Topic recommendation method and device, equipment, storage medium and vehicle

Abstract

The invention relates to a topic recommendation method and device, equipment, a storage medium and a vehicle, and the method comprises the steps: obtaining historical dialogue information which comprises historical question information and historical answer information; multiple paths of input information are extracted from the historical dialogue information, the multiple paths of input information comprise at least two items of historical problem information, historical dialogue information and keywords, each item serves as one path of input information, and the keywords are extracted from the historical dialogue information; and taking the multi-path input information as the input of the trained large model, calling the trained large model to carry out multi-path recall, and outputting the target recommendation topic, compared with the prior art, the topic recommendation method provided by the invention depends on the emergence capability of the large model and does not need to depend on a large amount of marked data to carry out training, so that the topic recommendation efficiency is improved. The candidate topics with high quality can be generated under the condition of data scarcity, the diversity of topic recommendation is improved, and the application field of the topic recommendation method is expanded.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technologies, and in particular, to a topic recommendation method, apparatus, device, storage medium, and vehicle. Background Art

[0002] With the rapid development of the network, network information has increased sharply, and various topics will gather in the network within a certain period of time or range.

[0003] In the prior art, topic recommendation is performed through traditional intelligent recommendation algorithms. The defect of topic recommendation through traditional intelligent recommendation algorithms is that a large amount of labeled data is required for training, complex context information cannot be captured, the application of traditional intelligent recommendation algorithms in the case of scarce data is limited, and the quality of the recommended topics is low.

[0004] Therefore, there is an urgent need for a topic recommendation method to solve the above problems. Summary of the Invention

[0005] In order to solve the above technical problems or at least partially solve the above technical problems, the present disclosure provides a topic recommendation method, apparatus, device, storage medium, and vehicle to improve the quality of topic recommendation.

[0006] In a first aspect, an embodiment of the present disclosure provides a topic recommendation method, including:

[0007] Obtain historical conversation information, where the historical conversation information includes historical question information and historical answer information;

[0008] Extract multiplex input information from the historical conversation information, where the multiplex input information includes at least two of the historical question information, the historical conversation information, and keywords, and each item serves as one path of input information, where the keywords are extracted from the historical conversation information;

[0009] Use the multiplex input information as the input of a trained large model, and call the trained large model for multiplex recall to output a target recommended topic.

[0010] In some embodiments, using the multiplex input information as the input of a trained large model, and calling the trained large model for multiplex recall to output a target recommended topic includes:

[0011] Use the historical question information, the historical conversation information, and the keywords as each path of input of the trained large model, and respectively output multiple first candidate topics, multiple second candidate topics, and multiple third candidate topics through the emergence ability of the trained large model;

[0012] Deduplicate the multiple first candidate topics, the multiple second candidate topics, and the multiple third candidate topics to obtain multiple recommended topics;

[0013] Sort the multiple recommended topics, and obtain the target recommended topic through the sorting result of the multiple recommended topics.

[0014] In some embodiments, deduplicating the multiple first candidate topics, the multiple second candidate topics, and the multiple third candidate topics to obtain multiple recommended topics includes:

[0015] Perform internal deduplication on the multiple first candidate topics, the multiple second candidate topics, and the multiple third candidate topics to obtain multiple candidate topics;

[0016] Compare the multiple candidate topics with the historical conversation information, and perform external deduplication on the multiple candidate topics to obtain multiple recommended topics.

[0017] In some embodiments, performing internal deduplication on the multiple first candidate topics, the multiple second candidate topics, and the multiple third candidate topics to obtain multiple candidate topics includes:

[0018] Traverse the multiple first candidate topics, the multiple second candidate topics, and the multiple third candidate topics, calculate the duplicate similarity between any two candidate topics, and eliminate the candidate topics with a duplicate similarity greater than the first similarity threshold to obtain multiple candidate topics.

[0019] In some embodiments, comparing the multiple candidate topics with the historical conversation information, and performing external deduplication on the multiple candidate topics to obtain multiple recommended topics includes:

[0020] Calculate the semantic similarity between the multiple candidate topics and the historical conversation information, and eliminate the candidate topics with a semantic similarity greater than the second similarity threshold to obtain multiple recommended topics.

[0021] In some embodiments, sorting the multiple recommended topics, and obtaining the target recommended topic through the sorting result of the multiple recommended topics includes:

[0022] Calculate the priorities of the multiple recommended topics;

[0023] Sort the multiple recommended topics according to the magnitudes of the priorities to obtain the first sorting result of the multiple recommended topics, and obtain the target recommended topic through the first sorting result.

[0024] In some embodiments, calculating the priorities of the multiple recommended topics includes:

[0025] Obtain a preset annotation information table, where the preset annotation information table includes annotation information, the field of the annotation information, and the weight of the field;

[0026] For each of the multiple recommended topics, calculate the priority of the recommended topic according to the preset annotation information table with weights.

[0027] In some embodiments, sorting the multiple recommended topics, and obtaining a target recommended topic through the sorting result of the multiple recommended topics, includes:

[0028] Sort the multiple recommended topics according to the relevance between the multiple recommended topics and the historical conversation information and the degree of interest of the user, obtain a second sorting result of the multiple recommended topics, and obtain a target recommended topic through the second sorting result.

[0029] In some embodiments, sorting the multiple recommended topics according to the relevance between the multiple recommended topics and the historical conversation information, the association between the multiple recommended topics and the historical conversation information, and the degree of interest of the user, includes:

[0030] Extract a first entity list from the historical conversation information;

[0031] For each of the recommended topics, extract a second entity list from the recommended topic;

[0032] Calculate the intersection entities of the first entity list and the second entity list;

[0033] Annotate any of the intersection entities according to the degree of user interest to obtain weight annotation information of the intersection entities;

[0034] Calculate the scores of the multiple recommended topics by weighting with the weight annotation information of the intersection entities;

[0035] Sort the multiple recommended topics according to the magnitudes of the scores.

[0036] In a second aspect, an embodiment of the present disclosure provides a topic recommendation device, including:

[0037] An acquisition module, configured to acquire historical conversation information, where the historical conversation information includes historical question information and historical answer information;

[0038] An extraction module, configured to extract multi-channel input information from the historical conversation information, where the multi-channel input information includes at least two of the historical question information, the historical conversation information, and keywords, and each item is used as one channel of input information, where the keywords are extracted from the historical conversation information;

[0039] A calling module, configured to use the multiplexed input information as the input of a trained large model, call the trained large model for multiplexed recall, and output a target recommended topic.

[0040] In a third aspect, an embodiment of the present disclosure provides an electronic device, including:

[0041] A memory;

[0042] A processor; and

[0043] A computer program;

[0044] Wherein, the computer program is stored in the memory and is configured to be executed by the processor to implement the method as described in the first aspect.

[0045] In a fourth aspect, an embodiment of the present disclosure provides a computer-readable storage medium, on which a computer program is stored, and the computer program is executed by a processor to implement the method as described in the first aspect.

[0046] In a fifth aspect, an embodiment of the present disclosure further provides a vehicle, including: the topic recommendation device as described in the second aspect; or, the electronic device as described in the third aspect; or, the computer-readable storage medium as described in the fourth aspect.

[0047] The topic recommendation method, device, equipment, storage medium and vehicle provided by the embodiments of the present disclosure obtain historical conversation information, which includes historical question information and historical answer information; extract multiplexed input information from the historical conversation information, and the multiplexed input information includes at least two of historical question information, historical conversation information, and keywords, and each item is used as one-way input information, where the keywords are extracted from the historical conversation information; use the multiplexed input information as the input of a trained large model, call the trained large model for multiplexed recall, and output a target recommended topic. Compared with the prior art, the topic recommendation method provided by the embodiments of the present disclosure extracts multiplexed input information from the historical conversation information, relies on the emerging ability of the large model for multiplexed recall, does not need to rely on a large amount of labeled data for training, can generate candidate topics with high quality in the case of scarce data, improves the diversity of topic recommendation, and expands the application field of the topic recommendation method. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The drawings here are incorporated into the specification and constitute a part of this specification, showing embodiments consistent with the present disclosure, and are used together with the specification to explain the principles of the present disclosure.

[0049] To more clearly illustrate the technical solutions in the embodiments of the present disclosure or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, for those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0050] Figure 1 Flowchart of the topic recommendation method provided by the embodiments of the present disclosure;

[0051] Figure 2 Schematic diagram of an application scenario provided by the embodiments of the present disclosure;

[0052] Figure 3 Flowchart of the topic recommendation method provided by the embodiments of the present disclosure;

[0053] Figure 4 Schematic diagram of an application scenario provided by the embodiments of the present disclosure;

[0054] Figure 5 Schematic diagram of the structure of the topic recommendation device provided by the embodiments of the present disclosure;

[0055] Figure 6 Schematic diagram of the structure of the electronic device provided by the embodiments of the present disclosure. Specific implementation manners

[0056] In order to more clearly understand the above objects, features and advantages of the present disclosure, the following will further describe the solutions of the present disclosure. It should be noted that, without conflict, the embodiments of the present disclosure and the features in the embodiments can be combined with each other.

[0057] Many specific details are set forth in the following description to facilitate a thorough understanding of the present disclosure, but the present disclosure can also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only a part of the embodiments of the present disclosure, rather than all the embodiments.

[0058] The embodiments of the present disclosure provide a topic recommendation method, and the following will introduce this method in combination with specific embodiments.

[0059] Figure 1 Flowchart of the topic recommendation method provided by the embodiments of the present disclosure. This method can be executed by a trained large model or a topic recommendation device. The topic recommendation device can be implemented in software and / or hardware, and can be configured in an electronic device, such as a server or a terminal. Specifically, the terminal includes a mobile phone, a computer, a tablet computer, etc. In addition, this method can be applied to Figure 2The application scenario shown includes multiple input information 20, a large model 21 that has completed training, and a target recommended topic 22. It can be understood that the topic recommendation method provided by the embodiments of the present disclosure can also be applied in other scenarios.

[0060] The following combines Figure 2 the application scenario shown to introduce Figure 1 the topic recommendation method shown. The specific steps included in this method are as follows:

[0061] S101. Obtain historical conversation information, where the historical conversation information includes historical question information and historical answer information.

[0062] The topic recommendation device obtains historical conversation information. This historical conversation information can be historical conversation information obtained through user communication or historical conversation information obtained through human-machine interaction. Among them, human-machine interaction can specifically be a question-and-answer between a person and a voice assistant, or historical conversation information obtained through machine-to-machine interaction. This embodiment does not make a limitation. Subsequently, the historical conversation information obtained through human-machine interaction, that is, the voice assistant's reply to the user, will be used as an example for elaboration. Among them, the user is represented as user, and the voice assistant is represented as assistant.

[0063] Among them, the historical conversation information includes historical question information and historical answer information. Specifically, the historical question information is not limited to historical question information and can also be historical content information; the historical answer information is not limited to historical answer information and can also be historical reply information. It can be understood that the historical answer information is the information obtained by replying to the historical question information, and the historical question information and the historical answer information are relevant.

[0064] Exemplarily, the historical question information can be: "What is Neptune?"

[0065] Exemplarily, the historical answer information can be: "Neptune is the planet farthest from the sun among the eight major planets in the solar system. Its mass is 17 times that of the Earth, but its volume only ranks fourth. The atmosphere of Neptune is mainly composed of hydrogen and helium, and also contains trace amounts of methane, which is one of the reasons for the planet to appear blue. Its name comes from Neptune, the god of the sea in Roman mythology."

[0066] Exemplarily, the historical conversation information includes historical question information and historical answer information. Therefore, the historical conversation information can be "What is Neptune?" and "Neptune is the planet farthest from the sun among the eight major planets in the solar system. Its mass is 17 times that of the Earth, but its volume only ranks fourth. The atmosphere of Neptune is mainly composed of hydrogen and helium, and also contains trace amounts of methane, which is one of the reasons for the planet to appear blue. Its name comes from Neptune, the god of the sea in Roman mythology."

[0067] S102. Extract multiple input information from the historical conversation information. The multiple input information includes at least two of the historical question information, the historical conversation information, and keywords. Each item serves as one path of input information, where the keywords are extracted from the historical conversation information.

[0068] After obtaining the historical conversation information, the topic recommendation device extracts keywords from the historical conversation information, and then extracts multiple input information 20 from the historical conversation information. The multiple input information includes at least two of the historical question information, the historical conversation information, and keywords. Each item can serve as one path of input information. Exemplarily, the multiple input information can include any two of the historical question information, the historical conversation information, and keywords. For example, the multiple input information includes the historical question information and the historical conversation information, or the historical conversation information and keywords, or the historical question information and keywords. The multiple input information can also include the historical question information, the historical conversation information, and keywords.

[0069] It can be understood that in some other embodiments, the multiple input information can also include at least two of the historical question information, the historical answer information, the historical conversation information, and keywords. Each item can serve as one path of input information. Since the historical conversation information includes the historical question information and the historical answer information, the historical question information and the historical answer information are only counted as one item.

[0070] The keywords are extracted from the historical conversation information, that is, the keywords are extracted from the historical question information and / or the historical answer information. Exemplarily, the keywords can be "Neptune", "Solar System", "planet", etc.

[0071] S103. Use the multiple input information as the input of the trained large model, and call the trained large model for multiple recalls to output the target recommended topic.

[0072] A large model refers to a neural network model with ultra-large-scale parameters and a machine learning model with extremely powerful computing resources. It can process massive amounts of data and complete various complex tasks, such as natural language processing, image recognition, etc. The large model contains billions of parameters, and the size of the large model can reach hundreds of GB or even larger, which enables the large model to have strong expressive and learning capabilities. The large model will learn multiple different NLP tasks together, such as machine translation, text summarization, question answering systems, etc., which can enable the large model to learn a more extensive and generalized language understanding ability.

[0073] The topic recommendation device uses the multiple input information 20 as the input of the trained large model 21, and calls the trained large model 21 for multiple recalls to output the target recommended topic 22.

[0074] Specifically, the topic recommendation device inputs the multiplexed input information 20 into the trained large model 21, and calls the trained large model 21 for multiplexed recall, that is, calls the trained large model 21 to construct a multi-task prompt and outputs the target recommended topic 22.

[0075] In the embodiment of the present disclosure, historical dialogue information is obtained, and the historical dialogue information includes historical question information and historical answer information; multiplexed input information is extracted from the historical dialogue information, and the multiplexed input information includes at least two of historical question information, historical dialogue information, and keywords, and each item is used as one-way input information, where the keywords are extracted from the historical dialogue information; the multiplexed input information is used as the input of the trained large model, and the trained large model is called for multiplexed recall to output the target recommended topic. Compared with the prior art, the topic recommendation method provided in the embodiment of the present disclosure extracts multiplexed input information from the historical dialogue information, relies on the emergent ability of the large model for multiplexed recall, does not need to rely on a large amount of labeled data for training, can generate high-quality candidate topics in the case of scarce data, improves the diversity of topic recommendation, and expands the application field of the topic recommendation method.

[0076] Specifically, S103 can be implemented by the method as Figure 3 shown. As Figure 3 shown, the specific steps included in S103 are as follows:

[0077] S301: Respectively use the historical question information, the historical dialogue information, and the keywords as each-way input of the trained large model, and respectively output multiple first candidate topics, multiple second candidate topics, and multiple third candidate topics through the emergent ability of the trained large model.

[0078] The topic recommendation device respectively uses the historical question information, the historical dialogue information, and the keywords as each-way input of the trained large model, and respectively outputs multiple first candidate topics, multiple second candidate topics, and multiple third candidate topics through the emergent ability of the trained large model. Specifically, the topic recommendation device uses the historical question information as one-way input of the trained large model and outputs multiple first candidate topics through the emergent ability of the trained large model; the topic recommendation device uses the historical dialogue information as one-way input of the trained large model and outputs multiple second candidate topics through the emergent ability of the trained large model; the topic recommendation device uses the keywords as one-way input of the trained large model and outputs multiple third candidate topics through the emergent ability of the trained large model.

[0079] Exemplarily, as Figure 4As shown in the figure, the multi-channel input information 20 includes historical question information 401, historical conversation information 402, and keywords 403. The topic recommendation device takes the historical question information 401, historical conversation information 402, and keywords 403 as the three-channel inputs of the trained large model respectively, performs multi-channel recall, and outputs multiple first candidate topics 411, multiple second candidate topics 412, and multiple third candidate topics 413 respectively through the emergence ability of the trained large model.

[0080] Exemplarily, when the multi-channel input information is historical question information, historical conversation information, and keywords, the relationship between the input information, Prompt (hint) examples, and candidate topics is shown in Table 1:

[0081] Table 1

[0082]

[0083] As can be seen from Table 1, according to the above historical question information, the trained large model is called to construct a multi-task prompt "For the question: "{}", if you are an expert in this field, propose 3 questions that are the most creative and different from the original question, and the number of words in each question does not exceed 15". The generated first candidate topics include: "How far is Neptune from the Earth", "How did Neptune get its name", "What is the revolution period of Neptune", and the generation of this topic recommendation is related to the direct intention of the question.

[0084] According to the above historical conversation information, the trained large model is called to construct a multi-task prompt "user: {}, asisstant: {}. If you are an expert in this field, please propose 3 questions that are the most creative, different from what the user said, and not answered by what the assistant said based on the above conversation between the user and the assistant, and the number of words in each question does not exceed 15". The generated candidate topics include "Who discovered Neptune", "What other planets are named after Roman mythology", "How thick is the atmosphere of Neptune", and the generation of this topic recommendation is related to the direct intention of the question or the detailed content of the answer.

[0085] According to the above keywords, the trained large model is called to construct a multi-task prompt "If you are an extremely capable question expert, please first extract the most important keyword from the following question "{}", and then propose 3 questions that are the most creative and different from the original question based on this most important keyword, and the number of words in each question does not exceed 15". The generated candidate topics include "Which planet is Neptune in the solar system", "What is Neptune mainly composed of", "Who is the largest planet in the solar system", and the generation of this topic recommendation is related to the extended content of the keywords.

[0086] S302. Deduplicate the multiple first candidate topics, the multiple second candidate topics, and the multiple third candidate topics to obtain multiple recommended topics.

[0087] The topic recommendation device deduplicates the multiple first candidate topics 411, the multiple second candidate topics 412, and the multiple third candidate topics 413 to obtain multiple recommended topics.

[0088] Optionally, deduplicating the multiple first candidate topics, the multiple second candidate topics, and the multiple third candidate topics to obtain multiple recommended topics includes: internally deduplicating the multiple first candidate topics, the multiple second candidate topics, and the multiple third candidate topics to obtain multiple candidate topics; comparing the multiple candidate topics with the historical conversation information, and externally deduplicating the multiple candidate topics to obtain multiple recommended topics.

[0089] Specifically, the topic recommendation device internally deduplicates the multiple first candidate topics 411, the multiple second candidate topics 412, and the multiple third candidate topics 413 to obtain multiple candidate topics; compares the multiple candidate topics with the historical conversation information 402, and externally deduplicates the multiple candidate topics to obtain multiple recommended topics.

[0090] Optionally, internally deduplicating the multiple first candidate topics, the multiple second candidate topics, and the multiple third candidate topics to obtain multiple candidate topics includes: traversing the multiple first candidate topics, the multiple second candidate topics, and the multiple third candidate topics, calculating the duplicate similarity between any two candidate topics, and removing the candidate topics with a duplicate similarity greater than the first similarity threshold to obtain multiple candidate topics.

[0091] Specifically, the topic recommendation device traverses the multiple first candidate topics 411, the multiple second candidate topics 412, and the multiple third candidate topics 413, calculates the duplicate similarity between any two candidate topics, and removes the candidate topics with a duplicate similarity greater than the first similarity threshold to obtain multiple candidate topics. The first similarity threshold can be set according to the actual situation and is not limited in this embodiment. It can be understood that any two candidate topics can be any two candidate topics among the multiple first candidate topics 411, any two candidate topics among the multiple second candidate topics 412, or any two candidate topics among the multiple third candidate topics 413; or one candidate topic each from the multiple first candidate topics 411 and the multiple second candidate topics 412, one candidate topic each from the multiple first candidate topics 411 and the multiple third candidate topics 413, or one candidate topic each from the multiple second candidate topics 412 and the multiple third candidate topics 413.

[0092] Exemplarily, any two candidate topics can be "How far is Neptune from the Earth?" and "Who discovered Neptune?". Calculate the duplicate similarity of these two candidate topics. When the duplicate similarity of these two candidate topics is greater than the first similarity threshold, randomly eliminate any one of these two candidate topics. For example, eliminate the candidate topic "How far is Neptune from the Earth?" or eliminate the candidate topic "Who discovered Neptune?". Traverse multiple first candidate topics 411, multiple second candidate topics 412, and multiple third candidate topics 413 to obtain multiple candidate topics.

[0093] Optionally, compare the multiple candidate topics with the historical conversation information, and perform external deduplication on the multiple candidate topics to obtain multiple recommended topics, including: calculating the semantic similarity between the multiple candidate topics and the historical conversation information, and eliminating the candidate topics with a semantic similarity greater than the second similarity threshold to obtain multiple recommended topics.

[0094] Semantic similarity has many important application scenarios. For example, it is used for semantic recall in a retrieval system or as a feature for refined ranking. The semantic similarity model is used to extract multiple candidate topics with a semantic similarity greater than a preset threshold.

[0095] Specifically, the topic recommendation device calculates the semantic similarity between multiple candidate topics and historical conversation information, and eliminates the candidate topics with a semantic similarity greater than the second similarity threshold to obtain multiple recommended topics. This second similarity threshold can be set according to the actual situation. Specifically, it can be the same as the first similarity threshold or different from the first similarity threshold. This embodiment does not make a limitation.

[0096] Optionally, calculate the semantic similarity between the multiple candidate topics and the historical conversation information, and eliminate the candidate topics with a semantic similarity greater than the second similarity threshold to obtain multiple recommended topics, including: calculating the first semantic similarity between the multiple candidate topics and the historical question information, and eliminating the candidate topics with a first semantic similarity greater than the second similarity threshold to obtain multiple target candidate topics; calculating the second semantic similarity between the multiple target candidate topics and the historical answer information, and eliminating the target candidate topics with a second semantic similarity greater than the second similarity threshold to obtain multiple recommended topics.

[0097] Specifically, the topic recommendation device calculates the first semantic similarity between multiple candidate topics and historical question information 401, and eliminates the candidate topics with a first semantic similarity greater than the second similarity threshold to obtain multiple target candidate topics; calculates the second semantic similarity between multiple target candidate topics and historical answer information, and eliminates the target candidate topics with a second semantic similarity greater than the second similarity threshold to obtain multiple recommended topics.

[0098] Specifically, the topic deduplication is shown in Table 2:

[0099] Table 2

[0100]

[0101] S303. Sort the multiple recommended topics, and obtain a target recommended topic based on the sorting results of the multiple recommended topics.

[0102] The topic recommendation device sorts the multiple recommended topics, and obtains a target recommended topic based on the sorting results of the multiple recommended topics.

[0103] Optionally, sorting the multiple recommended topics and obtaining a target recommended topic based on the sorting results of the multiple recommended topics includes: calculating the priorities of the multiple recommended topics; sorting the multiple recommended topics according to the magnitudes of the priorities to obtain a first sorting result of the multiple recommended topics, and obtaining a target recommended topic through the first sorting result.

[0104] Specifically, the topic recommendation device calculates the priorities of the multiple recommended topics; sorts the multiple recommended topics according to the magnitudes of the priorities to obtain a first sorting result of the multiple recommended topics, and obtains a target recommended topic through the first sorting result. Generally, the first sorting result of the multiple recommended topics is sorted in descending order of priorities. Therefore, by obtaining a target recommended topic through the first sorting result, this target recommended topic is generally the first recommended topic in the first sorting result.

[0105] Optionally, calculating the priorities of the multiple recommended topics includes: obtaining a preset annotation information table, where the preset annotation information table includes annotation information, the field of the annotation information, and the weight of the field; for each of the multiple recommended topics, calculating the priority of the recommended topic by weighting according to the preset annotation information table.

[0106] Specifically, the topic recommendation device obtains a preset annotation information table, where the preset annotation information table includes annotation information, the field of the annotation information, and the weight of the field. Among them, the weight of the annotation information field can indicate the degree of user interest; for each of the multiple recommended topics, calculating the priority of the recommended topic by weighting according to the preset annotation information table.

[0107] Optionally, sorting the multiple recommended topics and obtaining a target recommended topic based on the sorting results of the multiple recommended topics includes: sorting the multiple recommended topics according to the relevance between the multiple recommended topics and the historical conversation information and the degree of user interest to obtain a second sorting result of the multiple recommended topics, and obtaining a target recommended topic through the second sorting result.

[0108] Specifically, the topic recommendation device sorts multiple recommended topics according to the relevance between the multiple recommended topics and the historical conversation information, as well as the degree of user interest, to obtain a second sorting result of the multiple recommended topics, and obtains a target recommended topic through the second sorting result. The second sorting result is sorted according to the context relevance and correlation between the multiple recommended topics and the historical conversation information, as well as the degree of user interest.

[0109] Optionally, sorting the multiple recommended topics according to the relevance between the multiple recommended topics and the historical conversation information, the relevance between the multiple recommended topics and the historical conversation information, and the degree of user interest includes: extracting a first entity list from the historical conversation information; for each recommended topic, extracting a second entity list from the recommended topic; calculating the intersection entities of the first entity list and the second entity list; annotating any of the intersection entities according to the user interest degree to obtain weight annotation information of the intersection entities; calculating the scores of the multiple recommended topics by weighting through the weight annotation information of the intersection entities; sorting the multiple recommended topics according to the magnitudes of the scores.

[0110] An entity refers to an extracted entity. In this embodiment, an entity refers to a named entity, that is, an entity that is extracted and has characteristics.

[0111] Specifically, the topic recommendation device extracts a first entity list from the historical conversation information; for each recommended topic, extracts a second entity list from the recommended topic; calculates the intersection entities of the first entity list and the second entity list; annotates any of the intersection entities according to the user interest degree to obtain weight annotation information of the intersection entities; calculates the scores of the multiple recommended topics by weighting through the weight annotation information of the intersection entities; sorts the multiple recommended topics according to the magnitudes of the scores.

[0112] Exemplarily, the initial weights of the entities are all 1, and the weights of the intersection entities are increased according to the degree of user interest (or it can also be understood as the importance of the entities). For example, if the intersection entity is Neptune and the weight gain of Neptune is 3, then the weighted scores of the following recommended topics are shown in Table 3. Sort the multiple recommended topics according to the magnitudes of the scores. For example, "How thick is the atmosphere of Neptune?" and "How far is Neptune from the Earth?" are tied for first place, and "Who discovered Neptune?" and "What is Neptune mainly composed of?" are tied for second place.

[0113] Table 3

[0114] Recommended Topics Score Who Discovered Neptune? Neptune 3 How Thick is Neptune's Atmosphere? Neptune 3 + Atmosphere 1 = 4 What is Neptune Mainly Composed of? Neptune 3 How Far is Neptune from Earth? Neptune 3 + Earth 1 = 4

[0115] The specific topic sorting is shown in Table 4:

[0116] Table 4

[0117]

[0118] In an embodiment of the present disclosure, historical conversation information is obtained, where the historical conversation information includes historical question information and historical answer information; multiplexed input information is extracted from the historical conversation information, and the multiplexed input information includes at least two of historical question information, historical conversation information, and keywords, with each item serving as one path of input information, where the keywords are extracted from the historical conversation information; the historical question information, historical conversation information, and keywords are respectively used as each path of input to a trained large model, and multiple first candidate topics, multiple second candidate topics, and multiple third candidate topics are respectively output through the emergence ability of the trained large model; the multiple first candidate topics, multiple second candidate topics, and multiple third candidate topics are traversed, the repeated similarity between any two candidate topics is calculated, and the candidate topics with a repeated similarity greater than a first similarity threshold are removed to obtain multiple candidate topics; the first semantic similarity between the multiple candidate topics and the historical question information is calculated, and the candidate topics with a first semantic similarity greater than a second similarity threshold are removed to obtain multiple target candidate topics; the second semantic similarity between the multiple target candidate topics and the historical answer information is calculated, and the target candidate topics with a second semantic similarity greater than the second similarity threshold are removed to obtain multiple recommended topics; a preset annotation information table is obtained, and the preset annotation information table includes annotation information, the field of the annotation information, and the weight of the field; for each of the multiple recommended topics, the priority of the recommended topic is calculated by weighting according to the preset annotation information table; the multiple recommended topics are sorted according to the magnitude of the priority to obtain a first sorting result of the multiple recommended topics, and a target recommended topic is obtained through the first sorting result. Compared with the prior art, the topic recommendation method provided by the present disclosure relies on the emergence ability of the large model, does not require a large amount of labeled data for training, can generate high-quality candidate topics in the case of scarce data, improves the diversity of topic recommendation, expands the application field of the topic recommendation method, is not restricted by fixed rules or historical behavior recommendations, and has a high degree of personalization of the recommendation results.

[0119] Figure 5 It is a schematic structural diagram of a topic recommendation device provided by an embodiment of the present disclosure. The topic recommendation device may be the terminal described in the above embodiment, or the topic recommendation device may be a component or assembly in the terminal. The topic recommendation device provided by an embodiment of the present disclosure may execute the processing flow provided by the embodiment of the topic recommendation method, as Figure 5 shown, the topic recommendation device 50 includes:

[0120] An acquisition module 51, configured to acquire historical conversation information, where the historical conversation information includes historical question information and historical answer information;

[0121] An extraction module 52, configured to extract multi-channel input information from the historical conversation information, where the multi-channel input information includes at least two of the historical question information, the historical conversation information, and keywords, and each item serves as one channel of input information, and the keywords are extracted from the historical conversation information;

[0122] An invocation module 53, configured to use the multi-channel input information as the input of the trained large model, invoke the trained large model for multi-channel recall, and output a target recommended topic.

[0123] Optionally, the invocation module 53 is further configured to use the historical question information, the historical conversation information, and the keywords as each channel of input of the trained large model respectively, and respectively output a plurality of first candidate topics, a plurality of second candidate topics, and a plurality of third candidate topics through the emergence ability of the trained large model; remove duplicates from the plurality of first candidate topics, the plurality of second candidate topics, and the plurality of third candidate topics to obtain a plurality of recommended topics; sort the plurality of recommended topics, and obtain the target recommended topic through the sorting result of the plurality of recommended topics.

[0124] Optionally, the invocation module 53 is further configured to perform internal deduplication on the plurality of first candidate topics, the plurality of second candidate topics, and the plurality of third candidate topics to obtain a plurality of candidate topics; compare the plurality of candidate topics with the historical conversation information, and perform external deduplication on the plurality of candidate topics to obtain a plurality of recommended topics.

[0125] Optionally, the invocation module 53 is further configured to traverse the plurality of first candidate topics, the plurality of second candidate topics, and the plurality of third candidate topics, calculate the repeated similarity between any two candidate topics, and remove the candidate topics with a repeated similarity greater than the first similarity threshold to obtain a plurality of candidate topics.

[0126] Optionally, the invocation module 53 is further configured to calculate the semantic similarity between the plurality of candidate topics and the historical conversation information, and remove the candidate topics with a semantic similarity greater than the second similarity threshold to obtain a plurality of recommended topics.

[0127] Optionally, the invocation module 53 is further configured to calculate the priorities of the plurality of recommended topics; sort the plurality of recommended topics according to the magnitudes of the priorities to obtain a first sorting result of the plurality of recommended topics, and obtain the target recommended topic through the first sorting result.

[0128] Optionally, the invocation module 53 is further configured to obtain a preset annotation information table, where the preset annotation information table includes annotation information, the field of the annotation information, and the weight of the field; for each of the multiple recommended topics, calculate the priority of the recommended topic according to the preset annotation information table with weighting.

[0129] Optionally, the invocation module 53 is further configured to sort the multiple recommended topics according to the relevance between the multiple recommended topics and the historical conversation information and the degree of interest of the user, obtain a second sorting result of the multiple recommended topics, and obtain a target recommended topic through the second sorting result.

[0130] Optionally, the invocation module 53 is further configured to extract a first entity list from the historical conversation information; for each of the recommended topics, extract a second entity list from the recommended topic; calculate the intersection entities of the first entity list and the second entity list; annotate any of the intersection entities according to the degree of interest of the user to obtain the weight annotation information of the intersection entities; calculate the scores of the multiple recommended topics according to the weight annotation information of the intersection entities; and sort the multiple recommended topics according to the magnitudes of the scores.

[0131] Figure 5 The topic recommendation device in the illustrated embodiment can be used to execute the technical solutions of the above-mentioned topic recommendation method embodiments, and its implementation principle and technical effects are similar, which will not be elaborated here.

[0132] Figure 6 The following is a schematic structural diagram of an electronic device provided by an embodiment of the present disclosure. This electronic device can be the terminal described in the above embodiment. The electronic device provided by the embodiment of the present disclosure can execute the processing flow provided by the topic recommendation method embodiment, as Figure 6 shown, the electronic device 60 includes: a memory 61, a processor 62, a computer program, and a communication interface 63; wherein, the computer program is stored in the memory 61 and is configured to be executed by the processor 62 to perform the above-mentioned topic recommendation method.

[0133] In addition, an embodiment of the present disclosure further provides a computer-readable storage medium, on which a computer program is stored, and the computer program is executed by a processor to implement the topic recommendation method described in the above embodiment.

[0134] In addition, an embodiment of the present disclosure further provides a vehicle, which includes the topic recommendation device described in the above embodiment; or the electronic device described in the above embodiment; or the computer-readable storage medium described in the above embodiment.

[0135] It should be noted that in this text, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the said element.

[0136] The above are only specific embodiments of the present disclosure, enabling those skilled in the art to understand or implement the present disclosure. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present disclosure. Therefore, the present disclosure will not be limited to these embodiments described herein, but rather will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A topic recommendation method, characterized in that, The method includes: Obtaining historical conversation information, where the historical conversation information includes historical question information and historical answer information; Extracting multi-channel input information from the historical conversation information, where the multi-channel input information includes at least two of the historical question information, the historical conversation information, and keywords, with each item serving as one channel of input information, and where the keywords are extracted from the historical conversation information; Using the multi-channel input information as the input to a trained large model, calling the trained large model for multi-channel recall, and outputting a target recommended topic.

2. The method according to claim 1, wherein Using the multi-channel input information as the input to a trained large model, calling the trained large model for multi-channel recall, and outputting a target recommended topic, including: Using the historical question information, the historical conversation information, and the keywords as each channel of input to the trained large model, and respectively outputting multiple first candidate topics, multiple second candidate topics, and multiple third candidate topics through the emergence ability of the trained large model; Removing duplicates from the multiple first candidate topics, the multiple second candidate topics, and the multiple third candidate topics to obtain multiple recommended topics; Sorting the multiple recommended topics, and obtaining the target recommended topic through the sorting result of the multiple recommended topics.

3. The method according to claim 2, wherein Removing duplicates from the multiple first candidate topics, the multiple second candidate topics, and the multiple third candidate topics to obtain multiple recommended topics, including: Removing duplicates internally from the multiple first candidate topics, the multiple second candidate topics, and the multiple third candidate topics to obtain multiple candidate topics; Comparing the multiple candidate topics with the historical conversation information, and removing duplicates externally from the multiple candidate topics to obtain multiple recommended topics.

4. The method according to claim 3, characterized in that, Removing duplicates internally from the multiple first candidate topics, the multiple second candidate topics, and the multiple third candidate topics to obtain multiple candidate topics, including: Traversing the multiple first candidate topics, the multiple second candidate topics, and the multiple third candidate topics, calculating the repeated similarity between any two candidate topics, and removing the candidate topics with a repeated similarity greater than the first similarity threshold to obtain multiple candidate topics.

5. The method according to claim 3, characterized in that, Comparing the multiple candidate topics with the historical conversation information, and removing duplicates externally from the multiple candidate topics to obtain multiple recommended topics, including: Calculating the semantic similarity between the multiple candidate topics and the historical conversation information, and removing the candidate topics with a semantic similarity greater than the second similarity threshold to obtain multiple recommended topics.

6. The method according to claim 2, wherein Sorting the multiple recommended topics, and obtaining the target recommended topic through the sorting result of the multiple recommended topics, including: Calculating the priorities of the multiple recommended topics; Sorting the multiple recommended topics according to the magnitudes of the priorities to obtain a first sorting result of the multiple recommended topics, and obtaining the target recommended topic through the first sorting result.

7. The method according to claim 6, characterized in that, Calculating the priorities of the multiple recommended topics, including: Obtaining a preset annotation information table, where the preset annotation information table includes annotation information, the field of the annotation information, and the weight of the field; For each of the multiple recommended topics, calculate the priority of the recommended topic according to the preset annotation information table with weights.

8. The method according to claim 2, wherein Sort the multiple recommended topics, and obtain the target recommended topic through the sorting results of the multiple recommended topics, including: Sort the multiple recommended topics according to the relevance between the multiple recommended topics and the historical conversation information and the degree of user interest, obtain the second sorting result of the multiple recommended topics, and obtain the target recommended topic through the second sorting result.

9. The method according to claim 8, characterized in that, Sort the multiple recommended topics according to the relevance between the multiple recommended topics and the historical conversation information, the relevance between the multiple recommended topics and the historical conversation information, and the degree of user interest, including: Extract the first entity list from the historical conversation information; For each of the recommended topics, extract the second entity list from the recommended topic; Calculate the intersection entities of the first entity list and the second entity list; Annotate any of the intersection entities according to the degree of user interest to obtain the weight annotation information of the intersection entities; Calculate the scores of the multiple recommended topics by weighting according to the weight annotation information of the intersection entities; Sort the multiple recommended topics according to the magnitude of the scores.

10. A topic recommendation device, characterized in that, The device includes: An acquisition module, configured to acquire historical conversation information, where the historical conversation information includes historical question information and historical answer information; An extraction module, configured to extract multi-channel input information from the historical conversation information, where the multi-channel input information includes at least two of the historical question information, the historical conversation information, and keywords, and each item is used as one channel of input information, where the keywords are extracted from the historical conversation information; An invocation module, configured to use the multi-channel input information as the input of a trained large model, and invoke the trained large model for multi-channel recall to output a target recommended topic.

11. An electronic device, characterized in that, Includes: A memory; A processor; And A computer program; Wherein, the computer program is stored in the memory and is configured to be executed by the processor to implement the method according to any one of claims 1-9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method according to any one of claims 1-9.

13. A vehicle, characterized in that, Includes: The topic recommendation device according to claim 10; Or, the electronic device according to claim 11; Or, the computer-readable storage medium according to claim 12.

Images