Open question answering method and device for peach production knowledge based on migration context network

Abstract

The present invention provides a method and device for open question-answering of knowledge about peach production using a migration context network, the method comprising: inputting the acquired question text into a trained question intention classification network to obtain question semantic information with classification labels; inputting the question semantic information into a trained question similarity matching network to obtain at least one matching question corresponding to the question text; determining a matching answer corresponding to the question text from a pre-constructed question and answer database according to the matching question, and feeding back the matching answer to the user. The method and device for open question-answering of knowledge about peach production using a migration context network provided by the present invention effectively constructs a production knowledge base through knowledge migration, and improves the accuracy and robustness of production knowledge question-answering in a three-dimensional manner through data mining and clarification, question intention classification, question semantic understanding, and answer matching, and can satisfy the user's acquisition of production knowledge to the greatest extent.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and device for open question-answering of peach production knowledge in a migration context network. Background Art

[0002] With the continuous development of information technology computing power, the field of artificial intelligence has developed rapidly, especially intelligent question answering in natural language processing, which plays an important role in medical, legal, and Internet-related customer service. For the industrial development process of any agricultural product, the experience and technology throughout production, processing, and sales determine the actual harvest of farmers.

[0003] At present, agricultural technical experts provide production knowledge services through on-site services, technical training, online video courseware production and remote services, which can solve the production problems of some farmers. However, the number of experts and their energy are limited, and they cannot provide guidance services quickly and accurately for the massive demand during critical farming seasons. Therefore, using artificial intelligence services and platform applications to let machines learn knowledge like agricultural experts and distribute it to users through cloud services to solve the problems of untimely expert responses and limited knowledge service efficiency has become the mainstream choice.

[0004] However, in terms of data preprocessing, the processing model with high accuracy is highly complex and usually requires a high-performance cloud server (GPU Cloud Computing, GPU). The high cost requirement makes it difficult for applications to adapt to the needs of the agricultural field. In terms of the volume of the knowledge base, a large amount of data collection, screening, cleaning, and storage are required manually, which is time-consuming and labor-intensive, and requires special personnel to perform. Moreover, the question-and-answer system in the vertical field is usually difficult to meet the farmers' question-and-answer habits. It is too rigid in application and often gives "irrelevant answers", which reduces the user's interest in the question-and-answer system. In addition, the service model of most question-and-answer systems is relatively simple, and the cross-platform service is not ideal. Summary of the invention

[0005] The present invention provides a method and device for open question-answering of peach production knowledge based on a migration context network, which are used to solve or partially solve the defects of the prior art.

[0006] In a first aspect, the present invention provides an open question-answering method for peach production knowledge using a migration context network, comprising: inputting an acquired question text into a trained question intention classification network to acquire question semantic information with a classification label; inputting the question semantic information into a trained question similarity matching network to acquire at least one matching question corresponding to the question text; determining a matching answer corresponding to the question text from a pre-constructed question-answering database based on the matching question, and feeding back the matching answer to a user.

[0007] According to a method for answering open questions about peach production knowledge based on a migration context network provided by the present invention, before inputting the acquired question text into a trained question intention classification network, the method comprises:

[0008] Multi-source data collection is performed to construct an agricultural knowledge base; the corpus of the agricultural knowledge base is expanded using a pre-trained generative adversarial network to obtain an enhanced agricultural knowledge base; directional mining and knowledge clustering are performed on the enhanced agricultural knowledge base to obtain a target industry optimized dictionary; a pre-constructed question intention classification network is pre-trained using the target industry optimized dictionary to obtain the trained question intention classification network.

[0009] According to a method for open question answering of peach production knowledge based on a migration context network provided by the present invention, the enhanced agricultural knowledge base is subjected to directional mining and knowledge clustering to obtain an optimized dictionary for the target industry, including:

[0010] The enhanced agricultural knowledge base is cleaned to obtain a target industry knowledge base; seed words in the target industry knowledge base are extracted based on left-right entropy and mutual information; the pre-built deep learning model is trained with the seed words as annotation samples, and the trained deep learning model is used to iteratively mine the target industry knowledge base to construct a target industry dictionary; all the corpora in the target industry dictionary are aggregated using a human-machine assisted clustering annotation method to obtain the target industry optimized dictionary.

[0011] According to a method for answering open questions about peach production knowledge using a transfer context network provided by the present invention, the question intention classification network is constructed based on a convolutional neural network model; and the pre-constructed question intention classification network is pre-trained using the target industry optimization dictionary, including:

[0012] A plurality of corpus samples are randomly selected from the target industry optimized dictionary, and corresponding classification labels are annotated for each of the corpus samples; a combination of each of the corpus samples and the corresponding classification labels is used as a training sample to obtain a plurality of training samples; and a pre-constructed question intention classification network is pre-trained using the plurality of training samples.

[0013] According to a method for answering open questions about peach production knowledge based on a migration context network provided by the present invention, the step of inputting the question semantic information into a trained question similarity matching network to obtain at least one matching question corresponding to the question text comprises:

[0014] A knowledge-enhanced semantic representation network ERNIE model is selected as a teacher model, and a text convolutional network TextCNN model is selected as a student model; the ERNIE model and the TextCNN model are connected through a knowledge distillation layer to build a teacher-student network as the question similarity matching network; the question semantic information is input into the teacher-student network to obtain a matching question output by the teacher-student network.

[0015] According to the open question answering method of peach production knowledge based on the migration context network provided by the present invention, before inputting the question semantic information into the teacher-student network, the method further includes:

[0016] The ERNIE model is pre-trained using a public question-answering dataset from the entire Internet until a first preset number of training times is reached; and the pre-trained ERNIE model is fine-tuned using the target industry optimization dictionary until the training results converge or a second preset number of training times is reached.

[0017] According to the open question-answering method of peach production knowledge based on the migration context network provided by the present invention, after feeding back the matching answer to the user, the method further includes:

[0018] Receive feedback information from the user on the matching answer; if the feedback information meets a preset condition, use the question text as a sample and the matching answer as a label corresponding to the question text to construct a feedback sample; use the sentence intention classification network and the question similarity matching network as an overall network, and use the feedback sample to train the overall network.

[0019] In a second aspect, the present invention also provides an open question-answering device for peach production knowledge based on a migration context network, comprising: a semantic classification unit, used to input the acquired question text into a trained question intention classification network to obtain question semantic information with a classification label; a question matching unit, used to input the question semantic information into a trained question similarity matching network to obtain at least one matching question corresponding to the question text; and an answer matching unit, used to determine a matching answer corresponding to the question text from a pre-constructed question and answer database based on the matching question, and feed back the matching answer to a user.

[0020] In the third aspect, the present invention also provides an open question-and-answer system for peach production knowledge in a migration context network, comprising: a speech recognition device, a search server, a central control business scheduling server, a Redis server, an API interface calling server, and the above-mentioned open question-and-answer device for peach production knowledge in the migration context network.

[0021] The speech recognition device implements speech recognition services and is used to receive the question voice input by the user to generate the question text; the search server is implemented based on Lucene and adopts a RESTful web interface to provide a distributed multi-user full-text search engine; the central control business scheduling server builds a connection between the Redis server and the API interface calling server to implement the scheduling of different types of business data; the Redis server is used to provide high-availability distributed services through a master-slave data replication method, and adopts two persistence solutions, RDB and AOF, to save memory data to a hard disk; the memory data includes a question and answer database and a target industry optimization dictionary; the API interface calling server is used to provide a connection between the speech recognition device and the big peach production knowledge open question and answer device, and to provide an Internet data interface calling service for the big peach production knowledge open question and answer system.

[0022] In a fourth aspect, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the open question-answering method for peach production knowledge based on the migration context network as described in any one of the above are implemented.

[0023] In a fifth aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the open question-answering method for peach production knowledge based on the migration context network as described in any one of the above.

[0024] The open question-answering method and device for peach production knowledge based on the migration context network provided by the present invention effectively constructs a production knowledge base through knowledge migration, and three-dimensionally improves the accuracy and robustness of production knowledge question-answering through data mining and clarification, question intent classification, question semantic understanding, and answer matching, thereby satisfying users' acquisition of production knowledge to the greatest extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0026] Figure 1 It is one of the flow charts of the open question-answering method for peach production knowledge provided by the present invention;

[0027] Figure 2 This is the second flow chart of the open question-answering method for peach production knowledge provided by the present invention;

[0028] Figure 3 It is a schematic diagram of the operation of the question intention classification network provided by the present invention;

[0029] Figure 4 It is a flow chart of implementing corpus control based on a generative adversarial network provided by the present invention;

[0030] Figure 5 It is a schematic diagram of a knowledge distillation process provided by the present invention;

[0031] Figure 6 It is a structural schematic diagram of the open question-answering device for peach production knowledge provided by the present invention;

[0032] Figure 7 It is a structural schematic diagram of the open question-answering system for peach production knowledge provided by the present invention;

[0033] Figure 8 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are 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.

[0035] It should be noted that in the description of the embodiments of the present invention, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "include one..." do not exclude the existence of other identical elements in the process, method, article or device including the elements. The orientation or position relationship indicated by the terms "upper", "lower" and the like is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation of the present invention. Unless otherwise clearly specified and limited, the terms "installed", "connected" and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected, or indirectly connected through an intermediate medium, or it can be the internal communication of two elements. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0036] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are generally of one type, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" in the specification and claims represents at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0037] The core elements of open Q&A on production knowledge, such as open Q&A on peach production knowledge, include the following aspects: the accuracy and sufficiency of relevant corpus, the accuracy and real-time response of the Q&A system, the cross-platform consistency of professional knowledge, etc. The purpose of this invention is mainly to reduce unnecessary manual participation in the process of Q&A system model construction and service, reduce the carrier cost in the model training process, and then meet the knowledge acquisition needs of most scenarios in the agricultural production process through multi-type platform services.

[0038] Combine the following Figure 1-Figure 8 The invention describes a method and device for open question-answering of knowledge about peach production based on a migration context network provided by an embodiment of the invention.

[0039] Figure 1This is one of the flow charts of the open question-answering method for peach production knowledge provided by the present invention, such as Figure 1 As shown, including but not limited to the following steps:

[0040] Step 101: input the obtained question text into the trained question intention classification network to obtain the question semantic information with classification labels;

[0041] Step 102: inputting the question semantic information into a trained question similarity matching network to obtain at least one matching question corresponding to the question text;

[0042] Step 103: Determine a matching answer corresponding to the question text from a pre-built question-answer database according to the matching question, and feed back the matching answer to the user.

[0043] In order to clearly demonstrate the production knowledge open question and answer method provided by the present invention, the subsequent embodiments are all explained using agricultural production scenarios, especially open questions and answers on peach production knowledge related to peach planting. This is not to be regarded as a specific limitation on the scope of protection of the present invention. For example, the field of this knowledge open question and answer may also involve mechanical processing, traffic operation, and other human-computer intelligent interaction scenarios related to artificial intelligence. In the agricultural production scenario, it may involve open questions and answers on production knowledge related to all agricultural products such as peach planting, citrus planting, and poultry (livestock) breeding.

[0044] The open question-and-answer method for peach production knowledge in a migration context network provided by the present invention can effectively address the problems in agricultural production scenarios, such as low intelligent utilization rate of multi-source knowledge data, slow data value empowerment, serious data distribution islands, and uneven service quality, which result in the forms of production knowledge services being messy, and the timeliness and accuracy of services being difficult to guarantee. This in turn results in low usage stickiness among users who have little contact with the Internet, especially agricultural intelligent systems, and even erroneous feedback that has a negative impact on related work in the production process.

[0045] Figure 2 This is the second flow chart of the open question-answering method for peach production knowledge provided by the present invention, such as Figure 2 As shown, after receiving the voice input of the user's question about production knowledge, the question voice can be converted into question text with the help of open source speech-to-text software.

[0046] In view of the complexity and diversity of questions involved in agricultural production knowledge question-answering scenarios, such as peach varieties, pests and diseases, growth environment, sales prices and other related scenarios, the present invention uses the question intention classification network to solve the complexity and diversity of user questions in the question-answering process by inputting the question text into a pre-trained question intention classification network. For classification at different levels, the idea of ​​local perception is used to effectively extract key features in the question text to reduce the complexity of question recognition and effectively improve the accuracy of user dialogue intention classification.

[0047] After the question text is classified by the question intention classification network, the question semantic information with classification labels is output.

[0048] Assume that the question text corresponding to the question voice input by the user is: "Hello, may I ask, how many months ago did the peach trees I planted develop gummosis? How do I usually prevent and control it?", and input this question text into the pre-trained question intention classification network, then it can be determined that the question semantic information of its classification label includes: "How many months ago did gummosis of peach trees usually appear, and how do I prevent and control it?", and its corresponding label can be at least one of "peach tree", "gummosis", and "prevention and control". The extracted question semantic information can more clearly know the true intention of the user's question.

[0049] Furthermore, after the question semantic information is obtained, it is input into a pre-trained question similarity matching network to obtain at least one matching question corresponding thereto.

[0050] Since the user's spoken expression is flexible and changeable, even in different contexts, the same question semantic information actually corresponds to different questions. The present invention introduces a question similarity matching network based on the word hiding principle to accurately understand the real semantics contained in each question semantic information under different expressions.

[0051] Considering that commonly used deep learning models such as convolutional recurrent neural network (RNN), bidirectional long short-term memory neural network (Bi-directional Long Short-Term Memory, Bi-LSTM) often have problems such as difficulty in labeling and information leakage, the present invention uses the knowledge enhanced semantic representation model (Enhanced Representation from knowledge Integration, ERNIE) with semantic understanding characteristics as the basis for training and constructing a question similarity matching network, adopts the dialogue language model (DLM) for semantic modeling, and learns the implicit relationship of the corpus. By pre-training and fine-tuning the pre-built ERNIE model as a trained question similarity matching network, by modeling prior semantic knowledge such as entity concepts in massive data, learning to display the semantic relationship of the dialogue, and modeling through prior semantic knowledge units, the semantic representation ability is enhanced, and the accuracy of semantic understanding is further improved.

[0052] Specifically, after the question semantic information is input into the trained question similarity matching network, at least one matching question with the highest matching degree is automatically selected from the pre-built question and answer database.

[0053] Figure 3 : is a schematic diagram of the operation of the question intention classification network provided by the present invention, such as Figure 3 As shown, when the question semantic information is: "How many months ago does gummosis of peach trees usually appear, and how to prevent and control it?", it is input into the question similarity matching network, and by obtaining keywords in the question semantic information, such as "gumosis" and "how many months ago", at least one matching question outputted by it can be obtained: "How to prevent and control gummosis of peach trees?", and another matching question outputted by it may be: "When to prevent and control gummosis of peach trees?", or another matching question outputted by it may be: "What are the drugs for preventing and controlling gummosis of peach trees?", etc.

[0054] Furthermore, after obtaining matching questions related to the user's question, a matching answer corresponding to each matching question can be automatically found from a pre-built question-answer database based on the matching questions.

[0055] Finally, all matching responses are fed back to the user.

[0056] As an optional feedback method, the matching response can be converted into a corresponding response voice output through open source text-to-speech software for the user to accept and refer to.

[0057] The open question-answering method and device for peach production knowledge based on the migration context network provided by the present invention effectively constructs a production knowledge base through knowledge migration, and three-dimensionally improves the accuracy and robustness of production knowledge question-answering through data mining and clarification, question intent classification, question semantic understanding, and answer matching, thereby satisfying users' acquisition of production knowledge to the greatest extent.

[0058] Based on the content of the above embodiment, as an optional embodiment, before inputting the acquired question text into the trained question intention classification network, the following steps are included:

[0059] Collect multi-source data and build an agricultural knowledge base;

[0060] Using a pre-trained generative adversarial network to expand the agricultural knowledge base to obtain an enhanced agricultural knowledge base;

[0061] Directed mining and knowledge clustering are performed on the enhanced agricultural knowledge base to obtain a target industry optimization dictionary;

[0062] The pre-constructed question intention classification network is pre-trained using the target industry optimized dictionary to obtain the trained question intention classification network.

[0063] The open question-answering method for peach production knowledge provided by the present invention, before being officially put into operation, also includes: collecting multi-source data, and constructing an agricultural knowledge base from all the collected multi-source data, and expanding the collected data based on a sequence adversarial network; then, using directional mining and knowledge clustering, constructing a professional dictionary for solving the target industry to reduce the size of the knowledge base. In addition, it also includes using the corpus in the acquired professional dictionary to pre-train the question intention classification network and the question similarity matching network to improve the recognition robustness, recognition efficiency and recognition accuracy of each network.

[0064] The following is an example of open questions and answers on peach production knowledge. The above steps of collecting multi-source data and building an agricultural knowledge base include but are not limited to:

[0065] Through extensive data mining through different channels, corpus was collected according to the knowledge type of peach production to build an agricultural knowledge base.

[0066] Basically, Scrapy is an application framework suitable for crawling website data and extracting structured data. The present invention uses the Scrapy framework to implement a crawler to collect and clean data on knowledge-sharing platforms, forums, communities, and question-and-answer websites to form a widely covered agricultural knowledge base.

[0067] Furthermore, the present invention also records multiple elements on a standard question-answering platform, including the time and place of question reporting, so as to provide basic corpus support for seasonal and periodic common questions.

[0068] Importantly, since the user's conversation and question data have daily and colloquial expression characteristics, the present invention collects such data as basic data to lay the foundation for subsequent semantic understanding of the question.

[0069] Figure 4 is a flow chart of implementing corpus control based on a generative adversarial network provided by the present invention, such as Figure 4 As shown, in order to solve the defects of the limited amount of existing corpus and the time-consuming and laborious collection, the present invention uses a pre-trained generative adversarial network, such as a trained sequence adversarial network (Sequence Generative Adversarial Networks, seqGAN) on the original agricultural knowledge base (i.e. Figure 4 The sample set created by seqGAN and the original real sample set together form an enhanced agricultural knowledge base (i.e. Figure 4 ).

[0070] like Figure 4 As shown, the present invention can further combine knowledge distillation with seqGAN to reduce the requirements for hardware carrier performance and effectively lower the threshold for model deployment of seqGAN.

[0071] For example, a recurrent neural network (RNN) can be used as the generation model of seqGAN, and a convolutional neural network (CNN) can be used as the judgment model of seqGAN.

[0072] When constructing and training the generative model RNN, the knowledge distillation method can be adopted, that is, by pre-training a teacher model with a large number of parameters and strong learning ability, and then passing the features learned after pre-training to the student model with a smaller number of parameters and weaker learning ability, in the actual process of running corpus expansion, it generates a fake sample set.

[0073] On the one hand, a part of the corpus (for example, 20%) in the generated fake sample set is directly added to the manufactured sample set; on the other hand, the discriminant model CNN is used to discriminate the remaining part of the corpus in the fake sample set, and correctly add it to the manufactured sample set; finally, all the corpus in the manufactured sample set and all the corpus in the real sample set are combined to form a synthetic dataset as the final enhanced agricultural knowledge base.

[0074] Furthermore, the present invention performs directional mining and knowledge clustering on the enhanced agricultural knowledge base to obtain a target industry optimized dictionary, so as to delete invalid corpus, repeated meanings and other corpus that may appear in the enhanced agricultural knowledge base.

[0075] The open question-answering method for peach production knowledge provided by the present invention forms similar data samples through a GAN network using the idea of ​​game theory to increase the size of the training basic corpus in order to avoid the singleness of some types of agricultural corpus, especially the rare knowledge and the low sample data coverage caused by the data source and the collection method during the collection process.

[0076] As an optional embodiment, the present invention provides an open question-answering method for peach production knowledge, wherein the enhanced agricultural knowledge base is subjected to directional mining and knowledge clustering to obtain a target industry optimization dictionary, including:

[0077] Clean the enhanced agricultural knowledge base to obtain the target industry knowledge base;

[0078] Extracting seed words in the target industry knowledge base based on left-right entropy and mutual information;

[0079] Using the seed vocabulary as a labeled sample, training a pre-built deep learning model, and using the trained deep learning model to iteratively mine the target industry knowledge base to build a target industry dictionary;

[0080] A human-machine assisted clustering annotation method is used to aggregate all corpora in the target industry dictionary to obtain the target industry optimized dictionary.

[0081] (1) The acquired enhanced agricultural knowledge base related to the peach industry is cleaned, including but not limited to: filtering out chat corpus or corpus irrelevant to the peach industry to reduce its interference with the open question and answer of peach production knowledge, and finally obtaining a streamlined target industry knowledge base (herein referred to as the peach knowledge base).

[0082] The screening method may be a manual marking method or an existing automatic identification and marking method, which is not specifically limited in the present invention.

[0083] (2) By means of left-right entropy and mutual information, seed words are selected from all the corpus in the Datao knowledge base. The acquisition of seed words can play a key role in the output quality of subsequent iterative mining, and can be carried out by manual labeling.

[0084] For example, corpus related to peach planting, picking, selling, and storage can be set as seed vocabulary, which can be nouns, verbs, and prepositions. This can be set according to different actual industries.

[0085] Among them, mutual information reflects the degree of mutual dependence between two variables, such as: binary mutual information refers to the amount of correlation between two events (recorded as event X and event Y). The higher the value of mutual information, the higher the correlation between event X and event Y, and the greater the possibility that event X and event Y form a phrase; conversely, the lower the mutual information value, the lower the correlation between event X and event Y, and the greater the possibility that there is a phrase boundary between event X and event Y.

[0086] Information entropy is a measure of the uncertainty of a random variable, which can be expressed as follows:

[0087] Generally speaking, assuming that event X is a random variable that takes a finite number of values ​​(or X is a probability field of a finite number of discrete events), and the probability that event X takes the value x is P(x), then the entropy of event X can be calculated from this probability. No specific limitation is imposed on this definition and invention.

[0088] Left and right entropy refers to the entropy of the left and right boundaries of a multi-word expression. Taking the calculation of left entropy as an example, the information entropy of all possible words and word frequencies on the left side of a multi-word word is calculated and then summed.

[0089] The step of extracting seed words in the target industry knowledge base based on left-right entropy and mutual information may be:

[0090] (1) First, sort the words in any corpus that needs to be calculated in the target industry knowledge base (called the target domain vocabulary) and other domain vocabulary by frequency, that is, calculate the left and right entropy and mutual information of each domain word and other domain words, and sort them according to the calculation results.

[0091] (2) Import the target domain vocabulary and other domain vocabulary into the same data table, and divide the ranking values ​​of the same corpus in the two tables.

[0092] (3) Extract words from the target domain vocabulary by setting a threshold based on the rank ratio.

[0093] Seed words are extracted by the above-mentioned ranking ratio method. For example, the top 5,000 words in frequency order shared by the target field and other fields are used as seed words.

[0094] (3) Sowing seeds and iterative mining: After obtaining the seed vocabulary in the above way, the seed vocabulary is used as the labeled sample, and deep learning models such as CRF++ and Bi-LSTM are used to perform multiple rounds of iterative mining to form a target industry dictionary (here it can be called the big peach industry dictionary). While ensuring the accurate identification of agricultural knowledge subjects, it also improves the sustainability of the intelligent question-answering robot's dialogue.

[0095] Furthermore, in order to solve the data redundancy caused by repeated similar questions and answers in the big peach industry dictionary, or the inconsistency of user experience, a human-machine assisted clustering labeling method is adopted to effectively aggregate all the knowledge (corpus) involved in the target industry dictionary to reduce the scale of the target industry dictionary.

[0096] Specifically, select the current set of data from the target industry dictionary, and set the problems that can be classified into one category in this set of data as the central problem of the current problem, and set the current central problem. Among them, the central problem refers to the problem that can best describe the largest proportion of this set of data).

[0097] Taking into account that the directional mining of corpus is the core method to ensure the effectiveness of the corpus and is also the basis for question-answering model training, the open question-answering method of peach production knowledge based on the migration context network provided by the present invention proposes to ensure the identification of agricultural entities to the greatest extent by means of agricultural corpus cleaning, left and right entropy and mutual information seed vocabulary acquisition and seed sowing, and multi-round iterative mining using a deep learning model.

[0098] As an optional embodiment, the present invention provides a method for obtaining the target industry optimized dictionary by aggregating all corpora in the target industry dictionary using a human-machine assisted clustering annotation method, including but not limited to the following steps:

[0099] (1) Domain experts formulate classification standards and annotation specifications and provide samples as the first data;

[0100] (2) Cluster all the corpus in the target industry dictionary using the first data as the center point of the cluster, select the second data to train the labelers, and use the first data to test the labelers. If the test is passed, the next step of labeling can be carried out;

[0101] (3) Using the first data and the second data as training sets, all unclassified corpora are classified again. The corpora with high confidence can be directly adopted and added to the training data set to retrain the classifier;

[0102] (4) Select the data that is most worth labeling from the target industry dictionary, distribute it to labelers for labeling, and add the obtained labeling results to the training set to retrain the classifier;

[0103] (5) Iterate steps (3) and (4) until the accuracy of the classifier reaches a preset threshold.

[0104] (6) Use the classifier to complete the collection of all corpora in the target industry dictionary.

[0105] Based on the content of the above embodiment, as an optional embodiment, the question intention classification network is constructed based on CNN;

[0106] The pre-training of the pre-built question intention classification network using the target industry optimized dictionary includes:

[0107] Randomly select a plurality of corpus samples from the target industry optimization dictionary, and label each of the corpus samples with a corresponding classification label;

[0108] Taking a combination of each corpus sample and a corresponding classification label as a training sample, to obtain a plurality of training samples;

[0109] The pre-constructed question intention classification network is pre-trained using the multiple training samples.

[0110] In view of the complexity and diversity of issues involved in the production knowledge question-answering scenario related to peach production, such as peach varieties, pests and diseases, growth environment, sales price, etc., the present invention effectively extracts the key features in the question through the idea of ​​local perception, and adopts the CNN model as the basic model framework of the question intention classification network to achieve accurate classification of the user's better dialogue intention.

[0111] In the model pre-training stage, more than 150,000 training corpora are randomly selected from the target industry optimization dictionary. The training corpora are used as corpus samples, and human-computer collaborative data annotation is used to set classification labels for each corpus sample to complete the definition and training of the initial CNN model parameters. Based on the actual dialogue intent classification effect analysis, the initial CNN model is continuously iteratively optimized to obtain the question intent classification network.

[0112] Based on the content of the above embodiment, as an optional embodiment, the step of inputting the question semantic information into a trained question similarity matching network to obtain at least one matching question corresponding to the question text includes:

[0113] Select the knowledge-enhanced semantic representation network ERNIE model as the teacher model, and the text convolutional network TextCNN model as the student model;

[0114] Connecting the ERNIE model and the TextCNN model through a knowledge distillation layer to build a teacher-student network as the question similarity matching network;

[0115] The question semantic information is input into the teacher-student network to obtain a matching question output by the teacher-student network.

[0116] Figure 5 This is a schematic diagram of a knowledge distillation process provided by the present invention, such as Figure 5As shown, in order to effectively improve the model training efficiency and reduce the complexity of the model, the present invention adopts the knowledge distillation method to construct a question similarity matching network, which is mainly manifested in:

[0117] Use large and complex neural networks, such as the ERNIE model with large parameters and strong learning ability, as the teacher model. By training the ERNIE model and passing the features obtained from its training to small neural networks, such as the TextCNN model with small parameters and weak learning ability, the model is effectively compressed while reducing the requirements for hardware carrier performance, effectively lowering the threshold for model deployment.

[0118] The open question-answering method for peach production knowledge based on the transfer context network provided by the present invention can effectively reduce the high complexity disadvantage brought by the high-precision model, and utilizes the prior knowledge of the ERNIE model as the input of the student model through the principle of knowledge distillation, so that the low-complexity student model has similar capabilities to the teacher model, making it possible to deploy the deep network on the user side.

[0119] Based on the content of the above embodiment, as an optional embodiment, before inputting the question semantic information into the teacher-student network, the method further includes:

[0120] Using a public question-answering dataset from the entire Internet, pre-training the ERNIE model until a first preset number of training times is reached;

[0121] The pre-trained ERNIE model is fine-tuned using the target industry optimization dictionary until the training results converge or a second preset number of training times is reached.

[0122] The ERNIE model learns the semantic relations of the real world by modeling prior semantic knowledge such as entity concepts in massive data. Specifically, the ERNIE model masks semantic units such as words and entities to enable the model to learn the semantic representation of complete concepts. Compared with BERT, which learns raw language signals, the ERNIE model directly models prior semantic knowledge units, enhancing the model's semantic representation capabilities.

[0123] Since the user's spoken language is flexible and changeable, it is a great challenge for the question similarity matching network to match the correct matching question according to the recognized question semantic information. To solve this problem, the present invention uses the ERNIE model with semantic understanding characteristics, adopts DLM for semantic modeling, and learns the implicit relationship of the corpus. The specific training steps include:

[0124] (1) Pre-training stage: Introducing multi-source data knowledge, using about 1 billion pieces of public question-and-answer data from the entire Internet, and training the ERNIE model on a high-performance GPU cluster. The number of training steps can be set to 2 million.

[0125] The training corpus of the ERNIE model introduces multi-source data knowledge. In addition to modeling encyclopedia articles, it can also learn from news information and forum dialogue data. Take forum dialogue modeling as an example:

[0126] Learning about dialogue data is an important way to express semantics. Often, the same responses correspond to similar queries. Based on this assumption, the ERINE model uses DLM to model the query-response dialogue structure, takes dialogue pairs as input, introduces Dialogue Embedding to identify the roles of the dialogue, and uses Dialogue Response Loss to learn the implicit relationship of the dialogue. This method of modeling further improves the semantic representation ability of the ERINE model.

[0127] (2) Fine-tuning stage: Based on the pre-trained ERNIE model, its learning context is transferred and the ERNIE model is fine-tuned using a corpus of 100,000-level questions and answers in the big peach knowledge dictionary corresponding to the big peach production scenario, so as to significantly improve the semantic understanding generalization ability in the big peach question and answer scenario.

[0128] The open question-answering method for peach production knowledge based on the migration context network provided by the present invention uses an iteratively optimized ERNIE model to model prior semantic knowledge such as entity concepts in massive data, learn to display the semantic relationship of dialogue, and model it through prior semantic knowledge units, thereby enhancing the semantic representation ability. In the flexible and changeable question expression scenarios of agricultural question-answering, it has a high semantic understanding accuracy.

[0129] In addition, in order to understand the differences in semantics implied in questions in different contexts, we first use large-scale public network corpus to obtain richer semantic information through the ERNIE model based on the word hiding principle, which can further improve the accuracy of semantic understanding for Chinese.

[0130] Based on the content of the above embodiment, as an optional embodiment, after feeding back the matching response to the user, the method further includes:

[0131] receiving feedback information from the user regarding the matching response;

[0132] In the case where the feedback information meets the preset conditions, the question text is used as a sample, and the matching answer is used as a label corresponding to the question text to construct a feedback sample;

[0133] The sentence intention classification network and the question similarity matching network are used as an overall network, and the feedback samples are used to train the overall network.

[0134] The open question-and-answer method for peach production knowledge in the migration context network provided by the present invention provides a function of marking the answers fed back after the user asks questions during the service process, and divides the user feedback information into: very satisfied, satisfied, average, dissatisfied and very dissatisfied (which can be judged according to the user's score), and uses the feedback sample corresponding to the user's feedback information as one of the consideration factors for a new iteration.

[0135] For issues corresponding to very satisfied and satisfied feedback, the training corpus is stored in the database. For issues corresponding to general, dissatisfied and very dissatisfied feedback, the processing status of each link is checked. According to the feedback content, analysis is conducted from aspects such as speech recognition accuracy, corpus adequacy, matching model algorithm and parameters to improve the generalization of the model.

[0136] The open question-answering method for peach production knowledge based on the migration context network provided by the present invention provides a user feedback mechanism in the testing phase and the platform operation phase, effectively annotates historical questions and answers, and realizes cyclic iterative training of the model under the joint action of the platform and the user.

[0137] For example, the open question-answering method of peach knowledge oriented to the migration context network plans and practices the processing mechanisms and models of each link, allocates their weights and makes effective connections, forms a systematic question-answering service, and provides a solid guarantee for the robustness and generalization of agricultural question-answering.

[0138] Figure 6 : is a schematic diagram of the structure of the open question-answering device for peach production knowledge provided by the present invention, such as Figure 6 As shown, it mainly includes: a semantic classification unit 61, a question matching unit 62 and a response matching unit 63, wherein:

[0139] The semantic classification unit 61 is mainly used to input the obtained question text into the trained question intention classification network to obtain the question semantic information with classification labels; the question matching unit 62 is mainly used to input the question semantic information into the trained question similarity matching network to obtain at least one matching question corresponding to the question text; the answer matching unit 63 is mainly used to determine the matching answer corresponding to the question text from a pre-built question and answer database based on the matching question, and feed back the matching answer to the user.

[0140] The open question-and-answer device for peach production knowledge based on the migration context network provided by the present invention effectively constructs a production knowledge base through knowledge migration, and three-dimensionally improves the accuracy and robustness of production knowledge question-and-answer through methods such as data mining and clarification, question intent classification, question semantic understanding, and answer matching, thereby satisfying users' acquisition of production knowledge to the greatest extent.

[0141] Figure 7Schematic diagram of the structure of the open question-answering system for peach production knowledge provided by the present invention. Figure 7 As shown, it mainly includes: a voice recognition device, a search server (such as ElasticSearch), a central control business scheduling server, a Redis server, an API interface calling server and the above-mentioned peach production knowledge open question-and-answer device, wherein:

[0142] The speech recognition device can realize speech recognition service, and is used to receive the question voice input by the user to generate the question text; the ElasticSearch is implemented based on Lucene, and adopts RESTful web interface to provide a distributed multi-user full-text search engine; the central control business scheduling server is used to realize the scheduling of different types of business data by building a connection between the Redis server and the API interface calling server; the Redis server is used to provide high-availability distributed services through the master-slave data replication method, and adopts two persistence solutions, RDB and AOF, to save the memory data to the hard disk; the memory data includes a question and answer database and a target industry optimization dictionary; the API interface calling server is used to provide a connection between the speech recognition device and the big peach production knowledge open question and answer device, and to provide an Internet data interface calling service for the big peach production knowledge open question and answer system.

[0143] Specifically, the trained question intent classification network and question similarity matching network are integrated into the open question-answering system for peach production knowledge, providing services for the platform, APP, and ROS robot operating system, and providing API service interfaces for agricultural intelligent system developers. Platform services achieve natural language services that are close to human communication through the following aspects:

[0144] (1) Speech recognition: Integrate the speech recognition services publicly available on major platforms to convert speech into text and string type data structures;

[0145] (2) ElasticSearch service: Based on Lucene, it uses the RESTful web interface to provide a distributed multi-user full-text search engine with fast indexing capabilities and scalability, greatly improving data crawling efficiency.

[0146] Among them, Elasticsearch is a search server based on Lucene. It provides a distributed multi-user full-text search engine based on a RESTful web interface. Elasticsearch is developed in Java and released as open source under the terms of the Apache license. It is a popular enterprise-level search engine.

[0147] (3) Central control business dispatch service: dispatches different types of business data, including professional peach production corpus, weather, tourism, market prices and chat services, making users more natural during the conversation process.

[0148] (4) Redis service: It provides highly available distributed services through the master-slave data replication method, and uses the RDB and AOF strategies to save memory data to the hard disk to ensure data persistence and fast access speed, effectively supporting high-concurrency question-and-answer data requests.

[0149] (5) API interface call service: Provide Internet data interface call service and output the answers processed by the system.

[0150] The open question-and-answer system for peach production knowledge in the migration context network provided by the present invention starts from the actual scene of agricultural production, considers the user's requirements for the timeliness and accuracy of knowledge feedback in various operation scenes, especially during the busy farming season, and combines the application of the effectiveness of agricultural prior knowledge to propose an open question-and-answer system for peach production knowledge in the migration context network. Starting from voice or text questions, the overall construction of the peach production question-and-answer model is realized through the steps of question-and-answer pair corpus construction, intent classification, question similarity matching, answer accuracy feedback, text-to-speech conversion, etc. Redis servers are used to provide high-speed, secure, persistent, and highly concurrent data services, ElasticSearch is used to provide distributed, highly scalable, and highly real-time data search and analysis, and the API interface call service is used to provide automatic storage of source question-and-answer pairs, providing solid support for platform services.

[0151] Moreover, at the model training level, a method combining the ERNIE model, knowledge distillation, and GAN was proposed, which ensured the reliability of systematic services from the perspectives of model accuracy, service carrier convenience, and corpus expansion efficiency.

[0152] During the specific implementation process, the central control business dispatch service effectively dispatches agricultural corpus in vertical fields as well as weather, market, travel and daily chat corpus close to life, ensuring a more natural human-computer interaction method.

[0153] In addition, the system provides scalable services in terms of corpus sources, training sample annotations, and platform service forms. In particular, user feedback on matching responses can be recorded during the interaction process, providing the most effective guarantee for the iterative upgrade of the accuracy of the knowledge base.

[0154] It should be noted that the open question-and-answer device and system for peach production knowledge provided in the embodiments of the present invention can execute the open question-and-answer method for peach production knowledge based on the migration context network described in any of the above embodiments during specific operation, which will not be elaborated in this embodiment.

[0155] Figure 8 is a schematic diagram of the structure of the electronic device provided by the present invention, such as Figure 8 As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830 and a communication bus 840, wherein the processor 810, the communication interface 820 and the memory 830 communicate with each other through the communication bus 840. The processor 810 may call the logic instructions in the memory 830 to execute the open question-answering method of peach production knowledge of the migration context network, the method comprising: inputting the obtained question text into the trained question intention classification network to obtain the question semantic information with the classification label; inputting the question semantic information into the trained question similarity matching network to obtain at least one matching question corresponding to the question text; determining the matching answer corresponding to the question text from the pre-built question and answer database according to the matching question, and feeding back the matching answer to the user.

[0156] In addition, the logic instructions in the above-mentioned memory 830 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0157] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the open question-answering method of peach production knowledge based on the migration context network provided by the above methods. The method includes: inputting the obtained question text into a trained question intention classification network to obtain question semantic information with classification labels; inputting the question semantic information into a trained question similarity matching network to obtain at least one matching question corresponding to the question text; determining a matching answer corresponding to the question text from a pre-constructed question and answer database based on the matching question, and feeding back the matching answer to the user.

[0158] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the open question-answering method for peach production knowledge based on the migration context network provided in the above-mentioned embodiments, the method comprising: inputting the acquired question text into a trained question intention classification network to obtain question semantic information with a classification label; inputting the question semantic information into a trained question similarity matching network to obtain at least one matching question corresponding to the question text; determining a matching answer corresponding to the question text from a pre-constructed question-answering database based on the matching question, and feeding back the matching answer to the user.

[0159] The device embodiments described above are merely illustrative, 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. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0160] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0161] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for answering open questions about peach production knowledge based on a migration context network, characterized in that: include: Input the obtained question text into the trained question intention classification network to obtain the question semantic information with classification labels; Inputting the question semantic information into a trained question similarity matching network to obtain at least one matching question corresponding to the question text; Determine a matching answer corresponding to the question text from a pre-built question-answer database according to the matching question, and feed back the matching answer to the user; Before inputting the acquired question text into the trained question intention classification network, the following steps are included: Collect multi-source data to build an agricultural knowledge base; use a pre-trained generative adversarial network to expand the corpus of the agricultural knowledge base to obtain an enhanced agricultural knowledge base; perform directional mining and knowledge clustering on the enhanced agricultural knowledge base to obtain a target industry optimization dictionary; use the target industry optimization dictionary to pre-train a pre-built question intention classification network to obtain the trained question intention classification network; The method of performing directional mining and knowledge clustering on the enhanced agricultural knowledge base to obtain a target industry optimization dictionary includes: The enhanced agricultural knowledge base is cleaned to obtain a target industry knowledge base; seed words in the target industry knowledge base are extracted based on left-right entropy and mutual information; the pre-built deep learning model is trained with the seed words as annotation samples, and the trained deep learning model is used to iteratively mine the target industry knowledge base to construct a target industry dictionary; all the corpora in the target industry dictionary are aggregated using a human-machine assisted clustering annotation method to obtain the target industry optimized dictionary.

2. The open question answering method of peach production knowledge based on migration context network according to claim 1 is characterized in that: The question intention classification network is constructed based on a convolutional neural network model; The pre-training of the pre-built question intention classification network using the target industry optimized dictionary includes: Randomly select a plurality of corpus samples from the target industry optimization dictionary, and label each of the corpus samples with a corresponding classification label; Taking a combination of each corpus sample and a corresponding classification label as a training sample, to obtain a plurality of training samples; The pre-constructed question intention classification network is pre-trained using the multiple training samples.

3. The open question-answering method for peach production knowledge based on the migration context network according to claim 1 is characterized in that: The step of inputting the question semantic information into a trained question similarity matching network to obtain at least one matching question corresponding to the question text includes: Select the knowledge-enhanced semantic representation network ERNIE model as the teacher model, and the text convolutional network TextCNN model as the student model; Connecting the ERNIE model and the TextCNN model through a knowledge distillation layer to build a teacher-student network as the question similarity matching network; The question semantic information is input into the teacher-student network to obtain a matching question output by the teacher-student network.

4. The open question-answering method for peach production knowledge based on the migration context network according to claim 3 is characterized in that: Before inputting the question semantic information into the teacher-student network, the method further includes: Using a public question-answering dataset from the entire Internet, pre-training the ERNIE model until a first preset number of training times is reached; The pre-trained ERNIE model is fine-tuned using the target industry optimization dictionary until the training results converge or a second preset number of training times is reached.

5. The open question-answering method for peach production knowledge based on the migration context network according to claim 1 is characterized in that: After feeding back the matching response to the user, the method further includes: receiving feedback information from the user regarding the matching response; In the case where the feedback information meets the preset conditions, the question text is used as a sample, and the matching answer is used as a label corresponding to the question text to construct a feedback sample; The sentence intention classification network and the question similarity matching network are used as an overall network, and the feedback samples are used to train the overall network.

6. A peach production knowledge open question-answering device based on a migration context network, characterized in that: include: A semantic classification unit is used to input the obtained question text into the trained question intention classification network to obtain the question semantic information with classification labels; A question matching unit, used to input the question semantic information into a trained question similarity matching network to obtain at least one matching question corresponding to the question text; An answer matching unit, configured to determine a matching answer corresponding to the question text from a pre-built question and answer database according to the matching question, and feed back the matching answer to the user; Before inputting the acquired question text into the trained question intention classification network, the following steps are included: Collect multi-source data to build an agricultural knowledge base; use a pre-trained generative adversarial network to expand the corpus of the agricultural knowledge base to obtain an enhanced agricultural knowledge base; perform directional mining and knowledge clustering on the enhanced agricultural knowledge base to obtain a target industry optimization dictionary; use the target industry optimization dictionary to pre-train a pre-built question intention classification network to obtain the trained question intention classification network; The method of performing directional mining and knowledge clustering on the enhanced agricultural knowledge base to obtain a target industry optimization dictionary includes: The enhanced agricultural knowledge base is cleaned to obtain a target industry knowledge base; seed words in the target industry knowledge base are extracted based on left-right entropy and mutual information; the pre-built deep learning model is trained with the seed words as annotation samples, and the trained deep learning model is used to iteratively mine the target industry knowledge base to construct a target industry dictionary; all the corpora in the target industry dictionary are aggregated using a human-machine assisted clustering annotation method to obtain the target industry optimized dictionary.

7. A peach production knowledge open question answering system based on a migration context network, characterized in that: include: A speech recognition device, a search server, a central control business scheduling server, a Redis server, an API interface calling server, and an open question-and-answer device for peach production knowledge as described in claim 6; The speech recognition device implements a speech recognition service and is used to receive a question voice input by a user to generate the question text; The search server is implemented based on Lucene and uses a RESTful web interface to provide a distributed multi-user full-text search engine; The central control business scheduling server establishes a connection between the Redis server and the API interface calling server to implement the scheduling of different types of business data; The Redis server is used to provide high-availability distributed services through a master-slave data replication method, and uses two persistence solutions, RDB and AOF, to save memory data to a hard disk; the memory data includes a question-and-answer database and a target industry optimization dictionary; The API interface calling server is used to provide a connection between the voice recognition device and the big peach production knowledge open question and answer device, and to provide an Internet data interface calling service for the big peach production knowledge open question and answer system.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the open question-answering method for peach production knowledge based on the migration context network as described in any one of claims 1 to 5 are implemented.

Images