CN110674279A

CN110674279A - Question-answer processing method, device, equipment and storage medium based on artificial intelligence

Info

Publication number: CN110674279A
Application number: CN201910980232.0A
Authority: CN
Inventors: 冉邱; 林衍凯; 李鹏; 周杰
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2019-10-15
Filing date: 2019-10-15
Publication date: 2020-01-10

Abstract

The invention provides a question-answer processing method, a question-answer processing device, electronic equipment and a storage medium based on artificial intelligence; the method comprises the following steps: respectively coding the question information and the document information through a coding model in the question-answering processing model to obtain coding information corresponding to the question information and coding information corresponding to the document information; establishing a directed graph of the coding information of the question information and the coding information of the document information through a reasoning model in the question-answering processing model to obtain the directed graph comprising the question information and the document information; carrying out numerical reasoning processing on the directed graph comprising the problem information and the document information through a reasoning model to obtain document information containing numerical relationships; and carrying out prediction processing on the document information containing the numerical relation through a prediction model in the question-answering processing model to obtain answers corresponding to the question information. By the method and the device, accurate answers can be obtained according to the problems needing numerical reasoning.

Description

Question-answer processing method, device, equipment and storage medium based on artificial intelligence

Technical Field

The present invention relates to artificial intelligence natural language processing technology, and is especially artificial intelligence based question and answer processing method and apparatus, electronic device and storage medium.

Background

Natural Language Processing (NLP) is an important direction in the fields of computer science and artificial intelligence, and enables effective communication between people and computers using natural Language. Natural language processing is a science integrating linguistics, computer science and mathematics. Therefore, the field relates to natural language, namely the language used by people daily, so that the field is closely related to linguistics. Natural language processing techniques typically include text processing, semantic understanding, machine translation, robotic question and answer, knowledge mapping, and the like.

The question-answering processing system is one of important applications in the field of natural language processing, and is widely applied to a dialog system, a customer service system, intelligent hardware and the like, namely the question-answering processing system is a basic component of complex systems.

However, for questions requiring numerical reasoning, the current question-answering processing system cannot obtain accurate answers according to the questions.

Disclosure of Invention

The embodiment of the invention provides a question-answer processing method and device based on artificial intelligence, electronic equipment and a storage medium, which can obtain accurate answers according to questions needing numerical reasoning.

The technical scheme of the embodiment of the invention is realized as follows:

the embodiment of the invention provides a question-answer processing method based on artificial intelligence, which comprises the following steps:

respectively encoding question information and document information through an encoding model in a question-answering processing model to obtain encoding information corresponding to the question information and encoding information corresponding to the document information;

establishing a directed graph for the coding information of the question information and the coding information of the document information through a reasoning model in the question-answering processing model to obtain the directed graph comprising the question information and the document information;

performing numerical reasoning processing on the directed graph comprising the problem information and the document information through the reasoning model to obtain document information containing numerical relationships;

and carrying out prediction processing on the document information containing the numerical relation through a prediction model in the question-answering processing model to obtain an answer corresponding to the question information.

The embodiment of the invention provides a question-answering processing device based on artificial intelligence, which comprises:

the coding module is used for coding question information and document information respectively through a coding model in the question-answering processing model to obtain coding information corresponding to the question information and coding information corresponding to the document information;

the reasoning module is used for establishing a directed graph for the coding information of the question information and the coding information of the document information through a reasoning model in the question-answering processing model to obtain the directed graph comprising the question information and the document information;

and the prediction module is used for performing prediction processing on the document information containing the numerical relation through a prediction model in the question-answering processing model to obtain an answer corresponding to the question information.

In the above technical solution, the apparatus further includes:

the processing module is used for determining the weight of the keywords in the question information according to the keywords in the question information;

determining the relevancy between the keywords in the question information and each document to be recalled in a plurality of documents to be recalled;

based on the weight of the keyword, carrying out weighted summation on the relevancy between the keyword and each document to be recalled respectively to obtain the relevancy score of the question information and the document to be recalled;

and sequencing the plurality of documents to be recalled based on the relevance scores of the question information and each document to be recalled, and determining the document to be recalled with the highest relevance score as the document information corresponding to the question information.

In the above technical solution, the encoding module is further configured to perform first encoding on the problem information and the document information respectively through the encoding model to obtain intermediate encoding information corresponding to the problem information and intermediate encoding information corresponding to the document information;

and respectively carrying out secondary coding on the intermediate coding information corresponding to the problem information and the intermediate coding information corresponding to the document information to obtain the coding information corresponding to the problem information and the coding information corresponding to the document information.

In the above technical solution, the encoding module is further configured to map the problem information and the document information to a vector space through the encoding model, so as to obtain intermediate encoding information corresponding to the problem information and intermediate encoding information corresponding to the document information;

and coding the intermediate coding information corresponding to the question information and the intermediate coding information corresponding to the document information through an attention mechanism to obtain coding information of the question information containing the document information and coding information of the document information containing the question information.

In the above technical solution, the inference module is further configured to perform dimension transformation on the coding information of the problem information and the coding information of the document information through the inference model, respectively, to obtain transformed coding information of the problem information and transformed coding information of the document information;

performing numerical value extraction processing on the converted coding information of the question information and the converted coding information of the document information to obtain numerical values in the question information and the document information;

and establishing a directed graph according to the extracted numerical values to obtain the directed graph comprising the problem information and the document information.

In the above technical solution, the inference module is further configured to

Establishing a numerical value node corresponding to the numerical value according to the extracted numerical value;

determining the size relationship of adjacent numerical nodes based on the size of each numerical node;

determining the node type of the adjacent numerical node based on the node type of each numerical node;

establishing side information of the adjacent numerical value nodes according to the size relationship of the adjacent numerical value nodes and the node types of the adjacent numerical value nodes;

and establishing a directed graph comprising the problem information and the document information according to the numerical value nodes and the side information of the adjacent numerical value nodes.

In the above technical solution, the inference module is further configured to perform at least one numerical inference process on the directed graph including the problem information and the document information through the inference model to obtain an inferred numerical node;

and splicing the inferred numerical nodes and the document information to obtain document information containing numerical relationships.

In the above technical solution, the inference module is further configured to determine a weight of each numerical node in the directed graph according to a parameter learnable matrix and a deviation vector in the inference model;

performing the following processing for each of the numerical nodes:

determining neighbor numerical node information of the numerical node according to the side information of the numerical node and the neighbor numerical node in the directed graph and the weight of the neighbor numerical node, and

and fusing neighbor numerical node information of the numerical node and self information of the numerical node to obtain the inferred numerical node.

In the above technical solution, the inference module is further configured to obtain side information of the numerical node and a corresponding neighbor numerical node from the directed graph;

and multiplying the weight of the neighbor numerical node, the conversion matrix and the side information to obtain neighbor numerical node information of the numerical node.

and analyzing the side information to obtain the node type of the numerical node, the type of the neighbor numerical node and the size relation between the numerical node and the neighbor numerical node.

In the above technical solution, the inference module is further configured to multiply a parameter learnable matrix in the inference model by the information of the numerical node itself, and add the multiplied information of the numerical node itself to the information of the neighboring numerical node of the numerical node and the deviation vector in the inference model to obtain an updated numerical node;

and carrying out nonlinear transformation on the updated numerical value nodes through an activation function in the reasoning model to obtain the deduced numerical value nodes.

In the above technical solution, the prediction module is further configured to perform screening processing on the document information including the numerical relationship through the prediction model to obtain an answer type corresponding to the question information;

and performing prediction processing on the document information containing the numerical relation according to the answer type to obtain an answer corresponding to the question information.

In the above technical solution, the apparatus further includes:

the training module is used for screening the document information samples containing the numerical value relationship through the prediction model to obtain the answer type probability of the corresponding question information samples;

performing prediction processing on a document information sample containing numerical relation according to the answer type probability to obtain an answer probability corresponding to the question information sample;

constructing a loss function of the question-answer processing model according to the answer type probability and the answer probability;

and updating the parameters of the question-answering processing model until the loss function converges.

The embodiment of the invention provides question-answering processing equipment based on artificial intelligence, which comprises:

a memory for storing executable instructions;

and the processor is used for realizing the question-answering processing method based on artificial intelligence provided by the embodiment of the invention when the executable instructions stored in the memory are executed.

The embodiment of the invention provides a storage medium, which stores executable instructions and is used for causing a processor to execute so as to realize the question-answering processing method based on artificial intelligence provided by the embodiment of the invention.

The embodiment of the invention has the following beneficial effects:

the method comprises the steps of performing numerical reasoning on a directed graph to enable document information to contain numerical relationships, and performing prediction processing on the document information containing the numerical relationships to obtain accurate answers; by predicting the document information containing the numerical relation, answers can be extracted from the document information and displayed to the user, so that the situation that the original document with a longer length is directly presented to the user is avoided, and the user experience is improved.

Drawings

FIG. 1 is a schematic diagram of an application scenario of an artificial intelligence-based question-answering processing system 10 according to an embodiment of the present invention;

FIG. 2 is a schematic structural diagram of an artificial intelligence-based question answering processing device 500 according to an embodiment of the present invention;

FIG. 3 is a schematic structural diagram of an artificial intelligence-based question-answering processing device 555 according to an embodiment of the present invention;

FIGS. 4-8 are schematic flow charts of the artificial intelligence-based question-answering processing method provided by the embodiment of the invention;

fig. 9 is a schematic diagram of an overall structure of a question-answering processing model according to an embodiment of the present invention;

fig. 10 is a diagram illustrating that different NumGNN layer numbers have different numerical reasoning capabilities according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail with reference to the accompanying drawings, the described embodiments should not be construed as limiting the present invention, and all other embodiments obtained by a person of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

In the description that follows, references to the terms "first", "second", and the like, are intended only to distinguish similar objects and not to indicate a particular ordering for the objects, it being understood that "first", "second", and the like may be interchanged under certain circumstances or sequences of events to enable embodiments of the invention described herein to be practiced in other than the order illustrated or described herein.

Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to be limiting of the invention.

Before further detailed description of the embodiments of the present invention, terms and expressions mentioned in the embodiments of the present invention are explained, and the terms and expressions mentioned in the embodiments of the present invention are applied to the following explanations.

1) Recall (Recall): relevant documents are retrieved from a document repository.

2) Machine Reading Comprehension (MRC): given a document and a question, the machine predicts an answer to the question based on the document. For a pull machine-read understanding, the answer is typically a segment of a known document, and the MRC model predicts the answer by predicting the location of the answer in the known document for the start word and the stop word.

3) Directed graph: representing the relationship from item to item, a directed graph may be represented by ordered triples (v (D), a (D), ψ D), where ψ D is the correlation function, which is the ordered pair of elements for which each element in a (D) corresponds to v (D).

4) Graph Neural Network (GNN): a neural network acting directly on a graph structure mainly processes data of a non-Euclidean space structure (graph structure). Have an input order that ignores nodes; in the calculation process, the representation of the node is influenced by the neighbor nodes around the node, and the connection of the graph is unchanged; the representation of graph structure enables graph-based reasoning. In general, a graph neural network consists of two modules: the system comprises a propagation Module (propatiionmodule) and an Output Module (Output Module), wherein the propagation Module is used for transmitting information between nodes in the graph and updating the state, and the Output Module is used for defining an objective function according to different tasks based on vector representation of the nodes and edges of the graph. The graph neural network has: graph Convolutional Neural Networks (GCNs), Gated Graph Neural Networks (GGNNs), and Graph attention Neural Networks based on attention mechanism (GAT).

In order to solve at least the above technical problems of the related art, embodiments of the present invention provide a question-answering processing method and apparatus based on artificial intelligence, an electronic device, and a storage medium, which can obtain an accurate answer according to a problem requiring numerical reasoning. An exemplary application of the artificial intelligence based question-answering processing device provided by the embodiment of the present invention is described below, where the artificial intelligence based question-answering processing device provided by the embodiment of the present invention may be a server, for example, a server deployed in a cloud, and provides an accurate answer to a user according to questions and document information that need numerical reasoning and are provided by other devices or users, for example, the server obtains questions and document information that need numerical reasoning in a knowledge question-answering application (for example, a few age groups account for more than 7% of a total population) and a question that needs numerical reasoning in a weather inquiry application (for example, which city is a second-hot city today) and document information and the like according to other devices, establishes a directed graph according to the question information and the document information, and performs numerical reasoning on the directed graph to obtain document information including a numerical relationship, and carrying out prediction processing on the document information containing the numerical relationship to obtain an accurate answer (for example, the answer is '5' corresponding to a question needing numerical reasoning in knowledge question-answering application, and the answer is 'Wuhan City' corresponding to a question needing numerical reasoning in weather question-answering application); the system can also be various types of user terminals such as a notebook computer, a tablet computer, a desktop computer, a mobile device (e.g., a mobile phone, a personal digital assistant) and the like, for example, a handheld terminal, and an accurate answer is obtained according to a question which needs numerical reasoning and is input on the handheld terminal by a user and a document information image, and the question is displayed on a display interface of the handheld terminal.

Referring to fig. 1 by way of example, fig. 1 is a schematic view of an application scenario of an artificial intelligence-based question-answering processing system 10 according to an embodiment of the present invention, a terminal 200 is connected to a server 100 through a network 300, and the network 300 may be a wide area network or a local area network, or a combination of both.

The terminal 200 can be used to acquire questions and document information requiring numerical reasoning, for example, when a user inputs questions and document information requiring numerical reasoning through the input interface, the terminal automatically acquires the questions and document information requiring numerical reasoning and sends the questions and document information to the server after the input is completed.

In some embodiments, the terminal 200 locally executes the question-answer processing method based on artificial intelligence provided by the embodiments of the present invention to complete questions and document information that need numerical reasoning, and obtain accurate answers, for example, an Application (APP) such as a question-answer APP is installed on the terminal 200, a user inputs the questions and document information that need numerical reasoning in the question-answer APP, and the terminal 200 obtains accurate answers corresponding to the questions and document information according to the questions and document information that need numerical reasoning, and displays the accurate answers on the display interface 210 of the terminal 200.

In some embodiments, the terminal 200 may also send the questions and document information that need numerical reasoning and are input by the user on the terminal 100 to the server 100 through the network 300, and invoke the question-answer processing function based on artificial intelligence and provided by the server 100, the server 100 obtains accurate answers to the questions and document information through the question-answer processing method based on artificial intelligence and provided by the embodiments of the present invention, for example, a question-answer APP is installed on the terminal 200, the user inputs the questions and document information that need numerical reasoning and are input in the question-answer APP, the terminal sends the questions and document information that need numerical reasoning to the server 100 through the network 300, after receiving the questions and document information that need numerical reasoning and are input by the user, the server 100 obtains accurate answers to the questions and document information according to the questions and document information that need numerical reasoning and returns the answers to the question-answer APP, the answer is displayed on the display interface 210 of the terminal 200, or the server 100 directly answers.

Continuing to describe the structure of the artificial intelligence based question-answering processing device provided by the embodiment of the present invention, the artificial intelligence based question-answering processing device may be various terminals, such as a mobile phone, a computer, etc., or may be the server 100 shown in fig. 1.

Referring to fig. 2, fig. 2 is a schematic structural diagram of an artificial intelligence based question-answering processing apparatus 500 according to an embodiment of the present invention, and the artificial intelligence based question-answering processing apparatus 500 shown in fig. 2 includes: at least one processor 510, memory 550, at least one network interface 520, and a user interface 530. The various components in artificial intelligence based image translation device 500 are coupled together by a bus system 540. It is understood that the bus system 540 is used to enable communications among the components. The bus system 540 includes a power bus, a control bus, and a status signal bus in addition to a data bus. For clarity of illustration, however, the various buses are labeled as bus system 540 in fig. 2.

The Processor 510 may be an integrated circuit chip having Signal processing capabilities, such as a general purpose Processor, a Digital Signal Processor (DSP), or other programmable logic device, discrete gate or transistor logic device, discrete hardware components, or the like, wherein the general purpose Processor may be a microprocessor or any conventional Processor, or the like.

The user interface 530 includes one or more output devices 531 enabling presentation of media content, including one or more speakers and/or one or more visual display screens. The user interface 530 also includes one or more input devices 532, including user interface components to facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

The memory 550 may comprise volatile memory or nonvolatile memory, and may also comprise both volatile and nonvolatile memory. The non-volatile Memory may be a Read Only Memory (ROM), and the volatile Memory may be a Random Access Memory (RAM). The memory 550 described in connection with embodiments of the invention is intended to comprise any suitable type of memory. Memory 550 optionally includes one or more storage devices physically located remote from processor 510.

In some embodiments, memory 550 can store data to support various operations, examples of which include programs, modules, and data structures, or subsets or supersets thereof, as exemplified below.

An operating system 551 including system programs for processing various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks;

a network communication module 552 for communicating to other computing devices via one or more (wired or wireless) network interfaces 520, exemplary network interfaces 520 including: bluetooth, wireless compatibility authentication (WiFi), and Universal Serial Bus (USB), etc.;

a display module 553 for enabling presentation of information (e.g., a user interface for operating peripherals and displaying content and information) via one or more output devices 531 (e.g., a display screen, speakers, etc.) associated with the user interface 530;

an input processing module 554 to detect one or more user inputs or interactions from one of the one or more input devices 532 and to translate the detected inputs or interactions.

In some embodiments, the artificial intelligence based question-answering processing apparatus provided by the embodiments of the present invention may be implemented by combining hardware and software, and by way of example, the artificial intelligence based question-answering processing apparatus provided by the embodiments of the present invention may be a processor In the form of a hardware decoding processor, which is programmed to execute the artificial intelligence based question-answering processing method provided by the embodiments of the present invention, for example, the processor In the form of the hardware decoding processor may employ one or more Application Specific Integrated Circuits (ASICs), DSPs, Programmable Logic Devices (PLDs), Complex Programmable Logic Devices (CPLDs), Field Programmable Gate Arrays (FPGAs), or other electronic elements.

In other embodiments, the artificial intelligence based question-answering processing apparatus provided by the embodiment of the present invention may be implemented in software, and fig. 2 shows an artificial intelligence based question-answering processing apparatus 555 stored in a memory 550, which may be software in the form of programs and plug-ins, and includes a series of modules including an encoding module 5551, an inference module 5552, a prediction module 5553, a processing module 5554, and a training module 5555; the encoding module 5551, the reasoning module 5552, the prediction module 5553, the processing module 5554, and the training module 5555 are used to implement the artificial intelligence-based question-answering processing method provided by the embodiment of the invention.

As can be understood from the foregoing, the question-answering processing method based on artificial intelligence provided in the embodiments of the present invention may be implemented by various types of question-answering processing devices based on artificial intelligence, such as an intelligent terminal and a server.

The question-answer processing method based on artificial intelligence provided by the embodiment of the invention is described below by combining with the exemplary application and implementation of the server provided by the embodiment of the invention. Referring to fig. 3 and fig. 4, fig. 3 is a schematic structural diagram of an artificial intelligence-based question-answering processing apparatus 555 according to an embodiment of the present invention, which shows a question-answering processing flow, and fig. 4 is a schematic flow diagram of an artificial intelligence-based question-answering processing method according to an embodiment of the present invention, which is described with reference to the steps shown in fig. 4.

In step 101, question information and document information are encoded by the encoding model in the question-answering processing model, and encoding information corresponding to the question information and encoding information corresponding to the document information are obtained.

The problem information and the document information can be respectively coded after the problem information and the document information are received by the server, and coding information corresponding to the problem information and coding information corresponding to the document information are obtained, so that a directed graph can be established in the following process.

In some embodiments, before encoding the question information and the document information respectively, a step 105 of obtaining the question information and the document information is further included. Referring to fig. 5, fig. 5 is an optional flowchart provided in an embodiment of the present invention, and in some embodiments, fig. 5 illustrates that step 105 may be implemented by step 1051 to step 1054 illustrated in fig. 5.

In step 1051, the weights of the keywords in the question information are determined based on the keywords in the question information.

The user can input the problem information on an input interface of the terminal, after the input is completed, the terminal can forward the problem information to the server, and after the server receives the problem information, the server can perform word segmentation processing on the problem information to obtain keywords in the problem information, so that the weight of the keywords in the problem information is determined. The server can search the documents to be recalled in the database according to the keywords of the problem information, count the number of the documents containing the keywords in the documents to be recalled, and calculate the ratio of the number of the documents containing the keywords in the documents to be recalled to the total number of the documents to be recalled to obtain the weight of the keywords in the problem information. When the frequency of the keywords in the question information appearing in the document to be recalled is higher, the more important the keywords indicating the question information are, that is, the more important the weight of the keywords of the question information is.

In step 1052, a degree of relevance between the keyword in the question information and each of the plurality of documents to be recalled is determined.

After obtaining the keywords in the question information, the server may determine the relevancy between the keywords and each document to be recalled in the plurality of documents to be recalled according to the keywords in the question information and the documents to be recalled. The server can determine a first ratio between the length of the document to be recalled and the average length of the document, determine the first ratio as the relative length of the document to be recalled, determine the frequency of the keywords in the question information appearing in the document to be recalled and a second ratio between the relative length of the document to be recalled, and generate the correlation between the keywords and the text to be recalled according to the mapping relationship between the second ratio and the correlation and the second ratio.

In step 1053, based on the weights of the keywords, the relevancy between the keywords and each document to be recalled is weighted and summed to obtain the relevancy scores between the question information and the documents to be recalled.

After the server obtains the weight of the key words and the relevancy between the key words and the document to be recalled, the weight of the key words and the relevancy between the key words and the document to be recalled can be weighted and summed to obtain the relevancy score between the question information and the document to be recalled.

In step 1054, based on the relevance scores of the question information and each of the documents to be recalled, the plurality of documents to be recalled are ranked, and the document to be recalled with the highest relevance score is determined as the document information corresponding to the question information.

After the server obtains the relevance scores of the question information and the documents to be recalled, the documents to be recalled are sorted in a descending order based on the relevance scores of the question information and the documents to be recalled to obtain the documents to be recalled with the highest relevance scores, and the documents to be recalled with the highest relevance scores are determined as the document information corresponding to the question information, so that the document information is correspondingly coded in the following.

In some embodiments, the step of encoding the question information and the document information respectively by using the encoding model in the question-answering processing model to obtain the encoding information corresponding to the question information and the encoding information corresponding to the document information includes: respectively carrying out first coding on the problem information and the document information through a coding model to obtain intermediate coding information corresponding to the problem information and intermediate coding information corresponding to the document information; and respectively carrying out secondary encoding on the intermediate encoding information corresponding to the problem information and the intermediate encoding information corresponding to the document information to obtain the encoding information corresponding to the problem information and the encoding information corresponding to the document information.

In order to obtain accurate encoded information corresponding to the question information and encoded information corresponding to the document information, the question information and the document information may be first encoded by an encoding model to obtain intermediate encoded information corresponding to the question information and intermediate encoded information corresponding to the document information, and then the intermediate encoded information corresponding to the question information and the intermediate encoded information corresponding to the document information may be second encoded to obtain accurate encoded information corresponding to the question information and encoded information corresponding to the document information.

Referring to fig. 6, fig. 6 is an optional flowchart provided in an embodiment of the present invention, and in some embodiments, fig. 6 shows that step 101 may be implemented by steps 1011 to 1012 shown in fig. 6.

In step 1011, the question information and the document information are mapped to the vector space by the coding model, and intermediate coding information corresponding to the question information and intermediate coding information corresponding to the document information are obtained.

In order to facilitate the question answering processing on the question information and the document information, the question information and the document information can be mapped to a vector space through a coding model to obtain intermediate coding information corresponding to the question information and intermediate coding information corresponding to the document information.

In step 1012, the intermediate code information corresponding to the question information and the intermediate code information corresponding to the document information are encoded by the attention mechanism, and the code information of the question information including the document information and the code information of the document information including the question information are obtained.

In order to fuse the question information and the document information to a certain extent, the intermediate code information corresponding to the question information and the intermediate code information corresponding to the document information may be encoded by an attention mechanism (attention layer) to obtain the code information of the question information including the document information and the code information of the document information including the question information.

As an example, referring to fig. 3, by searching a database according to question information in the processing module 5554 of the artificial intelligence-based question-answering processing apparatus 555, document information corresponding to the question information is obtained, and the document information and the question information are input to the encoding module 5551, and the encoding model in the encoding module 5551 obtains encoding information of the question information including the document information and encoding information of the document information including the question information by performing encoding operations twice on the document information and the question information, respectively.

In step 102, a directed graph including the question information and the document information is obtained by performing directed graph establishment processing on the coding information of the question information and the coding information of the document information through an inference model in the question-answering processing model.

For the questions needing numerical reasoning, a directed graph can be established for question information and document information through a reasoning model in a question-answering processing model, so that numerical reasoning can be carried out subsequently according to the directed graph, and therefore document information containing numerical relationships can be obtained, and effective answers for the questions needing numerical reasoning can be obtained according to the document information containing the numerical relationships.

In some embodiments, the step of performing directed graph building processing on the coding information of the question information and the coding information of the document information through an inference model in the question-and-answer processing model to obtain a directed graph including the question information and the document information includes: performing dimensionality transformation on the coding information of the problem information and the coding information of the document information through an inference model to obtain the transformed coding information of the problem information and the transformed coding information of the document information; performing numerical value extraction processing on the coded information of the converted question information and the coded information of the converted document information to obtain numerical values in the question information and the document information; and establishing a directed graph according to the extracted numerical values to obtain the directed graph comprising the problem information and the document information.

The set-up of the directed graph cannot be performed because the dimension of the encoded information of the problem information and the encoded information of the document information may not be suitable for the inference module. Therefore, the dimension transformation can be respectively carried out on the coding information of the problem information and the coding information of the document information, so that the dimension of the transformed coding information of the problem information and the transformed coding information of the document information is suitable for the inference module, and the nonlinear transformation can be carried out on the transformed coding information of the problem information and the transformed coding information of the document information, so as to fit the coding information of the problem information and the coding information of the document information. Extracting the numerical values in the transformed problem information and the transformed document information, and establishing a directed graph according to the extracted numerical values so as to carry out numerical reasoning according to the directed graph subsequently.

In some embodiments, establishing a directed graph according to the extracted numerical values, and obtaining the directed graph including the question information and the document information, includes: establishing a numerical node corresponding to the numerical value according to the extracted numerical value; determining the size relationship of adjacent numerical nodes based on the size of each numerical node; determining the node type of the adjacent numerical node based on the node type of each numerical node; establishing side information of adjacent numerical value nodes according to the size relationship of the adjacent numerical value nodes and the node types of the adjacent numerical value nodes; and establishing a directed graph comprising problem information and document information according to the numerical value nodes and the side information of the adjacent numerical value nodes.

For each extracted value, a corresponding value node is established, for example, a value 2 and a value 3, and a value node corresponding to 2 and 3 is established, respectively. And determining the magnitude relation of adjacent numerical nodes according to the magnitude of each numerical node, for example, if the numerical node 1 is adjacent to the numerical node 2, and the numerical value of the numerical node 1 is greater than the numerical value of the numerical node 2, it is determined that the numerical node 1 points to the numerical node 2. And determining the node type of the adjacent numerical nodes based on the node type of each numerical node, wherein for example, if the numerical node 1 is adjacent to the numerical node 2, the numerical node 1 belongs to the problem information, and the numerical node 2 belongs to the document information, the node types of the numerical node and the numerical node 2 are determined to belong to (q-p). After the size relationship of the adjacent numerical value nodes and the node types of the adjacent numerical value nodes are determined, the size relationship of the adjacent numerical value nodes and the node types of the adjacent numerical value nodes are determined as the side information of the adjacent numerical value nodes. And constructing a directed graph comprising the problem information and the document information according to each numerical value node and the side information.

As an example, referring to fig. 3, the inference model in the inference module 5552 may perform a process of creating a directed graph on the encoded information of the question information and the encoded information of the document information, so as to obtain a directed graph including the question information and the document information.

In step 103, a numerical reasoning process is performed on the directed graph including the question information and the document information through a reasoning model, so as to obtain document information including a numerical relationship.

In order to incorporate a numerical relationship with the question information in the document information, numerical reasoning can be performed on the directed graph including the question information and the document information by using a reasoning model.

Referring to fig. 7, fig. 7 is an optional flowchart diagram provided in an embodiment of the present invention, and in some embodiments, fig. 7 shows that step 103 may be implemented by step 1031 to step 1032 shown in fig. 7.

In step 1031, the directed graph including the question information and the document information is subjected to at least one numerical inference process through an inference model, so as to obtain an inferred numerical node.

Because numerical reasoning is carried out once, the relationship between adjacent numerical nodes can be deduced, and the problem of judgment is solved according to the adjacent numerical nodes. However, for the relationship between a plurality of numerical nodes, a plurality of numerical inferences need to be performed, so that the numerical inference can be performed at least once on the directed graph including the problem information and the document information through the inference model to obtain the inferred numerical nodes.

In some embodiments, performing at least one numerical inference process on a directed graph including question information and document information through an inference model to obtain an inferred numerical node includes: determining the weight of each numerical value node in the directed graph according to the parameter learnable matrix and the deviation vector in the reasoning model; performing the following processing for each of the numerical nodes:

and determining neighbor numerical node information of the numerical node according to the side information of the numerical node and the neighbor numerical node in the directed graph and the weight of the neighbor numerical node, and fusing the neighbor numerical node information of the numerical node and the self information of the numerical node to obtain the inferred numerical node.

It should be noted that, performing numerical reasoning once on the directed graph includes three steps, which are respectively: 1) determining the weight of the numerical node; 2) determining neighbor numerical node information of the numerical node; 3) and fusing the information of the numerical node.

In some embodiments, determining neighbor numerical node information of the numerical node according to the edge information of the numerical node and neighbor numerical nodes in the directed graph and the weights of the neighbor numerical nodes includes: acquiring side information of the numerical node and the corresponding neighbor numerical node from the directed graph; and multiplying the weight of the neighbor numerical node, the conversion matrix and the side information to obtain the neighbor numerical node information of the numerical node.

The numerical node is not only related to the numerical node per se but also related to the neighbor numerical node in the reasoning process, so that numerical reasoning can be carried out through the neighbor numerical node. The neighbor numerical node information of the numerical node is obtained by multiplying the weight of the neighbor numerical node, the conversion matrix and the side information, wherein the conversion matrix is a matrix with a specific relationship and can be obtained by training.

In some embodiments, obtaining side information of the numerical node and the corresponding neighbor numerical node from the directed graph includes: acquiring side information of the numerical node and the corresponding neighbor numerical node from the directed graph; and analyzing the side information to obtain the node type of the numerical node, the type of the neighbor numerical node and the size relation between the numerical node and the neighbor numerical node.

It should be noted that the side information includes a node type of the numerical node, a type of the neighbor numerical node, and a size relationship between the numerical node and the neighbor numerical node, and therefore, the side information acquired from the directed graph includes the node type of the numerical node, the type of the neighbor numerical node, and the size relationship between the numerical node and the neighbor numerical node.

In some embodiments, fusing neighbor numerical node information of the numerical node and self information of the numerical node to obtain the inferred numerical node, including: multiplying the parameter learnable matrix in the inference model by the self information of the numerical node, and adding the multiplied self information of the numerical node with the neighbor numerical node information of the numerical node and the deviation vector in the inference model to obtain an updated numerical node; and carrying out nonlinear transformation on the updated numerical value nodes through an activation function in the reasoning model to obtain the inferred numerical value nodes.

In order to integrate the self information of the numerical node into the neighbor numerical node information of the numerical node, the self information of the numerical node is multiplied by a parameter learnable matrix in the reasoning model, and the multiplied self information of the numerical node is added with the neighbor numerical node information of the numerical node and a deviation vector in the reasoning model, so that the updated neighbor numerical node information and the updated self information of the numerical node are integrated. Wherein, the parameter learnable matrix and the deviation vector can be obtained by training.

In step 1032, the inferred numerical nodes and the document information are spliced to obtain document information containing numerical relationships.

Because the inferred numerical nodes may only contain numerical relationships, in order to contain non-numerical text segments, the inferred numerical nodes and the document information may be spliced to obtain the document information containing the numerical relationships.

As an example, referring to fig. 3, a directed graph including question information and document information may be numerically inferred by an inference model in the inference module 5552, to obtain document information including a numerical relationship, and the document information including the numerical relationship is passed to the prediction model 5553.

In step 104, the document information containing the numerical relationship is subjected to prediction processing through the prediction model in the question-answering processing model, so as to obtain the answer corresponding to the question information.

After the server obtains the document information containing the numerical relationship, the document information containing the numerical relationship can be subjected to prediction processing through a prediction model in the question-answer processing model, so that answers corresponding to the question information are obtained, and the problem of numerical reasoning is solved.

In some embodiments, the obtaining the answer corresponding to the question information by performing prediction processing on the document information containing the numerical relationship through a prediction model in the question-answer processing model includes: screening the document information containing the numerical relation through a prediction model to obtain an answer type corresponding to the question information; and carrying out prediction processing on the document information containing the numerical relation according to the answer type to obtain the answer corresponding to the question information.

The answer types corresponding to the question information are 4 types at present, which are respectively: 1) the answer of the document fragment class is the fragment in the document, and the probability is the product of the probabilities of the positions of the starting word and the ending word of the fragment; 2) the answer of question segment is the segment in the question, and the probability is the product of the probabilities of the initial word and the final word of the segment; 3) counting answers are obtained through counting, and the forecasting model models the answers into multi-classification questions with the numerical values of 0-9; 4) the arithmetic expression is constructed by extracting all numbers in a document, assigning a positive value or a negative value or 0 to each number, and adding the numbers to obtain the answer. In order to obtain an accurate answer more quickly, the document information containing the numerical relationship may be first screened to obtain an answer type corresponding to the question information. And then, carrying out prediction processing on the document information containing the numerical relation according to the answer type to obtain the answer corresponding to the question information.

For example, referring to fig. 3, the document information including the numerical relationship transmitted by the inference model may be predicted by the prediction model in the prediction module 5553, so as to obtain the answer corresponding to the question information.

In some embodiments, a description is given to training of a question-and-answer processing model, referring to fig. 8, based on fig. 4, fig. 8 is a schematic flow chart of a question-and-answer processing method based on artificial intelligence provided in an embodiment of the present invention, and in step 106, a document information sample containing a numerical relationship is subjected to a screening process by a prediction model to obtain an answer type probability of a corresponding question information sample; performing prediction processing on the document information samples containing the numerical relationship according to the answer type probability to obtain the answer probability of the corresponding question information samples; constructing a loss function of the question-answer processing model according to the answer type probability and the answer probability; and updating parameters of the question-answering processing model until the loss function converges.

Step 106 does not have an obvious sequence from step 101 to step 104. When the server determines the value of the loss function of the question-answering processing model based on the probability that the word belongs to the answer type and the answer probability, whether the value of the loss function exceeds a preset threshold value or not can be judged, when the value of the loss function exceeds the preset threshold value, an error signal of the question-answering processing model is determined based on the loss function, error information is reversely propagated in the question-answering processing model, and model parameters of all layers are updated in the propagation process.

Describing backward propagation, inputting training sample data into an input layer of a neural network model, passing through a hidden layer, finally reaching an output layer and outputting a result, which is a forward propagation process of the neural network model, wherein because the output result of the neural network model has an error with an actual result, an error between the output result and the actual value is calculated and is propagated backward from the output layer to the hidden layer until the error is propagated to the input layer, and in the process of backward propagation, the value of a model parameter is adjusted according to the error; and continuously iterating the process until convergence, wherein the question-answering processing model belongs to the neural network model.

As an example, referring to fig. 3, the question-answering processing model may be trained based on a loss function of the question-answering processing model by a prediction model in the training module 5553 in the artificial intelligence based question-answering processing apparatus 555.

So far, the question-answering processing method based on artificial intelligence provided by the embodiment of the present invention has been described in conjunction with the exemplary application and implementation of the server provided by the embodiment of the present invention, and the following continues to describe a scheme for implementing question-answering processing based on artificial intelligence in cooperation with each module in the question-answering processing device 555 based on artificial intelligence provided by the embodiment of the present invention.

The encoding module 5551 is configured to encode question information and document information respectively through an encoding model in a question-answering processing model to obtain encoding information corresponding to the question information and encoding information corresponding to the document information;

the reasoning module 5552 is configured to perform, through a reasoning model in the question-answering processing model, establishment processing of a directed graph on the coding information of the question information and the coding information of the document information to obtain a directed graph including the question information and the document information;

the prediction module 5553 is configured to perform prediction processing on the document information including the numerical relationship through a prediction model in the question-and-answer processing model to obtain an answer corresponding to the question information.

In the above technical solution, the question-answering processing device 555 based on artificial intelligence further includes:

the processing module 5554 is configured to determine a weight of a keyword in the question information according to the keyword in the question information; determining the relevancy between the keywords in the question information and each document to be recalled in a plurality of documents to be recalled; based on the weight of the keyword, carrying out weighted summation on the relevancy between the keyword and each document to be recalled respectively to obtain the relevancy score of the question information and the document to be recalled; and sequencing the plurality of documents to be recalled based on the relevance scores of the question information and each document to be recalled, and determining the document to be recalled with the highest relevance score as the document information corresponding to the question information.

In the above technical solution, the encoding module 5551 is further configured to perform first encoding on the problem information and the document information respectively through the encoding model to obtain intermediate encoding information corresponding to the problem information and intermediate encoding information corresponding to the document information; and respectively carrying out secondary coding on the intermediate coding information corresponding to the problem information and the intermediate coding information corresponding to the document information to obtain the coding information corresponding to the problem information and the coding information corresponding to the document information.

In the above technical solution, the encoding module 5551 is further configured to map the problem information and the document information to a vector space through the encoding model, so as to obtain intermediate encoding information corresponding to the problem information and intermediate encoding information corresponding to the document information; and coding the intermediate coding information corresponding to the question information and the intermediate coding information corresponding to the document information through an attention mechanism to obtain coding information of the question information containing the document information and coding information of the document information containing the question information.

In the above technical solution, the inference module 5552 is further configured to perform dimension transformation on the coding information of the question information and the coding information of the document information through the inference model, respectively, to obtain transformed coding information of the question information and transformed coding information of the document information; performing numerical value extraction processing on the converted coding information of the question information and the converted coding information of the document information to obtain numerical values in the question information and the document information; and establishing a directed graph according to the extracted numerical values to obtain the directed graph comprising the problem information and the document information.

In the above technical solution, the inference module 5552 is further configured to establish a numerical node corresponding to the numerical value according to the extracted numerical value; determining the size relationship of adjacent numerical nodes based on the size of each numerical node; determining the node type of the adjacent numerical node based on the node type of each numerical node; establishing side information of the adjacent numerical value nodes according to the size relationship of the adjacent numerical value nodes and the node types of the adjacent numerical value nodes; and establishing a directed graph comprising the problem information and the document information according to the numerical value nodes and the side information of the adjacent numerical value nodes.

In the above technical solution, the inference module 5552 is further configured to perform at least one numerical inference process on the directed graph including the problem information and the document information through the inference model to obtain an inferred numerical node; and splicing the inferred numerical nodes and the document information to obtain document information containing numerical relationships.

In the above technical solution, the inference module 5552 is further configured to determine a weight of each numerical node in the directed graph according to a parameter learnable matrix and a deviation vector in the inference model;

performing the following processing for each of the numerical nodes:

and determining neighbor numerical node information of the numerical node according to the side information of the numerical node and neighbor numerical nodes in the directed graph and the weight of the neighbor numerical node, and fusing the neighbor numerical node information of the numerical node and the self information of the numerical node to obtain the inferred numerical node.

In the above technical solution, the inference module 5552 is further configured to obtain side information of the numerical node and a corresponding neighbor numerical node from the directed graph; and multiplying the weight of the neighbor numerical node, the conversion matrix and the side information to obtain neighbor numerical node information of the numerical node.

In the above technical solution, the inference module 5552 is further configured to obtain side information of the numerical node and a corresponding neighbor numerical node from the directed graph; and analyzing the side information to obtain the node type of the numerical node, the type of the neighbor numerical node and the size relation between the numerical node and the neighbor numerical node.

In the above technical solution, the inference module 5552 is further configured to multiply the parameter learnable matrix in the inference model by the self information of the numerical node, and add the multiplied self information of the numerical node, the neighbor numerical node information of the numerical node, and the deviation vector in the inference model to obtain an updated numerical node; and carrying out nonlinear transformation on the updated numerical value nodes through an activation function in the reasoning model to obtain the deduced numerical value nodes.

In the above technical solution, the prediction module 5553 is further configured to perform screening processing on the document information including the numerical relationship through the prediction model to obtain an answer type corresponding to the question information; and performing prediction processing on the document information containing the numerical relation according to the answer type to obtain an answer corresponding to the question information.

the training module 5555 is configured to perform screening processing on the document information samples containing numerical relationships through the prediction model to obtain answer type probabilities of the corresponding question information samples; performing prediction processing on a document information sample containing numerical relation according to the answer type probability to obtain an answer probability corresponding to the question information sample; constructing a loss function of the question-answer processing model according to the answer type probability and the answer probability; and updating the parameters of the question-answering processing model until the loss function converges.

In the following, an exemplary application of the embodiments of the present invention in a practical application scenario will be described.

For a pull-type machine-readable understanding, the answer is typically a snippet in a known document. The QANT model is one of models with good effect on the extraction type MRC, the QANT model uses an embedded representation layer and a multilayer encoder layer to map documents and questions into a vector space, then the vector representations of the documents and the questions are interacted, namely, an attention layer obtains the document representation with known questions, then the multilayer modeling encoder layer is used for processing the document representation with known questions, finally, a prediction module is used for predicting the probability of the positions of starting words and ending words, and the product of the probabilities of the starting words and the ending words is used as the probability of corresponding document fragments.

The NAQANet model is a machine reading understanding model based on the QANet model, and can classify answers into 4 types: 1) document snippet, 2) question snippet, 3) count, 4) arithmetic expression. The N AQANet model can predict the 4 types of answers with separate network structures, respectively, to obtain corresponding answers. The question-answering processing model of the embodiment of the invention is a machine reading understanding model based on the NAQANT model, and can further improve the numerical reasoning capability of comparison and sequencing types.

When the machine reads and understands, when financial news, scientific articles and the like are encountered, numerical reasoning, such as addition, subtraction, sequencing and counting, is generally required, and the numerical reasoning capability of the existing machine reading and understanding model is weak. The NAQANet model can divide answer types into 1) document fragments, 2) question fragments, 3) counting and 4) arithmetic expressions aiming at the problems with weak numerical reasoning capability, and respectively predicts the answer types, so that counting and addition and subtraction operation numerical reasoning can be performed to a certain degree. However, the NAQANet model does not explicitly perform numerical reasoning of comparison or sort type, and thus the processing effect is poor for the problem requiring the reasoning. The question-answering processing model in the embodiment of the invention can construct a directed graph according to the magnitude relation among numerical values in a machine-read text on the basis of an NAQANT model, and adopts a graph neural network to carry out multi-step reasoning on the directed graph, so that the question-answering processing model can answer the questions needing the reasoning, such as 1) numerical value comparison or sorting type questions, 2) numerical value condition type questions (numbers in the questions are used as value regions or threshold values, and the like). The embodiment of the present invention illustrates an example in which numerical comparison is required, and the answer can be inferred from the relevant part in the document, as shown in table 1, for the second question in table 1, it is necessary to know which age group accounts for more than 7% of the total population to count:

TABLE 1

The question-answering processing method based on artificial intelligence provided by the embodiment of the invention is used as a machine reading understanding technology, and can be applied to a general application scene of MRC. MRC techniques may be applied in the following application scenarios:

1) the method is applied to a search engine, for example, when a search request of a user is a specific question and the search engine retrieves the most relevant article, the MRC technology can be used to extract the answer from the article and display the answer to the user instead of directly presenting a long original text to the user, so as to improve the user experience.

2) When the method is applied to the chat robot, for example, when a user communicates with the chat robot, the user may ask some questions of the chat robot, which are generally various and large in number, and if answers are prepared in advance, it is difficult to cover all possible questions of the user. In this case, when a question of the user is received, the most relevant document may be retrieved first, and then the MRC technique may be used to extract the corresponding answer from the document.

However, in the above application scenario, if a question needs to be compared or ranked for reasoning to some extent, the existing MRC technology is prone to give wrong answers, and the accuracy of the answers to the question can be improved by using the artificial intelligence-based question-answering processing method according to the embodiment of the present invention.

The question-answering processing model in the embodiment of the invention mainly comprises (1) a coding model, (2) an inference model and (3) a prediction model, the overall structure is shown as fig. 9, and fig. 9 is an overall structure schematic diagram of the question-answering processing model provided by the embodiment of the invention.

(1) Coding model

The coding model uses a QANT model or a coding component in an NAQANT model to code documents and questions into a vector space, and the calculation formula is shown as formulas (1) and (2):

Q＝QANet-Emb-Enc(q) (1)

P＝QANet-Emb-Enc(p) (2)

wherein Q represents question information, P represents document information, QANet-Emb-Enc () represents an embedded representation layer and a multi-layer encoder layer (stacked encoder blocks) in a QANet model, Q represents information in which the question information is encoded in a vector space, and P represents information in which the document information is encoded in the vector space.

After the document and the question are coded into the vector space, the known question representation of the document and the known question representation of the document are calculated, and the calculation formulas are shown as formulas (3) and (4):

wherein QANT-Att () represents the attention layer (context-request entry) in the QANT model,

a question representation (question information containing document information) indicating that the document is known,

a document representation (document information containing question information) indicating that the question is known.

(2) Inference model

Inputting the document known question representation output by the coding model and the document representation with known question into the reasoning model, and according to the document known question representation and the document with known questionFile representation building directed graph

The node V in the directed graph is a number in the document information and the problem information, the edge E represents the size relationship between the corresponding two numbers, and then the GNN is used for reasoning the directed graph, and the calculation formulas are shown as formulas (5), (6) and (7):

wherein, W^MFor parameter learnable matrix, U is node representation (value node after inference) corresponding to digit after inference, QANT-Mod-Enc is modeling encoder layer (stacked model encoder blocks) in QANT model, and pairOrFitting is performed by performing a nonlinear transformation. U only contains the representation corresponding to the number, and in order to process the answer corresponding to the document fragment type containing the non-number words, U can be spliced to M^PTo obtain a document expression M0 with a known numerical relationship, the calculation formula is shown in formulas (8), (9), (10):

M₀＝QANet-Mod-Enc(M′₀′) (10)

wherein, 0 represents a zero vector and,

representing the ith word in the document, I (i) representing the index corresponding to the number in the document,

column i, [ ·; a]Representing a matrix splice, W₀For parameters to learn matrices, b₀Is a deviation vector.

(3) Prediction model

The predictive model divides the answers into 4 types and each uses a separate network layer to calculate the probability of the answer under different types. Wherein, 4 answer types are as follows:

1) answer to document fragment class: the answer is a segment in the document, and the probability is the product of the probabilities of the positions of the starting word and the ending word of the segment;

2) answer to question fragment class: the answer is a segment in the question, and the probability is the product of the probabilities of the positions of the starting word and the ending word of the segment;

3) counting answers: the answers are obtained by counting, and the prediction model models the answers into a multi-classification problem with the numerical value of 0-9;

4) arithmetic expression class answer: the answer is the calculation result of the arithmetic expression, and the arithmetic expression construction method is that all numbers in the document are extracted firstly, then positive and negative values or 0 are given to each number, and finally the numbers are added to obtain the answer.

Wherein, the probability of the answer type is calculated by adopting an additional network layer. When training the question-answering processing model, the final answer probability is the joint probability under all feasible types, and the calculation formula is shown as formula (11):

∑_typePr(type)Pr(answer|type) (11)

where Pr (type) represents the probability of the answer type, and Pr (answer | type) represents the probability of the answer under the known answer type.

When the predicted answer is made, the prediction model firstly determines the most possible answer type, and then predicts the most possible answer under the answer of the type.

The construction of the directed graph is further illustrated below:

the embodiment of the invention takes all the numbers appearing in the documents and the problems as the nodes in the directed graph. V for node set of numbers in corresponding document^PV for representing and problem number node assembly^QIf so, all nodes are represented as V ═ V^Q∪V^PThe number corresponding to node v is denoted as n (v). The embodiment of the invention considers two sides:

1) greater than edge

For two nodes v_i,v_jE.g. V, if n (V)_i)＞n(v_j) Then from v_iDirection v_jDirected edge of

A directed graph is added (as shown by the solid arrows in the block diagram of fig. 9).

2) Less than or equal to the edge

For two nodes v_i,v_jE.g. V, if n (V)_i)≤n(v_j) Then there is an edge

A directed graph is added (as shown by the dashed arrows in the block diagram of fig. 9).

Wherein,

and

are complementary. However, a number may appear multiple times in a document and generally represents a different fact, so embodiments of the present invention add a node to each occurrence of the same numberPotential confusion is avoided and both edges are used to encode equality information between nodes.

Numerical reasoning is further illustrated below:

the inference model in the embodiment of the invention can use a graph neural network to carry out inference on the constructed directed graph.

To V^PEach node in the node representation is represented by M^PThe corresponding vector is initialized, and the calculation formula is shown as the formula (12):

wherein,

to represent

An index into a document. In the same way, for V^QWhere each node representation may be represented by M^QThe corresponding vector in (1) is initialized.

The inference model can perform at least one step of numerical inference. Wherein, the numerical reasoning in one step comprises the following steps:

1) and calculating the node relevance. Only a portion of the numbers in the documents and questions are relevant to answer the questions, and the inference model calculates a weight for each numerical node to reduce the influence of the irrelevant numerical nodes, the node weights being calculated as shown in equation (13):

α_i＝sigmiod(W_vv[i]+b_v) (13)

wherein, W_vFor parameters to learn matrices, b_vIs a deviation vector, α_iIs the degree of correlation of the numerical node, wherein sigmoid is an activation function.

2) And (5) information is spread. Each numerical node plays a role in the inference process not only determined by itself but also related to the neighbor numerical nodes of the numerical node, so that the inference model propagates information from each numerical node to all its neighbor numerical nodes to make inferences. Since the roles of the numbers in the documents and questions may be different, the inference model distinguishes different edges, and the calculation formula of the information representation of the numerical nodes is shown as formula (14):

wherein,representing a numerical node v_iInformation of (r)_jiIndicating an assigned edge e_jiIn the context of (a) or (b),

a transformation matrix representing a specific relationship, N_iRepresenting a numerical node v_iIs determined. For each edge e_ji，r_jiIt is determined by two factors, respectively: (1) the numerical magnitude relationship, i.e., greater than or less than or equal to; (2) node types, the numbers corresponding to the two numeric nodes on an edge, are all from the question (q-q), or are all from the document (p-p), or one from the question and one from the document (q-p), or one from the document and one from the question (p-q). The formula expression is shown as formula (15):

r_ij∈{＞,≤}×{q-q,p-p,q-p,p-q} (15)

3) the node represents an update. Information obtained in the second step

Only information of neighbor numerical nodes is contained, and the method needs to be implemented

And the information of the numerical node is fused. The numerical node update formula is shown as equation (16):

wherein v is_i' denotes an updated numerical node, W_fRepresenting a parameter learnable matrix, b_fRepresenting the deviation vector, ReLU is a modified linear unit activation function. The whole process of the one-step numerical reasoning is expressed by a formula, and the calculation formula is shown as a formula (17):

wherein v' represents a numerical node after the numerical Reasoning in one Step, v represents a numerical node before the numerical Reasoning, and reading-Step represents a numerical Reasoning process. Through one-step numerical reasoning, the reasoning model can only deduce the relation between adjacent nodes. However, for some tasks (e.g., ranking), the inference model needs to infer relationships between multiple nodes, and therefore, multiple steps of numerical inference are required. The multi-step numerical reasoning formula is shown in equation (18):

v^t＝Reasoning-Step(v^t-1) (18)

wherein t represents the number of numerical reasoning, and when t-step numerical reasoning is needed, the steps 1) -3) are repeated in sequence, so as to obtain a numerical node v after t-step numerical reasoning^t，v^tThe subsequent splicing process can be continued as U in the formula (7).

Embodiments of the present invention evaluate a model by two metrics, Exact Match (EM) and number-centered F1 scores, respectively, where the DROP dataset is a common MRC dataset.

Comparing the model in the embodiment of the invention with several baseline models, wherein the baseline models comprise a semantic analysis model, a traditional MRC model and a numerical MRC model:

A. semantic analysis model

1) SynDep: a neural semantic analysis model (KDG) based on a sentence representation of Stanford dependencies;

2) OpenIE: the neural semantic analysis model is expressed based on open information extracted sentences;

3) SRL: the neural semantic analysis model is expressed based on sentences marked by semantic roles;

B. traditional MRC model

1) BiDAF: an MRC model, which encodes questions and articles using a bidirectional attention flow network;

2) QANT: representing problems and paragraphs using convolution and attention as building blocks of the encoder;

3) BERT: a pre-trained bidirectional Transformer (Transformer) -based language model;

C. numerical MRC model

1) NAQANT: a digital version of the QANT model;

2) NAQANT +: the enhanced version of NAQANet, may further consider real numbers (e.g., "2.5"), richer arithmetic expressions, data enhancement, etc.

The model (NumNet model) in the example of the present invention was compared with other baselines on the DROP dataset for performance, and the comparison results are shown in table 2:

TABLE 2

As can be seen from table 2, compared with a model based on semantic analysis, a traditional MRC model, and a digital MRC model, the model in the embodiment of the present invention can make full use of comparison information between numbers in a problem and a document, and therefore, better results are obtained on both a development set and a test set of a DROP data set.

The present embodiment demonstrates the effect of different GNN structures on the DROP development set through table 3, the performance of the different GNN structures is shown in table 3, the "compare", "number", and "all" correspond to the compare-like question subset, the number-type answer subset, and the entire development set, respectively, "-question number" indicates that the number in the question is not included in the graph, and "- ≦ side" and "- > side" indicate that the ≦ side and the > side, respectively, are not used.

TABLE 3

As can be seen from table 3, NumGNN is a significant improvement in both EM and F1 scores, particularly in comparison to conventional GNN. The result shows that for answering comparison questions, the numerical reasoning can be effectively assisted by considering comparison information among numbers.

The number of layers of the NumGNN represents the numerical reasoning capability of the model in the embodiment of the invention, and the K-layer model has the capability of K-step numerical reasoning. As shown in fig. 10, fig. 10 is a schematic diagram of different numbers of layers of the umGNN with different numerical reasoning capabilities according to the embodiment of the present invention:

(1) the 2-layer NumNet performs best on the comparative problem. Generally, a comparison class problem requires only two inferences (e.g., "who is the second older player of MLB, XX or MM;

(2) as some numerical problems require reasoning on more characters in documents, the performance of the NumNet model on the overall development set is continuously improved as the number of GNN layers increases. However, further studies have shown that the performance gain is not stable when K.gtoreq.4.

Embodiments of the present invention illustrate numerical reasoning in MRC by way of example, as shown in table 4:

TABLE 4

From table 4, it can be seen that for the first case, the NAQANet + model gives a wrong prediction, since NAQANet + cannot distinguish which of 10.1% and 56.2% is larger, whereas NAQANet + model gives a wrong prediction for "which age group is smaller: 18 years or under 18-24 years? The problem of "gives the same prediction. For the second case, NAQANet + cannot identify that the site target of the second length is 22 codes and gives a wrong prediction. For both cases, the NumNet model can give the correct answer through numerical reasoning, indicating the effectiveness of the NumNet model.

By performing error analysis on random samples of the NumNet prediction, it can be found that:

(1) the NumNet model can correctly answer about 76% of the sorting/comparison problems, and the NumNet model shows that the NumNet model realizes numerical reasoning capability to a certain extent;

(2) in answering incorrect classification/comparison questions, most (26%) questions are answered in multiple non-contiguous segments (e.g., line 1 in table 5), and the second largest question (19%) is a question that involves comparison with an intermediate number that is not present in the document/question (e.g., line 1 in table 5), but is obtained by counting or arithmetic operations, table 5 being a typical example of error.

TABLE 5

In summary, embodiments of the present invention encode numbers as nodes and relationships between them as edges by NumGNN. Through one-step reasoning, comparison and identification of numerical conditions in questions can be performed, and through multi-step reasoning, further ranking of questions and answers can be performed.

Machine reading comprehension problems requires digital reasoning skills such as addition, quadratic pull, sorting and counting. However, these skills are not available for most existing MRC models. According to the embodiment of the invention, through the Num Net model, explicit digital reasoning is carried out when an article is read. Specifically, the NumNet model encodes numerical relationships between numbers in a problem into a graph as a topological structure of the graph, and controls a numerically-aware graph neural network to numerically infer the graph, so that the NumNet model has great advantages in a DROP data set and is superior to other baseline models.

The embodiment of the invention can improve the effect of the MRC model in answering the questions needing comparison or sequencing type numerical reasoning, and can accurately answer the questions, thereby expanding the types of the questions which can be answered by the MRC model.

The model can take the NAQANT model as a basic model, and an inference module which can carry out comparison or sequencing type numerical value inference is added on the basis of the NAQANT model, so that the question-answering processing model in the embodiment of the invention can better process MRC problems needing the type inference. The basic model can also be other existing MRC models, and is not limited to the NAQANet model.

Under the condition of allowing the performance of the model to be reduced, the directed graph structure constructed by the reasoning module of the question-and-answer processing model in the embodiment of the invention can only consider the numbers in the document, can only use one magnitude relation edge, and can not distinguish whether the numbers corresponding to the numerical nodes come from the question or the document in the information propagation step of one-step numerical reasoning.

In summary, the embodiment of the present invention performs a series of processing through the question-answering processing model according to the question information and the document information to obtain the answer corresponding to the question information, and has the following beneficial effects:

The above description is only an example of the present invention, and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, and improvement made within the spirit and scope of the present invention are included in the protection scope of the present invention.

Claims

1. A question-answer processing method based on artificial intelligence is characterized by comprising the following steps:

2. The method according to claim 1, wherein the encoding of question information and document information by an encoding model in a question-and-answer processing model to obtain encoded information corresponding to the question information and encoded information corresponding to the document information comprises:

respectively carrying out first coding on the problem information and the document information through the coding model to obtain intermediate coding information corresponding to the problem information and intermediate coding information corresponding to the document information;

3. The method according to claim 1 or 2, wherein the encoding model in the question-answering processing model is used for encoding question information and document information respectively to obtain encoding information corresponding to the question information and encoding information corresponding to the document information, and the method comprises the following steps:

mapping the problem information and the document information to a vector space through the coding model to obtain intermediate coding information corresponding to the problem information and intermediate coding information corresponding to the document information;

4. The method according to claim 1 or 2, wherein the performing, by an inference model in the question-answering processing model, a directed graph building process on the coded information of the question information and the coded information of the document information to obtain a directed graph including the question information and the document information includes:

performing dimensionality transformation on the coding information of the problem information and the coding information of the document information through the reasoning model to obtain transformed coding information of the problem information and transformed coding information of the document information;

5. The method according to claim 4, wherein the creating a directed graph according to the extracted numerical values to obtain the directed graph including the question information and the document information comprises:

6. The method according to claim 1 or 2, wherein the numerically reasoning the directed graph including the question information and the document information through the reasoning model to obtain document information including a numerical relationship comprises:

performing numerical reasoning processing on the directed graph comprising the problem information and the document information at least once through the reasoning model to obtain a numerical node after reasoning;

7. The method according to claim 6, wherein the performing at least one numerical inference process on the directed graph including the question information and the document information through the inference model to obtain an inferred numerical node comprises:

determining the weight of each numerical value node in the directed graph according to the parameter learnable matrix and the deviation vector in the reasoning model;

performing the following processing for each of the numerical nodes:

8. The method according to claim 7, wherein determining neighbor numerical node information of the numerical node according to the side information of the numerical node and neighbor numerical nodes in the directed graph and weights of the neighbor numerical nodes comprises:

acquiring the side information of the numerical value node and the corresponding neighbor numerical value node from the directed graph;

9. The method of claim 8, wherein the obtaining the side information of the numerical node and the corresponding neighbor numerical node from the directed graph comprises:

10. The method according to claim 7, wherein the fusing the neighbor numerical node information of the numerical node and the self information of the numerical node to obtain the inferred numerical node comprises:

multiplying a parameter learnable matrix in the reasoning model by the information of the numerical node, and adding the multiplied information of the numerical node with the neighbor numerical node information of the numerical node and the deviation vector in the reasoning model to obtain an updated numerical node;

11. The method according to any one of claims 1 to 10, wherein the obtaining of the answer corresponding to the question information by performing prediction processing on the document information including the numerical relationship through a prediction model in the question-and-answer processing model comprises:

screening the document information containing the numerical relationship through the prediction model to obtain an answer type corresponding to the question information;

12. The method according to any one of claims 1 to 10, further comprising:

screening the document information samples containing numerical relationships through the prediction model to obtain answer type probabilities of the corresponding question information samples;

13. An artificial intelligence-based question-answering processing apparatus, characterized in that the apparatus comprises:

14. An artificial intelligence-based question-answering processing apparatus, characterized in that the apparatus comprises:

a memory for storing executable instructions;

a processor for implementing the artificial intelligence based question-answering method of any one of claims 1 to 12 when executing the executable instructions stored in the memory.

15. A storage medium storing executable instructions for causing a processor to perform the artificial intelligence based question answering method according to any one of claims 1 to 12 when executed.