Intelligent question answering method and device, equipment, storage medium and program product

Abstract

The present disclosure provides a kind of intelligent question answering method, device, equipment, storage medium and program product, it is related to big language model, artificial intelligence technical field such as intelligent question answering.The method comprises: the intent recognition of the question information of user is carried out in combination with the special dictionary of target field, and the intent recognition result is obtained;Recall the candidate answer information set matched with question information in the pre-set question and answer knowledge base based on the intent recognition result;Determine the target answer information corresponding to the question information based on the similarity of candidate answer information in candidate answer information set and question information.The present embodiment discloses that, in target field, the intent recognition of user is carried out in combination with natural language understanding algorithm, by understanding user intent, finding corresponding knowledge base and recalling corresponding target answer information, it can be more accurately generated answer information for the question information of user, improve the adaptability of intelligent question answering in specific field, and can be used for timely interaction with user.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, specifically to the field of artificial intelligence technology such as large language models and intelligent question answering, and particularly to an intelligent question answering method, device, equipment, storage medium, and program product. Background Technology

[0002] With the development of artificial intelligence (AI) technology, various parties have begun to adopt intelligent AI systems to replace human operation and maintenance. These systems understand user questions, find answer templates in the knowledge base, and return them to the user to answer their questions. However, the question-answering capabilities of these intelligent AI systems are more suitable for general scenarios. In some specific professional scenarios, these intelligent AI systems still have many problems, which trouble relevant practitioners. Summary of the Invention

[0003] This disclosure provides an intelligent question-answering method, apparatus, device, storage medium, and program product.

[0004] In a first aspect, embodiments of this disclosure propose an intelligent question-answering method, comprising: performing intent recognition on user question information by combining a proprietary dictionary of the target domain to obtain intent recognition results; recalling a set of candidate answer information that matches the question information in a preset question-answering knowledge base based on the intent recognition results; and determining the target answer information corresponding to the question information based on the similarity between the candidate answer information in the candidate answer information set and the question information.

[0005] Secondly, embodiments of this disclosure propose an intelligent question-answering device, comprising: an intent recognition module configured to perform intent recognition on a user's question information by combining a proprietary dictionary of the target domain, thereby obtaining an intent recognition result; a candidate answer generation module configured to recall a set of candidate answer information that matches the question information from a preset question-answering knowledge base based on the intent recognition result; and a target answer generation module configured to determine the target answer information corresponding to the question information based on the similarity between the candidate answer information in the candidate answer information set and the question information.

[0006] Thirdly, embodiments of this disclosure provide an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to implement the intelligent question-answering method as described in any implementation of the first aspect.

[0007] Fourthly, embodiments of this disclosure provide a non-transitory computer-readable storage medium storing computer instructions that enable a computer to implement the intelligent question-answering method as described in any implementation of the first aspect.

[0008] Fifthly, embodiments of this disclosure provide a computer program product including a computer program that, when executed by a processor, can implement the intelligent question-answering method as described in any implementation of the first aspect.

[0009] The intelligent question-answering method, apparatus, device, storage medium, and program product provided in this disclosure, in a specific target domain, combine natural language understanding algorithms to identify user intent. By understanding user intent and searching the corresponding knowledge base to recall the corresponding target answer information, it can more accurately generate answer information for user questions, improve the adaptability of intelligent question answering in a specific domain, and can be used for timely interaction with users.

[0010] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0011] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0012] Figure 1 This is an exemplary system architecture to which this disclosure can be applied;

[0013] Figure 2 A flowchart of an intelligent question-answering method provided in this embodiment of the disclosure;

[0014] Figure 3A A flowchart illustrating the process of performing intent recognition on user query information by combining a proprietary dictionary of the target domain to obtain intent recognition results, as provided in this embodiment of the disclosure;

[0015] Figure 3B A schematic diagram illustrating the principle of using a user intent recognition model combined with a domain-specific dictionary to recognize user intent in an embodiment of this disclosure;

[0016] Figure 4 A flowchart illustrating the process of recalling a set of candidate answer information that matches the question information in a preset question-answering knowledge base based on intent recognition results, as provided in this embodiment of the disclosure;

[0017] Figure 5AA flowchart illustrating the process of using vector recall to retrieve a second set of answer information that matches the question information from a preset question-and-answer knowledge base, as provided in this embodiment of the disclosure;

[0018] Figure 5B A schematic diagram illustrating the principle of using a recall model for recall provided in this embodiment of the disclosure;

[0019] Figure 6A A flowchart illustrating the process of determining the target answer information corresponding to the question information based on the similarity between candidate answer information and question information in a candidate answer information set, as provided in this embodiment of the disclosure;

[0020] Figure 6B A schematic diagram illustrating the principle of fine-ranking using a fine-ranking model, provided for embodiments of this disclosure;

[0021] Figure 7 A flowchart of another intelligent question-answering method provided in this disclosure embodiment;

[0022] Figure 8 A flowchart illustrating yet another intelligent question-answering method provided in this disclosure embodiment;

[0023] Figure 9A A schematic diagram illustrating an intelligent question-answering method in an application scenario provided by an embodiment of this disclosure;

[0024] Figure 9B A flowchart illustrating an intelligent question-answering method in an application scenario provided by an embodiment of this disclosure;

[0025] Figure 10 A structural block diagram of an intelligent question-answering device provided in an embodiment of this disclosure;

[0026] Figure 11 This is a schematic diagram of the structure of an electronic device suitable for performing an intelligent question-answering method, provided as an embodiment of this disclosure. Detailed Implementation

[0027] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding; these should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description. It should be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0028] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0029] Figure 1 An exemplary system architecture 100 is shown, in which embodiments of the intelligent question-answering methods, apparatuses, electronic devices, and computer-readable storage media of this disclosure can be applied.

[0030] like Figure 1 As shown, system architecture 100 may include terminal devices 101, 102, and 103, a network 104, and a server 105. Network 104 serves as the medium for providing communication links between terminal devices 101, 102, and 103 and server 105. Network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0031] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various applications for enabling information communication between the terminal devices 101, 102, and 103 and server 105 can be installed. These applications include instant messaging applications.

[0032] Terminal devices 101, 102, and 103 and server 105 can be either hardware or software. When terminal devices 101, 102, and 103 are hardware, they can be various electronic devices with displays, including but not limited to smartphones, tablets, laptops, and desktop computers. When terminal devices 101, 102, and 103 are software, they can be installed in the aforementioned electronic devices, and can be implemented as multiple software programs or software modules, or as a single software program or software module; no specific limitation is made here. When server 105 is hardware, it can be implemented as a distributed server cluster composed of multiple servers, or as a single server. When server 105 is software, it can be implemented as multiple software programs or software modules, or as a single software program or software module; no specific limitation is made here.

[0033] Server 105 can provide various services through its built-in applications. It should be noted that services such as instant messaging and intelligent question answering can be obtained from terminal devices 101, 102, and 103 via network 104, or they can be pre-stored locally on server 105 through various means. Therefore, when server 105 detects that this data is already stored locally, it can choose to retrieve this data directly from the local storage. In this case, the exemplary system architecture 100 may not include terminal devices 101, 102, and 103 and network 104.

[0034] Since intelligent question answering may require significant computing resources and capabilities, the intelligent question answering methods provided in the subsequent embodiments of this disclosure are generally executed by a server 105 with strong computing power and abundant computing resources. Correspondingly, the intelligent question answering device is also generally located within the server 105. However, it should also be noted that when terminal devices 101, 102, and 103 also possess sufficient computing power and resources, they can also perform the aforementioned calculations performed by the server 105 through their installed instant messaging applications, thereby outputting the same results as the server 105. Especially when multiple terminal devices with different computing capabilities exist simultaneously, if the instant messaging application determines that its terminal device has strong computing power and abundant remaining computing resources, it can allow the terminal device to perform the aforementioned calculations, thereby appropriately reducing the computing pressure on the server 105. Accordingly, the intelligent question answering device can also be located within terminal devices 101, 102, and 103. In this case, the exemplary system architecture 100 may also exclude the server 105 and the network 104.

[0035] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0036] Please refer to Figure 2 , Figure 2 A flowchart of an intelligent question-answering method provided in this disclosure embodiment, wherein process 200 includes the following steps:

[0037] Step 201: Combine the user's question information with a domain-specific dictionary to perform intent recognition and obtain intent recognition results.

[0038] This step is intended for the implementer of the intelligent question-answering method (e.g., Figure 1The server 105 shown obtains the user's question information and, in conjunction with a domain-specific dictionary, performs intent recognition on the question information to obtain the user's intent recognition result. In this embodiment, the user's question information refers to information initiated by the user for consultation, inquiry, or request. This question information can be text information entered by the user through typing or touchscreen handwriting on a terminal device, voice information entered by the user through an audio device such as a microphone, or screenshots uploaded by the user after taking a screenshot using some functional software. This embodiment is not limited to these. Correspondingly, if the question information is text information, the executing entity can directly perform intent recognition based on the text form of the question information; if the question information is voice information, the executing entity can apply speech recognition technology, which is well-known in the field of natural language processing, to identify the corresponding text information from the voice information, and then perform intent recognition based on the identified text information; if the question information is a screenshot image, the executing entity can apply technologies such as Optical Character Recognition (OCR) to identify the corresponding text information from the screenshot image, and then perform intent recognition based on the identified text information.

[0039] Furthermore, in this embodiment, the executing entity combines a domain-specific dictionary with the user's query information to perform intent recognition. The main reason for this approach is that when using traditional intent recognition techniques to analyze and process user intent, inaccurate intent recognition may occur when the words or phrases in the user's query information involve specific professional domain terminology. Therefore, in this embodiment, the executing entity can combine a domain-specific dictionary with intent recognition to improve the accuracy of user intent recognition for that target domain. The target domain refers to various professional fields with proprietary vocabulary, such as logistics, e-commerce, finance, smart hardware, cloud computing, etc., and this embodiment is not limited to these. A proprietary dictionary is a database specifically used to store professional vocabulary and terminology for a corresponding field. A proprietary dictionary can provide detailed explanations and definitions of vocabulary within a specific field or topic, helping users quickly and accurately understand and use these words.

[0040] Step 202: Based on the intent recognition results, recall a set of candidate answer information that matches the question information from the preset question-answering knowledge base.

[0041] This step aims to have the aforementioned executing entity recall candidate answer information matching the question information from a pre-defined knowledge base based on the intent recognition results, and combine them into a candidate answer information set. A knowledge base is a resource in Natural Language Processing (NLP) used to store and represent structured or semi-structured knowledge information. The knowledge base provides rich background knowledge for NLP tasks, helping the system to more accurately understand the meaning and context of the input text. By querying and utilizing information in the knowledge base, NLP systems can perform complex reasoning and decision-making, thereby generating more reasonable and accurate outputs. Information in the knowledge base can help NLP models better adapt to new domains and tasks, improving the generalization ability of NLP models.

[0042] Recall is an important concept in information retrieval, referring to identifying the most relevant parts to the user's needs from a large candidate set. In this embodiment, recall refers to finding a set of candidate answers from the knowledge base that match the user's question.

[0043] Step 203: Determine the target answer information corresponding to the question information based on the similarity between the candidate answer information in the candidate answer information set and the question information.

[0044] This step aims to have the executing entity calculate the similarity between the candidate answer information in the candidate answer information set and the question information, and determine the target answer information corresponding to the question information based on the similarity. For example, the executing entity may select the candidate answer information with the highest similarity as the target answer information corresponding to the question information. In practical applications, the target answer information can also be selected based on the similarity as needed, and this embodiment is not limited thereto.

[0045] The intelligent question-answering method provided in this disclosure combines natural language understanding algorithms to identify user intent in a specific target domain. By understanding user intent and searching the corresponding knowledge base to retrieve relevant target answer information, it can more accurately generate answer information for user questions, improve the adaptability of intelligent question answering in specific domains, and can be used for timely interaction with users.

[0046] In some optional implementations of this embodiment, please refer to Figure 3A , Figure 3A The flowchart provided in this disclosure illustrates the process of combining a proprietary dictionary of the target domain to perform intent recognition on user query information and obtain intent recognition results. Figure 2Step 201 in the illustrated process 200 provides a specific implementation. Other steps in process 200 remain unchanged, and a new complete embodiment is obtained by replacing step 201 with the specific implementation provided in this embodiment. Step 201, which combines a domain-specific dictionary to perform intent recognition on the user's query information and obtain the intent recognition result, mainly includes:

[0047] Step 301: Perform word segmentation on the query information using a proprietary dictionary to obtain multiple first word units. The executing entity uses a word segmentation algorithm to segment the query information to obtain multiple first word units. In this embodiment, word segmentation is performed using a proprietary dictionary, changing the original word segmentation method that only decomposes each single Chinese character into a single word unit, and incorporating proprietary words from the target domain's proprietary dictionary. For example, the word "XY" is a word composed of two single characters X and Y, but this word is not a general word, but is used as a word in a specific domain. For this word "XY", a general word segmentation algorithm would decompose "XY" into two word units, "X" and "Y". However, in this embodiment, by combining the proprietary dictionary of this specific domain, "XY" is set as a single word unit, thereby improving the accuracy and adaptability of word segmentation within this specific domain.

[0048] Step 302: Generate an initial vector representation for each first word based on its index number. The executing entity can obtain the index number of each first word by searching the vocabulary, and then obtain the initial vector representation of each word using that index number. The vocabulary is used as a lexicon in natural language processing to help machines understand and process human language information. Furthermore, a unique index number can be assigned to each word in the vocabulary to improve retrieval efficiency.

[0049] Step 303: Calculate the final vector representation of the first word element based on the initial vector representation. The executing entity can calculate the final vector representation of the first word element based on the initial vector representation of each word element.

[0050] Step 304: Perform linear calculation based on the final vector representation of the first word element to obtain the target category information of the query information. The executing entity can perform linear calculation based on the final vector representation of the first word element to obtain the target category information of the query information.

[0051] Step 305: Obtain the intent recognition result based on the target category information. After determining the target category information of the question, the executing entity can obtain the intent recognition result of the question based on the target category information.

[0052] In practical implementation, at least one step in the above execution process, steps 302-305, can be executed by a pre-trained user intent recognition model. In this embodiment, the loss function used during the training of the user intent recognition model includes loss weights corresponding to sample category information. This loss function is shown in formula (1).

[0053]

[0054] Where n represents the number of samples, m represents the number of categories, and w c It is the weight of sample class c, y ij z is the true label of the i-th sample belonging to the j-th class. ij It is the numerical value of the i-th sample belonging to the j-th class in the model output. For each element z of the input vector ik Perform exponentiation Then, sum the results of all the exponents and normalize them.

[0055] The loss function first considers each element z ij This is transformed into a probability value between 0 and 1, where the sum of the probabilities of all elements is 1. This represents the probability distribution of the input vector belonging to each class. Then, the true label y is calculated. ij The difference between the original cross-entropy loss function and the probability distribution is considered. A smaller difference indicates a more accurate prediction. In this embodiment, the loss function adds weights to the loss for each category on top of the original cross-entropy loss. This is because in practical applications, user comments are diverse, and the distribution of each category is uneven. By adjusting the category weights of the cross-entropy loss function, corresponding loss weights are assigned to each sample category, effectively mitigating the impact of class imbalance in the sample dataset on the output results.

[0056] For example, an initial vector representation of each first word can be generated based on the index number of each first word using a pre-trained intent recognition model. Correspondingly, the current calculation process based on the final vector representation of the first word implemented in step 304 can be implemented by adding a linear layer to the output side of the intent recognition model.

[0057] Please refer to Figure 3B , Figure 3B This is a schematic diagram illustrating the principle of using a user intent recognition model combined with a domain-specific dictionary to recognize user intent in an embodiment of this disclosure. In this example, the target domain is logistics. The user's question is: "I want to send my package via XY Express, are there any coupons?" (XY can refer to the name of the logistics company providing the express service, or the type of express delivery, etc.).

[0058] Step 1: After obtaining the query information, the executing entity performs word segmentation on the query information using a specialized dictionary in the logistics field. The query "I want to send XY Express, are there any coupons?" is broken down into the word units "I", "want", "send", "XY", "express", ",", "have", "discount", "coupon", "what", and "?". Here we can see that, because a specialized dictionary in the logistics field is used, "XY" is not split into two word units, but rather treated as a single word unit. This processing result is clearly more consistent with the actual situation of the scenario and can provide a more accurate word unit foundation for subsequent intent understanding.

[0059] Step 2: For each segmented word, an initial vector representation is generated based on its index sequence number in the vocabulary. In this example, the vector is mapped to the corresponding index sequence number in the vocabulary to generate a 512-dimensional vector. Furthermore, for each word, a high-dimensional (768-dimensional) initial vector representation is generated.

[0060] Step 3: Calculate the final vector representation of the first word based on the initial vector representation. This is achieved through multi-layer hidden layer computation (transformer structure) in the intermediate process of the user intent recognition model, yielding the final vector representation of the first word.

[0061] Step 4: Through a linear layer, linear calculations are performed based on the vector result of the first word element to achieve dimensionality reduction. This yields classification results corresponding to three categories (service + consultation, express delivery consultation, and other questions), thus obtaining the intent recognition result corresponding to the user's question. In this example, the most likely intent recognition result is that the question belongs to the "express delivery consultation" category.

[0062] Through the above process, professional terms are added to the vocabulary as new lexical units to ensure that professional terms can be used as complete lexical units as input to the user intent recognition model during the word segmentation process. This can help the user intent recognition model encode and understand professional terms.

[0063] In some optional embodiments of the present disclosure, please refer to Figure 4 , Figure 4 The flowchart illustrates the process by which, based on intent recognition results, a set of candidate answer information matching the question information is retrieved from a preset question-answering knowledge base, as provided in this embodiment of the disclosure. Figure 2Step 202 in the illustrated process 200 provides a specific implementation method. Other steps in process 200 are not adjusted; a new complete embodiment is obtained by replacing step 202 with the specific implementation method provided in this embodiment. Step 202, the process of recalling a set of candidate answer information matching the question information in a preset question-answering knowledge base based on the intent recognition result, mainly includes:

[0064] Step 401: Use inverted index recall to retrieve the first set of answer information that matches the question information from the preset question-answer knowledge base.

[0065] Step 402: Use vector recall to retrieve a set of second answer information that matches the question information from a pre-defined question-and-answer knowledge base.

[0066] Step 403: Use rule-based recall to retrieve a set of third-party answer information that matches the question information from a pre-defined question-and-answer knowledge base.

[0067] Step 404: Integrate the first answer information set, the second answer information set, and the third answer information set to obtain the candidate answer information set.

[0068] In this embodiment, a multi-path recall mode is adopted in the recall stage, mainly including inverted index recall, vector recall, and rule-based recall. The input of the recall stage is the user's question text, and the output is a set of candidate templates. Since the dataset to be filtered in the recall stage is large (the entire knowledge base), a fast computation speed is required, but the accuracy does not need to be too high.

[0069] Among them, inverted recall: The calculation method for inverted recall is shown in formula (2):

[0070]

[0071] The IDF (Inverse Document Frequency) is calculated as shown in formula (3), where n(q) i ) is a term containing q i Number of documents:

[0072]

[0073] This retrieval method is widely used in document retrieval systems. The score is obtained by multiplying two parts: TF (term frequency) and IDF (inverse document frequency), where f(q) = f(t). i D) represents term q iThe frequency of occurrence in document D, |D| is the length of document D, avgdl is the average length of all documents, k1 and b are constant parameters, here k1 is set to 1.2 and b to 0.75, and N is the total number of documents in the document set. The method used in this embodiment is a general algorithm, and the principle will not be elaborated here. In addition, before calculating the BM25 score, words without practical meaning in the stop word list (such as interjections, modal particles, etc.) can be filtered out to avoid interfering with the calculation of the BM25 score.

[0074] Rule-based recall: Rule-based recall is mainly adjusted in real time based on business logic, without setting fixed algorithm logic. For example, it can be based on hot words or marketing terms at a certain moment. As long as the term is matched, it will be recalled.

[0075] Vector Recall: Vector recall involves calculating the similarity between the user's question text and all answer templates in the knowledge base to obtain the template with the highest similarity to the user's question text, and then recalling that template. In this embodiment, cosine similarity can be used to calculate the similarity, and the formula for calculating cosine similarity is shown in formula (4):

[0076]

[0077] Here, a and b represent the two vectors for which similarity is to be calculated. In this embodiment, a and b represent the vector representations of the user's question and the answer template in the knowledge base, respectively.

[0078] It should be understood that the cosine similarity is only an example to illustrate the method of calculating similarity. In practical applications, other methods such as Euclidean distance can also be used to calculate similarity as needed. This embodiment is not limited to this.

[0079] After obtaining the corresponding recall results through three different recall modes, they are integrated to obtain the candidate answer information set. This integration process can be set according to actual needs, such as taking the union or intersection of the three results, etc., and this embodiment is not limited to this.

[0080] In practical implementation, at least one step in the above execution process, steps 401-404, can be executed using a pre-trained text similarity matching model. In this embodiment, the loss function used in the training process of the text similarity matching model is shown in formula (5):

[0081]

[0082] Where N represents the number of samples, q represents the user's question text, and d + Represents a relevant answer template, d - Represents an unrelated answer template, where ∈ is a small constant (ensuring the initial state d).+ (Closer to q). The meaning of this loss function can be interpreted as follows: through model training, the user's question text q is made to correlate with the positively related template d. + The correlation between the user's inquiry text q and the negatively correlated template d is increasing. - The correlation between them is decreasing.

[0083] In some optional embodiments of this disclosure, such as Figure 5A and Figure 5B As shown, step 402 above, the process of using vector recall to retrieve a second set of answer information that matches the question information from a preset question-and-answer knowledge base, mainly includes:

[0084] Step 501: Combine a proprietary dictionary to perform word segmentation on the answer templates in the preset question-and-answer knowledge base to obtain multiple second word elements.

[0085] Step 502: Vectorize the multiple first word elements to obtain multiple first word element vectors. In this embodiment, vectorizing the multiple first word elements can be achieved through a text vectorization model to obtain the multiple first word element vectors.

[0086] Step 503: Vectorize the multiple second word elements to obtain multiple second word element vectors. In this embodiment, vectorizing the multiple second word elements can be achieved through a text vectorization model to obtain the multiple second word element vectors.

[0087] Step 504: Calculate the similarity between each first word vector and the multiple second word vectors. For example, the executing entity can use similarity calculation methods such as cosine similarity or Euclidean distance to calculate the similarity between each first word vector and the multiple second word vectors.

[0088] Step 505: Obtain a second answer information set matching the question information based on the calculated similarity results. After calculating the similarity between each first word vector and multiple second word vectors, the executing entity can obtain a second answer information set matching the question information based on the calculated similarity results. For example, for the user's question "I want to send my package via XY Express, are there any coupons?", after the above process, the retrieved second answer information is "XY Express welfare group, welfare coupons from 75% off: Click to claim coupon, place an order now, fill in the information, and the courier will pick up the package!"

[0089] In practical applications, the second answer information set obtained through vector recall can be a collection containing multiple second answer information sets. That is, it can be sorted according to the similarity results, and the second answer information sets of the top preset number (e.g., 100, 1000, or 10000) can be recalled to form the second answer information set.

[0090] In this embodiment, the above process can also be called the recall phase. The model in the recall phase has a faster computation speed but lower accuracy. The model structure built for recall is as follows: Figure 5B As shown, during the recall phase, the user question text and the knowledge base answer template text of the model are calculated into vectors separately, and then crossed near the output. This can significantly reduce the number of calculation parameters, but will result in some performance loss.

[0091] In some optional embodiments of the present disclosure, please refer to Figure 6A , Figure 6A This is a flowchart illustrating the process of determining the target answer information corresponding to the question information based on the similarity between candidate answer information and question information in the candidate answer information set, as provided in this embodiment of the disclosure. Figure 2 Step 203 in the illustrated process 200 provides a specific implementation. Other steps in process 200 are not adjusted; a new complete embodiment is obtained by replacing step 203 with the specific implementation provided in this embodiment. Step 203, the process of determining the target answer information corresponding to the question information based on the similarity between candidate answer information in the candidate answer information set and the question information, mainly includes:

[0092] Step 601: Combine a proprietary dictionary to perform word segmentation on the candidate answer information in the candidate answer information set to obtain multiple third word elements.

[0093] Step 602: Vectorize the multiple first word elements to obtain multiple first word element vectors. In this embodiment, vectorizing the multiple first word elements can be achieved through a text vectorization model to obtain the multiple first word element vectors.

[0094] Step 603: Vectorize the multiple third-word units to obtain multiple third-word unit vectors. In this embodiment, vectorizing the multiple third-word units can be achieved through a text vectorization model to obtain the multiple third-word unit vectors.

[0095] Step 604: Perform vector cross operation on multiple first word vectors and multiple third word vectors, so that each word vector receives information from other words during the calculation in the input stage.

[0096] Step 605: Calculate the similarity between each first word vector and the multiple third word vectors based on the vector cross-operation results. For example, the executing entity can use similarity calculation methods such as cosine similarity or Euclidean distance to calculate the similarity between each first word vector and the multiple third word vectors.

[0097] Step 606: Sort the candidate answer information based on the calculated similarity results to obtain the sorting results. After calculating the similarity between each first word vector and multiple third word vectors, the executing entity can sort them according to the calculated similarity results to obtain the sorting results.

[0098] Step 607: Determine the target answer information for the corresponding question based on the ranking results. Based on the ranking results, the executing entity can further sort and filter the recall results obtained from steps 401-404 to obtain more accurate output results. For example, for the user's question "I want to send my package via XY Express, are there any coupons?", after the above process, the second recalled answer information is "XY Express welfare group, welfare coupons from 75% off: Click to claim coupon, place an order now, fill in the information, and the courier will pick up the package!"

[0099] In this embodiment, the above process can also be called the fine-ranking stage. The purpose of the fine-ranking stage is to reorder the candidate answer template set obtained in the recall stage. Since the number of candidate samples in this stage is much smaller than that in the recall stage, but higher accuracy is required, a more complex model computation structure is needed. That is to say, the model in the recall stage has a faster computation speed but lower accuracy; the model in the fine-ranking stage has a larger computational load and a relatively slower computation speed, but higher accuracy. The model structure built for fine-ranking is as follows: Figure 6B As shown, the overall structure is similar to that of a recall (e.g.) Figure 5B As shown), the underlying model structure is different. In the fine-ranking stage, after acquiring the word units, the model begins vector cross-combining, meaning each word vector receives information from other word units during the input stage computation. This leads to a large amount of additional parameter calculation, but the results are better; while in the recall stage (as shown...), the underlying model... Figure 5B The user question text and the knowledge base answer template text (as shown) will be vectorized separately, and finally crossed near the output end. This can greatly reduce the calculation parameters, but will lose some performance. The loss function in the fine ranking stage is the same as shown in formula (4).

[0100] In this embodiment, through the above process, in the process of matching and filtering answers based on the user's intent recognition results, the accuracy of similarity matching is ensured through two stages: recall and fine ranking. The model in the recall stage and the model in the fine ranking stage combine the advantages of the intent understanding model with the proprietary dictionary unique to the target domain for fine-tuning, and the model structure is modified according to the task type. That is, the stage where word vectors intersect is different in the recall and fine ranking stages, so as to improve the accuracy and precision of the model's output results.

[0101] Please refer to Figure 7 , Figure 7A flowchart of another intelligent question-answering method provided in this disclosure embodiment, wherein process 700 includes the following steps:

[0102] Step 701: Obtain initial information input by the user. In this embodiment, the initial information of the user refers to information initiated by the user for consultation, inquiry, or request, etc. This initial information may be text information input by the user through typing or touch screen handwriting on a terminal device, or voice information input by the user through audio devices such as microphones, or screenshots uploaded by the user after taking screenshots through some functional software, etc. This embodiment is not limited to these.

[0103] Step 702: In response to the initial information containing sensitive words, the initial information is intercepted, and / or a reminder message is sent to the user. In response to the initial information not containing sensitive words, the initial information is determined to be a question message. Sensitive words generally refer to words that have specific meanings in a certain context and may lead to discomfort or adverse consequences. These words may be considered sensitive or inappropriate, or even violate public order and good morals, due to factors such as culture, politics, religion, morality, or social norms. Therefore, in this embodiment, the executing entity filters the initial information input by the user for sensitive words, determines whether the initial information contains sensitive words, and if the initial information contains sensitive words, it needs to be intercepted; or a reminder message is sent to the user to warn the user that the sent information should not contain sensitive words; or the initial information is intercepted and a reminder message is sent to the user. If the initial information does not contain sensitive words, the initial information is determined to be a question message, and the subsequent process is executed.

[0104] Optionally, in addition to intercepting the user's initial information, the executing entity may also process the initial information to ensure that other compliant information sent by the user can be transmitted. For example, content involving sensitive words in the initial information may be replaced with an asterisk "*" or other symbols, while other content that does not involve sensitive words may be displayed. This embodiment is not limited to this.

[0105] By filtering and selecting user-input information through the above process, the legality, compliance, and security of information transmission can be guaranteed.

[0106] Step 703: Combine the user's question information with a domain-specific dictionary to perform intent recognition and obtain the intent recognition result.

[0107] Step 704: Based on the intent recognition results, recall a set of candidate answer information that matches the question information from the preset question-answering knowledge base.

[0108] Step 705: Determine the target answer information corresponding to the question information based on the similarity between the candidate answer information in the candidate answer information set and the question information.

[0109] The above steps 703-705 and as follows Figure 2 The steps 201-203 shown are the same. For the same parts, please refer to the corresponding parts of the previous embodiment. They will not be repeated here.

[0110] Please refer to Figure 8 , Figure 8 A flowchart of another intelligent question-answering method provided in this disclosure embodiment, wherein process 800 includes the following steps:

[0111] Step 801: Combine the user's question information with a domain-specific dictionary to perform intent recognition and obtain the intent recognition result.

[0112] Step 802: Based on the intent recognition results, recall a set of candidate answer information that matches the question information from the preset question-answering knowledge base.

[0113] Step 803: Determine the target answer information corresponding to the question information based on the similarity between the candidate answer information in the candidate answer information set and the question information.

[0114] The above steps 801-803 and as follows Figure 2 The steps 201-203 shown are the same. For the same parts, please refer to the corresponding parts of the previous embodiment. They will not be repeated here.

[0115] Step 804: Obtain pre-configured prompt words, which are used to represent the role information and rule information of the large language model.

[0116] In this embodiment, after determining the target answer information corresponding to the user's question, the executing entity can further optimize the target answer information by calling a large language model. The main process includes: pre-configuring prompt words for the large language model, which are used to represent the role information and rule information of the large language model. For example:

[0117] (Please role-play as a customer service representative for XY Express. Your replies should be as brief as possible, but polite and courteous, with a cute and refreshing tone, and should follow these rules:)

[0118] 1. If the user does not express an intention to ask a question, or lacks contextual information, do not give a direct answer; simply engage in friendly interaction.

[0119] 2. You do not have customer service authority and cannot directly assist customer service with any procedures.

[0120] 3. You can only answer general questions, not personalized questions from users. You can only give vague answers or direct users to consult the 'XY Express' mini-program.

[0121] 4. If you lack the relevant knowledge, please guide users to consult XY Express customer service or contact the corresponding deliveryman for inquiries. For questions about claiming coupons, please guide users to claim coupons through the 'XY Deliveryman' mini-program.

[0122] 5. If a user sends a welcome message, guide them to the 'XY Xiaoge' mini-program to claim a coupon.

[0123] By setting the above prompts, the large language model can be guided to generate text output that meets expectations.

[0124] Step 805: Adjust the target answer information based on the prompt words using a large language model. Based on the configured prompt words, the large language model can adjust the target answer information to better align the output with the rules and characteristics of the roles represented by the prompt words.

[0125] Step 806: Feedback the adjusted target answer information to the user. The adjusted target answer information is more suitable for the required scenario and role. For example, as mentioned above, as a courier customer service representative, in response to a user's question, "I want to send my package via XY Express, are there any coupons?", after the above process, the adjusted target answer information would be: "XY Express benefits group, coupons starting at 75% off: Click to claim your coupon, place your order now, fill in your information, and the courier will pick up your package!" In terms of content and tone, this is more in line with the characteristics of a courier customer service representative, making the reply more natural and the language style more vivid.

[0126] In the embodiments of this disclosure, the intelligent question-answering method described in any of the above embodiments or their implementations is mainly implemented through an execution subject, and further combined with a large language model. However, in practical applications, problems or scenarios that the execution subject or the large language model cannot handle may be encountered. To avoid this situation, in some optional embodiments of this disclosure, the intelligent question-answering method may further include: adding auxiliary service link information to the target answer information and feeding it back to the user. The auxiliary service link information is used to redirect to a preset consultation link. The preset consultation link may be, for example, a consultation webpage or consultation interface corresponding to human customer service, through which human customer service provides further answers to the user's questions.

[0127] To enhance understanding, this disclosure also provides a specific implementation scheme based on a particular application scenario, as shown in process 900 in Figure 9. Process 900 mainly includes:

[0128] Step 901: User 91 obtains the user's original message 93 (e.g., "I want to send XY Express, are there any coupons?") through the relevant application (e.g., instant messaging application) on the terminal device 92.

[0129] Step 902: Server 94 obtains the original message 93, determines whether the original message is a text type message. If it is text type data, it directly proceeds to the next process; if it is not text type data, it converts it to text type data and proceeds to the next process.

[0130] Step 903: Server 94 determines whether the text content of the original message contains prohibited words and / or sensitive words. If the text content contains sensitive information involving illegality, violation, politics, war, etc., the message is blocked or the user who spoke is warned; if it does not contain sensitive words, the process proceeds to the next step.

[0131] Step 904: Classify the text content of the user's original message using the "User Intent Recognition Model" to determine that the user's intention is to provide express delivery information 95, and then pass the text content of the user's original message to the next process.

[0132] Step 905: Using the "text similarity matching model," retrieve the answer to the text content of the user's original message from the question-and-answer knowledge base. If the answer exists, return the template for the corresponding question answer; if the answer does not exist, return a rejection template, such as guiding the user to consult the relevant staff. Then, pass the returned template to the next process.

[0133] Step 906: Use the large model in conjunction with the user's original message to adjust the response template returned in the previous stage. This is done because although the template contains answers to the user's questions, the language may be stiff and rigid, which greatly affects the user experience. Therefore, the large model is used to optimize the language style.

[0134] Step 907: After obtaining the user's reply content 96 (e.g., "XY Express Welfare Group, coupons up to 75% off: Click to claim coupon, place order now, fill in the information, and the courier will pick up your package!"), depending on the specific situation, a link to human customer service will be attached at the end of the reply content as a backup strategy to avoid confusion caused by the model failing to accurately answer the user's question.

[0135] Further reference Figure 10 As an implementation of the methods shown in the above figures, this disclosure provides an embodiment of an intelligent question-answering device, which is similar to... Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0136] like Figure 10As shown, the intelligent question-answering device 1000 of this embodiment may include: an intent recognition module 1001, a candidate answer generation module 1002, and a target answer generation module 1003. The intent recognition module 1001 is configured to perform intent recognition on the user's question information by combining a proprietary dictionary of the target domain, thereby obtaining an intent recognition result. The candidate answer generation module 1002 is configured to recall a set of candidate answer information that matches the question information from a preset question-answering knowledge base based on the intent recognition result. The target answer generation module 1003 is configured to determine the target answer information corresponding to the question information based on the similarity between the candidate answer information in the candidate answer information set and the question information.

[0137] In this embodiment, the specific processing of the intent recognition module 1001, the candidate answer generation module 1002, and the target answer generation module 1003 in the intelligent question-answering device 1000, and the resulting technical effects, can be referred to respectively. Figure 2 The relevant descriptions of steps 201-203 in the corresponding embodiments will not be repeated here.

[0138] In some optional implementations of this embodiment, the intent recognition module 1001 includes: a first word element generation submodule, configured to perform word segmentation processing on the question information using a proprietary dictionary to obtain multiple first word elements; an initial vector generation submodule, configured to generate an initial vector representation for each first word element based on its index number; a final vector generation submodule, configured to calculate the final vector representation of the first first word element based on the initial vector representation; a target category information acquisition submodule, configured to perform linear calculation based on the final vector representation of the first first word element to obtain the target category information of the question information; and an intent recognition result generation submodule, configured to obtain the intent recognition result based on the target category information.

[0139] In some optional implementations of this embodiment, the initial vector generation submodule is further configured to generate an initial vector representation of each first word based on the index number of each first word through a pre-trained intent recognition model; wherein, during the training process of the intent recognition model, the loss function used includes loss weights corresponding to sample category information.

[0140] In some optional implementations of this embodiment, the candidate answer generation module 1002 includes: a first answer information set acquisition submodule, configured to use inverted index recall to recall a first answer information set matching the question information from a preset question-and-answer knowledge base; a second answer information set acquisition submodule, configured to use vector recall to recall a second answer information set matching the question information from a preset question-and-answer knowledge base; a third answer information set acquisition submodule, configured to use rule recall to recall a third answer information set matching the question information from a preset question-and-answer knowledge base; and a candidate answer information set generation submodule, configured to integrate the first, second, and third answer information sets to obtain the candidate answer information set.

[0141] In some optional implementations of this embodiment, the second answer information set acquisition submodule includes: a second word element generation unit, configured to perform word segmentation processing on the answer template in the preset question-and-answer knowledge base using a proprietary dictionary to obtain multiple second word elements; a first word element vector generation unit, configured to vectorize the multiple first word elements to obtain multiple first word element vectors; a second word element vector generation unit, configured to vectorize the multiple second word elements to obtain multiple second word element vectors; a similarity calculation unit, configured to calculate the similarity between each first word element vector and the multiple second word element vectors; and a second answer information set acquisition unit, configured to obtain a second answer information set matching the question information based on the calculated similarity results.

[0142] In some optional implementations of this embodiment, the target answer generation module 1003 includes: a third word element acquisition submodule, configured to perform word segmentation processing on candidate answer information in the candidate answer information set using a proprietary dictionary to obtain multiple third word elements; a first word element vector generation submodule, configured to vectorize multiple first word elements to obtain multiple first word element vectors; a third word element vector generation submodule, configured to vectorize multiple third word elements to obtain multiple third word element vectors; a vector cross operation submodule, configured to perform vector cross operation on multiple first word element vectors and multiple third word element vectors; a similarity calculation submodule, configured to calculate the similarity between each first word element vector and multiple third word element vectors based on the vector cross operation results; a sorting submodule, configured to sort the candidate answer information based on the calculated similarity results to obtain a sorting result; and a target answer information generation submodule, configured to determine the target answer information corresponding to the question information based on the sorting result.

[0143] In some optional implementations of this embodiment, the intelligent question-answering device further includes: an initial information acquisition module, configured to acquire initial information input by the user; and a processing module, configured to intercept the initial information and / or provide feedback reminder information to the user in response to the initial information containing sensitive words.

[0144] In some optional implementations of this embodiment, the intelligent question-answering device further includes: a prompt word acquisition module, configured to acquire pre-configured prompt words, which represent the role information and rule information of the large language model; an answer information adjustment module, configured to adjust the target answer information based on the prompt words using the large language model; and an answer feedback module, configured to feed back the adjusted target answer information to the user.

[0145] In some optional implementations of this embodiment, the intelligent question-answering device further includes: an auxiliary information adding module, configured to add auxiliary service link information to the target answer information and provide feedback to the user, wherein the auxiliary service link information is used to jump to a preset consultation link.

[0146] This embodiment exists as a device embodiment corresponding to the above method embodiment. The intelligent question answering device provided in this embodiment, in a specific target domain, combines natural language understanding algorithms to identify user intent. By understanding user intent and searching the corresponding knowledge base to recall the corresponding target answer information, it can more accurately generate answer information for user question information, improve the adaptability of intelligent question answering in a specific domain, and can be used for timely interaction with users.

[0147] According to embodiments of this disclosure, this disclosure also provides an electronic device, the electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to implement the intelligent question-answering method described in any of the above embodiments when executed.

[0148] According to embodiments of this disclosure, this disclosure also provides a readable storage medium storing computer instructions that enable a computer to implement the intelligent question-answering method described in any of the above embodiments when executed.

[0149] According to embodiments of this disclosure, this disclosure also provides a computer program product that, when executed by a processor, can implement the intelligent question-answering method described in any of the above embodiments.

[0150] Figure 11A schematic block diagram of an example electronic device 1100 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0151] like Figure 11 As shown, device 1100 includes a computing unit 1101, which can perform various appropriate actions and processes according to a computer program stored in read-only memory (ROM) 1102 or a computer program loaded from storage unit 1108 into random access memory (RAM) 1103. The RAM 1103 may also store various programs and data required for the operation of device 1100. The computing unit 1101, ROM 1102, and RAM 1103 are interconnected via bus 1104. Input / output (I / O) interface 1105 is also connected to bus 1104.

[0152] Multiple components in device 1100 are connected to I / O interface 1105, including: input unit 1106, such as keyboard, mouse, etc.; output unit 1107, such as various types of monitors, speakers, etc.; storage unit 1108, such as disk, optical disk, etc.; and communication unit 1109, such as network card, modem, wireless transceiver, etc. Communication unit 1109 allows device 1100 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0153] The computing unit 1101 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1101 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1101 performs the various methods and processes described above, such as intelligent question-answering methods. For example, in some embodiments, the intelligent question-answering method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 1108. In some embodiments, part or all of the computer program may be loaded and / or installed on device 1100 via ROM 1102 and / or communication unit 1109. When the computer program is loaded into RAM 1103 and executed by the computing unit 1101, one or more steps of the intelligent question-answering method described above may be performed. Alternatively, in other embodiments, the computing unit 1101 may be configured to perform an intelligent question-answering method by any other suitable means (e.g., by means of firmware).

[0154] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0155] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0156] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0157] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0158] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with embodiments of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0159] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and Virtual Private Server (VPS) services, such as high management difficulty and weak business scalability.

[0160] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0161] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. An intelligent question-answering method, comprising: By combining a domain-specific dictionary with the user's query information, intent recognition is performed to obtain intent recognition results; Based on the intent recognition results, a set of candidate answer information that matches the question information is recalled from a preset question-and-answer knowledge base; The target answer information corresponding to the question information is determined based on the similarity between the candidate answer information in the candidate answer information set and the question information.

2. The intelligent question-answering method according to claim 1, wherein, The process of combining a domain-specific dictionary with user query information to perform intent recognition, and obtaining intent recognition results, includes: The query information is segmented using the proprietary dictionary to obtain multiple first word elements; An initial vector representation of each first word is generated based on its index number. The final vector representation of the first word element is calculated based on the initial vector representation. Linear calculations are performed based on the final vector representation of the first word element to obtain the target category information of the question information; The intent recognition result is obtained based on the target category information.

3. The intelligent question-answering method according to claim 2, wherein, The step of generating an initial vector representation for each first word based on its index number includes: An initial vector representation of each first word is generated based on the index number of each first word through a pre-trained intent recognition model. The loss function used in the training process of the intent recognition model includes loss weights corresponding to sample category information.

4. The intelligent question-answering method according to claim 2, wherein, The step of recalling a set of candidate answer information that matches the question information from a preset question-answering knowledge base based on the intent recognition result includes: The inverted index recall is used to recall the first set of answer information that matches the question information in the preset question and answer knowledge base; Vector recall is used to retrieve a second set of answer information that matches the question information from the preset question-and-answer knowledge base; The rule-based recall function retrieves a third set of answer information that matches the question information from the preset question-and-answer knowledge base. The first answer information set, the second answer information set, and the third answer information set are integrated to obtain the candidate answer information set.

5. The intelligent question-answering method according to claim 4, wherein, The step of using vector recall to retrieve a second set of answer information that matches the question information from the preset question-answering knowledge base includes: The answer templates in the preset question-and-answer knowledge base are segmented using the proprietary dictionary to obtain multiple second word elements; The plurality of first word elements are vectorized to obtain a plurality of first word element vectors; The plurality of second word elements are vectorized to obtain a plurality of second word element vectors; Calculate the similarity between each of the first word vectors and the plurality of second word vectors respectively; Based on the calculated similarity results, a second set of answer information matching the question information is obtained.

6. The intelligent question-answering method according to claim 2, wherein, The step of determining the target answer information corresponding to the question information based on the similarity between the candidate answer information in the candidate answer information set and the question information includes: The candidate answer information in the candidate answer information set is segmented using the proprietary dictionary to obtain multiple third word elements; The plurality of first word elements are vectorized to obtain a plurality of first word element vectors; The multiple third word elements are vectorized to obtain multiple third word element vectors; Perform vector cross operation on the plurality of first word vectors and the plurality of third word vectors; The similarity between each first word vector and the plurality of third word vectors is calculated by combining the results of vector cross-operation. The candidate answer information is sorted based on the calculated similarity results to obtain the sorting results; Based on the sorting results, the target answer information corresponding to the question information is determined.

7. The intelligent question-answering method according to claim 1, further comprising: Obtain initial information input by the user; In response to the initial information containing sensitive words, the initial information is intercepted, and / or a reminder message is sent to the user.

8. The intelligent question-answering method according to any one of claims 1-7, the method further comprising: Obtain pre-configured prompt words, which are used to represent the role information and rule information of the large language model; The target answer information is adjusted based on the prompt words using the large language model. The adjusted target answer information will be fed back to the user.

9. The intelligent question-answering method according to any one of claims 1-7, the method further comprising: Add auxiliary service link information to the target answer information and provide feedback to the user. The auxiliary service link information is used to redirect to a preset consultation link.

10. An intelligent question-answering device, comprising: The intent recognition module is configured to combine a domain-specific dictionary to perform intent recognition on the user's question information and obtain the intent recognition result. The candidate answer generation module is configured to recall a set of candidate answer information that matches the question information from a preset question-and-answer knowledge base based on the intent recognition result; The target answer generation module is configured to determine the target answer information corresponding to the question information based on the similarity between the candidate answer information in the candidate answer information set and the question information.

11. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the intelligent question-answering method according to any one of claims 1-9.

12. A non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the intelligent question-answering method according to any one of claims 1-9.

13. A computer program product comprising a computer program that, when executed by a processor, implements the steps of the intelligent question-answering method according to any one of claims 1-9.

Images