Problem-solving method, device, medium and equipment based on artificial intelligence

Abstract

The present application discloses a problem-solving method, device, medium and equipment based on artificial intelligence, which belongs to the field of artificial intelligence. The method trains a problem-solving model, and during the model training process, constructs a question triple based on the contrastive learning idea for the training set; wherein, a question triple includes a basic question, a positive sample with a solution expression similar to the basic question, and a negative sample with a solution expression different from the basic question; then, each type of question in the question triple is encoded and decoded, and an expression tree of each type of question in the question triple is generated; then, the first loss function and the second loss function are constructed respectively in combination with the expression tree output by the problem-solving model; the first loss function is based on the contrastive learning idea, and the second loss function is used to measure the degree of difference between the predicted solution expression and the standard solution expression. When performing a problem-solving task based on the trained problem-solving model, a solution expression with a high accuracy rate can be generated.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and in particular to a problem-solving method, device, medium and equipment based on artificial intelligence. Background Art

[0002] Automatic question answering is a very important research direction in the field of artificial intelligence (AI). In recent years, with the development of theories such as natural language processing and machine learning and the popularization of online education, this research direction has become a new hot spot.

[0003] Automatic problem solving includes automatic problem solving of word problems. As word problems are very important for developing students’ reading comprehension and logical thinking abilities, they are widely used in examination papers, homework and daily exercises, and have always been the focus and difficulty of mathematics teaching.

[0004] For automatic problem solving of application questions, how to improve the accuracy of problem solving is the key, because this will seriously affect the quality of problem solving, and then affect the question-answering experience. Therefore, how to perform high-quality automatic problem solving based on artificial intelligence technology has become an urgent problem to be solved in this field. Summary of the invention

[0005] The embodiment of the present application provides a method, device, medium and equipment for solving problems based on artificial intelligence, which can significantly improve the accuracy of solving problems. The technical solution is as follows:

[0006] On the one hand, a problem-solving method based on artificial intelligence is provided, the method comprising:

[0007] Acquire a training set, the training set comprising a plurality of question triples; wherein one of the question triples comprises a basic question, a positive sample of the basic question, and a negative sample, the solution expression of the positive sample is similar to the solution expression of the basic question, and the solution expression of the negative sample is different from the solution expression of the basic question;

[0008] For any question triple in the training set, an encoder based on the problem-solving model encodes each type of question in the question triple to obtain encoding vectors of each type of question in the question triple;

[0009] The decoder based on the problem-solving model decodes the encoding vectors of each type of question in the question triple to obtain an expression tree of each type of question in the question triple; wherein the expression tree describes the predicted solution expressions of each type of question in a tree structure;

[0010] Based on the expression trees of various types of questions in the multiple question triples, a first loss function is constructed; wherein the first loss function is used to reduce the distance between the encoding vector of the basic question in any question triple and the encoding vector of the positive sample, and increase the distance between the encoding vector of the basic question in any question triple and the encoding vector of the negative sample;

[0011] Constructing a second loss function based on the expression tree of at least one type of question in the multiple question triples and the corresponding standard solution expression;

[0012] Training the problem-solving model based on the first loss function and the second loss function;

[0013] Obtain a target question, and generate a solution expression for the target question based on the trained problem-solving model.

[0014] On the other hand, there is provided a problem-solving device based on artificial intelligence, the device comprising:

[0015] The acquisition module is configured to acquire a training set, wherein the training set includes a plurality of question triples; wherein one of the question triples includes a basic question, a positive sample of the basic question, and a negative sample, wherein the solution expression of the positive sample is similar to the solution expression of the basic question, and the solution expression of the negative sample is different from the solution expression of the basic question;

[0016] An encoding module is configured to encode, for any one of the question triples in the training set, each type of question in the question triple based on an encoder of a problem-solving model, to obtain encoding vectors of each type of question in the question triple;

[0017] A decoding module is configured to decode the encoding vectors of each type of question in the question triple based on the decoder of the question-solving model to obtain expression trees of each type of question in the question triple; wherein the expression tree describes the predicted solution expressions of each type of question in a tree structure;

[0018] A training module is configured to construct a first loss function based on expression trees of various types of questions in the multiple question triples; wherein the first loss function is used to reduce the distance between the encoding vector of the basic question in any question triple and the encoding vector of the positive sample, and increase the distance between the encoding vector of the basic question in any question triple and the encoding vector of the negative sample;

[0019] The training module is further configured to construct a second loss function based on the expression tree of at least one type of question in the plurality of question triples and the corresponding standard solution expression;

[0020] The training module is further configured to train the problem-solving model based on the first loss function and the second loss function;

[0021] The generation module is configured to obtain a target question and generate a solution expression for the target question based on the trained problem-solving model.

[0022] In some possible implementations, the training module is further configured to:

[0023] For any question triple, obtain the node vector of the first expression tree of the basic question in the question triple, the node vector of the second expression tree of the positive sample in the question triple, and the node vector of the third expression tree of the negative sample in the question triple; wherein the node vector refers to the feature vector of the node in the corresponding expression tree;

[0024] Obtaining a first cosine similarity between a node vector of the first expression tree and a node vector of the second expression tree;

[0025] Obtaining a second cosine similarity between the node vector of the first expression tree and the node vector of the third expression tree;

[0026] Determining a loss value of the question triple based on the first cosine similarity and the second cosine similarity;

[0027] The first loss function is constructed based on the loss values ​​of the multiple question triplets.

[0028] In some possible implementations, the solution expression of the positive example is similar to the solution expression of the basic question, including:

[0029] The solution expression of the positive example is partially the same as the solution expression of the basic question; or,

[0030] The solution expression of the positive example is exactly the same as the solution expression of the basic question.

[0031] In some possible implementations, the acquisition module is further configured to:

[0032] In response to the existence of a subtree in the expression tree of the first question being consistent with the first expression tree of the basic question, taking the first question as a positive example of the basic question;

[0033] In response to the expression tree of the second question and the first expression tree of the basic question having the same number of nodes and different operators, the second question is used as a negative sample of the basic question.

[0034] In some possible implementations, the training module is further configured to:

[0035] Constructing a target loss function based on the first loss function, the second loss function and a loss weight, wherein the loss weight is used to measure the importance of the second loss function;

[0036] The problem-solving model is trained with minimizing the target loss function as the training goal.

[0037] In some possible implementations, the training module is further configured to:

[0038] The second loss function is constructed based on the first expression trees and corresponding standard solution expressions of the basic questions in the multiple question triples, and the second expression trees and corresponding standard solution expressions of the positive examples in the multiple question triples.

[0039] On the other hand, a computer device is provided, comprising a processor and a memory, wherein the memory stores at least one program code, and the at least one program code is loaded and executed by the processor to implement the above-mentioned artificial intelligence-based problem-solving method.

[0040] On the other hand, a computer-readable storage medium is provided, in which at least one program code is stored, and the at least one program code is loaded and executed by a processor to implement the above-mentioned artificial intelligence-based problem-solving method.

[0041] On the other hand, a computer program product or a computer program is provided, which includes a computer program code, and the computer program code is stored in a computer-readable storage medium. A processor of a computer device reads the computer program code from the computer-readable storage medium, and the processor executes the computer program code, so that the computer device executes the above-mentioned artificial intelligence-based problem-solving method.

[0042] The problem-solving solution provided in the embodiment of the present application will train a problem-solving model. During the model training process, a question triple based on the contrastive learning idea is constructed for the training set; wherein a question triple includes a basic question, a positive sample and a negative sample of the basic question, the solution expression of the positive sample is similar to the solution expression of the basic question, and the solution expression of the negative sample is different from the solution expression of the basic question; then, the encoder and decoder based on the problem-solving model respectively encode and decode each type of question in the question triple to generate an expression tree for each type of question in the question triple; then, the problem-solving model is trained based on the multi-task learning training method. In detail, the first loss function and the second loss function are constructed respectively in combination with the expression tree output by the problem-solving model; wherein, the first loss function is used to reduce the distance between the encoding vector of the basic question and the encoding vector of the positive sample in the question triple, and increase the distance between the encoding vector of the basic question and the encoding vector of the negative sample in the question triple; since the contrastive learning idea is introduced into the model training process, the encoding vectors of similar solution expressions are brought closer in the feature space, and the encoding vectors of different solution expressions are pulled farther in the feature space, so the trained problem-solving model can better understand the solution mode of the question, that is, better understand the quantitative relationship in the question, and thus can greatly improve the accuracy of problem-solving. In addition, the second loss function is used to measure the degree of difference between the predicted solution expression and the standard solution expression. The second loss function can prompt the model to better understand the semantics of the question, and further ensure the accuracy of problem-solving. Therefore, when performing problem-solving tasks based on the trained problem-solving model, a solution expression with a high accuracy rate can be generated. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0044] Figure 1 It is a schematic diagram of an implementation environment involved in an artificial intelligence-based problem-solving method provided in an embodiment of the present application;

[0045] Figure 2 is a schematic diagram of an application scenario of a problem-solving task provided in an embodiment of the present application;

[0046] Figure 3 is a schematic diagram of a visualization result of an encoding vector of a question provided in an embodiment of the present application;

[0047] Figure 4 It is a flow chart of a problem-solving method based on artificial intelligence provided in an embodiment of the present application;

[0048] Figure 5 It is a schematic diagram of the architecture of a problem-solving model provided in an embodiment of the present application;

[0049] Figure 6 It is a structural schematic diagram of an artificial intelligence-based problem-solving device provided in an embodiment of the present application;

[0050] Figure 7 It is a structural diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0051] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0052] In this application, the terms "first", "second", etc. are used to distinguish the same or similar items with substantially the same role and function. It should be understood that there is no logical or temporal dependency between "first", "second", and "nth", nor is there a limitation on the quantity and execution order. It should also be understood that although the following description uses the terms first, second, etc. to describe various elements, these elements should not be limited by the terms.

[0053] These terms are only used to distinguish one element from another element. For example, without departing from the scope of various examples, the first element can be referred to as the second element, and similarly, the second element can also be referred to as the first element. Both the first element and the second element can be elements, and in some cases, can be separate and different elements.

[0054] Here, at least one means one or more than one, for example, at least one element can be one element, two elements, three elements, or any other integer greater than or equal to one. And multiple means two or more than two, for example, multiple elements can be two elements, three elements, or any other integer greater than or equal to two.

[0055] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions. For example, the initial text and interest tags involved in this application are obtained with full authorization.

[0056] The problem-solving solution provided in the embodiments of the present application involves artificial intelligence technology.

[0057] Among them, artificial intelligence is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines so that machines have the functions of perception, reasoning and decision-making.

[0058] Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields, including both hardware-level and software-level technologies. The basic technologies of artificial intelligence generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics and other technologies. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0059] Natural Language Processing (NLP) is an important direction in the fields of computer science and artificial intelligence. It studies various theories and methods that can achieve effective communication between people and computers using natural language. Natural language processing is a science that integrates linguistics, computer science, and mathematics. Therefore, research in this field will involve natural language, that is, the language people use in daily life, so it is closely related to the study of linguistics. Natural language processing technology usually includes text processing, semantic understanding, machine translation, robot question answering, knowledge graph and other technologies.

[0060] Machine Learning (ML) is a multi-disciplinary subject that involves probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, teaching learning, representation learning and other technologies.

[0061] With the research and advancement of artificial intelligence technology, artificial intelligence technology has been studied and applied in many fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned driving, automatic driving, drones, robots, smart medical care, smart customer service, etc. I believe that with the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0062] The problem-solving solutions provided in the embodiments of the present application may involve technologies such as natural language processing and machine learning of artificial intelligence, which are specifically illustrated by the following embodiments.

[0063] Some key terms or abbreviations involved in the embodiments of the present application are first introduced below.

[0064] Contrastive Learning (CL): is a type of representation learning that learns the feature representation of a sample by comparing the data with its positive and negative samples in the feature space. In other words, the goal of contrastive learning is to learn a feature space in which similar pairs of samples are close to each other, while different pairs of samples are far apart.

[0065] AI automatic solution: In the present application embodiment, it specifically refers to the machine learning model understanding the problem statement of the word problem and then generating a solution expression. Exemplarily, the above word problem is an elementary mathematics word problem, such as a primary school mathematics word problem.

[0066] The implementation environment of the problem-solving solution provided in the embodiment of the present application is introduced below.

[0067] The problem-solving method based on artificial intelligence provided in the embodiment of the present application is applied to a computer device. For example, Figure 1 1 is a schematic diagram of an implementation environment involved in an artificial intelligence-based problem-solving method provided in an embodiment of the present application. The implementation environment includes: a first computer device 101 and a second computer device 102. The first computer device 101 is used to train a problem-solving model, and the second computer device 102 is used to perform a problem-solving task using the trained problem-solving model.

[0068] In some possible implementations, the first computer device 101 and the second computer device 102 may also be the same device, that is, after completing the model training on the device, the trained problem-solving model is used to perform the problem-solving task on the device, and this application does not impose any restrictions on this.

[0069] In some other possible implementations, the first computer device 101 is represented as a server, and the second computer device 102 is represented as a user terminal. Among them, the server can be an independent physical server, or a server sample cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. The user terminal includes but is not limited to mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle-mounted terminals, aircraft, etc., but is not limited to this. Among them, the computer can include tablet computers, laptops and desktop computers. The user terminal and the server can be directly or indirectly connected via wired or wireless communication, and this application is not limited here.

[0070] In some other possible implementations, based on the above implementation environment, the trained problem-solving model can be applied to Figure 2 The problem-solving scenario, automatic grading scenario, and question bank scenario are shown in the figure. Figure 2 As shown, given a math word problem, the trained problem-solving model generates the corresponding solution expression, which is then directly applied in the problem-solving scenario; for the automatic correction scenario, the student's answer expression can be automatically judged based on the solution expression output by the model, that is, automatic correction can be achieved based on the trained problem-solving model. For the question bank scenario, the question bank can be expanded and the answer verified based on the trained problem-solving model. For example, first determine whether the given math word problem already exists in the question bank; in response to the math word problem existing in the question bank, the answer saved in the question bank can be verified according to the solution expression output by the model; in addition, a manual review step can be added, that is, manually review whether the answer to the question bank is really wrong, and modify the answer to the question bank based on the solution expression output by the model in the case of an error; in response to the math word problem not existing in the question bank, the solution formula output by the model is used to expand the question bank. In this case, a manual review step can also be added, that is, manually review whether the solution expression output by the model is correct, and add it to the question bank if it is correct.

[0071] In some other possible implementations, the problem-solving model proposed in the embodiment of the present application adopts an encoder-decoder architecture. In addition, considering that the same quantitative relationship also exists in word problems of different topics, the embodiment of the present application applies comparative learning to the problem-solving task of word problems. In detail, the vector representation of the problem output by the encoder of the decoding model (referred to as the encoding vector in this article) is visualized and analyzed. Among them, Figure 3 The visualization results based on the tSNE algorithm are given. Figure 3 It can be seen from the figure that if the solution expressions are similar, even if the two questions are expressed differently, the encoding vectors of most questions will tend to cluster together.

[0072] for example, Figure 3 Although the problem A and problem B have different expressions, their solutions are the same (both are n1-n2), so their encoding vectors will also show a trend of clustering. Figure 3 The solution expressions for Problem C and Problem D are different. One solution expression is n1*n2 and the other solution expression is n1 / n2, so the encoding vectors of the two are also far apart.

[0073] It can be concluded that the solution expression of the question reflects the solution mode of the question, and its solution expression also represents the quantitative relationship in the question. In addition, the encoding vector of the question will be clustered according to the solution expression, and the encoding vector far away from the cluster center tends to generate an erroneous solution expression. Inspired by this, the embodiment of the present application applies comparative learning to the solution of word problems. By finding similar solution expressions, the encoding vectors of similar solution expressions are pulled closer in the feature space, and the encoding vectors of different solution expressions are pulled away in the feature space, to support the problem-solving model to better understand the solution mode of the question, that is, to truly understand the quantitative relationship in the question, thereby improving the accuracy of problem solving.

[0074] The solution to the problem proposed in the embodiment of the present application is described in detail through the following implementation method.

[0075] Figure 4 is a flowchart of an artificial intelligence-based problem-solving method provided in an embodiment of the present application, and the execution subject of the problem-solving method is a computer device. Figure 4 , the method comprises the following steps.

[0076] 401. A computer device obtains a training set, which includes multiple question triplets; wherein a question triplet includes a basic question, a positive sample of the basic question, and a negative sample, the solution expression of the positive sample is similar to the solution expression of the basic question, and the solution expression of the negative sample is different from the solution expression of the basic question.

[0077] The first point to be explained is that, in order to distinguish the expression trees of basic questions, positive samples and negative samples in the following text, the expression trees of basic questions are collectively referred to as the first expression tree, the expression trees of positive samples are collectively referred to as the second expression tree, and the expression trees of negative samples are collectively referred to as the third expression tree. The expression tree describes the solution expression in a tree structure, and the leaf nodes in the expression tree indicate the operation numbers, and the non-leaf nodes indicate the operation symbols.

[0078] The introduction of the contrastive learning method requires the construction of a set of question triplets that can be compared, also known as application question triplets. In the embodiment of the present application, the question triplets are represented by the symbol (p, p + ,p - ) refers to, where question p is the basic question, p + is a positive sample, p - As mentioned above, the contrastive learning method compares the basic question p and the positive sample p + The encoding vector of the feature space is pulled closer and the basic question p and the negative sample p - The encoding vectors are pulled further away in the feature space, making the clustering effect of the problem-solving model on the entire training set more obvious and reducing the proportion of incorrect solution expressions generated by the problem-solving model.

[0079] In some possible implementations, the solution expression of the positive example sample is similar to the solution expression of the basic question, including the following two cases: the solution expression of the positive example sample is partially identical to the solution expression of the basic question; or, the solution expression of the positive example sample is exactly the same as the solution expression of the basic question.

[0080] For the above positive examples, the simplest way to collect them is to collect questions with the same solution expressions as the basic questions and use them as positive examples. In addition, the same quantitative relationship as question p also exists in some complex questions. Figure 3 Take question A in the example, for another question E "Xiao Ming has N1 yuan, and he spent N2 yuan on records. How many games can he buy for N3 yuan?", there is still a sub-problem when solving this question "How much money does Xiao Ming have left?" Although the solution formula of the whole question is (N1-N2) / N3, N1-N2 represents the solution expression of this sub-problem, and the solution expression of this sub-problem is completely consistent with the solution expression of question A, that is, the solution expression of question A is partially the same as the solution expression of question E. Therefore, it is also necessary to obtain this type of data as a positive example sample in the sample data acquisition stage. That is, when obtaining positive example samples, the embodiment of the present application does not pay attention to the specific expression of the question, but only obtains this type of data based on whether the solution expression is consistent.

[0081] In some other possible implementations, in the sample data acquisition stage, the solution expression of the acquired question can be first parsed and converted into an expression tree; in the process of constructing a question triple, the basic question is first determined. After the basic question is determined, if there is a subtree in the expression tree of a question that is consistent with the first expression tree of the basic question, then the question is regarded as a positive sample of the basic question. If the expression tree of a question has the same number of nodes as the first expression tree of the basic question but the operators are inconsistent, then the question is regarded as a negative sample of the basic question. That is, the method provided in the embodiment of the present application also includes: in response to the existence of a subtree in the expression tree of the first question that is consistent with the first expression tree of the basic question, the first question is regarded as a positive sample of the basic question; in response to the expression tree of the second question and the first expression tree of the basic question having the same number of nodes and different operators, the second question is regarded as a negative sample of the basic question. Among them, the first question and the second question generally refer to the questions acquired in the sample data acquisition stage.

[0082] For example, see Figure 5 , positive sample p + The expression tree tree + There is a subtree in the expression tree that is consistent with the expression tree of the basic question p, and the expression tree of the basic question p is consistent with the negative sample p. - The expression tree tree - Although they have the same number of nodes, the operators are different, one is a multiplication sign and the other is a addition sign.

[0083] The second point that needs to be explained is that a basic question may have multiple positive samples or multiple negative samples. When constructing question triples, more than one question triplet can be constructed for the same basic question, that is, the basic questions in these question triplets are the same, but have different positive samples or negative samples. This application does not impose any restrictions here.

[0084] In some other possible implementations, the training set includes, in addition to multiple question triples, corresponding standard (correct) solution expressions. That is, the training data used in the embodiment of the present application is questions and their solution expressions. Exemplarily, any question triple in the training set corresponds to three standard solution expressions, which are the standard solution expression of the basic question, the standard solution expression of the positive sample, and the standard solution expression of the negative sample.

[0085] 402. For any question triple in the training set, the computer device encodes each type of question in the question triple based on the encoder of the problem-solving model to obtain the encoding vector of each type of question in the question triple; the decoder based on the problem-solving model decodes the encoding vector of each type of question in the question triple to obtain the expression tree of each type of question in the question triple.

[0086] In some possible implementations, the problem-solving model adopts an encoder-decoder architecture. The encoder is used to understand the semantics of the question, also known as the question semantic encoder; the encoder is used to encode the input question to obtain the encoding vector of the question; and the encoding vector is further input into the decoder to guide the decoder to generate a solution expression. In other words, the input of the problem-solving model is a complete question, and the output of the problem-solving model is a solution expression. The solution expression output by the problem-solving model is collectively referred to as the predicted solution expression in this article.

[0087] In some possible implementations, see Figure 5 , the encoder of the problem-solving model is BERT (Bidirectional Encoder Representation from Transformers), that is, BERT is used as the question semantic encoder, and the tree decoder is applied to obtain the solution expression. It should be noted that, Figure 5 As shown in the figure, the tree decoder actually generates an expression tree, and the expression tree generated by the tree decoder describes the predicted solution expressions of each type of question in the question triple in a tree structure. Afterwards, by traversing the expression tree generated by the tree decoder, the predicted solution expressions of each type of question in the question triple can be obtained.

[0088] Figure 5 The figure shows the processing process of the problem solving model for any question triple. For example, the question triple is the question triple (p, p + ,p - ) as an example, see Figure 5 First, based on BERT, we analyze the basic questions p and positive samples p + And negative sample p - Encode to obtain the encoding vector h of the basic question p and the positive sample p + The encoding vector h + And negative sample p - The encoding vector h - ; Then, based on the tree decoder, the encoding vector h and the encoding vector h + and the encoding vector h - Decode them separately to get the expression tree tree of the basic question p and the positive sample p +The expression tree tree + And negative sample p - The expression tree tree - By traversing the expression tree tree, we can get the predicted solution expression eq. + Then we can get the predicted solution expression eq + , traverse the expression tree tree - Then we can get the predicted solution expression eq - .

[0089] BERT is essentially a feature extraction model composed of a bidirectional Transformer (transformer). In an embodiment of the present application, BERT includes N layers of feature encoders with the same structure stacked in sequence. The value of N is a positive integer, for example, the value of N can be 12 or 24, that is, the BERT model can include 12 layers of feature encoders (BERT-base model) or 24 layers of feature encoders (BERT-large model), and the present application does not limit this. It should be noted that the feature encoder here refers to the encoder of the Transformer. Since BERT uses a multi-layer Transformer to learn the topic, and the Transformer can read the topic in a sampling one-time reading manner, it can learn the contextual relationship of the topic more accurately, and have a deeper understanding of the semantics, so that it can accurately extract features from the topic. Therefore, BERT has better task processing effects than other models for processing natural language processing tasks.

[0090] The tree decoder decodes the encoding vector of the question to establish a tree structure relationship between the numbers in the question. The following introduces the workflow of the tree decoder in conjunction with question F "Xiao Ming is packing his biscuits into bags, and each bag contains 6 biscuits. If he has 23 chocolate biscuits and 25 cookies, how many bags does he need?"

[0091] For question F, the tree decoder first determines the root target vector based on the encoding vector of the question, which represents the final goal of question F, that is, how many packages are needed; then, according to the root target vector, it obtains relevant context information, such as "a bag contains 6 cookies, 23 chocolate cookies, and 25 cookies". Next, the context information and the root target vector are used to predict the token corresponding to the root node position. The corresponding token predicted for question F should be the operator " / ". Since the predicted token is an operator, not an operation number, it indicates that it needs to be further decomposed into two sub-goals. The two sub-goals are "how many cookies are there in total" and "how many cookies can be packed in a bag". Then, continue to predict the token corresponding to the first sub-goal. Since the predicted token is the operator "+", it needs to be further decomposed; according to this process, the expression tree is constructed from top to bottom through repeated repetitions.

[0092] It should be noted that the embodiment of the present application adopts a multi-task learning training method to optimize the problem-solving model, wherein the task objectives include problem-solving objectives and comparative learning objectives. To put it another way, during the training process, the embodiment of the present application constructs a comparative learning loss function and a problem-solving loss function. In this article, the comparative learning loss function is also referred to as the first loss function, and the problem-solving loss function is also referred to as the second loss function. The construction process of the two loss functions is described below through the following steps 403 and 404, respectively.

[0093] 403. The computer device constructs a first loss function based on the expression trees of various types of questions in multiple question triples; wherein the first loss function is used to reduce the distance between the encoding vector of the basic question in any question triple and the encoding vector of the positive sample, and increase the distance between the encoding vector of the basic question in any question triple and the encoding vector of the negative sample.

[0094] In some possible implementations, based on expression trees of various types of questions in multiple question triples, a first loss function is constructed, including but not limited to the following steps:

[0095] 4031. For any question triple, obtain the node vector of the first expression tree of the basic question in the question triple, the node vector of the second expression tree of the positive example in the question triple, and the node vector of the third expression tree of the negative example in the question triple.

[0096] It should be noted that the node vector mentioned here refers to the feature vector of the node in the corresponding expression tree (also called node embedding); for any node in the expression tree, the corresponding node embedding includes the information of the entire subtree with the node as the root node. The node vector is calculated when the tree decoder generates the expression tree.

[0097] 4032. Obtain a first cosine similarity between the node vector of the first expression tree and the node vector of the second expression tree.

[0098] The symbol α refers to the first cosine similarity, the symbol e refers to the node vector of the first expression tree, and the symbol e + Taking the node vector representing the second expression tree as an example, the first cosine similarity is obtained by the following formula.

[0099]

[0100] 4033. Obtain the second cosine similarity between the node vector of the first expression tree and the node vector of the third expression tree.

[0101] The symbol β refers to the second cosine similarity, and the symbol e - Taking the node vector representing the third expression tree as an example, the second cosine similarity is obtained by the following formula.

[0102]

[0103] 4034. Determine a loss value of the question triple based on the first cosine similarity and the second cosine similarity; and construct a first loss function based on the loss values ​​of multiple question triples.

[0104] Among them, the loss value of the triplet of this question is max(0, η+α-β), η is a hyperparameter, and the max(,) function is used to return the maximum value in a set of values.

[0105] In some possible implementations, for the entire training set, the first loss function is the sum of the loss values ​​of all the question triplets in the training set, expressed as L eq Refers to the first loss function, and takes D as an example to refer to the training set, then

[0106] 404. The computer device constructs a second loss function based on the expression tree of at least one type of question in the multiple question triples and the corresponding standard solution expression.

[0107] In an embodiment of the present application, the second loss function is constructed based on the standard solution expression of the question and the solution expression predicted by the decoder of the problem-solving model. Among them, the second loss function compares the solution expression predicted by the decoder with the standard solution expression, that is, measures the degree of difference between the two, thereby calculating the gradient of the model (the direction of model optimization). Exemplarily, in response to the decoder being a tree decoder, the predicted solution expression is described in the form of an expression tree.

[0108] In some possible implementations, the embodiment of the present application constructs a second loss function based on the expression tree of at least one type of question in the training set and the corresponding standard solution expression tree. For example, Figure 5 The second loss function is calculated using the basic questions and positive samples in the training set. That is, the second loss function is constructed based on the first expression tree and the corresponding standard solution expression of the basic questions in the training set, the second expression tree and the corresponding standard solution expression of the positive samples.

[0109] In addition, the second loss function can also be calculated using the basic questions in the training set. That is, the second loss function is constructed based on the first expression tree of the basic questions in the training set and the corresponding standard solution expression. In addition, the second loss function can also be calculated using all the questions in the training set. That is, the second loss function is constructed based on the first expression tree of the basic questions in the training set and the corresponding standard solution expression, the second expression tree of the positive sample and the corresponding standard solution expression, and the third expression tree of the negative sample and the corresponding standard solution expression.

[0110] Exemplarily, each node in the expression tree corresponds to a prediction token, wherein the prediction token is an operation symbol or an operation number that appears in the prediction solution expression. The tree decoder predicts a token corresponding to a node at each step, and the prediction token is determined based on a prediction probability vector, which gives the probability of each item in the candidate operation item appearing at the node, and the prediction token is the operation item indicated by the maximum probability value in the prediction probability vector. Among them, the candidate operation items include both operation symbols such as addition, subtraction, multiplication and division, and numbers that appear in the question, and this application does not limit this.

[0111] Taking the second loss function as the negative log-likelihood loss function as an example, the standard solution expression can be disassembled first, and then the negative log-likelihood is calculated item by item with the nodes in the expression tree generated by the problem-solving model, and finally the negative log-likelihood of all questions participating in the calculation of the first loss function is accumulated to obtain the second loss function. Exemplarily, when calculating the negative log-likelihood, the negative log-likelihood can be calculated based on the true probability vector of each item in the standard solution expression and the predicted probability vector corresponding to the node in the expression tree, and this application is not limited here.

[0112] 405. The computer device trains a problem-solving model based on the first loss function and the second loss function.

[0113] In some possible implementations, the problem-solving model is trained based on the first loss function and the second loss function, including: constructing a target loss function based on the first loss function, the second loss function and the loss weight, the loss weight being used to measure the importance of the second loss function, that is, used to measure the importance of contrastive learning; and training the problem-solving model by minimizing the target loss function as the training goal.

[0114] Among them, as the training continues, the second loss function prompts the problem-solving model to understand the scene or semantic information in the question, and the first loss function supports the problem-solving model to understand the problem-solving mode from questions with similar solution expressions, so that the trained problem-solving model can achieve higher problem-solving accuracy.

[0115] 406. The computer device obtains a target question and generates a solution expression for the target question based on the trained problem-solving model.

[0116] After the problem-solving model is trained, the problem-solving task can be performed based on the trained problem-solving model. For example, the target question is any mathematical word problem to be solved.

[0117] The problem-solving solution provided in the embodiment of the present application will train a problem-solving model. During the model training process, a question triple based on the contrastive learning idea is constructed for the training set; wherein a question triple includes a basic question, a positive sample and a negative sample of the basic question, the solution expression of the positive sample is similar to the solution expression of the basic question, and the solution expression of the negative sample is different from the solution expression of the basic question; then, the encoder and decoder based on the problem-solving model respectively encode and decode each type of question in the question triple to generate an expression tree for each type of question in the question triple; then, the problem-solving model is trained based on the multi-task learning training method. In detail, the first loss function and the second loss function are constructed respectively in combination with the expression tree output by the problem-solving model; wherein, the first loss function is used to reduce the distance between the encoding vector of the basic question in the question triple and the encoding vector of the positive sample, and increase the distance between the encoding vector of the basic question in the question triple and the encoding vector of the negative sample; due to the introduction of the contrastive learning idea into the model training process, the encoding vectors of similar solution expressions are pulled closer in the feature space, and the encoding vectors of different solution expressions are pulled away in the feature space, so the trained problem-solving model can better understand the problem-solving mode, that is, better understand the quantitative relationship in the problem, and thus can greatly improve the problem-solving accuracy. In addition, the second loss function is used to measure the degree of difference between the predicted solution expression and the standard solution expression. The second loss function can prompt the model to better understand the semantics of the question and further ensure the problem-solving accuracy. Therefore, when performing the problem-solving task based on the trained problem-solving model, a solution expression with a higher accuracy rate can be generated. In other words, the problem-solving model training method provided in the embodiment of the present application can improve the generalization ability and problem-solving ability of the model.

[0118] For example, the present application embodiment was experimented on the Chinese data set Math23k and the English data set MathQA. The problem-solving model provided by the present application embodiment can significantly improve the accuracy of problem-solving. On Math23k, the problem-solving accuracy rate was increased from 80.5% to 83.9% compared with the related art. On MathQA, the problem-solving accuracy rate was increased from 73.5% to 76.3% compared with the related art. At the same time, the problem-solving model provided by the present application embodiment was applied to the automatic correction in the field of education to verify the expressions of students' answers. The correction accuracy rate was increased from 46.7% to 69% compared with the related art.

[0119] In some other possible implementations, the generalization and problem-solving ability of the problem-solving model can be improved in a variety of ways: in terms of data, a high-quality and diverse data set is constructed for training the problem-solving model; in terms of the solution architecture, a verification network can be added to verify the solution expression predicted by the problem-solving model; for example, for a certain question, the difference between the corresponding predicted solution expression and the standard solution expression is verified based on the verification network; illustratively, the verification network can be based on the feature vector of the predicted solution expression and the feature vector of the standard solution expression, and the verification process can be finally converted into a binary classification problem, that is, the verification network finally gives a verification result of whether the two are consistent. In terms of training methods, you can try to use pre-training and fine-tuning to train the problem-solving model. For example, most of the parameters of the initial model have been pre-trained by pre-training. After obtaining this initial model, the initial model is fine-tuned based on the obtained training set, and then the problem-solving model required to perform the problem-solving task is obtained.

[0120] Figure 6 It is a structural schematic diagram of an artificial intelligence-based problem-solving device provided in an embodiment of the present application.

[0121] See also Figure 6 , the device comprises:

[0122] The acquisition module 601 is configured to acquire a training set, wherein the training set includes a plurality of question triples; wherein one of the question triples includes a basic question, a positive sample of the basic question, and a negative sample, wherein the solution expression of the positive sample is similar to the solution expression of the basic question, and the solution expression of the negative sample is different from the solution expression of the basic question;

[0123] The encoding module 602 is configured to encode, for any one of the question triples in the training set, each type of question in the question triple using an encoder based on the problem-solving model to obtain encoding vectors for each type of question in the question triple;

[0124] The decoding module 603 is configured to decode the encoding vectors of each type of question in the question triple based on the decoder of the question-solving model to obtain expression trees of each type of question in the question triple; wherein the expression tree describes the predicted solution expressions of each type of question in a tree structure;

[0125] The training module 604 is configured to construct a first loss function based on the expression trees of the various types of questions in the multiple question triples; wherein the first loss function is used to reduce the distance between the encoding vector of the basic question in any question triple and the encoding vector of the positive sample, and increase the distance between the encoding vector of the basic question in any question triple and the encoding vector of the negative sample;

[0126] The training module 604 is further configured to construct a second loss function based on the expression tree of at least one type of question in the plurality of question triples and the corresponding standard solution expression;

[0127] The training module 604 is further configured to train the problem-solving model based on the first loss function and the second loss function;

[0128] The generation module 605 is configured to obtain a target question and generate a solution expression for the target question based on the trained problem-solving model.

[0129] The problem-solving solution provided in the embodiment of the present application will train a problem-solving model. During the model training process, a question triple based on the contrastive learning idea is constructed for the training set; wherein a question triple includes a basic question, a positive sample and a negative sample of the basic question, the solution expression of the positive sample is similar to the solution expression of the basic question, and the solution expression of the negative sample is different from the solution expression of the basic question; then, the encoder and decoder based on the problem-solving model respectively encode and decode each type of question in the question triple to generate an expression tree for each type of question in the question triple; then, the problem-solving model is trained based on the multi-task learning training method. In detail, the first loss function and the second loss function are constructed respectively in combination with the expression tree output by the problem-solving model; wherein, the first loss function is used to reduce the distance between the encoding vector of the basic question and the encoding vector of the positive sample in the question triple, and increase the distance between the encoding vector of the basic question and the encoding vector of the negative sample in the question triple; since the contrastive learning idea is introduced into the model training process, the encoding vectors of similar solution expressions are brought closer in the feature space, and the encoding vectors of different solution expressions are pulled farther in the feature space, so the trained problem-solving model can better understand the solution mode of the question, that is, better understand the quantitative relationship in the question, and thus can greatly improve the accuracy of problem-solving. In addition, the second loss function is used to measure the degree of difference between the predicted solution expression and the standard solution expression. The second loss function can prompt the model to better understand the semantics of the question, and further ensure the accuracy of problem-solving. Therefore, when performing problem-solving tasks based on the trained problem-solving model, a solution expression with a high accuracy rate can be generated.

[0130] In some possible implementations, the training module 604 is further configured to:

[0131] For any question triple, obtain the node vector of the first expression tree of the basic question in the question triple, the node vector of the second expression tree of the positive sample in the question triple, and the node vector of the third expression tree of the negative sample in the question triple; wherein the node vector refers to the feature vector of the node in the corresponding expression tree;

[0132] Obtaining a first cosine similarity between a node vector of the first expression tree and a node vector of the second expression tree;

[0133] Obtaining a second cosine similarity between the node vector of the first expression tree and the node vector of the third expression tree;

[0134] Determining a loss value of the question triple based on the first cosine similarity and the second cosine similarity;

[0135] The first loss function is constructed based on the loss values ​​of the multiple question triplets.

[0136] In some possible implementations, the solution expression of the positive example is similar to the solution expression of the basic question, including:

[0137] The solution expression of the positive example is partially the same as the solution expression of the basic question; or,

[0138] The solution expression of the positive example is exactly the same as the solution expression of the basic question.

[0139] In some possible implementations, the acquisition module 601 is further configured to:

[0140] In response to the existence of a subtree in the expression tree of the first question being consistent with the first expression tree of the basic question, taking the first question as a positive example of the basic question;

[0141] In response to the expression tree of the second question and the first expression tree of the basic question having the same number of nodes and different operators, the second question is used as a negative sample of the basic question.

[0142] In some possible implementations, the training module 604 is further configured to:

[0143] Constructing a target loss function based on the first loss function, the second loss function and a loss weight, wherein the loss weight is used to measure the importance of the second loss function;

[0144] The problem-solving model is trained with minimizing the target loss function as the training goal.

[0145] In some possible implementations, the training module 604 is further configured to:

[0146] The second loss function is constructed based on the first expression trees and corresponding standard solution expressions of the basic questions in the multiple question triples, and the second expression trees and corresponding standard solution expressions of the positive examples in the multiple question triples.

[0147] All the above optional technical solutions can be arbitrarily combined to form optional embodiments of the present disclosure, and will not be described in detail here.

[0148] It should be noted that: when the problem-solving device based on artificial intelligence provided in the above embodiment performs a problem-solving task, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the problem-solving device based on artificial intelligence is divided into different functional modules to complete all or part of the functions described above. In addition, the problem-solving device based on artificial intelligence provided in the above embodiment and the problem-solving method embodiment based on artificial intelligence belong to the same concept. The specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0149] Figure 7 It is a structural diagram of a computer device 700 provided in an embodiment of the present application. The computer 700 may be a server. The computer device 700 may have relatively large differences due to different configurations or performances, and may include one or more processors (Central Processing Units, CPU) 701 and one or more memories 702, wherein at least one program code is stored in the memory 702, and the at least one program code is loaded and executed by the processor 701 to implement the problem-solving method based on artificial intelligence provided by the above-mentioned various method embodiments. Of course, the computer device 700 may also have components such as a wired or wireless network interface, a keyboard, and an input and output interface for input and output, and the computer device 700 may also include other components for implementing device functions, which will not be repeated here.

[0150] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including a program code, and the program code can be executed by a processor in a computer device to complete the problem-solving method based on artificial intelligence in the above embodiment. For example, the computer-readable storage medium can be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), a magnetic tape, a floppy disk, and an optical data storage device.

[0151] In an exemplary embodiment, a computer program product or a computer program is also provided, which includes a computer program code, and the computer program code is stored in a computer-readable storage medium. A processor of a computer device reads the computer program code from the computer-readable storage medium, and the processor executes the computer program code, so that the computer device executes the above-mentioned artificial intelligence-based problem-solving method.

[0152] In some possible implementations, the computer program involved in the embodiments of the present application may be deployed and executed on a computer device, or on multiple computer devices located at one location, or on multiple computer devices distributed at multiple locations and interconnected by a communication network. Multiple computer devices distributed at multiple locations and interconnected by a communication network may constitute a blockchain system.

[0153] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0154] The above description is only an optional embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A problem-solving method based on artificial intelligence, characterized in that: The method comprises: Acquire a training set, the training set comprising a plurality of question triples; wherein one of the question triples comprises a basic question, a positive sample of the basic question, and a negative sample, the solution expression of the positive sample is similar to the solution expression of the basic question, and the solution expression of the negative sample is different from the solution expression of the basic question; For any question triple in the training set, an encoder based on the problem-solving model encodes each type of question in the question triple to obtain encoding vectors of each type of question in the question triple; The decoder based on the problem-solving model decodes the encoding vectors of each type of question in the question triple to obtain an expression tree of each type of question in the question triple; wherein the expression tree describes the predicted solution expressions of each type of question in a tree structure; Based on the expression trees of various types of questions in the multiple question triples, a first loss function is constructed; wherein the first loss function is used to reduce the distance between the encoding vector of the basic question in any question triple and the encoding vector of the positive sample, and increase the distance between the encoding vector of the basic question in any question triple and the encoding vector of the negative sample; Constructing a second loss function based on the expression tree of at least one type of question in the multiple question triples and the corresponding standard solution expression; Training the problem-solving model based on the first loss function and the second loss function; Obtain a target question, and generate a solution expression for the target question based on the trained problem-solving model.

2. The method according to claim 1, characterized in that The constructing a first loss function based on expression trees of various types of questions in the multiple question triples includes: For any question triple, obtain the node vector of the first expression tree of the basic question in the question triple, the node vector of the second expression tree of the positive sample in the question triple, and the node vector of the third expression tree of the negative sample in the question triple; wherein the node vector refers to the feature vector of the node in the corresponding expression tree; Obtaining a first cosine similarity between a node vector of the first expression tree and a node vector of the second expression tree; Obtaining a second cosine similarity between the node vector of the first expression tree and the node vector of the third expression tree; Determining a loss value of the question triple based on the first cosine similarity and the second cosine similarity; The first loss function is constructed based on the loss values ​​of the multiple question triplets.

3. The method according to claim 1, characterized in that The solution expression of the positive example is similar to the solution expression of the basic question, including: The solution expression of the positive example is partially the same as the solution expression of the basic question; or, The solution expression of the positive example is exactly the same as the solution expression of the basic question.

4. The method according to claim 1 or 3, characterized in that: The method further comprises: In response to the existence of a subtree in the expression tree of the first question being consistent with the first expression tree of the basic question, taking the first question as a positive example of the basic question; In response to the expression tree of the second question and the first expression tree of the basic question having the same number of nodes and different operators, the second question is used as a negative sample of the basic question.

5. The method according to claim 1, characterized in that The step of training the problem-solving model based on the first loss function and the second loss function includes: Constructing a target loss function based on the first loss function, the second loss function and a loss weight, wherein the loss weight is used to measure the importance of the second loss function; The problem-solving model is trained with minimizing the target loss function as the training goal.

6. The method according to claim 1, characterized in that The constructing a second loss function based on the expression tree of at least one type of question in the multiple question triples and the corresponding standard solution expression includes: The second loss function is constructed based on the first expression trees and corresponding standard solution expressions of the basic questions in the multiple question triples, and the second expression trees and corresponding standard solution expressions of the positive examples in the multiple question triples.

7. A problem-solving device based on artificial intelligence, characterized in that: The device comprises: The acquisition module is configured to acquire a training set, wherein the training set includes a plurality of question triples; wherein one of the question triples includes a basic question, a positive sample of the basic question, and a negative sample, wherein the solution expression of the positive sample is similar to the solution expression of the basic question, and the solution expression of the negative sample is different from the solution expression of the basic question; An encoding module is configured to encode, for any one of the question triples in the training set, each type of question in the question triple based on an encoder of a problem-solving model, to obtain encoding vectors of each type of question in the question triple; A decoding module is configured to decode the encoding vectors of each type of question in the question triple based on the decoder of the question-solving model to obtain expression trees of each type of question in the question triple; wherein the expression tree describes the predicted solution expressions of each type of question in a tree structure; A training module is configured to construct a first loss function based on expression trees of various types of questions in the multiple question triples; wherein the first loss function is used to reduce the distance between the encoding vector of the basic question in any question triple and the encoding vector of the positive sample, and increase the distance between the encoding vector of the basic question in any question triple and the encoding vector of the negative sample; The training module is further configured to construct a second loss function based on the expression tree of at least one type of question in the plurality of question triples and the corresponding standard solution expression; The training module is further configured to train the problem-solving model based on the first loss function and the second loss function; The generation module is configured to obtain a target question and generate a solution expression for the target question based on the trained problem-solving model.

8. A computer device, characterized in that: The device includes a processor and a memory, wherein the memory stores at least one program code, and the at least one program code is loaded and executed by the processor to implement the artificial intelligence-based problem-solving method as claimed in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: At least one program code is stored in the storage medium, and the at least one program code is loaded and executed by the processor to implement the artificial intelligence-based problem-solving method as claimed in any one of claims 1 to 6.

10. A computer program product, characterized in that The computer program product includes a computer program code, which is stored in a computer-readable storage medium. A processor of a computer device reads the computer program code from the computer-readable storage medium, and the processor executes the computer program code, so that the computer device executes the artificial intelligence-based problem-solving method as described in any one of claims 1 to 6.

Images