WO2021135455A1

WO2021135455A1 - Semantic recall method, apparatus, computer device, and storage medium

Info

Publication number: WO2021135455A1
Application number: PCT/CN2020/118454
Authority: WO
Inventors: 骆迅
Original assignee: 平安科技（深圳）有限公司
Priority date: 2020-05-13
Filing date: 2020-09-28
Publication date: 2021-07-08
Also published as: CN111767375A

Abstract

Provided is a semantic recall method, belonging to the field of artificial intelligence, comprising: upon receiving online query data, obtaining, on the basis of a sentence vector generator, an online sentence vector corresponding to the online query data (S200); obtaining a stored candidate sentence vector (S300); matching the online sentence vector and the candidate sentence vector on the basis of a sentence vector assembler to obtain the similarity between the online sentence vector and the candidate sentence vector (S400); sorting candidate sentence vectors in descending order according to similarity, and returning, as a correct answer, a first-ranked answer of a candidate question corresponding to the candidate sentence vector (S500). The method is such that without changing the accuracy of an original model, the representation layer and output layer of a conventional model are split into a sentence vector generator and an assembler, increasing the concurrency processed by the model, and improving the processing efficiency of the model when processing corpus data and the accuracy of matching of questions and answers.

Description

Semantic recall method, device, computer equipment and storage medium

This application claims the priority of a Chinese patent application filed with the Chinese Patent Office on May 13, 2020, the application number is 202010402690.9, and the invention title is "Semantic Recall Method, Device, Computer Equipment, and Storage Medium", the entire content of which is incorporated by reference In this application.

Technical field

This application relates to the field of artificial intelligence technology, in particular to a semantic recall method, device, computer equipment and storage medium.

Background technique

At present, the semantic recall model is widely used in AI question answering systems. With the development of technology, AI question answering systems are used in more and more places to replace manual question answering to improve processing efficiency. Among them, the semantic recall model is mainly based on traditional deep learning models, such as CNN, LSTM, and ESTM models.

However, with the rapid development of the information age, the corpus data that the model needs to process is also increasing, the accuracy is getting higher and higher, and the coverage is getting wider. The inventor realizes that the current semantic recall model cannot efficiently process a large amount of corpus data when processing a large amount of corpus data. Its slow training speed, long convergence time and large memory footprint ultimately lead to the efficiency of the semantic recall model in processing corpus data. Low technical problems.

Summary of the invention

The purpose of the embodiments of the present application is to propose a semantic recall method, device, computer equipment, and storage medium, aiming to solve the technical problem of low efficiency of semantic recall model processing corpus data.

In order to solve the above technical problems, an embodiment of the present application provides a semantic recall method, which adopts the following technical solutions:

A semantic recall method includes the following steps:

When receiving online query data, obtain the online sentence vector corresponding to the online query data based on the sentence vector generator;

Obtain the stored candidate sentence vector;

Matching the online sentence vector and the candidate sentence vector based on the sentence vector splicer to obtain the similarity between the online sentence vector and the candidate sentence vector;

The candidate sentence vectors are sorted in descending order according to the similarity, and the answer to the candidate question corresponding to the first-ranked candidate sentence vector is returned as the correct answer.

In order to solve the above technical problems, an embodiment of the present application also provides a semantic recall device, which adopts the following technical solutions:

The first obtaining module is configured to obtain the online sentence vector corresponding to the online query data based on the sentence vector generator when the online query data is received;

The second obtaining module is used to obtain the stored candidate sentence vector;

A splicing module, configured to match the online sentence vector and the candidate sentence vector based on a sentence vector splicer to obtain the similarity between the online sentence vector and the candidate sentence vector;

The sorting module is configured to sort the candidate sentence vectors in descending order according to the similarity, and return the answer to the candidate question corresponding to the candidate sentence vector ranked first as the correct answer.

In order to solve the above technical problems, an embodiment of the present application also provides a computer device, including a memory and a processor, and computer-readable instructions stored in the memory and capable of running on the processor, and the processor executes The computer-readable instructions implement the steps of the semantic recall method as described below:

Obtain the stored candidate sentence vector;

In order to solve the above technical problems, embodiments of the present application also provide a computer-readable storage medium, the computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, the following Steps of semantic recall method:

Obtain the stored candidate sentence vector;

The details of one or more embodiments of the present application are presented in the following drawings and description, and other features and advantages of the present application will become apparent from the description, drawings and claims.

The above semantic recall method, device, computer equipment and storage medium, when online query data is received, the online query data is the input sentence, and the online query data corresponding to the online query data is obtained based on the sentence vector generator. Sentence vector, the online sentence vector is the vector form data corresponding to the online query data; obtain the stored candidate sentence vector, where the candidate sentence vector is the sentence vector corresponding to the candidate question stored in the database in advance; based on the sentence vector The splicer matches the online sentence vector and the candidate sentence vector to obtain the similarity between the online sentence vector and the candidate sentence vector; the candidate sentence vector can be screened according to the similarity, so as to filter out the The candidate sentence vector that best matches the online sentence vector. Specifically, the candidate sentence vectors are sorted in descending order according to the similarity, and the answer to the candidate question corresponding to the candidate sentence vector ranked first is returned as the correct answer. As a result, the technical problem that the semantic recall model is inefficient in processing corpus data is solved.

Description of the drawings

In order to explain the solution in this application more clearly, the following will briefly introduce the drawings used in the description of the embodiments of the application. Obviously, the drawings in the following description are some embodiments of the application. Ordinary technicians can obtain other drawings based on these drawings without creative work.

Figure 1 is an exemplary system architecture diagram to which the present application can be applied;

Figure 2 is a flowchart of an embodiment of a semantic recall method;

Figure 3 is a schematic diagram of a sentence vector generator;

Figure 4 is a schematic diagram of a sentence vector splicer;

Fig. 5 is a schematic structural diagram of an embodiment of a semantic recall device according to the present application;

Fig. 6 is a schematic structural diagram of an embodiment of a computer device according to the present application.

Reference signs: semantic recall device 600, first acquisition module 610, second acquisition module 620, splicing module 630, sorting module 640.

Detailed ways

Unless otherwise defined, all technical and scientific terms used herein have the same meanings as commonly understood by those skilled in the technical field of the application; the terms used in the specification of the application herein are only for describing specific embodiments. The purpose is not to limit the application; the terms "including" and "having" in the specification and claims of the application and the above-mentioned description of the drawings and any variations thereof are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order.

The reference to "embodiments" herein means that a specific feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art clearly and implicitly understand that the embodiments described herein can be combined with other embodiments.

In order to make the objectives, technical solutions, and advantages of this application clearer, the following further describes the application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of this application.

As shown in FIG. 1, the system architecture 100 may include

terminal devices

101, 102, 103, a network 104, and a server 105. The network 104 is used to provide a medium for communication links between the

terminal devices

101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired, wireless communication links, or fiber optic cables, and so on.

The user can use the

terminal devices

101, 102, and 103 to interact with the server 105 through the network 104 to receive or send messages and so on. Various communication client applications, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, and social platform software, may be installed on the

terminal devices

101, 102, and 103.

The

terminal devices

101, 102, and 103 may be various electronic devices with display screens and supporting web browsing, including but not limited to smart phones, tablet computers, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III, dynamic Video experts compress standard audio layer 3), MP4 (Moving Picture Experts Group Audio Layer IV, dynamic image experts compress standard audio layer 4) players, laptop portable computers and desktop computers, etc.

The server 105 may be a server that provides various services, for example, a background server that provides support for pages displayed on the

terminal devices

101, 102, and 103.

It should be noted that the semantic recall method provided in the embodiments of the present application is generally executed by the server/terminal, and correspondingly, the semantic recall device is generally installed in the server/terminal device.

It should be understood that the numbers of terminals, networks, and servers in FIG. 1 are merely illustrative. There can be any number of terminal devices, networks, and servers according to implementation needs.

Continuing to refer to FIG. 2, a flowchart of an embodiment of the semantic recall method according to the present application is shown. The semantic recall method includes the following steps:

Step S200, when receiving online query data, obtain an online sentence vector corresponding to the online query data based on the sentence vector generator;

Online query data is real-time query data received online. When the online query data is received, the online sentence vector corresponding to the online query data is obtained based on the sentence vector generator. Wherein, the obtained online sentence vector is the sentence vector corresponding to the online query data. Specifically, when online query data is received, the online query data is a sentence, the sentence is input to the tokenizer layer in the sentence vector generator, and the word in the online query data is id based on the tokenizer layer Transformation means converting each word in the sentence into the format of ID. Then the ID is passed through the embedding layer to obtain the word vector corresponding to each word in the online query data. When the word vector is obtained, convolution processing is performed on the word vector to obtain the online sentence vector corresponding to the current online query data.

The sentence vector generator is an independent model structure for processing online query data. The traditional deep learning model usually includes a representation layer and an output layer. The representation layer and output layer of the traditional deep learning model are separated, and the part of the representation layer is separated. As a sentence vector generator, the corresponding sentence vector generator is obtained. Take the CNN model as an example. In the CNN model, the sentence vector generator is shown in Figure 3.

It can be seen from Figure 3 that in the model, q1(char) represents the input layer q1 of the sentence, that is, online query data, and then through the embedding layer to obtain the word vector corresponding to each word in the online query data, the word vector passes (Conv+GlobalMaxPooling)*3, that is, the three-layer convolutional neural network performs convolution processing to obtain the convolution result, where conv is convolution and GlobalMaxPooling is global pooling. Concat splices the obtained convolution results, and outputs the spliced result to obtain the online sentence vector corresponding to the online query data. Among them, because three layers of convolution are performed, the results of each layer of convolution must be spliced. The purpose of multi-layer convolution is to make the obtained data more accurate. Therefore, other models may not include concat.

Step S300: Obtain a stored candidate sentence vector;

The candidate sentence vector is pre-stored in the database, and the candidate sentence vector is obtained and stored in advance by the sentence vector generator for the candidate question. In the question answering system, candidate questions are obtained in advance, and the sentence vector of the candidate question is generated offline through the offline sentence vector generator. When the candidate sentence vector corresponding to the candidate question is obtained, the candidate sentence vector is stored in the database in.

Step S400, matching the online sentence vector and the candidate sentence vector based on the sentence vector splicer to obtain the similarity between the online sentence vector and the candidate sentence vector;

When the candidate sentence vector and the online sentence vector are obtained, the similarity between the candidate sentence vector and the online sentence vector is calculated based on the vector splicer. Specifically, the difference feature vectors of the candidate sentence vector and the online sentence vector in different measurement dimensions are calculated, and finally the difference feature vectors in different measurement dimensions are combined and spliced to obtain the final difference between the candidate sentence vector and the online sentence vector Feature vector. When the final difference feature vector is obtained, regularization processing is performed on the difference feature vector to obtain the similarity between the online sentence vector and the candidate sentence vector.

Take the CNN model as an example. In the CNN model, the sentence vector splicer is shown in Figure 4. It can be seen from Figure 4 that in the model, q1 (real-time) represents the online sentence vector, q2 (offline) represents the candidate sentence vector. When the online sentence vector and candidate sentence vector are obtained, the online sentence vector and candidate sentence The vector is input to Diff+Mul+Max; Diff+Mul+Max calculates the difference feature vector of the online sentence vector and candidate sentence vector from the three measurement dimensions of subtraction, multiplication and maximum, thereby obtaining the online sentence The difference feature vector of the vector and the candidate sentence vector in the three dimensions; concat splices the difference feature vector calculated in the three measurement dimensions to obtain the final difference feature vector; input the final difference feature vector to 3*( Dense+BatchNormalization+Relu+Dropout) to make it regularize the final difference feature vector obtained by splicing. Then input the result of the regularization process into Sigmoid, Sigmoid is the activation function, and use

Indicates that the result of the regularization process is passed through the activation function to obtain the similarity between the online sentence vector and the candidate sentence vector.

Step S500: Sort the candidate sentence vectors in descending order according to the similarity, and return the answer to the candidate question corresponding to the candidate sentence vector ranked first as the correct answer.

When determining the similarity between the online sentence vector and the candidate sentence vector, the candidate questions corresponding to the candidate sentence vector are sorted in descending order according to the similarity, that is, sorted from large to small. The answer to the candidate question corresponding to the candidate sentence vector with the highest similarity of the sentence vector on the line is selected as the correct answer. Use the correct answer as the correct answer for online query data and push it to the user interface.

In this embodiment, it is realized that the representation layer and output layer of the traditional model are separated into a sentence vector generator and a splicer without changing the accuracy of the original model. When the sentence vector is obtained, only a single The sentence vector generator processes the data, and then through the sentence vector splicer, the processed data and the candidate sentence vector are spliced, without the need for the overall model structure, which increases the amount of concurrency of model processing and improves the model’s ability to process corpus Data processing efficiency and accuracy of question and answer matching. And can be applied to a variety of different types of models, with mobility and high scalability. This application belongs to the field of artificial intelligence technology and has good performance in both machine learning and deep learning.

In some embodiments of the present application, step S200 includes:

Obtain the word vector of the online query data based on the sentence vector generator;

Multi-layer convolution processing is performed on the word vector to obtain the online sentence vector of the online query data.

The online query data is a single sentence, where the word vector is the vector corresponding to each word in the single sentence. When the online query data is received, each word in the online query data is idized according to the word frequency, TF-IDF (term frequency—inverse document frequency) and other characteristics in the online query data to obtain the The ID corresponding to each word in the online query data. Then, based on the mbedding (embedding) layer, the word vector corresponding to each ID in the online query data is obtained. In the embedding layer, there is a mapping relationship between the ID and the word vector. When the ID of each word is obtained, the embedding layer is passed The word vector corresponding to each word in the online query data can be obtained.

When the word vector corresponding to each word in the online query data is obtained, all the word vectors included in the online query data are input to the convolutional neural network, and the word vector is processed based on the convolutional neural network to obtain convolution As a result, the convolution result is a set of sentence vectors corresponding to the online query data. However, a set of sentence vectors cannot fully reflect the characteristic information of the current online query data. Therefore, all the word vectors obtained in the online query data are subjected to multi-layer convolution processing to obtain multiple sets of convolution results. The multiple sets of convolution results obtained are spliced together, and the final result obtained is the online sentence vector corresponding to the online query data.

In this embodiment, the online sentence vector of the online query data is obtained according to the word vector. There is no need for a complete model structure. Only a sentence vector generator is needed to obtain the corresponding online sentence vector, which improves The model's processing efficiency on corpus data has further improved the concurrency of model processing.

In some embodiments of the present application, based on the sentence vector generator described above, acquiring the word vector of the online query data includes:

Based on the tag analysis layer of the sentence vector generator, IDize each word in the online query data to obtain an ID corresponding to each word in the online query data;

The ID is feature-encoded based on the embedding layer of the sentence vector generator to obtain a word vector corresponding to each word in the online query data.

The token analysis layer is the tokenizer layer, and each word in the received online query data can be IDized according to the tokenizer layer. Specifically, when online query data is received, the word frequency, tfidf and other characteristics of the online query data are obtained. Based on the characteristics, the tokenizer layer can ID each word in the online query data, for example, The ID of the word division with the word frequency of 5 is 001. After the ID of each word in the online query data is completed in the tokenizer layer, the ID of each word obtained is input to the embedding layer, that is, the embedding layer. The Embedding layer determines the word vector corresponding to each word according to the ID, that is, based on the embedding layer, the ID of each word is feature-coded, and the mapping between each word and the multi-dimensional space is determined, thereby obtaining the currently input online query data The word vector corresponding to each word in.

In this embodiment, the analysis and extraction of online query data according to the tag analysis layer and the embedding layer are realized, the efficiency and accuracy of online query data analysis are improved, and the corresponding matching data obtained from online query data is further improved. (That is, the correct answer) efficiency and accuracy.

In some embodiments of the present application, performing multi-layer convolution processing on the word vector to obtain the online sentence vector of the online query data includes:

Performing multi-layer convolution processing on the word vector based on a convolutional neural network to obtain semantic features corresponding to the online query data;

The semantic features obtained each time are spliced together to obtain the online sentence vector of the online query data.

When the word vector corresponding to each word in the online query data is obtained, it is determined that the online query data is based on the semantic feature of the word, and the semantic feature is the logical representation based on the word in the online query data. Through a convolutional neural network (such as a CNN three-layer convolutional neural network), the semantic features of the online query data can be extracted based on the obtained word vector. Specifically, the word vector of each word in the online query data is subjected to convolution processing through the convolutional neural network, and the convolution result obtained is the semantic feature of the online query data based on the word, and the semantic feature is also A set of vectors. Multi-layer convolution is performed on the word vector through the convolutional neural network, and the semantic features obtained each time are spliced together to obtain the online sentence vector corresponding to the online query data. For example, a three-layer convolutional neural network is used to perform three-layer convolution on all word vectors in the online query data, and the result of the three-layer convolution, that is, the semantic feature, is spliced together, and the output is the online The online sentence vector corresponding to the query data.

In this embodiment, the online sentence vector corresponding to the online query data is obtained by stitching according to semantic features, which improves the accuracy of obtaining the online sentence vector corresponding to the online query data, and further improves the The accuracy of the vector matching to get the correct answer.

In some embodiments of the present application, step S300 includes:

Obtain candidate questions stored in the question library;

Perform offline calculation on the candidate question based on the sentence vector generator to obtain a candidate sentence vector corresponding to the candidate question.

Candidate questions are pre-collected questions, and the candidate questions are pre-stored in the question library. When the candidate question is obtained, the candidate sentence vectors are calculated one by one for all candidate questions in the question library. Specifically, when the candidate question is obtained, the candidate question is calculated offline based on the sentence vector generator, and the calculation process is the same as that of the online sentence vector. However, the candidate sentence vector can also be calculated offline based on the sentence vector generator without a network connection; for online sentence vectors, the sentence vector generator only performs real-time calculations on the received online questions. For example, {How is diabetes formed: [0.76,0.54,0.77,...,0.65,0.23,0.13], how should diabetes be treated: [0.12,0.25,0.65,...,0.11,0.86,0.92]}, where, " How does diabetes form?" "How should diabetes be treated?" The numbers after the two sentences are the sentence vectors that correspond to these two sentences. When the two questions are candidate questions, the questions are stored in the form of candidate sentence vectors.

In this embodiment, the calculation of the candidate sentence vector of the candidate question is realized, and the pre-calculation and storage of the candidate sentence vector saves the matching time during question and answer matching, and improves the efficiency of obtaining answers.

In some embodiments of the present application, after the above step of performing offline calculation on the candidate question based on the sentence vector generator to obtain the candidate sentence vector corresponding to the candidate question, the method further includes:

Acquiring unique identification information corresponding to each candidate sentence vector;

According to the identification information, the candidate sentence vector is stored in a database in a dictionary in association with the candidate question.

When the candidate sentence vector is obtained, the candidate sentence vector is stored in the form of a dictionary. Specifically, each candidate sentence vector corresponds to unique identification information, and the candidate sentence vector and its corresponding candidate question are associated and stored according to the identification information. When extracting the candidate sentence vector and the corresponding candidate question, it can be directly extracted based on the identification information.

In this embodiment, pre-storage of candidate sentence vectors in the form of a dictionary is realized, which further improves the extraction efficiency of candidate sentence vectors during matching, and saves the duration of question and answer matching.

In some embodiments of the present application, the semantic recall method further includes:

Calculating the difference feature vectors of the online sentence vector and the candidate sentence vector in the three measurement dimensions of multiplication, subtraction, and maximum;

Splicing the difference feature vectors in the three measurement dimensions together to obtain the final difference feature vector;

Performing regularization processing on the final difference feature vector to obtain a processing result;

Performing function processing on the processing result to obtain the similarity between the online sentence vector and the candidate sentence vector.

When the online sentence vector and the candidate sentence vector are obtained, the difference feature vectors of the online sentence vector and the candidate sentence vector in the three measurement dimensions of multiplication, subtraction and maximum are calculated respectively. Among them, the multiplication is to perform dot multiplication on the online sentence vector and the candidate sentence vector, and the result obtained is the difference feature vector of the online sentence vector and the candidate sentence vector in the multiplication measurement dimension; the subtraction is the online sentence The vector and the candidate sentence vector are subtracted, and the result is the difference feature vector of the line sentence vector and the candidate sentence vector in the subtraction measurement dimension; the maximum value is the maximum value of the line sentence vector and the candidate sentence vector , The maximum value obtained is the feature vector of the difference between the line sentence vector and the candidate sentence vector in the maximum measurement dimension. The difference feature values corresponding to the three measurement dimensions of the multiplication, subtraction, and maximum value are spliced together to obtain the final difference feature vector of the online sentence vector and the candidate sentence vector. Among them, the measurement dimension includes, but is not limited to, three measurement dimensions of multiplication, subtraction, and maximum value, and may also include measurement dimensions such as minimum value.

When the final difference feature vector is obtained, the difference feature vector is regularized, and subjected to dense layer dimensionality reduction and activation function sigmoid processing. Among them, the sigmoid function can map the variable to between 0 and 1, which can be obtained An output result is a probability value from 0 to 1. According to the probability value, the similarity between the online sentence vector and the candidate sentence vector is measured; if the probability is greater than 0.5, it is determined that the online sentence vector and the candidate sentence vector are similar, otherwise, they are not similar.

In this embodiment, the splicing and matching of online sentence vectors and candidate sentence vectors is realized, and the processing of the entire model is also not required, which improves the processing efficiency of the model, and the candidate sentence vector with the highest matching degree is determined through the similarity output, and further This greatly improves the accuracy of obtaining answers to questions.

Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing relevant hardware through computer-readable instructions, which can be stored in a computer-readable storage medium. When the computer-readable instructions are executed, they may include the processes of the above-mentioned method embodiments. Among them, the aforementioned storage medium may be a non-volatile storage medium such as a magnetic disk, an optical disc, a read-only memory (Read-Only Memory, ROM), or a random access memory (Random Access Memory, RAM), etc.

It should be understood that although the various steps in the flowchart of the drawings are displayed in sequence as indicated by the arrows, these steps are not necessarily performed in sequence in the order indicated by the arrows. Unless explicitly stated in this article, the execution of these steps is not strictly limited in order, and they can be executed in other orders. Moreover, at least part of the steps in the flowchart of the drawings may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the order of execution is also It is not necessarily performed sequentially, but may be performed alternately or alternately with at least a part of other steps or sub-steps or stages of other steps.

With further reference to FIG. 5, as an implementation of the method shown in FIG. 2, this application provides an embodiment of a semantic recall device. The device embodiment corresponds to the method embodiment shown in FIG. Used in various electronic devices.

As shown in FIG. 5, the semantic recall device 600 in this embodiment includes:

The first obtaining module 610 is configured to obtain the online sentence vector corresponding to the online query data based on the sentence vector generator when the online query data is received;

Wherein, the first obtaining module 610 includes:

The first obtaining unit is configured to obtain the word vector of the online query data based on the sentence vector generator;

The first processing unit is configured to perform multi-layer convolution processing on the word vector to obtain the online sentence vector of the online query data.

The first acquiring unit further includes:

The second processing unit is configured to perform ID processing on each word in the online query data based on the tag analysis layer of the sentence vector generator to obtain an ID corresponding to each word in the online query data;

The third processing unit is configured to perform feature encoding on the ID based on the embedding layer of the sentence vector generator to obtain a word vector corresponding to each word in the online query data.

The first processing unit further includes:

A fourth processing unit, configured to perform multi-layer convolution processing on the word vector based on a convolutional neural network to obtain semantic features corresponding to the online query data;

The first splicing unit is used to splice the semantic features obtained each time together to obtain the online sentence vector of the online query data.

It can be seen from Figure 3 that in the model, q1(char) represents the input sentence q1, that is, online query data, and then through the embedding layer to obtain the word vector corresponding to each word in the online query data, the word vector passes ( Conv+GlobalMaxPooling)*3, that is, the three-layer convolutional neural network performs convolution processing to obtain the convolution result, where conv is convolution and GlobalMaxPooling is global pooling. Concat splices the obtained convolution results, and outputs the spliced result to obtain the online sentence vector corresponding to the online query data. Among them, because three layers of convolution are performed, the results of each layer of convolution must be spliced. The purpose of multi-layer convolution is to make the obtained data more accurate. Therefore, other models may not include concat.

The second obtaining module 620 is configured to obtain stored candidate sentence vectors;

Wherein, the second obtaining module 620 includes:

The second obtaining unit is used to obtain candidate questions stored in the question library;

The first calculation unit is configured to perform offline calculation on the candidate question based on the sentence vector generator to obtain a candidate sentence vector corresponding to the candidate question.

The third acquiring unit is configured to acquire the unique identification information corresponding to each candidate sentence vector;

The storage unit is configured to store the candidate sentence vector in a dictionary in the form of a dictionary in association with the candidate question in a database according to the identification information.

The splicing module 630 is configured to match the online sentence vector and the candidate sentence vector based on a sentence vector splicer to obtain the similarity between the online sentence vector and the candidate sentence vector;

Wherein, the splicing module includes:

The second calculation unit is used to calculate the difference feature vector of the online sentence vector and the candidate sentence vector in the three measurement dimensions of multiplication, subtraction, and maximum;

The second splicing unit is used to splice the difference feature vectors in the three measurement dimensions together to obtain the final difference feature vector;

A fifth processing unit, configured to perform regularization processing on the final difference feature vector to obtain a processing result;

The sixth processing unit is configured to perform function processing on the processing result to obtain the similarity between the online sentence vector and the candidate sentence vector.

The sorting module 640 is configured to sort the candidate sentence vectors in descending order according to the similarity, and return the answer to the candidate question corresponding to the first-ranked candidate sentence vector as the correct answer.

In order to solve the above technical problems, the embodiments of the present application also provide computer equipment. Please refer to FIG. 6 for details. FIG. 6 is a block diagram of the basic structure of the computer device in this embodiment.

The computer device 6 includes a memory 61, a processor 62, and a network interface 63 that communicate with each other through a system bus. It should be pointed out that only the computer device 6 with components 61-63 is shown in the figure, but it should be understood that it is not required to implement all the illustrated components, and more or fewer components may be implemented instead. Among them, those skilled in the art can understand that the computer device here is a device that can automatically perform numerical calculation and/or information processing in accordance with pre-set or stored instructions. Its hardware includes, but is not limited to, a microprocessor, a dedicated Integrated Circuit (Application Specific Integrated Circuit, ASIC), Programmable Gate Array (Field-Programmable Gate Array, FPGA), Digital Processor (Digital Signal Processor, DSP), embedded equipment, etc.

The computer device may be a computing device such as a desktop computer, a notebook, a palmtop computer, and a cloud server. The computer device can interact with the user through a keyboard, a mouse, a remote control, a touch panel, or a voice control device.

The memory 61 includes at least one type of readable storage medium, the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static memory Random access memory (SRAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), programmable read only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. The computer-readable storage medium may be non-volatile or volatile. In some embodiments, the memory 61 may be an internal storage unit of the computer device 6, such as a hard disk or a memory of the computer device 6. In other embodiments, the memory 61 may also be an external storage device of the computer device 6, such as a plug-in hard disk equipped on the computer device 6, a smart media card (SMC), a secure digital (Secure Digital, SD) card, Flash Card, etc. Of course, the memory 61 may also include both the internal storage unit of the computer device 6 and its external storage device. In this embodiment, the memory 61 is generally used to store an operating system and various application software installed in the computer device 6, such as computer-readable instructions of a semantic recall method. In addition, the memory 61 can also be used to temporarily store various types of data that have been output or will be output.

In some embodiments, the processor 62 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chips. The processor 62 is generally used to control the overall operation of the computer device 6. In this embodiment, the processor 62 is configured to execute computer-readable instructions or process data stored in the memory 61, for example, computer-readable instructions for executing the semantic recall method.

The network interface 63 may include a wireless network interface or a wired network interface, and the network interface 63 is generally used to establish a communication connection between the computer device 6 and other electronic devices.

In this embodiment, the computer device realizes that without changing the accuracy of the original model, the representation layer and the output layer of the traditional model are split into a sentence vector generator and a splicer, respectively. When obtaining the sentence vector , Only need to process the data through a single sentence vector generator, and then use the sentence vector splicer to splice the processed data and candidate sentence vectors, without the need for the overall model structure, which increases the amount of concurrency of model processing. The processing efficiency of the model in processing corpus data and the accuracy of question and answer matching. And can be applied to a variety of different types of models, with mobility and high scalability.

This application also provides another implementation manner, that is, to provide a computer-readable storage medium that stores a semantic recall process, and the semantic recall process can be executed by at least one processor to enable all The at least one processor executes the steps of the semantic recall method as described above.

In this embodiment, the computer-readable storage medium realizes that without changing the accuracy of the original model, the representation layer and the output layer of the traditional model are split into a sentence vector generator and a splicer, respectively. When using sentence vectors, only a single sentence vector generator is needed to process the data, and then a sentence vector splicer is used to splice the processed data and candidate sentence vectors, without the need for the overall model structure, which improves the concurrency of model processing. This improves the model’s processing efficiency and the accuracy of Q&A matching when processing corpus data. And can be applied to a variety of different types of models, with mobility and high scalability.

Through the description of the above implementation manners, those skilled in the art can clearly understand that the above-mentioned embodiment method can be implemented by means of software plus the necessary general hardware platform, of course, it can also be implemented by hardware, but in many cases the former is better.的实施方式。 Based on this understanding, the technical solution of this application essentially or the part that contributes to the existing technology can be embodied in the form of a software product, and the computer software product is stored in a storage medium (such as ROM/RAM, magnetic disk, The optical disc) includes several instructions to make a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) execute the methods described in the various embodiments of the present application.

Obviously, the embodiments described above are only a part of the embodiments of the present application, rather than all of the embodiments. The drawings show preferred embodiments of the present application, but do not limit the patent scope of the present application. This application can be implemented in many different forms. On the contrary, the purpose of providing these examples is to make the understanding of the disclosure of this application more thorough and comprehensive. Although this application has been described in detail with reference to the foregoing embodiments, for those skilled in the art, it is still possible for those skilled in the art to modify the technical solutions described in each of the foregoing specific embodiments, or equivalently replace some of the technical features. . All equivalent structures made by using the contents of the description and drawings of this application, directly or indirectly used in other related technical fields, are similarly within the scope of patent protection of this application.

Claims

A semantic recall method includes the following steps:

When receiving online query data, obtain the online sentence vector corresponding to the online query data based on the sentence vector generator;

Obtain the stored candidate sentence vector;

Matching the online sentence vector and the candidate sentence vector based on the sentence vector splicer to obtain the similarity between the online sentence vector and the candidate sentence vector;

The candidate sentence vectors are sorted in descending order according to the similarity, and the answer to the candidate question corresponding to the first-ranked candidate sentence vector is returned as the correct answer.
The semantic recall method according to claim 1, wherein the step of obtaining the online sentence vector corresponding to the online query data by the sentence-based vector generator comprises:

Obtain the word vector of the online query data based on the sentence vector generator;

Multi-layer convolution processing is performed on the word vector to obtain the online sentence vector of the online query data.
The semantic recall method according to claim 2, wherein the step of obtaining the word vector of the online query data based on the sentence vector generator comprises:

Based on the tag analysis layer of the sentence vector generator, IDize each word in the online query data to obtain an ID corresponding to each word in the online query data;

The ID is feature-encoded based on the embedding layer of the sentence vector generator to obtain a word vector corresponding to each word in the online query data.
The semantic recall method according to claim 2, wherein the step of performing multi-layer convolution processing on the word vector to obtain the online sentence vector of the online query data comprises:

Performing multi-layer convolution processing on the word vector based on a convolutional neural network to obtain semantic features corresponding to the online query data;

The semantic features obtained each time are spliced together to obtain the online sentence vector of the online query data.
The semantic recall method according to claim 1, wherein the step of obtaining the stored candidate sentence vector comprises:

Obtain candidate questions stored in the question library;

Perform offline calculation on the candidate question based on the sentence vector generator to obtain a candidate sentence vector corresponding to the candidate question.
The semantic recall method according to claim 5, wherein after the step of performing offline calculation on the candidate question based on the sentence vector generator to obtain the candidate sentence vector corresponding to the candidate question, the method further comprises:

Acquiring unique identification information corresponding to each candidate sentence vector;

According to the identification information, the candidate sentence vector is stored in a database in a dictionary in association with the candidate question.
The semantic recall method according to any one of claims 1 to 6, wherein the sentence vector-based splicer matches the online sentence vector and the candidate sentence vector to obtain the online sentence vector and the candidate sentence vector The steps of similarity of sentence vectors include:

Calculating the difference feature vectors of the online sentence vector and the candidate sentence vector in the three measurement dimensions of multiplication, subtraction, and maximum;

Splicing the difference feature vectors in the three measurement dimensions together to obtain the final difference feature vector;

Performing regularization processing on the final difference feature vector to obtain a processing result;

Performing function processing on the processing result to obtain the similarity between the online sentence vector and the candidate sentence vector.
A semantic recall device includes:

The first obtaining module is configured to obtain the online sentence vector corresponding to the online query data based on the sentence vector generator when the online query data is received;

The second obtaining module is used to obtain the stored candidate sentence vector;

A splicing module, configured to match the online sentence vector and the candidate sentence vector based on a sentence vector splicer to obtain the similarity between the online sentence vector and the candidate sentence vector;

The sorting module is configured to sort the candidate sentence vectors in descending order according to the similarity, and return the answer to the candidate question corresponding to the candidate sentence vector ranked first as the correct answer.
A computer device includes a memory and a processor, wherein computer-readable instructions are stored in the memory, and when the processor executes the computer-readable instructions, the steps of the semantic recall method as described below are implemented:

When receiving online query data, obtain the online sentence vector corresponding to the online query data based on the sentence vector generator;

Obtain the stored candidate sentence vector;

Matching the online sentence vector and the candidate sentence vector based on the sentence vector splicer to obtain the similarity between the online sentence vector and the candidate sentence vector;

The candidate sentence vectors are sorted in descending order according to the similarity, and the answer to the candidate question corresponding to the first-ranked candidate sentence vector is returned as the correct answer.
9. The computer device according to claim 9, wherein the step of obtaining the online sentence vector corresponding to the online query data by the sentence-based vector generator comprises:

Obtain the word vector of the online query data based on the sentence vector generator;

Multi-layer convolution processing is performed on the word vector to obtain the online sentence vector of the online query data.
The computer device according to claim 10, wherein the step of obtaining the word vector of the online query data by the sentence vector generator comprises:

Based on the tag analysis layer of the sentence vector generator, IDize each word in the online query data to obtain an ID corresponding to each word in the online query data;

The ID is feature-encoded based on the embedding layer of the sentence vector generator to obtain a word vector corresponding to each word in the online query data.
The computer device according to claim 10, wherein the step of performing multi-layer convolution processing on the word vector to obtain the online sentence vector of the online query data comprises:

Performing multi-layer convolution processing on the word vector based on a convolutional neural network to obtain semantic features corresponding to the online query data;

The semantic features obtained each time are spliced together to obtain the online sentence vector of the online query data.
The computer device according to claim 9, wherein the step of obtaining the stored candidate sentence vector comprises:

Obtain candidate questions stored in the question library;

Perform offline calculation on the candidate question based on the sentence vector generator to obtain a candidate sentence vector corresponding to the candidate question.
The computer device according to claim 13, wherein after the step of performing offline calculation on the candidate question based on the sentence vector generator to obtain the candidate sentence vector corresponding to the candidate question, the method further comprises:

Acquiring unique identification information corresponding to each candidate sentence vector;

According to the identification information, the candidate sentence vector is stored in a database in a dictionary in association with the candidate question.
The computer device according to any one of claims 9 to 14, wherein the sentence vector-based splicer matches the online sentence vector and the candidate sentence vector to obtain the online sentence vector and the candidate sentence The steps of vector similarity include:

Calculating the difference feature vectors of the online sentence vector and the candidate sentence vector in the three measurement dimensions of multiplication, subtraction, and maximum;

Splicing the difference feature vectors in the three measurement dimensions together to obtain the final difference feature vector;

Performing regularization processing on the final difference feature vector to obtain a processing result;

Performing function processing on the processing result to obtain the similarity between the online sentence vector and the candidate sentence vector.
A computer-readable storage medium having computer-readable instructions stored thereon, and when the computer-readable instructions are executed by a processor, the steps of the semantic recall method as described below are realized:

When receiving online query data, obtain the online sentence vector corresponding to the online query data based on the sentence vector generator;

Obtain the stored candidate sentence vector;

Matching the online sentence vector and the candidate sentence vector based on the sentence vector splicer to obtain the similarity between the online sentence vector and the candidate sentence vector;

The candidate sentence vectors are sorted in descending order according to the similarity, and the answer to the candidate question corresponding to the first-ranked candidate sentence vector is returned as the correct answer.
15. The computer-readable storage medium according to claim 16, wherein the step of obtaining the online sentence vector corresponding to the online query data by the sentence-based vector generator comprises:

Obtain the word vector of the online query data based on the sentence vector generator;

Multi-layer convolution processing is performed on the word vector to obtain the online sentence vector of the online query data.
18. The computer-readable storage medium according to claim 17, wherein the step of obtaining the word vector of the online query data based on the sentence vector generator comprises:

Based on the tag analysis layer of the sentence vector generator, IDize each word in the online query data to obtain an ID corresponding to each word in the online query data;

The ID is feature-encoded based on the embedding layer of the sentence vector generator to obtain a word vector corresponding to each word in the online query data.
18. The computer-readable storage medium according to claim 17, wherein the step of performing multi-layer convolution processing on the word vector to obtain the online sentence vector of the online query data comprises:

Performing multi-layer convolution processing on the word vector based on a convolutional neural network to obtain semantic features corresponding to the online query data;

The semantic features obtained each time are spliced together to obtain the online sentence vector of the online query data.
The computer-readable storage medium according to claim 16, wherein the step of obtaining the stored candidate sentence vector comprises:

Obtain candidate questions stored in the question library;

Perform offline calculation on the candidate question based on the sentence vector generator to obtain a candidate sentence vector corresponding to the candidate question.