A problem processing method, apparatus and device

Abstract

The application provides a question processing method, device and equipment. The method comprises the following steps: if a user input text question obtained includes a to-be-completed abbreviation, M first full names corresponding to the to-be-completed abbreviation and business features corresponding to each first full name are obtained by querying a search database; if M is greater than 1, N candidate first full names are selected from the M first full names based on the business features corresponding to the to-be-completed abbreviation and the business features corresponding to each first full name, wherein N is a positive integer; a target full name is selected from the N candidate first full names; the target full name is used to replace the to-be-completed abbreviation in the user input text question, so as to obtain a target text question; and a question answer corresponding to the target text question is determined. Through the technical scheme, the user intention can be accurately recognized, an accurate and reliable question answer can be obtained, and the calculation resources can be saved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a problem-solving method, apparatus, and device. Background Technology

[0002] An intelligent agent is an artificial intelligence entity designed for a specific industry. It perceives environmental information, autonomously decides actions based on these perceptions, and executes those actions to achieve pre-defined goals. Intelligent agents can be implemented as software programs or hardware. They are deeply integrated with industry-specific expertise and tools, providing customized problem-solving capabilities. Intelligent agents are widely used in industries such as transportation and customer service, processing user queries through integrated professional knowledge, such as customer service dialogues or data analysis tasks.

[0003] When an agent answers a current text question, it needs to input the current text question and historical text questions (i.e., dialogue history information, such as text questions from the previous 20 rounds) into a large model. The large model then answers the current text question based on the historical text questions and outputs the answer.

[0004] However, in the above approach, if the current text question includes abbreviations, the large model cannot support variant expressions, i.e., it cannot understand the intent of the abbreviations, thus resulting in the inability to output the correct answer to the question.

[0005] In summary, the variant representations in the current text question lead to biases in the large model's understanding of user intent, resulting in an inaccurate and unreliable answer—that is, an incorrect answer. Furthermore, the large model consumes significant computational resources to process these variant representations, wasting resources and exhibiting poor performance. Summary of the Invention

[0006] This application provides a problem-solving method, the method comprising: If the user input text question includes an abbreviation to be completed, then by querying the search database, M first full names corresponding to the abbreviation to be completed and the business features corresponding to each first full name are obtained; If M is greater than 1, then based on the business characteristics corresponding to the abbreviation to be completed and the business characteristics corresponding to each first full name, N candidate first full names are selected from the M first full names, where N is a positive integer; Select the target full name from the N candidate first full names; replace the abbreviation to be completed in the user input text question with the target full name to obtain the target text question; Determine the answer to the question corresponding to the target text question.

[0007] This application provides a problem-solving apparatus, the apparatus comprising: The query module is used to search the database by querying the database if the acquired user input text question includes an abbreviation to be completed, and to obtain M first full names corresponding to the abbreviation to be completed and the business features corresponding to each first full name; The acquisition module is used to select N candidate first full names from the M first full names, where N is a positive integer, based on the business characteristics corresponding to the abbreviation to be completed and the business characteristics corresponding to each first full name, if M is greater than 1; and to select the target full name from the N candidate first full names. The processing module is used to replace the abbreviation to be completed in the user input text question with the full name of the target to obtain the target text question; and to determine the question answer corresponding to the target text question.

[0008] This application provides an electronic device, including: a processor and a machine-readable storage medium, the machine-readable storage medium storing machine-executable instructions executable by the processor; the processor is used to execute the machine-executable instructions to implement the problem-solving method executable in the example above of this application.

[0009] This application provides a computer program product, which includes a computer program that, when executed by a processor, implements the problem-solving method described above in this application.

[0010] This application provides a machine-readable storage medium storing machine-executable instructions that can be executed by a processor; wherein the processor is used to execute the machine-executable instructions to implement the problem-solving method described in the above example of this application.

[0011] As can be seen from the above technical solutions, in this embodiment, if the user input text question includes an abbreviation to be completed, then M first full names and their corresponding business features are obtained by querying the search database. Based on the business features corresponding to the abbreviation to be completed and the business features corresponding to each first full name, N candidate first full names are selected from the M first full names. The target full name is selected from the N candidate first full names, and the abbreviation to be completed is replaced by the target full name to obtain the target text question (the target text question is used as the current text question). The answer to the target text question is then determined. In this way, the abbreviation to be completed can be replaced with the target full name. Even if the large model cannot support variant expressions, it can still accurately understand the intent and output the correct question answer. By inputting the target text question into the large model, the large model can accurately identify the user's intent and obtain an accurate and reliable question answer. Moreover, the large model does not need to consume a lot of computing resources to process variant expressions in the target text question, saving computing resources and achieving good processing performance. Attached Figure Description

[0012] Figure 1This is a flowchart illustrating a problem-solving method in one embodiment of this application; Figure 2 This is a schematic diagram of a large-scale user problem processing system according to one embodiment of this application; Figure 3 This is a flowchart illustrating the noun matching process in one embodiment of this application; Figure 4 This is a schematic diagram of the processing procedure of the noun matching layer in one embodiment of this application; Figure 5 This is a schematic diagram of the structure of a problem processing device according to one embodiment of this application; Figure 6 This is a hardware structure diagram of an electronic device according to one embodiment of this application. Detailed Implementation

[0013] This application proposes a problem-solving method that can be applied to electronic devices. The electronic device can be any device supporting intelligent agents, or any device supporting large models, such as personal computers, laptops, smartphones, IoT devices, cloud devices, servers, etc. There is no limitation on the type of electronic device. See also... Figure 1 The diagram shown illustrates a problem-solving method, which includes: Step 101: If the user input text question includes an abbreviation to be completed, then query the database to obtain the M first full names corresponding to the abbreviation to be completed and the business features corresponding to each first full name.

[0014] Step 102: If M is greater than 1, then based on the business characteristics corresponding to the abbreviation to be completed and the business characteristics corresponding to each first full name, select N candidate first full names from the M first full names, where N is a positive integer.

[0015] Step 103: Select the target full name from the N candidate first full names, and replace the abbreviation to be completed in the user input text question with the target full name to obtain the target text question.

[0016] Step 104: Determine the answer to the question corresponding to the target text question.

[0017] In one example, the target full name is selected from N candidate first full names, including but not limited to: if N is greater than 1, the user-input text question and the N candidate first full names are input into a large model to obtain the initial confidence score of each candidate first full name. For each candidate first full name, a screening score is determined based on its initial confidence score and its dynamic weight; the dynamic weight is positively correlated with the number of successful matches, which represents the number of times the candidate first full name is used as the target full name. The target full name is then selected from the N candidate first full names based on the screening score of each candidate first full name.

[0018] In one example, determining the screening score of the candidate first full name based on its initial confidence level and dynamic weight can include, but is not limited to: adjusting the initial confidence level of the candidate first full name based on configured calibration coefficients and entity discrimination to obtain an adjusted confidence level; determining the screening score of the candidate first full name based on the adjusted confidence level and dynamic weight; wherein, the entity discrimination can be determined based on the largest and second-largest initial confidence levels among the N initial confidence levels corresponding to the N candidate first full names.

[0019] Among them, the screening score of the candidate first full name can be positively correlated with the adjusted confidence level; the screening score of the candidate first full name can be positively correlated with the dynamic weight of the candidate first full name.

[0020] In one example, after selecting N candidate first full names from M first full names, the recall rate can be determined based on the N initial confidence scores corresponding to the N candidate first full names. Specifically, the number of initial confidence scores greater than the confidence threshold is counted, and the ratio of this number to N is determined as the recall rate. If the recall rate is greater than the configured recall rate threshold, the operation of selecting the target full name from the N candidate first full names is performed.

[0021] In one example, after determining the recall rate based on N initial confidence scores corresponding to N candidate first full names, if the recall rate is not greater than the recall rate threshold, the user input text question can be input into a conditional random field model to obtain entity features of multiple characters in the user input text question. Based on the entity features of multiple characters, characters are selected from the multiple characters as the abbreviation to be completed through the conditional random field model, and the abbreviation to be completed is input into the large model. The large model determines the second full name corresponding to the abbreviation to be completed. The second full name is used to replace the abbreviation to be completed in the user input text question to obtain the target text question.

[0022] In one example, entity features may include, but are not limited to, at least one of character-level features, lexical-level features, semantic-level features, and positional features; wherein, character-level features may include features of the current character, the characters preceding the current character, and the characters following the current character; lexical-level features may include part-of-speech tagging features and / or dictionary matching features of the current character; semantic-level features may include semantic features of the current character; and positional features may include features of the relative position of the current character in the user input text question.

[0023] As can be seen from the above technical solutions, in this embodiment, if the user input text question includes an abbreviation to be completed, then M first full names and their corresponding business features are obtained by querying the search database. Based on the business features corresponding to the abbreviation to be completed and the business features corresponding to each first full name, N candidate first full names are selected from the M first full names. The target full name is selected from the N candidate first full names, and the abbreviation to be completed is replaced by the target full name to obtain the target text question (the target text question is used as the current text question). The answer to the target text question is then determined. In this way, the abbreviation to be completed can be replaced with the target full name. Even if the large model cannot support variant expressions, it can still accurately understand the intent and output the correct question answer. By inputting the target text question into the large model, the large model can accurately identify the user's intent and obtain an accurate and reliable question answer. Moreover, the large model does not need to consume a lot of computing resources to process variant expressions in the target text question, saving computing resources and achieving good processing performance.

[0024] The technical solutions described above in the embodiments of this application will be explained below in conjunction with specific application scenarios.

[0025] When answering a current text question, both the current and historical text questions need to be input into a large model. The large model then answers the question based on these two text questions and outputs the answer. However, in this method, the current text question input to the large model contains abbreviations (i.e., variant expressions), leading to a bias in the large model's understanding of the user's intent and resulting in an inaccurate and unreliable answer. Furthermore, the large model consumes significant computational resources to process these variant expressions, wasting hardware resources and resulting in poor processing performance.

[0026] The large model adopts an end-to-end natural language processing framework, which has the following problems when facing complex business scenarios: Insufficient entity retrieval accuracy: The large model relies on fuzzy text matching, which makes it difficult to handle the precise entity query requirements in business scenarios. It cannot support variant representations and is prone to missed detections due to the lack of entity normalization, resulting in the loss of key information. In other words, the large model cannot pay attention to variant representations when processing problems (abbreviation).

[0027] To address the aforementioned findings, this application proposes a problem-solving method that leverages large models (such as pre-trained language models based on the Transformer architecture, possessing natural language understanding and generation capabilities) to achieve entity recognition and entity normalization within large-model dialogue systems. This method can be applied to scenarios such as intelligent customer service and knowledge-based question answering, as well as professional business scenarios (i.e., scenarios requiring high-precision entity recognition) such as bidding, financial consulting, and government affairs Q&A. This embodiment's problem-solving method improves upon the following issue: Missed detection due to entity variants: The inability to effectively handle entity variant expressions in business scenarios (such as company abbreviations, project aliases, etc.) leads to missed detection of key information. Entity normalization is the process of mapping entity variant expressions in business scenarios to standard entities, such as mapping abbreviations to full names. Entity normalization is used to solve the missed detection problem caused by entity variant expressions.

[0028] See Figure 2 As shown in the embodiments of this application, a four-layer collaborative large-model user question processing system is proposed. This system can include a noun matching layer, a multi-turn dialogue layer, a long-term memory layer, and a path planning layer. The noun matching layer implements the noun matching process, enabling standardization and accurate retrieval of business entities. This solves the problem of insufficient entity retrieval accuracy and effectively handles variant representations of entities in business scenarios (such as company abbreviations and project aliases), avoiding the omission of key information. The multi-turn dialogue layer implements the multi-turn dialogue process, ensuring the continuity of long-term dialogue context and solving the problem of continuity breaks in multi-turn dialogues. It retains early key parameters in long dialogues, preventing interruptions in business processing flows. The long-term memory layer implements the long-term memory process, providing enhanced memory storage and retrieval of business metadata. This solves the problem of memory being disconnected from business scenarios and binds long-term memory to business parameters, resulting in high relevance of historical information retrieval. The path planning layer implements the path planning process, dynamically mapping intent recognition to business processes. This solves the problem of intent routing mismatch with business topology and adapts to dynamic business links.

[0029] The noun matching layer, multi-turn dialogue layer, long-term memory layer, and path planning layer can achieve context pass-through through cross-layer state synchronization interfaces, forming a closed-loop collaborative mechanism. For the noun matching process, noun matching refers to the process of identifying and standardizing business entities (such as company names and project numbers) from user input. For the multi-turn dialogue process, it refers to a continuous dialogue management mechanism spanning multiple interaction rounds, requiring the maintenance of context consistency. For the long-term memory process, it refers to the persistent storage and efficient retrieval of historical dialogues and business parameters, transcending the limitations of a large model context window. For the path planning process, it refers to the ability to dynamically select and execute multiple business processing flows based on user intent.

[0030] In this embodiment, the processing of the noun matching layer is described. The processing of the multi-turn dialogue layer, long-term memory layer and path planning layer will not be described in this embodiment.

[0031] For example, in the noun matching layer, to address the issue of missed detections caused by entity variants, a dual-channel entity parsing architecture is adopted. Dual-channel entity parsing is a hybrid retrieval mechanism consisting of an exact matching channel and a semantic normalization channel, which work together.

[0032] See Figure 2 As shown, the dual-channel entity parsing architecture refers to the noun matching layer employing an exact matching channel and a semantic normalization channel, using a collaborative strategy that prioritizes exact matching and supplements it with semantic normalization. A dynamic switching mechanism connects the two retrieval channels. For example, the exact matching channel acts as the primary channel, responsible for quickly processing entity queries with standard expressions. When the recall rate is higher than a threshold (e.g., 0.3), the exact matching result from the exact matching channel is directly used to ensure query efficiency. The semantic normalization channel acts as an auxiliary channel, automatically activated when the recall rate is not higher than the threshold. The semantic normalization channel maps entity variant expressions to standard entities through a large model, resolving the issue of missed detections of variant expressions.

[0033] The noun matching layer can also include an index enhancement channel, which acts as a feedback channel, writing the normalized results back to the search database with dynamic weights, forming a closed-loop learning mechanism of "identification-normalization-enhancement". The channel switching mechanism adopts a soft switching strategy. When the recall rate is close to the threshold, two channels are executed in parallel, and the optimal result is selected through a confidence calibration function to avoid performance fluctuations caused by hard switching.

[0034] Through the collaborative working mechanism of the dual-channel entity parsing architecture, it is ensured that while maintaining high query efficiency, it can effectively handle entity variant representations, achieving the best balance between retrieval accuracy and query efficiency.

[0035] For the noun matching layer, abbreviation completion can be performed on the user-input text question to obtain the corresponding target text question. For example, noun matching can be performed on the user-input text question to obtain the target text question, which can be the current text question. The target text question can then be processed to obtain the answer. Alternatively, the target text question and historical text questions can be input into a larger model, which can then process the target text question based on the historical text questions to obtain the corresponding answer. There are no restrictions on this processing method. For example, see [link to relevant documentation]. Figure 3 The diagram shown illustrates the noun matching process, which may include the following steps: Step 301: Obtain the user input text question. The user input text question is the question entered by the user.

[0036] For example, when a user inputs a question, the question could be in text format, which we'll call the user-input text question. Alternatively, the question could be in speech format, which we can convert to text format and call the text question. Or, the question could be in image format, which we can convert to text format and call the text question. Of course, these are just a few examples; the key is to obtain the user-input text question.

[0037] For example, after receiving the user's input text question, a word segmentation operation can be performed on the user's input text question to obtain multiple words (also called characters or keywords) of the user's input text question. There are no restrictions on this word segmentation process. If there are abbreviations to be completed among the multiple words, then step 302 is executed; if there are no abbreviations to be completed among the multiple words, then the user's input text question is taken as the target text question, and the noun matching process ends. Alternatively, the user's input text question can be directly taken as the target text question, and the noun matching process can end, without determining whether there are abbreviations to be completed among the multiple words. This process will not be elaborated further.

[0038] Step 302: If the user input text question includes an abbreviation to be completed, then query the database to obtain M first full names corresponding to the abbreviation to be completed, as well as the business features corresponding to each first full name.

[0039] In one example, a search database can be pre-maintained, including a mapping between abbreviations and full names. For instance, users can configure this mapping in the search database. Alternatively, an algorithm can be used to obtain the mapping, and this mapping can be configured in the search database; there are no restrictions on this approach.

[0040] When configuring the mapping between abbreviations and full names, one abbreviation may correspond to one full name or multiple full names (i.e., each full name can be used for the abbreviation). Furthermore, when configuring the mapping between abbreviations and full names, for each full name, you can also configure the corresponding business characteristics (structured business characteristics), such as project type, amount, time, etc. There are no restrictions on these business characteristics; they can be configured according to actual needs. Business characteristics indicate what kind of business the full name refers to, such as a full name for a project type, a full name for an amount, or a full name for a time.

[0041] For example, information corresponding to business characteristics can also be called business parameters. Business parameters are parameters that are of critical significance in a specific business scenario, such as tender number, project type, amount range, and time point. Business parameters can be a key focus, and it is necessary to ensure that business parameters are not lost during long conversations.

[0042] In one example, after obtaining the abbreviation to be completed, the database can be searched using the abbreviation to obtain M first full names corresponding to the abbreviation and the business characteristics corresponding to each first full name, where M can be a positive integer. If the M first full names are a single first full name, then this first full name is directly used as the target full name, and the subsequent operation of selecting the target full name is not performed. If M is greater than 1, that is, the M first full names are multiple first full names, then the subsequent operation of selecting the target full name is performed, and step 303 is executed.

[0043] Furthermore, if the search database does not contain a first full name corresponding to the abbreviation to be completed, then proceed directly to the next step 310, that is, determine the full name corresponding to the abbreviation to be completed through the large model.

[0044] Step 303: If M is greater than 1, then based on the business characteristics corresponding to the abbreviation to be completed and the business characteristics corresponding to each first full name, select N candidate first full names from the M first full names, where N is a positive integer.

[0045] In one example, the business characteristics (which can be one or more) corresponding to the abbreviation to be completed can be determined. For instance, by analyzing the user's input text question to obtain the user intent, the business characteristics corresponding to that user intent can be used as the business characteristics corresponding to the abbreviation to be completed. For example, if the user intent is a bidding intent, the business characteristics corresponding to the bidding intent could include project type, amount, and time.

[0046] In this case, when configuring the business characteristics corresponding to the full name in the search database, the business characteristics corresponding to the full name can also be configured based on the user's intent. For example, if a certain full name is for bidding intent, then the business characteristics corresponding to that full name are the business characteristics corresponding to the bidding intent.

[0047] In one example, the business characteristics corresponding to the abbreviation to be completed can be determined. For instance, by analyzing the user's input text question, the business characteristics of the user's input text question can be obtained, and these business characteristics can be used as the business characteristics corresponding to the abbreviation to be completed. For example, if the user's input text question includes item type, amount, and time, then the business characteristics corresponding to the abbreviation to be completed can include item type, amount, and time.

[0048] In one example, based on the business characteristics corresponding to the abbreviation to be completed and the business characteristics corresponding to each first full name, if the business characteristics corresponding to the first full name match the business characteristics corresponding to the abbreviation to be completed, then the first full name is considered a candidate first full name; if the business characteristics corresponding to the first full name do not match the business characteristics corresponding to the abbreviation to be completed, then the first full name is not considered a candidate first full name. Based on the above processing, N candidate first full names can be selected from M first full names, where N is a positive integer and N is not greater than M.

[0049] For example, if the business characteristics corresponding to the abbreviation to be completed include project type, amount, and time, then if the business characteristic corresponding to the first full name is project type, the business characteristics corresponding to the first full name match the business characteristics corresponding to the abbreviation to be completed, and the first full name is taken as a candidate first full name, and so on.

[0050] In summary, N candidate first full names can be selected from M candidate first full names, where N can be a positive integer. If the N candidate first full names are a single candidate first full name, then this single candidate first full name is directly used as the target first full name, and the subsequent operation of selecting the target full name is not performed. If N is greater than 1, that is, the N candidate first full names are multiple candidate first full names, then the subsequent operation of selecting the target full name is performed, and step 304 is executed.

[0051] Furthermore, if there is no candidate first full name that meets the requirements among the M first full names (i.e., N is 0), then proceed directly to the next step 310, that is, determine the full name corresponding to the abbreviation to be completed through the large model.

[0052] In the above process, the process of querying the database to obtain M first full names corresponding to the abbreviation to be completed, as well as the business characteristics corresponding to each first full name, can also be called the primary index. The primary index is used for database query operations. The process of selecting N candidate first full names from the M first full names based on the business characteristics corresponding to the abbreviation to be completed and the business characteristics corresponding to each first full name can also be called the secondary index. The secondary index is used for indexing operations based on key business fields (business characteristics).

[0053] Step 304: If N is greater than 1, input the user-input text question and N candidate first full names into the large model to obtain the initial confidence score of each candidate first full name. The initial confidence score of a candidate first full name represents the degree of confidence in its association with the user-input text question.

[0054] In one example, the user inputs a text question and candidate first full names into a large model. The model analyzes the degree of matching (i.e., correlation) between the candidate first full names and the user input text question, and outputs the initial confidence score of the candidate first full name. A higher initial confidence score indicates a stronger match between the candidate first full name and the user input text question, and a higher probability that the candidate first full name is the target full name. Conversely, a lower initial confidence score indicates a weaker match between the candidate first full name and the user input text question, and a lower probability that the candidate first full name is the target full name.

[0055] Step 305: For each candidate first full name, determine the adjusted confidence level of the candidate first full name based on its initial confidence level, the configured calibration coefficient, and the entity discrimination.

[0056] For example, based on the configured calibration coefficients and entity discrimination, the initial confidence of the candidate first full name is adjusted to obtain the adjusted confidence of the candidate first full name. The adjusted confidence can be understood as the optimized confidence, also known as the disambiguation confidence. The entity discrimination is determined based on the largest and second-largest initial confidence among the N initial confidences corresponding to the N candidate first full names.

[0057] For example, the adjusted confidence score of a candidate first full name can represent the degree of confidence in the association between the candidate first full name and the user's input text question; it is a normalized entity recognition confidence score. For instance, the initial confidence score may contain errors. By optimizing the initial confidence score to obtain the adjusted confidence score, it can accurately reflect the degree of confidence in the association between the candidate first full name and the user's input text question.

[0058] The calibration coefficient can be configured according to actual needs. It can be a value in the range [0, 1]. The calibration coefficient achieves the best balance between recall and precision on the test set, and there are no restrictions on it, such as 0.2.

[0059] Entity discriminability, also known as the discriminability between candidate entities (candidate first universal names), is determined when a large model outputs N initial confidence scores for N candidate first universal names. The entity discriminability can be determined based on the highest and second-highest initial confidence scores among these N initial confidence scores. For example, the entity discriminability can be calculated using the following formula: context_ambiguity = 1 - (max_similarity - second_max_similarity). In this formula, context_ambiguity represents the entity discriminability, max_similarity represents the highest initial confidence score among the N initial confidence scores, and second_max_similarity represents the second-highest initial confidence score among the N initial confidence scores. A higher entity discriminability value indicates lower discriminability between candidate entities.

[0060] Based on the initial confidence level, calibration coefficient, and entity discrimination, the adjusted confidence level can be determined using the following formula: calibrated_confidence = base_confidence × (1 - λ × context_ambiguity). calibrated_confidence represents the adjusted confidence level of the candidate first full name, base_confidence represents the initial confidence level of the candidate first full name, and λ represents the calibration coefficient, such as 0.2.

[0061] Step 306: For each candidate first full name, determine the screening score of the candidate first full name based on the adjusted confidence level and the dynamic weight of the candidate first full name.

[0062] For example, the screening score for the candidate's first full name can be positively correlated with the adjusted confidence level, and the screening score for the candidate's first full name can be positively correlated with the dynamic weight. For example, the screening score can be determined based on the product of the adjusted confidence level and the dynamic weight, or the adjusted confidence level and the dynamic weight can be weighted to obtain the screening score; there are no restrictions on this.

[0063] In one example, the dynamic weight for the candidate first full name represents the probability that the candidate first full name will be selected as the target full name. That is, the larger the dynamic weight, the greater the probability that the candidate first full name will be selected as the target full name. When storing the candidate first full name in the search database, the initial value of the dynamic weight is a fixed value, such as 0.6. In subsequent use, the dynamic weight can be adjusted, for example, based on the number of successful matches. Therefore, the dynamic weight is positively correlated with the number of successful matches, and the number of successful matches represents the number of times the candidate first full name has been used as the second full name.

[0064] For example, the initial dynamic weight of the candidate first full name is 0.6, and the dynamic weight is determined by the following formula: w = 0.6 + 0.4 × (1 - 1 / (1+n)). In the above formula, n represents the number of successful matches. Obviously, when the number of successful matches is 0, the dynamic weight of the candidate first full name is 0.6; when the number of successful matches is 1, the dynamic weight of the candidate first full name is 0.8; when the number of successful matches is 3, the dynamic weight of the candidate first full name is 0.9, and so on. As the number of successful matches increases, the dynamic weight also increases.

[0065] Step 307: Based on the screening score of each candidate first full name, select the target full name from the N candidate first full names. In this way, the target full name can be selected from the N candidate first full names.

[0066] For example, based on the screening score of each candidate first full name, the candidate first full name with the highest screening score can be used as the target full name. Another example: based on the screening score of each candidate first full name, if the highest screening score is greater than a threshold t1, then the candidate first full name with the highest screening score is used as the target full name. If the highest screening score is not greater than the threshold t1, then the difference between the highest screening score and the second highest screening score is calculated. If this difference is greater than a threshold t2, then the candidate first full name with the highest screening score is used as the target full name. If this difference is not greater than the threshold t2, then the user is guided to select the target full name from N candidate first full names. The above are just examples; the ability to select the target full name from N candidate first full names is sufficient, and there are no restrictions on this.

[0067] Step 308: Determine the recall rate based on the N initial confidence scores corresponding to the N candidate first full names, or determine the recall rate based on the N adjusted confidence scores corresponding to the N candidate first full names. The implementation methods are similar; taking the determination of recall rate based on N initial confidence scores as an example. Based on this, determine whether the recall rate is higher than the recall rate threshold. If yes, proceed to step 309; otherwise, proceed to step 310.

[0068] In one example, before step 305, it can be determined whether the recall rate is higher than the recall rate threshold. If so, steps 305-307 are executed; otherwise, steps 305-307 are not executed.

[0069] To determine recall, we can count the number of initial confidence levels (adjusted confidence levels) greater than a certain confidence threshold (configurable according to actual needs). The ratio of this number to the total number of initial confidence levels N can be used as the recall rate. For example, if 2 out of 4 initial confidence levels are greater than the confidence threshold, the recall rate is 0.5; if 1 initial confidence level is greater than the confidence threshold, the recall rate is 0.25. The recall threshold can be configured according to actual needs, such as 0.3.

[0070] Step 309: Replace the abbreviation to be completed in the user input text question with the full name of the target to obtain the target text question. That is, the target text question includes the full name of the target, not the abbreviation to be completed.

[0071] If the search database also includes attribute information corresponding to the full name of the target, such as credit code, then this attribute information can be added to the user's input text question (either before or after the full name of the target) to obtain the target text question, meaning the target text question also includes this attribute information.

[0072] In summary, the abbreviation to be completed in the user-input text question is replaced with the full name of the target and the credit code, resulting in the target text question. That is, the target text question can include the full name of the target and the credit code.

[0073] After obtaining the target text question, it can be input into a large model. The model will process the target text question and output the corresponding answer. For example, the target text question and historical text questions can be input into the model. Historical text questions can be text questions preceding the target text question and can include multiple historical fragments. The large model outputs the answer based on the target text question and historical text questions, without imposing any restrictions on the question processing procedure.

[0074] Step 310: Input the user input text question into the Conditional Random Field (CRF) model, obtain the entity features of multiple characters in the user input text question through the CRF model, and select characters from the multiple characters as the abbreviation to be completed through the CRF model based on the entity features of multiple characters.

[0075] In one example, the CRF (Conditional Random Field) model is a statistical modeling method for sequence labeling. In this embodiment, the CRF model can be used for entity boundary detection, that is, to accurately locate the abbreviation to be completed from the user-input text question. For example, the CRF model can be optimized using the L-BFGS (Limited memory Broyden Fletcher Goldfarb Shanno, an efficient algorithm for unconstrained optimization problems). The model complexity can be controlled by a regularization parameter (such as C=0.1) to prevent overfitting. There are no restrictions on this CRF model.

[0076] After inputting the user-input text question into a Conditional Random Field (CRF) model, the CRF model can obtain entity features of multiple characters (i.e., words or keywords) in the user-input text question. Based on these entity features, the CRF model can select one or more characters from the multiple characters as the abbreviation to be completed, and there are no restrictions on the selection process of this abbreviation.

[0077] For each character (denoted as the current character), the entity features of the current character may include, but are not limited to, at least one of character-level features, lexical-level features, semantic-level features, and positional features.

[0078] The character-level features of the current character are the features of the current character, the characters preceding the current character, and the characters following the current character. For example, based on the current character, the two characters preceding the current character, and the two characters following the current character (a total of 5 character windows), the character-level features are the type features of these characters (Chinese / number / symbol), that is, the features used to indicate whether these characters are Chinese, numbers, or symbols.

[0079] The lexical-level features of the current character can be part-of-speech tagging features (such as the part-of-speech tagging result) and / or dictionary matching features (such as the dictionary matching flag). For example, if the part-of-speech tagging result of the current character is determined based on a general word segmentation tool, this part-of-speech tagging result can be used as a lexical-level feature. Alternatively, if based on a domain-specific dictionary, it can be determined whether the current character matches the domain-specific dictionary, resulting in a dictionary matching flag, which indicates whether the current character matches or does not match the domain-specific dictionary.

[0080] The semantic-level features of the current character can be the semantic features of the current character. For example, the Conditional Random Field model can include BERT (Bidirectional Encoder Representations from Transformers), which extracts the semantic-level features of the current character.

[0081] The positional feature of the current character can be the feature of the relative position of the current character in the user input text question, that is, the relative position of the current character in the sentence, such as the start position, middle position, and end position.

[0082] Step 311: Input the abbreviation to be completed into the large model, and determine the second full name corresponding to the abbreviation to be completed through the large model; replace the abbreviation to be completed in the user input text question with the second full name to obtain the target text question, that is, the target text question includes the second full name, not the abbreviation to be completed.

[0083] For example, after determining the abbreviation to be completed using a conditional random field model, the abbreviation can be input into a larger model, which will then process it to obtain the full name (denoted as the second full name). Alternatively, after inputting the abbreviation into the larger model, the model can also output the second full name and its corresponding attribute information, such as a credit code.

[0084] Based on this, the abbreviation to be completed in the user input text question can be replaced with the second full name to obtain the target text question. Furthermore, if the large model also outputs attribute information corresponding to the abbreviation to be completed, such as a credit code, then this attribute information can be added to the user input text question (either before or after the second full name) to obtain the target text question.

[0085] After obtaining the target text question, it can be input into the large model, which will process the target text question and output the corresponding answer.

[0086] In one example, after obtaining the second full name corresponding to the abbreviation to be completed, the mapping relationship between the abbreviation to be completed and the second full name can be updated in the search database for subsequent queries of the full name corresponding to the abbreviation to be completed. Alternatively, the mapping relationship between the abbreviation to be completed, the second full name, and attribute information can be updated in the search database for subsequent queries of the full name and attribute information corresponding to the abbreviation to be completed.

[0087] The database can be updated with the mapping between the abbreviation to be completed and the second full name. It can also be updated with the business characteristics corresponding to the second full name, such as project type, amount, time, etc.

[0088] The system updates the mapping relationship between the abbreviation to be completed and the second full name in the search database. It can also update the dynamic weight corresponding to the second full name in the search database, such as the initial value of the dynamic weight being 0.6.

[0089] In one example, see Figure 4 The diagram illustrates the processing flow of the noun matching layer. The user-input text question undergoes primary and secondary indexing in the exact matching channel, and recall is calculated. The primary index involves querying the database to obtain M first full names corresponding to the abbreviation to be completed, along with the corresponding business features. The secondary index involves selecting N candidate first full names from the M first full names based on the business features corresponding to the abbreviation to be completed and the business features corresponding to each first full name. Calculating recall involves inputting the user-input text question and the N candidate first full names into the large model, obtaining the initial confidence score for each candidate first full name, and determining the recall rate based on the initial confidence scores of the N candidate first full names.

[0090] After obtaining the recall rate, it can be determined whether the recall rate is higher than the recall rate threshold.

[0091] If so, the target text question is determined based on the matching results of the exact matching channel. For example, for each candidate full name, the adjusted confidence of the candidate full name is determined based on its initial confidence, calibration coefficient, and entity discrimination; the screening score of the candidate full name is determined based on its adjusted confidence and dynamic weight; the target full name is selected from N candidate full names based on the screening score of each candidate full name; and the target text question is obtained by replacing the abbreviation to be completed in the user input text question with the target full name.

[0092] If not, the user-input text question undergoes entity normalization in the semantic normalization channel, mapping variant representations to standard entities. For example, entity features of multiple characters in the user-input text question are obtained through a Conditional Random Field (CRF) model. Based on these entity features, a character is selected from the multiple characters as the abbreviation to be completed using the CRF model. The second full name corresponding to the abbreviation to be completed is determined through a large model; the second full name is then used to replace the abbreviation to be completed in the user-input text question to obtain the target text question.

[0093] After the semantic normalization channel finishes processing, in the index enhancement channel, the mapping relationship between the abbreviation to be completed and the second full name is updated in the search database. The business features corresponding to the second full name are updated in the search database, and the dynamic weights corresponding to the second full name are updated in the search database.

[0094] In one example, at the noun matching layer, besides using a joint retrieval mechanism combining a search database and a large model, regular expressions and knowledge graphs can also be employed. Regular expressions are used to extract the abbreviations to be completed in the user's input text (e.g., company names such as "*Company" or "*Group"). Then, the extracted results are aligned with nodes in a pre-built knowledge graph (e.g., an enterprise relationship graph) to obtain the full name corresponding to the abbreviation to be completed. Finally, the full name is used to replace the abbreviation in the user's input text.

[0095] In one example, the system collaboration workflow can be as follows: User input text question: When was the bidding for XX Jiaotong's ZH-2024-038 project opened? At the noun matching layer: The recall rate of the exact matching channel is lower than 0.3, activating the semantic normalization channel; the CRF model identifies "XX Transportation Investment" as an entity (i.e., the abbreviation to be completed), and the large model normalizes "XX Transportation Investment" to "XX Province XXX Infrastructure Construction Group Co., Ltd.", where "XX" in the above example represents the omission of the actual words, and the large model outputs the unified social credit code as "91000000000000000B". Based on this, the large model can accurately return that "The ZH-2024-038 project was opened for bidding on March 15, 2024."

[0096] In an example, for the embodiment of the bidding scenario, it can show the specific application in the bidding scenario. The user inputs the text question: When was the ZH-2024-038 project of XX Transportation Investment opened for bidding? Processing at the noun matching layer: Exact matching channel: Query "XX Transportation Investment" based on the hierarchical index of the search database, and the recall rate is 0.25 (lower than the threshold of 0.3), triggering the semantic normalization channel.

[0097] Processing at the semantic normalization channel: The CRF model identifies "XX Transportation Investment" as an entity, and the feature extraction results are as follows: Character-level features: The first 2 characters are "X", "X", and the last 2 characters are "Jiao", "Tou", and the character types are all Chinese; Lexical-level features: The word segmentation result is ["XX", "Transportation Investment"], and the词性标注 is [ns, n]; Semantic-level features: The first 50 dimensions of the BERT embedding, and the vector values after PCA dimensionality reduction are [0.32, -0.15,..., 0.07].[[]] [[ID=]]

[0098] The large model performs entity normalization: Normalize "XX Transportation Investment" to "XX Province XXX Infrastructure Construction Group Co., Ltd.", and the large model outputs the unified social credit code as "91000000000000000B".

[0099] In the process of calculating the disambiguation confidence: base_confidence = 0.85, context_ambiguity = 0.32, calibrated_confidence (adjusted confidence) = 0.85×(1 - 0.2×0.32)=0.796.

[0100] The large model accurately returns that "The ZH-2024-038 project was opened for bidding on March 15, 2024."

[0101] In one example, a performance demonstration is presented for a 50-round dialogue scenario, simulating the complete process of a user conducting 50 rounds of bidding consultation. The dialogue scenario describes the user starting with a project inquiry and gradually delving into multiple stages, including qualification requirements, bidding procedures, and bid bond payment. Key business parameters include: project number (ZH-2024-038), project type (EPC), amount range (123.456 million yuan), and bid opening time (March 15, 2024). The first round inputs the project number, and the 50th round inquires, "How should the bid bond for the previously mentioned 123.456 million yuan EPC project be paid?". Based on this, the processing flow is as follows: System output: The large model accurately answers the question, "The deposit for EPC project ZH-2024-038 must be paid before March 10, 2024, and the amount is 2% of the contract price, i.e., RMB 2,469,120." Based on the above processing method, it can successfully identify and retain all key parameters, accurately answering the user's question.

[0102] In one example, the test dataset is constructed as follows: Business scenario data: Real bidding data from 2020-2024, obtained from the public service platform for bidding and tendering, includes 15,682 bidding projects and information on 28,431 companies. Dialogue simulation data: Based on real user consultation records, 5 domain experts simulated 50 rounds of dialogue scenarios, resulting in 2,350 dialogue samples. Evaluation metrics: Entity retrieval F1 score: Evaluates the accuracy and recall rate of entity recognition; 50-round dialogue completion rate: Evaluates the system's ability to retain key parameters in long dialogue scenarios; Long-tail question response time: Evaluates the efficiency in handling complex business questions; Key parameter retention rate: Evaluates the retention effect of key business parameters in multi-round dialogues. Business scenario comparison: Bidding and tendering scenario: Focuses on key parameters such as company name, bidding number, and amount; Financial compliance scenario: Focuses on company code, transaction amount, and time range; Government consultation scenario: Focuses on policy number, scope of application, and execution time. By comparing performance in different business scenarios, the universality of this embodiment is verified.

[0103] Comparative testing revealed that the technical solution in this embodiment significantly improves entity recognition accuracy: the F1 score for entity retrieval increased from 0.72 to 0.94, and the false negative rate decreased by 76.4%, effectively solving the problem of false negatives in variant representations. Performance breakthroughs in long dialogue scenarios: in a 50-turn dialogue scenario, task completeness reached 92%, a significant improvement over the fixed window solution (61% completion rate), with a key parameter retention rate of 92.7%. Response efficiency was greatly improved: the response time for long-tail questions was reduced from 8.3s to 3.1s, reducing user waiting time. Resource consumption was significantly reduced: through dynamic context compression, the consumption of large model tokens was reduced by 58.3%, lowering computational resource costs.

[0104] In this embodiment, a dual-channel entity parsing architecture is adopted in the noun matching layer, which cascades the exact matching channel and the semantic normalization channel to solve the problem of entity variant representation. An entity normalization protocol (such as a company name-unified credit code mapping table) is defined; a dynamic weight allocation strategy is constructed, with the initial weight w=0.6 for new mapping relationships, increasing by w=0.6+0.4×(1-1 / (1+n)) with the number of successful matches n; the calibration function for the adjusted confidence is: calibrated_confidence=base_confidence×(1-λ×context_ambiguity), λ=0.2.

[0105] A conditional random field entity boundary detection model is proposed, integrating character-level, lexical-level, and semantic-level features for multi-granularity entity recognition. The design incorporates a context window (two characters before and after the word segmentation), word segmentation results, and pre-trained word vectors, improving entity boundary recognition accuracy to 96.3%, a 9.7% improvement over single-feature models. A multi-round to single-round prompt word reconstruction technique is also proposed, based on a semantically preserved reconstruction mechanism using a key information pool. This retains key business parameters while maintaining semantic coherence, achieving a 98.2% retention rate of business intent for the reconstructed prompt words.

[0106] Based on the same concept as the above method, this application proposes a problem-solving apparatus, see [link to previous application]. Figure 5 The diagram shown is a structural schematic of the problem-solving device, which may include: Query module 51 is used to query the database through the abbreviation to be completed if the user input text question includes an abbreviation to be completed. This query module 51 obtains M first full names corresponding to the abbreviation to be completed and the business features corresponding to each first full name. The acquisition module 52 is used to select N candidate first full names from the M first full names, where N is a positive integer, based on the business characteristics corresponding to the abbreviation to be completed and the business characteristics corresponding to each first full name, if M is greater than 1; and select the target full name from the N candidate first full names. Processing module 53 is used to replace the abbreviation to be completed in the user input text question with the full name of the target to obtain the target text question; and to determine the question answer corresponding to the target text question.

[0107] In one example, when the acquisition module 52 selects the target full name from the N candidate first full names, it is specifically used to: if N is greater than 1, input the user input text question and the N candidate first full names into the large model to obtain the initial confidence of each candidate first full name; For each candidate full name, a screening score is determined based on its initial confidence level and dynamic weight. Specifically, the initial confidence level of the candidate full name is adjusted based on a configured calibration coefficient and entity discrimination ratio to obtain an adjusted confidence level. The screening score is then determined based on the adjusted confidence level and the dynamic weight of the candidate full name. The dynamic weight is positively correlated with the number of successful matches, where the number of successful matches represents the number of times the candidate full name is used as the target full name. The entity discrimination ratio is determined based on the maximum and second-largest initial confidence levels among the N initial confidence levels corresponding to the N candidate full names. The screening score of the candidate full name is positively correlated with the adjusted confidence level and the dynamic weight of the candidate full name. The target full name is selected from N candidate full names based on the screening score of each candidate full name.

[0108] In one example, the acquisition module 52 is further configured to determine the recall rate based on N initial confidence levels corresponding to the N candidate first full names after selecting N candidate first full names; wherein, the number of initial confidence levels greater than the confidence threshold is counted, and the ratio of this number to N is determined as the recall rate; if the recall rate is greater than the configured recall rate threshold, the acquisition module selects the target full name from the N candidate first full names; The acquisition module 52 is further configured to, if the recall rate is not greater than the recall rate threshold, input the user input text question into a conditional random field model to obtain entity features of multiple characters in the user input text question; based on the entity features of the multiple characters, select characters from the multiple characters as abbreviated names to be completed through the conditional random field model, and input the abbreviated names to be completed into a large model; and determine the second full name corresponding to the abbreviated names to be completed through the large model. The processing module 53 is further configured to replace the abbreviation to be completed in the user input text question with the second full name to obtain the target text question.

[0109] In one example, the entity features include at least one of character-level features, lexical-level features, semantic-level features, and positional features; wherein, the character-level features include features of the current character, the characters preceding the current character, and the characters following the current character; the lexical-level features include part-of-speech tagging features and / or dictionary matching features of the current character; the semantic-level features include semantic features of the current character; and the positional features include features of the relative position of the current character in the user-input text question.

[0110] Based on the same concept as the above method, this application proposes an electronic device, see [link to previous application]. Figure 6 As shown, the electronic device includes a processor 61 and a machine-readable storage medium 62, the machine-readable storage medium 62 storing machine-executable instructions that can be executed by the processor 61; the processor 61 is used to execute the machine-executable instructions to implement the problem-solving method disclosed in the above example of this application.

[0111] Based on the same application concept as the above method, this application embodiment also provides a machine-readable storage medium storing a plurality of computer instructions, which, when executed by a processor, can implement the problem-solving method disclosed in the above examples of this application.

[0112] The aforementioned machine-readable storage medium can be any electronic, magnetic, optical, or other physical storage device that can contain or store information, such as executable instructions, data, etc. For example, machine-readable storage media can be: RAM (Random Access Memory), volatile memory, non-volatile memory, flash memory, storage drives (such as hard disk drives), solid-state drives, any type of storage disk (such as optical discs, DVDs, etc.), or similar storage media, or combinations thereof.

[0113] Based on the same concept as the methods described above, this application also provides a computer program product, which may include a computer program. When executed by a processor, the computer program implements the problem-solving method disclosed in the examples above.

[0114] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, embodiments of this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0115] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A problem-solving method, characterized in that, The method includes: If the user input text question includes an abbreviation to be completed, then by querying the search database, M first full names corresponding to the abbreviation to be completed and the business features corresponding to each first full name are obtained; If M is greater than 1, then based on the business characteristics corresponding to the abbreviation to be completed and the business characteristics corresponding to each first full name, N candidate first full names are selected from the M first full names, where N is a positive integer; Select the target full name from the N candidate first full names; replace the abbreviation to be completed in the user input text question with the target full name to obtain the target text question; Determine the answer to the question corresponding to the target text question.

2. The method according to claim 1, characterized in that, The step of selecting the target full name from the N candidate first full names includes: If N is greater than 1, then input the user-input text question and the N candidate first full names into the large model to obtain the initial confidence of each candidate first full name; For each candidate first full name, a screening score is determined based on the initial confidence level and the dynamic weight of the candidate first full name; the dynamic weight is positively correlated with the number of successful matches, and the number of successful matches represents the number of times the candidate first full name is used as the target full name; The target full name is selected from N candidate full names based on the screening score of each candidate full name.

3. The method according to claim 2, characterized in that, The selection score for the candidate first full name is determined based on its initial confidence level and dynamic weight, including: The initial confidence level of the candidate first full name is adjusted based on the configured calibration coefficients and entity discrimination, resulting in an adjusted confidence level. The screening score of the candidate first full name is determined based on the adjusted confidence level and the dynamic weight of the candidate first full name. The entity discrimination is determined based on the maximum and second-largest initial confidence levels among the N initial confidence levels corresponding to the N candidate first full names. The screening score of the candidate first full name is positively correlated with the adjusted confidence level and positively correlated with the dynamic weight of the candidate first full name.

4. The method according to any one of claims 1-3, characterized in that, After selecting N candidate first universal names from the M first universal names, the method further includes: The recall rate is determined based on the N initial confidence levels corresponding to the N candidate first full names; wherein, the number of initial confidence levels greater than the confidence threshold is counted, and the ratio of this number to N is determined as the recall rate; If the recall rate is greater than the configured recall rate threshold, then the operation of selecting the target full name from the N candidate first full names is performed.

5. The method according to claim 4, characterized in that, After determining the recall rate based on the N initial confidence scores corresponding to the N candidate first full names, the method further includes: If the recall rate is not greater than the recall rate threshold, then the user input text question is input into a conditional random field model to obtain entity features of multiple characters in the user input text question; Based on the entity features of the multiple characters, a character is selected from the multiple characters as the abbreviation to be completed using the conditional random field model, and the abbreviation to be completed is input into the large model; The second full name corresponding to the abbreviation to be completed is determined by a large model; the abbreviation to be completed in the user input text question is replaced by the second full name to obtain the target text question.

6. The method according to claim 5, characterized in that, The entity features include at least one of character-level features, lexical-level features, semantic-level features, and positional features; wherein, the character-level features include features of the current character, the characters preceding the current character, and the characters following the current character; the lexical-level features include part-of-speech tagging features and / or dictionary matching features of the current character; the semantic-level features include semantic features of the current character; and the positional features include features of the relative position of the current character in the user input text question.

7. A problem-solving device, characterized in that, The device includes: The query module is used to search the database by querying the database if the acquired user input text question includes an abbreviation to be completed, and to obtain M first full names corresponding to the abbreviation to be completed and the business features corresponding to each first full name; The acquisition module is used to select N candidate first full names from the M first full names, where N is a positive integer, based on the business characteristics corresponding to the abbreviation to be completed and the business characteristics corresponding to each first full name, if M is greater than 1; and to select the target full name from the N candidate first full names. The processing module is used to replace the abbreviation to be completed in the user input text question with the full name of the target to obtain the target text question; and to determine the question answer corresponding to the target text question.

8. The apparatus according to claim 7, characterized in that, When the acquisition module selects the target full name from the N candidate first full names, it is specifically used for: If N is greater than 1, then input the user-input text question and the N candidate first full names into the large model to obtain the initial confidence of each candidate first full name; For each candidate full name, a screening score is determined based on its initial confidence level and dynamic weight. Specifically, the initial confidence level of the candidate full name is adjusted based on a configured calibration coefficient and entity discrimination ratio to obtain an adjusted confidence level. The screening score is then determined based on the adjusted confidence level and the dynamic weight of the candidate full name. The dynamic weight is positively correlated with the number of successful matches, where the number of successful matches represents the number of times the candidate full name is used as the target full name. The entity discrimination ratio is determined based on the maximum and second-largest initial confidence levels among the N initial confidence levels corresponding to the N candidate full names. The screening score of the candidate full name is positively correlated with the adjusted confidence level and the dynamic weight of the candidate full name. The target full name is selected from N candidate full names based on the screening score of each candidate full name.

9. The apparatus according to claim 7 or 8, characterized in that, The acquisition module is further configured to determine the recall rate based on N initial confidence levels corresponding to the N candidate first full names after selecting N candidate first full names; wherein, the number of initial confidence levels greater than the confidence threshold is counted, and the ratio of this number to N is determined as the recall rate; if the recall rate is greater than the configured recall rate threshold, the acquisition module selects the target full name from the N candidate first full names; The acquisition module is further configured to, if the recall rate is not greater than the recall rate threshold, input the user input text question into a conditional random field model to obtain entity features of multiple characters in the user input text question; based on the entity features of the multiple characters, select characters from the multiple characters as abbreviated names to be completed through the conditional random field model, and input the abbreviated names to be completed into a large model; and determine the second full name corresponding to the abbreviated names to be completed through the large model. The processing module is further configured to replace the abbreviation to be completed in the user input text question with the second full name to obtain the target text question.

10. An electronic device, characterized in that, include: A processor and a machine-readable storage medium, the machine-readable storage medium storing machine-executable instructions that can be executed by the processor; The processor is configured to execute machine-executable instructions to implement the method of any one of claims 1-6.

Images