Question and answer method and system

Abstract

The invention discloses a question and answer method and system, and the method comprises the following steps: obtaining user input, and recognizing the user input; after identification of user input is completed, training of a vertical model is carried out; after the training of the model is completed, constructing a knowledge base; after the knowledge base is constructed, intention recognition is carried out; after intention recognition is carried out, semantic expansion is carried out on user input, and knowledge recall is carried out; generating an answer according to a knowledge recall result; and performing voice synthesis on the generated answer. The answering system which can meet complex scenes and is good in interaction and high in reply quality is constructed, the working efficiency is improved, and the working cost is saved.

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular, to a question-answering method and a system thereof. Background Art

[0002] With the continuous development of artificial intelligence technology, the question-answering system is also constantly improving and perfecting. Through the question-answering system, users' questions can be answered in a short time, and the core points of the questions can be accurately understood. The question-answering system can be applied in many fields, such as Taobao customer service and government customer service. Taking government customer service as an example, this customer service can effectively solve a large number of manual consultation and complaint problems, release human resources, and reduce the cost of government services. At the same time, it can respond to the public's demands in a timely manner, give suggestions and processing methods, and improve the public's satisfaction with government services. The current government intelligent customer service solution mainly builds question-answer pairs of common questions in advance, and answers the public's demands based on traditional NLP models and rule matching. Traditional government intelligent customer service relies too much on the construction of question-answer pairs and the matching of rules. The applicable scenarios are relatively simple and cannot meet complex scenarios. In addition, the real-time and fluency of voice interaction are not high. With the continuous construction of digital government, various livelihood demands are increasing, the scenarios are becoming more complex, and the quality requirements for responses and interactive experience are getting higher and higher. The traditional government intelligent customer service system has obvious shortcomings and cannot make accurate and real-time responses. It usually needs to be transferred to manual processing. Due to the limited number of business personnel, it is difficult to meet the massive needs of the public. It also brings a huge workload to the business handling personnel, and the overall handling efficiency needs to be further improved.

[0003] Therefore, how to provide a question-and-answer method that can save question-and-answer time and reduce work costs has become an urgent problem to be solved in this field. Summary of the invention

[0004] The present application proposes a question-answering method, comprising the following steps: obtaining user input and identifying the user input; after completing the identification of the user input, training a vertical model; after completing the training of the model, constructing a knowledge base; after completing the construction of the knowledge base, performing intent recognition; after performing intent recognition, semantically expanding the user input and performing knowledge recall; based on the result of the knowledge recall, generating an answer; and performing speech synthesis on the generated answer.

[0005] As above, user input recognition is performed through a real-time streaming recognition model, an offline hot word model, and a punctuation prediction model.

[0006] As mentioned above, the process of identifying the user also includes correcting the user input.

[0007] As mentioned above, the training of the vertical model includes: constructing a data set in a specified field, and using the data set to perform supervised instruction fine-tuning on the vertical model.

[0008] As mentioned above, the construction of the knowledge base includes the following sub-steps: performing sub-library storage to complete fine-grained knowledge division; performing unsupervised vector domainization to complete the domainization of the vector model.

[0009] A question-answering system specifically includes: a recognition unit, a training unit, a construction unit, an intention recognition unit, a knowledge recall unit, an answer generation unit and a synthesis unit; the recognition unit is used to obtain user input and recognize the user input; the training unit is used to train a vertical model after completing the recognition of the user input; the construction unit is used to build a knowledge base; the intention recognition unit is used to perform intention recognition; the knowledge recall unit is used to semantically expand the user input and recall knowledge; the answer generation unit is used to generate an answer based on the result of knowledge recall; the synthesis unit is used to perform speech synthesis on the generated answer.

[0010] As above, the recognition unit performs user input recognition through a real-time streaming recognition model, an offline hot word model and a punctuation prediction model.

[0011] As mentioned above, the process of identifying the user by the identification unit further includes correcting the user input.

[0012] As mentioned above, the training unit performs training of the vertical model including: constructing a data set in a specified field, and using the data set to perform supervised instruction fine-tuning on the vertical model.

[0013] As mentioned above, the construction unit constructs the knowledge base including the following sub-steps: performing sub-library storage to complete fine-grained knowledge division; performing unsupervised vector domainization to complete the domainization of the vector model.

[0014] This application has the following beneficial effects:

[0015] (1) This application uses artificial intelligence technologies such as natural language processing, speech recognition and synthesis to build an answering system that can meet complex scenarios, has good interaction and high response quality, improves people's satisfaction with government services, and reduces the workload of business personnel.

[0016] (2) The complete intelligent reply process established by this application, from speech recognition, verticalization of large language model fields, knowledge base construction, intent recognition, knowledge recall, answer generation, and speech synthesis, greatly compensates for the shortcomings of low answer quality and low real-time performance in traditional replies. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0018] Figure 1 is a flowchart of a question-and-answer method provided according to an embodiment of the present application;

[0019] Figure 2 is a flow chart for performing speech recognition according to an embodiment of the present application;

[0020] Figure 3 It is a schematic diagram of a knowledge base construction method provided according to an embodiment of the present application;

[0021] Figure 4 is a flowchart for performing intent recognition according to an embodiment of the present application;

[0022] Figure 5 is a flowchart of performing knowledge recall according to an embodiment of the present application;

[0023] Figure 6 is a flowchart for performing speech synthesis according to an embodiment of the present application;

[0024] Figure 7 It is a schematic diagram of the internal structure of the question-answering system provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] The following is a clear and complete description of the technical solutions in the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0026] This application is based on natural language processing technology and uses artificial intelligence technology to identify the intention of user input. It enhances the knowledge of the big model through the knowledge base and injecting domain knowledge, improves the professional ability of the big model in the field, and thus provides high-quality answers to questions input by users.

[0027] Embodiment 1

[0028] like Figure 1 As shown, this embodiment provides a question-answering method, which specifically includes the following steps:

[0029] Step S110: Obtain user input and identify the user input.

[0030] The user input may be text input or voice input. If it is text input, no conversion is required and the next step can be directly performed.

[0031] If the user input is voice input, it is necessary to recognize the voice input, convert the voice input into text input, and then proceed to the next step. Among them, speech recognition is mainly based on key links such as acoustic feature extraction, acoustic model establishment, language model application and model training. It is a key step in achieving human-computer interaction. Existing methods mainly rely on vocal tract models, template matching, and the use of recurrent neural network architecture to automatically extract features to improve recognition accuracy.

[0032] The real-time speech input recognition used in this embodiment mainly includes a real-time streaming recognition model, an offline hot word model, and a punctuation prediction model. It can output the recognition results in real-time streaming, and can also use the hot word model to transcribe the text with high precision at the end of a sentence, and output text with punctuation.

[0033] Furthermore, this embodiment uses speech analysis technologies such as ASR and TTS to recognize the speech input by the user in real time and automatically correct errors.

[0034] The application field of this embodiment can be the government affairs field. Since there are a large number of specific words in the government affairs field, in order to more accurately identify these words, it also includes building a hot word dictionary in the government affairs field to improve the recognition ability of the model, and dynamically adjust and update the hot word list to perform real-time optimization according to scene requirements. After voice recognition, the hot word dictionary can be used to correct the results.

[0035] like Figure 2 As shown, the specific process of performing speech recognition provided by this embodiment, wherein the input speech is first received from the client through the message queue and processed every 60 milliseconds. Then, the real-time speech endpoint detection model performs real-time speech activity detection, identifies non-silent segments, and passes the results to the speech recognition real-time model every 600 milliseconds for real-time speech recognition. The recognized text is sent to the message queue at intervals of 600 milliseconds and then transmitted back to the client. In addition, a non-streaming model for speech recognition and a punctuation prediction model are used for more accurate offline processing, and the speech activation detection node is coordinated to ensure that the recognized text is corrected and finally output.

[0036] Step S120: After completing the recognition of the user input, the vertical model is trained.

[0037] The vertical model refers to a general large-scale machine learning model (referred to as the big model), which can be oriented to a specific field or task, and has excellent predictive performance in this field or task.

[0038] The implementation of vertical models requires a hierarchical structure, dividing the model training and prediction tasks into several layers, and using different model structures and parameters to complete the training of large vertical models. This requires a large amount of data and computing resources, and also requires model compression and accuracy optimization to reduce the size and computational complexity of the model. Compared with traditional general machine learning models, vertical models can better meet the needs of specific fields or tasks, and usually have higher prediction performance and accuracy.

[0039] Specifically, by constructing a dataset in a specified field, which contains data in the specified field, the dataset is used to perform supervised instruction fine-tuning on the vertical large model, injecting corresponding domain knowledge into the vertical model, and improving the standardization and accuracy of question and answering in the specified field.

[0040] Supervised instruction fine-tuning includes further fine-tuning based on the trained language model using labeled specific task data, so that the model has the ability to follow instructions. Here, the fine-tuning model is performed by constructing a knowledge question and answer pair in the government affairs field.

[0041] For example, by further fine-tuning the big model based on the general big model through the government affairs data set, the big model is injected with knowledge from the government affairs field, the knowledge comprehension ability of the big model is strengthened, and the big model can give more professional and standardized answers.

[0042] The designated field may be a government affairs field, and the data set of the designated field may be a data set of the government affairs field.

[0043] Step S130: After completing the model training, construct the knowledge base.

[0044] This embodiment improves on the traditional knowledge base construction and proposes a more accurate knowledge base construction method, such as Figure 3 shown.

[0045] First, we use sub-database storage to complete the fine-grained knowledge division. We store the knowledge of different business lines in a specified field (such as the government field) separately to avoid the decrease in retrieval accuracy caused by too much mixed knowledge. Therefore, different business knowledge has its corresponding knowledge base.

[0046] Secondly, we fine-tune the unsupervised vector model. Specifically, based on the principle that the vectors obtained after the same sentence passes through the dropout layer are semantically similar but slightly different, the different vectors obtained after the same sentence passes through the dropout layer are used as positive sample pairs, and the vectors of different sentences are used as negative sample pairs. The vector model can be trained without labeling the sentences, thus achieving unsupervised vector model fine-tuning and completing the domainization of the vector model.

[0047] Since the vector model is pre-trained on a general data set, vectorizing data with high verticality in the government field will reduce the recall rate. Therefore, this embodiment uses the SimCSE unsupervised method to fine-tune the vector model on the government field data set to enhance the performance of the vector model in the professional field, and uses unsupervised training to avoid consuming a lot of manpower and material resources.

[0048] The vector model is a common model in the prior art, and the specific structure will not be described in detail here.

[0049] Step S140: After the knowledge base is constructed, intent recognition is performed.

[0050] In the process of constructing the above knowledge base, different knowledge bases are constructed for different business data. Therefore, in this step, intent recognition is used to determine the knowledge base that needs to be retrieved.

[0051] Since user input may have multiple intentions or unclear intentions in actual use, it is necessary to further analyze the user input and correctly judge the user intentions.

[0052] This embodiment uses a large model to teach a small model to infer to determine user intent, which not only makes up for the shortcomings of low accuracy of traditional intent recognition models in complex scenarios, but also improves the problems of slow speed and high cost of using models with ultra-large parameters.

[0053] Specifically, the conventional approach to using a large model for intent recognition is to guide the model to reason step by step through the prompt word thinking chain method to generate an answer. However, for a model with a small number of parameters (0.1B-0.5B), its ability to follow the prompt words is poor and its reasoning ability is weak. This embodiment uses the Finetuning-COT method, such as Figure 4 The figure shows a flowchart for intent recognition. Specifically, a model with super large parameters is used to generate intent recognition data with a thought chain reasoning process, and a variety of different thought chain reasoning data are generated for the same problem to enhance the reasoning diversity of the small model. The small model is fine-tuned through the generated data set, and the thinking and reasoning capabilities of the large model are introduced into the small model to enhance the recognition effect of the small model on complex intents, thereby obtaining the final intent recognition result.

[0054] Step S150: After the intent is recognized, semantic expansion is performed on the user input to perform knowledge recall.

[0055] This embodiment improves on the traditional text recall and expands the semantic space of user input through a large model, such as Figure 5As shown in the figure, it is a flowchart for knowledge recall. Several similar questions are generated based on the original user input, and then the rewritten questions are retrieved and recalled separately. The RRF algorithm is used to rearrange the multi-way recalled documents, filter out documents that are not highly relevant to the questions, and improve the quality of the recalled documents.

[0056] Step S160: Generate answers based on the results of knowledge recall.

[0057] Although documents with low relevance to user input have been filtered in the knowledge recall process in the above steps, there are still some relevant documents with low relevance or that do not contain answers to questions, which will have a negative effect on the generation of the large model. Therefore, this embodiment improves the problem that the large model is easily disturbed by irrelevant text during the generation process, and improves the quality of the large model's responses to user questions.

[0058] Among them, since the large model with hundreds of billions of parameters has super text understanding and generation capabilities, it can filter out the influence of irrelevant text to a large extent and generate the final answer. Therefore, this embodiment first merges the relevant text and the irrelevant text in a certain proportion, generates an answer with a thinking chain through the large model with hundreds of billions of parameters, constructs a data set to fine-tune the model that injects knowledge in the government affairs field, improves the model's ability to filter irrelevant text in the answering process, and generates more useful answers, that is, completes the answer generation.

[0059] Step S170: Perform speech synthesis on the generated answer.

[0060] This embodiment adopts the KAN-TTS solution, which consists of a text front-end, an acoustic model, and a vocoder. The front-end module uses a model combined with rules to flexibly process texts in various scenarios, while the back-end module uses SAM-BERT+HIFIGAN to provide highly expressive streaming synthesis effects. Based on KAN-TTS, in order to be applied to intelligent customer service question-and-answer scenarios, a new timestamp function is added, word-level timestamps are provided, openai's whisper is used for timestamp prediction, and a sentence-level streaming TTS interface is added to improve real-time performance.

[0061] like Figure 6 The flowchart of speech synthesis provided by this embodiment is shown in FIG. Specifically, starting from text input, linguistic features are generated through text normalization, word segmentation, disambiguation, and polyphone processing. Then, acoustic features are generated using a duration model and an acoustic model, and finally speech is synthesized through a vocoder.

[0062] Embodiment 2

[0063] like Figure 7As shown, this embodiment provides a question-answering system, which specifically includes: an identification unit 701, a training unit 702, a construction unit 703, an intention identification unit 704, a knowledge recall unit 705, an answer generation unit 706 and a synthesis unit 707.

[0064] The identification unit 701 is used to obtain user input and identify the user input.

[0065] The user input may be text input or voice input. If it is text input, no conversion is required and the next step can be directly performed.

[0066] If it is voice input, it is necessary to recognize the voice input, convert the voice input into text input, and then proceed to the next step. Among them, voice recognition is mainly based on key links such as acoustic feature extraction, acoustic model establishment, language model application and model training. It is a key step in achieving human-computer interaction. Existing methods mainly rely on vocal tract models, template matching, and the use of recurrent neural network architecture to automatically extract features to improve recognition accuracy.

[0067] The real-time speech input recognition used in this embodiment mainly includes a real-time streaming recognition model, an offline hot word model, and a punctuation prediction model. It can output the recognition results in real-time streaming, and can also use the hot word model to transcribe the text with high precision at the end of a sentence, and output text with punctuation.

[0068] Furthermore, this embodiment uses speech analysis technologies such as ASR and TTS to recognize the speech input by the user in real time and automatically correct errors.

[0069] The application field of this embodiment can be the government affairs field. Since there are a large number of specific words in the government affairs field, in order to more accurately identify these words, it also includes building a hot word dictionary in the government affairs field to improve the recognition ability of the model, and dynamically adjust and update the hot word list to perform real-time optimization according to scene requirements. After voice recognition, the hot word dictionary can be used to correct the results.

[0070] like Figure 2 As shown, the specific process of performing speech recognition provided by this embodiment, wherein the input speech is first received from the client through the message queue and processed every 60 milliseconds. Then, the real-time speech endpoint detection model performs real-time speech activity detection, identifies non-silent segments, and passes the results to the speech recognition real-time model every 600 milliseconds for real-time speech recognition. The recognized text is sent to the message queue at intervals of 600 milliseconds and then transmitted back to the client. In addition, a non-streaming model for speech recognition and a punctuation prediction model are used for more accurate offline processing, and the speech activation detection node is coordinated to ensure that the recognized text is corrected and finally output.

[0071] The training unit 702 is used to train the vertical model after completing the recognition of the user input.

[0072] The vertical model refers to a general large-scale machine learning model (referred to as the big model), which can be oriented to a specific field or task, and has excellent predictive performance in this field or task.

[0073] The implementation of vertical models requires a hierarchical structure, dividing the model training and prediction tasks into several layers, and using different model structures and parameters to complete the training of large vertical models. This requires a large amount of data and computing resources, and also requires model compression and accuracy optimization to reduce the size and computational complexity of the model. Compared with traditional general machine learning models, vertical models can better meet the needs of specific fields or tasks, and usually have higher prediction performance and accuracy.

[0074] Specifically, by constructing a dataset in a specified field, which contains data in the specified field, the dataset is used to perform supervised instruction fine-tuning on the vertical large model, injecting corresponding domain knowledge into the vertical model, and improving the standardization and accuracy of question and answering in the specified field.

[0075] For example, by further fine-tuning the big model based on the general big model through the government affairs data set, the big model is injected with knowledge from the government affairs field, the knowledge comprehension ability of the big model is strengthened, and the big model can give more professional and standardized answers.

[0076] The designated field may be a government affairs field, and the data set of the designated field may be a data set of the government affairs field.

[0077] The construction unit 703 is used to construct a knowledge base after completing the training of the model.

[0078] This embodiment improves on the traditional knowledge base construction and proposes a more accurate knowledge base construction method, such as Figure 3 shown.

[0079] First, we use sub-database storage to complete the fine-grained knowledge division. We store the knowledge of different business lines in a specified field (such as the government field) separately to avoid the decrease in retrieval accuracy caused by too much mixed knowledge. Therefore, different business knowledge has its corresponding knowledge base.

[0080] Secondly, unsupervised vector domainization is performed to complete the domainization of the vector model. Since the vector model is pre-trained on a general data set, vectorizing data with high verticality such as the government affairs field will reduce the recall rate. Therefore, this embodiment uses the SimCSE unsupervised method to fine-tune the vector model on the government affairs field data set to enhance the performance of the vector model in the professional field, and trains it through an unsupervised method to avoid consuming a lot of manpower and material resources.

[0081] The vector model is a common model in the prior art, and the specific structure will not be described in detail here.

[0082] The intention recognition unit 704 is used to perform intention recognition after completing the construction of the knowledge base.

[0083] In the process of constructing the above knowledge base, different knowledge bases are constructed for different business data. Therefore, in this step, intent recognition is used to determine the knowledge base that needs to be retrieved.

[0084] Since user input may have multiple intentions or unclear intentions in actual use, it is necessary to further analyze the user input and correctly judge the user intentions.

[0085] This embodiment uses a large model to teach a small model to infer to determine user intent, which not only makes up for the shortcomings of low accuracy of traditional intent recognition models in complex scenarios, but also improves the problems of slow speed and high cost of using models with ultra-large parameters.

[0086] Specifically, the conventional approach to using a large model for intent recognition is to guide the model to reason step by step through the prompt word thinking chain method to generate an answer. However, for a model with a small number of parameters (0.1B-0.5B), its ability to follow the prompt words is poor and its reasoning ability is weak. This embodiment uses the Finetuning-COT method, such as Figure 4 The figure shows a flowchart for intent recognition. Specifically, a model with super large parameters is used to generate intent recognition data with a thought chain reasoning process, and a variety of different thought chain reasoning data are generated for the same problem to enhance the reasoning diversity of the small model. The small model is fine-tuned through the generated data set, and the thinking and reasoning capabilities of the large model are introduced into the small model to enhance the recognition effect of the small model on complex intents, thereby obtaining the final intent recognition result.

[0087] The knowledge recall unit 705 is used to perform semantic expansion on the user input and perform knowledge recall after intent recognition.

[0088] This embodiment improves on the traditional text recall and expands the semantic space of user input through a large model, such as Figure 5As shown in the figure, it is a flowchart for knowledge recall. Several similar questions are generated based on the original user input, and then the rewritten questions are retrieved and recalled separately. The RRF algorithm is used to rearrange the multi-way recalled documents, filter out documents that are not highly relevant to the questions, and improve the quality of the recalled documents.

[0089] The answer generation unit 706 is used to generate answers according to the results of knowledge recall.

[0090] Although documents with low relevance to user input have been filtered in the knowledge recall process in the above steps, there are still some relevant documents with low relevance or that do not contain answers to questions, which will have a negative effect on the generation of the large model. Therefore, this embodiment improves the problem that the large model is easily disturbed by irrelevant text during the generation process, and improves the quality of the large model's responses to user questions.

[0091] Among them, since the large model with hundreds of billions of parameters has super text understanding and generation capabilities, it can filter out the influence of irrelevant text to a large extent and generate the final answer. Therefore, this embodiment first merges the relevant text and the irrelevant text in a certain proportion, generates an answer with a thinking chain through the large model with hundreds of billions of parameters, constructs a data set to fine-tune the model that injects knowledge in the government affairs field, improves the model's ability to filter irrelevant text in the answering process, and generates more useful answers, that is, completes the answer generation.

[0092] The synthesis unit 707 is used to perform speech synthesis on the generated answer.

[0093] This embodiment adopts the KAN-TTS solution, which consists of a text front-end, an acoustic model, and a vocoder. The front-end module uses a model combined with rules to flexibly process texts in various scenarios, while the back-end module uses SAM-BERT+HIFIGAN to provide highly expressive streaming synthesis effects. Based on KAN-TTS, in order to be applied to intelligent customer service question-and-answer scenarios, a new timestamp function is added, word-level timestamps are provided, openai's whisper is used for timestamp prediction, and a sentence-level streaming TTS interface is added to improve real-time performance.

[0094] like Figure 6 The flowchart of speech synthesis provided by this embodiment is shown in FIG. Specifically, starting from text input, linguistic features are generated through text normalization, word segmentation, disambiguation, and polyphone processing. Then, acoustic features are generated using a duration model and an acoustic model, and finally speech is synthesized through a vocoder.

[0095] This application has the following beneficial effects:

[0096] (1) This application uses artificial intelligence technologies such as natural language processing, speech recognition and synthesis to build an answering system that can meet complex scenarios, has good interaction and high response quality, improves people's satisfaction with government services, and reduces the workload of business personnel.

[0097] (2) The complete intelligent reply process established by this application, from speech recognition, verticalization of large language model fields, knowledge base construction, intent recognition, knowledge recall, answer generation, and speech synthesis, greatly compensates for the shortcomings of low answer quality and low real-time performance in traditional replies.

[0098] Although the present application is described with reference to examples, this is for illustrative purposes only and is not intended to limit the present application, and changes, additions and / or deletions to the embodiments may be made without departing from the scope of the present application.

[0099] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A question-answering method, characterized in that: The following steps are involved: Obtain user input and identify the user input; After completing the recognition of user input, the vertical model is trained; After completing the model training, the knowledge base is constructed; After completing the construction of the knowledge base, perform intent recognition; After intent recognition, semantic expansion of user input is performed to recall knowledge; Generate answers based on the results of knowledge recall; Perform speech synthesis on the generated answers.

2. The question-answering method according to claim 1, wherein: User input recognition is performed through real-time streaming recognition models, offline hot word models, and punctuation prediction models.

3. The question-answering method according to claim 1, wherein: The process of user identification also includes correcting the user input.

4. The question-answering method according to claim 1, wherein: Training a vertical model includes: building a data set in a specified field, and using the data set to perform supervised instruction fine-tuning on the vertical model.

5. The question-answering method according to claim 1, wherein: The construction of the knowledge base includes the following sub-steps: Perform sub-database storage to complete fine-grained knowledge division; Perform unsupervised vector domainization and complete the domainization of the vector model.

6. A question-answering system, characterized in that: Specifically include: Recognition unit, training unit, construction unit, intention recognition unit, knowledge recall unit, answer generation unit, and synthesis unit; The recognition unit is used to obtain user input and recognize the user input; A training unit, used to train the vertical model after completing the recognition of the user input; Construction unit, used to construct the knowledge base; An intention recognition unit, used for performing intention recognition; A knowledge recall unit, used to perform semantic expansion on user input and perform knowledge recall; An answer generation unit, used for generating answers according to the result of knowledge recall; The synthesis unit is used to perform speech synthesis on the generated answers.

7. The question-answering system according to claim 6, wherein: The recognition unit recognizes user input through a real-time streaming recognition model, an offline hot word model, and a punctuation prediction model.

8. The question-answering system according to claim 6, wherein: The process of the recognition unit recognizing the user also includes correcting the user input.

9. The question-answering system according to claim 6, wherein: The training unit performs training of the vertical model including: constructing a data set in a specified field, and using the data set to perform supervised instruction fine-tuning on the vertical model.

10. The question-answering system according to claim 6, wherein: The construction unit constructs the knowledge base including the following sub-steps: Perform sub-database storage to complete fine-grained knowledge division; Perform unsupervised vector domainization and complete the domainization of the vector model.

Images