KR102567937B1 - Method using inductive tree-structured neural model based on natural language processing for solving math word problems and system thereof - Google Patents

Abstract

A method for solving a descriptive mathematical problem using an inductive mathematical tree model based on natural language processing is disclosed. According to one aspect of the present invention, the tokenization unit tokenizes a narrative mathematical problem for problem solving, and the solver converts the tokenized mathematical problem into a formula tree of a binary tree logic structure according to an inductive problem solving sequence. Provided is a method for solving a descriptive mathematical problem using an inductive formula tree model based on natural language processing, including a step of performing substitution and solution.

Description

Method and system for solving descriptive mathematical problems using an inductive formula tree model based on natural language processing

The present invention relates to a method and system for solving a descriptive mathematical problem using an inductive mathematical tree model based on natural language processing.

Artificial intelligence (AI) technology is a technology for implementing human cognitive ability, learning ability, reasoning ability, language ability, etc. into a computer. After the advent of deep learning and machine learning, artificial intelligence (AI) technology It has developed rapidly, especially in visual intelligence, and in relation to linguistic intelligence, it has succeeded in developing a human-like text analysis model through extensive prior learning.

However, in terms of reasoning intelligence that thinks and judges like humans through human descriptive language and context understanding, great achievements have not been made. Acquisition of knowledge is still limited.

Therefore, research that uses artificial intelligence to understand human language and solve descriptive mathematical problems through mathematical thinking is a technology that combines common sense understanding-based reasoning, natural language processing, and graph reasoning, and complex intelligence technologies, and is widely used in the artificial intelligence industry. It can be said to be a very challenging task.

Republic of Korea Patent Publication No. 10-2020-0003329 (2020.01.09. Publication)

The present invention provides a method and system for solving a descriptive mathematical problem using an inductive mathematical tree model based on natural language processing.

According to one aspect of the present invention, the tokenization unit tokenizes a narrative mathematical problem for problem solving, and the solver converts the tokenized mathematical problem into a formula tree of a binary tree logic structure according to an inductive problem solving sequence. Provided is a method for solving a descriptive mathematical problem using an inductive formula tree model based on natural language processing, including a step of performing substitution and solution.

The tokenizing step may include converting the math problem into a plurality of tokens based on spacing, but separately converting numbers in the math problem into NUM tokens, and storing information about the NUM tokens as a NUM list. .

The storing as a NUM list may include storing the numbers converted into NUM tokens as a first NUM list, and storing the positions of the NUM tokens in the math problem as a second NUM list.

The step of replacing and solving with an expression tree is a step of generating an embedding result by the embedding unit pre-processing the token and then embedding it using a pre-learning language model, and encoding the embedding result by the encoding unit using a recurrent neural network to obtain a first encoding result. It may include generating a second encoding result by extracting a part for a number from the first encoding result.

In the step of replacing and solving with an equation tree, after the step of generating the second encoding result, the prediction generation unit analyzes the first encoding result and the second encoding result, and the node values of the nodes constituting the expression tree according to the prefix notation. The method may further include solving the math problem by generating a formula tree downward while determining a number or operator that becomes a number or an operator, and sequentially merging a plurality of nodes while recursing upward along the formula tree.

Solving the mathematical problem includes calculating embedding scores for embedding numbers and operators from the first encoding result and the second encoding result, and taking a logistic regression function on the embedding score to obtain a number or operator with the highest embedding score. A step of determining the node value may be included.

In the step of solving a mathematical problem, after the step of determining a number or operator as a node value, if the node value is determined as an operator, the step of generating a pair of child nodes and determining the node values of the child nodes is further added. can include

The step of determining the node value of the child node may include extracting a context value through contextual understanding of the node value of the child node using the first encoding result, the second encoding result, and the attention mechanism, and the predetermined node value. and determining the node value of the child node by combining the context value and the step of determining the node value of the child node.

In the step of solving the mathematical problem, after the step of determining the node value of the pair of child nodes, when the node value of any one of the pair of child nodes is determined as a number and a leaf node is formed, any one of the pair of child nodes and determining a node value of another one of the pair of child nodes through a node value of a parent node of the child node.

In the step of solving the mathematical problem, after the step of determining the value of another node, when the value of the other one of the pair of child nodes is determined as a number, the pair of child nodes and the parent node are merged. A step of sequentially merging a plurality of node values may be further included.

According to the present invention, a descriptive mathematical problem can be effectively solved using an inductive mathematical tree model based on natural language processing.

1 is a flowchart illustrating a method for solving a descriptive mathematical problem using an inductive formula tree model based on natural language processing according to an embodiment of the present invention.
2 is a schematic process diagram illustrating a method for solving a descriptive mathematical problem using an inductive formula tree model based on natural language processing according to an embodiment of the present invention.
3 is a schematic process diagram showing merging after generating a formula tree according to node value determination in a descriptive mathematical problem solving method using an inductive formula tree model based on natural language processing according to an embodiment of the present invention.
4 is a configuration diagram illustrating a descriptive mathematical problem solving system using an inductive formula tree model based on natural language processing according to an embodiment of the present invention.

Since the present invention can apply various transformations and have various embodiments, specific embodiments will be illustrated in the drawings and described in detail in the detailed description. However, it should be understood that this is not intended to limit the present invention to specific embodiments, and includes all transformations, equivalents, and substitutes included in the spirit and scope of the present invention. In describing the present invention, if it is determined that a detailed description of related known technologies may obscure the gist of the present invention, the detailed description will be omitted.

Terms such as first and second may be used to describe various components, but the components should not be limited by the terms. These terms are only used for the purpose of distinguishing one component from another.

Terms used in this application are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly dictates otherwise. In this application, the terms "include" or "have" are intended to designate that there is a feature, number, step, operation, component, part, or combination thereof described in the specification, but one or more other features It should be understood that the presence or addition of numbers, steps, operations, components, parts, or combinations thereof is not precluded.

Hereinafter, a descriptive mathematical problem solving method and system 100 using an inductive mathematical tree model based on natural language processing according to the present invention will be described in detail with reference to the accompanying drawings. In the description with reference to the accompanying drawings, the same or Corresponding components are assigned the same reference numerals, and overlapping descriptions thereof will be omitted.

First, a method for solving a descriptive mathematical problem using an inductive formula tree model based on natural language processing according to an embodiment of the present invention will be described.

1 is a flowchart illustrating a method for solving a descriptive mathematical problem using an inductive formula tree model based on natural language processing according to an exemplary embodiment.

According to this embodiment, as shown in FIG. 1, the tokenization unit 110 tokenizes the descriptive math problem for problem solving (S110), and the solution unit 120 tokenized math A method for solving a descriptive mathematical problem using an inductive equation tree model based on natural language processing is provided, including the step of replacing the problem with an equation tree of a binary tree logical structure according to an inductive problem solving sequence (S120).

According to this embodiment, a descriptive mathematical problem can be effectively solved using an inductive mathematical tree model based on natural language processing.

Hereinafter, each step of a descriptive mathematical problem solving method using an inductive formula tree model based on natural language processing according to the present embodiment will be described with reference to FIGS. 1 to 4 .

domestic 1) Jung Young-han, Recurrent Neural Network Exploration Using a Sentence-type Mathematical Problem Solving Model, KAIST Mathematical Sciences Master's Thesis, 2018
2) Woo Chang-hyeop, Kwon Ga-jin, Automatic Solving of Korean Mathematics Sentence Problem, Proceedings of Korean Language and Information Processing Conference, 2018
3) Kyung Seo Ki, Dong Geon Lee, Gahgene Gweon, KoTAB: Korean Template-Based Arithmetic Solver with BERT, 2020
4) https://github.com/monologig/KoELECTRA
5) Sangkyu Lim, Kyungseo Ki, Bugeun Kim, Gajin Kwon, KoEPT: Automatic Solving of Korean Mathematical Sentence Problems Using Transformer-based Generative Model, 2021 Oversea 6) https://math-qa.github.io/math-QA/
7) https://github.com/hendrycks/math/
8) https://github.com/arkilpatel/SVAMP
9) https://paperswithcode.com/task/math-word-problem-solving
10) A. Patel, et. al, Are NLP Models really able to Solver Simple Math Word Problems?, https://arxiv.org/abs/2103.07191
11) J. Zhang, el. al, Graph-to-Tree Learning for solving math word problems, Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2020
12) Z. Xie and S. Sun, A Goal-Driven Tree-Structured Neural Model for Math Word Problems, Proceedings of the 28th International Joint Conference on Artificial Intelligence

Looking at existing studies or data sets on solving descriptive mathematical problems using artificial intelligence technology in Table 1 above, most of them are approaches that change descriptive sentences into one or two equations made up of simple 4 arithmetic operations. Difficulty solving problems that cross bounds or require sequential reasoning.

Therefore, in order to solve various types of descriptive mathematical problems, it will be said that expansion of existing research and new approaches are required. Accordingly, a model capable of recognizing and performing various mathematical functions and operations based on arithmetic operations through this embodiment want to provide

Regarding the mechanism of the natural language processing-based inductive formula tree model according to this embodiment, a person said, “When I put the same number of 74 batteries in a box, there are 8 boxes, and there are 2 left. Find out how many pieces are in a box.” How many did you put in the box?), 2) I will pay attention to the information needed to get the correct answer (74 batteries were put in the same number in the box, so there are 8 boxes and 2 left).

More specifically, a person 1) 'there are 74 batteries in total', 2) 'because the number of remaining batteries in the box is 2', 3) 'subtract the number of remaining batteries from the total number of batteries, 4) the total box If you divide by the number of , you will think that 'the result of 4) is the correct answer', and this process can be represented as a formula tree structure as shown in Figure 1 below.

[Figure 1]

As described above, this model is a model made with the same or similar structure to the mechanism that solves a problem by inductively decomposing the goal of the problem when a person solves a mathematical problem. To help understand the present embodiment by explaining and providing specific details or examples thereof.

In step 110, the tokenization unit 110 may tokenize the narrative mathematical problem to solve the problem.

Here, tokenization is one of text preprocessing processes, through which material data for solving descriptive mathematical problems can be prepared.

An example of tokenization of open-ended math problems is shown in Table 2 below.

problem There are 3 boxes of 6 ping-pong balls each. If you distribute these ping-pong balls equally to 9 people, how many will each person get?
(Answer formula: 6*3/9) Tokenization [(['Ping-pong balls', 'NUM', 'each', 'in', 'there', 'box', 'NUM', 'box', 'there is', 'this', 'ping-pong balls ', 'NUM', 'to a person', 'equally', 'divide', 'if given,', 'one', 'one', 'how many', 'each', 'have', 'will it?'] , [/* N0 N1 N2], ['6', '3', '9'], [1, 6, 11])]

Specifically, the tokenizing step (S110) converts the mathematical problem into a plurality of tokens based on spaces, but converts the numbers in the mathematical problem separately into NUM tokens (S111), and the information about the NUM token into a NUM list It may include a step of storing as (S112).

As shown in Table 2 above, the numbers (6, 3, 9) of the given problem are converted to NUM tokens ('NUM'), and information about NUM tokens is ['6', '3', '9'], It can be stored as a NUM list like [1, 6, 11].

The step of storing as a NUM list (S112) includes the step of storing the number converted into a NUM token as a first NUM list (S112a), and the step of storing the position of the NUM token in a mathematical problem as a second NUM list (S112b) can include

As shown in Table 2 above, the numbers (6, 3, 9) converted to NUM tokens are stored as the first NUM list as ['6', '3', '9'], and in the mathematical problem of NUM tokens Positions can be stored as a second NUM list, such as [1, 6, 11].

The tokenizing step (S110) is, after the step of converting to NUM tokens (S111), if the math problem is provided for learning, converting the number in the correct answer formula provided with the math problem into NUM tokens (S114) In addition, after the tokenizing step (S110), a step of substituting the converted correct answer formula into a formula tree (S130) may be further included in order to progress learning through comparison of formula trees.

As shown in Table 2 above, the numbers in the correct answer formula provided with the learning math problem are converted to NUM tokens as in [/ * N0 N1 N2], and the converted answer formula [N0*N1/N2] is It can be replaced with an equation tree so that it can be compared with the equation tree inferred by this model for the learning progress of .

In the tokenizing step (S110), after the step of converting the number in the correct answer formula into a NUM token (S114), a constant that appears more than a set number of times in a plurality of math problems as a number included in the correct answer formula but not included in the math problem A step of storing the constant as a list so that it can be used when solving problems (S115) may be further included.

For example, if a mathematical problem for learning is given, “Find the circumference of a circle with a radius of 3”, the circumference of 3.14 is included in the correct answer formula but not included in the mathematical problem, and is a kind of constant (generate_num(extra_num)). Assuming that it appears more than a set number of times (specifically, 2 to 6 times, preferably 4 times) in a math problem, it is stored as a constant list (generate_num(extra_num) list) and can be used for problem solving using this model later.

2 is a schematic process diagram illustrating a method for solving a descriptive mathematical problem using an inductive formula tree model based on natural language processing according to the present embodiment.

In step 120, the solver 120 may solve the tokenized mathematical problem by replacing it with a formula tree of a binary tree logical structure according to an inductive problem solving sequence.

That is, the solver 120 may replace the tokenized mathematical problem with an equation tree having a binary tree logic structure as shown in Figure 1 to solve the tokenized mathematical problem according to an inductive problem solving sequence such as the above-described example in which a person solves a problem.

The step of replacing the solution with an equation tree (S120) is the step of generating an embedding result by the embedding unit 130 pre-processing the token and then embedding it using a pre-learning language model to generate an embedding result (S121), and the encoding unit 140 ) Encodes the embedding result using a recurrent neural network to generate a first encoding result (encoder outputs), but extracts a part for numbers from the first encoding result to generate a second encoding result (number encoder outputs) (S122 ) may be included.

Here, as the pre-learning language model, ELECTRA (efficiently learning an encoder that classifies token replacements accurately) may be specifically used, and more specifically, KoELECTRA, a pre-learning language model pre-trained with Korean data based on the ELECTRA model, may be used. .

ELECTRA is a natural language processing model learned by looking at a token from a generator and discriminating whether it is a true token or a false token from a discriminator.

In the test of this model, koelectra-base-v3-discriminator was specifically used as a pretraining language model.

In addition, the recurrent neural network used for encoding may specifically be a bi-directional long short-term memory (LSTM), and more specifically, it may be a bi-directional LSTM with two layers and a hidden size of 512 there is.

In the step of generating an embedding result (S121), the token is re-divided into morpheme units and padded according to the set length (S121a), the padded token includes a NUM token, but a plurality of settings each having an indexing value It may include matching words (vocab) and converting them into indexing values (S121b), and embedding the indexing values according to the embedding size of the pre-learning language model (S121c).

3 is a schematic process diagram showing merging after generating a formula tree according to node value determination in a descriptive mathematical problem solving method using an inductive formula tree model based on natural language processing according to the present embodiment.

In the step of replacing and solving with an equation tree (S120), after the step of generating the second encoding result (S122), the prediction generating unit 150 analyzes the first encoding result and the second encoding result to obtain a prefix notation. Further comprising a step (S123) of generating a formula tree downward while determining a number or an operator that becomes a node value of a node constituting the formula tree and solving a mathematical problem by sequentially merging a plurality of nodes while recursing upward along the formula tree. can do.

Specifically, the step of solving the mathematical problem (S123) is the step of calculating an embedding score (score) for the embedding of numbers and operators from the first encoding result and the second encoding result (S123a), and logistic regression on the embedding score A step of determining a number or operator having the highest embedding score as a node value by taking a function (S123b).

Here, the logistic regression function may specifically be a log softmax function, and may take the log softmax function and then take the argmax function.

In the step of solving the mathematical problem (S123), after the step of determining a number or operator as a node value (S123b), if the node value is determined as an operator, a pair of child nodes are generated and the node values of the child nodes are generated again. A step of determining (S123c) may be further included.

Specifically, the step of determining the node value of the child node (S123c) is through contextual understanding of the node value of the child node using the first encoding result, the second encoding result, and the attention ((tree) attention) mechanism. It may include extracting a context value (S123d), and determining a node value of a child node by combining a predetermined node value and a context value (S123e).

The step of solving the mathematical problem (S123) is, after the step of determining the node value of the child node (S123c), when the node value of any one of the pair of child nodes is determined as a number and a leaf node is formed, a pair of A step of determining the node value of another one of the pair of child nodes through one of the child nodes of and the node value of the parent node for the child node may be further included (S124).

In this case, as described above, the first encoding result, the second encoding result, and the context value obtained through the contextual understanding of the node value of the other one of the pair of child nodes are extracted and combined using the Attention mechanism to obtain a pair of The value of another node among child nodes can be determined.

In the step of solving the mathematical problem (S123), after the step of determining another node value (S124), when the node value of the other one of the pair of child nodes is determined as a number, the pair of child nodes and A step of sequentially merging a plurality of node values starting with merging of parent nodes (S125) may be further included.

Accordingly, the solution of the provided mathematical problem or the inference of the formula tree may be completed and output.

Hereinafter, in this model that utilizes the binary tree structure as described above, in addition to the binary operators for the four arithmetic operations, operator extensions are prepared to enable operations using unary operators or polynomial operators for more than three terms (operands, variables, and arguments). to explain the content.

In the step of determining the node value of the child node (S123c), when the operator determined as the node value is a unary operator, a node value of any one of a pair of child nodes may be determined as a number set as a dummy.

Here, it can be understood that the number set as a dummy only occupies a place as a node value for simple condition satisfaction and does not affect the operation.

In this way, by mapping unary operators to a binary tree structure, operators usable in this model can be expanded, and accordingly, unary operators such as fac (factorial (!), factorial) can be specifically introduced into this model. there is.

Specifically, for example, the operation of fac(10)(10!) can be mapped to a binary tree structure as shown in Figure 2 below.

[Figure 2]

In the step of determining the node value of the child node (S123c), when the operator determined as the node value is a polynomial operator other than a binary operator, a pair of child nodes are generated by overlapping a pair of child nodes and become a connection node. Any one of the node values can be determined as a unary operator or a binary operator that becomes a unit operator for the operation of a polynomial operator.

In this way, by mapping polynomial operators to a binary tree structure, the operators available for this model can be expanded, and accordingly, the polynomial operators min (minimum value), max (maximum value), sum (sum), len (operand ( Variable, argument) length, self-defined function), gen (n-digit number generation function, self-defined function), etc. can be introduced into this model.

Specifically, for example, the operation of min(1,2,3,4) can be mapped to a binary tree structure as shown in Figure 3 below.

[Figure 3]

On the other hand, in the case of binary operators, nCr (combination), nPr (permutation), nHr (redundant combination) for the number of cases such as ^ (power), @ (quotient), # (remainder), etc., in addition to operators for four arithmetic operations etc. can be additionally introduced into this model.

Hereinafter, in this model that utilizes a binary tree structure, when numbers or letters (names, etc.) that become operands (variables, arguments) of operators are listed in a mathematical problem, a method is described so that they can be used as operands for operators. let it do

The tokenizing step ( S110 ) may further include storing a plurality of characters or a plurality of numbers listed in the mathematical problem as a character list or a number list so that they can be used as operands for an operator ( S113 ).

In this way, a plurality of letters or a plurality of numbers enumerated in a mathematical problem may be stored as a letter list or a number list to be used in the formula reasoning process for the mathematical problem.

For example, the math problem "There are one red marble, one blue marble, and one yellow marble. Find the number of ways in which 2 (NUM, N1) of these marbles can be placed into two boxes of the same shape." If given, 'red marble, blue marble, yellow marble' is stored as the character list NAME LIST SO, and the formula can be inferred as nCr(len(S0), N1) using this.

Next, a descriptive mathematical problem solving system 100 using an inductive formula tree model based on natural language processing according to an embodiment of the present invention will be described.

4 is a block diagram showing a descriptive mathematical problem solving system 100 using an inductive formula tree model based on natural language processing according to the present embodiment.

According to this embodiment, as shown in FIG. 4, the tokenization unit 110 tokenizes a narrative mathematical problem for problem solving, and a binary tree logical structure according to an inductive problem solving sequence for the tokenized mathematical problem. A descriptive mathematical problem solving system 100 using an inductive formula tree model based on natural language processing including a solver 120 that performs a solution by substituting a formula tree is provided.

According to this embodiment, a descriptive mathematical problem can be effectively solved using an inductive mathematical tree model based on natural language processing.

Hereinafter, each component of the descriptive mathematical problem solving system 100 using the natural language processing-based inductive equation tree model according to the present embodiment will be described with reference to FIGS. It is assumed to follow the contents of the descriptive mathematical problem solving method using the tree model.

The tokenization unit 110 may tokenize a narrative mathematical problem to solve the problem.

Specifically, the tokenization unit 110 may convert a math problem into a plurality of tokens based on spacing, but separately convert numbers in the math problem into NUM tokens, and store information about the NUM tokens as a NUM list.

The tokenizing unit 110 may store the numbers converted into NUM tokens as a first NUM list, and may store the positions of the NUM tokens in a math problem as a second NUM list.

When the math problem is provided for learning, the tokenization unit 110 converts the numbers in the correct answer formula provided with the math problem into NUM tokens, and formulates the converted answer formula for learning progress through comparison of formula trees can be replaced with a tree.

The tokenization unit 110 may store constants that appear more than a set number of times in a plurality of mathematical problems as numbers included in the correct answer formula but not included in the mathematical problem as a constant list so that they can be used in solving the problem later.

The solver 120 may solve the tokenized mathematical problem by replacing it with a formula tree of a binary tree logic structure according to an inductive problem solving sequence.

The embedding unit 130 pre-processes the token and embeds it using a pre-learning language model to generate an embedding result, and the encoding unit 140 encodes the embedding result using a recurrent neural network to generate a first encoding result (encoder). outputs), but a second encoding result (number encoder outputs) may be generated by extracting a part for a number from the first encoding result.

The embedding unit 130 re-divides the token into morpheme units, pads the token according to the set length, and matches the padded token with a plurality of set words (vocab) including a NUM token and each having an indexing value. Indexing values, and indexing values can be embedded according to the embedding size of the pre-learning language model.

The prediction generation unit 150 (including the prediction unit 160 and the generation unit 170) analyzes the first encoding result and the second encoding result and becomes the node value of the node constituting the formula tree according to the prefix notation. Mathematical problems can be solved by generating a formula tree downward while determining numbers or operators, and recursing upward along the formula tree, merging multiple nodes sequentially.

Specifically, the prediction unit 160 calculates an embedding score for the embedding of numbers and operators from the first encoding result and the second encoding result, takes a logistic regression function on the embedding score, and calculates the number having the highest embedding score. Alternatively, operators can be determined as node values.

When the node value is determined by the operator by the predictor 160, the generator 170 may generate a pair of child nodes, and the predictor 160 may determine the node value of the child node again.

Specifically, the prediction unit 160 extracts a context value that has undergone contextual understanding of node values of child nodes using a first encoding result, a second encoding result, and an attention ((tree) attention) mechanism. And, a node value of a child node may be determined by combining a predetermined node value and a context value.

When the node value of any one of the pair of child nodes is determined as a number to form a leaf node, the prediction unit 160 determines one of the pair of child nodes and the node value of the parent node for the child node. The node value of the other of the child nodes of the pair can be determined.

When the value of the other one of the pair of child nodes is determined by the predictor 160 as a number, the generation unit 170 starts by merging the pair of child nodes and the parent node to generate a plurality of nodes. Node values can be sequentially merged.

When the operator determined as a node value by the predictor 160 is a unary operator, the predictor 160 may determine a node value of any one of a pair of child nodes as a set number as a dummy.

When the operator determined as the node value by the prediction unit 160 is a polynomial operator other than a binary operator, the generation unit 170 overlaps and creates a pair of child nodes, but the prediction unit 160 connects A node value of any one of a pair of child nodes serving as a node may be determined as a unary operator or a binary operator serving as a unit operator for the operation of a polynomial operator.

The tokenization unit 110 may store a plurality of letters or numbers listed in a mathematical problem as a letter list or a number list so that they can be used as operands for operators.

Meanwhile, the components of the above-described embodiment can be easily grasped from a process point of view. That is, each component can be identified as each process. In addition, the process of the above-described embodiment can be easily grasped from the viewpoint of components of the device.

In addition, the technical contents described above may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer readable medium. The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination. Program commands recorded on the medium may be specially designed and configured for the embodiments or may be known and usable to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. - includes hardware devices specially configured to store and execute program instructions, such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of program instructions include high-level language codes that can be executed by a computer using an interpreter, as well as machine language codes such as those produced by a compiler. A hardware device may be configured to act as one or more software modules to perform the operations of the embodiments and vice versa.

In the above, one embodiment of the present invention has been described, but those skilled in the art can add, change, delete, or add components within the scope not departing from the spirit of the present invention described in the claims. The present invention can be variously modified and changed by the like, and this will also be said to be included within the scope of the present invention.

100: A descriptive mathematical problem solving system using an inductive mathematical tree model based on natural language processing
110: tokenization unit
120: solver
130: embedding unit
140: encoding unit
150: prediction generating unit
160: prediction unit
170: generating unit

Claims (10)

Tokenizing, by a tokenization unit, a narrative mathematical problem to solve the problem; and
A step of solving by a solution unit by substituting the tokenized math problem into a formula tree of a binary tree logic structure according to an inductive problem solving sequence,
The tokenizing step is
converting the mathematical problem into a plurality of tokens based on spaces, but separately converting numbers in the mathematical problem into NUM tokens; and
Storing information about the NUM token as a NUM list;
The step of solving by substituting into the equation tree,
generating, by an embedding unit, an embedding result by pre-processing the token and embedding it using a pre-learning language model; and
An encoding unit encoding the embedding result using a recurrent neural network to generate a first encoding result, and generating a second encoding result by extracting a part for a number from the first encoding result,
The step of solving by substituting into the equation tree,
After generating the second encoding result,
A prediction generation unit analyzes the first encoding result and the second encoding result and generates the formula tree downward while determining a number or operator that becomes a node value of a node constituting the formula tree according to prefix notation, and generates the formula tree A descriptive mathematical problem solving method using an inductive formula tree model based on natural language processing, further comprising recursively upward along and sequentially merging a plurality of nodes to solve the mathematical problem.
delete According to claim 1,
In the step of storing the NUM list,
storing the number converted into the NUM token as a first NUM list; and
A descriptive mathematical problem solving method using an inductive formula tree model based on natural language processing, comprising the step of storing a position of the NUM token in the mathematical problem as a second NUM list.
delete delete According to claim 1,
To solve the mathematical problem,
Calculating embedding scores for embedding numbers and operators from the first encoding result and the second encoding result; and
Descriptive mathematical problem solving method using a natural language processing-based inductive formula tree model, comprising taking a logistic regression function on the embedding score and determining a number or operator having the highest embedding score as a node value.
According to claim 6,
To solve the mathematical problem,
After determining the number or operator as a node value,
When the node value is determined by the operator, a method for solving a descriptive mathematical problem using a natural language processing-based inductive formula tree model, further comprising generating a pair of child nodes and determining node values of the child nodes again.
According to claim 7,
The step of determining the node value of the child node,
extracting a context value through contextual understanding of the node value of the child node using the first encoding result, the second encoding result, and an attention mechanism; and
A method for solving a descriptive mathematical problem using an inductive formula tree model based on natural language processing, comprising determining a node value of the child node by combining a predetermined node value and the context value.
According to claim 7,
To solve the mathematical problem,
After determining the node value of the child node,
When a node value of any one of the pair of child nodes is determined by a number and a leaf node is formed, the pair of the child nodes is determined through one of the pair of child nodes and the node value of a parent node for the child node. A method for solving a descriptive mathematical problem using an inductive formula tree model based on natural language processing, further comprising determining a node value of another one of the nodes.
According to claim 9,
To solve the mathematical problem,
After the step of determining the other node value,
When the node value of another one of the pair of child nodes is determined by a number, natural language processing further comprising sequentially merging a plurality of node values, starting with merging the pair of child nodes and the parent node. A method for solving open-ended mathematical problems using an inductive equation tree model based on