JP6854486B2 - Judgment device, judgment method, and program - Google Patents

Description

The present invention relates to a determination device, a determination method, and a program, and more particularly to a determination device, a determination method, and a program for improving the calculation speed of fact prediction of the Knowledge Graph.

The knowledge graph G = (E, F) describes knowledge in relation to each other and represents a set of knowledge in a labeled directed graph. FIG. 5 shows an example of the Knowledge Graph. An entity corresponds to a node, a relationship corresponds to a label of an edge, and an edge is a triplet [s; r; o] of an entity s, o and its relationship r, and is also called a fact. In Knowledge Graph G, F is a set of facts [s; r; o], and E is a set of entities e (any entity e can be either s or o).

Relationship prediction on the Knowledge Graph is automatic prediction of the existence of facts. In general, not all necessary facts are registered in the Knowledge Graph, and necessary facts may be missing, for example, the facts shown by the dotted line in FIG. Automatically compensating for such lack of facts based on some score function is an important technique in the construction of knowledge graphs and information retrieval systems.

Designing a highly accurate and high-speed relationship prediction model is an important issue, but Non-Patent Document 1 proposes a model called a neural tensor network (NTNs). This model models the score for the fact [s; r; o] by a mathematical formula such as the following equation (1).

Given a certain unknown fact [s; r; o], if the score of the equation (1) is high, it can be determined to be a correct fact, and conversely, if it is low, it can be determined to be an incorrect fact.

は、関係ｒに対するｎ次元のベクトル、

は、関係ｒに対する３次テンソル、

は、関係ｒに対する行列、

は、エンティティｓ、ｏに対するｄ次元ベクトル、

は、シグモイド関数やｔａｎｈ関数のような非線形関数である。 In equation (1)

Is an n-dimensional vector for the relationship r,

Is a cubic tensor for the relationship r,

Is the matrix for the relation r,

Is a d-dimensional vector for entities s, o,

Is a non-linear function such as a sigmoid function or a tanh function.

Further, the dimension numbers d and n are arbitrarily given. Generally, a number of dimensions smaller than the number of relationships and entities is set.

と表記する）を、下記式（２）で最適となるように学習する。 In NTNs, Knowledge Graph G is received as input, and the above vectors, matrices and tensors (collectively model parameters)

(Indicated as) is learned so as to be optimal by the following equation (2).

は、負例事実の集合である。 here,

Is a set of negative facts.

を設定する。この学習には、例えば、確率的勾配降下法（非特許文献１）が用いられる。 Equation (2) is a learning standard called margin maximum learning, and is a model so that the score difference between the correct fact [s; r; o] and the negative fact [s';r';o'] is as large as possible. Parameters

To set. For this learning, for example, a stochastic gradient descent method (Non-Patent Document 1) is used.

"Reasoning with Neural Tensor Networks for Knowledge Base Completion", Richard Socher, Danqi Chen, Christopher D. Manning, Andrew Y.Ng, Proceedings of the NIPS, 2013.

のテンソル部分に、

の空間使用量を必要とする。また、そのスコア計算にも、

を必要とする。そのため、ＮＴＮｓは、次元数ｄ、ｎが増大すると、空間使用量も増大し、計算速度が低下するという問題が指摘されている。 In the above NTNs model,

In the tensor part of

Requires space usage. Also, for the score calculation

Need. Therefore, it has been pointed out that NTNs have a problem that when the number of dimensions d and n increases, the amount of space used also increases and the calculation speed decreases.

The present invention has been made in view of the above points, and a determination device, a determination method, and a determination method capable of determining whether an unknown fact is correct or incorrect even in a knowledge graph having a large order with high accuracy and high speed. And the purpose of providing the program.

The determination device according to the present invention is a determination device that determines whether or not the facts of the knowledge graph, which is a labeled directed graph describing a plurality of facts representing the relationships between entities, are correct, and inputs the facts to be determined. A neural network model for calculating the input unit to be accepted and the score used for determining whether or not the knowledge graph is correct as a fact, and each matrix included in the cubic tensor corresponding to the relationship is a real symmetric matrix. Alternatively, it is configured to include a determination unit for determining whether or not the input fact is correct as the fact of the knowledge graph by using a neural network model as a real normal matrix.

Further, the determination method according to the present invention is a determination method for determining whether or not the facts of the knowledge graph, which is a labeled directed graph describing a plurality of facts representing the relationships between entities, are correct, and the input unit determines. A neural network model for calculating a score used for determining whether or not the facts of the target are input and whether or not the determination unit is correct as the facts of the knowledge graph, and the third-order tensor corresponding to the relationship is used. It includes a step of determining whether or not the input fact is correct as the fact of the knowledge graph by using a neural network model in which each included matrix is a real symmetric matrix or a real normal matrix.

According to the determination device and the determination method according to the present invention, the input unit receives the input of the fact to be determined, and the determination unit calculates the score used for determining whether or not the fact is correct in the knowledge graph. Is the input fact correct as the fact of the knowledge graph in the network model using the neural network model in which each matrix included in the cubic tensor corresponding to the relationship is a real symmetric matrix or a real normal matrix? Judge whether or not.

In this way, it is a neural network model for calculating the score used for determining whether or not the facts of the knowledge graph are correct, and each matrix included in the cubic tensor corresponding to the relationship is set to a real symmetric matrix or a real normal. By determining whether the input facts are correct as the facts of the knowledge graph using a neural network model as a matrix, it is possible to determine whether the unknown facts are correct or incorrect even in a knowledge graph having a large order. It can be judged with high accuracy and high speed.

Further, the neural network model according to the present invention can be expressed by the following equation.

は関係ｒに対するｎ次元のベクトルであり、

は関係ｒに対する行列であり、

はエンティティｓ、ｏに対するｄ次元ベクトルであり、

における

は、実対称行列を直交行列を用いてスペクトル分解することにより得られる対角行列であり、

は非線形関数であり、

は、

を含むモデルパラメータである。 However, s and o are entities, and r is the relationship between s and o.

Is an n-dimensional vector for the relationship r

Is a matrix for the relation r

Is a d-dimensional vector for entities s, o,

In

Is a diagonal matrix obtained by decomposing a real symmetric matrix using an orthogonal matrix.

Is a non-linear function

Is

It is a model parameter including.

Further, the neural network model according to the present invention can be expressed by the following equation.

は関係ｒに対するｎ次元のベクトルであり、

は関係ｒに対する行列であり、

はエンティティｓ、ｏに対する２ｄ次元ベクトルであり、

における

は、実正規行列をユニタリ行列を用いてスペクトル分解することにより得られる複素数を対角成分に持つ対角行列であり、

は非線形関数であり、

は、

を含むモデルパラメータである。 However, s and o are entities, and r is the relationship between s and o.

Is an n-dimensional vector for the relationship r

Is a matrix for the relation r

Is a 2d dimensional vector for entities s, o,

In

Is a diagonal matrix having complex numbers as diagonal components obtained by spectrally decomposing a real normal matrix using a unitary matrix.

Is a non-linear function

Is

It is a model parameter including.

The program according to the present invention is a program for functioning as each part of the above-mentioned determination device.

The model parameter learning device according to the present invention is based on an input unit that accepts input of a knowledge graph, which is a labeled directed graph that describes a plurality of facts representing relationships between entities, and the knowledge graph input to the input unit. , To calculate a negative case data generation unit that generates negative case data including at least one negative case fact, and a score used for determining whether or not the fact to be determined is a correct fact in the knowledge graph. The knowledge graph and the negative example data generation unit generate a neural network model in which each matrix included in the cubic tensor corresponding to the relationship is a real symmetric matrix or a real normal matrix. It is configured to include a model learning unit for learning using the above-mentioned negative example data.

Further, in the model parameter learning method according to the present invention, the input unit receives input of a knowledge graph which is a labeled directed graph describing a plurality of facts representing relationships between entities, and the negative example data generation unit is described above. Based on the knowledge graph input to the input unit, the step of generating negative example data including at least one negative example fact and the fact that the model learning unit determines are the correct facts in the knowledge graph. A neural network model for calculating a score used for determining whether or not there is a neural network, in which each matrix included in the cubic tensor corresponding to the above relationship is a real symmetric matrix or a real normal matrix. The knowledge graph includes a step of learning using the negative example data generated by the negative example data generation unit.

According to the model parameter learning device and the model parameter learning method according to the present invention, the input unit receives input of a knowledge graph which is a labeled directed graph describing a plurality of facts representing relationships between entities, and a negative example data generation unit. However, based on the knowledge graph input to the input unit, negative example data including at least one fact that is a negative example is generated.

Then, the model learning unit is a neural network model for calculating a score used for determining whether or not the fact to be determined is a correct fact in the knowledge graph, and is included in the cubic tensor corresponding to the relationship. A neural network model in which each matrix is a real symmetric matrix or a real normal matrix is learned using a knowledge graph and negative example data generated by a negative example data generation unit.

In this way, it is a neural network model for calculating the score used for determining whether or not it is a correct fact in the knowledge graph, and each matrix included in the cubic tensor corresponding to the relationship is a real symmetric matrix or a real. By learning the neural network model as a normal matrix using the knowledge graph and the negative example data generated by the negative example data generator, whether the unknown fact is correct or incorrect even in the knowledge graph with many orders. It is possible to learn model parameters for determining with high accuracy and high speed.

According to the determination device, determination method, and program according to the present invention, it is possible to determine whether an unknown fact is correct or incorrect with high accuracy and high speed even in a knowledge graph having a large order.

It is a block diagram which shows the structure of the model parameter learning apparatus which concerns on embodiment of this invention. It is a flowchart which shows the model parameter learning processing routine of the model parameter learning apparatus which concerns on embodiment of this invention. It is a block diagram which shows an example of the structure of the determination apparatus which concerns on embodiment of this invention. It is a flowchart which shows the determination processing routine of the determination apparatus which concerns on embodiment of this invention. It is a figure which shows the example of the knowledge graph.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.

<Principle of model parameter learning device according to the embodiment of the present invention>
First, the principle of the embodiment of the present invention will be described.

に制限を設けることで、新しいニューラルネットワークモデルを２つ提案し、空間使用量と計算速度の問題を解決したことである。ここでは、２つのモデルを、それぞれニューラルスペクトラルテンソルネットワーク１、２（ＮＳＴＮｓ−１，２；Ｎｅｕｒａｌ Ｓｐｅｃｔｒａｌ Ｔｅｎｓｏｒ Ｎｅｔｗｏｒｋｓ−１，２）と呼ぶ。 The result in the embodiment of the present invention is each slice (matrix) of the tensor portion in the NTNs model.

By setting a limit on, we proposed two new neural network models and solved the problems of space usage and calculation speed. Here, the two models are referred to as neural spectral tensor networks 1 and 2 (NSTNs-1,2; Natural Spectral Tensor Networks-1,2), respectively.

を使って、下記式（３）として対角化できることが知られている。 << NSTNs-1 model >>
In the NSTNs-1 model, each matrix (slice) of the tensor part in the equation (1) is limited to a real symmetric matrix. Real symmetric matrix is orthogonal matrix by spectral decomposition

It is known that can be diagonalized as the following equation (3) using.

に対して、テンソル部分の各スライスを下記式（４）として変形することできる。 From equation (3) to equation (1)

On the other hand, each slice of the tensor portion can be transformed by the following equation (4).

の空間使用量を、

とすることができる。 By transforming as equation (4), the tensor

Space usage of

Can be.

が対角成分のみから成るため、実際は

の空間使用量に制限することができる。そのため、ＮＴＮｓの空間使用量、計算速度の問題を緩和することができる。 That is, each slice

Is actually composed of only diagonal components

Can be limited to the amount of space used. Therefore, the problems of the space usage of NTNs and the calculation speed can be alleviated.

The formula of the NSTNs-1 model can be expressed as the following formula (5).

は、関係ｒに対するｎ次元のベクトル、

は、関係ｒに対する３次テンソル（ただし、

に対して

は対角行列）、

は、関係ｒに対する行列、

は、エンティティｓ、ｏに対するｄ次元ベクトル、

は、シグモイド関数やｔａｎｈ関数のような非線形関数とする。 In equation (5)

Is an n-dimensional vector for the relationship r,

Is a cubic tensor for the relation r (provided

Against

Is a diagonal matrix),

Is the matrix for the relation r,

Is a d-dimensional vector for entities s, o,

Is a non-linear function such as a sigmoid function or a tanh function.

、及び、行列

のパラメータをまとめて、モデルパラメータ

とする。 Also vector

, And the matrix

The parameters of the model are summarized.

And.

を使って、下記式（６）として対角化できることが知られている。 << NSTNs-2 model >>
In the second model, each matrix of the tensor part in the equation (1) is limited to a real normal matrix. A real normal matrix is a class matrix with less restrictions than a real symmetric matrix, and can handle asymmetric relationships. In addition, the real symmetric matrix is also a unitary matrix by spectral decomposition.

It is known that can be diagonalized as the following equation (6) using.

は、ユニタリ行列、

は、複素数を対角成分に持つ対角行列とする。 here,

Is a unitary matrix,

Is a diagonal matrix having complex numbers as diagonal components.

と、非線形関数

とは、同一の記号（γ）を一部に用いて表されているが、異なるものである。 Here, the diagonal matrix

And the nonlinear function

Is expressed using the same symbol (γ) in part, but is different.

は、複素数の実部だけをとる操作であるとする。 In addition, it should be noted.

Is an operation that takes only the real part of a complex number.

及び、

を用いることができる。これにより、式（１）の

に対して、テンソル部分の各スライスを下記式（７）として変形することで、テンソル

の空間使用量を

とすることができる。 Also, as an operation to convert a complex matrix to a real matrix

as well as,

Can be used. As a result, the formula (1)

On the other hand, by transforming each slice of the tensor part as the following equation (7), the tensor

Space usage

Can be.

Similar to the NSTNs-1 model, this model can alleviate the problems of space usage and calculation speed of NTNs. On the other hand, it is more expressive than the NSTNs-1 model and is suitable for modeling asymmetric relationships that appear in the Knowledge Graph. This is because the NSTNs-1 model considers only symmetric relationships in the tensor portion.

Finally, the equation of the NSTNs-2 model can be expressed as the following equation (8).

は、関係ｒに対するｎ次元のベクトル、

は、関係ｒに対する３次テンソル（ただし、

に対して、

は、２×２ブロック対角行列）、

は、関係ｒに対する行列、

は、エンティティｓ、ｏに対する２ｄ次元ベクトル、

は、シグモイド関数やｔａｎｈ関数のような非線形関数とする。 In equation (8)

Is an n-dimensional vector for the relationship r,

Is a cubic tensor for the relation r (provided

Against

Is a 2x2 block diagonal matrix),

Is the matrix for the relation r,

Is a 2d vector for entities, o,

Is a non-linear function such as a sigmoid function or a tanh function.

、及び、行列

のパラメータをまとめて、モデルパラメータ

とする。 Also vector

, And the matrix

The parameters of the model are summarized.

And.

<< Learning model parameters >>

を最適化して学習する。 Model parameters of NSTNs-1 and 2 according to the following equation (9)

Optimize and learn.

は、Ｌ１正則化項であり、モデルパラメータの値を０に近づける効果を持ち、オーバーフィッティングの問題を回避できる（参考文献１）。 here,

Is an L1 regularization term, which has the effect of bringing the value of the model parameter close to 0, and can avoid the problem of overfitting (Reference 1).

[Reference 1]
"Regression Shrinkage and Selection via the Lasso", Tibshirani Robert, Journal of the Royal Statistical Society, Series B (Methodological), 267-288, 1996.

は、関係やエンティティなど複数種類のベクトル、行列パラメータを持つが、個々にＬ１正則化を書くと冗長となるため、まとめて

として記述している。上記の式（９）の最適化には、例えば、確率的勾配降下法（非特許文献１）を用いることができる。 Model parameters

Has multiple types of vector and matrix parameters such as relationships and entities, but if you write L1 regularization individually, it will be redundant, so all together.

It is described as. For the optimization of the above equation (9), for example, a stochastic gradient descent method (Non-Patent Document 1) can be used.

<< Use of model parameters >>

Using the model parameters learned by Eq. (9), the score of the NSTNs-1 model (Equation (5)) or NSTNs-2 model (Equation (8)) for the input fact F is calculated. Predict the presence or absence of facts based on the calculated score. For example, the presence or absence of facts is determined based on whether or not a predetermined threshold value is exceeded.

を必要としたが、式（５）では、上記の通り空間使用量が、

であるため、式（５）におけるスコア計算も

に抑えることができる。 Equation (1) is used in score calculation.

However, in equation (5), the amount of space used is as described above.

Therefore, the score calculation in the formula (5) is also

Can be suppressed to.

と同様に、これを式（４）において変形した

もｄ次元ベクトルであるため、この点においてスコア計算の速度に変化はない。
よって、式（５）では、式（１）に比してスコア計算の高速化を実現することができることがわかる。同様に、式（８）も式（１）に比してスコア計算の高速化を実現することができる。 Also, in equation (1)

Similarly, this was modified in equation (4).

Is also a d-dimensional vector, so there is no change in the speed of score calculation at this point.
Therefore, it can be seen that the equation (5) can realize a higher speed of score calculation than the equation (1). Similarly, the equation (8) can realize a higher speed of score calculation than the equation (1).

Therefore, by using NSTNs, it is possible to determine whether an unknown fact is correct or incorrect with higher accuracy and higher speed than NTNs. As a result, it leads to the construction of a knowledge graph (addition of new facts) and the improvement of the accuracy of information retrieval.

<Structure of model parameter learning device according to the embodiment of the present invention>
The configuration of the model parameter learning device according to the embodiment of the present invention will be described with reference to FIG. FIG. 1 is a block diagram showing a configuration of a model parameter learning device according to an embodiment of the present invention.

The model parameter learning device 10 is composed of a computer including a CPU, a RAM, and a ROM that stores a program for executing an action selection learning processing routine described later, and is functionally configured as shown below. ing.

As shown in FIG. 1, the model parameter learning device 10 according to the present embodiment includes an input unit 100, a calculation unit 200, and an output unit 300.

The input unit 100 accepts input of a knowledge graph which is a labeled directed graph describing a plurality of facts representing relationships between entities.

Specifically, the input unit 100 receives the input knowledge graph G = (E, F).

Then, the input unit 100 passes the input knowledge graph G to the negative example data generation unit 210, and passes the input model to the model selection unit 220.

The calculation unit 200 is a neural network model for calculating a score used for determining whether or not the fact to be determined is a correct fact in the Knowledge Graph, and each matrix included in the cubic tensor corresponding to the relationship. Is learned as a real symmetric matrix or a real normal matrix using a plurality of facts described by the knowledge graph and negative example data.

Specifically, the calculation unit 200 includes a negative example data generation unit 210, a model selection unit 220, and a model learning unit 230.

The negative example data generation unit 210 generates negative example data including at least one fact that is a negative example based on the knowledge graph input to the input unit 100.

、及び、任意のエンティティ

に対して、［ｅ，ｒ，ｏ］、及び、［ｓ，ｒ，ｅ］の事実を生成し、サンプリングしたものを負例事実の集合

に追加する。負例の定義、及び負例の生成方法は、任意の方法を採用することができ、例えば、非特許文献１の負例サンプリング法を用いることができる。 Specifically, the negative example data generation unit 210 includes all the facts included in the set F of the input knowledge graph G.

, And any entity

On the other hand, the facts of [e, r, o] and [s, r, e] are generated, and the sampled one is a set of negative facts.

Add to. Any method can be adopted as the definition of the negative example and the method of generating the negative example, and for example, the negative example sampling method of Non-Patent Document 1 can be used.

The model selection unit 220 accepts the user's selection and, as a function used in the model learning unit 230, is a neural network for calculating a score used for determining whether or not the fact to be determined is a correct fact in the knowledge graph. A neural network model is selected in which each matrix included in the cubic tensor corresponding to the relationship is a real symmetric matrix or a real normal matrix.

として、上記式（５）（ＮＳＴＮｓ−１モデル）、及び式（８）（ＮＳＴＮｓ−２モデル）の何れかを、選択する。例えば、モデル選択部２２０は、入力部１００に入力されたユーザの選択に応じて、学習するモデルを選択する。 Specifically, the model selection unit 220 is a function used in the model learning unit 230.

As the above formula (5) (NSTNs-1 model) and formula (8) (NSTNs-2 model) are selected. For example, the model selection unit 220 selects a model to be learned according to the user's selection input to the input unit 100.

Further, the model selection unit 220 may be configured so that the NTNs model of the formula (1) can be selected.

Then, the model selection unit 220 passes the selected model to the model learning unit 230.

The model learning unit 230 learns the neural network model selected by the model selection unit 220 using a plurality of facts described by the knowledge graph and the negative example data generated by the negative example data generation unit 210.

の各パラメータを、モデル選択部２２０により取得したＮＳＴＮｓモデルに応じて初期化する等の必要な処理を行う。 Specifically, first, the model learning unit 230 has model parameters.

Perform necessary processing such as initializing each parameter of the above according to the NSTNs model acquired by the model selection unit 220.

を最適化して学習する。最適化には、例えば、確率的勾配降下法（非特許文献１）を用いることができる。 Next, the model learning unit 230 uses the model parameters of the NSTNs model (NSTNs-1 model or NSTNs-2 model) selected by the model selection unit 220 according to the equation (9).

Optimize and learn. For optimization, for example, a stochastic gradient descent method (Non-Patent Document 1) can be used.

を最適化して学習してもよい。ここで、式（１）のモデルパラメータ

を式（９）にしたがって最適化して学習した場合には、Ｌ１正則化項を用いるため、オーバーフィッティングの問題を回避することが可能である。 When the equation (1) (NTNs model) is selected, the model parameters are according to any of the equations (2) and (9).

May be optimized and learned. Here, the model parameter of the equation (1)

When is learned by optimizing according to the equation (9), the problem of overfitting can be avoided because the L1 regularization term is used.

とを、出力部３００に渡す。 Then, the model learning unit 230 uses the selected model and the learned model parameters.

Is passed to the output unit 300.

を出力する。 The output unit 300 is a model parameter learned by the model learning unit 230.

Is output.

<Operation of model parameter learning device according to the embodiment of the present invention>
FIG. 2 is a flowchart showing a model parameter learning processing routine according to the embodiment of the present invention.

When the knowledge graph is input to the input unit 100, the model parameter learning device 10 executes the model parameter learning processing routine shown in FIG.

First, in step S100, the input unit 100 accepts input of a knowledge graph which is a labeled directed graph describing a plurality of facts representing relationships between entities.

In step S110, the negative example data generation unit 210 generates negative example data including at least one fact that is a negative example based on the knowledge graph input in step S100.

In step S120, the model selection unit 220 accepts the user's selection and calculates a score used for determining whether or not the fact to be determined is a correct fact in the knowledge graph as a function used by the model learning unit 230. For this purpose, a neural network model is selected in which each matrix included in the cubic tensor corresponding to the relationship is a real symmetric matrix or a real normal matrix.

In step S130, the model learning unit 230 initializes each parameter of the model parameter of the neural network model selected in step S120.

In step S140, the neural network model selected in step S120 is learned using the plurality of facts described by the knowledge graph and the negative example data generated by the negative example data generation unit 210 in step S110.

In step S150, the model parameters learned in step S140 are output.

As described above, according to the model parameter learning device according to the present embodiment, it is a neural network model for calculating a score used for determining whether or not it is a correct fact in a knowledge graph, and corresponds to a relationship. Since a neural network model in which each matrix included in the cubic tensor is a real symmetric matrix or a real normal matrix is learned using a knowledge graph and negative example data generated by the negative example data generator, the order is large. Also in the knowledge graph, it is possible to learn model parameters for determining whether an unknown fact is correct or incorrect with high accuracy and at high speed.

<Structure of the determination device according to the embodiment of the present invention>
The configuration of the determination device according to the embodiment of the present invention will be described with reference to FIG. FIG. 3 is a block diagram showing a configuration of a determination device according to an embodiment of the present invention.

The determination device 20 is composed of a computer including a CPU, a RAM, and a ROM that stores a program for executing a determination processing routine described later, and is functionally configured as shown below.

As shown in FIG. 3, the determination device 20 according to the present embodiment includes an input unit 400, a calculation unit 500, and an output unit 600.

The input unit 400 accepts the input of facts to be determined as to whether or not the facts of the Knowledge Graph are correct.

Specifically, the input unit 400 accepts facts (s, r, o) to be determined as to whether or not the facts of the Knowledge Graph are correct.

Then, the input unit 400 passes the input facts (s, r, o) to the determination unit 510.

The calculation unit 500 determines whether or not the input facts are correct as the facts of the Knowledge Graph.

Specifically, the calculation unit 500 includes a determination unit 510 and a model storage unit 520.

The determination unit 510 is a neural network model for calculating a score used for determining whether or not the knowledge graph is correct as a fact, and each matrix included in the cubic tensor corresponding to the relationship is a real symmetric matrix or a real symmetric matrix. Using a neural network model as a real normal matrix, it is determined whether or not the input fact is correct as a knowledge graph fact.

を取得する。ここで、判定に用いるニューラルネットワークモデルとして、ＮＴＮｓモデル（式（１））、ＮＳＴＮｓ−１モデル（式（５））、又はＮＳＴＮｓ−２モデル（式（８）））のうち、モデルパラメータ学習装置１０によって学習済みのモデルを取得する。学習済みのモデルが複数ある場合には、任意に選択できるように構成されてもよい。 Specifically, first, the determination unit 510 is a neural network model learned by the model parameter learning device 10 from the model storage unit 520.

To get. Here, as the neural network model used for the determination, among the NTNs model (Equation (1)), NSTNs-1 model (Equation (5)), or NSTNs-2 model (Equation (8)), the model parameter learning device Acquire the trained model by 10. When there are a plurality of trained models, they may be configured so that they can be arbitrarily selected.

を用いて、スコアを算出する。 Next, the determination unit 510 acquires the acquired neural network model for the input facts (s, r, o).

Is used to calculate the score.

Then, the determination unit 510 determines that the input facts (s, r, o) are correct as the facts of the Knowledge Graph when the calculated score is equal to or higher than the predetermined threshold value. Further, the determination unit 510 determines that the input facts (s, r, o) are not correct as the facts of the Knowledge Graph when the calculated score is not equal to or higher than the predetermined threshold value.
Then, the determination unit 510 outputs the determination result to the output unit 600.

The output unit 600 outputs the determination result of the determination unit 510.

<Operation of the determination device according to the embodiment of the present invention>
FIG. 4 is a flowchart showing a determination processing routine according to the embodiment of the present invention.

When a fact is input to the input unit 400, the determination device 20 executes the determination processing routine shown in FIG.

First, in step S200, the input unit 400 accepts the input of facts.

In step S210, the determination unit 510 acquires the neural network model learned by the model parameter learning device 10 from the model storage unit 520.

In step S220, the determination unit 510 calculates a score for the fact input in step S200 using the neural network model acquired in step S210.

In step S230, the determination unit 510 determines that the input fact is correct as the fact of the Knowledge Graph when the score calculated in step S220 is equal to or higher than the predetermined threshold value. Further, the determination unit 510 determines that the input fact is not correct as a fact of the Knowledge Graph when the calculated score is not equal to or more than a predetermined threshold value.

In step S240, the output unit 600 outputs the determination result of the determination unit 510.

As described above, according to the determination device according to the present embodiment, it is determined whether or not the input fact is correct as the fact of the Knowledge Graph by using the neural network model learned by the model parameter learning device 10. Therefore, even in a knowledge graph having a large order, it is possible to determine whether an unknown fact is correct or incorrect with high accuracy and high speed.

The present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

Further, in the present specification, although the program has been described as a pre-installed embodiment, it is also possible to provide the program by storing it in a computer-readable recording medium.

10 Model parameter learning device 20 Judgment device 100 Input unit 200 Calculation unit 210 Negative example data generation unit 220 Model selection unit 230 Model learning unit 300 Output unit 400 Input unit 500 Calculation unit 510 Judgment unit 520 Model storage unit 600 Output unit

Claims (5)

It is a judgment device that judges whether or not it is correct as a fact of a knowledge graph, which is a labeled directed graph that describes a plurality of facts representing relationships between entities.
An input unit that accepts the input of facts to be judged,
A neural network model for calculating a score used for determining whether or not the knowledge graph is correct as a fact, and each matrix included in the cubic tensor corresponding to the relationship is referred to as a real symmetric matrix or a real normal matrix. Using the neural network model, a determination unit that determines whether or not the input fact is correct as the fact of the knowledge graph, and
Judgment device including. The determination device according to claim 1, wherein the neural network model is represented by the following equation.

However, s and o are entities, and r is the relationship between s and o.

Is an n-dimensional vector for the relationship r

Is a matrix for the relation r

Is a d-dimensional vector for entities s, o,

In

Is a diagonal matrix obtained by decomposing a real symmetric matrix using an orthogonal matrix.

Is a non-linear function

Is

It is a model parameter including. The determination device according to claim 1, wherein the neural network model is represented by the following equation.

However, s and o are entities, and r is the relationship between s and o.

Is an n-dimensional vector for the relationship r

Is a matrix for the relation r

Is a 2d dimensional vector for entities s, o,

In

Is a diagonal matrix having complex numbers as diagonal components obtained by spectrally decomposing a real normal matrix using a unitary matrix.

Is a non-linear function

Is

It is a model parameter including. It is a judgment method for judging whether or not it is correct as a fact of a knowledge graph, which is a labeled directed graph that describes a plurality of facts representing relationships between entities.
The step in which the input unit accepts the input of the fact to be judged,
A neural network model for calculating a score used by the determination unit to determine whether or not the knowledge graph is correct as a fact, and each matrix included in the cubic tensor corresponding to the relationship is a real symmetric matrix or Using a neural network model as a real normal matrix, the step of determining whether or not the input fact is correct as the fact of the knowledge graph, and
Judgment method including. A program for causing a computer to function as each part of the determination device according to any one of claims 1 to 3.

Images