Task processing method, electronic device, storage medium and computer program product

Abstract

The application discloses a task processing method, an electronic device, a storage medium and a computer program product. The method comprises: obtaining an after-sales question and answer task to be processed, the after-sales question and answer task being used to request an answer corresponding to an after-sales question; determining one or more fault keywords in the after-sales question by using a first model, the first model being capable of identifying a first vector of a domain dimension associated with the after-sales question, and the one or more fault keywords being obtained by using at least the first vector; and obtaining the answer by using the one or more fault keywords and a second model, the second model being a model constructed based on first fault sample data, and the first fault sample data containing text data with a fixed structure. The technical scheme provided by the application can meet the question and answer demand of a user in a question and answer scene in a small or emerging field.

Description

Technical Field

[0001] This application relates to the field of deep learning technology, and in particular to a task processing method, electronic device, storage medium, and computer program product. Background Technology

[0002] Driven by technology, the cloud computing industry is booming. As enterprises deepen their application of cloud computing and cloud desktop technologies, the number of users continues to rise. However, users often encounter various malfunctions or unexpected problems during use.

[0003] To address the aforementioned issues, intelligent question-answering systems have been proposed in related technologies. As an important branch of natural language processing, intelligent question-answering systems can integrate various technologies such as information retrieval, knowledge representation, machine learning, and deep learning (specifically involving core areas such as knowledge graphs, natural language understanding, and deep learning) to provide corresponding answers to users' questions.

[0004] However, when facing niche or emerging fields (such as cloud computing), the relevant intelligent question-answering systems struggle to deeply understand and accurately respond to unique terminology, architecture, and professional questions, thus failing to meet users' question-answering needs. Summary of the Invention

[0005] To address the related technical problems, embodiments of this application provide a task processing method, an electronic device, a storage medium, and a computer program product.

[0006] The technical solution of this application embodiment is implemented as follows: This application provides a task processing method, including: Obtain pending after-sales Q&A tasks, wherein the after-sales Q&A tasks are used to request answers to after-sales questions; Using a first model, one or more fault keywords in the after-sales problem are determined. The first model is able to identify a first vector of the domain dimension associated with the after-sales problem and obtain the one or more fault keywords using at least the first vector. The answer is obtained by using one or more fault keywords and a second model, wherein the second model is a model constructed based on first fault sample data, and the first fault sample data contains text data with a fixed structure.

[0007] This application also provides an electronic device, including: a processor and a memory for storing a computer program that can run on the processor, wherein when the processor runs the computer program, it executes the steps of any of the above-described task processing methods.

[0008] This application also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of any of the above-described task processing methods.

[0009] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the above-described task processing methods.

[0010] The task processing method, apparatus, electronic device, storage medium, and computer program product provided in this application embodiment include: acquiring a post-sales Q&A task to be processed, wherein the post-sales Q&A task is used to request answers corresponding to post-sales questions; using a first model to determine one or more fault keywords in the post-sales question, wherein the first model can identify a first vector of a domain dimension associated with the post-sales question, and at least use the first vector to obtain the one or more fault keywords; using the one or more fault keywords and a second model to obtain the answer, wherein the second model is a model constructed based on first fault sample data, and the first fault sample data contains text data with a fixed structure. The technical solution provided in this application embodiment, for users' post-sales questions, introduces a domain embedding vector (i.e., a first vector) into a language model (i.e., the first model), enabling the language model to fully learn domain knowledge and enhance its understanding of post-sales questions, thereby accurately extracting fault keywords to determine the user's intent. Then, a semantic network model (i.e., the second model) is used to match the fault keywords and generate corresponding answers, which can improve the accuracy of question answering. Thus, it can meet users' question-and-answer needs in niche or emerging fields. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating a task processing method according to an embodiment of this application; Figure 2 This is a schematic diagram of the network structure of the first model in the embodiment of this application; Figure 3 This is a schematic diagram of the structure of a first fault sample data according to an embodiment of this application; Figure 4 This is a flowchart illustrating an intelligent question-answering system, which is an application example of this application. Figure 5 This is a schematic diagram of the structure of a task processing device according to an embodiment of this application; Figure 6 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0012] The present application will now be described in further detail with reference to the accompanying drawings and embodiments.

[0013] In the after-sales service of cloud computing products, the handling mechanism for standardized issues mainly relies on simple string matching based on customer input, and the solutions provided are usually pre-set, fixed scripts. However, this mechanism is unsuitable for non-standardized issues (i.e., complex or professional faults). It not only fails to provide users with accurate and effective solutions but also damages user trust in the responses. The handling mechanism for non-standardized issues primarily depends on human customer service, initiating a lengthy internal workflow. Specifically, after encountering a fault, a user first calls customer service, who records the problem and forwards it to the operations and maintenance team. The operations and maintenance personnel then investigate and locate the root cause of the problem, contact relevant personnel for root cause analysis and repair, and finally provide a solution to the user. This cumbersome process results in a long feedback cycle, leading to decreased user satisfaction and experience. Furthermore, it significantly increases labor costs, resulting in a generally low level of after-sales service quality for cloud computing products.

[0014] In related technologies, intelligent question-answering systems typically focus on general domains (such as healthcare, e-commerce, or broad customer service). For niche or long-tail domains (which can also be understood as highly specialized sub-fields, or niche and diverse sub-fields), these intelligent question-answering systems cannot meet users' needs for in-depth or precise question answering. This is because: 1. For the above-mentioned domains, there is a lack of large-scale and high-quality datasets that can be directly applied. 2. Large models in intelligent question-answering systems have performance limitations in understanding the technical terminology, complex concepts, and contextual scenarios specific to these domains. 3. Large models require significant computational overhead during training and inference, which creates obstacles to implementation in resource-constrained specialized vertical fields.

[0015] Based on this, various embodiments of this application propose a domain-aware embedding space enhancement strategy, which introduces domain vectors into the input representation of the language model, enabling domain knowledge to be deeply integrated into the language model, thereby improving the accuracy of identifying fault keywords (i.e., entity fault keywords). Furthermore, a semantic network model is used to match fault keywords and generate corresponding answers. This improves the accuracy of question-answering in niche or emerging question-answering scenarios, thus enhancing the performance and usability of intelligent question-answering capabilities.

[0016] This application provides a task processing method applied to electronic devices, such as... Figure 1 As shown, the method includes: Step 101: Obtain the after-sales Q&A task to be processed, wherein the after-sales Q&A task is used to request the answer to the after-sales question; Step 102: Using a first model, determine one or more fault keywords in the after-sales problem. The first model is able to identify a first vector of the domain dimension associated with the after-sales problem and obtain the one or more fault keywords using at least the first vector. Step 103: Using the one or more fault keywords and the second model, obtain the answer. The second model is a model built based on the first fault sample data, which contains text data with a fixed structure.

[0017] In practical applications, the electronic device may include a device capable of performing steps 101 to 103, such as a server, laptop, tablet, or desktop computer. This application embodiment does not limit the type of electronic device, as long as its function is implemented. Furthermore, the first model may be called a language model or a pre-trained language model, and it must at least be capable of semantic analysis and entity recognition to extract keywords from the question. This application embodiment does not limit the name of the first model, as long as its function is implemented.

[0018] In practical applications, before step 101, the electronic device can construct the first model and train the first model.

[0019] Based on this, in some optional embodiments, the method may further include: Obtain the first fault sample data and the after-sales Q&A sample data; Using the first fault sample data, a first set is generated, which contains fault labels corresponding to one or more domains to which the first fault sample data belongs; Using the first set and the after-sales Q&A sample data, a first sample vector is generated, the first sample vector containing a domain dimension vector associated with the after-sales Q&A sample data; The first model is trained using the first sample vector.

[0020] In practical applications, the first fault sample data can be understood as a structured fault report (such as a fault report in Portable Document Format (PDF) format). This fault report can contain highly structured fields, such as time fields, attribution fields, or processing fields associated with the fault, providing a foundation for subsequent data extraction. The after-sales Q&A sample data can be understood as textual corpus data from niche or long-tail fields (such as cloud computer after-sales service). This after-sales Q&A sample data can include one or more of the following (or at least one): user manuals, frequently asked questions, fault reports, technical support forum posts, user feedback and complaints, and repair records.

[0021] In practical applications, the electronic device can obtain the first fault sample data from the relevant database and perform clustering processing on the first fault sample data to generate an initial first set; wherein, the initial first set can be understood as a set of tags corresponding to the one or more fields. When the one or more fields include the cloud platform after-sales field, the first set can include fault tags such as hardware failure, network connection abnormality, and identity verification abnormality.

[0022] For example, if the first fault sample data contains an attribution field (specifically referred to as the root cause of the fault) associated with the fault, the electronic device can perform K-means clustering on the attribution field (e.g., with a cluster size of 5) to obtain the initial first set, wherein the attribution field can reflect the fault type in the first fault sample data.

[0023] In practical applications, after obtaining the initial first set, overlapping tags in the initial first set can be merged through manual review to obtain the final first set; wherein, the first set can be understood as a standardized domain tag set (which can be represented as...). , This is represented by the number of fault tags. Represented as the first Fault labels corresponding to each field.

[0024] For example, by introducing domain experts for manual review, the tags "login timeout" and "authentication failure" in the initial first set can be merged into the tag "authentication exception".

[0025] In practical applications, to improve the accuracy of the after-sales Q&A sample data, data augmentation (i.e., introducing domain experts as technical assistance) can be introduced to improve data quality and quantity, thereby ensuring the generalization ability of the first model. Furthermore, to ensure the real-time nature of the after-sales Q&A sample data, the electronic device can periodically acquire users' consultation scripts as sample data, so that the acquired data can be used to update the first model subsequently.

[0026] In practical applications, the electronic device can also preprocess the after-sales Q&A sample data to improve data quality.

[0027] For example, when the after-sales Q&A sample data is text corpus data in the field of cloud computer after-sales service, the electronic device can use the Jieba word segmentation tool to load dictionary data (specifically including terms in the fields of virtual graphics processor (vGPU) or cloud desktop protocol) in the text corpus data to solve the problem of splitting out-of-vocabulary words (such as identifying GPU-Zone as a single entity); use HanLP to perform part-of-speech tagging on the text corpus data to expand domain-specific word class tags (such as / hw_error (hardware error) or / net_fail (network failure) etc.) and optimize stop word filtering rules. At the same time, it retains negative words (such as not, not, etc.) and degree words (such as severe or frequent, etc.) to avoid semantic distortion (such as splitting the input "cloud computer is completely unresponsive" into "cloud computer", "not", "responsive"); and use entity boundary detection algorithms to merge compound terms in the text corpus data to treat misidentified fields as a whole (such as treating the GPU (Graphics Processing Unit) initialization timeout as a whole).

[0028] In practical applications, after preprocessing the after-sales Q&A sample data, the electronic device can use the preprocessed after-sales Q&A sample data and the first set to generate the first sample vector; wherein, the first sample vector can be understood as a vector corresponding to the domain embedding generated based on a combination of data-driven and human knowledge.

[0029] For example, the electronic device can process the pre-processed after-sales Q&A sample data using Word2Vec to obtain a corresponding word vector set; the word vector set is then grouped using the fault tags contained in the first set to obtain vector subsets corresponding to different fault tags. Each vector subset contains vectors corresponding to one or more terms (such as GPU passthrough failure or vGPU initialization error) associated with a fault tag (e.g., virtualization hardware anomaly); for different fault tags, the average value of all vectors in the corresponding vector subset is calculated, and the result is used as the initial first sample vector; the value can be calculated using the following formula... Related domain dimension vector : (1) in, Indicated as fault label The corresponding vector subset, This is represented as a vector within the corresponding vector subset. It is expressed as the number of vectors in a vector subset.

[0030] In practical applications, the Bidirectional Encoder Representations from Transformers (BERT) model, based on a deep bidirectional self-attention mechanism, can simultaneously aggregate the global context of target words during semantic space construction, dynamically fusing bidirectional contextual information while calculating word vectors. This characteristic enables the BERT model to exhibit better intent recognition performance when parsing complex semantic relationships (such as referential resolution or polysemy determination), thus becoming the core inference engine of related intelligent question-answering systems. However, when facing niche or emerging fields, the BERT model struggles to effectively understand the high-dimensional conceptual topologies of vertical domains (such as specialized concepts like GPU virtualization fault isolation or distributed storage hot migration thresholds in cloud computing after-sales scenarios).

[0031] To address the aforementioned issues, the electronic device can incorporate the first sample vector into the third-order embedding architecture of the relevant BERT model to construct the network structure of the first model. In other words, the electronic device can introduce the first sample vector into the input representation of the BERT model as a new domain embedding.

[0032] Based on this, in some optional embodiments, training the first model using the first sample vector includes: Using the after-sales Q&A sample data, a second sample vector, a third sample vector, and a fourth sample vector are generated respectively. The second sample vector contains a word-dimensional vector associated with the after-sales Q&A sample data, the third sample vector contains a segment-dimensional vector associated with the after-sales Q&A sample data, and the fourth sample vector contains a position-dimensional vector associated with the after-sales Q&A sample data. The first model is trained using the first sample vector, the second sample vector, the third sample vector, and the fourth sample vector.

[0033] In practical applications, the second sample vector can be understood as the vector corresponding to the word embedding, which can reflect the semantic information of the words; the third sample vector can be understood as the vector corresponding to the segment embedding, which can reflect the sentences in the after-sales Q&A sample data; and the fourth sample vector can be understood as the vector corresponding to the position embedding, which can reflect the position information of the words.

[0034] In practical applications, after constructing the network structure of the first model, the electronic device can convert the after-sales Q&A sample data into four types of vector representations (which can be represented as...). ), that is, the first sample vector and the second sample vector (which can be represented as The third sample vector (which can be represented as) ) and the fourth sample vector (which can be represented as This allows the first model to be trained using four types of vector representations.

[0035] In practical applications, to optimize the collaborative training of the four types of vector representations, the electronic device can divide the training process of the first model (which can also be understood as the adjustment process of model parameters) into three stages to dynamically adjust the learning rate of different network layers, thereby achieving gradual transfer of domain knowledge. This avoids domain knowledge being overwhelmed by general semantics or triggering catastrophic forgetting.

[0036] Based on this, in some optional embodiments, in the first stage of model training, the parameters of the network layers corresponding to the word dimension, segment dimension and position dimension in the first model and all the parameters of the encoding layer are frozen, and the parameters of the network layer corresponding to the domain dimension are updated. In the second stage of model training, some parameters of the encoding layer in the first model are unfrozen, and the parameters of the network layer corresponding to the domain dimension are continuously updated. In the third stage of model training, the parameters of the network layers corresponding to the word dimension, segment dimension, and position dimension in the first model, as well as the remaining parameters of the encoding layer, are unfrozen, and the parameters of the network layers corresponding to the domain dimension and some parameters of the encoding layer are continuously updated.

[0037] In practical applications, before model training, the electronic device can pre-train the first model using the after-sales Q&A sample data. This allows the first model to initially learn the language logic and terminology system specific to one or more domains, thus laying the foundation for subsequent model training. For example, the electronic device can use the after-sales Q&A sample data to take the masking strategy in the first model as a pre-training task; wherein the masking strategy tends to select domain terms from the after-sales Q&A sample data, such as GPU virtualization, Virtual Desktop Infrastructure (VDI) protocols, or cloud desktops.

[0038] In practical applications, during the first stage (also known as the domain embedding fusion stage, such as the first two iterations of model training), the electronic device can freeze the parameters of the network layers corresponding to the word dimension, segment dimension, and position dimension in the first model, as well as all the parameters of the encoding layer (which can be expressed as Transformer in English) (which can also be understood as all the weights of the backbone encoding layer). It only updates the parameters of the network layer (also known as the domain embedding layer) and the parameters of the classification layer (which can also be understood as the output layer) corresponding to the domain dimension, so that the first model can initially learn how to activate the first sample vector based on the current task (which can be understood as the task of entity recognition of the after-sales Q&A sample data), thereby making the first sample vector in a state of adaptation to the current task.

[0039] In practical applications, during the second stage (also known as the feature adjustment stage), the electronic device can unfreeze some parameters of the coding layer and continue to update the parameters of the network layer and the classification layer corresponding to the domain dimension, so that some network layers of the first model can adjust the semantic representation of the domain dimension based on the first sample vector; wherein, when the coding layer of the first model is composed of multiple coding layers stacked together, the parameters of the last layer or the last two coding layers can be unfrozen.

[0040] In practical applications, during the third stage (also known as the overall optimization stage), the electronic device can use a differential learning rate to unfreeze the parameters of the network layers corresponding to the word, segment, and position dimensions in the first model, as well as the remaining parameters of the encoding layer. It can also continuously update the parameters of the network layers corresponding to the domain dimension and some parameters of the encoding layer. That is, a smaller learning rate (such as 5e-6) is used to adjust the parameters of the lower-level network layers so that the lower-level network layers maintain general language knowledge, while a larger learning rate (such as 5e-5) is used to adjust the parameters of the higher-level network layers and the network layers corresponding to the domain dimension so that the higher-level network layers and the network layers corresponding to the domain dimension can better adapt to the task requirements associated with the domain.

[0041] For example, such as Figure 2As shown, assuming the first model includes an encoding layer (i.e., Transformer), a network layer corresponding to the domain dimension (i.e., domain embedding layer), a network layer corresponding to the word dimension (i.e., word embedding layer), a network layer corresponding to the segment dimension (i.e., segment embedding layer), a network layer corresponding to the position dimension (i.e., position embedding layer), an input layer, and an output layer (also called an entity recognition classification layer or classification layer), based on the network structure of the first model, the electronic device can divide the model training process (also called the fine-tuning process) into three stages to dynamically adjust the learning rate of different network layers, thereby achieving gradual transfer of domain knowledge. Specifically, in the first stage (i.e., the initial stage of fine-tuning), all parameters of the encoding layer, the parameters of the network layers corresponding to the word dimension, segment dimension, and position dimension are frozen, and only the parameters of the network layers corresponding to the domain dimension and the parameters of the higher-level output layers are updated. In the second stage (i.e., the network layers corresponding to the domain dimension are initially stable), the parameters of the last or last two encoding layers are unfrozen, and at the same time, the parameters of the network layers corresponding to the domain dimension and the parameters of the higher-level output layers continue to be updated. In the third stage (i.e., the later stage of fine-tuning), all network layers of the first model are unfrozen, and differential learning rates are used to adjust the parameters of different network layers.

[0042] In this embodiment of the application, by dividing the training process of the first model into different stages, the learning rate of different network layers can be adjusted at different stages. In this way, a smooth transition from general language knowledge to domain knowledge is achieved, effectively avoiding catastrophic forgetting.

[0043] In practical applications, during model training, the electronic device can further optimize the parameters of the first model by combining a loss function and an optimizer, so that the first model has a convergence direction. The optimizer is used to update the first model to minimize the loss function, and the optimizer may include AdamW (i.e., an optimizer that adds weight decay (L2 regularization) to Adam). Additionally, the loss function may include a cross-entropy loss function, which at least measures the difference between the probability distribution of the predicted label and the probability distribution of the true label. This loss function can be expressed as: (2) in, This represents the one-hot encoding of the actual label; This is represented as the probability distribution of the predicted labels; This is represented as element-wise multiplication.

[0044] In practical applications, before step 101, the electronic device also needs to construct the second model.

[0045] Based on this, in some optional embodiments, the method may further include: Obtain the data of the first fault sample; Information is extracted from the first fault sample data to obtain a first extraction result. The first extraction result includes one or more of the following: one or more first entities, the association information between the one or more first entities, and the attribute information corresponding to the one or more first entities. The second model is constructed using the first extraction result.

[0046] In practical applications, the first fault sample data can contain highly structured fields and corresponding field values. Specifically, it can include one or more of the following fields: fault description, fault related terms, fault root cause, responsible party, solution, and knowledge management attribute fields. The fault description records the fault phenomenon in text form; the fault related terms provide keywords (also known as key terms) associated with matching and retrieval; the fault root cause (which can include primary and secondary classifications) indicates the fault type; the responsible party (which can be the fault optimization department and the responsible department) indicates the entity responsible for handling the fault; the solution (which can be the fault investigation process and handling measures) records the solution for handling the fault; and the knowledge management attribute (which can include whether it recurs and whether there are common risks) provides tags for knowledge reuse or risk warning. These structured fields ensure the consistency and comprehensiveness of the data source, enabling the construction of a second model with high topological reasoning capabilities.

[0047] For example, such as Figure 3 As shown, assuming the first fault sample data acquired is a fault report in PDF format (i.e., it not only contains basic spatiotemporal fields such as fault occurrence time and fault recovery time, but also defines fault attribution and processing chain), the electronic device can preprocess the PDF fault report to convert it into text data with a fixed structure (which can also be understood as structured text data). This text data can contain one or more field names and corresponding field values. Specifically, firstly, the electronic device uses the pyMuPDF library to parse the raw text stream in the PDF fault report. Next, by parsing the underlying drawing instructions of the PDF, it identifies the cell boundaries of the table to correctly associate text spanning rows (such as fault time information) and columns (such as fault descriptions) with the corresponding headings. Then, based on a predefined list of key field names (such as the primary category of fault root cause or responsible department), it locates and extracts the valid content within adjacent cells, filtering out unstructured noise (which can also be understood as redundant information) such as table frame lines and irrelevant headers and footers. This ensures the integrity of information extraction from the PDF fault report.

[0048] In practical applications, for the first fault sample data, the electronic device can convert the fixed-structure text data in the first fault sample data into entities and entity relationships required by the second model, thereby obtaining the first extraction result; wherein, based on the inherent logic of the first fault sample data, the association information between the one or more first entities can be obtained; and based on the field values ​​in the first fault sample data, the attribute information corresponding to the one or more first entities or the one or more first entities can be obtained.

[0049] For example, when the first fault sample data includes fault description and fault-related keyword fields, the electronic device can use annotation tools to extract a first entity associated with the keywords from the field values ​​corresponding to the fault description and fault-related keyword fields; when the first fault sample data includes a fault root cause field, the electronic device can extract a first entity associated with the fault root cause from the field values ​​corresponding to the fault root cause field; when the first fault sample data includes a responsible party field, the electronic device can extract a first entity associated with the responsible party from the field values ​​corresponding to the responsible party field; when the first fault sample data includes a solution field, the electronic device can extract a first entity associated with the solution from the field values ​​corresponding to the solution field.

[0050] For example, when the first fault sample data contains fields for fault root cause, responsible party, and solution, the electronic device can automatically associate the field values ​​corresponding to the fault root cause (specifically, it can include the field values ​​corresponding to the first-level fault root cause (hardware problem) and the field values ​​corresponding to the second-level fault root cause (GPU failure)), the field values ​​corresponding to the responsible party (specifically, it can be the virtualization hardware group), and the field values ​​corresponding to the solution field (specifically, it can be upgrading the graphics card driver) to obtain the association information between the three first entities.

[0051] In practical applications, manual verification can be introduced to sample and review the first extraction results to ensure the accuracy of the terms in the first extraction results (such as distinguishing between stuttering and delay) and the correctness of the relationship links (such as ensuring that the first entity associated with the solution and the first entity associated with the root cause of the failure are accurately associated).

[0052] In practical applications, after the sampling review is completed, the electronic device can use the first sampling result to define the entities, entity relationships and attributes of the second model, and then construct the second model.

[0053] For example, regarding the entities in the second model, the electronic device can define six types of entities: fault cause (with a unique fault cause ID), fault keyword (with a unique fault keyword ID), primary root cause (with a unique primary root cause ID), secondary root cause (with a unique secondary root cause ID), responsible department (with a unique responsible department ID), and fault solution (with a unique fault solution ID). Regarding the entity relationships in the second model, the electronic device can define five types of entity relationships: fault cause and fault solution (one fault cause corresponds to multiple fault solutions), fault cause and fault keyword (one fault cause corresponds to multiple fault keywords), fault cause and primary root cause (one fault cause corresponds to one primary root cause), fault cause and secondary root cause (one fault cause corresponds to one secondary root cause), and fault cause and responsible department (one fault cause corresponds to one responsible department). Regarding the attributes of the second model, the electronic device can define fault causes, which can specifically be recurring or share common risks. After defining the entities, entity relationships, and attributes of the second model, the electronic device can create corresponding entities and attributes in the Neo4j graph database and convert the first extraction result into a CSV file for import into the database. This completes the construction of the second model.

[0054] In practical applications, new entities or entity relationships may emerge over time. Therefore, the electronic device can update the second model through reinforcement learning.

[0055] Accordingly, in some optional embodiments, the method further includes: Obtain the second fault sample data; Information is extracted from the second fault sample to obtain a second extraction result. The second extraction result includes one or more of the following: one or more second entities, the association information between the one or more second entities, and the attribute information corresponding to the one or more second entities. The first extraction result and the second extraction result are matched to obtain a first matching result, which represents the similarity between the first extraction result and the second extraction result. The second model is updated using the first matching result.

[0056] In practical applications, the second fault sample data can be understood as newly generated structured text data, and the second fault sample data is at least partially different from the first fault sample data.

[0057] In practical applications, the electronic device can obtain the second fault sample data from the relevant database and convert the fixed-structure text data in the second fault sample data into entities and entity relationships to obtain the second extraction result; wherein, based on the inherent logic of the second fault sample data, the association information between the one or more second entities can be obtained; and based on the field values ​​in the second fault sample data, the attribute information corresponding to the one or more second entities or the one or more second entities can be obtained.

[0058] In practical applications, after obtaining the second extraction result, the electronic device can match (or compare) the second extraction result with the first extraction result to determine whether similar entities, entity relationships, or attributes exist. If the similarity represented by the first matching result is less than a preset first threshold (the value can be set as needed, and this embodiment does not limit this), the electronic device can consider that there are no similar entities, entity relationships, or attributes. In this case, the second extraction result can be used to update the second model, that is, the second extraction result can be updated into the second model. If the similarity represented by the first matching result is greater than or equal to the preset first threshold, the electronic device can consider that there are similar entities, entity relationships, or attributes. In this case, the similar entities, entity relationships, or attributes and the second extraction result can be used to partially update the second model.

[0059] In practical applications, to ensure the accuracy of the second model, the electronic device can also update the second model based on user feedback. For example, the electronic device can monitor user feedback or automatically detected error information. If negative feedback, error information, or improvement suggestions are received, erroneous entities are identified, and entities, entity relationships, or attributes in the second model are added or corrected to achieve iterative updates to the second model. This enables continuous optimization of the second model, thereby ensuring the accuracy of the problem-solving results.

[0060] In practical applications, after the construction of the first model and the second model is completed, the electronic device can receive after-sales Q&A tasks input by the user.

[0061] It should be noted that, to improve the robustness of the input content, the electronic device can be equipped with a real-time verification and error recovery mechanism for the input content. This involves calculating the similarity between the current after-sales Q&A task and historical after-sales Q&A tasks in the local database. If the similarity is less than a preset second threshold (which can be set as needed, for example, 0.4), then ambiguous input content in the after-sales Q&A task (such as a black screen on a computer) is identified, and guidance suggestions are automatically generated (e.g., please specify the device type: cloud computer or local computer). Simultaneously, users can correct the after-sales Q&A task via voice (e.g., voice input of corrected content to replace erroneous fields in the original content).

[0062] In practical applications, after obtaining the after-sales Q&A task, the electronic device can use the first model to extract entities from the after-sales question (also known as entity recognition) to obtain one or more fault keywords. That is, the electronic device can use the extracted entities as one or more fault keywords, and then use the second model to match the one or more fault keywords to obtain the answer, which can be understood as a solution to the after-sales question.

[0063] Based on this, in some optional embodiments, obtaining the answer using the one or more fault keywords and the second model includes: The one or more fault keywords and fault-related information associated with the second model are matched to obtain a second matching result, which represents the similarity between the one or more fault keywords and the fault-related information. The answer is obtained using the matching results.

[0064] In practical applications, the fault-related information may include one or more of the entities, entity relationships, and attributes in the second model.

[0065] In practical applications, by matching the one or more fault keywords with the fault-related information, the electronic device can determine one or more of the entities, entity relationships, and attributes associated with the one or more fault keywords in the fault-related information, that is, one or more of the entities, entity relationships, and attributes with a similarity greater than or equal to a preset third threshold (the value can be set as needed, and this application embodiment does not limit this). Then, using one or more of the associated entities, entity relationships, and attributes, the answer is obtained.

[0066] For example, the electronic device can match one or more fault keywords with the fault-related information to obtain a matching entity (specifically, the root cause of the fault, the responsible party, and the solution); by feeding back the matching entity to relevant personnel, the relevant personnel can quickly troubleshoot and locate the fault, thereby feeding back the answer to the user.

[0067] The task processing method provided in this application includes: obtaining a post-sales Q&A task to be processed, wherein the post-sales Q&A task is used to request answers corresponding to post-sales questions; using a first model to determine one or more fault keywords in the post-sales question, wherein the first model can identify a first vector of a domain dimension associated with the post-sales question, and at least use the first vector to obtain the one or more fault keywords; using the one or more fault keywords and a second model to obtain the answer, wherein the second model is a model constructed based on first fault sample data, and the first fault sample data contains text data with a fixed structure. The technical solution provided in this application, for users' post-sales questions, introduces a domain embedding vector (i.e., a first vector) into a language model (i.e., the first model), enabling the language model to fully learn domain knowledge and enhance its understanding of post-sales questions, thereby accurately extracting fault keywords to determine the user's intent. Then, a semantic network model (i.e., the second model) is used to match the fault keywords and generate corresponding answers, which can improve the accuracy of question answering. Thus, it can meet users' question-and-answer needs in niche or emerging fields.

[0068] The following section provides a more detailed description of this application with reference to application examples.

[0069] In the application examples of this application, an intelligent question-answering system integrating an adaptive semantic network and domain features is proposed for the cloud computer after-sales service field; specifically, such as... Figure 4 As shown, the BERT model (i.e., the first model mentioned above) in the intelligent question-answering system analyzes and identifies entities in the user input information (i.e., the after-sales question-answering task mentioned above), extracts fault keywords, and then matches the fault keywords with the cloud computer fault semantic network (i.e., the second model mentioned above) in the intelligent question-answering system to obtain the matched fault-related information (i.e., the answer mentioned above) and feeds it back to the maintenance personnel. The maintenance personnel generate the result based on the matched fault-related information and feed it back to the user. Among them, the cloud computer fault semantic network is constructed based on existing structured fault records (such as fault records in PDF format) and unstructured text resources (such as historical input speech-to-text). The input representation of the BERT model at least includes a domain embedding vector (i.e., the first vector mentioned above).

[0070] Here, to avoid relying on semantic prior knowledge trained on large-scale open-domain corpora and to address the issues of insufficient terminology understanding and domain knowledge gaps in language models within the cloud computer after-sales service domain, a domain-aware embedding space enhancement was performed on the BERT model. Specifically, a learnable domain embedding vector was introduced as the input representation on top of the original three-level network structure of BERT, and the BERT model was pre-trained using a domain sample dataset. Then, an adaptive hierarchical fine-tuning algorithm (also known as parameter efficiency optimization fine-tuning) was employed to divide the fine-tuning process of the BERT model into three stages to dynamically adjust the parameters of different network layers, thereby ensuring that domain knowledge is deeply integrated into the BERT model. This effectively enhances the BERT model's contextual understanding ability and problem-solving accuracy in cloud computer after-sales scenarios, achieving a dual improvement in performance and usability.

[0071] In addition, by accumulating rich heterogeneous information (i.e., existing structured fault records and unstructured text resources), a cloud computer fault semantic network is constructed, providing graph database support for a wide range of application scenarios. This enables the cloud computer fault semantic network to accurately and efficiently retrieve content from the knowledge base and intelligently recommend solutions to related problems, thereby replacing the cumbersome traditional path of users directly seeking human customer service.

[0072] Thus, the aforementioned intelligent question-and-answer system can provide online, real-time professional answers to questions, effectively reducing the workload of manual labor, lowering costs, and significantly improving the efficiency and quality of after-sales service.

[0073] To implement the task processing method of this application embodiment, this application embodiment also provides a task processing device, which is installed on an electronic device, such as... Figure 5 As shown, the device includes: The acquisition unit 501 is used to acquire the after-sales Q&A task to be processed, wherein the after-sales Q&A task is used to request the answer to the after-sales question; The first processing unit 502 is configured to use a first model to determine one or more fault keywords in the after-sales problem. The first model is capable of identifying a first vector of the domain dimension associated with the after-sales problem and obtaining the one or more fault keywords using at least the first vector. The second processing unit 503 is used to obtain the answer by utilizing the one or more fault keywords and the second model, wherein the second model is a model constructed based on the first fault sample data, and the first fault sample data contains text data with a fixed structure.

[0074] In some optional embodiments, the acquisition unit 501 is further configured to acquire the first fault sample data and after-sales Q&A sample data; The first processing unit 502 is further configured to generate a first set using the first fault sample data, the first set containing fault labels corresponding to one or more domains to which the first fault sample data belongs; generate a first sample vector using the first set and the after-sales Q&A sample data, the first sample vector containing a domain dimension vector associated with the after-sales Q&A sample data; and train the first model using the first sample vector.

[0075] In some optional embodiments, the first processing unit 502 is configured to generate a second sample vector, a third sample vector, and a fourth sample vector using the after-sales Q&A sample data, wherein the second sample vector contains a word-dimensional vector associated with the after-sales Q&A sample data, the third sample vector contains a segment-dimensional vector associated with the after-sales Q&A sample data, and the fourth sample vector contains a position-dimensional vector associated with the after-sales Q&A sample data; and to train the first model using the first sample vector, the second sample vector, the third sample vector, and the fourth sample vector.

[0076] In some optional embodiments, the first processing unit 502 is configured to, in the first stage of model training, freeze the parameters of the network layers corresponding to the word dimension, segment dimension, and position dimension in the first model, as well as all the parameters of the encoding layer, and update the parameters of the network layer corresponding to the domain dimension; in the second stage of model training, unfreeze some of the parameters of the encoding layer in the first model, and continuously update the parameters of the network layer corresponding to the domain dimension; in the third stage of model training, unfreeze the parameters of the network layers corresponding to the word dimension, segment dimension, and position dimension in the first model, as well as the remaining parameters of the encoding layer, and continuously update the parameters of the network layer corresponding to the domain dimension and some of the parameters of the encoding layer.

[0077] In some optional embodiments, the acquisition unit 501 is further configured to acquire the first fault sample data; The second processing unit 503 is further configured to extract information from the first fault sample data to obtain a first extraction result, wherein the first extraction result includes one or more of the following: one or more first entities, association information between the one or more first entities, and attribute information corresponding to the one or more first entities; and to construct the second model using the first extraction result.

[0078] In some optional embodiments, the acquisition unit 501 is further configured to acquire second fault sample data; The second processing unit 503 is further configured to extract information from the second fault sample to obtain a second extraction result, wherein the second extraction result includes one or more of the following: one or more second entities, association information between the one or more second entities, and attribute information corresponding to the one or more second entities; match the first extraction result and the second extraction result to obtain a first matching result, wherein the first matching result characterizes the similarity between the first extraction result and the second extraction result; and update the second model using the first matching result.

[0079] In some optional embodiments, the second processing unit 503 is further configured to match the one or more fault keywords with fault-related information associated with the second model to obtain a second matching result, wherein the second matching result characterizes the similarity between the one or more fault keywords and the fault-related information; and to obtain the answer using the matching result.

[0080] In practical applications, the acquisition unit 501 can be implemented by the communication interface in the task processing device; the first processing unit 502 and the second processing unit 503 can be implemented by the processor in the task processing device.

[0081] It should be noted that the task processing device provided in the above embodiments is only illustrated by the division of the above-described program units. In practical applications, the above processing can be assigned to different program units as needed, that is, the internal structure of the device can be divided into different program units to complete all or part of the processing described above. In addition, the task processing device and the task processing method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the method embodiments, which will not be repeated here.

[0082] Based on the hardware implementation of the above program modules, and in order to implement the task processing method of the embodiments of this application, the embodiments of this application also provide an electronic device, such as... Figure 6 As shown, the electronic device 600 includes: The communication interface 601 enables interaction with other devices; The processor 602 is connected to the communication interface 601 to enable interaction with other devices and to execute the methods provided by one or more of the above-mentioned technical solutions when running a computer program; The computer program is stored in memory 603.

[0083] Specifically, the communication interface 601 is used to acquire after-sales Q&A tasks to be processed, and the after-sales Q&A tasks are used to request answers to after-sales questions; The processor 602 is configured to use a first model to determine one or more fault keywords in the after-sales problem, the first model being able to identify a first vector of a domain dimension associated with the after-sales problem and obtain the one or more fault keywords using at least the first vector; and to obtain the answer using the one or more fault keywords and a second model, the second model being a model constructed based on first fault sample data, the first fault sample data containing text data with a fixed structure.

[0084] In some optional embodiments, the communication interface 601 is further used to acquire the first fault sample data and after-sales Q&A sample data; The processor 602 is further configured to generate a first set using the first fault sample data, the first set containing fault labels corresponding to one or more domains to which the first fault sample data belongs; generate a first sample vector using the first set and the after-sales Q&A sample data, the first sample vector containing a domain dimension vector associated with the after-sales Q&A sample data; and train the first model using the first sample vector.

[0085] In some optional embodiments, the processor 602 is configured to generate a second sample vector, a third sample vector, and a fourth sample vector using the after-sales Q&A sample data, wherein the second sample vector contains a word-dimensional vector associated with the after-sales Q&A sample data, the third sample vector contains a segment-dimensional vector associated with the after-sales Q&A sample data, and the fourth sample vector contains a position-dimensional vector associated with the after-sales Q&A sample data; and to train the first model using the first sample vector, the second sample vector, the third sample vector, and the fourth sample vector.

[0086] In some optional embodiments, the processor 602 is configured to, in the first stage of model training, freeze the parameters of the network layers corresponding to the word dimension, segment dimension, and position dimension in the first model, as well as all the parameters of the encoding layer, and update the parameters of the network layer corresponding to the domain dimension; in the second stage of model training, unfreeze some of the parameters of the encoding layer in the first model, and continuously update the parameters of the network layer corresponding to the domain dimension; in the third stage of model training, unfreeze the parameters of the network layers corresponding to the word dimension, segment dimension, and position dimension in the first model, as well as the remaining parameters of the encoding layer, and continuously update the parameters of the network layer corresponding to the domain dimension and some of the parameters of the encoding layer.

[0087] In some optional embodiments, the communication interface 601 is further configured to acquire the first fault sample data; The processor 602 is further configured to extract information from the first fault sample data to obtain a first extraction result, wherein the first extraction result includes one or more of the following: one or more first entities, association information between the one or more first entities, and attribute information corresponding to the one or more first entities; and to construct the second model using the first extraction result.

[0088] In some optional embodiments, the communication interface 601 is further used to acquire second fault sample data; The processor 602 is further configured to extract information from the second fault sample to obtain a second extraction result, wherein the second extraction result includes one or more of the following: one or more second entities, association information between the one or more second entities, and attribute information corresponding to the one or more second entities; match the first extraction result and the second extraction result to obtain a first matching result, wherein the first matching result characterizes the similarity between the first extraction result and the second extraction result; and update the second model using the first matching result.

[0089] In some optional embodiments, the processor 602 is configured to match the one or more fault keywords with fault-related information associated with the second model to obtain a second matching result, the second matching result representing the similarity between the one or more fault keywords and the fault-related information; and to obtain the answer using the matching result.

[0090] It should be noted that the specific processing procedures of the communication interface 601 and the processor 602 can be understood with reference to the above method.

[0091] Of course, in practical applications, the various components in electronic device 600 are coupled together through bus system 604. It can be understood that bus system 604 is used to realize the connection and communication between these components. In addition to a data bus, bus system 604 also includes a power bus, a control bus, and a status signal bus. However, for the sake of clarity, in... Figure 6 The general designated all buses as Bus System 604.

[0092] The memory 603 in this embodiment is used to store various types of data to support the operation of the electronic device 600. Examples of such data include any computer program used to operate on the electronic device 600.

[0093] The methods disclosed in the embodiments of this application can be applied to the processor 602, or implemented by the processor 602. The processor 602 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in the processor 602 or by instructions in the form of software. The processor 602 may be a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 602 can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor, etc. The steps of the methods disclosed in the embodiments of this application can be directly reflected as being executed by a hardware decoding processor, or being executed by a combination of hardware and software modules in the decoding processor. The software modules may be located in a storage medium, which is located in the memory 603. The processor 602 reads the information in the memory 603 and combines its hardware to complete the steps of the aforementioned method.

[0094] In an exemplary embodiment, the electronic device 600 may be implemented by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers (MCUs), microprocessors, or other electronic components to perform the aforementioned method.

[0095] It is understood that the memory (memory 603) in this embodiment of the application can be volatile memory or non-volatile memory, or both. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), ferromagnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM); the magnetic surface memory can be disk storage or magnetic tape storage. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Synchronous Static Random Access Memory (SSRAM), Dynamic Random Access Memory (DRAM), Synchronous Dynamic Random Access Memory (SDRAM), Double Data Rate Synchronous Dynamic Random Access Memory (DDRSDRAM), Enhanced Synchronous Dynamic Random Access Memory (ESDRAM), SyncLink Dynamic Random Access Memory (SLDRAM), and Direct Rambus Random Access Memory (DRRAM).The memories described in the embodiments of this application are intended to include, but are not limited to, these and any other suitable types of memories.

[0096] In an exemplary embodiment, this application also provides a storage medium, namely a computer storage medium, specifically a computer-readable storage medium, such as a memory 603 storing a computer program, which can be executed by the processor 602 of the electronic device 600 to complete the steps of the aforementioned task processing method. The computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, Flash Memory, magnetic surface memory, optical disc, or CD-ROM.

[0097] In an exemplary embodiment, this application also provides a computer program product, including a computer program that can be executed by a processor 602 of an electronic device 600 to complete the steps of the aforementioned task processing method.

[0098] It should be noted that terms such as "first" and "second" are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0099] Furthermore, the technical solutions described in the embodiments of this application can be combined arbitrarily without conflict.

[0100] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application.

Claims

1. A task processing method, characterized in that, The method includes: Obtain pending after-sales Q&A tasks, wherein the after-sales Q&A tasks are used to request answers to after-sales questions; Using a first model, one or more fault keywords in the after-sales problem are determined; wherein, the first model is capable of identifying a first vector of the domain dimension associated with the after-sales problem, and obtaining the one or more fault keywords using at least the first vector; The answer is obtained by using one or more fault keywords and a second model, wherein the second model is a model built based on first fault sample data, and the first fault sample data contains text data with a fixed structure. The method further includes: The first model is trained using a first sample vector, a second sample vector, a third sample vector, and a fourth sample vector. The first sample vector contains a domain dimension vector associated with the after-sales Q&A sample data; the second sample vector contains a word dimension vector associated with the after-sales Q&A sample data; the third sample vector contains a segment dimension vector associated with the after-sales Q&A sample data; and the fourth sample vector contains a position dimension vector associated with the after-sales Q&A sample data. In the first stage of model training, the parameters of the network layers corresponding to the word, segment, and position dimensions, as well as all parameters of the encoding layer, are frozen, while the parameters of the network layers corresponding to the domain dimensions are updated. In the second stage of model training, some parameters of the encoding layer in the first model are unfrozen, and the parameters of the network layers corresponding to the domain dimensions are continuously updated. In the third stage of model training, the parameters of the network layers corresponding to the word, segment, and position dimensions, as well as the remaining parameters of the encoding layer, are unfrozen, while the parameters of the network layers corresponding to the domain dimensions and some parameters of the encoding layer are continuously updated.

2. The method according to claim 1, characterized in that, The method further includes: Obtain the first fault sample data and the after-sales Q&A sample data; Using the first fault sample data, a first set is generated, which contains fault labels corresponding to one or more domains to which the first fault sample data belongs; The first sample vector is generated using the first set and the after-sales Q&A sample data.

3. The method according to claim 2, characterized in that, The method further includes: Using the after-sales Q&A sample data, the second sample vector, the third sample vector, and the fourth sample vector are generated respectively.

4. The method according to claim 1, characterized in that, The method further includes: Obtain the data of the first fault sample; Information is extracted from the first fault sample data to obtain a first extraction result. The first extraction result includes one or more of the following: one or more first entities, the association information between the one or more first entities, and the attribute information corresponding to the one or more first entities. The second model is constructed using the first extraction result.

5. The method according to claim 4, characterized in that, The method further includes: Obtain the second fault sample data; Information is extracted from the second fault sample to obtain a second extraction result. The second extraction result includes one or more of the following: one or more second entities, the association information between the one or more second entities, and the attribute information corresponding to the one or more second entities. The first extraction result and the second extraction result are matched to obtain a first matching result, which represents the similarity between the first extraction result and the second extraction result. The second model is updated using the first matching result.

6. The method according to claim 4, characterized in that, The process of obtaining the answer using one or more fault keywords and the second model includes: The one or more fault keywords and fault-related information associated with the second model are matched to obtain a second matching result, which represents the similarity between the one or more fault keywords and the fault-related information. The answer is obtained using the matching results.

7. An electronic device, characterized in that, include: The processor and the memory used to store computer programs that can run on the processor. When the processor is used to run the computer program, it performs the steps of the method according to any one of claims 1 to 6.

8. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images