A conversation dataset construction system for training user item preference recommendations

Abstract

This invention provides a conversation dataset construction system for training user item preference recommendations. The system includes a text conversation generation module, a voice conversation generation module, and a conversation data quality assessment module. These modules construct text-based and voice-based conversations for the user and virtual assistant, and assess conversation data quality. This completes the generation of text conversation content, selects the optimal role based on the user's situation, and ultimately aligns the generated data for a comprehensive assessment of quality. This provides a novel conversational approach, improving the accuracy of recommendations during conversations.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence dialogue recommendation, and in particular to a dialogue dataset construction system for training user item preference recommendations, an electronic device, and a readable storage medium. Background Art

[0002] Conversational Recommendation Systems (CRS) have recently garnered significant attention in the field of recommender systems (RS). The inherent advantages of CRS, such as the ability to capture user preferences through interactive dialogue and provide intuitive explanations for recommendations, align with the growing trend of explainable artificial intelligence (AI). To date, most research on CRS has focused on text-based conversations between users and virtual assistants (TCRS). However, TCRS are not user-friendly for some users, such as those with visual impairments or limited writing and reading skills.

[0003] Compared to text, voice conversations are more direct, natural, convenient, and time-saving. Using voice input in CRS (i.e., Voice-based Convertional Recommendation Systems, VCRS) can bring a series of benefits. First, the user experience is more direct, engaging, and efficient because users are less likely to be distracted or lost browsing products. With VCRS, users can ask exactly what they want and place orders conveniently. Second, VCRS are more conversational in nature and allow for the inclusion of objective deep features and subjective features in users' queries. Third, voice input can convey more information than text input, such as user age, gender, and accent, which enables VCRS to provide more accurate and transparent recommendations. Finally, VCRS are more user-friendly and easy to use, especially for individuals who have difficulty writing and reading.

[0004] In addition to providing a user-friendly and convenient way, voice conversations also convey more information (such as age, gender, accent, and emotional state) than plain text conversations. In addition, extracting this auxiliary information from speech is necessary and more reliable for two reasons:

[0005] 1) Some auxiliary information can only be obtained from speech, such as accent and emotional state. It is difficult to obtain such detailed and dynamic information through user registration in the back-end system.

[0006] 2) While text can capture some auxiliary information (e.g., gender, age), it often suffers from missing data and false information, i.e., user registration information is randomly entered or has restricted access. In contrast, human voices are honest in conveying such information, especially when engaging in voice-based conversations.

[0007] Therefore, in order to effectively utilize the characteristics of text and voice data, it is necessary to construct a dialogue dataset suitable for training and testing user item preference recommendation models to ensure the recommendation accuracy of the model.

[0008] It can be seen that the shortcomings of the prior art and the reasons for these shortcomings are:

[0009] First, text-based conversational recommendation systems face difficulties in understanding user intent and the complexity of conversation management, especially in dealing with ambiguous expressions and maintaining consistency across multiple rounds of conversations.

[0010] Second, the accuracy of personalized recommendations and user experience are affected by insufficient data, slow system response, and sentiment recognition errors, which may lead to poor recommendation results and reduced user satisfaction. Summary of the Invention

[0011] In view of the above problems, the present invention is proposed to provide a conversation dataset construction system, an electronic device and a readable storage medium for training user item preference recommendations that overcome the above problems or at least partially solve the above problems.

[0012] The present invention provides a system for constructing a conversation dataset for training user item preference recommendations, comprising:

[0013] A text conversation generation module is configured to generate conversation templates for the user and the virtual assistant, respectively, based on the project topic; the conversation templates include a question template and a reply template; select a target conversation template from the conversation templates for each round of conversation between the user and the virtual assistant, respectively; analyze user-project interaction data to determine the importance ranking of each conversation feature, and generate at least one round of questions using the target conversation template and the conversation features, in descending order of importance, and obtain real-term feature values ​​of the conversation features, thereby obtaining a text conversation between the user and the virtual assistant;

[0014] A speech dialogue generation module, which uses an end-to-end text-to-speech model to convert text dialogues between the user and the virtual assistant into speech dialogues;

[0015] The conversation data quality assessment module is used to evaluate the quality of text and voice conversations between the user and virtual assistant.

[0016] Optionally, each conversation template is configured with a preset tag;

[0017] The text dialogue generation module is also used for:

[0018] In each round of conversation, a conversation template is randomly selected from the user side or the virtual assistant side, and a corresponding conversation template is selected from the other side according to a predefined tag compatibility mode to form the conversation template matched in each round of conversation; the predefined tag compatibility mode is compatible with the conversation templates with the same preset tags in the user side and the virtual assistant side.

[0019] Optionally, the inquiry template corresponds to different questioning methods, including a selection-based inquiry template and a query-based inquiry template;

[0020] The text dialogue generation module is further configured to select inquiry templates with different questioning methods for the virtual assistant according to the current dialogue stage; if the current dialogue stage is the question-asking stage, the selection-based inquiry template is selected for the virtual assistant.

[0021] Optionally, the text dialogue generation module is further configured to:

[0022] Clean and feature-engineer user-project interaction data to obtain conversation features;

[0023] The LightGBM model based on decision tree is used to analyze the importance of each dialogue feature and obtain the importance score of each dialogue feature;

[0024] The conversation features are ranked according to their importance scores to obtain importance ranking results of the conversation features.

[0025] Optionally, the text dialogue generation module is further configured to:

[0026] Determine the number of question rounds based on the questioning method of the inquiry template selected for the virtual assistant;

[0027] If the inquiry template selected for the virtual assistant side is a query-based inquiry template, the number of rounds of questioning is one;

[0028] If the inquiry template selected for the virtual assistant is a selection-based inquiry template, the number of question rounds is determined according to the time point when the feature value of the real item is successfully obtained.

[0029] Optionally, the voice dialogue generation module is used to:

[0030] A first end-to-end text-to-speech model is used to convert text conversations on the virtual assistant end into speech conversations; the first end-to-end text-to-speech model is trained using the LJ Speech dataset; the LJ Speech dataset contains different short audio clips of a single speaker.

[0031] Optionally, the voice dialogue generation module is used to:

[0032] A second end-to-end text-to-speech model is used to convert text conversations on the user side into voice conversations that match the basic personal information of different users; the second end-to-end text-to-speech model is trained using the Vctk dataset; the Vctk dataset contains voice data with different accents from different speakers.

[0033] Optionally, the conversation data quality assessment module is used to:

[0034] The quality of the textual conversations between the user and the virtual assistant is measured using fine-grained conversation evaluation metrics, using the same evaluations conducted on multiple real-world human conversation datasets as a benchmark. The dimensions of the fine-grained conversation evaluation metrics include comprehension, relevance, coherence, interactivity, interest, diversity, naturalness, accuracy, informativeness, conciseness, fluency, consistency, trustworthiness, politeness, supportiveness, innovation, emotional expression, and inquiry.

[0035] For the text conversation fragments extracted from the text conversation between the user end and the virtual assistant end, an end-to-end text-to-speech model is used to generate the voice conversation between the user end and the virtual assistant end, and a two-stage TTS model is used to generate voice data. Based on the user survey results, the quality of the generated audio fragments is evaluated and compared in terms of audio naturalness, speaker age recognition accuracy, and speaker gender recognition accuracy.

[0036] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the electronic device executes the computer program, it loads a conversation dataset construction system for training user item preference recommendations as described in any one of the embodiments of the present invention.

[0037] The present invention also provides a readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the system for constructing a conversation dataset for training user item preference recommendations as described in any one of the embodiments of the present invention is loaded.

[0038] The present invention includes the following advantages:

[0039] The present invention's conversation dataset construction system for training user item preference recommendations generates textual conversation content by constructing text-based and voice-based conversations between the user and virtual assistant, and generating conversation data quality assessments. It then selects appropriate roles based on the user's situation to optimize the role match, and ultimately conducts a comprehensive assessment of the generated quality. This provides a new conversational approach and improves the accuracy of recommendations during conversations. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 The present invention provides a system for constructing a conversation data set for training user item preference recommendations. DETAILED DESCRIPTION

[0041] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] Reference Figure 1 , shows a structural block diagram of a system for constructing a conversation dataset for training user item preference recommendations provided in an embodiment of the present invention, which may specifically include the following modules:

[0043] A text conversation generation module is configured to generate conversation templates for the user and the virtual assistant, respectively, based on the project topic; the conversation templates include a question template and a reply template; select a target conversation template from the conversation templates for each round of conversation between the user and the virtual assistant, respectively; analyze user-project interaction data to determine the importance ranking of each conversation feature, and generate at least one round of questions using the target conversation template and the conversation features, in descending order of importance, and obtain real-term feature values ​​of the conversation features, thereby obtaining a text conversation between the user and the virtual assistant;

[0044] A speech dialogue generation module, which uses an end-to-end text-to-speech model to convert text dialogues between the user and the virtual assistant into speech dialogues;

[0045] The conversation data quality assessment module is used to evaluate the quality of text and voice conversations between the user and virtual assistant.

[0046] In one embodiment of the present invention, each dialogue template is configured with a preset tag;

[0047] The text dialogue generation module is also used for:

[0048] In each round of conversation, a conversation template is randomly selected from the user side or the virtual assistant side, and a corresponding conversation template is selected from the other side according to a predefined tag compatibility mode to form the conversation template matched in each round of conversation; the predefined tag compatibility mode is compatible with the conversation templates with the same preset tags in the user side and the virtual assistant side.

[0049] In one embodiment of the present invention, the query template corresponds to different questioning methods, including a selection-based query template and a query-based query template;

[0050] The text dialogue generation module is further configured to select inquiry templates with different questioning methods for the virtual assistant according to the current dialogue stage; if the current dialogue stage is the question-asking stage, the selection-based inquiry template is selected for the virtual assistant.

[0051] In one embodiment of the present invention, the text dialogue generation module is further configured to:

[0052] Clean and feature-engineer user-project interaction data to obtain conversation features;

[0053] The LightGBM model based on decision tree is used to analyze the importance of each dialogue feature and obtain the importance score of each dialogue feature;

[0054] The conversation features are ranked according to their importance scores to obtain importance ranking results of the conversation features.

[0055] In one embodiment of the present invention, the text dialogue generation module is further configured to:

[0056] Determine the number of question rounds based on the questioning method of the inquiry template selected for the virtual assistant;

[0057] If the inquiry template selected for the virtual assistant side is a query-based inquiry template, the number of rounds of questioning is one;

[0058] If the inquiry template selected for the virtual assistant is a selection-based inquiry template, the number of question rounds is determined according to the time point when the feature value of the real item is successfully obtained.

[0059] In one embodiment of the present invention, the voice dialogue generation module is used to:

[0060] A first end-to-end text-to-speech model is used to convert text conversations on the virtual assistant end into speech conversations; the first end-to-end text-to-speech model is trained using the LJ Speech dataset; the LJ Speech dataset contains different short audio clips of a single speaker.

[0061] In one embodiment of the present invention, the voice dialogue generation module is used to:

[0062] A second end-to-end text-to-speech model is used to convert text conversations on the user side into voice conversations that match the basic personal information of different users; the second end-to-end text-to-speech model is trained using the Vctk dataset; the Vctk dataset contains voice data with different accents from different speakers.

[0063] In one embodiment of the present invention, the conversation data quality assessment module is configured to:

[0064] The quality of the textual conversations between the user and the virtual assistant is measured using fine-grained conversation evaluation metrics, using the same evaluations conducted on multiple real-world human conversation datasets as a benchmark. The dimensions of the fine-grained conversation evaluation metrics include comprehension, relevance, coherence, interactivity, interest, diversity, naturalness, accuracy, informativeness, conciseness, fluency, consistency, trustworthiness, politeness, supportiveness, innovation, emotional expression, and inquiry.

[0065] For the text conversation fragments extracted from the text conversation between the user end and the virtual assistant end, an end-to-end text-to-speech model is used to generate the voice conversation between the user end and the virtual assistant end, and a two-stage TTS model is used to generate voice data. Based on the user survey results, the quality of the generated audio fragments is evaluated and compared in terms of audio naturalness, speaker age recognition accuracy, and speaker gender recognition accuracy.

[0066] In the present invention, for each user-item interaction record, the goal is to simulate multiple rounds of dialogue by asking the user for their preferences for features related to the item until the item is successfully recommended. The application scenarios of the dialogue recommendation of the present invention are very broad. The goal of the system design is to capture the user's preference information during the interaction with any type of item or topic (such as services, content, experience, etc.). Therefore, this dialogue recommendation solution is applicable to various scenarios that require understanding user preferences and making personalized recommendations. In the process of introducing the solution, the present invention uses item features as an example to illustrate how the system captures user preferences through multiple rounds of dialogue, but the dialogue recommendation of the present invention is not limited to the recommendation product scenario.

[0067] In the process of generating textual dialogues between the user and the virtual assistant, the first step is to generate dialogue templates for the user and the virtual assistant, respectively. To ensure the naturalness and diversity of the dialogues, different query and response templates need to be generated.

[0068] The user-side template primarily simulates user behavior and reactions during conversations. This template includes the types of questions users might ask, how they respond, and their preferred choices in different scenarios. Its purpose is to provide the system with diverse user interaction modes, enabling simulations of conversations as close to real users as possible during model testing and training. The virtual assistant-side template defines the system's response patterns when interacting with users. This template includes the virtual assistant's responses to user inquiries, the suggestions it provides, and how it adjusts its next questions or suggestions based on user feedback. Its goal is to ensure the virtual assistant's responses are logical, accurate, and personalized. The two differ in role positioning and interaction goals: the user-side template focuses on simulating diverse user behaviors, while the virtual assistant-side template focuses on optimizing the system's responses. By combining these two templates, we can ensure the rationality and richness of the conversation generation process.

[0069] Especially for virtual assistants, when given each item feature, the present invention will change the way of asking questions, either directly asking the user's favorite item features (based on the inquiry, tag identifier tag-id: 00), or providing single or multiple choices (based on the choice, tag identifier tag-id: 01).

[0070] Furthermore, the present invention generates templates for both the virtual assistant and the user regarding item recommendations and opening / closing conversations. All of these templates under different tag-ids can be generated by ChatGPT, a large language model (LLM) released by OpenAI. ChatGPT's training largely follows InstrucGPT, which utilizes supervised training with human expert labels for question-answering conversations and reinforcement learning with an expert-driven reward model.

[0071] After generating conversation templates for the user and virtual assistant, it's necessary to select the appropriate template for each conversation between the virtual assistant and the user. Due to the diversity of templates, conversations with mismatched questions and answers are inevitable. To avoid these awkward situations and increase conversational coherence, the present invention has devised the following strategies:

[0072] Specifically, the present invention first adds a tag-id to each template on the virtual assistant and the user, and defines compatible matching patterns for these tags. Only templates with the same tag-id on the virtual assistant and the user are compatible. Therefore, in each round of the conversation, the present invention randomly samples a template from each of the virtual assistant and the user, guided by the predefined compatible pattern. This ensures the rationality of the generated conversation.

[0073] Specifically, during the question-posing phase, the present invention provides the virtual assistant with a greater probability of sampling a choice-based template (tag-id: 01) rather than a query-based template (tag-id: 00). This is because users typically do not initially have a clear intent when interacting with a CRS. In other words, the CRS requires multiple rounds of dialogue to elicit user preferences.

[0074] After providing the selected template, it is necessary to determine the best question to ask. In other words, it is necessary to select features from the conversation so that the system can quickly identify user preferences and ultimately make better recommendations. For example, in the product recommendation scenario, the system determines the characteristics of the item being asked about, and in the functional service scenario, the system determines the functional characteristics. This invention designs the following strategies:

[0075] Specifically, it is necessary to determine the order in which the feature types are asked, because different types of features may have different impacts on the user's final decision. If the feature types with greater influence can be asked earlier, the user's preferences can be inferred more quickly. Therefore, the present invention uses the decision tree-based method LightGBM to calculate the importance of different feature types. Feature types with higher importance values ​​obtained by LightGBM are closer to the root, indicating that they have a greater impact on the user's final decision, and therefore should be asked at an earlier stage. The resulting decision tree will eventually participate in the text-based dialogue generation process to determine the order in which the feature types are asked.

[0076] When determining the order of feature types, this paper uses LightGBM, a decision tree-based gradient boosting framework, to calculate the importance of each feature. The specific steps are as follows:

[0077] 1. Data preprocessing: First, the user interaction data will be cleaned and feature engineered to ensure the quality and consistency of the features.

[0078] 2. Training model: The processed data is then input into the LightGBM model for training. During the training process, the model constructs a decision tree based on the importance of the features and outputs the importance score of each feature.

[0079] 3. Feature sorting: Based on the importance scores output by the LightGBM model, features are sorted from high to low. Features with higher scores generally have a greater impact on user decisions. These features will be closer to the root of the decision tree, indicating that they contribute more to the model prediction.

[0080] 4. Conversation optimization: Using the ranking results of these feature importance scores, the conversation system prioritizes asking users questions about high-importance features, thereby capturing user preferences more quickly and accurately and optimizing recommendation results.

[0081] Specifically, the present invention requires determining the number of turns for each queried feature type. If the selected virtual assistant template is query-based, the present invention only requires one turn to directly obtain the actual item feature value of the queried feature type from the user. If it is selection-based, the number of turns depends on when the actual feature value can be sampled. To ensure the rationality of the conversation, the present invention adheres to the interaction distribution, that is, the feature value of the queried feature type when sampling the feature value based on the selected template.

[0082] After generating a text conversation between the user and the virtual assistant, the generated text conversation can be converted into a voice conversation, ensuring that the system can generate realistic voice interaction scenarios and better simulate the user experience. Through advanced technical means, the user experience and the capabilities of the dialogue system are improved, while promoting the widespread application and development of the technology.

[0083] Specifically, text-to-speech dialogue generation is achieved. The present invention utilizes the end-to-end text-to-speech model VITS to complete text-to-speech dialogue generation. On the virtual assistant side, a VITS model trained with the LJ Speech dataset is used. The LJ Speech dataset may contain short audio clips from a single speaker. On the user side, a VITS model trained with the Vctk dataset is used. The Vctk dataset may contain speech data from 109 English speakers with different accents. For each text-based dialogue, an appropriate speaker is matched based on the user's gender and age to convert it into a speech dialogue.

[0084] The quality of the constructed conversation datasets can be evaluated separately for the generated text conversations and speech conversations.

[0085] To automatically and systematically measure the quality of generated text-based dialogues, we employ the fine-grained evaluation of dialogue (FED) metric, an automatic evaluation metric using the large-scale pre-trained model DialoGPT. Using the FED metric has three advantages: 1) it does not rely on real responses, 2) it does not require any training data, and 3) it can measure both fine-grained dialogue quality (e.g., correctness, comprehensibility, and interestingness) and overall dialogue quality (e.g., coherence, consistency, and diversity) at both the turn level. To ensure the validity of our dataset, we used the same evaluation of three real-world human dialogue datasets (i.e., ReDial, OpenDialKG, and INSPIRED) as a benchmark.

[0086] Specifically, we use a human conversation dataset (e.g., ReDial) as a benchmark and calculate the performance gap between it and two other datasets (e.g., OpenDialKG and INSPIRED). Generally speaking, the smaller the performance gap with the benchmark, the better the quality of the corresponding dataset.

[0087] Specifically, the results obtained by FED in 18 aspects at the turn and dialogue levels (understanding, relevance, coherence, interactivity, interest, diversity, naturalness, accuracy, informativeness, conciseness, fluency, consistency, trustworthiness, politeness, supportiveness, innovation, emotional expression, and inquiry) are averaged as the final performance score: "Average" refers to the average score of the three benchmark datasets; "Gap-R" represents the absolute performance gap between the other two models.

[0088] To evaluate the quality of the generated speech dialogues, we compared the speech data generated by end-to-end VITS with those generated by three two-stage TTS models (i.e., Text2Mel+Vocoders): Tacotron 2+WaveGAN (TacoW), Transformer+WaveGAN (TranW), and FastSpeech 2+WaveGAN (FastW).

[0089] Specifically, we first randomly sampled 10 text snippets from the text-based conversations generated on the selected dataset. We then used VITS and three baseline TTS models to generate corresponding audio snippets. Finally, we obtained 20 sets of audio clips, each associated with a single text snippet and four audio clips generated by the four TTS models.

[0090] Specifically, based on user research, the quality of the four synthesized audio clips in each group was evaluated from three different perspectives:

[0091] (1) Obtaining the mean opinion score (MOS), i.e., the naturalness of the audio clip, which is divided into 1-5 levels (bad, poor, average, good, excellent);

[0092] (2) Use different age-stage options to determine the speaker's age;

[0093] (3) Use three options to determine whether the speaker is male, female, or other.

[0094] We recruited multiple participants to answer each of the 20 sets of questions. To ensure fair evaluation without introducing personal bias, we randomly selected four audio clips generated by four different TTS models from each set to prevent participants from over- or under-rating certain audio clips.

[0095] Based on the same inventive concept, another embodiment of the present invention provides an electronic device, including a memory, a processor, and a computer program stored on the memory, wherein the processor executes the computer program to load a conversation dataset construction system for training user item preference recommendations as described in any one of the embodiments of the present invention.

[0096] Specifically, the electronic device includes: a memory and a processor, the memory and the processor are connected via a bus communication, the memory stores a computer program, and the computer program can be run on the processor, thereby loading the conversation dataset construction system for training user item preference recommendations as described in any one of the first aspects of the embodiments of the present invention.

[0097] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage. Alternatively, the memory may be at least one storage device located away from the processor.

[0098] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.

[0099] Based on the same inventive concept, another embodiment of the present invention provides a computer-readable storage medium having a computer program / instruction stored thereon. When the computer program / instruction is executed by a processor, the system for constructing a conversation dataset for training user item preference recommendations as described in any one of the first aspects of the embodiments of the present invention is loaded.

[0100] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0101] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, devices, electronic devices, storage media, or computer program products. Therefore, embodiments of the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Furthermore, embodiments of the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CDROM, optical storage, etc.) containing computer-usable program code.

[0102] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

[0103] Finally, it should be noted that, in this document, relational terms such as first and second, etc., are used only 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 "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or terminal device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or terminal device that includes the element.

[0104] The above is a detailed introduction to a conversation dataset construction system, electronic device and readable storage medium for training user item preference recommendations provided by the present invention. Specific examples are used herein to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method and core idea of ​​the present application. At the same time, for those skilled in the art, according to the idea of ​​the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present application. The above embodiments are only preferred embodiments for fully illustrating the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or changes made by those skilled in the art based on the present invention are all within the scope of protection of the present invention.

Claims

1. A conversation dataset construction system for training user item preference recommendation, characterized by: The system comprises: A text conversation generation module is configured to generate conversation templates for the user and the virtual assistant, respectively, based on the project topic; the conversation templates include a question template and a reply template; select a target conversation template from the conversation templates for each round of conversation between the user and the virtual assistant, respectively; analyze user-project interaction data to determine the importance ranking of each conversation feature, and generate at least one round of questions using the target conversation template and the conversation features, in descending order of importance, and obtain real-term feature values ​​of the conversation features, thereby obtaining a text conversation between the user and the virtual assistant; A speech dialogue generation module, which uses an end-to-end text-to-speech model to convert text dialogues between the user and the virtual assistant into speech dialogues; A conversation data quality assessment module, used to assess the quality of text and voice conversations between the user and virtual assistant. The voice dialogue generation module is used to: A second end-to-end text-to-speech model is used to convert the text conversation of the user end into a voice conversation that matches the basic personal information of different users; the second end-to-end text-to-speech model is trained using a Vctk dataset; the Vctk dataset contains speech data of different speakers with different accents; The dialogue data quality assessment module is used to: Using fine-grained conversation evaluation metrics to measure the conversation quality of the text conversation between the user and the virtual assistant, and using the same evaluation metrics on multiple real-world human conversation datasets as a benchmark; For the text conversation fragments extracted from the text conversation between the user end and the virtual assistant end, an end-to-end text-to-speech model is used to generate the voice conversation between the user end and the virtual assistant end, and a two-stage TTS model is used to generate voice data. Based on the user survey results, the generated audio fragments are evaluated and compared in multiple dimensions.

2. The system for constructing a conversation dataset for training user item preference recommendations according to claim 1, characterized in that: Each conversation template is configured with a preset tag; The text dialogue generation module is also used for: In each round of dialogue, a dialogue template is randomly selected from the user side or the virtual assistant side, and a corresponding dialogue template is selected from the other side according to the predefined tag compatibility mode to form the dialogue template matching each round of dialogue; The predefined tag compatibility mode is that the dialogue templates with the same preset tags in the user terminal and the virtual assistant terminal are compatible.

3. The system for constructing a conversation dataset for training user item preference recommendation according to claim 1, characterized in that: The query templates correspond to different questioning methods, including selection-based query templates and query-based query templates; The text dialogue generation module is further configured to select inquiry templates with different questioning methods for the virtual assistant according to the current dialogue stage; if the current dialogue stage is the question-asking stage, the selection-based inquiry template is selected for the virtual assistant.

4. The system for constructing a conversation dataset for training user item preference recommendations according to claim 1, characterized in that: The text dialogue generation module is also used for: Clean and feature-engineer user-project interaction data to obtain conversation features; The LightGBM model based on decision tree is used to analyze the importance of each dialogue feature and obtain the importance score of each dialogue feature; The conversation features are ranked according to their importance scores to obtain importance ranking results of the conversation features.

5. The system for constructing a conversation dataset for training user item preference recommendation according to claim 3, characterized in that: The text dialogue generation module is also used for: Determine the number of question rounds based on the questioning method of the inquiry template selected for the virtual assistant; If the inquiry template selected for the virtual assistant side is a query-based inquiry template, the number of rounds of questioning is one; If the inquiry template selected for the virtual assistant is a selection-based inquiry template, the number of question rounds is determined according to the time point when the feature value of the real item is successfully obtained.

6. The system for constructing a conversation dataset for training user item preference recommendation according to claim 1, characterized in that: The voice dialogue generation module is used to: A first end-to-end text-to-speech model is used to convert text conversations on the virtual assistant end into speech conversations; the first end-to-end text-to-speech model is trained using the LJ Speech dataset; the LJ Speech dataset contains different short audio clips of a single speaker.

7. The system for constructing a conversation dataset for training user item preference recommendation according to claim 1, characterized in that: The dialogue data quality assessment module is used to: The quality of the textual conversations between the user and the virtual assistant is measured using fine-grained conversation evaluation metrics, using the same evaluations conducted on multiple real-world human conversation datasets as a benchmark. The dimensions of the fine-grained conversation evaluation metrics include comprehension, relevance, coherence, interactivity, interest, diversity, naturalness, accuracy, informativeness, conciseness, fluency, consistency, trustworthiness, politeness, supportiveness, innovation, emotional expression, and inquiry. For the text conversation fragments extracted from the text conversation between the user end and the virtual assistant end, an end-to-end text-to-speech model is used to generate the voice conversation between the user end and the virtual assistant end, and a two-stage TTS model is used to generate voice data. Based on the user survey results, the quality of the generated audio fragments is evaluated and compared in terms of audio naturalness, speaker age recognition accuracy, and speaker gender recognition accuracy.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the electronic device executes the computer program, it loads the conversation dataset construction system for training user item preference recommendation according to any one of claims 1 to 7.

9. A readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the system for constructing a conversation dataset for training user item preference recommendation according to any one of claims 1 to 7 is loaded.

Images