Question and answer control method and device based on question decomposition

Abstract

The invention discloses a question and answer control method and device based on question decomposition. The method comprises the steps of obtaining an initial question input by a user; analyzing the initial problem to obtain a plurality of key components corresponding to the initial problem; according to each key component, based on a preset knowledge base and a probability model, solving each key component to obtain an answer corresponding to each key component; and based on all the answers, generating a final answer of the initial question through a natural language generation technology. According to the method, the complex questions are analyzed into the single questions which can be processed, and finally the single questions which can be processed are summarized, so that the complexity of the questions is converted into the relevance of answers, the difficulty of question answering is reduced, and the accuracy of solving the complex questions is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of knowledge question answering, and in particular to a question answering control method and device based on question decomposition. Background Art

[0002] With the popularization of human-computer interaction, human-computer language intelligent interaction often occurs when dealing with complex problems. When interacting with complex problems, on the one hand, the user's natural language expression is affected by individual language habits, culture, etc., and complex questions have more complex question types and component semantics, which leads to errors in the understanding of the problem and the analysis of the key components of the problem. On the other hand, methods based on semantic parsing are heavily dependent on the design of logical forms and parsing algorithms, and it is difficult to cover various complex query types (such as multi-hop reasoning, constraint relations, and numerical operations). Complex problems involving more relations and topics will greatly increase the possible search space for parsing, thereby reducing parsing efficiency, and ultimately resulting in less accurate answers given by the question-answering system, thereby reducing the user experience of the question-answering system.

[0003] Therefore, there is an urgent need for a question-answering control strategy based on question decomposition to solve the problem of low accuracy in answering complex questions. Summary of the invention

[0004] The embodiment of the present invention provides a question-answering control method and device based on question decomposition to solve the problem of low accuracy in answering complex questions.

[0005] In order to solve the above problems, an embodiment of the present invention provides a question-answering control method and device based on question decomposition, including:

[0006] Get the initial question from the user;

[0007] Parsing the initial question to obtain several key components corresponding to the initial question;

[0008] According to each key component, each key component is searched based on the preset knowledge base, and based on the search results, each key component is solved through the knowledge graph and probability model to obtain the answer corresponding to each key component;

[0009] According to all the answers, connectives related to each answer are generated through natural language generation technology, and based on the answers and the connectives, a final answer to the initial question is generated.

[0010] As an improvement of the above solution, the initial question is parsed to obtain several key components corresponding to the initial question, including:

[0011] Decomposing the initial problem by using natural language processing technology and problem understanding technology to obtain a number of decomposed problems;

[0012] Perform syntactic analysis on each decomposed question to obtain the key components corresponding to each decomposed question.

[0013] As an improvement of the above scheme, according to each key component, each key component is searched based on a preset knowledge base, and based on the search results, each key component is solved through a knowledge graph and a probability model to obtain the answer corresponding to each key component, including:

[0014] For each key component, a knowledge search is performed on the current key component through a preset knowledge base, and the search results are judged;

[0015] If the search is successful, the knowledge retrieved from the current key component is logically inferred based on the knowledge graph to obtain the answer corresponding to the current key component; wherein the knowledge graph is used to associate knowledge;

[0016] If the search fails, the probability of the relevant knowledge of the current key component is calculated through the preset probability model, and the answer to the current key component is obtained based on the probability calculation result.

[0017] As an improvement of the above solution, the method generates, based on all the answers, a connecting word related to each answer through natural language generation technology, and generates a final answer to the initial question based on the answers and the connecting words, including:

[0018] Through natural language generation technology, the language corresponding to each key component is generated to obtain the connecting words corresponding to each answer;

[0019] By generating a model, the answers corresponding to each key component are screened to obtain the candidate answers for each decomposed question, and the candidate answers are combined into a set of candidate answers for the initial question;

[0020] Calculating the relevance of the answer set to be selected;

[0021] The final answer to the initial question is obtained by combining a set of candidate answers with a correlation greater than a preset value with the connecting words corresponding to the candidate answers in the set of candidate answers.

[0022] As an improvement of the above solution, the knowledge base includes structured data and unstructured data; wherein the structured data is data stored in a database, and the unstructured data is unprocessed text data.

[0023] Accordingly, an embodiment of the present invention further provides a question-answering control device based on question decomposition, comprising: a data acquisition module, a data parsing module, an answer solving module and a result generating module;

[0024] The data acquisition module is used to acquire the initial question input by the user;

[0025] The data analysis module is used to analyze the initial question and obtain several key components corresponding to the initial question;

[0026] The answer solving module is used to search for each key component based on a preset knowledge base, and solve each key component based on the search results through a knowledge graph and a probability model to obtain the answer corresponding to each key component;

[0027] The result generation module is used to generate connecting words related to each answer according to all the answers through natural language generation technology, and generate a final answer to the initial question based on the answers and the connecting words.

[0028] As an improvement of the above solution, the data parsing module includes: a decomposition unit and an analysis unit;

[0029] The decomposition unit is used to decompose the initial problem through natural language processing technology and problem understanding technology to obtain a number of decomposed problems;

[0030] The analysis unit is used to perform syntactic analysis on each decomposed question to obtain key components corresponding to each decomposed question.

[0031] As an improvement of the above solution, the answer-solving module includes: a retrieval unit, a logic reasoning unit and a probability calculation unit;

[0032] The retrieval unit is used to perform knowledge retrieval on the current key component through a preset knowledge base for each key component and judge the retrieval result;

[0033] The logic reasoning unit is used to perform logic reasoning on the knowledge retrieved from the current key component based on the knowledge graph if the retrieval is successful, so as to obtain the answer corresponding to the current key component; wherein the knowledge graph is used to associate knowledge;

[0034] The probability calculation unit is used to perform probability calculation of relevant knowledge on the current key component through a preset probability model if the retrieval fails, and obtain the answer to the current key component based on the probability calculation result.

[0035] As an improvement of the above solution, the result generation module includes: a natural language generation unit, a screening unit, an association unit and an answer generation unit;

[0036] The natural language generation unit is used to generate language for the answer corresponding to each key component through natural language generation technology to obtain the connecting words corresponding to each answer;

[0037] The screening unit is used to screen the answers corresponding to each key component by generating a model, obtain the candidate answers for each decomposed question, and combine them into a candidate answer set for the initial question;

[0038] The association unit is used to calculate the association degree of the answer set to be selected;

[0039] The answer generation unit is used to combine the candidate answer set with a correlation greater than a preset value with the connecting words corresponding to the candidate answers in the candidate answer set to obtain a final answer to the initial question.

[0040] As an improvement of the above solution, the knowledge base includes structured data and unstructured data; wherein the structured data is data stored in a database, and the unstructured data is unprocessed text data.

[0041] As can be seen from the above, the present invention has the following beneficial effects:

[0042] The present invention provides a question-answering control method based on problem decomposition, which obtains an initial question input by a user; parses the initial question to obtain several key components corresponding to the initial question; solves each key component based on a preset knowledge base and probability model to obtain an answer corresponding to each key component; and generates a final answer to the initial question based on all the answers through natural language generation technology. The present invention parses the initial question to decompose the question into multiple processable key components, answers the key components through a knowledge base and a probability model, and generates a final answer to the answer based on natural language generation technology. The present invention parses complex problems into single processable problems, and finally aggregates the single processable problems, thereby realizing the transformation of the complexity of the problem into the relevance of the answer, thereby reducing the difficulty of problem answering and greatly improving the accuracy of solving complex problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flowchart of a question-answering control method based on question decomposition provided by an embodiment of the present invention;

[0044] Figure 2 is a structural diagram of a question-answering control device based on question decomposition provided by an embodiment of the present invention;

[0045] Figure 3 It is a schematic diagram of the structure of a terminal device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] Embodiment 1

[0048] See also Figure 1 , Figure 1 is a flow chart of a question-answering control method based on question decomposition provided by an embodiment of the present invention, such as Figure 1 As shown, this embodiment includes steps 101 to 104, and each step is specifically as follows:

[0049] Step 101: Obtain an initial question input by a user.

[0050] Step 102: Analyze the initial question to obtain several key components corresponding to the initial question.

[0051] In this embodiment, the initial question is parsed to obtain several key components corresponding to the initial question, including:

[0052] Decomposing the initial problem by using natural language processing technology and problem understanding technology to obtain a number of decomposed problems;

[0053] Perform syntactic analysis on each decomposed question to obtain the key components corresponding to each decomposed question.

[0054] In a specific embodiment, syntactic analysis or semantic role labeling techniques are used to identify key components in the question, including subject, predicate and object.

[0055] In a specific embodiment, through natural language processing, question classification and question understanding, the questions raised by users are parsed, their intentions and semantics are understood, the field or type of the question is determined, complex questions are decomposed into a series of simpler questions, and the key components of the question are found.

[0056] In a specific embodiment, natural language processing includes text processing, grammatical and semantic analysis, word vector representation, sequence modeling, language model, machine translation, sentiment analysis, text production, question-answering system and speech recognition and synthesis. Text processing is to segment sentences into words or symbols, restore vocabulary to its basic form, remove common words in the text that are not very meaningful for analysis, and identify the part of speech of each word in the text. Grammar and semantic analysis is to analyze the dependency between words in a sentence and identify the relationship between a predicate and its arguments in a sentence. Word vector representation is to represent text as a collection of words, ignoring the order, and measuring the importance of words in a document. Map vocabulary to high-dimensional vector space to capture semantic relationships between vocabulary. Sequence modeling is used for sequence labeling tasks to process sequence data and capture time dependencies. Improve RNN to solve long-term dependency problems. Language models predict the probability of the next word based on the previous n-1 words and use neural networks to predict the probability of the next word. Machine translation is based on statistical models and uses neural networks for translation. Sentiment analysis uses predefined rules and vocabularies to determine the emotional tendency of text. Text production generates text using predefined templates. Question answering systems retrieve the most relevant answers from a pre-prepared knowledge base and use deep learning models to generate answers.

[0057] Step 103: According to each key component, each key component is searched based on a preset knowledge base, and based on the search results, each key component is solved through a knowledge graph and a probability model to obtain the answer corresponding to each key component.

[0058] In this embodiment, according to each key component, each key component is searched based on a preset knowledge base, and based on the search results, each key component is solved through a knowledge graph and a probability model to obtain an answer corresponding to each key component, including:

[0059] For each key component, a knowledge search is performed on the current key component through a preset knowledge base, and the search results are judged;

[0060] If the search is successful, the knowledge retrieved from the current key component is logically inferred based on the knowledge graph to obtain the answer corresponding to the current key component; wherein the knowledge graph is used to associate knowledge;

[0061] If the search fails, the probability of the relevant knowledge of the current key component is calculated through the preset probability model, and the answer to the current key component is obtained based on the probability calculation result.

[0062] In this embodiment, the knowledge base includes structured data and unstructured data; wherein the structured data is data stored in a database, and the unstructured data is unprocessed text data.

[0063] In a specific embodiment, the probability calculation of the relevant knowledge of the current key component is performed through a preset probability model, and based on the probability calculation result, the answer of the current key component is obtained, specifically:

[0064] 1. Key components and knowledge retrieval:

[0065] Suppose we have a key component "X", which represents a specific scientific concept or phenomenon. In order to solve this key component, we first perform knowledge retrieval through a preset knowledge base. The knowledge base can be a large database that contains various scientific concepts, definitions, formulas, experimental data, etc.

[0066] Knowledge retrieval process:

[0067] Enter key ingredient "X" into the knowledge base.

[0068] The knowledge base returns all entries or records related to "X".

[0069] Search result judgment:

[0070] If the search results contain clear and direct information, such as definitions, formulas, or experimental data, the search is considered successful.

[0071] If there is no clear information in the search results, or the information is too vague or uncertain, the search is judged to have failed.

[0072] 2. Logical reasoning based on knowledge graph:

[0073] If the retrieval is successful, we will use the knowledge graph for logical reasoning. Knowledge graph is a technology that represents entities, relationships, and attributes in the real world in the form of a graph. It can help us understand and analyze complex concepts and relationships.

[0074] Knowledge graph applications:

[0075] In the knowledge graph, find entities and relationships that are directly related to the key ingredient "X".

[0076] Based on these entities and relationships, conduct logical reasoning and draw conclusions or answers about "X".

[0077] Logical reasoning example:

[0078] Assume that the key ingredient "X" is "the law of refraction of light".

[0079] In the knowledge graph, find entities related to the "law of refraction of light", such as "incident light", "refracted light", "normal", and "refractive index".

[0080] Make logical inferences based on the relationships between these entities, such as "the angle between the incident light ray and the normal is equal to the angle between the refracted light ray and the normal multiplied by the refractive index."

[0081] Finally, we come to the conclusion or answer about the "law of refraction of light".

[0082] 3. Probability calculation based on probability model:

[0083] If the search fails, we will use the preset probability model to calculate the probability of relevant knowledge. The probability model can be a model based on statistical methods such as Bayesian theorem and Markov chain.

[0084] Probabilistic model selection:

[0085] According to the nature of the key component "X" and the contextual information, a suitable probability model is selected.

[0086] For example, if “X” is a key component associated with a random event, we can choose Bayes’ theorem to calculate the probability of the associated event.

[0087] Probability calculation process:

[0088] Define random events and variables related to the key ingredient "X".

[0089] Calculate the probabilities of these random events based on the formulas and assumptions of the probability model.

[0090] To better illustrate, here is an example:

[0091] 1. Choice of probability model:

[0092] In the probability calculation unit, you can choose a variety of probability models for calculation. The specific choice depends on the nature of the problem and the type of probability to be solved. The following are some common probability models:

[0093] Numerical probability model: often used to solve numerical calculation problems. This model can only obtain an approximate solution to the problem, and the accuracy of the solution generally increases with the increase of calculation time. For example, when solving numerical problems such as definite integrals, the value of the definite integral can be estimated by the frequency of occurrence of random events through random simulation methods.

[0094] Monte Carlo model: When the problem to be solved is the probability of a random event, or the expected value of a random variable, the Monte Carlo model can be used. This model estimates the probability of a random event based on its frequency of occurrence through a large number of simulation experiments, or obtains certain numerical characteristics of the random variable.

[0095] 2. Probability calculation based on key components:

[0096] For example, if there is no directly related information about the key component "X" in the knowledge base, we need to use the probability calculation unit to solve the probability of the knowledge related to "X".

[0097] Determine the inputs to the probabilistic model:

[0098] In this case, we can choose the Monte Carlo model as the probability model.

[0099] The inputs are the possible events or states associated with "X" and their distribution in the population.

[0100] Perform probability calculations:

[0101] Based on the Monte Carlo model, we can design a series of simulation experiments to estimate the probability of knowledge related to "X".

[0102] For example, we could simulate many random events related to "X" and record the number of times these events occur.

[0103] By calculating the ratio of the number of times these events occur to the total number of simulations, we can get an approximate probability of knowledge related to "X".

[0104] Interpret the probability results:

[0105] The probability results obtained can be used to explain the likelihood of knowledge related to "X".

[0106] For example, if the calculated probability of an event associated with "X" is 0.6, then we can assume that the event is likely to occur in most cases.

[0107] 3. The answer corresponding to each probability:

[0108] In the probability calculation unit, each probability value corresponds to one or more possible answers. These answers are calculated based on the probability model and reflect the likelihood of knowledge related to the key component "X".

[0109] High probability answer:

[0110] If the probability of an answer being related to the knowledge of the key ingredient "X" is high (e.g., greater than 0.5), then we can consider the answer to be more reliable.

[0111] In practical applications, we can give priority to these high-probability answers.

[0112] Low probability answer:

[0113] If a certain answer has a low probability of knowledge related to the key ingredient “X” (e.g., less than 0.5), then we can consider this answer to be unlikely.

[0114] However, this does not mean that these low probability answers have no value. In some cases, these answers may provide additional information or clues related to the key ingredient "X".

[0115] In a specific embodiment, structured and unstructured data, where structured data is information in a database and unstructured data is text, articles, etc., use knowledge graphs to organize and associate different types of information so that the system can perform complex reasoning. The system also needs to access external APIs or data sources to obtain the latest or field-specific information.

[0116] In a specific embodiment, logical reasoning includes deductive reasoning, inductive reasoning, abductive reasoning and default reasoning. Deductive reasoning is to derive a necessary conclusion from a set of premises. If the premise is true, then the conclusion must be true. Commonly used deductive reasoning methods include proofs in propositional logic and predicate logic. Inductive reasoning is to derive general conclusions from specific examples. Inductive reasoning is usually not necessary, but based on observation and experience. Abductive reasoning is to derive the most likely cause or explanation from the observed phenomenon. Abductive reasoning is often used for diagnosis and troubleshooting. Default reasoning is to deal with uncertainty and default assumptions. Default reasoning allows the system to make reasonable assumptions in the absence of sufficient information. Probabilistic reasoning includes precise reasoning and approximate reasoning. Reasoning is performed by calculating precise probability values. When the computational complexity of precise reasoning is too high, approximate methods are used to estimate probabilities.

[0117] Step 104: Based on all the answers, generate the connecting words related to each answer through natural language generation technology, and generate the final answer to the initial question based on the answers and the connecting words.

[0118] In this embodiment, according to all the answers, a connecting word related to each answer is generated by natural language generation technology, and based on the answers and the connecting words, a final answer to the initial question is generated, including:

[0119] Through natural language generation technology, the language corresponding to each key component is generated to obtain the connecting words corresponding to each answer;

[0120] By generating a model, the answers corresponding to each key component are screened to obtain the candidate answers for each decomposed question, and the candidate answers are combined into a set of candidate answers for the initial question;

[0121] Calculating the relevance of the answer set to be selected;

[0122] The final answer to the initial question is obtained by combining a set of candidate answers with a correlation greater than a preset value with the connecting words corresponding to the candidate answers in the set of candidate answers.

[0123] In a specific embodiment, a preliminary answer is generated for each sub-question, the answer is directly extracted from the retrieved information, and a generative model is used to generate the answer. The retrieval and generation methods are combined to improve the accuracy and coherence of the answer. Answer integration is to sort and screen multiple candidate answers, select the most relevant and accurate answer (that is, the correlation degree described in the present invention is greater than a preset value), ensure that there is no contradiction between the answers to the various sub-questions, maintain consistency, remove repeated or redundant information, and ensure that the answer is concise and clear.

[0124] In a specific embodiment, natural language generation includes template filling, rule-based methods, statistical methods and deep learning-based methods. Template filling is to use predefined templates to fill in answers. The rule-based method is to use linguistic rules and templates to generate natural language text. The statistical-based method is to use statistical models to generate natural language text. The deep learning-based method is to use deep learning models, including RNN, LSTM, Transformer, to generate coherent natural language text.

[0125] It should be noted that the present invention also includes: user interaction, including dialogue management and feedback mechanism, interacting with users based on the answer produced in step S4, supporting multiple rounds of dialogue, allowing users to further clarify questions or provide more information, adjusting answers based on user feedback, and improving accuracy and satisfaction.

[0126] Among them, dialogue management includes rule-based dialogue management, statistics-based dialogue management, machine learning-based dialogue management, and end-to-end dialogue management, which use predefined rules and conditions to manage dialogues. This method is more common in simple dialogue systems, but it is difficult to handle complex dialogue scenarios. Using statistical models to learn the best dialogue strategy can handle more complex dialogue scenarios, but requires a large amount of labeled data. Using machine learning models to learn dialogue strategies can handle more complex dialogue scenarios and can self-optimize with more data. It learns dialogue strategies directly from raw dialogue data without explicit state tracking and strategy selection. This method simplifies the dialogue management process, but its performance in complex dialogue scenarios may be limited.

[0127] In addition, the present invention continuously learns and optimizes, including continuous learning, performance evaluation, and update maintenance. Based on each interaction with the user in step S5, the system learns from it, gradually improves the problem-solving ability, regularly evaluates system performance, including accuracy and response speed, regularly updates the knowledge base and model, and expands new information. Update maintenance mainly involves regularly updating system documents, including configuration manuals, operation guides, and troubleshooting manuals, maintaining a knowledge base, recording solutions and best practices for common problems for reference by team members, and recording all system changes, including updates, upgrades, and configuration changes, for tracking and rollback.

[0128] For better explanation, the following application scenarios are provided for illustration: intelligent medical consultation system;

[0129] S1. The user inputs a question through voice or text, such as: "I always feel tired recently and occasionally dizzy. What could be the reason?" The system first uses natural language processing technology to parse the user's input and understand that the intention is to ask about the possible causes of health problems. Then, the system classifies the question into the "health consultation" field and breaks it down into several sub-questions, such as: "What are the common causes of fatigue?", "What are the possible causes of dizziness?", etc. At the same time, the system identifies the key components in the question, including symptoms such as "fatigue" and "dizziness";

[0130] S2. The system retrieves medical knowledge related to fatigue and dizziness, such as common causes, diagnostic criteria, and treatment methods, from structured data (such as medical databases). At the same time, the system also uses unstructured data (such as medical literature, expert opinions, etc.) to enrich the knowledge base. In addition, the system also visualizes the relationship between different symptoms, diseases, and treatment plans through knowledge graphs to understand the problem more intuitively;

[0131] S3. Based on existing medical knowledge and rules, the system performs logical reasoning to analyze the possible causes of fatigue and dizziness. For uncertain information, the system uses probability models to evaluate the possibility of different causes and gives corresponding probability values. In addition, for specific types of problems (such as disease diagnosis), the system can also use deep learning models to make predictions and improve the accuracy of diagnosis;

[0132] S4. The system integrates the answers to each sub-question to form a complete answer to the original question. For example, the system may answer: "Based on the symptoms you provided, fatigue and dizziness may be caused by anemia, low blood pressure, or lack of sleep. It is recommended that you undergo further examination to confirm the diagnosis." At the same time, the system uses natural language generation technology to convert the results of machine processing into a language form that is easy for humans to understand, ensuring that the generated answer is consistent with the context of the question;

[0133] S5. The system supports multiple rounds of dialogue, allowing users to further clarify questions or provide more information. For example, a user may ask: "Which department of the hospital should I go to for examination?" The system adjusts the answer based on the user's feedback and gives corresponding suggestions. In addition, the system also collects user satisfaction and opinions through a feedback mechanism in order to continuously optimize services;

[0134] S6. The system learns new medical knowledge and user habits from each interaction, gradually improving its problem-solving capabilities. For example, the system may find that a certain disease is more common in a certain population, thereby adjusting the reasoning algorithm to improve the accuracy of the diagnosis. At the same time, the system also regularly evaluates performance (such as accuracy and response speed) and updates the knowledge base and model based on the evaluation results to expand emerging information and respond to new challenges.

[0135] See also Figure 2 , Figure 2 2 is a schematic diagram of the structure of a question-answering control device based on question decomposition provided by an embodiment of the present invention, comprising: a data acquisition module 201, a data analysis module 202, an answer solving module 203 and a result generating module 204;

[0136] The data acquisition module is used to acquire the initial question input by the user;

[0137] The data analysis module is used to analyze the initial question and obtain several key components corresponding to the initial question;

[0138] The answer solving module is used to search for each key component based on a preset knowledge base, and solve each key component based on the search results through a knowledge graph and a probability model to obtain the answer corresponding to each key component;

[0139] The result generation module is used to generate connecting words related to each answer according to all the answers through natural language generation technology, and generate a final answer to the initial question based on the answers and the connecting words.

[0140] As an improvement of the above solution, the data parsing module includes: a decomposition unit and an analysis unit;

[0141] The decomposition unit is used to decompose the initial problem through natural language processing technology and problem understanding technology to obtain a number of decomposed problems;

[0142] The analysis unit is used to perform syntactic analysis on each decomposed question to obtain key components corresponding to each decomposed question.

[0143] As an improvement of the above solution, the answer-solving module includes: a retrieval unit, a logic reasoning unit and a probability calculation unit;

[0144] The retrieval unit is used to perform knowledge retrieval on the current key component through a preset knowledge base for each key component and judge the retrieval result;

[0145] The logic reasoning unit is used to perform logic reasoning on the knowledge retrieved from the current key component based on the knowledge graph if the retrieval is successful, so as to obtain the answer corresponding to the current key component; wherein the knowledge graph is used to associate knowledge;

[0146] The probability calculation unit is used to perform probability calculation of relevant knowledge on the current key component through a preset probability model if the retrieval fails, and obtain the answer to the current key component based on the probability calculation result.

[0147] As an improvement of the above solution, the result generation module includes: a natural language generation unit, a screening unit, an association unit and an answer generation unit;

[0148] The natural language generation unit is used to generate language for the answer corresponding to each key component through natural language generation technology to obtain the connecting words corresponding to each answer;

[0149] The screening unit is used to screen the answers corresponding to each key component by generating a model, obtain the candidate answers for each decomposed question, and combine them into a candidate answer set for the initial question;

[0150] The association unit is used to calculate the association degree of the answer set to be selected;

[0151] The answer generation unit is used to combine the candidate answer set with a correlation greater than a preset value with the connecting words corresponding to the candidate answers in the candidate answer set to obtain a final answer to the initial question.

[0152] As an improvement of the above solution, the knowledge base includes structured data and unstructured data; wherein the structured data is data stored in a database, and the unstructured data is unprocessed text data.

[0153] This embodiment obtains an initial question input by a user; parses the initial question to obtain several key components corresponding to the initial question; solves each key component based on a preset knowledge base and probability model to obtain an answer corresponding to each key component; and generates a final answer to the initial question based on all the answers through natural language generation technology. The present invention parses the initial question to break the problem down into multiple processable key components, solves the key components through a knowledge base and a probability model, and generates a final answer to the solved answers based on natural language generation technology. The present invention parses complex problems into single processable problems, and finally aggregates the single processable problems, thereby converting the complexity of the problem into the relevance of the answer, thereby reducing the difficulty of problem solving and greatly improving the accuracy of solving complex problems.

[0154] Embodiment 2

[0155] See also Figure 3 , Figure 3 It is a schematic diagram of the structure of a terminal device provided in one embodiment of the present invention.

[0156] A terminal device of this embodiment includes: a processor 301, a memory 302, and a computer program stored in the memory 302 and executable on the processor 301. When the processor 301 executes the computer program, the steps of the above-mentioned question-answering control method based on question decomposition in the embodiment are implemented, for example: Figure 1 Alternatively, when the processor executes the computer program, the functions of each module in the above-mentioned device embodiments are implemented, for example: Figure 2 All modules of the question-answering control device based on question decomposition are shown.

[0157] In addition, an embodiment of the present invention further provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the question-answering control method based on question decomposition as described in any of the above embodiments.

[0158] Those skilled in the art will understand that the schematic diagram is merely an example of a terminal device and does not constitute a limitation on the terminal device. The terminal device may include more or fewer components than shown in the diagram, or a combination of certain components, or different components. For example, the terminal device may also include input and output devices, network access devices, buses, etc.

[0159] The processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor 301 is the control center of the terminal device, and uses various interfaces and lines to connect various parts of the entire terminal device.

[0160] The memory 302 can be used to store the computer program and / or module. The processor 301 implements various functions of the terminal device by running or executing the computer program and / or module stored in the memory and calling the data stored in the memory 302. The memory 302 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area can store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0161] Wherein, if the module / unit integrated in the terminal device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium, etc.

[0162] It should be noted that the device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art may understand and implement it without paying any creative effort.

[0163] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A question-answering control method based on question decomposition, characterized in that: include: Get the initial question from the user; Parsing the initial question to obtain several key components corresponding to the initial question; According to each key component, each key component is searched based on the preset knowledge base, and based on the search results, each key component is solved through the knowledge graph and probability model to obtain the answer corresponding to each key component; According to all the answers, connectives related to each answer are generated through natural language generation technology, and based on the answers and the connectives, a final answer to the initial question is generated.

2. The question-answering control method based on question decomposition according to claim 1, characterized in that: The initial question is parsed to obtain several key components corresponding to the initial question, including: Decomposing the initial problem by using natural language processing technology and problem understanding technology to obtain a number of decomposed problems; Perform syntactic analysis on each decomposed question to obtain the key components corresponding to each decomposed question.

3. The question-answering control method based on question decomposition according to claim 2 is characterized in that: According to each key component, each key component is searched based on a preset knowledge base, and based on the search results, each key component is solved through a knowledge graph and a probability model to obtain an answer corresponding to each key component, including: For each key component, a knowledge search is performed on the current key component through a preset knowledge base, and the search results are judged; If the search is successful, the knowledge retrieved from the current key component is logically inferred based on the knowledge graph to obtain the answer corresponding to the current key component; wherein the knowledge graph is used to associate knowledge; If the search fails, the probability of the relevant knowledge of the current key component is calculated through the preset probability model, and the answer to the current key component is obtained based on the probability calculation result.

4. The question-answering control method based on question decomposition according to claim 3 is characterized in that: The method of generating a linking word related to each answer based on all the answers by using natural language generation technology, and generating a final answer to the initial question based on the answers and the linking words, includes: Through natural language generation technology, the language corresponding to each key component is generated to obtain the connecting words corresponding to each answer; By generating a model, the answers corresponding to each key component are screened to obtain the candidate answers for each decomposed question, and the candidate answers are combined into a set of candidate answers for the initial question; Calculating the relevance of the answer set to be selected; The final answer to the initial question is obtained by combining a set of candidate answers with a correlation greater than a preset value with the connecting words corresponding to the candidate answers in the set of candidate answers.

5. The question-answering control method based on question decomposition according to claim 4 is characterized in that: The knowledge base includes structured data and unstructured data; wherein the structured data is data stored in a database, and the unstructured data is unprocessed text data.

6. A question-answering control device based on question decomposition, characterized in that: include: Data acquisition module, data parsing module, answer solving module and result generation module; The data acquisition module is used to acquire the initial question input by the user; The data analysis module is used to analyze the initial question and obtain several key components corresponding to the initial question; The answer solving module is used to search for each key component based on a preset knowledge base, and solve each key component based on the search results through a knowledge graph and a probability model to obtain the answer corresponding to each key component; The result generation module is used to generate connecting words related to each answer according to all the answers through natural language generation technology, and generate a final answer to the initial question based on the answers and the connecting words.

7. The question-answering control device based on question decomposition according to claim 6, characterized in that: The data parsing module includes: a decomposition unit and an analysis unit; The decomposition unit is used to decompose the initial problem through natural language processing technology and problem understanding technology to obtain a number of decomposed problems; The analysis unit is used to perform syntactic analysis on each decomposed question to obtain key components corresponding to each decomposed question.

8. The question-answering control device based on question decomposition according to claim 7, characterized in that: The answer-solving module includes: a retrieval unit, a logic reasoning unit and a probability calculation unit; The retrieval unit is used to perform knowledge retrieval on the current key component through a preset knowledge base for each key component and judge the retrieval result; The logic reasoning unit is used to perform logic reasoning on the knowledge retrieved from the current key component based on the knowledge graph if the retrieval is successful, so as to obtain the answer corresponding to the current key component; wherein the knowledge graph is used to associate knowledge; The probability calculation unit is used to perform probability calculation of relevant knowledge on the current key component through a preset probability model if the retrieval fails, and obtain the answer to the current key component based on the probability calculation result.

9. The question-answering control device based on question decomposition according to claim 8, characterized in that: The result generation module includes: a natural language generation unit, a screening unit, an association unit and an answer generation unit; The natural language generation unit is used to generate language for the answer corresponding to each key component through natural language generation technology to obtain the connecting words corresponding to each answer; The screening unit is used to screen the answers corresponding to each key component by generating a model, obtain the candidate answers for each decomposed question, and combine them into a candidate answer set for the initial question; The association unit is used to calculate the association degree of the answer set to be selected; The answer generation unit is used to combine the candidate answer set with a correlation greater than a preset value with the connecting words corresponding to the candidate answers in the candidate answer set to obtain a final answer to the initial question.

10. The question-answering control device based on question decomposition according to claim 9, characterized in that: The knowledge base includes structured data and unstructured data; wherein the structured data is data stored in a database, and the unstructured data is unprocessed text data.

Images