CN116956934A - Task processing method, device, equipment and storage medium - Google Patents
Task processing method, device, equipment and storage medium Download PDFInfo
- Publication number
- CN116956934A CN116956934A CN202310754367.1A CN202310754367A CN116956934A CN 116956934 A CN116956934 A CN 116956934A CN 202310754367 A CN202310754367 A CN 202310754367A CN 116956934 A CN116956934 A CN 116956934A
- Authority
- CN
- China
- Prior art keywords
- task
- information
- model
- processing
- predefined
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000003860 storage Methods 0.000 title claims abstract description 23
- 238000003672 processing method Methods 0.000 title abstract description 21
- 238000012545 processing Methods 0.000 claims abstract description 226
- 238000000034 method Methods 0.000 claims abstract description 150
- 238000003058 natural language processing Methods 0.000 claims abstract description 28
- 238000012549 training Methods 0.000 claims description 78
- 238000000605 extraction Methods 0.000 claims description 49
- 230000003993 interaction Effects 0.000 claims description 27
- 238000010801 machine learning Methods 0.000 claims description 27
- 238000004590 computer program Methods 0.000 claims description 18
- 230000009471 action Effects 0.000 claims description 13
- 238000011049 filling Methods 0.000 claims description 5
- 238000013473 artificial intelligence Methods 0.000 abstract description 21
- 238000005065 mining Methods 0.000 description 44
- 230000008569 process Effects 0.000 description 28
- 238000005516 engineering process Methods 0.000 description 27
- 238000013135 deep learning Methods 0.000 description 17
- 230000000694 effects Effects 0.000 description 14
- 238000004458 analytical method Methods 0.000 description 13
- 230000008901 benefit Effects 0.000 description 13
- 238000010586 diagram Methods 0.000 description 13
- 238000013528 artificial neural network Methods 0.000 description 12
- 230000001965 increasing effect Effects 0.000 description 11
- 238000005457 optimization Methods 0.000 description 11
- 238000013461 design Methods 0.000 description 10
- 238000013519 translation Methods 0.000 description 10
- 239000000284 extract Substances 0.000 description 9
- 230000002829 reductive effect Effects 0.000 description 9
- 241000282414 Homo sapiens Species 0.000 description 7
- 238000011161 development Methods 0.000 description 7
- 230000003203 everyday effect Effects 0.000 description 7
- 238000013459 approach Methods 0.000 description 6
- 230000006835 compression Effects 0.000 description 6
- 238000007906 compression Methods 0.000 description 6
- 238000011160 research Methods 0.000 description 6
- 238000004422 calculation algorithm Methods 0.000 description 5
- 230000008451 emotion Effects 0.000 description 5
- 239000013598 vector Substances 0.000 description 5
- 230000009286 beneficial effect Effects 0.000 description 4
- 230000002354 daily effect Effects 0.000 description 4
- 230000006870 function Effects 0.000 description 4
- 230000007246 mechanism Effects 0.000 description 4
- 238000013139 quantization Methods 0.000 description 4
- 238000012552 review Methods 0.000 description 4
- 230000000007 visual effect Effects 0.000 description 4
- 230000001133 acceleration Effects 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 238000004891 communication Methods 0.000 description 3
- 238000010276 construction Methods 0.000 description 3
- 230000014509 gene expression Effects 0.000 description 3
- 230000002452 interceptive effect Effects 0.000 description 3
- 230000002787 reinforcement Effects 0.000 description 3
- 239000007787 solid Substances 0.000 description 3
- 230000003190 augmentative effect Effects 0.000 description 2
- 230000006399 behavior Effects 0.000 description 2
- 230000033228 biological regulation Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000004140 cleaning Methods 0.000 description 2
- 230000001419 dependent effect Effects 0.000 description 2
- 208000019622 heart disease Diseases 0.000 description 2
- 230000000670 limiting effect Effects 0.000 description 2
- 238000012423 maintenance Methods 0.000 description 2
- 238000003062 neural network model Methods 0.000 description 2
- 238000013138 pruning Methods 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 238000007619 statistical method Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 102100033814 Alanine aminotransferase 2 Human genes 0.000 description 1
- 101710096000 Alanine aminotransferase 2 Proteins 0.000 description 1
- 206010063385 Intellectualisation Diseases 0.000 description 1
- 241000070023 Phoenicopterus roseus Species 0.000 description 1
- 238000012356 Product development Methods 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000002457 bidirectional effect Effects 0.000 description 1
- QVFWZNCVPCJQOP-UHFFFAOYSA-N chloralodol Chemical compound CC(O)(C)CC(C)OC(O)C(Cl)(Cl)Cl QVFWZNCVPCJQOP-UHFFFAOYSA-N 0.000 description 1
- 230000001427 coherent effect Effects 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 239000003814 drug Substances 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000007667 floating Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000006698 induction Effects 0.000 description 1
- 238000013140 knowledge distillation Methods 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 230000005055 memory storage Effects 0.000 description 1
- 230000003340 mental effect Effects 0.000 description 1
- 238000013508 migration Methods 0.000 description 1
- 230000005012 migration Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000036651 mood Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000008520 organization Effects 0.000 description 1
- 230000036961 partial effect Effects 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 238000011084 recovery Methods 0.000 description 1
- 238000009877 rendering Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 238000000547 structure data Methods 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 238000013526 transfer learning Methods 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/903—Querying
- G06F16/9032—Query formulation
- G06F16/90332—Natural language query formulation or dialogue systems
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/903—Querying
- G06F16/90335—Query processing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/903—Querying
- G06F16/9038—Presentation of query results
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Databases & Information Systems (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Computing Systems (AREA)
- Machine Translation (AREA)
Abstract
The application discloses a task processing method, device, equipment and storage medium, and relates to the technical field of artificial intelligence. The method comprises the following steps: acquiring description information and sample data of a first task, wherein the first task is a task to be processed related to natural language processing, the description information is used for indicating the processing requirement of the first task, and the sample data is used for indicating the processing object of the first task; extracting at least one key point information of the first task from the description information, wherein the key point information is used for indicating the processing key point of the first task; determining task prompt information corresponding to a first task according to the sample data and at least one key point information, wherein the task prompt information is used for representing the processing requirement and the processing object of the first task; and generating a processing result of the first task according to the task prompt information through the task processing model. The method improves the precision of the generated processing result.
Description
Technical Field
The embodiment of the application relates to the technical field of artificial intelligence, in particular to a task processing method, device, equipment and storage medium.
Background
With the development of artificial intelligence technology, natural language processing technology is gradually maturing. In the natural language processing technology, the intention of the user can be understood according to the description information of the user, and corresponding results are given.
In the related art, a corresponding processing result is generated through a pre-training large model according to description information input by a user. The pre-training large model is a more general language model, such as a ChatGPT model, and can generate corresponding processing results aiming at any description information of a user.
However, in the above related art, the pre-trained large model is a general language model, and its understanding ability for a user-specified task is relatively weak, which results in lower accuracy of processing results generated for the user's descriptive information for the specified task.
Disclosure of Invention
The embodiment of the application provides a task processing method, device, equipment and storage medium. The technical scheme is as follows:
according to an aspect of an embodiment of the present application, there is provided a task processing method, including:
acquiring description information and sample data of a first task, wherein the first task is a task to be processed related to natural language processing, the description information is used for indicating the processing requirement of the first task, and the sample data is used for indicating a processing object of the first task;
Extracting at least one gist information of the first task from the description information, wherein the gist information is used for indicating a processing gist of the first task;
determining task prompt information corresponding to the first task according to the sample data and the at least one key point information, wherein the task prompt information is used for representing the processing requirement and the processing object of the first task;
and generating a processing result of the first task according to the task prompt information through a task processing model.
According to an aspect of an embodiment of the present application, there is provided a task processing device including:
the data acquisition module is used for acquiring description information and sample data of a first task, wherein the first task is a task to be processed related to natural language processing, the description information is used for indicating the processing requirement of the first task, and the sample data is used for indicating the processing object of the first task;
the information extraction module is used for extracting at least one key point information of the first task from the description information, wherein the key point information is used for indicating the processing key point of the first task;
the information determining module is used for determining task prompt information corresponding to the first task according to the sample data and the at least one key point information, and the task prompt information is used for representing the processing requirement and the processing object of the first task;
And the result generation module is used for generating a processing result of the first task according to the task prompt information through a task processing model.
According to an aspect of an embodiment of the present application, there is provided a computer device including a processor and a memory, the memory having stored therein a computer program that is loaded and executed by the processor to implement the above-described method.
According to an aspect of an embodiment of the present application, there is provided a computer-readable storage medium having stored therein a computer program loaded and executed by a processor to implement the above-described method.
According to an aspect of an embodiment of the present application, there is provided a computer program product comprising a computer program stored in a computer readable storage medium. The processor of the computer device reads the computer program from the computer-readable storage medium, and the processor executes the computer program so that the computer device performs the above-described method.
The technical scheme provided by the embodiment of the application can have the following beneficial effects:
at least one key point information is extracted from the description information aiming at the first task, and task prompt information corresponding to the first task is determined by combining sample data of the first task. That is, the processing requirement of the first task is clarified by extracting the gist, and the description information which is not easy to be understood by the task processing model is converted into the task prompt information which is easy to be understood by the task processing model. Therefore, the method and the device are beneficial to the task processing model to quickly understand the processing requirement corresponding to the first task, and based on the processing requirement, the accuracy of the processing result generated by the task processing model on the first task is more accurate, and the processing requirement of the first task is more met.
Drawings
FIG. 1 is a schematic illustration of an implementation environment for an embodiment of the present application;
FIG. 2 is a schematic diagram of a task processing method according to an embodiment of the present application;
FIG. 3 is a schematic diagram of a task processing method according to another embodiment of the present application;
FIG. 4 is a schematic diagram of a task processing method according to another embodiment of the present application;
FIG. 5 is a flow chart of a task processing method provided by one embodiment of the present application;
FIG. 6 is a flow chart of a task processing method provided by another embodiment of the present application;
FIG. 7 is a flow chart of a task processing method provided by another embodiment of the present application;
FIG. 8 is a block diagram of a task processing method provided by one embodiment of the present application;
FIG. 9 is a block diagram of a task processing device provided by one embodiment of the present application;
FIG. 10 is a block diagram of a task processing device provided in another embodiment of the present application;
FIG. 11 is a block diagram of a computer device according to one embodiment of the present application.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the present application more apparent, the embodiments of the present application will be described in further detail with reference to the accompanying drawings.
Before describing the technical scheme of the application, a few background technical knowledge related to the application is described. The following related technologies may be optionally combined with the technical solutions of the embodiments of the present application, which all belong to the protection scope of the embodiments of the present application. Embodiments of the present application include at least some of the following.
Artificial intelligence (Artificial Intelligence, AI for short) is a theory, method, technique, and application system that simulates, extends, and extends human intelligence using a digital computer or a machine controlled by a digital computer, perceives the environment, obtains knowledge, and uses the knowledge to obtain optimal results. In other words, artificial intelligence is an integrated technology of computer science that attempts to understand the essence of intelligence and to produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence, i.e. research on design principles and implementation methods of various intelligent machines, enables the machines to have functions of sensing, reasoning and decision.
The artificial intelligence technology is a comprehensive subject, and relates to the technology with wide fields, namely the technology with a hardware level and the technology with a software level. Artificial intelligence infrastructure technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, mechatronics, and the like. The artificial intelligence software technology mainly comprises natural language processing technology, machine learning/deep learning and other directions.
Machine Learning (ML) is a multi-domain interdisciplinary, and involves multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, and the like. It is specially studied how a computer simulates or implements learning behavior of a human to acquire new knowledge or skills, and reorganizes related knowledge structures to continuously improve own performance. Machine learning is the core of artificial intelligence, a fundamental approach to letting computers have intelligence, which is applied throughout various areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, induction learning, teaching learning, and the like.
Deep Learning (DL) is a new research direction in the field of Machine Learning (ML), and is introduced into Machine Learning to make it closer to the original goal, i.e., artificial intelligence. Deep learning is the inherent regularity and presentation hierarchy of learning sample data, and the information obtained during such learning is helpful in interpreting data such as text, images and sounds. Its final goal is to have the machine have analytical learning capabilities like a person, and to recognize text, image, and sound data. Deep learning is a complex machine learning algorithm that achieves far greater results in terms of speech and image recognition than prior art. Deep learning has achieved many results in search technology, data mining, machine learning, machine translation, natural language processing, multimedia learning, speech, recommendation and personalization techniques, and other related fields. The deep learning makes the machine imitate the activities of human beings such as audio-visual and thinking, solves a plurality of complex pattern recognition problems, and makes the related technology of artificial intelligence greatly advanced.
Natural language processing (Nature Language Processing, NLP for short) is an important direction in the fields of computer science and artificial intelligence. It is studying various theories and methods that enable effective communication between a person and a computer in natural language. Natural language processing is a science that integrates linguistics, computer science, and mathematics. Thus, the research in this field will involve natural language, i.e. language that people use daily, so it has a close relationship with the research in linguistics. Natural language processing techniques typically include text processing, semantic understanding, machine translation, robotic questions and answers, knowledge graph techniques, and the like.
With research and advancement of artificial intelligence technology, research and application of artificial intelligence technology is being developed in various fields, such as common smart home, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned, autopilot, unmanned, robotic, smart medical, smart customer service, virtual Reality (VR), augmented Reality (Augmented Reality, AR), games, virtual persons, digital persons, etc., and it is believed that with the development of technology, artificial intelligence technology will be applied in more fields and with increasing importance value.
The scheme provided by the embodiment of the application relates to the technologies of artificial intelligence such as machine learning, natural language processing and the like, and is specifically described by the following embodiment.
Before describing the technical scheme of the application, some nouns related to the application are explained. The following related explanations may be optionally combined with the technical solutions of the embodiments of the present application, which all belong to the protection scope of the embodiments of the present application. Embodiments of the present application include at least some of the following.
1. Pre-training large model (PTM): refers to deep learning models that are pre-trained on large-scale data sets, typically containing billions of parameters and tens of millions of samples. In some embodiments, the pre-training model, also called a kerbstone model, a large model, refers to a deep neural network (Deep neural network, abbreviated as DNN) with large parameters, which is trained on massive unlabeled data, PTM extracts common features on the data by utilizing the function approximation capability of the large-parameter DNN, and is suitable for downstream tasks through fine tuning (fine tune), efficient fine tuning (PEFT), prompt-tuning and other technologies. Therefore, the pre-training model can achieve ideal effects in a small sample (Few-shot) or Zero sample (Zero-shot) scene. PTM can be classified according to the data modality of the process into a language model (ELMO, BERT, GPT), a visual model (swin-transducer, viT, V-MOE), a speech model (VALL-E), a multi-modal model (ViBERT, CLIP, flamingo, gato), etc., wherein a multi-modal model refers to a model that builds a representation of the characteristics of two or more data modalities. The pre-training model is an important tool for outputting Artificial Intelligence Generation Content (AIGC), and can also be used as a general interface for connecting a plurality of specific task models. The purpose of the pre-trained large model is to extract a generic language representation by learning patterns and rules in a large-scale dataset for fine-tuning and migration learning in various natural language processing tasks. Pretraining large models are often pretrained in an unsupervised learning manner, such as a self-encoder, language model, etc. In the pre-training stage, the model extracts general language representations, such as word vectors, sentence vectors and the like, by learning language structures and semantic information in a large-scale dataset. During the fine-tuning and transition learning phase, the model may be fine-tuned over a small amount of annotation data to accommodate specific natural language processing tasks, such as text classification, emotion analysis, machine translation, etc. Pre-training large models have achieved great success in the field of natural language processing, such as BERT (Bidirectional Encoder Representation from Transformers, language representation model), GPT-2, XLNet, etc., models, which have led to advanced performance in various natural language processing tasks. The pre-training large model has the advantage that the universal language representation can be extracted by utilizing language structure and semantic information in a large-scale data set, so that the generalization capability and performance of the model are improved.
Chatgpt: is a natural language processing technology based on a large-scale pre-training language model, and is developed by OpenAI. It is one of the most advanced dialog generation models at present, and can generate high-quality, coherent and logical natural language dialogs. The core of the ChatGPT is a neural network model based on a transducer architecture, and the neural network model can be pre-trained by using a large-scale corpus, so that rich language structure and semantic information can be learned. It has the following advantages: (1) a pre-training model: chatGPT is developed based on a large-scale Pre-training language model (GPT for short), and can be Pre-trained by using a large-scale corpus, so that abundant language structures and semantic information can be learned. This allows the ChatGPT to generate a high quality, consistent, logical natural language dialogue with high language understanding and generating capabilities. (2) context awareness: the ChatGPT can generate a dialog from the context information, i.e., can generate subsequent dialog content from the previous dialog content. This allows the ChatGPT to generate a more consistent, natural conversation while also better understanding the user's intent and needs. (3) diversity generation: the ChatGPT can generate various dialogue contents, namely, can generate various replies, thereby increasing the diversity and the interestingness of the dialogue. This allows the ChatGPT to better meet the needs and interests of the user. (4) The ChatGPT can generate dialogue content according to the requirements and preferences of users, namely, can control the theme, emotion, mood and other aspects of the dialogue. This allows the ChatGPT to better meet the personalized needs and scenarios of the user. The technical scheme provided by the embodiment of the application can be realized by using the ChatGPT, or not by using the ChatGPT, but by using a medium-small model.
3. Task mining: task mining through natural language descriptions is a natural language processing technique that aims to automatically identify tasks that a user needs to complete from natural language descriptions provided by the user and convert them into computer-executable tasks. The technology can help the user to complete various tasks more conveniently, and improves the user experience and efficiency. Meanwhile, the system can help enterprises and organizations to better understand the demands of users and provide better services and support. The application scene of task mining is very extensive, such as intelligent customer service, intelligent assistant, intelligent house, etc. In the intelligent customer service field, users can describe own problems through voice or words, and the system can automatically recognize the intention of the users and provide corresponding solutions. In the field of intelligent assistants, users can describe their own needs through voice or words, and the system can automatically recognize the intention of the user and help the user to complete corresponding tasks. In the field of intelligent home, a user can describe own requirements through voice or words, and the system can automatically recognize the intention of the user and control corresponding home equipment. In short, task mining based on user descriptions is a very useful natural language processing technology, which can help users to complete various tasks more conveniently, improve user experience and efficiency, and help enterprises and organizations to understand user demands better and provide better services and support.
4. Finally, illustrative explanation is made on nine basic tasks (4.1-4.9) related to natural language processing technology in the embodiment of the application, and the nine basic tasks almost completely cover the knowledge breadth of human understanding on natural language. The predefined tasks in the embodiments described below include, but are not limited to, at least one of these nine basic tasks.
4.1 language model: a language model is a model for predicting the next word or character. It can predict the probability of the next word or character from the previous word or character. The language model may be used for tasks such as automatic text generation, machine translation, speech recognition, etc. Common language models include n-gram models, recurrent neural network (Recurrent Neural Network, RNN) models, transducer models, and the like.
4.2 part-of-speech tagging: part of speech tagging is the process of tagging each word in text as its part of speech. The part-of-speech tagging can be used for text classification, information extraction, and other tasks. Common part-of-speech tagging methods include rule-based methods, statistical-based methods, and deep learning-based methods.
4.3 named entity recognition: named entity recognition is the process of recognizing entities in text (e.g., person names, place names, organization, etc.). Named entity recognition can be used for tasks such as information extraction, question-answering systems and the like. Common named entity recognition methods include rule-based methods, statistical-based methods, and deep learning-based methods.
4.4 syntax analysis: syntactic analysis is the process of analyzing sentence structure in text. Syntactic analysis can be used for text classification, information extraction, etc. Common syntactic analysis methods include rule-based methods, statistical-based methods, and deep learning-based methods.
4.5 semantic analysis: semantic analysis is the process of analyzing the meaning in text. Semantic analysis can be used for tasks such as emotion analysis, question and answer systems and the like. Common semantic analysis methods include rule-based methods, statistical-based methods, and deep learning-based methods.
4.6 machine translation: machine translation is the process of translating one language into another. Machine translation may be used for cross-language communication, text translation, and like tasks. Common machine translation methods include rule-based methods, statistical-based methods, and deep learning-based methods.
4.7 text classification: text classification is the process of classifying text into different categories. Text classification may be used for spam filtering, emotion analysis, and other tasks. Common text classification methods include rule-based methods, statistical-based methods, and deep learning-based methods.
4.8, information extraction: information extraction is the process of extracting useful information from text. The information extraction can be used for the tasks of knowledge graph construction, question-answering system and the like. Common information extraction methods include rule-based methods, statistical-based methods, and deep learning-based methods.
4.9 question and answer system: the question-answering system is a process of finding an answer from text and returning the answer to a user according to a question posed by the user. The question and answer system can be used for intelligent customer service, intelligent assistant and other tasks. Common question-answering system methods include rule-based methods, statistical-based methods, and deep learning-based methods.
Referring to fig. 1, a schematic diagram of an implementation environment of an embodiment of the present application is shown. The implementation environment of the scheme can comprise: a terminal device 10 and a server 20.
The terminal device 10 includes, but is not limited to, a mobile phone, a tablet computer, an intelligent voice interaction device, a game console, a wearable device, a multimedia playing device, a PC (Personal Computer ), a vehicle-mounted terminal, an intelligent home appliance, and the like. A client of a target application (e.g., a game application) may be installed in the terminal device 10. Alternatively, the target application may be an application that needs to be downloaded and installed, or may be a point-and-use application, which is not limited in the embodiment of the present application.
In the embodiment of the application, the target application program can be any one of a task processing application program, a search type application program, a query type application program, a social type application program and a simulation program. In addition, the tasks that need to be handled are different for different applications. To the extent that the target application is a task processing application, it may be an application tailored to the enterprise or individual, the task processing application may respond to user descriptive information for a particular type of task and output processing results for that task based on given sample data. Alternatively, the terminal device 10 has a client running the above-described target application. In other embodiments, the target application may also be a web page, and on the terminal device 10 running the web page, in response to the description information of the user for the specific task, and in combination with the sample data, the web page may output the processing result corresponding to the specific task.
The server 20 is used to provide background services for clients of target applications in the terminal device 10. For example, the server 20 may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or may be a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDNs (Content Delivery Network, content delivery networks), and basic cloud computing services such as big data and artificial intelligence platforms, but not limited thereto.
The terminal device 10 and the server 20 can communicate with each other via a network. The network may be a wired network or a wireless network.
In the method provided by the embodiment of the application, the execution main body of each step can be computer equipment, and the computer equipment refers to electronic equipment with data calculation, processing and storage capabilities. The computer device may be the terminal device 10 in fig. 1 or the server 20.
Before the technical scheme provided by the embodiment of the application is introduced, some task mining methods and corresponding defects in the related art are briefly described. Task mining based on user descriptions and provided data is a complex natural language processing task that needs to be implemented in combination with a variety of techniques and methods. The different techniques and methods have respective advantages and disadvantages, and a proper method needs to be selected according to specific application scenes and requirements.
1. Rule-based methods. This approach uses manually designed rules to identify the user's intent and task, while task mining is performed in conjunction with the data provided. For example, rules such as regular expressions, finite state machines, etc. may be used to extract key information in the user description and then task mining is performed in conjunction with the provided data. The advantage of this approach is simplicity and intelligibility, but requires a lot of manual design and maintenance rules, and it is difficult to cope with complex natural language descriptions and data. Disadvantages: rule-based method: information in text is identified and extracted by manually defining rules. He has stronger interpretability mainly by a manually defined method, can be applied to various texts, comprises structured and unstructured data, and has wider application range. However, the manual operation is very dependent, so that knowledge and experience in many professional fields are required, a lot of time and effort are required, and each set of rules is generated for a specific task and has no mobility.
2. Statistical-based methods. This approach uses machine learning algorithms to learn the user's intent and tasks while performing task mining in conjunction with the data provided. For example, naive Bayes, support vector machines, etc. algorithms may be used to learn the relationships between user descriptions and provided data, followed by task mining. The method has the advantages of automatically learning rules and features, being applicable to complex natural language description and data, but requiring a large amount of labeling data and feature engineering. Disadvantages: statistical-based methods: it identifies and extracts information in text by statistical analysis of large amounts of data. The method has high automation degree and stronger expansibility and portability, but because the statistical method is usually a black box model, the intrinsic mechanism of the model is difficult to interpret and understand, high-quality data is required, the effect of the model is easy to influence, and the conditions have certain difficulty in practical application scenes.
3. A method based on a graph neural network. This approach uses a graph neural network to learn the relationships between user descriptions and provided data, and then performs task mining. For example, algorithms such as a graph convolution neural network, a graph annotation neural network, etc. may be used to learn the relationships between user descriptions and provided data, and then perform task mining. The method has the advantages of automatically learning rules and features, being applicable to complex natural language description and data, and processing graph structure data, but requiring a large amount of computing resources and data. Disadvantages: the method based on the graph neural network comprises the following steps: it recognizes and extracts information in text by building a graph structure on the text and learning and inferring using a graph neural network. The method can be expanded and improved by increasing the number of nodes and edges, changing model parameters and the like so as to adapt to different tasks and applications. But require a large amount of computing resources to train and infer the model and therefore require several times more computing and storage power.
4. A method based on knowledge graph. This method uses knowledge maps to represent the relationships between user descriptions and provided data, and then performs task mining. For example, ontology techniques and the like can be used to construct knowledge maps, and then the user description and provided data are mapped into the knowledge maps for task mining. The method has the advantages that abundant domain knowledge and semantic information can be utilized, the accuracy and efficiency of task mining are improved, but a great deal of knowledge engineering and maintenance are required. Disadvantages: the method based on the knowledge graph comprises the following steps: it recognizes and extracts information in text by constructing and using knowledge maps. The method can effectively utilize the existing knowledge base, and can effectively dig out different tasks by adding entities and relations, but the construction difficulty of the knowledge graph is high, people with professional field knowledge are required to construct and maintain, and the later iteration cost is high.
The technical scheme provided by the embodiment of the application can take ChatGPT as an example to explain how to utilize a large pre-training model to perform task mining by disassembling the points to form a task template, and hopefully has the advantages of a rule-based method, a statistical-based method, a graph neural network-based method and a knowledge graph-based method, and brings promotion in the following aspects: (1) faster product development speed: the pre-training model is used for reducing the time and cost of model development, and can be directly used for task mining because the model is pre-trained on large-scale data, so that the development time of the model is greatly reduced, and the effect is realized rapidly. (2) more accurate capture requirements: the ChatGPT is used for task mining, and related information of a context can be effectively utilized through self-correction of multiple rounds of conversations, so that user requirements can be recognized more accurately, and a task structure can be generated better. (3) better resource utilization: zero sample learning can be realized by using ChatGPT, more training samples of a mobile phone are not needed, and the resource utilization rate is greatly improved. (4) better generalization ability: task mining using ChatGPT can better handle complex language tasks such as generating text, machine translation, text classification, emotion analysis, etc. In a word, the performance and the robustness of the model can be improved by using the ChatGPT to perform task mining, so that the requirements of users can be better completed, and the effect of each downstream task can be conveniently improved. In order to achieve the above purpose, the application provides three steps of task mining based on ChatGPT: 1. the description and partial data of the user are comprehensively judged through the ChatGPT, and six major points of actions, targets, entities, interaction modes, constraints and limitations and abnormal processing are extracted to serve as key bases of tasks. 2. And generating a prompt (prompt word) of a task template through six major points, wherein the prompt is used as a matching sample of a task library, and meanwhile, the data expansion of fewer samples is facilitated in the later stage. 3. The template is matched with the offline task library again by using the ChatGPT, mapped to the basic task, meanwhile, for some complex situations of multitasking, a prompt optimization strategy is adopted to realize the integrated expression of the multitasking. Therefore, the full process of the task which is rapidly realized under the assistance of the large model is realized while the task under the complex scene is excavated. And the task prompt information obtained after the process is used as the input of a large-scale pre-training model to complete a full-automatic task process and realize closed loop. In the whole process, complex information is filtered out, the influence of the pre-training model on noise information is reduced, so that the accuracy of a result is ensured, the whole process is ensured to be greatly improved in efficiency and effect under the help of strong knowledge capacity of the pre-training model.
In the technical scheme provided by the embodiment of the application, the technical scheme provided by the embodiment of the application can be realized by adopting a large-scale pre-training model such as ChatGPT, and at the moment, because the ChatGPT is a relatively perfect pre-training model, the accuracy and precision of task mining and task processing can be effectively ensured by continuing the technical scheme provided by the embodiment of the application on the basis. In this case, the ChatGPT is considered to include at least one of the following point extraction model, matching model, and task processing model. Correspondingly, a large-scale pre-training model like ChatGPT can be omitted, and a plurality of small and medium-sized models can be designed to realize the technical scheme provided by the embodiment of the application, namely the following key point extraction model, the matching model and the task processing model are all small and medium-sized models. At the moment, the technical scheme of the application is realized by adopting the small and medium models, so that the use cost of the models can be effectively reduced, and the models of all parts are independent, so that the model is relatively simple to train and use. Therefore, on the basis of ensuring the use cost, the relation between the precision of the processing result output by the model and the use requirement (the requirement of a user on the precision) is balanced. In summary, the technical solution provided by the embodiment of the present application is not limited to what type of model is specifically used.
Furthermore, the embodiment of the application adopts a mode of ChatGPT+small and medium model to carry out task mining and task processing. In some embodiments, when the middle-size and small-size models are adopted to perform task mining, the steps adopted in the following embodiments are used to determine task prompt information corresponding to the first task according to sample data and at least one key point information. Considering that a developer designs a universal task template aiming at a first task, the task template can be suitable for any task related to natural language processing, although task prompt information formed by assembling different gist information after filling the universal task template can characterize different task requirements. However, considering the design of the general task template, there is a high probability that the task prompt information for the first task is not smooth. In order to enable task prompt information to better represent task demands, in the embodiment of the application, chatGPT is adopted to further optimize the expression of the task prompt information. The specific flow is as follows: and extracting at least one key point information in the description information by using the key point extraction model, filling the key point information into a task template, and assembling to form task prompt information. After the task prompt information is obtained, on one hand, the basic task which is most matched with the basic task is determined from nine basic tasks, and the task processing model is finely tuned by utilizing a task database of the basic task which is most matched with the basic task. On the other hand, the ChatGPT is used as a task template optimizer, the task prompt information successfully assembled is used as the input of the ChatGPT, the task templates corresponding to the nine basic tasks are used as conditions, and the task templates are also provided for the ChatGPT, so that the ChatGPT can learn the context-consistent semantic information in the task templates corresponding to the nine basic tasks, and accordingly the task prompt information corresponding to the first task is optimized, and the optimized task prompt information is obtained. Optionally, the optimized task prompt information is input into the task processing model after fine adjustment, and a processing result corresponding to the first task is obtained. Through the mode of the key point extraction model plus ChatGPT plus task processing model, the task prompt information can be linked smoothly, the readability is stronger, and therefore the task prompt information after optimization is processed by the task processing model, and more accurate processing results are obtained.
The specific task mining and task processing procedures in the present application are explained with reference to the following examples.
Referring to fig. 2, a schematic diagram of a task processing method according to an embodiment of the application is shown.
In the fields of machine learning and deep learning, firstly, the intention of a user is precisely mined through language description of the user, so that the existing task library is selected and matched to complete the construction of a downstream task template, and a proper task model is trained. However, in the actual process, because of different background texts of users, description and provided data may have ambiguity, tasks may involve complicated processes and other problems, however, most models cannot achieve the effect of GPT4, and tasks cannot be accurately mined directly from such random and diverse information, so that the task mining technology has great challenges. In order to solve the problem, the technical scheme provided by the embodiment of the application starts to try to understand the intention and task of the user by using a pre-trained large-scale language model (also called as a key point extraction model in the scheme), and uses the pre-trained large-scale language model to understand the intention and task of the user, so that the method can adapt to the conditions under various different scenes, greatly improves the robustness of the model, and further helps the user to better solve the practical problem. Meanwhile, after the task is precisely excavated, the extracted task template can be used for inputting a pre-training model (the scheme can also be called a task processing model or a task model realization), and the task requirements of a user are completed by means of strong knowledge features.
As shown in 210 of fig. 2, in the present application, description information and sample data of a first task are first acquired, and at least one gist information is acquired from the description information. And then, filling the point information and the sample data into a task template to obtain task prompt information corresponding to the first task. The task prompt information corresponding to the first task is passed through the task processing model 200 to obtain a processing result of the first task (for a specific explanation, see the following embodiment).
Fig. 3 is a schematic diagram of a task processing method according to another embodiment of the application. As shown in 300 of fig. 3, the description information for the first task is input into the gist extraction model to obtain at least one gist, the at least one gist is filled into the task template, meanwhile, the sample data is combined to obtain task prompt information, and the obtained task prompt information is input into the task processing model to obtain a processing result of the first task. At this time, the gist extraction model and the task processing model are both machine learning models after training.
Referring to fig. 4, a schematic diagram of a task processing method according to another embodiment of the application is shown. As shown in 400 of fig. 4, the description information for the first task is input into the gist extraction model to obtain at least one gist, and the at least one gist is filled into the task template, and meanwhile, the sample data is combined to obtain the task prompt information. At this time, the task prompt information is matched with the task templates corresponding to the nine basic tasks respectively to obtain matched basic tasks, and after the matched basic tasks are obtained, the task processing model is finely tuned by utilizing the task database of the matched basic tasks, so that the task processing model after fine tuning can have a correct task processing direction. And inputting the task prompt information of the first task into the task processing model after fine adjustment to obtain a processing result of the first task. The key point extraction model is a machine learning model with pre-training.
The embodiment of the application provides a task mining method for performing task mining based on a large-scale pre-training model (key point extraction model), which is used for identifying and extracting information in a text by pre-training on large-scale data and then fine-tuning on a specific task. And further processes the mined tasks using the task processing model. The proposal at least comprises the following advantages: (1) high efficiency. The method based on the large-scale pre-training model can be used for pre-training by utilizing large-scale data, so that a better effect is obtained, and meanwhile, fine adjustment is performed on specific tasks, so that the method can be rapidly adapted to different tasks and applications. (2) high precision. The method based on the large-scale pre-training model can utilize large-scale data to perform pre-training, so that better generalization capability and precision are obtained, and fine adjustment is performed on a specific task, so that the accuracy of the model can be further improved. (3) high mobility. The method based on the large-scale pre-training model can transfer the learned model and parameters to other tasks and applications so as to improve the efficiency and accuracy of the model. (4) scalability is strong. The method based on the large-scale pre-training model can be expanded and improved by increasing the data volume, changing the model parameters and the like so as to adapt to different tasks and applications. (5) has a strong interpretability. Methods based on large-scale pre-trained models can interpret and understand the internal mechanisms and inference processes of the model by visualizing the model and parameters. In summary, the technical scheme provided by the embodiment of the application has the advantages of high efficiency, high precision, strong mobility, strong expandability, strong interpretability and the like.
Referring to fig. 5, a flowchart of a task processing method according to an embodiment of the application is shown. The main execution body of each step of the method may be the terminal device 10 described above or the server 20. In the following method embodiments, for convenience of description, only the execution subject of each step is described as "computer device". The method may comprise at least one of the following steps (510-540):
step 510, obtaining description information and sample data of a first task, wherein the first task is a task to be processed related to natural language processing, the description information is used for indicating a processing requirement of the first task, and the sample data is used for indicating a processing object of the first task.
Description information: it is understood that the processing requirements for the first task entered by the user. The embodiment of the application is not limited to the content, format and the like of the description information. The user can use text, voice, pictures and the like as descriptive information. In some embodiments, the descriptive information is obtained based on text, voice, picture, etc. information entered by the user. Alternatively, when the user input is text, the text input by the user is directly used as descriptive information. Optionally, when the input of the user is voice, the voice input by the user is converted into text by a mode of converting the voice into text, and the text is used as the description information. Optionally, when the input of the user is a picture, the description information is acquired from the picture input by the user by means of picture content extraction. In some embodiments, the requirements of different types of users or businesses are different, and the corresponding descriptive information is different.
Sample data: a processing object for indicating a first task. Optionally, the sample data has a correspondence with the first task. For example, there is first sample data, then there is a first task based on the sample data, and then there is descriptive information for the first task. This means that when the sample data changes, the first task also changes. Accordingly, when the first task is changed, the corresponding sample data also needs to be provided.
The following exemplary description is made with respect to the above-described first task, description information, and sample data.
Scene 1: the insurance company receives the insurance report information every day. The first task may be "determine the number, address, reason for reporting case" of the mobile phone of the claimant from the insurance report information. At this time, the description information may be "hello, i is an insurance company salesman, i can receive many insurance report information every day, i need to have to review these information, then extract the mobile phone number, address and reason of the report person, and check the report with the client step by step according to these information. Because a lot of information needs to be contacted every day, a lot of time is wasted in this respect, you need to help me pick up these, and send me in the form of a table with the order of the cell phone number, address, reason, and for the cell phone number in which the information does not meet the rules, you want to tell me and mark out, if you cannot recognize that the mark is empty. Accordingly, the sample data is "daily received insurance report information". For example, one sample data is "I are administrative staff of Shenzhen xx electronic technology Co., ltd., I am an employee, and sudden heart disease is taken during working hours, and the staff is sent to Shenzhen xx people's hospitals to die after treatment is not effective. By proposal, our policy number is xxxxxxx and my contact is 123456789. The insured person: shenzhen City xx electronic technology Co. The reason for the danger: sudden heart disease. Treatment hospital: shenzhen City xx people Hospital. Time of risk: when working. Number of policy: xxxxxxx. Contact phone: 123456789".
Scene 2: a large number of patients in hospitals reserve for medical treatment every day, and the number of on-duty doctors needs to be arranged according to the number of reserved medical treatment people in each department every day. The first task may be "determine the scheduled number of visits in each department from patients scheduled to visit each day". The descriptive information may be "hello, i is a medical staff, i have a lot of patient appointment information every day in our hospital, i need to review this information, and extract the name, age and department of appointment, i need to count this information. Please help me count the appointment department, the name of the patient appointment and the age of the patient appointment and transmit the information to me in text form. Accordingly, the sample data is "appointment information of patients daily". Illustratively, one sample data is "I are xxx, age xx years, appointment x month x day 8 a.m. to you hospital department. Patient number: xxx. Patient name: xxx. Age of patient: xx. Appointment time of visit: 8 points in the morning on the x month and the x day. Appointment department of medical science: internal medicine.
Scene 3: in an on-line examination scenario, multiple examinations exist, and a large number of answer sheets are provided for each examination and submitted by the answer sheets, so that a batch of paper teachers need to be allocated to review the electronic answer sheets. First task: and distributing the electronic answer sheets to be read for the teacher according to the teaching information of the batch teacher. At this time, the descriptive information may be "hello, i is an examination holder, my holds xx examination in 6.7-6.9, and involves multiple subjects, i need to distribute a batch of paper teachers for all the test takers' electronic papers. Please help me distribute all the test takers' electronic answer sheets according to the teaching information of the various volume teacher. Note that at the same time, batch time needs to be allocated, and batch tasks cannot be scheduled during rest time. The electronic answer sheet numbers, the serial numbers of the batch teacher and the batch time are used as column names, the column names are transmitted to me in the form of a table, and for the electronic answer sheet numbers which do not accord with rules in the information, you hope to tell me and mark out, if the mark is empty, the mark is not recognized. At this time, the sample data is "the electronic answer sheet number and the subject corresponding to the electronic answer sheet, the name, serial number, and the teaching information of the batch teacher".
As can be seen from the above-described scenario, sample data corresponding to a task needs to be acquired for different task requirements. If only a task is required and there is no sample data (processing object of the task), there is a high possibility that the task cannot be processed or the processing result of the task is inaccurate. Generally, the technical scheme provided by the embodiment of the application can be understood to design task processing models for scenes with different requirements. For example, in the above-mentioned insurance scenario, since sample data (insurance report information) is obtained, any problem in the insurance report information may be a first task, and the task processing model has an ability to process the first task. Accordingly, if there is no insurance report information, but only the description information of the first task, the task processing model may also obtain the processing result, but inevitably results in an inaccurate processing result. Therefore, in the technical scheme provided by the embodiment of the application, the first task is solved by combining the description information of the first task with the processing object corresponding to the first task, so that different requirements of different scenes can be met, customized requirements are realized, and the precision of the generated result is ensured.
Step 520, extracting at least one gist information of the first task from the description information, wherein the gist information is used for indicating a processing gist of the first task.
In a practical scenario, the available information mainly includes a description of the user's own needs, data introduction, and a part of the provided data samples, and the part of the information is often disordered and lacks regularity. Meanwhile, due to the background of different users and different description modes, if task extraction is directly carried out on the information, the information is easily affected by noise data, and meanwhile, the effect of a model is also affected. Therefore, the existing information needs to be summarized in a key point extraction mode to obtain the needed beneficial information. In order to extract key information from the user description and sample data, to better understand the user's needs and design solutions. The embodiment of the application designs the following points.
In some embodiments, the at least one point information includes at least one of: action information for indicating the operation to be completed by the first task; target information for indicating a result that the first task is expected to reach; entity information indicating a role in which the first task is desired to be performed; interaction mode information used for indicating interaction modes between the user and the task processing model; constraint restriction information indicating rules or restrictions to be adhered to when executing the first task; the exception handling information is used for indicating an exception condition and a corresponding handling mode when the first task is executed.
In some embodiments, the action information is also referred to as an action, and refers to an operation or behavior that a user needs to perform, such as searching, rewriting, evaluating, etc.
In some embodiments, the goal information is also referred to as a goal, which refers to a result or purpose that the user desires to achieve for the first task, such as querying for certain specific information, completing certain specific things, etc.
In some embodiments, entity information, also referred to as an entity, refers to an object or character that a user needs to operate, such as to act as a mental doctor, authoring assistant, etc.
In some embodiments, the interaction style information is also referred to as an interaction style, which refers to the way a user interacts with a system (task processing model), such as a command line, outputting a specific code, symbol, etc.
In some embodiments, constraint limit information, also referred to as constraints and limits, refers to rules or constraints that a user needs to follow when completing a task, such as restrictions on output formats, restrictions on answer scope, and so forth.
In some embodiments, the exception handling information is also called exception handling, and refers to an exception condition that may be encountered by a user when completing a task and a corresponding handling manner, such as error prompt, recovery mechanism, and the like.
The extracted gist information is shown below for each of the above-described scenes 1, 2, and 3.
For the above-described scene 1, the extracted gist includes at least one of the following. Action information: reading insurance report information; target information: extracting the mobile phone number, address and reason of the case from the given insurance case information; entity information: you need to act as an insurance salesman; interaction mode information: outputting a result table form; constraint limit information: the mobile phone number, address, reason sequence is column name, output is form of table; exception handling information: the mark is empty for unrecognized requests.
For the above-described scene 2, the extracted gist includes at least one of the following. Action information: review patient appointment information; target information: extracting a scheduled visit department, a patient name of the scheduled visit and a patient age of the scheduled visit from scheduled visit information of a given patient every day; entity information: you need to be a healthcare worker; interaction mode information: outputting a result text form; constraint limit information: the method is free; exception handling information: and no.
For the above-described scene 3, the extracted gist includes at least one of the following. Action information: distributing batch volume teachers for the electronic answer sheets of all examinees; target information: distributing all the electronic answer sheets of the examinees and distributing batch time according to the batch teacher's information in the subjects, the names and serial numbers of the batch teacher and the information in the batch corresponding to the given electronic answer sheet numbers and the electronic answer sheets; entity information: you need to act as an examination host; interaction mode information: outputting a result table form; constraint limit information: taking the sequence of the electronic answer sheet number, the batch teacher serial number and the batch time as column names, and not arranging batch tasks in rest time; exception handling information: the mark is empty for unrecognized requests.
It is not difficult to find that although the embodiment of the present application designs six major points, in the case where a part of the points cannot be extracted, an empty string may be set to characterize that the relevant content has not been extracted for the point information. The embodiment of the application is not limited to the mode of how to extract the key point information from the description information, and can extract the key point information from the description information in a machine learning mode.
In the embodiment of the application, the above key point extraction can filter out the disordered information, so that the method has the following advantages. (1) Under the condition of a large amount of information, the key content of the information can be quickly acquired through key point extraction, and the long, repeated and irrelevant information reading is avoided, so that the information acquisition efficiency is improved. (2) The information understanding process is simpler, more visual and easier to digest, so that the information absorbing and memorizing effects are improved. (3) The information understanding process is simpler, more visual and easier to digest, so that the information absorbing and memorizing effects are improved. On the basis of obtaining six main points, a solid foundation can be provided for the subsequent flow.
And 530, determining task prompt information corresponding to the first task according to the sample data and at least one key point information, wherein the task prompt information is used for representing the processing requirement and the processing object of the first task.
In some embodiments, the task suggestion information is determined based on descriptive information of the first task.
In some embodiments, the task hint information includes: task role information indicating a role in which the first task is desired to be executed; task summary information for indicating an operation that the first task needs to complete and a desired result; and the task instruction information is used for instructing the sample data provided by adopting a specified interaction mode to be processed under the specified requirement. Alternatively, the type of the task prompt information may be a task prompt word.
In some embodiments, the task role information is determined according to entity information in the gist information. In some embodiments, the entity information directly constitutes task role information. Taking scenario 1 above as an example, the task role information is "you need to be an insurance salesman". The other two scenarios, see scenario 1, are not described in detail.
In some embodiments, the task summary information is determined based on the action information in the gist information and the target information. In some embodiments, the action information and the target information are assembled in a format to obtain the task summary information. For example, for scenario 1 above, the task summary information is "you need to read insurance case information, and the mobile phone number, address, and reason for case report of the user are extracted from the given insurance case information. The other two scenarios, see scenario 1, are not described in detail.
In some embodiments, the task instruction information is determined based on the sample data from the interaction mode information, the constraint restriction information, and the exception handling information in the above-described gist information. In some embodiments, the sample data, the interaction mode information, the constraint limit information and the exception handling information are assembled in a certain format to obtain the task specification information. For example, for scenario 1 above, the task instruction information is "the given sample data is daily received insurance report information, your interaction is output results in tabular form, and there are the following requirements: the mobile phone number, address, reason sequence is column name, output is in form of table and marked empty for no recognition request. The other two scenarios, see scenario 1, are not described in detail.
In summary, the task role information, the task summary information, and the task instruction information are a combination of the extracted gist information and sample data. Optionally, the key point information is linked by using a certain linking word. Optionally, the linking words between the key point information are preset, or may be automatically generated, which is not limited by the present application.
In addition, through the steps 510 to 530, task mining of the first task based on the description information and the sample data is achieved, so that the mined task is a task that can be better processed by the model.
Step 540, generating a processing result of the first task according to the task prompt information through the task processing model.
The task processing model in the embodiment of the application is a machine learning model, and the training method of the task processing model is not limited. In some embodiments, the task processing model may be trained using a training sample set of a small number of samples, may be trained using reinforcement learning, and may be trained using contrast learning. Of course, the task processing model may also be fine-tuned using a task database of matched base tasks in the embodiments described below, see in particular the embodiments described below.
The application is also not limited with respect to the model architecture of the task processing model. The task processing model can be regarded as a black box, the input of the task processing model is task prompt information, and the output is the processing result of the first task. In some embodiments, the task processing model includes an encoder and a decoder, the task prompt information is encoded by the encoder, and the processing result is decoded from the task prompt information by the decoder. In some embodiments, the task processing model is a transducer model. In other embodiments, the task processing model is a large-scale pre-trained model, such as ChatGPT.
According to the technical scheme provided by the embodiment of the application, at least one key point information is extracted from the description information aiming at the first task, and the task prompt information corresponding to the first task is determined by combining the sample data of the first task. That is, the processing requirement of the first task is clarified by extracting the gist, and the description information which is not easy to be understood by the task processing model is converted into the task prompt information which is easy to be understood by the task processing model. Therefore, the method and the device are beneficial to the task processing model to quickly understand the processing requirement corresponding to the first task, and based on the processing requirement, the accuracy of the processing result generated by the task processing model on the first task is more accurate, and the processing requirement of the first task is more met.
Referring to fig. 6, a flowchart of a task processing method according to another embodiment of the present application is shown. The main execution body of each step of the method may be the terminal device 10 described above or the server 20. In the following method embodiments, for convenience of description, only the execution subject of each step is described as "computer device". The method may comprise at least one of the following steps (610-640):
In step 610, description information and sample data of a first task are obtained, the first task is a task to be processed related to natural language processing, the description information is used for indicating a processing requirement of the first task, and the sample data is used for indicating a processing object of the first task.
At step 620, at least one gist information of the first task is extracted from the description information through a gist extraction model, wherein the gist extraction model is a machine learning model with pre-training completed, and the gist information is used for indicating a processing gist of the first task.
The key point extraction model in the embodiment of the application is a machine learning model, and the training method of the key point extraction model is not limited. In some embodiments, the point extraction model may be trained using a training sample set of a small number of samples, may be trained using reinforcement learning, and may be trained using contrast learning.
The application is not limited to the model architecture of the point extraction model. The gist extraction model may be seen as a black box, the input of the gist extraction model being descriptive information and the output being at least one gist information. In some embodiments, the gist extraction model includes an encoder and a decoder, the description information is encoded by the encoder, and the at least one gist information is decoded from the description information by the decoder. In some embodiments, the point extraction model is a transducer model. In other embodiments, the point extraction model is a large-scale pre-trained model, such as ChatGPT.
In the technical scheme provided by the embodiment of the application, the key point extraction model is a machine learning model with pre-training. In the case where the gist extraction model is a large-scale pre-training model, first, accuracy of the gist extraction model and the task processing model can be improved. Task mining based on a large-scale pre-training model can utilize large-scale data to perform pre-training, so that better generalization capability and precision are obtained, and accuracy and efficiency of the key point extraction model and the task processing model can be improved. Secondly, the generalization capability of the task processing model can be improved. Task mining based on a large-scale pre-training model can utilize large-scale data to perform pre-training, so that the generalization capability of the model is improved, and different tasks and applications can be adapted. Furthermore, the interpretability of the task processing model is improved. Task mining based on a large-scale pre-training model can explain and understand the internal mechanism and the inference process of a task processing model through a visual model and parameters, so that the interpretability and the comprehensiveness of the task processing model are improved. Then, the efficiency of the gist extraction model and the task processing model can be improved. Task mining based on a large-scale pre-training model can utilize related pre-training models and parameters, so that training time and computing resources of the model are reduced, and efficiency and speed of the model are improved. Finally, the level of application intelligence is improved. Task mining based on a large-scale pre-training model can bring higher intelligent level to the application, and improve the competitiveness and user experience of the application.
Step 630, filling the sample data and at least one key point information into a task template to obtain task prompt information corresponding to the first task, wherein the task template is used for defining a standardized format of the task prompt information, and the task prompt information is used for representing a processing requirement and a processing object of the first task.
In some embodiments, the different first tasks correspond to a generic task template (a fixed task template), even if the first task changes, the task template does not change. Optionally, the task template is designed in advance. Optionally, the task template is "task role information: { entity information }; task summary information: { you need { action info }, want to implement { target info }; task instruction information: { given data is { input (format is consistent with offline database (sample data) } (there is no data, this part is no), your interaction is { interaction information }, and there are the following requirements: { constraint info } { exception handling info }. Optionally, the extracted key point information is directly filled into the task template, and sample data is filled into the task template. In some embodiments, the data may be selectively populated for some constraints and constraints of the descriptive information while populating the sample data. Alternatively, the generic task template is employed for scenario 1, scenario 2, and scenario 3 described above.
In other embodiments, different types of first tasks correspond to different task templates, and when the types of first tasks change, the task templates change. Alternatively, the task templates may be individually customized for different tasks at this time for the business or individual of the requirements. For example, a task template may be customized individually for an insurance company, a task template may be customized individually for a hospital, and a template may be customized individually for an examination host. At this time, the task template may be designed manually. Optionally, the above scenario 1, scenario 2, and scenario 3 correspond to different task templates.
And step 640, generating a processing result of the first task according to the task prompt information through the task processing model.
In the technical scheme provided by the embodiment of the application, a general task template is designed for at least one key point information, so that the key point information is directly filled into the task template after being extracted, and the task prompt information is automatically generated. Therefore, automation of the task realization mining is ensured. In addition, the method is a universal template, so that different task demands can be met, and even if the task templates are different, different task information prompt words can be still determined under the condition that the extracted key points are different, and different task demands are represented. Therefore, in the embodiment of the application, the common characteristics and the individual characteristics are combined when the task is excavated, so that different requirements can be met.
Of course, when task templates are designed for different types of tasks, diversity and flexibility of task mining in the embodiment of the application can be more embodied. Based on the task templates customized independently and the key point information determined based on the description information, the distinguishing strength of task demands and task demands is increased, so that the tasks are processed better, and the accuracy of task processing results is further improved.
Referring to fig. 7, a flowchart of a task processing method according to another embodiment of the present application is shown. The main execution body of each step of the method may be the terminal device 10 described above or the server 20. In the following method embodiments, for convenience of description, only the execution subject of each step is described as "computer device". The method may comprise at least one of the following steps (710-750):
step 710, obtaining description information and sample data of a first task, wherein the first task is a task to be processed related to natural language processing, the description information is used for indicating a processing requirement of the first task, and the sample data is used for indicating a processing object of the first task.
And step 720, extracting at least one key point information of the first task from the description information, wherein the key point information is used for indicating the processing key point of the first task.
Step 730, determining task prompt information corresponding to the first task according to the sample data and at least one key point information, where the task prompt information is used to characterize a processing requirement and a processing object of the first task.
Step 740, determining a first predefined task matched with the first task from at least one predefined task according to task prompt information corresponding to the first task and task template information corresponding to at least one predefined task respectively; each predefined task corresponds to a task type, and task template information corresponding to the predefined task is used for representing the task type corresponding to the predefined task.
In some embodiments, the predefined tasks include at least one of the nine major basic tasks described above. Optionally, a predefined task that matches the first task is matched from the nine basic tasks. Optionally, the nine major basic tasks correspond to nine task types. In other embodiments, since the nine basic tasks are more mature tasks, they correspond to a large amount of basic data (training data set). Optionally, task template information of the nine basic tasks is determined based on the basic data. Taking a predefined task as an example of a part-of-speech tagging task, task template information corresponding to the part-of-speech tagging task may be "we give a piece of text information { input }, please extract each word of the piece of text, and give the part-of-speech of each word, where the output format is { word: part of speech }). Optionally, the task template information corresponding to the nine large basic tasks may be summarized in the existing basic data, or may be summarized based on the basic data corresponding to the nine large tasks, and the source of the task template information corresponding to the predefined tasks is not limited. In other embodiments, the task template information is used to characterize a task type corresponding to a predefined task, and further, the task template information is also used to characterize input and output information of a task under the task type and intermediate processing flow information.
In some embodiments, for each of at least one predefined task, determining a degree of matching between task prompt information corresponding to the first task and task template information corresponding to the predefined task; and determining the predefined task corresponding to the maximum value of the matching degree as a first predefined task. Optionally, a first matching comparison is performed, and the first matching comparison directly finds out a predefined task corresponding to the maximum value of the matching degree from the total of M predefined tasks. Optionally, at least two matches are performed, the first match comparing a predefined task of K before matching is found out from a total of M predefined tasks. The second matching comparison uniquely identifies a best matching predefined task from the K predefined tasks. Wherein M is a positive integer, and K is a positive integer smaller than M. Alternatively, the first and second matching pairs employ different matching means.
In some embodiments, the matching means includes the following two.
First kind: calculating semantic similarity between task prompt information corresponding to a first task and task template information corresponding to a predefined task; the semantic similarity is determined as a match. Optionally, the semantic similarity is used for representing the similarity between the task prompt information corresponding to the first task and the task template information corresponding to the predefined task. The embodiment of the application is not limited to a way of calculating the semantic similarity. Optionally, the task prompt information corresponding to the first task and the task template information corresponding to the predefined task are respectively converted into feature vectors, and then distances between the feature vectors are calculated, where the distances may be cosine distances, euclidean distances, and the like. And regarding the distance between the feature vectors as the similarity between the task prompt information corresponding to the first task and the task template information corresponding to the predefined task.
Secondly, determining the matching degree through a matching model according to task prompt information corresponding to the first task and task template information corresponding to the predefined task; wherein the matching model is a machine learning model. The matching model is a pre-trained machine learning model for computing similarity. The input of the matching is task prompt information corresponding to the first task and task template information corresponding to the predefined task, and the output is the matching degree between the task prompt information corresponding to the first task and the predefined task. The application is not limited to the specific architecture of the matching model.
In other embodiments, when matching is performed, the sample data may not be filled into the task template, but only the information obtained after the essential point information is filled into the task template is considered as task prompt information, so as to match with the task template information corresponding to the predefined task, thereby obtaining a matching result. The matching mode does not match input data, but only matches specific task requirements, so that the matching quantity can be effectively reduced, the matching cost is reduced, and the matching efficiency is improved.
In some embodiments, the first predefined task that matches the task hints information for the first task in scenario 1 above is an information extraction task. In some embodiments, the first predefined task that matches the task cues of the first task in scenario 2 described above is a text classification task. In some embodiments, the text classification task that matches the task prompt for the first task in scenario 3 described above.
Step 750, adjusting parameters of the task processing model based on a task database corresponding to the first predefined task to obtain an adjusted task processing model; the task database corresponding to the first predefined task comprises model training data related to the first predefined task.
In the embodiment of the application, the parameters of the task processing model are adjusted by utilizing the task database corresponding to the first predefined task, which is also called fine tuning, and the first predefined task can be also called a matched basic task, a similar task and the like. The parameters of the task processing model are adjusted by utilizing the task database corresponding to the first predefined task, so that the processing direction of the task processing model is relatively correct, and the accuracy of the processing result is improved.
In some embodiments, parameters of the task processing model are adjusted by using model training data related to the first predefined task, which is included in a task database corresponding to the first predefined task, to obtain an adjusted task processing model. In some embodiments, the model training data comprises a training sample set. In some embodiments, the training samples include training samples and training tags. The model training data is optionally utilized to fine tune the task processing model. For example, when the first predefined task is a part-of-speech tagging task, a piece of model training data may be "input: the earth is round; and (3) outputting: earth-nouns, are-predicates, round-adjectives.
In some embodiments, the task database corresponding to the first predefined task further includes model parameters of a model related to the first predefined task. In some embodiments, inputting task prompt information corresponding to a first task into a task processing model to obtain a first output result of the task processing model; inputting task prompt information corresponding to the first task into a model related to the first predefined task to obtain a second output result of the model related to the first predefined task; and adjusting parameters of the task processing model by taking the difference value between the minimized first output result and the minimized second output result as a target to obtain an adjusted task processing model. In some embodiments, since the task database corresponding to the first predefined task further includes model parameters of the model related to the first predefined task, the model parameters are directly copied into another model to obtain a task model corresponding to the first predefined task, or a mature model of the first predefined task is directly taken for use. Optionally, inputting task prompt information corresponding to the first task into a task processing model to obtain a first output result of the task processing model; and inputting the task prompt information corresponding to the first task into a model related to the first predefined task to obtain a second output result of the model related to the first predefined task. And fine tuning the parameters of the task processing model in a comparison learning mode.
In other embodiments, there are multiple predefined tasks matched. For example, the first K predefined tasks of the match are taken. Optionally, parameter adjustment is performed on the task processing model by using task databases respectively corresponding to the K predefined tasks. In a practical scenario, one basic task often cannot sufficiently meet the needs of the user, so cooperation between multiple tasks needs to be considered. In the task matching process, if a plurality of basic tasks are found to be similar to the task prompt information of the first predefined task, the tasks are expected to be matched with each other, and the user is helped together in a multi-task flow mode. In some embodiments, the design of a multitasking auxiliary flow is aided by ChatGPT, which then needs to be done with an interactive multi-round dialog approach. In some embodiments, providing one-time generation of task cues and interactive generation are two different ways, differing in the way the task cues are generated and the quality of the results generated. The meaning of generating the answer once is that the ChatGPT can generate the complete answer once after receiving the input without multiple interactions. This means that the ChatGPT can directly generate a complete answer or text after one input without requiring multiple iterations or interactions by the user. For example, if a user enters a question, the ChatGPT may generate a complete answer at a time without requiring multiple iterations or interactions by the user. The one-time answer generation mode can improve the efficiency of the model and the experience of the user, and reduce the waiting time and the interaction times of the user. ChatGPT can make multiple rounds of conversations, i.e., continue to generate the next round of conversations based on the previous round of conversations. This means that ChatGPT can simulate the process of human conversation, making more complex and flexible interactions. The ChatGPT can generate corresponding answers or text according to user input, and can be adjusted and modified according to user feedback. This means that the ChatGPT can interact with the user in real time, and dynamically adjust and generate according to the needs and feedback of the user. For example, in a chat robot scenario, a user may input a question, the ChatGPT may generate an answer, then the user may feedback according to the content of the answer, and the ChatGPT may adjust and modify according to the user's feedback to generate a more accurate and reasonable answer. The interactive generation mode can improve the flexibility and adaptability of the model, and better meet the requirements and feedback of users. According to the embodiment of the application, on the basis of the process completion of the ChatGPT design, the realization of the user task is gradually completed in a multi-round dialogue mode according to the preset process, so that the multi-task requirement under a complex scene is realized, meanwhile, the correlation among tasks is effectively utilized, the context of task feedback is continuously optimized, the obtained processing result is continuously optimized, the user experience effect is improved, and the precision of the generated processing result is improved to a great extent.
Step 760, generating a processing result of the first task according to the task prompt information through the adjusted task processing model.
In some embodiments, after the task processing model is trimmed, an adjusted task processing model is obtained, and task prompt information is input to the trimmed task processing model to generate a processing result of the first task.
In the technical scheme provided by the embodiment of the application, the basic task almost completely covers the knowledge breadth of understanding natural language by human beings, so that a large amount of data and data are collected, a specific data format is summarized for the nine-big task of the basic, and a related data set is collected and downloaded to find a large amount of related task description. The information of the tasks with the mature knowledge system is arranged to be used as an offline task matching library, on the premise that task prompt information is generated through the above procedure, the information can be accurately and efficiently mapped into basic tasks in the offline task library only through a simple instruction of ChatGPT, and fine tuning training is carried out on a task processing model by utilizing a task database corresponding to the matched basic tasks, so that targeted iteration and optimization of the tasks are realized.
In summary, the task mining and task processing method provided in the embodiments of the present application may bring the following five improvements to the product side: (1) increasing the level of intellectualization of the product: task mining based on a large-scale pre-training model can utilize large-scale data to perform pre-training, so that better generalization capability and precision are obtained, a higher intelligent level can be brought to products, and the competitiveness of the products is improved. (2) enhancing the user experience of the product: task mining based on a large-scale pre-training model can provide more personalized and accurate service by analyzing the demands of users, so that the user experience of products is improved, and the viscosity and loyalty of the users are increased. (3) reducing the development cost of the product: the task mining based on the large-scale pre-training model can utilize the existing pre-training model and parameters, reduce the development cost and time of products, and simultaneously can be rapidly adapted to different tasks and applications by fine-tuning the model and parameters. (4) improving the data security of the product: the task mining based on the large-scale pre-training model can utilize the existing pre-training model and parameters, so that the dependence on user data is reduced, and the data security and privacy protection capability of the product are improved. (5) increase product scalability: task mining based on a large-scale pre-training model can be expanded and improved by means of increasing data volume, changing model parameters and the like so as to adapt to different tasks and applications, and therefore the expandability and adaptability of products are improved. In a word, task mining based on a large-scale pre-training model can bring higher intelligent level, better user experience, lower development cost and higher data security to the product side, and can help the product to obtain greater advantages in market competition.
The following outlines the optimization of the gist extraction model and the task processing model mentioned in the embodiments of the present application. (1) model structure optimization: the model can be optimized by changing the model structure, increasing the layer number, adjusting parameters and the like, and the accuracy and generalization capability of the model are improved. (2) data enhancement and cleaning: the data can be optimized by increasing the data quantity, improving the data quality, increasing the data diversity and the like, and the generalization capability and effect of the model are improved. (3) pretraining strategy optimization: the pre-training process can be optimized by changing the pre-training strategy, adding pre-training data, adjusting the pre-training parameters and the like, so that the generalization capability and effect of the model are improved. For example, the methods such as Lora, p-turn, etc., add more data, try different loss functions, and evaluate the index. (4) fine tuning strategy optimization: the fine tuning process can be optimized by changing the fine tuning strategy, adding fine tuning data, adjusting fine tuning parameters and the like, the accuracy and efficiency of the model are improved, for example, the data quality of a specific task is improved, the model is instructed to be adjusted by learning, a better regularization method is used and the like. (5) model compression and acceleration: the size and the speed of the model can be optimized through modes of model compression, pruning, quantization, acceleration and the like, the efficiency and the usability of the model are improved, and the model compression and quantization means that the model size is reduced and model reasoning is accelerated through the compression and quantization technology, so that the cost of the model in storage and calculation is reduced. Model compression typically includes pruning, low-rank decomposition, knowledge distillation, etc., and model quantization refers to converting floating point number parameters in a model to fixed point number or integer parameters, thereby reducing model size and speeding up model reasoning. In a word, aiming at task mining of a large-scale pre-training model, optimization can be performed from the directions of model structure optimization, data enhancement and cleaning, pre-training strategy optimization, fine tuning strategy optimization, model compression, acceleration and the like, so that the accuracy, generalization capability, efficiency and usability of the model are improved.
Referring to fig. 8, a flowchart of a task processing method according to another embodiment of the present application is shown. The main execution body of each step of the method may be the terminal device 10 described above or the server 20. In the following method embodiments, for convenience of description, only the execution subject of each step is described as "computer device". The method may comprise at least one of the following steps (810-860):
at step 810, at least one gist information is extracted from the description information of the first task.
Step 820, determining task description information of the first task in combination with the sample data of the first task.
In step 830, a predefined task that matches the first task is determined.
In some embodiments, after the first predefined task is determined, the parameter adjustment data of the last time is simultaneously retained in the task processing model, and the following step 840 is performed according to the parameter adjustment data of the last time in the task processing model. The parameter adjustment data at least comprises a first predefined task corresponding to the last model adjustment.
In step 840, whether the current task processing model has been trimmed using the task database of the matched predefined task.
If yes, go to step 860, if no, go to step 850.
At step 850, the current task processing model is trimmed using the task database of the matched predefined tasks.
Step 860, the task prompt information is input into the task processing model, and the processing result of the first task is obtained.
The following are examples of the apparatus of the present application that may be used to perform the method embodiments of the present application. For details not disclosed in the embodiments of the apparatus of the present application, please refer to the embodiments of the method of the present application.
Referring to fig. 9, a block diagram of a task processing device according to an embodiment of the present application is shown. The apparatus 900 may include: a data acquisition module 910, an information extraction module 920, an information determination module 930, and a result generation module 940.
The data acquisition module 910 is configured to acquire description information and sample data of a first task, where the first task is a task to be processed related to natural language processing, the description information is used to indicate a processing requirement of the first task, and the sample data is used to indicate a processing object of the first task.
An information extraction module 920, configured to extract at least one gist information of the first task from the description information, where the gist information is used to indicate a processing gist of the first task.
The information determining module 930 is configured to determine, according to the sample data and the at least one gist information, task prompt information corresponding to the first task, where the task prompt information is used to characterize a processing requirement and a processing object of the first task.
And a result generating module 940, configured to generate, according to the task prompt information, a processing result of the first task through a task processing model.
In some embodiments, the at least one point information includes at least one of: action information, which is used for indicating the operation which needs to be completed by the first task; target information for indicating a result that the first task is expected to reach; entity information indicating a role in which the first task is desired to be performed; interaction mode information used for indicating interaction modes between a user and the task processing model; constraint limit information indicating rules or constraint conditions to be adhered to when the first task is executed; the exception handling information is used for indicating the exception condition and the corresponding handling mode which are encountered when the first task is executed.
In some embodiments, the task hint information includes: task role information indicating a role in which the first task is desired to be executed; task summary information for indicating an operation that the first task needs to complete and a desired result; and the task instruction information is used for instructing the sample data provided by adopting a specified interaction mode to be processed under the specified requirement.
In some embodiments, the information determining module 930 is configured to populate the sample data and the at least one gist information into a task template to obtain task prompt information corresponding to the first task, where the task template is used to define a normalized format of the task prompt information.
In some embodiments, the information extraction module 920 is configured to extract at least one gist information of the first task from the description information through a gist extraction model, where the gist extraction model is a pre-trained machine learning model.
In some embodiments, as shown in fig. 10, the apparatus further comprises a task determination module 950, a model adjustment module 960.
The task determining module 950 is configured to determine, according to task prompt information corresponding to the first task and task template information corresponding to at least one predefined task, a first predefined task that is matched with the first task from the at least one predefined task; each predefined task corresponds to a task type, and task template information corresponding to the predefined task is used for representing the task type corresponding to the predefined task.
The model adjustment module 960 is configured to adjust parameters of the task processing model based on a task database corresponding to the first predefined task, so as to obtain an adjusted task processing model; the task database corresponding to the first predefined task comprises model training data related to the first predefined task.
And a result generating module 940, configured to generate, according to the task prompt information, a processing result of the first task through the adjusted task processing model.
In some embodiments, the task database corresponding to the first predefined task further includes model parameters of a model related to the first predefined task.
A model adjustment module 960, configured to adjust parameters of the task processing model by using model training data related to the first predefined task included in a task database corresponding to the first predefined task, so as to obtain the adjusted task processing model; or, inputting the task prompt information corresponding to the first task into the task processing model to obtain a first output result of the task processing model; inputting the task prompt information corresponding to the first task into a model related to the first predefined task to obtain a second output result of the model related to the first predefined task; and adjusting parameters of the task processing model with the aim of minimizing the difference between the first output result and the second output result to obtain the adjusted task processing model.
In some embodiments, the task determining module 950 is configured to determine, for each of the at least one predefined task, a degree of matching between task prompt information corresponding to the first task and task template information corresponding to the predefined task; and determining a predefined task corresponding to the maximum value of the matching degree as the first predefined task.
In some embodiments, the task determining module 950 is configured to calculate a semantic similarity between the task prompt information corresponding to the first task and the task template information corresponding to the predefined task; determining the semantic similarity as the matching degree; or determining the matching degree through a matching model according to the task prompt information corresponding to the first task and the task template information corresponding to the predefined task; wherein the matching model is a machine learning model.
Fig. 11 is a block diagram showing a structure of a computer device according to another exemplary embodiment of the present application.
In general, the computer device 1100 includes: a processor 1101 and a memory 1102.
The processor 1101 may include one or more processing cores, such as a 4-core processor, an 11-core processor, and the like. The processor 1101 may be implemented in at least one hardware form of DSP (Digital Signal Processing ), FPGA (Field Programmable Gate Array, field programmable gate array), PLA (Programmable Logic Array ). The processor 1101 may also include a main processor, which is a processor for processing data in an awake state, also called a CPU (Central Processing Unit ), and a coprocessor; a coprocessor is a low-power processor for processing data in a standby state. In some embodiments, the processor 1101 may integrate a GPU (Graphics Processing Unit, image processor) for rendering and drawing of content required to be displayed by the display screen. In some embodiments, the processor 1101 may also include an AI processor for processing computing operations related to machine learning.
Memory 1102 may include one or more computer-readable storage media, which may be tangible and non-transitory. Memory 1102 may also include high-speed random access memory, as well as non-volatile memory, such as one or more magnetic disk storage devices, flash memory storage devices. In some embodiments, a non-transitory computer readable storage medium in memory 1102 stores a computer program that is loaded and executed by processor 1101 to implement the methods provided by the various method embodiments described above.
Those skilled in the art will appreciate that the architecture shown in fig. 11 is not limiting as to the computer device 1100, and may include more or fewer components than shown, or may combine certain components, or employ a different arrangement of components.
In an exemplary embodiment, a computer readable storage medium is also provided, in which a computer program is stored which, when being executed by a processor, implements the above-mentioned method.
Alternatively, the computer-readable storage medium may include: ROM (Read-Only Memory), RAM (Random Access Memory ), SSD (Solid State Drives, solid state disk), or optical disk, etc. The random access memory may include, among other things, reRAM (Resistance Random Access Memory, resistive random access memory) and DRAM (Dynamic Random Access Memory ).
In an exemplary embodiment, a computer program product is also provided, the computer program product comprising a computer program stored in a computer readable storage medium. A processor of a computer device reads the computer program from the computer readable storage medium, and the processor executes the computer program so that the computer device performs the above-described method.
It should be noted that, in the present application, the collection and processing of related data (such as sample data, description information, etc.) should be strictly based on the requirements of relevant national laws and regulations during the application of the examples, so as to obtain the informed consent or independent consent of the personal information body, and develop the subsequent data use and processing actions within the authorized range of the laws and regulations and the personal information body.
It should be understood that references herein to "a plurality" are to two or more. "and/or", describes an association relationship of an association object, and indicates that there may be three relationships, for example, a and/or B, and may indicate: a exists alone, A and B exist together, and B exists alone. The character "/" generally indicates that the context-dependent object is an "or" relationship. In addition, the step numbers described herein are merely exemplary of one possible execution sequence among steps, and in some other embodiments, the steps may be executed out of the order of numbers, such as two differently numbered steps being executed simultaneously, or two differently numbered steps being executed in an order opposite to that shown, which is not limiting.
The foregoing description of the exemplary embodiments of the application is not intended to limit the application to the particular embodiments disclosed, but on the contrary, the intention is to cover all modifications, equivalents, and alternatives falling within the spirit and scope of the application.
Claims (13)
1. A method of task processing, the method comprising:
acquiring description information and sample data of a first task, wherein the first task is a task to be processed related to natural language processing, the description information is used for indicating the processing requirement of the first task, and the sample data is used for indicating a processing object of the first task;
extracting at least one gist information of the first task from the description information, wherein the gist information is used for indicating a processing gist of the first task;
determining task prompt information corresponding to the first task according to the sample data and the at least one key point information, wherein the task prompt information is used for representing the processing requirement and the processing object of the first task;
and generating a processing result of the first task according to the task prompt information through a task processing model.
2. The method of claim 1, wherein the at least one point information comprises at least one of:
Action information, which is used for indicating the operation which needs to be completed by the first task;
target information for indicating a result that the first task is expected to reach;
entity information indicating a role in which the first task is desired to be performed;
interaction mode information used for indicating interaction modes between a user and the task processing model;
constraint limit information indicating rules or constraint conditions to be adhered to when the first task is executed;
the exception handling information is used for indicating the exception condition and the corresponding handling mode which are encountered when the first task is executed.
3. The method of claim 2, wherein the task prompt information comprises:
task role information indicating a role in which the first task is desired to be executed;
task summary information for indicating an operation that the first task needs to complete and a desired result;
and the task instruction information is used for instructing the sample data provided by adopting a specified interaction mode to be processed under the specified requirement.
4. The method according to claim 1, wherein determining task prompt information corresponding to the first task according to the sample data and the at least one gist information includes:
And filling the sample data and the at least one key point information into a task template to obtain task prompt information corresponding to the first task, wherein the task template is used for defining a standardized format of the task prompt information.
5. The method of claim 1, wherein extracting at least one gist information of the first task from the description information comprises:
and extracting at least one key point information of the first task from the description information through a key point extraction model, wherein the key point extraction model is a machine learning model which is pre-trained.
6. The method according to any one of claims 1 to 5, wherein after determining the task prompt information corresponding to the first task according to the sample data and the at least one gist information, the method further includes:
determining a first predefined task matched with the first task from at least one predefined task according to task prompt information corresponding to the first task and task template information corresponding to at least one predefined task respectively; each predefined task corresponds to a task type, and task template information corresponding to the predefined task is used for representing the task type corresponding to the predefined task;
Based on a task database corresponding to the first predefined task, adjusting parameters of the task processing model to obtain an adjusted task processing model; the task database corresponding to the first predefined task comprises model training data related to the first predefined task;
the generating, by the task processing model, a processing result of the first task according to the task prompt information includes:
and generating a processing result of the first task according to the task prompt information through the adjusted task processing model.
7. The method of claim 6, wherein the task database corresponding to the first predefined task further includes model parameters of a model related to the first predefined task;
the step of adjusting the parameters of the task processing model based on the task database corresponding to the first predefined task to obtain an adjusted task processing model includes:
adjusting parameters of the task processing model by using model training data related to the first predefined task, which is included in a task database corresponding to the first predefined task, so as to obtain the adjusted task processing model;
Or alternatively, the first and second heat exchangers may be,
inputting task prompt information corresponding to the first task into the task processing model to obtain a first output result of the task processing model; inputting the task prompt information corresponding to the first task into a model related to the first predefined task to obtain a second output result of the model related to the first predefined task; and adjusting parameters of the task processing model with the aim of minimizing the difference between the first output result and the second output result to obtain the adjusted task processing model.
8. The method according to claim 6, wherein the determining, according to the task prompt information corresponding to the first task and task template information corresponding to at least one predefined task, a first predefined task that matches the first task from the at least one predefined task includes:
for each predefined task in the at least one predefined task, determining the matching degree between the task prompt information corresponding to the first task and the task template information corresponding to the predefined task;
and determining a predefined task corresponding to the maximum value of the matching degree as the first predefined task.
9. The method of claim 8, wherein determining a degree of matching between the task prompt information corresponding to the first task and the task template information corresponding to the predefined task comprises:
calculating semantic similarity between task prompt information corresponding to the first task and task template information corresponding to the predefined task; determining the semantic similarity as the matching degree;
or determining the matching degree through a matching model according to the task prompt information corresponding to the first task and the task template information corresponding to the predefined task; wherein the matching model is a machine learning model.
10. A task processing device, the device comprising:
the data acquisition module is used for acquiring description information and sample data of a first task, wherein the first task is a task to be processed related to natural language processing, the description information is used for indicating the processing requirement of the first task, and the sample data is used for indicating the processing object of the first task;
the information extraction module is used for extracting at least one key point information of the first task from the description information, wherein the key point information is used for indicating the processing key point of the first task;
The information determining module is used for determining task prompt information corresponding to the first task according to the sample data and the at least one key point information, and the task prompt information is used for representing the processing requirement and the processing object of the first task;
and the result generation module is used for generating a processing result of the first task according to the task prompt information through a task processing model.
11. A computer device comprising a processor and a memory, the memory having stored therein a computer program that is loaded and executed by the processor to implement the method of any of claims 1 to 9.
12. A computer readable storage medium, characterized in that the computer readable storage medium has stored therein a computer program, which is loaded and executed by a processor to implement the method of any of claims 1 to 9.
13. A computer program product, characterized in that it comprises a computer program stored in a computer readable storage medium, from which a processor reads and executes the computer program to implement the method according to any one of claims 1 to 9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310754367.1A CN116956934A (en) | 2023-06-25 | 2023-06-25 | Task processing method, device, equipment and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310754367.1A CN116956934A (en) | 2023-06-25 | 2023-06-25 | Task processing method, device, equipment and storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116956934A true CN116956934A (en) | 2023-10-27 |
Family
ID=88455672
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310754367.1A Pending CN116956934A (en) | 2023-06-25 | 2023-06-25 | Task processing method, device, equipment and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116956934A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117762499A (en) * | 2024-02-21 | 2024-03-26 | 腾讯科技（深圳）有限公司 | Task instruction construction method and task processing method |
-
2023
- 2023-06-25 CN CN202310754367.1A patent/CN116956934A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117762499A (en) * | 2024-02-21 | 2024-03-26 | 腾讯科技（深圳）有限公司 | Task instruction construction method and task processing method |
| CN117762499B (en) * | 2024-02-21 | 2024-05-28 | 腾讯科技（深圳）有限公司 | Task instruction construction method and task processing method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112131366B (en) | Method, device and storage medium for training text classification model and text classification | |
| US12001964B2 (en) | Artificial intelligence advisory systems and methods for behavioral pattern matching and language generation | |
| WO2023029506A1 (en) | Illness state analysis method and apparatus, electronic device, and storage medium | |
| CN112131393A (en) | Construction method of medical knowledge map question-answering system based on BERT and similarity algorithm | |
| CN112015917A (en) | Data processing method and device based on knowledge graph and computer equipment | |
| US11532387B2 (en) | Identifying information in plain text narratives EMRs | |
| US20220415203A1 (en) | Interface to natural language generator for generation of knowledge assessment items | |
| CN111651579B (en) | Information query method, device, computer equipment and storage medium | |
| Jing et al. | Knowledge-enhanced attentive learning for answer selection in community question answering systems | |
| CN114139546A (en) | Formulating problems using differences between sets of syntax trees and differences between sets of semantic trees | |
| CN118133971A (en) | Medical question-answering method and device based on large language model | |
| CN116956934A (en) | Task processing method, device, equipment and storage medium | |
| Berge et al. | Combining unsupervised, supervised and rule-based learning: the case of detecting patient allergies in electronic health records | |
| Gupta et al. | Fundamentals Of Chat GPT For Beginners Using AI | |
| EP3901875A1 (en) | Topic modelling of short medical inquiries | |
| CN117473057A (en) | Question-answering processing method, system, equipment and storage medium | |
| CN117149982A (en) | Question-answering processing method, device, equipment and storage medium based on artificial intelligence | |
| CN117131176A (en) | Interactive question-answering processing method and device, electronic equipment and storage medium | |
| CN116796840A (en) | Medical entity information extraction method, device, computer equipment and storage medium | |
| CN116956869A (en) | Text normalization method, device, electronic equipment and storage medium | |
| Huang et al. | Inferring Stressors from Conversation: Towards an Emotional Support Robot Companion | |
| Carl | Translation norms, translation behavior, and continuous vector space models | |
| Kulkarni et al. | Deep Reinforcement-Based Conversational AI Agent in Healthcare System | |
| CN111444338A (en) | Text processing device, storage medium and equipment | |
| Jamal | Automatic Construction of Ontology with Public-Domain Datasets for Personalized Tutoring with ECA |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication |