Generative large models, on the other hand, are generative models based on large-scale corpora. They are large-scale neural network models capable of generating, understanding, and reasoning about natural language in an end-to-end manner. By training on large amounts of text data, they can perform a wide range of tasks, including text summarization, translation, sentiment analysis, and more. With the continuous improvement of computer hardware performance and the optimization of deep learning algorithms, the development of large models is accelerating. The parameter scale of large models is constantly expanding, and the training time is increasing, but the performance is also improving. Large models are typically based on deep learning architectures, such as Transformers, enabling them to demonstrate impressive capabilities in various natural language processing tasks. Common large models include, but are not limited to, ChatGPT, GPT-4, ERNIE, and the Lingyi Bot.

[0027] Step 202: Determine positive and negative topics from at least one candidate topic.

[0028] In this embodiment, the execution entity determines positive and negative examples from a batch of generated candidate topics. Specifically, the execution entity can evaluate and score each generated candidate topic according to preset multi-dimensional evaluation criteria, thereby identifying candidate topics with high overall scores as positive examples and those with low overall scores as negative examples. Positive examples are samples for the model's positive learning, while negative examples are samples that the model needs to avoid and correct in the future. It should be noted that there are generally multiple positive and negative examples.

[0029] Here, the aforementioned execution entity will perform automated, multi-dimensional evaluation of candidate topics. These evaluation dimensions can include, but are not limited to: relevance evaluation, popularity evaluation, novelty evaluation, and quality evaluation. Relevance evaluation refers to assessing the correlation between the candidate topic and the signal data that generated it; semantic relevance models can be used for this evaluation. Popularity evaluation assesses the potential spread of the candidate topic; this can be based on a spread model of historically similar topics. Novelty evaluation assesses the distinguishability of the candidate topic from existing topics in the topic library, thus determining its novelty; vector distance and semantic repetition can be used for this evaluation. Quality evaluation assesses the topic's grammatical correctness, fluency, information content, and security; this can be determined using a series of classifiers or rules.

[0030] Step 203: Generate positive training sample pairs and negative training sample pairs based on multi-source signal data, positive example topics, and negative example topics.

[0031] In this embodiment, the execution entity generates positive training sample pairs and negative training sample pairs based on multi-source signal data, positive example topics, and negative example topics. Specifically, the execution entity determines the signal data from the multi-source signal data that generates positive example topics, uses it as positive example signal data, and pairs the positive example topics and positive example signal data as positive training sample pairs.

[0032] For negative example topics, the aforementioned execution entity determines the dimensions to be repaired based on the topic's score. Then, it corrects the negative example topic according to pre-defined repair rules corresponding to the dimensions to be repaired, resulting in a corrected topic. Since the repair rules correspond to the defect dimensions, correcting negative example topics according to the repair rules can avoid correction biases caused by general rules, ensuring the accuracy of the corrected topic.

[0033] The aforementioned execution entity will pair negative example topics with corrected topics as negative training sample pairs.

[0034] Step 204: Train the large language model using positive training sample pairs and negative training sample pairs to obtain the information generation model.

[0035] In this embodiment, the aforementioned execution entity trains the initial large model using positive and negative training sample pairs to obtain a trained information generation model. Specifically, the positive and negative training sample pairs are used as training samples to fine-tune the initial large model, enabling it to learn "what style and quality of topic should be generated when seeing a certain world signal." For example, when the input data is news channel data, the large model should generate rigorous topics that conform to news style and quality requirements.

[0036] It should be noted that the above model training process can be carried out periodically or triggered, so that the capabilities of large models can be continuously enhanced through continuous evaluation, selection and adaptation, just like biological evolution.

[0037] The information generation model training method disclosed herein first inputs collected multi-source signal data into a large language model to obtain at least one candidate topic; then, positive and negative examples are determined from the at least one candidate topic; subsequently, positive and negative training sample pairs are generated based on the multi-source signal data, positive and negative examples; finally, the large language model is trained using the positive and negative training sample pairs to obtain the information generation model. This method, through positive learning of positive training sample pairs, enables the model to solidify the generation logic of high-quality topics (such as keyword fusion, expression style, and domain adaptation); through negative error correction learning of negative training sample pairs, the model can accurately identify and avoid generation defects of low-quality topics (such as insufficient relevance, low novelty, and quality defects). Therefore, it improves the efficiency and accuracy of the information generation model in generating topic information and reduces the similar error generation rate of the information generation model. Furthermore, the training process uses actual generated candidate topics as samples and multi-dimensional evaluation results as optimization criteria. The sample pairs are strongly bound to actual business scenarios, eliminating the need for training with massive general corpora. Model capabilities can be upgraded simply by efficiently fine-tuning parameters, reducing the consumption of training resources. It also supports periodic or triggered iterative training, and the model capabilities evolve dynamically with hot trends and business needs, achieving a closed-loop self-evolution of generation-evaluation-training-optimization.

[0038] Furthermore, the collection, storage, use, processing, transmission, provision, and disclosure of any type of information, such as user personal information, involved in the technical solutions disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0039] Continue to refer to Figure 3 , Figure 3A flow 300 of a second embodiment of a training method for an information generation model according to the present disclosure is shown. The training method for the information generation model includes the following steps: Step 301: Preprocess the multi-source signal data to obtain a structured signal vector.

[0040] In this embodiment, the execution entity of the training method for the information generation model (e.g.) Figure 1 The server 105 shown will preprocess the multi-source signal data to obtain a structured signal vector. Since multi-source signal data is generally unstructured data, the aforementioned execution entity will first preprocess the multi-source signal data to obtain the processed structured signal vector.

[0041] In some optional implementations of this embodiment, step 301 includes: Step 3011: Use natural language processing techniques and feature extraction methods to extract multi-dimensional features from multi-source signal data.

[0042] The aforementioned execution entity will analyze the multi-source signal data using NLP technology, which may include methods such as keyword extraction and topic clustering. Then, feature extraction methods will be used to extract multi-dimensional features from the analyzed signal data. These multi-dimensional features may include, but are not limited to, semantic features, attribute features, and statistical features.

[0043] Step 3012: Based on the data sources of the multi-source signal data, the multi-dimensional features are fused to obtain a structured signal vector.

[0044] The aforementioned execution entity determines the data sources of the multi-source signal data, that is, it divides the multi-source signal data according to the data source to obtain the signal data corresponding to each data source. Then, it fuses the extracted multi-dimensional features to obtain a structured signal vector.

[0045] By relying on NLP technology and feature extraction methods, the features of multi-source signal data can be captured in all dimensions, realizing refined and multi-dimensional feature extraction of multi-source signal data.

[0046] In some optional implementations of this embodiment, step 3012 includes: fusing multi-dimensional features corresponding to signal data from the same data source to obtain a single-source signal vector; fusing features of all single-source signal vectors and standardizing the fused vectors to obtain a structured signal vector.

[0047] The aforementioned execution entity divides the multi-source signal data according to the data source, obtaining the signal data corresponding to each data source, that is, determining the signal data from the same data source, and fusing the multi-dimensional features corresponding to the signal data from the same data source to obtain multiple single-source signal vectors. Then, feature fusion is performed on all single-source signal vectors, and the fused vectors are standardized to obtain the structured signal vectors corresponding to the multi-source signal data.

[0048] This approach first aggregates and purifies the features of signals from the same source, enhancing the representational ability of single-source signals. Then, it fuses multi-source single vectors to construct a full-dimensional feature representation, thereby improving the comprehensiveness of structured signal vectors.

[0049] Step 302: Input the structured signal vector into the large language model and output at least one candidate topic.

[0050] In this embodiment, the aforementioned execution entity will input the structured signal vectors corresponding to the multi-source signal data into the basic large language model, so as to utilize the semantic understanding and topic generation capabilities of the basic large model to generate a batch of topic information, which will be recorded as candidate topics.

[0051] Step 303: Determine positive and negative examples from at least one candidate topic.

[0052] Step 304: Generate positive training sample pairs and negative training sample pairs based on multi-source signal data, positive example topics, and negative example topics.

[0053] Step 305: Train the large language model using positive training sample pairs and negative training sample pairs to obtain the information generation model.

[0054] Steps 303-305 are basically the same as steps 202-204 in the aforementioned embodiments. For specific implementation methods, please refer to the aforementioned description of steps 202-204, which will not be repeated here.

[0055] from Figure 3 It can be seen from this that, with Figure 2 Compared to the corresponding embodiments, the training method of the information generation model in this embodiment emphasizes the step of generating candidate topics using a large language model. This method generates structured signal vectors through preprocessing of multi-source signal data, achieving standardized conversion of multi-source heterogeneous data and thus improving the efficiency and accuracy of candidate topic generation.

[0056] Continue to refer to Figure 4 , Figure 4 A flow 400 of a third embodiment of a training method for an information generation model according to the present disclosure is shown. The training method for the information generation model includes the following steps: Step 401: Preprocess the multi-source signal data to obtain a structured signal vector.

[0057] Step 402: Input the structured signal vector into the large language model and output at least one candidate topic.

[0058] Steps 401-402 are basically the same as steps 301-302 in the aforementioned embodiments. For specific implementation methods, please refer to the aforementioned description of steps 301-302, which will not be repeated here.

[0059] Step 403: Evaluate the candidate topics according to the preset multi-dimensional evaluation criteria to obtain the evaluation results.

[0060] In this embodiment, the execution entity of the training method for the information generation model (e.g.) Figure 1 The server 105 shown will evaluate and score each generated candidate topic according to the preset multi-dimensional evaluation criteria, thereby obtaining the evaluation result, which is the evaluation score.

[0061] The evaluation dimensions here can include, but are not limited to: relevance evaluation, popularity evaluation, novelty evaluation, and quality evaluation. Relevance evaluation refers to assessing the correlation between candidate topics and the signal data that generated them, which can be done using a semantic relevance model. Popularity evaluation assesses the potential spread of candidate topics, which can be done using a spread model based on historically similar topics. Novelty evaluation assesses the distinguishability of candidate topics from topics in the existing topic library to determine their novelty, which can be done using vector distance and semantic repetition. Quality evaluation assesses the grammatical correctness, fluency, information content, and security of the topic, which can be determined using a series of classifiers or rules.

[0062] Step 404: Based on the evaluation results, determine the positive and negative examples from at least one candidate topic.

[0063] In this embodiment, the aforementioned execution entity determines positive and negative examples from multiple candidate topics based on the evaluation results. Specifically, candidate topics with high evaluation scores are designated as positive examples, and those with low scores are designated as negative examples. Positive examples are samples for the model's positive learning, while negative examples are samples that the model needs to avoid and correct in the future. It should be noted that there are generally multiple positive and negative examples.

[0064] Step 405: Generate positive training sample pairs and negative training sample pairs based on multi-source signal data, positive example topics, and negative example topics.

[0065] Step 406: Train the large language model using positive training sample pairs and negative training sample pairs to obtain the information generation model.

[0066] Steps 405-406 are basically the same as steps 203-204 in the aforementioned embodiments. For specific implementation methods, please refer to the aforementioned description of steps 203-204, which will not be repeated here.

[0067] from Figure 4 It can be seen from this that, with Figure 3 Compared with the corresponding embodiments, the training method of the information generation model in this embodiment emphasizes the steps of determining positive and negative topics, thereby quantifying and scoring candidate topics through a multi-dimensional quantitative evaluation system, improving the accuracy and comprehensiveness of the evaluation results of candidate topics, and screening positive and negative topics based on the quantitative evaluation results, thus realizing the hierarchical extraction of samples.

[0068] Continue to refer to Figure 5 , Figure 5 It shows Figure 4 The process 500 of the embodiment of step 403 includes: Step 501: Evaluate the semantic relevance between candidate topics and the signal data that generated the candidate topics according to the semantic dimension evaluation criteria, and obtain the semantic dimension evaluation results.

[0069] The multi-dimensional evaluation criteria include semantic dimension evaluation criteria and quality dimension evaluation criteria. The aforementioned implementing entity will evaluate the correlation between candidate topics and the signal data that generated candidate topics based on the semantic dimension evaluation criteria, thereby obtaining the semantic dimension evaluation results. Here, a semantic correlation model can be used to evaluate the correlation.

[0070] Step 502: Evaluate the quality of the candidate topics according to the quality dimension evaluation criteria to obtain the quality dimension evaluation results.

[0071] The aforementioned implementing entity will evaluate the quality of candidate topics according to quality dimension evaluation criteria. Here, a semantic relevance model can be used to evaluate relevance, thereby obtaining the quality dimension evaluation results. Quality evaluation refers to assessing the grammatical correctness, fluency, information content, and security of the topic, which can be determined through a series of classifiers or rules.

[0072] Step 503: Generate evaluation results based on the semantic dimension evaluation results and the quality dimension evaluation results.

[0073] The aforementioned implementing entity will combine the semantic dimension evaluation results and the quality dimension evaluation results to generate the final evaluation result.

[0074] It should be noted that in some optional implementations, the multi-dimensional evaluation criteria can also include: popularity evaluation and novelty evaluation. Popularity evaluation refers to assessing the potential spread of candidate topics, which can be based on the spread model of historically similar topics. Novelty evaluation refers to assessing the distinguishability of candidate topics from topics in the existing topic library, thereby determining the novelty of the candidate topics. Novelty can be assessed through vector distance and semantic repetition. Finally, the aforementioned execution entity will integrate the semantic dimension evaluation results, quality dimension evaluation results, popularity evaluation results, and novelty evaluation results to generate the final evaluation result.

[0075] In some optional implementations of this embodiment, step 503 includes: Step 5031: Generate a comprehensive score for the candidate topics based on the semantic dimension evaluation results and the quality dimension evaluation results.

[0076] Since both the semantic dimension evaluation result and the quality dimension evaluation result are individual scores, the aforementioned implementing entity will generate a comprehensive score for the candidate topics based on these two results. For example, the implementing entity can take the average of the semantic dimension evaluation results and the quality dimension evaluation results to generate a comprehensive score; or the implementing entity can perform a weighted calculation of the semantic dimension evaluation results and the quality dimension evaluation results to obtain a comprehensive score.

[0077] Step 5032: The extreme value normalization method is used to calibrate the comprehensive score of all candidate topics, and the calibrated comprehensive score is mapped to a preset interval to obtain the mapped score.

[0078] The aforementioned execution entity will use the extreme value normalization method to calibrate the comprehensive score of all candidate topics, thereby normalizing the score to the [0,1] interval, and then mapping the calibrated score to a preset interval, which is generally [0,100], thus obtaining the mapped score.

[0079] Step 5033: Determine the evaluation result based on the mapped score and the preset score threshold.

[0080] The aforementioned execution entity determines the evaluation result based on the relationship between the mapped score and a preset score threshold. This preset threshold can be set according to actual circumstances. The execution entity judges the magnitude of the mapping score and the preset score threshold to determine the evaluation result, and then identifies positive and negative topics from the candidate topics based on the evaluation result. This dual-dimensional (semantic and quality) fusion generates a comprehensive score, making the evaluation more aligned with the topic generation needs, thereby improving the scientific rigor and relevance of the evaluation.

[0081] By selectively assessing the relevance of candidate topics to signal data in a semantic dimension and independently evaluating the quality compliance of candidate topics themselves in a quality dimension, and then integrating the results of the two dimensions to generate the final evaluation result, a two-dimensional quantitative evaluation system that fits the core needs of topic generation has been constructed, which improves the accuracy, objectivity and relevance of the evaluation results of candidate topics to actual business.

[0082] Continue to refer to Figure 6 , Figure 6 A flowchart 600 of a fourth embodiment of a training method for an information generation model according to the present disclosure is shown. The training method for the information generation model includes the following steps: Step 601: Preprocess the multi-source signal data to obtain a structured signal vector.

[0083] Step 602: Input the structured signal vector into the large language model and output at least one candidate topic.

[0084] Step 603: Evaluate the candidate topics according to the preset multi-dimensional evaluation criteria to obtain the evaluation results.

[0085] Steps 601-603 are basically the same as steps 401-403 in the aforementioned embodiments. For specific implementation methods, please refer to the aforementioned description of steps 401-403, which will not be repeated here.

[0086] Step 604: In response to determining that the evaluation result is that the mapped score is greater than the first preset score threshold, the candidate topic is determined as a positive example topic.

[0087] In this embodiment, the execution entity of the training method for the information generation model (e.g.) Figure 1 If the server 105 (as shown) determines that the mapped score is greater than the first preset score threshold, the candidate topic is identified as a positive topic. The first preset score threshold is the threshold for judging high-quality topics, which can be set according to actual conditions, for example, to 70 points. That is, if the mapped score is greater than the first preset score threshold, the candidate topic is identified as a positive topic.

[0088] Step 605: In response to determining that the evaluation result is that the mapped score is less than the second preset score threshold, the candidate topic is determined as a negative example topic.

[0089] In this embodiment, if the execution entity determines that the mapped score is less than a second preset score threshold, it determines the candidate topic as a negative example topic. Here, the second preset score threshold is the judgment threshold for inferior topics, which can be set according to actual conditions, for example, to 60 points. The second preset score threshold is less than the first preset score threshold. That is, if the mapped score is less than the second preset score threshold, the candidate topic is determined as a negative example topic. Thus, positive and negative examples are determined based on the relationship between the score and the preset score threshold, improving the efficiency of determining positive and negative examples.

[0090] Step 606: Determine the positive example data corresponding to the positive example topic from the multi-source signal data.

[0091] In this embodiment, the aforementioned execution entity determines the positive example signal data for generating the positive example topic from the multi-source signal data.

[0092] Step 607: Correct the negative example topic based on preset rules to obtain the corrected topic.

[0093] In this embodiment, the execution entity determines the dimension to be corrected for the negative example topic based on its score; for example, the dimension to be corrected is the semantic dimension. Then, the semantic dimension to be corrected is corrected to obtain the corrected topic.

[0094] In some optional implementations of this embodiment, step 607 includes: determining the dimension to be corrected corresponding to the negative example topic according to the multi-dimensional evaluation criteria; correcting the negative example topic according to the correction rules corresponding to the dimension to be corrected, and obtaining the corrected topic.

[0095] The aforementioned execution entity determines the dimension to be corrected for the negative example topic based on its score; for example, the dimension to be corrected might be semantic. Then, it corrects the negative example topic according to pre-defined repair rules corresponding to the dimension to be corrected, resulting in a corrected topic. Since the repair rules correspond to the defect dimensions, correcting the negative example topic according to the repair rules avoids correction biases caused by general rules, ensuring the accuracy of the corrected topic.

[0096] Step 608: Use positive example topics and positive example data as positive training sample pairs, and use negative example topics and corrected topics as negative training sample pairs.

[0097] In this embodiment, the execution entity pairs positive example topics with positive example data to obtain positive training sample pairs. It then pairs negative example topics with corrected topics to obtain negative training sample pairs, which are also known as corrected sample pairs.

[0098] Step 609: Train the large language model using positive training sample pairs and negative training sample pairs to obtain the information generation model.

[0099] Step 609 is basically the same as step 204 in the aforementioned embodiment. For the specific implementation method, please refer to the aforementioned description of step 204, which will not be repeated here.

[0100] from Figure 6 It can be seen from this that, with Figure 4 Compared to the corresponding embodiments, the training method of the information generation model in this embodiment emphasizes the steps of generating positive training sample pairs and negative training sample pairs. By accurately matching positive example topics with corresponding multi-source signal positive example data and regularizing and correcting negative example topics, the corrected topics are obtained. Finally, positive training sample pairs (positive example data - positive example topic) and negative training sample pairs (negative example topic - corrected topic) are constructed, providing a high-quality, highly targeted and standardized training sample foundation for model training.

[0101] Continue to refer to Figure 7 , Figure 7 A flow 700 of an embodiment of an information generation method according to the present disclosure is shown. The information generation method includes the following steps: Step 701: Obtain the target data and target theme style.

[0102] In this embodiment, the execution subject of the information generation method (e.g.) Figure 1 The server 105 shown will acquire target data, which is the signal data for generating topic information. The aforementioned execution entity will also acquire the target topic style, which is the style information of the generated topic information specified by the user. The target topic style can be news style, social style, etc., but this embodiment does not specifically limit it.

[0103] Step 702: Input the target data and target theme style into the information generation model and output at least one target theme.

[0104] In this embodiment, the executing entity inputs the target data and target theme style into the information generation model, thereby outputting at least one target theme. The information generation model is trained using the method described in the preceding embodiments. That is, the information generation model generates multiple theme information items that conform to the target theme style based on the target data and target theme style.

[0105] The information generation method provided in this disclosure first acquires target data and a target theme style; then, it inputs the target data and the target theme style into an information generation model and outputs at least one target theme. Since the information generation model has the ability to perceive multi-source signal data and generate themes of various styles, using the information generation model to generate themes that conform to the target theme style can improve the efficiency of theme generation and its fit with the preset style, thereby improving the user experience and enhancing the targeting of the recommendation system.

[0106] Continue to refer to Figure 8 , Figure 8 A flow 800 of another embodiment of the information generation method according to the present disclosure is shown. The information generation method includes the following steps: Step 801: Obtain the target data and target theme style.

[0107] Step 801 is basically the same as step 701 in the aforementioned embodiment. For the specific implementation method, please refer to the aforementioned description of step 701, which will not be repeated here.

[0108] Step 802: Evaluate the target topic according to the preset multi-dimensional evaluation criteria and generate the evaluation score corresponding to the target topic.

[0109] In this embodiment, the execution subject of the information generation method (e.g.) Figure 1 The server 105 shown will evaluate the target topic according to preset multi-dimensional evaluation criteria and generate an evaluation score corresponding to the target topic. The evaluation dimensions here can be, but are not limited to: relevance evaluation, popularity evaluation, novelty evaluation, quality evaluation, etc. Relevance evaluation refers to evaluating the relevance between the target topic and the style of the target topic. Here, a semantic relevance model can be used to evaluate the relevance. Popularity evaluation refers to evaluating the potential spread popularity of the target topic. Here, popularity can be evaluated based on the spread model of historical similar topics. Novelty evaluation refers to evaluating the distinguishability of the target topic from the topics in the existing topic library, thereby determining the novelty of the target topic. Here, novelty can be evaluated through vector distance and semantic repetition. Quality evaluation refers to evaluating the grammatical correctness, fluency, information content, and security of the target topic. Here, a series of classifiers or rules can be used to make judgments.

[0110] In some optional implementations of this embodiment, step 802 includes: determining the weight value corresponding to each evaluation dimension in the multi-dimensional evaluation standard; and performing a weighted calculation based on the scores of each evaluation dimension and the weight value corresponding to each evaluation dimension to obtain the evaluation score.

[0111] In this implementation, the aforementioned executing entity determines the weight value corresponding to each evaluation dimension in the multi-dimensional evaluation standard, and then performs a weighted calculation based on the scores of each evaluation dimension and the weight value of each evaluation dimension to obtain the evaluation score.

[0112] For example, the multiple evaluation dimensions include relevance evaluation, popularity evaluation, novelty evaluation, and quality evaluation. The weight of relevance evaluation is the first weight, popularity evaluation is the second weight, novelty evaluation is the third weight, and quality evaluation is the fourth weight. Then, the above-mentioned implementing entity can perform weighted calculation based on the relevance score and the first weight, the popularity score and the second weight, the novelty evaluation and the third weight, and the quality score and the fourth weight to obtain the final evaluation score.

[0113] By assigning personalized weight values ​​to each evaluation dimension, the evaluation system is made to align with core business needs, thereby enhancing the relevance and scientific rigor of the evaluation.

[0114] Step 803: Input the target data and target theme style into the information generation model, and output at least one target theme and the corresponding evaluation score of the target theme.

[0115] In this embodiment, the aforementioned execution entity inputs the target data and target theme style into the information generation model, thereby outputting at least one target theme and the corresponding evaluation score of the target theme. That is, the information generation model not only outputs the generated theme, but also outputs the comprehensive evaluation score of each theme, and can also output the evaluation score of each dimension.

[0116] from Figure 8 It can be seen from this that, with Figure 7 Compared with the corresponding embodiments, the information generation method in this embodiment emphasizes the step of generating the evaluation score corresponding to the target topic. In this way, the evaluation score of the target topic is generated quantitatively through preset multi-dimensional evaluation standards, and the information generation model directly outputs the target topic and the corresponding evaluation score, which improves the efficiency of the entire chain of content generation and evaluation, and achieves accurate control of the quality of the target topic and a high degree of fit with business needs.

[0117] Further reference Figure 9 As an implementation of the methods shown in the above figures, this disclosure provides an embodiment of a training device for an information generation model, which is similar to... Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0118] like Figure 9As shown, the training device 900 for the information generation model in this embodiment includes: a topic generation module 901, a topic determination module 902, a training sample generation module 903, and a training module 904. The topic generation module 901 is configured to input collected multi-source signal data into a large language model to obtain at least one candidate topic; the topic determination module 902 is configured to determine positive and negative example topics from the at least one candidate topic; the training sample generation module 903 is configured to generate positive training sample pairs and negative training sample pairs based on the multi-source signal data, positive example topics, and negative example topics; and the training module 904 is configured to train the large language model using the positive and negative training sample pairs to obtain the information generation model.

[0119] In this embodiment, the specific processing of the topic generation module 901, topic determination module 902, training sample generation module 903, and training module 904 in the training device 900 of the information generation model, and the resulting technical effects, can be referred to respectively. Figure 2 The relevant descriptions of steps 201-204 in the corresponding embodiments will not be repeated here.

[0120] In some optional implementations of this embodiment, the topic generation module 901 includes: a preprocessing submodule configured to preprocess multi-source signal data to obtain a structured signal vector; and an output submodule configured to input the structured signal vector into a large language model and output at least one candidate topic.

[0121] In some optional implementations of this embodiment, the preprocessing submodule includes: an extraction unit configured to extract multi-dimensional features of multi-source signal data using natural language processing techniques and feature extraction methods; and a fusion unit configured to fuse the multi-dimensional features according to the data source of the multi-source signal data to obtain a structured signal vector.

[0122] In some optional implementations of this embodiment, the fusion unit is further configured to: fuse multi-dimensional features corresponding to signal data from the same data source to obtain a single-source signal vector; perform feature fusion on all single-source signal vectors and standardize the fused vector to obtain a structured signal vector.

[0123] In some optional implementations of this embodiment, the training device 900 for the information generation model further includes: an evaluation module configured to evaluate candidate topics according to preset multi-dimensional evaluation criteria to obtain evaluation results; and a topic determination module 902 including: a topic determination submodule configured to determine positive example topics and negative example topics from at least one candidate topic based on the evaluation results.

[0124] In some optional implementations of this embodiment, the multi-dimensional evaluation criteria include: a semantic dimension evaluation criterion and a quality dimension evaluation criterion; and the evaluation module includes: a semantic evaluation submodule, configured to evaluate the semantic correlation between candidate topics and the signal data that generates candidate topics according to the semantic dimension evaluation criterion, to obtain a semantic dimension evaluation result; a quality evaluation submodule, configured to evaluate the quality of candidate topics according to the quality dimension evaluation criterion, to obtain a quality dimension evaluation result; and a comprehensive evaluation submodule, configured to generate an evaluation result based on the semantic dimension evaluation result and the quality dimension evaluation result.

[0125] In some optional implementations of this embodiment, the comprehensive evaluation submodule is further configured to: generate a comprehensive score for candidate topics based on the semantic dimension evaluation results and the quality dimension evaluation results; calibrate the comprehensive scores of all candidate topics using the extreme value normalization method, and map the calibrated comprehensive scores to a preset interval to obtain the mapped scores; and determine the evaluation results based on the mapped scores and a preset score threshold.

[0126] In some optional implementations of this embodiment, the topic determination submodule is further configured to: determine a candidate topic as a positive topic in response to determining that the evaluation result is a score after mapping that is greater than a first preset score threshold; and determine a candidate topic as a negative topic in response to determining that the evaluation result is a score after mapping that is less than a second preset score threshold, wherein the first preset score threshold is greater than the second preset score threshold.

[0127] In some optional implementations of this embodiment, the training sample generation module 903 includes: a positive example determination submodule, configured to determine positive example data corresponding to a positive example topic from multi-source signal data; a correction submodule, configured to correct a negative example topic based on a preset rule to obtain a corrected topic; and a sample pair determination submodule, configured to use a positive example topic and positive example data as a positive training sample pair, and a negative example topic and the corrected topic as a negative training sample pair.

[0128] In some optional implementations of this embodiment, the correction submodule is further configured to: determine the dimension to be corrected corresponding to the negative example topic according to the multi-dimensional evaluation criteria; and correct the negative example topic according to the correction rules corresponding to the dimension to be corrected to obtain the corrected topic.

[0129] Further reference Figure 10 As an implementation of the methods shown in the above figures, this disclosure provides an embodiment of an information generation apparatus, which is similar to... Figure 7 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0130] like Figure 10As shown, the information generation device 1000 of this embodiment includes an acquisition module 1001 and an output module 1002. The acquisition module 1001 is configured to acquire target data and a target theme style; the output module 1002 is configured to input the target data and the target theme style into an information generation model and output at least one target theme. The information generation model is trained using the method described in the foregoing embodiments.

[0131] In this embodiment, the specific processing of the acquisition module 1001 and the output module 1002 in the information generation device 1000 and the resulting technical effects can be referred to respectively. Figure 7 The relevant descriptions of steps 701-702 in the corresponding embodiments will not be repeated here.

[0132] In some optional implementations of this embodiment, the information generation device 1000 further includes: an evaluation score generation module configured to evaluate the target topic according to a preset multi-dimensional evaluation standard and generate an evaluation score corresponding to the target topic; and an output module 1002 further configured to: input the target data and the target topic style into the information generation model and output at least one target topic and the evaluation score corresponding to the target topic.

[0133] In some optional implementations of this embodiment, the evaluation score generation module is further configured to: determine the weight value corresponding to each evaluation dimension in the multi-dimensional evaluation standard; and perform weighted calculation based on the scores of each evaluation dimension and the weight value corresponding to each evaluation dimension to obtain the evaluation score.

[0134] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0135] Figure 11 A schematic block diagram of an example electronic device 1100 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0136] Figure 11A schematic block diagram of an example electronic device 1100 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0137] like Figure 11 As shown, device 1100 includes a computing unit 1101, which can perform various appropriate actions and processes according to a computer program stored in read-only memory (ROM) 1102 or a computer program loaded into random access memory (RAM) 1103 from storage unit 1108. The RAM 1103 may also store various programs and data required for the operation of device 1100. The computing unit 1101, ROM 1102, and RAM 1103 are interconnected via bus 1104. Input / output (I / O) interface 1105 is also connected to bus 1104.

[0138] Multiple components in device 1100 are connected to I / O interface 1105, including: input unit 1106, such as keyboard, mouse, etc.; output unit 1107, such as various types of monitors, speakers, etc.; storage unit 1108, such as disk, optical disk, etc.; and communication unit 1109, such as network card, modem, wireless transceiver, etc. Communication unit 1109 allows device 1100 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0139] The computing unit 1101 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1101 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1101 performs the various methods and processes described above, such as training methods for information generation models or information generation methods. For example, in some embodiments, the training methods for information generation models or information generation methods can be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 1108. In some embodiments, part or all of the computer program can be loaded and / or installed on device 1100 via ROM 1102 and / or communication unit 1109. When the computer program is loaded into RAM 1103 and executed by the computing unit 1101, one or more steps of the training methods for information generation models or information generation methods described above can be performed. Alternatively, in other embodiments, the computing unit 1101 may be configured by any other suitable means (e.g., by means of firmware) to perform a training method or an information generation method for an information generation model.

[0140] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0141] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0142] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0143] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0144] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0145] Cloud computing refers to a technological system that enables access to elastic and scalable shared physical or virtual resources via a network. These resources can include servers, operating systems, networks, software, and storage devices, and can be deployed and managed in an on-demand, self-service manner. Cloud computing technology can provide efficient and powerful data processing capabilities for applications such as artificial intelligence and blockchain, as well as for model training.

[0146] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and Virtual Private Server (VPS) services, such as high management difficulty and weak business scalability.

[0147] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0148] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A training method for an information generation model, comprising: The collected multi-source signal data is input into a large language model to obtain at least one candidate topic; Determine positive and negative examples from the at least one candidate topic; Based on the multi-source signal data, the positive example topics, and the negative example topics, positive training sample pairs and negative training sample pairs are generated. The large language model is trained using the positive training sample pairs and the negative training sample pairs to obtain an information generation model.

2. The method according to claim 1, wherein, The process involves inputting the collected multi-source signal data into a large language model to obtain at least one candidate topic, including: The multi-source signal data is preprocessed to obtain a structured signal vector; The structured signal vector is input into the large language model, and at least one candidate topic is output.

3. The method according to claim 2, wherein, The preprocessing of the multi-source signal data to obtain a structured signal vector includes: Multi-dimensional features of the multi-source signal data are extracted using natural language processing techniques and feature extraction methods; The multi-dimensional features are fused based on the data sources of the multi-source signal data to obtain the structured signal vector.

4. The method according to claim 3, wherein, The step of fusing the multi-dimensional features based on the data sources of the multi-source signal data to obtain the structured signal vector includes: Multi-dimensional features corresponding to signal data from the same data source are fused to obtain a single-source signal vector; Feature fusion is performed on all single-source signal vectors, and the fused vectors are then standardized to obtain the structured signal vector.

5. The method according to claim 1, further comprising: The candidate topics are evaluated according to preset multi-dimensional evaluation criteria to obtain evaluation results; as well as The step of determining positive and negative topics from the at least one candidate topic includes: Based on the evaluation results, positive and negative examples are determined from the at least one candidate topics.

6. The method according to claim 5, wherein, The multi-dimensional evaluation criteria include: semantic dimension evaluation criteria and quality dimension evaluation criteria; and The evaluation of the candidate topics according to the preset multi-dimensional evaluation criteria to obtain the evaluation results includes: The semantic relevance between the candidate topics and the signal data that generated the candidate topics is evaluated according to the semantic dimension evaluation criteria to obtain the semantic dimension evaluation results. The quality of the candidate topics is evaluated according to the quality dimension evaluation criteria to obtain the quality dimension evaluation results; The evaluation results are generated based on the semantic dimension evaluation results and the quality dimension evaluation results.

7. The method according to claim 6, wherein, The step of generating the evaluation result based on the semantic dimension evaluation result and the quality dimension evaluation result includes: Based on the semantic dimension evaluation results and the quality dimension evaluation results, a comprehensive score is generated for the candidate topics; The extreme value normalization method is used to calibrate the comprehensive score of all candidate topics, and the calibrated comprehensive score is mapped to a preset interval to obtain the mapped score; The evaluation result is determined based on the mapped score and the preset score threshold.

8. The method according to claim 7, wherein, The step of determining positive and negative examples from the at least one candidate topics based on the evaluation results includes: In response to determining that the evaluation result is that the score after mapping is greater than a first preset score threshold, the candidate topic is determined as the positive example topic; In response to determining that the evaluation result is that the score after mapping is less than a second preset score threshold, the candidate topic is determined as the negative example topic, wherein the first preset score threshold is greater than the second preset score threshold.

9. The method according to claim 5, wherein, The step of generating positive training sample pairs and negative training sample pairs based on the multi-source signal data, the positive example topics, and the negative example topics includes: Determine the positive example data corresponding to the positive example topic from the multi-source signal data; The negative example topic is corrected based on preset rules to obtain the corrected topic; The positive example topics and the positive example data are used as the positive training sample pairs, and the negative example topics and the corrected topics are used as the negative training sample pairs.

10. The method according to claim 9, wherein, The step of correcting the negative example topic based on preset rules to obtain the corrected topic includes: Based on the multi-dimensional evaluation criteria, determine the dimensions to be corrected corresponding to the negative example topics; The negative example topic is corrected according to the correction rule corresponding to the dimension to be corrected, and the corrected topic is obtained.

11. An information generation method, comprising: Obtain the target data and target theme style; The target data and the target topic style are input into the information generation model, and at least one target topic is output. The information generation model is trained using the method described in any one of claims 1-10.

12. The method of claim 11, further comprising: The target topic is evaluated according to a preset multi-dimensional evaluation standard, and an evaluation score corresponding to the target topic is generated. as well as The step of inputting the target data and the target theme style into the information generation model and outputting at least one target theme includes: The target data and the target theme style are input into the information generation model, and at least one target theme and the corresponding evaluation score are output.

13. The method according to claim 12, wherein, The step of evaluating the target topic according to a preset multi-dimensional evaluation standard and generating an evaluation score corresponding to the target topic includes: Determine the weight value corresponding to each evaluation dimension in the multi-dimensional evaluation criteria; The evaluation score is obtained by weighting the scores of each evaluation dimension and the corresponding weight values ​​of each evaluation dimension.

14. A training device for an information generation model, comprising: The topic generation module is configured to input the collected multi-source signal data into the large language model to obtain at least one candidate topic; The topic determination module is configured to determine positive and negative topics from the at least one candidate topic; The training sample generation module is configured to generate positive training sample pairs and negative training sample pairs based on the multi-source signal data, the positive example topics, and the negative example topics; The training module is configured to train the large language model using the positive training sample pairs and the negative training sample pairs to obtain an information generation model.

15. An information generation apparatus, comprising: The acquisition module is configured to acquire the target data and the target theme style. The output module is configured to input the target data and the target theme style into the information generation model and output at least one target theme, wherein the information generation model is trained using the method described in any one of claims 1-10.

16. An electronic device comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-10 or 11-13.

17. A non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the method of any one of claims 1-10 or 11-13.

18. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-10 or 11-13.

Images