CN118520095A - Method and device for generating reply content based on language model - Google Patents

Abstract

The present disclosure provides a method and device for generating reply content based on a language model, the method for generating reply content comprising: obtaining current inquiry data of a user, and based on the current inquiry data, querying target knowledge associated with the current inquiry data from a preset knowledge base, wherein the target knowledge comprises first knowledge matching the current inquiry data, and second knowledge having a hierarchical relationship with the first knowledge, and source data for constructing the knowledge base comprises multiple pieces of knowledge having a hierarchical relationship; generating prompt information comprising the current inquiry data and the target knowledge; inputting the prompt information into a language model based on pre-training, and generating reply content corresponding to the prompt information by the language model, which can enhance the extensibility of the knowledge base while also enhancing the extensibility of the knowledge base retrieval, thereby enhancing the accuracy of the reply content generated by the language model based on the retrieval.

Description

Method and device for generating reply content based on language model

Technical Field

The present disclosure relates to the field of natural language processing, and in particular to a method and device for generating reply content based on a language model.

Background Art

The dialogue system is a human-computer interaction technology that aims to enable computers to have natural and fluent conversations with humans. The dialogue system is an important research direction in the field of artificial intelligence. It has important practical application value and wide universality. The dialogue system in related technologies has been applied in various fields such as intelligent customer service, intelligent assistants, and chatbots.

The dialogue systems of related technologies already have excellent text comprehension and dialogue capabilities. However, for existing knowledge-base-based dialogue systems, due to the poor scalability of the knowledge base, the accuracy of the response content generated by the dialogue system relying on the knowledge base during the dialogue needs to be improved.

Summary of the invention

In view of the above technical problems, the present disclosure provides a method and device for generating reply content based on a language model, and the technical solution is as follows:

According to a first aspect of an embodiment of this specification, a method for generating reply content based on a language model is provided, the method comprising:

Acquire the user's current query data, and based on the current query data, query the target knowledge associated with the current query data from a preset knowledge base, wherein the target knowledge includes first knowledge matching the current query data and second knowledge having a hierarchical relationship with the first knowledge, and the source data used to construct the knowledge base includes multiple pieces of knowledge having a hierarchical relationship;

generating prompt information including the current query data and the target knowledge;

The prompt information is input into a pre-trained language model, and the language model generates a reply content corresponding to the prompt information.

According to a second aspect of an embodiment of this specification, there is provided a device for generating reply content based on a language model, the device comprising:

A query module, used to: obtain the user's current query data, and based on the current query data, query the target knowledge associated with the current query data from a preset knowledge base, wherein the target knowledge includes first knowledge matching the current query data and second knowledge having a hierarchical relationship with the first knowledge, and the source data used to construct the knowledge base includes multiple pieces of knowledge having a hierarchical relationship;

A generation module is used to: generate prompt information including the current query data and the target knowledge; input the prompt information into a pre-trained language model, and generate a reply content corresponding to the prompt information by the language model.

The technical solutions provided by the embodiments of this specification may have the following beneficial effects:

In an embodiment of the present specification, target knowledge associated with the user's current query data is retrieved from a preset knowledge base, and the retrieved target knowledge is combined with the user's current query data to generate prompt information, which is input into a language model, and the language model generates reply content corresponding to the prompt information. In the reply content generation stage, external knowledge sources can be introduced through knowledge base retrieval to enhance the reply capability of the language model. At the same time, for knowledge sources with hierarchical relationships, a knowledge base that can contain the hierarchical relationship is pre-constructed. When searching for knowledge associated with the user's current query data in the knowledge base, the first knowledge matching the current query data can be directly queried, and the second knowledge that has a hierarchical relationship with the first knowledge can be further queried. While enhancing the scalability of the knowledge base, the scalability of the knowledge base retrieval is also enhanced, thereby enhancing the accuracy of the reply content generated by the language model based on the retrieval.

It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the technical solutions in the embodiments of this specification or related technologies, the drawings required for use in the embodiments or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this disclosure, and for ordinary technicians in this field, other drawings can also be obtained based on these drawings.

FIG1 is a schematic diagram of the structure of an interactive system provided by an embodiment of this specification;

FIG2 is a flow chart of a method for generating reply content according to an embodiment of the present specification;

FIG3 is a schematic diagram of a document structure of an embodiment of the present specification;

FIG4 is a schematic diagram of the Markdown format of an embodiment of the present specification;

FIG5 is a schematic diagram of the KB-Tree structure of an embodiment of the present specification;

FIG6 is a schematic diagram of the hierarchical relationship of knowledge nodes according to an embodiment of the present specification;

FIG7 is a schematic diagram of a specific application scenario of generating reply content according to an embodiment of the present specification;

FIG8 is a schematic diagram of the structure of a reply content generating device according to an embodiment of the present specification;

FIG. 9 is a schematic diagram of the structure of an electronic device according to an embodiment of the present specification.

DETAILED DESCRIPTION

Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with this specification. Instead, they are merely examples of devices and methods consistent with some aspects of this specification as detailed in the appended claims.

The terms used in this specification are for the purpose of describing specific embodiments only and are not intended to limit this specification. The singular forms "a", "the" and "the" used in this specification and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and/or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

It should be understood that although the terms first, second, third, etc. may be used in this specification to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of this specification, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

The dialogue system is a human-computer interaction technology that aims to enable computers to have natural and fluent conversations with humans. The dialogue system is an important research direction in the field of artificial intelligence, with important practical application value and wide universality. The importance of the dialogue system lies in that it can provide an efficient and convenient human-computer interaction method. Traditional human-computer interfaces, such as graphical user interfaces and command line interfaces, require users to learn specific operations and commands and follow fixed interaction modes. The dialogue system interacts through natural language, which is closer to the daily communication method of humans. It allows users to express their intentions, ask questions, and obtain information through natural dialogue without learning specific commands and operations.

Dialogue systems are widely used in real life and technology. In daily life, one of the application scenarios of dialogue systems - voice assistants have become a common tool that can help users complete various tasks, such as checking the weather, playing music, setting reminders, etc. In addition, dialogue systems are also widely used in the field of customer service, providing automated customer support and problem solving, reducing the burden on manual customer service.

In the field of technology, dialogue systems are widely used in intelligent robots, virtual personal assistants, intelligent recommendation systems, etc. Intelligent robots can have real-time conversations with users and perform designated tasks. For example, home assistant robots can communicate with users through dialogues and perform household management tasks. Virtual personal assistants (such as voice assistants on smartphones) can provide users with personalized services and suggestions. Intelligent recommendation systems can understand users' likes and preferences through dialogues, thereby providing personalized recommendation information.

In addition, dialogue systems are also widely used in education, medical care, finance, e-commerce and other fields. In the field of education, dialogue systems can serve as learning companions to provide personalized learning support and feedback. In the medical field, dialogue systems can be used for health consultation, disease diagnosis assistance, etc. In the financial field, dialogue systems can provide customer service, investment consulting and other services. In the field of e-commerce, dialogue systems can provide personalized product recommendations, shopping guidance and other services.

The dialogue system of related technologies has excellent text comprehension and dialogue capabilities. However, for the existing dialogue system based on knowledge base, due to the poor scalability of the knowledge base, the accuracy of the reply content generated by the dialogue system during the dialogue still needs to be improved. For example, the traditional question-answering robot based on the question-answer (QA) knowledge base requires a lot of expert experience to manually organize the document knowledge into question-answer pairs during the knowledge base construction stage. The labor cost of knowledge base maintenance is high and the timeliness is poor. In the user question stage, it can only reply based on the fixed answers corresponding to the relevant questions in the knowledge base where the question-answer pairs exist, and the reply is not scalable. In the reply stage, since the answers corresponding to the questions in the knowledge base are fixed and unchanging, the robot's reply has a fixed speech template and no reading comprehension ability.

Although existing dialogue systems have been applied in many fields, there are still some fields that do not involve the application of dialogue systems. Take the content security field as an example. With the diversification and complexity of business forms and scenarios in the content security field, the accumulated knowledge in the content security field (including rules, cases, interpretations, risk calendars, workflows, review mechanisms, liaison personnel, etc.) is also increasing. However, when business personnel in the current content security field need to obtain knowledge in the content security field, they often need to manually consult content security operators, who will manually reply to them, and the work efficiency is relatively low. Therefore, how to apply the dialogue system in the content security field so that business personnel (that is, users who initiate inquiries) can use the dialogue system to obtain the required content security knowledge and work guidelines more conveniently, accurately and intelligently without becoming experts in the content security field is also an urgent problem to be solved.

In response to the above problems, the embodiments of this specification provide a method for generating reply content based on a language model, which can enhance the extensibility of the knowledge base and the extensibility of the knowledge base search, thereby enhancing the accuracy of the reply content generated by the language model based on the search. In the actual application process, the user's current query data is first obtained, and based on the user's current query data, the target knowledge associated with the current query data is queried from the preset knowledge base, wherein the target knowledge includes the first knowledge matching the current query data, and the second knowledge having a hierarchical relationship with the first knowledge, and the source data used to construct the knowledge base includes multiple pieces of knowledge with a hierarchical relationship. After the target knowledge is queried, prompt information including the user's current query data and the target knowledge associated with the current query data is generated, and the generated prompt information is input into the pre-trained language model, and the language model generates the reply content corresponding to the prompt information.

In the embodiments of the present specification, the target knowledge associated with the user's current query data can be retrieved from a preset knowledge base, and the retrieved target knowledge and the user's current query data can be combined to generate prompt information, which is input into a language model, and the language model generates reply content corresponding to the prompt information. In the reply content generation stage, external knowledge sources can be introduced through knowledge base retrieval to enhance the reply capability of the language model. At the same time, for knowledge sources with hierarchical relationships, a knowledge base that can contain the hierarchical relationship is pre-constructed. When searching for knowledge associated with the user's current query data in the knowledge base, the first knowledge that matches the current query data can be directly queried, and the second knowledge that has a hierarchical relationship with the first knowledge can be further queried. While enhancing the scalability of the knowledge base, the scalability of the knowledge base retrieval is also enhanced, thereby enhancing the accuracy of the reply content generated by the language model based on the retrieval.

In an exemplary embodiment, please refer to Figure 1, which is an interactive system provided in an embodiment of this specification, wherein the interactive system includes a server 101 and at least one client 102. Exemplarily, the client 102 can access the server 101 through the network to use the services provided by the server 101, including but not limited to conversation services, product delivery services, product purchase services, reading services, audio and video playback services, or search services, etc.

The server 101 may be a program installed in a background device to provide services to users. Exemplarily, as shown in FIG1 , the background device may be a server, which may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

The client 102 may be a program installed in a user device to provide services to the user, and the client 102 includes but is not limited to an application APP, a web page, a small program, a plug-in or a component, etc. As shown in FIG1 , the user device includes but is not limited to a smart phone, a personal digital assistant, a tablet computer, a personal computer, a laptop computer, a virtual reality terminal device, an augmented reality terminal device, etc.

The method for generating reply content based on the language model provided in the embodiment of this specification can be executed by either the server 101 or the client 102, or by the cooperation of the two (each of which executes a part of the steps), and this embodiment does not impose any restrictions on this. Taking the server 101 executing the reply content generation method as an example, the server 101 can be deployed with a language model, the client 102 can send the user's current inquiry data to the server 101, and the server 101 can execute the reply content generation method provided in the embodiment of this specification to obtain the reply content, and then the reply content can be sent to the client 102 so that the client 102 can display the reply content; or the language model can be deployed in the client 102, and the reply content generation method can be executed by the client 102, and there is no specific limitation on this.

Next, the process of the reply content generation method is exemplified:

Please refer to FIG2 . The embodiment of this specification provides a method for generating reply content based on a language model. As shown in FIG2 , the method includes the following steps:

S201. Acquire current query data of a user, and based on the current query data, query target knowledge associated with the current query data from a preset knowledge base.

The target knowledge includes first knowledge matching the current query data and second knowledge having a hierarchical relationship with the first knowledge, and the source data used to construct the knowledge base includes multiple pieces of knowledge having a hierarchical relationship.

S202: Generate prompt information including the current query data and the target knowledge.

S203: input the prompt information into a pre-trained language model, and use the language model to generate a reply content corresponding to the prompt information.

As an example, a knowledge base in a certain vertical field can be constructed to implement a language model for automatic question and answering in the vertical field based on the knowledge base. In actual applications, the source data of some vertical field knowledge bases contain multiple pieces of knowledge with hierarchical relationships, such as the content security field. Taking the content security field as an example, the source data in this field includes data such as laws, regulations, rules, cases, work processes or review processes, and the multiple pieces of knowledge recorded in these data have a hierarchical relationship. It is worth noting that the above introduction to the vertical field using the content security field as an example is only an exemplary display. In actual applications, the knowledge base of this embodiment can also be various other vertical fields, and there is no specific limitation on this.

The above-mentioned knowledge base can be constructed in a variety of ways. As an example, one of the construction methods may include: the source data used to construct the knowledge base may include a document, the text in the document may be identified, the identified text is converted into data in Markdown format, and then a tree structure is constructed based on the data in Markdown format to obtain a knowledge base containing the tree structure. A data structure of a tree structure is constructed based on the Markdown storage format with hierarchical information, so as to build a knowledge structure with a hierarchical relationship in the knowledge base. It is worth noting that the above introduction to the construction method of the knowledge base is only an exemplary display. In actual applications, other construction implementations are not excluded and are not limited to this. Among them, the tree structure contains multiple nodes, each node stores a piece of knowledge, and the relationship between each node corresponds to the hierarchical relationship between each piece of knowledge in the source data.

As an example, the tree structure may be a KB-Tree data structure or other data structures, and there is no limitation on the specific implementation of the tree structure.

In conjunction with FIG. 3 to FIG. 5 , taking the above document as an example, a specific implementation method of identifying the text in the document of the Advertising Law, converting it into Markdown format, and then constructing the Markdown format of the Advertising Law into a KB-Tree data structure may include:

(1) Convert the Advertising Law document to Markdown format:

In this step, you need to convert the original format of the Advertising Law (which may be a Word document, PDF or other format) into Markdown format, which may involve the following operations:

① Title recognition: Identify the various titles in the document, such as chapter titles, section titles, etc., and format them using Markdown's title syntax. For example, use `#`, `##`, `###`, etc. to represent titles of different levels.

②Text formatting: Convert the main body of the text into paragraphs in Markdown format.

③List conversion: If there are lists in the document (such as a list of terms, a list of conditions, etc.), they need to be converted into an unordered list (using `-` or `*`) or an ordered list (using a number followed by `.`) supported by Markdown.

④ Links and pictures: If the document contains links or pictures, use Markdown's link and picture syntax to format them.

⑤Table creation: If there is table data in the document, use Markdown's table syntax to create the table.

(2) Construct KB-Tree based on Markdown structure:

① Determine the parent node: Take the highest level title of the text as the parent node of KB-Tree.

②Build child nodes: For each next-level title or content, create a child node of the parent node. For example, if there is a first-level title in the Markdown document and multiple second-level titles below it, then each second-level title should be a child node of the first-level title node.

③Process the main text content: For non-title main text content, you can add it as a leaf node under the corresponding parent node, or include it in the content attribute of the parent node.

④Add related attributes: If the Markdown document contains links, pictures or other special format content, these contents can be added to KB-Tree as node attributes.

⑤Maintain hierarchical relationships: Ensure that each node in the KB-Tree correctly reflects its hierarchical relationship in the Markdown document.

The above process is described below with reference to Figures 3 to 5:

First, it can be assumed that the advertising law document contains the structure shown in FIG. 3 , and the structure shown in FIG. 3 may be converted into the Markdown format as shown in FIG. 4 , and the KB-Tree constructed according to the Markdown format shown in FIG. 4 may be shown in FIG. 5 .

In practical applications, the specific implementation of constructing KB-Tree according to the Markdown format is not limited to the above method, and there is no specific limitation on this. The above document can have multiple specific implementations. As an example, the above document can be a DOC document, an Excel document or a PDF document. Each document records a type of knowledge. There is no limitation on the specific implementation of the document.

As an example, for the above-mentioned Excel document, the data originally stored in the table can be converted into plain text data through the column name conversion template; as another example, for the above-mentioned PDF document, the original image text can be extracted into plain text data through the PDF conversion capability of OCR.

As an example, the knowledge structure of different hierarchical relationships represented by the KB-Tree data structure of the above-mentioned knowledge base may include structures such as node index, title and text information, which are not specifically limited.

As an example, each source data can construct a tree structure, and the knowledge base can contain multiple tree structures constructed from multiple source data. These tree structures can serve as a subtree of the root node. The root node can be the entrance to the knowledge base and is not used to store specific knowledge.

As an example, after constructing the above tree structure and obtaining a knowledge base containing the tree structure, the knowledge base containing the tree structure can be synchronously stored in the intermediate storage medium OSS, and the stored knowledge base can be updated regularly by the online service to ensure the convenient update of knowledge in the knowledge base. The specific update method can be to obtain new source data regularly, continue to construct the tree structure of the new source data according to the above embodiment, and update the newly constructed tree structure to the knowledge base.

A pre-trained language model can be constructed in many ways. As an example, a base model can be selected first. The base model can be understood as a language model that has been pre-trained with a large amount of general data and has general capabilities. As another example, a large language model (LLM) can be selected as the above base model.

After selecting the base model, while retaining the base parameter capabilities of the original base model, the base model can be trained again using data from a specific field (technically referred to as "fine-tuning" the base model). As an example, the data in the specific field mentioned above can refer to the content security field mentioned in the above embodiment, or other vertical fields, which are not specifically limited.

There are many ways to retrain the selected base model, i.e., the pre-trained language model (i.e., "fine-tune" the selected base model). As an example, one of the training methods may include: obtaining historical conversation data, and generating data with different description methods for the user query data and the answer data corresponding to the user query data in the historical conversation data, respectively, to obtain a training data set, and then retraining the language model with at least the training data set. In order to prevent overfitting when retraining the language model, it is necessary to ensure the diversity of questions and corresponding answers when constructing the training data set, and introduce different description methods for data enhancement.

There are multiple ways to generate data with different description modes for user inquiry data in historical session data and reply data corresponding to the user inquiry data. As an example, one of the generation methods may include: obtaining a preset description template, and based on the preset description template, generating data with different semantic description modes for user inquiry data in historical session data and reply data corresponding to the user inquiry data, the preset description template including multiple preset description modes with different semantics for the same type of questions and corresponding answers to the questions. As another example, the questions and corresponding answers involved in the user inquiry data in the historical session data may be questions and corresponding answers of a defined category, and the types of questions and corresponding answers targeted by the preset description template may also be questions and corresponding answers of a defined category.

As another example, the definition category may mean a query for the definition of a proper noun in a specific field, or a query for the definition of other objects, and there is no specific limitation on this.

To illustrate the "definition type", take the same specific person A as an example, "Who is A", "Can you introduce A", "Please introduce A", "What does A do", "Please briefly describe A" are all inquiries with different semantics for A's definition information.

As another example, another method of generating data with different description methods for user inquiry data in historical conversation data and answer data corresponding to the user inquiry data may include: generating task information, which may be used to prompt a general pre-trained language model to generate different semantic description methods for the same question and the answer corresponding to the same question, and inputting the task information and the above-mentioned historical conversation data into the general pre-trained language model, so that the general pre-trained language model generates data with different semantic description methods for the user inquiry data in the historical conversation data and the answer data corresponding to the user inquiry data.

As another example, "general pre-trained language model" may refer to the language model mentioned in S203 of this specification, or may refer to other general language models, and there is no specific limitation on this.

As an example, the language model can be trained again based on the above training data set and supervised fine-tuning (SFT). SFT can be a process of constructing a supervised data set to retrain and fine-tune the parameters of the language model, so that the language model parameters can learn the knowledge and expression of related data. For example, a language model fine-tuning data set in the field of content security can be constructed (that is, the above training data set can be in the field of content security), and the data format can be adjusted based on the selection of the base model. The input and output forms of the model in the training process and the application stage are consistent. When training, the input of the language model is consistent with the form or structure of the prompt information in step 202, so the details can be seen in the subsequent description of the implementation example of the prompt information.

As an example, when the above language model is LLM, the above-mentioned re-training process of LLM can be performed based on the AdaLora structure and SFT, that is, the process of retraining and fine-tuning the language model. AdaLora can automatically allocate the number of fine-tunable parameters according to the importance of different layers and different types of parameters of the language model.

As an example, after the above-mentioned knowledge base and language model are constructed, the above-mentioned knowledge base and the above-mentioned language model can be used to generate reply content. The following is an exemplary description of the use stages of the above-mentioned knowledge base and the above-mentioned language model:

First, you need to get the user's current query data:

The user's current query data (query) may have multiple specific implementations. As an example, one specific implementation may include: the current query data may be text data; as another example, the text data may be keyword text, phrase text, or question text.

As another example, the user's current query data may also be image data, which may contain extractable text information.

As another example, the user's current inquiry data may also be voice data, which may contain extractable text information.

It is worth noting that the above introduction to the specific implementation of the user's current query data is only an exemplary display. In actual application, other specific implementations are not excluded and are not limited to this.

After obtaining the current query data, the target knowledge associated with the current query data is queried from the knowledge base based on the current query data:

Consider that some specific knowledge in the knowledge base may be time-sensitive, that is, the validity of some knowledge is time-limited. For example, when the knowledge is in the field of laws and regulations, some laws and regulations may have specific effective or expiration times. For example, Article X, Paragraph X of the Advertising Law will not take effect until after X month X day of X year. If only the content of the knowledge is considered when conducting knowledge query without considering the timeliness of the knowledge, invalid knowledge may be retrieved, which will make the response content inaccurate.

In view of this problem, there are multiple ways to query the target knowledge associated with the current query data from the preset knowledge base. As an example, one query method may include: at least one piece of knowledge in the preset knowledge base may be associated with standard time information, and the time-related text in the user's current query data may be identified, and the time-related text is converted into standard time information, and then the above target knowledge is queried based on the user's current query data and the converted standard time information. Since the language model itself in the related art does not have the ability to judge the current time, the language model may not be able to perceive what time and date the time-related text in the user's current query data specifically refers to, such as today, this week, and other fuzzy time expressions. The language model itself may not have the ability to judge the current time, such as what date "today" refers to. This embodiment can provide the time fuzzy recognition capability of the user's query, for example, converting the "today" mentioned in the user's query into the standard time information "April 1, 2024", so that the time-related text in the user's current query data can be better recognized and understood.

The knowledge in the preset knowledge base can be associated with standard time information in a variety of ways. As an example, the standard time information can be pre-carried in the knowledge; as another example, the standard time information can be associated with the knowledge in the process of storing the knowledge in the knowledge base. For example, a certain source data has an effective time range. For the tree structure constructed by the source data, all knowledge stored in the nodes need to be associated with standard time information. The root node of the tree structure can be associated with standard time information, and all child nodes under the root node are associated with standard time information. Alternatively, part of the knowledge in a source data may have an effective time range. For the tree structure constructed by the source data, the nodes where the knowledge that needs to be associated with standard time information is located are associated with standard time information. It is worth noting that the above introduction to the association method of standard time information is only an exemplary display. In practical applications, other association methods are not excluded and are not limited to this.

There can be many specific implementations of standard time information. As an example, the standard time information can be information including year, month, and day, such as "April 1, 2024"; it can also include only month or year information, such as "April" or "2024"; it can also be date information without year, for example, "April 1"; as another example, the standard time information can also include time information accurate to hours, minutes or seconds, such as "April 1, 2024 13:01:01". The standard time information can also be time information corresponding to the season, such as January to March (spring), April to June (summer), July to September (autumn), and October to December (winter). In actual applications, the standard time information does not exclude the existence of other specific implementations and is not limited to this.

Consider that some specific knowledge in the knowledge base may be related to sensitive people. For example, certain specific deeds, specific historical events, and specific judicial cases are related to specific sensitive people. If only the content of the knowledge is considered when performing knowledge query without considering the relevance of the knowledge to sensitive people, sensitive knowledge may be retrieved, making the content of the response more sensitive and posing a greater risk. In response to this problem, as an example, another way to query the target knowledge associated with the current inquiry data from a preset knowledge base may include: at least one piece of knowledge in the preset knowledge base may be associated with named entity information, and the naming information related to the named entity in the user's current inquiry data may be identified, and the above-mentioned target knowledge may be queried based on the user's current inquiry data and the naming information identified from the user's current inquiry data. In some dialogue scenarios, named entities are entity content that needs to be identified in the user's inquiry data, so a named entity recognition capability is provided to efficiently query the knowledge associated with the named entity contained in the inquiry data.

Named entity information can have multiple specific implementations. For example, named entity information can be character name information. Taking the content security field dialogue scenario as an example, in this scenario, it is necessary to focus on identifying the character name information contained in the user's query data to identify the name of a specific person, so as to quickly query the knowledge associated with the specific person and provide an efficient response. It is worth noting that the above introduction to the specific implementation of named entity information is only an exemplary display. In actual applications, other specific implementations are not excluded. For example, named entity information can also be information about other entity objects such as information about a specific object, and there is no specific limitation on this.

As an example, another way to query target knowledge associated with current query data from a preset knowledge base may include: at least one piece of knowledge in the preset knowledge base may be simultaneously associated with standard time information and named entity information, and time-related text and named entity-related naming information in the user's current query data may be identified, and the time-related text may be converted into standard time information, and finally the above-mentioned target knowledge may be queried based on the user's current query data, the converted standard time information, and the naming information identified from the user's current query data.

As an example, when querying the target knowledge associated with the current query data from the preset knowledge base, since the knowledge base contains a tree structure, each node in the tree structure stores a piece of knowledge. Starting from the root node of the tree structure, all nodes of the entire tree structure can be traversed in sequence to retrieve the target knowledge associated with the current query data.

As an example, another way to query target knowledge associated with current query data from a preset knowledge base may include: from the above-mentioned knowledge base, querying the first knowledge associated with the user's current query data at least in a vector similarity query manner, and the vector similarity between the first knowledge and the user's current query data is greater than or equal to a first threshold.

As an example, another method of querying target knowledge associated with current query data from a preset knowledge base may include: querying the first knowledge associated with the user's current query data from the above-mentioned knowledge base at least in a vector similarity query method and a character-level similarity query method, respectively, wherein the vector similarity between the first knowledge and the current query data is greater than or equal to the first threshold, and the character-level similarity between the first knowledge and the current query data is greater than or equal to the second threshold. Combining the two query methods of the vector similarity query method and the character-level similarity query method to form a fusion similarity query method that combines the two can improve query accuracy.

As an example, another method of querying target knowledge associated with current query data from a preset knowledge base may include: Based on the above-mentioned vector similarity query method, or the above-mentioned vector similarity query method combined with the above-mentioned character-level similarity query method, the keywords contained in the user's current query data can also be identified, and then the first knowledge matching the keywords contained in the current query data is queried from the above-mentioned knowledge base through the keyword regular matching query method. The keyword regular matching query method can improve the query accuracy of proper nouns in specific scenarios, such as the query accuracy of named entity information in the above-mentioned conversation scenario in the content security field; further, the keyword regular matching query method combined with the above-mentioned vector similarity query method and the above-mentioned character-level similarity query method can further improve the query accuracy.

As an example, the fusion query method can specifically be: the target knowledge includes: knowledge whose vector similarity with the current query data is greater than or equal to the first threshold, and whose character-level similarity with the current query data is greater than or equal to the second threshold, and also includes first knowledge that matches the keywords contained in the current query data.

The above-mentioned vector similarity query method can be implemented in many specific ways. As an example, the vector similarity query method can be a text vectorization (Text to Vector, Text2vec) query method. Text2vec is a neural network model for text vectorization representation, which can convert text into vector features, and then measure the correlation between the search user query and the retrieved knowledge through cosine similarity. The specific implementation of the vector similarity query method is not limited. As another example, the corresponding vector of each piece of knowledge in the above-mentioned knowledge base can be pre-calculated and stored in the knowledge base. After the user's current query data is obtained, the vector corresponding to the user's current query data can be calculated first, and the vector corresponding to the user's current query data can be compared with the vector corresponding to each piece of knowledge in the knowledge base to calculate the vector similarity.

The above-mentioned character-level similarity query method can be implemented in many specific ways. As an example, the character-level similarity query method can be a character-level similarity query algorithm (Best matching 25, BM25) query method. BM25 is an improved algorithm based on TF-IDF, which is used to measure the relevance between search terms and documents. In TF-IDF, TF is term frequency (TermFrequency), and IDF is inverse document frequency index (Inverse Document Frequency), which can be used as a representation feature of text information for information retrieval.

As an example, another method of querying the target knowledge associated with the current query data from the preset knowledge base may include: the above-mentioned knowledge base uses a tree structure to store multiple pieces of knowledge with a hierarchical relationship contained in the source data used to construct the knowledge base, and the first node storing the first knowledge matching the user's current query data can be queried from the tree structure, and then the second node having the same parent node as the first node is obtained from the tree structure, and the second knowledge stored in the second node is obtained. If the nodes to which the first knowledge and the second knowledge belong have the same parent node, the hierarchical relationship between the first knowledge and the second knowledge can be the same hierarchical relationship (same-level relationship), and the knowledge can be further associated and the knowledge retrieval query can be expanded through the topological relationship of the nodes at the same level of the nodes to which the knowledge matching the current query data belongs.

As an example, another method of querying target knowledge associated with current query data from a preset knowledge base may include: the above-mentioned knowledge base uses a tree structure to store multiple pieces of knowledge with a hierarchical relationship contained in the source data used to construct the knowledge base; as another example, the first node storing the first knowledge matching the user's current query data can be queried from the tree structure, and then the third node as the parent node of the first node is obtained from the tree structure, and the third knowledge stored in the third node is obtained; as another example, the first node storing the first knowledge matching the user's current query data can be queried from the tree structure, and then the fourth node as the child node of the first node is obtained from the tree structure, and the fourth knowledge stored in the fourth node is obtained; as another example, the first node storing the first knowledge matching the user's current query data can be queried from the tree structure, and then the third node as the parent node of the first node and the fourth node as the child node of the first node are obtained from the tree structure, and the third knowledge stored in the third node and the fourth knowledge stored in the fourth node are obtained. The third knowledge node is the parent node of the first knowledge node, and the fourth knowledge node is the child node of the first knowledge node. The topological relationship between the parent node or child node of the knowledge node matching the current query data can be used to further associate knowledge and expand knowledge retrieval queries.

Please refer to FIG6 , and the following is an exemplary introduction to the above knowledge and the hierarchical relationship of the nodes:

As an example, the second knowledge that has a hierarchical relationship with the first knowledge can be understood as the second knowledge and the first knowledge belonging to the same level, that is, the node to which the second knowledge and the first knowledge belong is a node of the same level with the same parent node; the second knowledge that has a hierarchical relationship with the first knowledge can also be understood as the second knowledge being the third knowledge whose node is the parent node of the node to which the first knowledge belongs, or the second knowledge can also be the fourth knowledge whose node is the child node of the node to which the first knowledge belongs, or the second knowledge can also include one or all of the three types of knowledge: the knowledge stored in the node of the same level of the node to which the first knowledge belongs (the second knowledge), the knowledge stored in the parent node of the node to which the first knowledge belongs (the third knowledge), and the knowledge stored in the child node of the node to which the first knowledge belongs (the fourth knowledge). It is worth noting that the above introduction to the specific implementation of the hierarchical relationship of knowledge and the nodes is only an exemplary display. In practical applications, other specific implementations are not excluded and are not limited to this.

After the target knowledge is found, prompt information including the user's current query data and the found target knowledge can be generated:

There are many ways to generate the above prompt information. As an example, one of the generating methods may include: based on the above target knowledge, the identity information corresponding to the above language model is obtained, and the identity information is used to instruct the language model to respond with a preset identity, and then generate prompt information containing the user's current query data, the above target knowledge, and the identity information corresponding to the above language model. The constructed prompt information may include identity role information to instruct the language model to respond with a specific identity role, thereby improving the accuracy of the language model's response to specific fields.

The above-mentioned preset identity can have multiple specific implementations. As an example, the above-mentioned preset identity can refer to the reply role identity corresponding to the field to which the target knowledge belongs; as another example, the field to which the target knowledge belongs can be the content security field; to give an illustrative example: when the field to which the target knowledge belongs is the content security field, the preset identity can be an audit assistant or a question-and-answer assistant in the content security field; as a more detailed example, when the field to which the target knowledge belongs is the advertising law field in the content security field, the preset identity can be an "advertising law expert identity"; therefore, there is no limitation on the specific implementation of the preset identity.

As an example, another way to generate the above-mentioned prompt information may include: generating prompt information including the user's current inquiry data, the above-mentioned target knowledge, the identity information corresponding to the above-mentioned language model and preset restricted answer information, wherein the restricted answer information is used to instruct the language model to reply within the range of replies represented by the target knowledge, which can limit the scope of the language model's reply, avoid the language model from fabricating reply content as much as possible, and further improve the accuracy of the reply.

The following is an example of the prompt information:

As an example, the prompt information may include one or more of the following information:

①The user's current query data:

For example, the text may be "Contents of the amendment to the Advertising Law by person A on date B" or "Whether behavior D performed on date C violates the Advertising Law" or similar content. In actual applications, it can be flexibly adjusted as needed, and this embodiment does not limit this.

②Target knowledge:

For example, the text may be as follows: "The revised content is: revisions have been made to Chapter X, Article X, Clause X of the Advertising Law. The content before the revision was: XX, and the content after the revision is: XXX" or "It violates the provisions of Chapter X, Article X, Clause X of the Advertising Law, and also violates the provisions of Chapter X, Article X, Clause X. Behavior D does not violate the above clause before date C, but violates the above clause after date C" and other similar content. In actual applications, it can be flexibly adjusted as needed, and this embodiment does not limit this.

③Identity information:

For example, the text may be "You are a question-answering assistant in the field of content security", "You are an auditing assistant in the field of content security", or "You are an advertising law expert" or similar content. In actual applications, it can be flexibly adjusted as needed, and this embodiment does not limit this.

④Preset restricted answer information:

For example, the text may be as follows: "Please answer based on the target knowledge in the prompt text", "Please answer within the scope of the specific clauses involved in the Advertising Law", or "Please answer within the scope of the revision date of the Advertising Law", etc. In actual applications, it can be flexibly adjusted as needed, and this embodiment does not limit this.

It is worth noting that the above-mentioned introduction to the method of generating the above-mentioned prompt information is only an exemplary display. In practical applications, other generation methods are not excluded and are not specifically limited to this.

After the prompt information is generated, the generated prompt information can be input into the above-mentioned pre-trained language model, and the language model generates the reply content corresponding to the prompt information.

Please refer to Figure 7. Figure 7 shows a reply content generation framework constructed based on the embodiments of this specification. As an example, the framework can be constructed based on the Retrieval-Augmented Generation (RAG) framework, which can be divided into three parts: knowledge base construction, knowledge retrieval (query), and large model reasoning generation. These three parts can all be deployed in the user's private server, which can effectively protect the privacy of user information and domain knowledge data.

Corresponding to the above method embodiment, the embodiment of this specification further provides a reply content generating device, as shown in FIG8 , the device may include:

The query module 801 is used to obtain the current query data of the user, and based on the current query data, query the target knowledge associated with the current query data from the preset knowledge base, wherein the target knowledge includes the first knowledge matching the current query data and the second knowledge having a hierarchical relationship with the first knowledge, and the source data used to construct the knowledge base includes multiple pieces of knowledge having a hierarchical relationship;

The generation module 802 is used to generate prompt information including the current query data and the target knowledge; input the prompt information into a pre-trained language model, and generate a reply content corresponding to the prompt information by the language model.

As an example, the query module 801 is specifically used to identify time-related text in the current query data and convert the time-related text into standard time information; and query the target knowledge based on the current query data and the converted standard time information.

As an example, at least one piece of knowledge in the knowledge base is associated with named entity information, and the query module 801 is specifically used to identify the naming information related to the named entity in the current query data, and query the target knowledge based on the current query data and the naming information.

As an example, the query module 801 is specifically configured to query the first knowledge associated with the current query data from the knowledge base at least in a vector similarity query manner, wherein the vector similarity between the first knowledge and the current query data is greater than or equal to a first threshold.

As an example, the query module 801 is specifically used to query the first knowledge from the knowledge base, at least using a vector similarity query method and a character level similarity query method, respectively, where the vector similarity between the first knowledge and the current query data is greater than or equal to the first threshold, and the character level similarity between the first knowledge and the current query data is greater than or equal to a second threshold.

As an example, the query module 801 is specifically used to identify the keywords contained in the current query data; and query the first knowledge matching the keywords contained in the current query data from the knowledge base through a keyword regular matching query method.

As an example, the query module 801 is specifically used to obtain identity information corresponding to the language model based on the target knowledge, and the identity information is used to instruct the language model to reply with a preset identity; and generate prompt information including the current query data, the target knowledge and the identity information.

As an example, the generation module 802 is specifically used to generate prompt information including the current query data, the target knowledge, the identity information and preset restricted answer information, and the restricted answer information is used to instruct the language model to respond within the range of responses represented by the target knowledge.

As an example, the training process of the language model at least includes:

Acquire historical conversation data, and generate data in different description modes for user inquiry data and answer data corresponding to the user inquiry data in the historical conversation data, to obtain a training data set;

The language model is trained again using at least the training data set.

As an example, the knowledge base uses a tree structure to store the multiple pieces of knowledge with a hierarchical relationship; the query module 801 is specifically used to query the first node that stores the first knowledge matching the current query data from the tree structure; obtain the second node that has the same parent node as the first node from the tree structure, and obtain the second knowledge stored in the second node.

As an example, the knowledge base uses a tree structure to store the multiple pieces of knowledge with a hierarchical relationship; the query module 801 is specifically used to query the first node that stores the first knowledge matching the current query data from the tree structure; obtain the third node that is the parent node of the first node from the tree structure, and obtain the third knowledge stored in the third node, and/or obtain the fourth node that is the child node of the first node from the tree structure, and obtain the fourth knowledge stored in the fourth node.

As an example, the source data includes a document; the above-mentioned knowledge base is constructed in the following manner: identifying text in the document, converting the identified text into data in Markdown format; constructing a tree structure based on the data in Markdown format to obtain a knowledge base containing the tree structure.

The embodiments of this specification also provide a computer program product, including a computer program, which, when executed by a processor, implements the reply content generation method described in any of the embodiments above.

The embodiment of this specification also provides an electronic device, as shown in FIG9 , the electronic device includes:

Processor 901;

Memory 902 for storing processor executable instructions;

The processor 901 is configured to implement the reply content generation method described in any of the above embodiments.

The embodiments of this specification also provide a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method for generating reply content described in any of the embodiments above is implemented.

For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, that is, they may be located in one place, or they may be distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this specification. Ordinary technicians in this field can understand and implement it without paying creative work.

The above embodiments can be applied to one or more computer devices, where the computer device is a device that can automatically perform numerical calculations and/or information processing according to pre-set or stored instructions. The hardware of the computer device includes but is not limited to a microprocessor, an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a digital signal processor (DSP), an embedded device, etc.

The computer device may be any electronic product that can perform human-computer interaction with a user, such as a personal computer, a tablet computer, a smart phone, a personal digital assistant (PDA), a game console, an interactive network television (IPTV), a smart wearable device, etc.

The computer device may also include a network device and/or a user device. The network device includes, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud consisting of a large number of hosts or network servers based on cloud computing.

The network where the computer device is located includes but is not limited to the Internet, a wide area network, a metropolitan area network, a local area network, a virtual private network (VPN), etc.

The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

The step division of the above methods is only for clear description. When implemented, they can be combined into one step or some steps can be split and decomposed into multiple steps. As long as they include the same logical relationship, they are all within the protection scope of this patent; adding insignificant modifications to the algorithm or process or introducing insignificant designs without changing the core design of the algorithm and process are all within the protection scope of this application.

The description of "specific examples" or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of this specification. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

Those skilled in the art will readily appreciate other embodiments of the specification after considering the specification and practicing the invention claimed herein. The specification is intended to cover any variations, uses or adaptations of the specification that follow the general principles of the specification and include common knowledge or customary techniques in the art that are not claimed in the specification. The specification and examples are to be considered exemplary only, and the true scope and spirit of the specification are indicated by the following claims.

It should be understood that the present description is not limited to the precise structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present description is limited only by the appended claims.

The above description is only a preferred embodiment of this specification and is not intended to limit this specification. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of this specification should be included in the scope of protection of this specification.

Claims (12)

1. A language model-based reply content generation method, comprising:

acquiring current query data of a user, and inquiring target knowledge associated with the current query data from a preset knowledge base based on the current query data, wherein the target knowledge comprises first knowledge matched with the current query data and second knowledge in hierarchical relationship with the first knowledge, and source data used for constructing the knowledge base comprises a plurality of pieces of knowledge in hierarchical relationship;

Generating prompt information containing the current query data and the target knowledge;

Inputting the prompt information into a pre-training-based language model, and generating reply content corresponding to the prompt information by the language model.

2. The method of claim 1, the knowledge base storing the plurality of knowledge in a hierarchical relationship in a tree structure;

The querying the target knowledge associated with the current query data from a preset knowledge base comprises the following steps:

Querying a first node from the tree structure that stores first knowledge matching the current query data;

and acquiring a second node which has the same father node as the first node from the tree structure, and acquiring second knowledge stored by the second node.

3. The method of claim 1 or 2, the knowledge base storing the plurality of knowledge in a hierarchical relationship in a tree structure;

The querying the target knowledge associated with the current query data from a preset knowledge base comprises the following steps:

Querying a first node from the tree structure that stores first knowledge matching the current query data;

and acquiring a third node serving as a father node of the first node from the tree structure, and acquiring third knowledge stored by the third node, and/or acquiring a fourth node serving as a child node of the first node from the tree structure, and acquiring fourth knowledge stored by the fourth node.

4. The method of claim 1, the source data comprising a document; the knowledge base is constructed by:

identifying the text in the document, and converting the identified text into data in a Markdown format;

Constructing a tree structure based on the data in the MarkDown format to obtain a knowledge base containing the tree structure;

and/or, at least one piece of knowledge in the knowledge base is associated with standard time information, the querying, from a preset knowledge base, the target knowledge associated with the current query data comprises:

Identifying time-related text in the current query data and converting the time-related text into standard time information;

inquiring the target knowledge based on the current inquiry data and the standard time information obtained by conversion;

And/or, at least one piece of knowledge in the knowledge base is associated with named entity information, the querying, from a preset knowledge base, the target knowledge associated with the current query data comprises:

and identifying naming information related to the naming entity in the current query data, and querying the target knowledge based on the current query data and the naming information.

5. The method of claim 1, the querying target knowledge associated with the current query data from a preset knowledge base, comprising:

And querying first knowledge associated with the current query data in a vector similarity query mode from the knowledge base, wherein the vector similarity of the first knowledge and the current query data is greater than or equal to a first threshold value.

6. The method of claim 5, wherein querying the target knowledge associated with the current query data from a preset knowledge base comprises:

And respectively inquiring the first knowledge from the knowledge base at least in a vector similarity inquiring mode and a character level similarity inquiring mode, wherein the vector similarity of the first knowledge and the current inquiring data is larger than or equal to the first threshold value, and the character level similarity of the first knowledge and the current inquiring data is larger than or equal to the second threshold value.

7. The method of claim 5, wherein querying the target knowledge associated with the current query data from a preset knowledge base, further comprises:

identifying keywords contained in the current query data;

And inquiring first knowledge matched with the keywords contained in the current inquiry data in the knowledge base in a keyword regular matching inquiry mode.

8. The method of claim 1, the generating a hint message containing the current query data and the target knowledge, comprising:

Based on the target knowledge, acquiring identity information corresponding to the language model, wherein the identity information is used for indicating the language model to answer with a preset identity;

And generating prompt information containing the current query data, the target knowledge and the identity information.

9. The method of claim 8, the generating a hint message containing the current query data, the target knowledge, and the identity information, comprising:

Generating prompt information comprising the current query data, the target knowledge, the identity information and preset limit answer information, wherein the limit answer information is used for indicating the language model to answer with an answer range characterized by the target knowledge.

10. The method of claim 1, the training process of the language model comprising at least:

Acquiring historical session data, and respectively generating data in different description modes aiming at user inquiry data and reply data corresponding to the user inquiry data in the historical session data to obtain a training data set;

the language model is retrained with at least the training dataset.

11. A language model-based reply content generation apparatus, the apparatus comprising:

A query module for: acquiring current query data of a user, and inquiring target knowledge associated with the current query data from a preset knowledge base based on the current query data, wherein the target knowledge comprises first knowledge matched with the current query data and second knowledge in hierarchical relationship with the first knowledge, and source data used for constructing the knowledge base comprises a plurality of pieces of knowledge in hierarchical relationship;

A generation module for: generating prompt information containing the current query data and the target knowledge; inputting the prompt information into a pre-training-based language model, and generating reply content corresponding to the prompt information by the language model.

12. The apparatus of claim 11, the knowledge base storing the plurality of knowledge in a tree structure; the query module is further configured to: querying a first node from the tree structure that stores first knowledge matching the current query data; acquiring a second node with the same father node as the first node from the tree structure, and acquiring second knowledge stored by the second node;

And/or, the query module is further configured to: querying a first node from the tree structure that stores first knowledge matching the current query data; acquiring a third node serving as a father node of the first node from the tree structure, acquiring third knowledge stored by the third node, acquiring a fourth node serving as a child node of the first node from the tree structure, and acquiring fourth knowledge stored by the fourth node;

And/or, the source data comprises a document; the knowledge base is constructed by: identifying the text in the document, and converting the identified text into data in a Markdown format; constructing a tree structure based on the data in the MarkDown format to obtain a knowledge base containing the tree structure;

And/or, at least one piece of knowledge in the knowledge base is associated with standard time information, and the query module is further configured to: identifying time-related text in the current query data and converting the time-related text into standard time information; inquiring the target knowledge based on the current inquiry data and the standard time information obtained by conversion;

And/or, at least one piece of knowledge in the knowledge base is associated with named entity information, and the query module is further used for: identifying naming information related to a naming entity in the current query data, and inquiring the target knowledge based on the current query data and the naming information;

And/or, the query module is further configured to: and querying first knowledge associated with the current query data in a vector similarity query mode from the knowledge base, wherein the vector similarity of the first knowledge and the current query data is greater than or equal to a first threshold value.