Natural language processing system using force-field calculation, natural language processing method and program

Abstract

To quantitatively evaluate a natural sentence.SOLUTION: A natural language processing system comprises: coordinate setting parts 23a and 23b that for each of a plurality of elements constituting a natural sentence to be analyzed, set virtual coordinates of the elements based on the distributional representation of the elements; energy calculating parts 24a and 24b that by considering the plurality of elements as a plurality of virtual particles placed on the virtual coordinates corresponding to each element and applying a potential function to the plurality of virtual particles to perform force-field calculation, calculate energy indicating an interaction between the virtual particles; and an evaluation value calculation part 28 that calculates an evaluation value indicating the interaction between the plurality of elements constituting the natural sentence to be analyzed.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present disclosure relates to techniques for processing natural language. [Background technology]

[0002] Conventionally, a patent information analysis device has been proposed which comprises: means for searching a set of patent documents based on a predetermined viewpoint; means for extracting a plurality of classification codes assigned to each of the patent documents belonging to the set obtained by the search to create a set of classification codes; means for selecting a classification code for numerical analysis from the set of classification codes; means for calculating the coordinates of the classification code for numerical analysis by numerical analysis; and means for calculating the coordinates of each of the patent documents based on the coordinates of the classification code for numerical analysis, and creating a map which represents the density of the patent documents based on their coordinates (see Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2015-207173 A Summary of the Invention [Problem to be solved by the invention]

[0004] Various techniques for evaluating natural language sentences have been proposed in the past. However, the conventional techniques have room for improvement in terms of quantitatively evaluating natural language sentences.

[0005] In view of the above problems, an object of the present disclosure is to quantitatively evaluate natural language. [Means for solving the problem]

[0006] One example of the present disclosure is a natural language processing system that uses force field calculation, including: a coordinate setting means that sets virtual coordinates of each of a plurality of elements constituting a natural sentence to be analyzed based on a distributed representation of the element; an energy calculation means that regards the plurality of elements as a plurality of virtual particles placed at the virtual coordinates corresponding to each element and calculates energy indicating interactions between the virtual particles by applying a potential function to the plurality of virtual particles to perform force field calculations; and an evaluation value calculation means that calculates an evaluation value indicating interactions between the plurality of elements constituting the natural sentence to be analyzed based on the energies calculated for the plurality of virtual particles.

[0007] Moreover, one example of the present disclosure includes a unit division means for dividing a natural sentence to be analyzed into a plurality of sentence units, an element coordinate setting means for setting, for each of a plurality of elements constituting each of the plurality of sentence units, virtual coordinates of the virtual particles when the element is regarded as a virtual particle based on a distributed representation of the element corresponding to the virtual particle, an inter-element energy calculation means for calculating, for each of the plurality of sentence units, energy indicating an interaction between the virtual particles corresponding to the plurality of elements by regarding the plurality of elements constituting the target sentence unit as a plurality of virtual particles placed at the virtual coordinates corresponding to each element and applying a potential function to the plurality of virtual particles to perform a force field calculation, and The natural language processing system uses force field calculation and includes: a unit coordinate setting means for setting virtual coordinates of a sentence unit based on the virtual coordinates corresponding to the plurality of elements constituting a chapter unit; an inter-unit energy calculation means for treating the plurality of sentence units as a plurality of virtual particles placed at the virtual coordinates corresponding to each sentence unit and calculating energy indicating an interaction between the virtual particles corresponding to the plurality of sentence units by applying a potential function to the plurality of virtual particles and performing force field calculation; and an evaluation value calculation means for calculating an evaluation value of the natural sentence to be analyzed based on the energy indicating the interaction between the virtual particles corresponding to the plurality of elements and the energy indicating the interaction between the virtual particles corresponding to the plurality of sentence units.

[0008] The present disclosure can be understood as an information processing device, a system, a method executed by a computer, or a program executed by a computer. The present disclosure can also be understood as such a program recorded on a recording medium readable by a computer or other device, machine, etc. Here, a recording medium readable by a computer, etc. refers to a recording medium that stores information such as data and programs by electrical, magnetic, optical, mechanical, or chemical action and can be read by a computer, etc. Effect of the Invention

[0009] According to the present disclosure, it becomes possible to quantitatively evaluate natural language sentences. [Brief description of the drawings]

[0010] [Figure 1] 1 is a schematic diagram showing a configuration of an information processing system according to a first embodiment. [Diagram 2] 1 is a diagram illustrating an outline of a functional configuration of an information processing device according to a first embodiment. [Diagram 3] FIG. 2 is a diagram showing a concept in which elements and sentence units constituting natural language sentences are regarded as virtual particles and energy is calculated in the first embodiment. [Figure 4] 4 is a flowchart showing the flow of a natural language analysis process according to the first embodiment. [Diagram 5] 3 is a schematic diagram showing a state in which a plurality of elements to which virtual coordinates are assigned are plotted in a three-dimensional space in the first embodiment. FIG. [Figure 6] FIG. 2 is a schematic diagram showing how a force field calculation is performed by applying a potential function between virtual particles corresponding to a plurality of elements constituting a natural text in the first embodiment. [Figure 7] FIG. 2 is a schematic diagram showing how a sentence unit including a plurality of elements is regarded as a virtual particle including virtual particles corresponding to the plurality of sentence units in the first embodiment. [Figure 8]FIG. 2 is a schematic diagram showing how a force field calculation is performed by applying a potential function between virtual particles corresponding to a plurality of sentence units constituting a natural sentence in the first embodiment. [Figure 9] FIG. 2 is a diagram illustrating an overview of a process for acquiring attributes of elements of natural language using a machine learning model in the first embodiment. [Figure 10] FIG. 11 is a diagram showing an overview of the relationship between the distance between virtual particles and energy in the case where a function for calculating an interatomic potential is adopted as the potential function in the first embodiment. [Figure 11] FIG. 11 is a diagram showing an example of parameters set in a potential function when a function for calculating an interatomic potential is adopted as the potential function in the first embodiment. [Figure 12] FIG. 11 is a diagram illustrating an outline of a functional configuration of an information processing device according to a second embodiment. [Figure 13] 13 is a flowchart showing the flow of a natural language analysis process according to a second embodiment. [Figure 14] FIG. 13 is a diagram illustrating an outline of the functional configuration of an information processing device according to a third embodiment. [Figure 15] 13 is a flowchart showing the flow of a natural language analysis process according to the third embodiment. [Figure 16] 13 is a flowchart showing the flow of a natural language analysis process related to a variation. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] <First embodiment> Hereinafter, an embodiment of an information processing device, method, and program according to the present disclosure will be described with reference to the drawings. However, the embodiment described below is merely an example of an embodiment, and the information processing device, method, and program according to the present disclosure are not limited to the specific configuration described below. In carrying out the present invention, a specific configuration according to the embodiment may be appropriately adopted, and various improvements and modifications may be made. The present invention can appropriately adopt at least a part of the configurations in the embodiments and variations described below.

[0012] In this embodiment, the technology according to the present disclosure is described in a patent value evaluation system or patent information analysis system for evaluating patent claims. However, the technology according to the present disclosure can be widely used in technologies for analyzing natural language, and the application of the present disclosure is not limited to the examples shown in the embodiment. For example, the technology according to the present disclosure may be implemented to evaluate advertising copy, business documents, or literary works.

[0013] <<System configuration>> 1 is a schematic diagram showing a configuration of an information processing system according to the present embodiment. The information processing system according to the present embodiment includes an information processing device 1 and one or more user terminals 9, which are connected to a network and capable of communicating with each other. A user inputs a natural sentence to be analyzed to the information processing device 1 via the user terminal 9, and causes the information processing device 1 to analyze and evaluate the natural sentence. However, the system configuration for implementing the technology according to the present disclosure is not limited to the example in the present embodiment. The information processing system for implementing the technology according to the present disclosure may be implemented by a single information processing device.

[0014] The information processing device 1 is a computer including a central processing unit (CPU) 11, a read only memory (ROM) 12, a random access memory (RAM) 13, a storage device 14 such as an electrically erasable and programmable read only memory (EEPROM) or a hard disk drive (HDD), a communication unit 15 such as a network interface card (NIC), etc. However, the specific hardware configuration of the information processing device 1 can be omitted, replaced, or added as appropriate depending on the embodiment. In addition, the information processing device 1 is not limited to a device consisting of a single housing. The information processing device 1 may be realized by multiple devices using so-called cloud or distributed computing technology, etc.

[0015] The user terminal 9 is a terminal device used by a user. The user terminal 9 is a computer equipped with a CPU, a ROM, a RAM, a storage device, a communication unit, an input device, an output device, and the like (not shown). However, the specific hardware configuration of the user terminal 9 can be omitted, replaced, or added as appropriate depending on the embodiment. Furthermore, the user terminal 9 is not limited to a device consisting of a single housing. The user terminal 9 may be realized by a plurality of devices using so-called cloud or distributed computing technology, etc. The user uses the services provided by the system according to this embodiment via these user terminals 9.

[0016] 2 is a diagram showing an outline of the functional configuration of the information processing device 1 according to this embodiment. The information processing device 1 functions as an information processing device including a unit dividing unit 21, an element identifying unit 22, an element coordinate setting unit 23a, an inter-element energy calculation unit 24a, a unit coordinate setting unit 23b, an inter-unit energy calculation unit 24b, and an evaluation value calculation unit 28, by a program recorded in a storage device 14 being read into a RAM 13 and executed by a CPU 11, and each piece of hardware included in the information processing device 1 is controlled. Note that in this embodiment and other embodiments described later, each function included in the information processing device 1 is executed by the CPU 11, which is a general-purpose processor, but some or all of these functions may be executed by one or more dedicated processors.

[0017] The unit division unit 21 divides the natural text to be analyzed (hereinafter, referred to as "natural text to be analyzed") into a plurality of text units. As described above, in this embodiment, an example will be described in which a patent claim is used as the natural text to be analyzed. In recent years, the need for intellectual property value evaluation for business purposes (negotiations) and internal decision-making has increased with changes in the social and economic environment due to changes in business models and the increased role of intangible assets, and in order to accurately perform patent value evaluation, which is one type of intellectual property value evaluation, it is necessary to evaluate patent claims. However, a method capable of quantitatively evaluating the technical scope specified in a patent claim has not yet been established. For this reason, in this embodiment, a new patent modeling is adopted by "element decomposition of technical scope" using natural language processing based on AI / ML (machine learning) technology and "interaction evaluation between elements (compatibility between elements)" using Classical MD (molecular dynamics) calculation (however, in this disclosure, kinetic energy is set to zero, so it is referred to as "force field calculation").

[0018] The element identification unit 22 performs a morpheme analysis of the natural sentence to be analyzed, extracts a plurality of morphemes constituting the natural sentence to be analyzed, and identifies the plurality of morphemes as a plurality of elements. However, in the technology according to the present disclosure, the objects treated as elements of natural sentences are not limited to morphemes. An element of natural sentences may be any part of natural sentences, and may be, for example, a paragraph, a clause, a phrase, a word, etc.

[0019] The element coordinate setting unit 23a sets virtual coordinates of a virtual particle when each of a plurality of elements constituting each of a plurality of sentence units is regarded as a virtual particle. In this embodiment, the element coordinate setting unit 23a inputs the element into a trained natural language model to obtain a distributed representation (vector representation) of the element. Then, the element coordinate setting unit 23a determines virtual coordinates of the virtual particle when each of the elements is regarded as a virtual particle by compressing the distributed representation of the element corresponding to the virtual particle to a predetermined dimension (e.g., three dimensions) by principal component analysis (PCA), and sets the determined virtual coordinates as the coordinates of the virtual particle.

[0020] The inter-element energy calculation unit 24a regards the multiple elements that make up the target sentence unit as multiple virtual particles placed at virtual coordinates corresponding to each element, and calculates potential energy (hereinafter simply referred to as "energy") that indicates the interaction between the virtual particles corresponding to the multiple elements by applying a potential function to the multiple virtual particles and performing a force field calculation.

[0021] The unit coordinate setting section 23b sets, for each of a plurality of sentence units constituting the natural sentence to be analyzed, virtual coordinates of the sentence unit based on virtual coordinates of a plurality of elements constituting the sentence unit.

[0022] The inter-unit energy calculation unit 24b regards multiple sentence units as multiple virtual particles placed at virtual coordinates corresponding to each sentence unit, and calculates energy indicating the interaction between the virtual particles corresponding to the multiple sentence units by applying a potential function to the multiple virtual particles and performing a force field calculation.

[0023] FIG. 3 is a diagram showing the concept of calculating energy by regarding elements and sentence units constituting natural sentences as virtual particles in this embodiment. For example, if any constituent element (sentence unit) in a patent claim is a character string of "D, which inputs A into B and determines C", the character string is decomposed into elements of [A], [for], [B], [to], [input], [do], [C], [for], [determine], [do], and [D] by morphological analysis, and these elements are regarded as virtual particles, and force field calculation between particles is performed. In addition, sentence units including these elements are also regarded as virtual particles, and force field calculation between particles is performed. In this manner, the energy related to the entire patent claim is calculated in this embodiment.

[0024] The evaluation value calculation unit 28 calculates an evaluation value of the natural text to be analyzed based on the energy indicating the interaction between virtual particles corresponding to a plurality of elements and the energy indicating the interaction between virtual particles corresponding to a plurality of sentence units. Here, since the lower the energy, the higher the stability, it is preferable to adopt a scale for the evaluation value such that the lower the energy, the higher the evaluation.

[0025] <<Process flow>> Next, a flow of processing executed by the information processing device according to the present embodiment will be described. Note that the specific contents and processing order of the processing described below are an example for implementing the present disclosure. The specific contents and processing order may be appropriately selected according to the embodiment of the present disclosure.

[0026] 4 is a flowchart showing the flow of the natural language analysis process according to this embodiment. The process shown in this flowchart is executed when a user inputs a natural language sentence to be evaluated (a natural language sentence to be analyzed) into the information processing device 1 and then issues an instruction to start the natural language analysis process.

[0027] In step S101, the natural text to be analyzed is divided into one or more text units. The unit division unit 21 divides the natural text to be analyzed into a plurality of text units. In this embodiment, an example will be described in which a patent claim is the natural text to be analyzed, and so-called constituent features are analyzed as text units. Here, various units can be adopted as units into which the text units are divided according to the embodiment, and the units into which the text units are divided are not limited to the examples in this embodiment. For example, the text units may be divided based on some delimiter inserted in the text, such as a line break, a paragraph, or a punctuation mark, or may be divided by analyzing the structure or flow of the entire text. For example, in the case of a patent claim, the constituent features are often divided by line breaks, so it is easy to divide each constituent feature defined by the creator by line breaks or paragraphs into each text unit. However, since there are also patent claims that are not clearly divided, it is preferable that the method of dividing into text units is appropriately selected according to the structure of the natural text to be analyzed. Thereafter, the process proceeds to step S102.

[0028] In step S102, a morphological analysis of the natural sentence to be analyzed is performed, and a plurality of elements (components) are extracted from the sentence unit. In this embodiment, the morphemes constituting the natural sentence are used as the elements of the natural sentence to be analyzed, and the element identification unit 22 performs a morphological analysis of the natural sentence to be analyzed to extract a plurality of morphemes (tokens) constituting the natural sentence to be analyzed, and identifies the plurality of morphemes included in the sentence unit as a plurality of elements. For example, in the case of a natural sentence of "D which inputs A into B and determines C", the element identification unit 22 extracts morphemes such as [A], [for], [B], [to], [input], [do], [C], [for], [determine], [do], and [D] from the natural sentence, and sets these as a plurality of elements included in the natural sentence. Here, an existing Tokenizer (for example, MeCab, etc.) may be used for the morphological analysis. After that, the process proceeds to step S103.

[0029] In steps S103 and S104, virtual coordinates are assigned to each of the multiple elements when they are regarded as virtual particles. First, a distributed representation (vector representation) of each element is acquired via an encoder. The element coordinate setting unit 23a inputs each of the multiple elements constituting the target sentence unit to an encoder (e.g., a trained natural language processing model) to acquire an N-dimensional distributed representation output from the model (step S103). For example, "r 0[A] =[0.1, 0.2, 0.9, 0.7...]" is obtained. Here, an existing trained natural language processing model (e.g., Word2Vec, FastText, Transformer-Encoder, etc.) may be used as the natural language processing model. In addition, although an example in which a trained natural language processing model is used as the encoder has been described in this embodiment, other encoders may also be used.

[0030] When the distributed representation of each of the elements constituting the target sentence unit is acquired, virtual coordinates are assigned to each element based on the acquired distributed representation. The element coordinate setting unit 23a sets virtual coordinates of each element by compressing the distributed representation obtained in step S103 for each of the elements constituting the target sentence unit to a predetermined dimension (e.g., three dimensions) by principal component analysis (step S104). For example, 0[A] =[0.1, 0.2, 0.9, 0.7...]” is called r [A] =[0.2, 0.1, 0.4]". In principal component analysis, the dimension is lowered so that objects located close to each other in the distributed representation (N dimensions) before compression are also located close to each other in a lower predetermined dimension (here, three dimensions) that is the compression target. Note that, in this embodiment, an example of compression to three dimensions will be described, but the number of dimensions of the compression destination is not limited to three.

[0031] 5 is a schematic diagram showing a state in which a plurality of elements to which virtual coordinates are assigned are plotted in a three-dimensional space in this embodiment. In this embodiment, each element included in the natural text to be analyzed is regarded as a classical virtual particle (hard sphere). After the virtual coordinates are assigned to the virtual particle, the process proceeds to step S105.

[0032] In step S105, for each of the multiple sentence units, energy indicating the interaction between the virtual particles when each of the multiple elements constituting the sentence unit is regarded as a virtual particle is calculated. Here, the potential energy between the elements in the sentence unit is calculated by applying an arbitrary potential function to the virtual coordinates of each element. The inter-element energy calculation unit 24a regards the multiple elements constituting the target sentence unit as multiple virtual particles placed at the virtual coordinates corresponding to each element, and calculates the energy indicating the interaction between the virtual particles corresponding to the multiple elements by applying the potential function to the multiple virtual particles and performing a force field calculation.

[0033] The following formula is a formula showing a molecular dynamics (MD) calculation, and shows that the internal energy of the system shown on the left side (corresponding to the Hamiltonian of the system) is the sum of the kinetic energy of the system shown on the right side (energy corresponding to the momentum p of each virtual particle) and the potential energy of the system (interaction energy corresponding to the position r of each virtual particle). In the technology shown in this embodiment, each of the multiple elements constituting a sentence unit is regarded as a virtual particle, and one sentence unit is regarded as one system to calculate the energy. However, the energy calculation in this embodiment differs from the molecular dynamics calculation in that it is a static force field calculation in which the kinetic energy is calculated as 0 (zero) (ignored), and the potential energy when one sentence unit is regarded as one system is treated as the internal energy of the system. For this reason, in this embodiment, the calculation performed by applying a potential function to multiple virtual particles is called a "force field calculation".

number

[0034] FIG. 6 is a schematic diagram showing a state in which a force field calculation is performed by applying a potential function between virtual particles corresponding to a plurality of elements constituting a natural sentence in this embodiment. In this embodiment, virtual springs are set between the virtual particles, and the system is treated as a harmonic oscillator. Here, there exists a most stable bond length between two particles (e.g., Lennard-Jones Potential), and there exists a most stable bond angle between three particles (e.g., Stillinger-Weber Potential). In the force field calculation, potential energies for these stable states are calculated using a potential function. A method for determining a potential function to be applied between virtual particles will be described later. When the energy for each of the plurality of sentence units is calculated, the process proceeds to step S106.

[0035] In step S106, virtual coordinates are assigned to each of the sentence units constituting the natural sentence to be analyzed, assuming that each of the sentence units is a virtual particle. Here, virtual coordinates are assigned to each sentence unit according to the virtual coordinates of each element included in the sentence unit.

[0036] FIG. 7 is a schematic diagram showing how a sentence unit including a plurality of elements is regarded as a virtual particle including virtual particles corresponding to the plurality of sentence units in this embodiment. The unit coordinate setting unit 23b sets virtual coordinates of each of the plurality of sentence units constituting the natural sentence to be analyzed, based on the virtual coordinates of the plurality of elements constituting the sentence unit. Various methods may be used to calculate the virtual coordinates assigned to the virtual particle corresponding to the sentence unit, but in this embodiment, the target sentence unit is coarse-grained, and a classical virtual particle (hard sphere) including all the elements included in the sentence unit is regarded as the virtual particle corresponding to the sentence unit. The virtual coordinates of the virtual particle corresponding to the sentence unit may be the coordinates of the center of a sphere based on the average of the virtual coordinates corresponding to the virtual particles of the plurality of elements, or may be the virtual coordinates of the center of gravity calculated in consideration of the coordinates of the virtual particles corresponding to the plurality of elements contained in the sphere or the weight according to the potential energy generated for each of the virtual particles corresponding to the plurality of elements. Here, for example, the sum of the potential energies calculated for all particle pairs including a virtual particle corresponding to a certain element may be adopted as the "potential energy generated for the virtual particle" corresponding to the element, and the weighting may be performed according to the ranking of this sum. More specifically, when there are virtual particles A, B, and C corresponding to elements, and the potential energies are high in the order of virtual particle pair AB, virtual particle pair BC, and virtual particle pair AC, the weight of virtual particle B, which has the largest total of potential energies calculated for all particle pairs including itself, may be set to the largest value. Then, the process proceeds to step S107.

[0037] In step S107, energy indicating interactions between virtual particles when each of a plurality of sentence units is regarded as a virtual particle is calculated for the entire natural sentence to be analyzed. Here, potential energy between coarse-grained sentence units in the natural sentence to be analyzed is calculated by applying an arbitrary potential function to the virtual coordinates of each sentence unit. The inter-unit energy calculation unit 24b regards a plurality of sentence units as a plurality of virtual particles placed at virtual coordinates corresponding to each sentence unit, and calculates energy indicating interactions between virtual particles corresponding to the plurality of sentence units by applying a potential function to the plurality of virtual particles and performing force field calculation.

[0038] The specific energy calculation method used here is roughly the same as that in step S105, except that the entire natural text (here, the patent claim) is regarded as one system, and the text unit is regarded as a particle, so the explanation is omitted. As in step S105, the kinetic energy is calculated as 0 (zero).

[0039] Fig. 8 is a schematic diagram showing how force field calculation is performed by applying a potential function between virtual particles corresponding to a plurality of sentence units constituting a natural sentence in this embodiment. As in the example of energy calculation between elements described with reference to Fig. 6, in this embodiment, virtual springs are set between the virtual particles, and the system is treated as a harmonic oscillator. In the force field calculation, potential energy for a stable state is calculated using a potential function. A method for determining a potential function to be applied between virtual particles will be described later. When the energy for the entire natural sentence to be analyzed is calculated, the process proceeds to step S108.

[0040] In step S108, an evaluation value of the entire natural sentence to be analyzed is calculated based on the potential energy of the natural sentence to be analyzed obtained by adding up the potential energy of each sentence unit and the potential energy between the sentence units. The evaluation value calculation unit 28 calculates the evaluation value of the natural sentence to be analyzed based on the energy indicating the interaction between virtual particles corresponding to a plurality of elements and the energy indicating the interaction between virtual particles corresponding to a plurality of sentence units. Here, the lower the energy, the more stable it is, so the lower the energy, the higher the evaluation. For example, the exponential of the total amount of potential energy may be used as the evaluation value. The following formula is an example of a formula for calculating the evaluation value based on the sum of the potential energies calculated for N sentence units.

number

[0041] However, the specific expression method and scale of the evaluation value are not limited to the examples in this disclosure. The evaluation value may be calculated according to the total amount of energy, and may be a numerical value, a rank, an estimated patent transfer price, or the like. When the evaluation value of the entire natural text to be analyzed is calculated, the process shown in this flowchart ends.

[0042] Next, a method for determining a potential function to be applied between virtual particles will be described. Although the method for determining the potential function is not limited to the specific example described in this embodiment, it is preferable that the potential function and parameters to be applied are determined according to a combination of at least any of the attributes of the natural text to be analyzed that corresponds to the virtual particle to which the potential function is to be applied, the corresponding element, and the corresponding sentence unit.

[0043] In this embodiment, an example will be described in which the inter-element energy calculation unit 24a or the inter-unit energy calculation unit 24b uses a patent classification code (e.g., the International Patent Classification (IPC) or FI used by the Japan Patent Office, etc.) as an attribute of an element or sentence unit corresponding to each of a plurality of virtual particles to which a potential function is applied. The method of identifying the patent classification code corresponding to an element or sentence unit is not limited, but in addition to a method of searching a list of patent classification codes, a method of inputting an element or sentence unit into a classifier (classification model) generated by subjecting a word, sentence, etc., to supervised machine learning of the relationship between the patent classification code and the word, sentence, etc., and the patent classification code, and obtaining the patent classification code as an output, may be adopted.

[0044] FIG. 9 is a diagram showing an overview of a process for acquiring attributes of elements of natural language using a machine learning model in this embodiment. In the example shown in FIG. 9, the word "payment" as element E1 and the word "neural network" as element E2 are input to a multi-label classifier, and one or more labels predicted by the classifier and the probability of each label are obtained as the output. In the example shown in FIG. 9, [···G06Q10 / 00, G06Q20 / 00, G06Q30 / 00···] are obtained as patent classification codes (labels) of element E1 "payment", and [···0.0, 0.7, 0.3···] are obtained as the probabilities of these patent classification codes. In addition, [···G06N3 / 02···] is obtained as the patent classification code (label) of element E2 "neural network", and [···0.9···] is obtained as the probability of these patent classification codes. When the prediction result is obtained by the classifier, the inter-element energy calculation unit 24a or the inter-unit energy calculation unit 24b reads out the potential function and parameters preset for the combination of the obtained patent classification codes and applies them to the target virtual particles. Note that, although Fig. 9 illustrates an example in which there are two virtual particles (2 bodies), the process flow is the same even when there are three or more virtual particles (3 bodies).

[0045] As the potential function, for example, the Lennard-Jones Potential function shown below for calculating an interatomic potential can be adopted. Here, the parameter ε indicates the depth of the potential well, and the parameter σ indicates the distance at which the potential value becomes zero. Note that, although an example in which a relatively simple function such as the Lennard-Jones Potential is adopted is described in this embodiment, the functions that can be adopted are not limited to the example below.

number

[0046] Next, a method for determining the parameters ε and σ when the above-mentioned function is used to calculate the interatomic potential will be described with reference to the drawings.

[0047] FIG. 10 is a diagram showing an outline of the relationship between the distance r between virtual particles and the energy V(r) when a function for calculating an interatomic potential is adopted as the potential function in this embodiment. In addition, in this figure, a graph is shown in which the exponent part (12,6) in the function is changed to (4,2) for ease of understanding. The dashed line C1 shows the relationship between the distance r between these virtual particles P1 and P2 and the energy V(r) when the virtual particle P1 (see FIG. 9) corresponding to the element E1 (in the example shown in FIG. 9, "Settlement") has the attribute A1 (in the example shown in FIG. 9, patent classification code G06Q20 / 00), and the virtual particle P2 (see FIG. 9) corresponding to the element E2 (in the example shown in FIG. 9, "Neural Network") has the attribute A2 (in the example shown in FIG. 9, patent classification code G06N3 / 02). The potential function used here is a function for calculating the interatomic potential as described above, and since the interatomic potential is determined by the energy of repulsive interaction and the energy of attractive interaction, both of which change according to distance, the point where the energy V(r) is a minimum value appears in the middle of the graph. And, since the lower the energy is, the more stable it is, as described above, the distance r where the energy V(r) is a minimum value is the distance where the evaluation is highest in the relationship between element E1 and element E2.

[0048] Fig. 11 is a diagram showing an example of parameters set in the potential function when a function for calculating an interatomic potential is adopted as the potential function in this embodiment. The inter-element energy calculation unit 24a or the inter-unit energy calculation unit 24b determines parameters to be applied to the potential function by referring to parameters defined in advance for a combination of attributes (attribute pair). When the referenced patent classification code is the most probable patent classification code (i.e., a representative label) for each virtual particle, for example, a combination of attribute A1 (patent classification code G06Q20 / 00) and attribute A2 (patent classification code G06N3 / 02), according to the example shown in Fig. 11, the inter-element energy calculation unit 24a or the inter-unit energy calculation unit 24b reads out parameters ε = 0.5 and σ = 1.0 as parameters defined in advance for the combination of these attributes, and applies these parameters to the potential function to calculate the energy.

[0049] The patent classification code referred to at this time may be only the patent classification code with the highest probability for each virtual particle (i.e., a representative label), or in addition to the representative patent classification code, a plurality of patent classification codes including patent classification codes with probabilities ranked second or lower may be referred to. The method of calculating the energy when patent classification codes with probabilities ranked second or lower are also referred to is not limited to the example in this embodiment, but below, a method of using a simple average or weighted average of parameters of the potential function when referring to patent classification codes ranked second or lower will be described.

[0050] Here, an example of a method for determining parameters of a potential function when calculating energy for a virtual particle P1 corresponding to element E1 "payment" by considering not only the attribute A1 (patent classification code G06Q20 / 00) with the highest probability but also the attribute A3 (patent classification code G06Q30 / 00) with the second highest probability. In Fig. 10, the dashed line C2 indicates the relationship between the distance r and the energy V(r) between the virtual particles P1 and P2 when the virtual particle P1 corresponding to element E1 "payment" has the attribute A3 (patent classification code G06Q30 / 00 in the example shown in Fig. 9) and the virtual particle P2 corresponding to element E2 "neural network" has the attribute A2 (patent classification code G06N3 / 02).

[0051] Here, if only the attribute A3 of the attributes of the virtual particle P1 is taken into consideration, the parameter ε=0.9 and the parameter σ=1.2, which are predefined for a combination of these attributes (see FIG. 11), are read out and these parameters are applied to the potential function to calculate the energy, but here, the attributes A1 and A3 of the attributes of the virtual particle P1 are taken into consideration. For this reason, values ​​calculated based on both the parameter ε=0.5 and the parameter σ=1.0 when the attribute of the virtual particle P1 is the attribute A1, and the parameter ε=0.9 and the parameter σ=1.2 when the attribute of the virtual particle P1 is the attribute A3 are adopted for the parameter ε and the parameter σ.

[0052] When a simple average is adopted as the method for determining the parameters, the parameters applied to the potential function between virtual particle P1 and virtual particle P2 are the simple averages of parameters ε=0.7 and σ=1.1 (see FIG. 11). When such simple average parameters are adopted, the curve showing the relationship between the distance r between the virtual particles and the energy V(r) is intermediate between the curve for the first attribute pair (in the example shown in the figure, the pair of attributes A1 and A2) and the curve for the second attribute pair (in the example shown in the figure, the pair of attributes A3 and A2) (see solid line C3 in FIG. 10).

[0053] On the other hand, a weighted average may be adopted as a method for determining the parameters. Various values ​​can be adopted as the weights when calculating the weighted average, but here, an example will be described in which the weighted average is calculated based on the probability output from the multi-label classifier for the target label. Specifically, the inter-element energy calculation unit 24a or the inter-unit energy calculation unit 24b calculates the parameters ε and σ using the following formulas so that the curve showing the relationship between the distance r between virtual particles and the energy V(r) is closer to the curve for the first attribute pair or the curve for the second attribute pair, whichever has a larger weight. Weighted average parameter = (first attribute pair parameter - second attribute pair parameter) * weighting factor + second attribute pair parameter Weighting factor = probability of first attribute pair / (probability of first attribute pair + probability of second attribute pair)

[0054] Applying the example parameters shown in Figure 11 to the above formula, we get Parameter ε=(0.5-0.9)*0.7+0.9 Parameter σ=(1.0-1.2)*0.7+1.2 and parameters ε=0.62 and σ=1.06, which are weighted averages, are adopted (see FIG. 11). The curve showing the relationship between the distance r between virtual particles and the energy V(r) is biased toward the curve with the greater weight between the curve for the first attribute pair and the curve for the second attribute pair (see solid line C4 in FIG. 10).

[0055] That is, in this embodiment, when two attributes G06Q20 / 00 and G06Q30 / 00 are estimated for an element constituting a natural sentence "payment", the final parameters between "payment" and "neural network" are set for the virtual particle pair "payment"-"neural network" by considering not only the parameters of attribute A1 (G06Q20 / 00)-attribute A2 (G06N3 / 2) but also the parameters of attribute A3 (G06Q30 / 00)-attribute A2 (G06N3 / 2). Here, an example in which two or more attributes are considered for one of the two virtual particles has been described, but the basic calculation method is the same when two or more attributes are considered for both of the two virtual particles. The basic calculation method is also the same when three or more attributes are considered for the virtual particles, or when multiple attributes are considered for three or more virtual particles. Here, as for the weight, only the probability (estimated result) of the attribute of one virtual particle may be used, or the product of the probabilities of the attributes of both virtual particles may be used as the weight.

[0056] According to the system of the present embodiment described above, it is possible to quantitatively evaluate natural language (in the example shown in the present embodiment, a patent claim). More specifically, according to the system of the present embodiment described above, it is possible to evaluate the stability of natural language by evaluating the interaction / compatibility between multiple elements that make up the natural language.

[0057] Second Embodiment In the first embodiment described above, an example in which the technology according to the present disclosure is used to evaluate patent claims has been described, but the technology according to the present disclosure can be used to evaluate any natural language, and the scope of application is not limited to the evaluation of patent claims. For example, the technology according to the present disclosure can be used to evaluate advertising copy. Note that the hardware configuration of the system according to this embodiment is roughly the same as that of the first embodiment, and therefore will not be described.

[0058] 12 is a diagram showing an outline of the functional configuration of an information processing device 1b according to this embodiment. In the information processing device 1b, a program recorded in the storage device 14 is read into the RAM 13 and executed by the CPU 11, and each piece of hardware included in the information processing device 1b is controlled, so that the information processing device 1b functions as an information processing device including an element identification unit 22, an element coordinate setting unit 23a, an inter-element energy calculation unit 24a, and an evaluation value calculation unit 28. Note that in this embodiment and other embodiments described later, each function included in the information processing device 1b is executed by the CPU 11, which is a general-purpose processor, but some or all of these functions may be executed by one or more dedicated processors.

[0059] The information processing device 1b according to this embodiment differs from the information processing device 1 according to the first embodiment in that the unit dividing unit 21, the unit coordinate setting unit 23b, and the inter-unit energy calculation unit 24b in the first embodiment are omitted, and the system that is the target of processing by the element coordinate setting unit 23a and the inter-element energy calculation unit 24a is the entire natural text to be analyzed, rather than the sentence units in the first embodiment. However, for natural text that is different from patent claims, such as advertising copy, the energy may be calculated by dividing it into multiple stages, such as sentence units and elements, as in the first embodiment, and an evaluation value may be obtained.

[0060] 13 is a flowchart showing the flow of the natural language analysis process according to this embodiment. The process shown in this flowchart is executed when a user inputs a natural language sentence to be evaluated (a natural language sentence to be analyzed) into the information processing device 1b and then issues an instruction to start the natural language analysis process.

[0061] In step S201, a morphological analysis of the natural sentence to be analyzed is performed, and a plurality of elements (components) are extracted from the natural sentence to be analyzed. In this embodiment, as in the first embodiment described above, the morphemes constituting the natural sentence are used as elements of the natural sentence to be analyzed, and the element identification unit 22 performs a morphological analysis of the natural sentence to be analyzed to extract a plurality of morphemes (tokens) constituting the natural sentence to be analyzed, and identifies the plurality of morphemes included in the natural sentence to be analyzed as a plurality of elements. For example, in the case of a natural sentence "You can get a lot of points," the element identification unit 22 extracts morphemes such as [points], [a lot], and [get] from the natural sentence, and sets these as a plurality of elements included in the natural sentence. After that, the process proceeds to step S202.

[0062] In steps S202 and S203, virtual coordinates are assigned to each of the elements when the elements are regarded as virtual particles. In this embodiment, as in the first embodiment described above, the element coordinate setting unit 23a inputs each of the elements constituting the natural text to be analyzed to an encoder to obtain an N-dimensional distributed representation output from the model (step S202). Then, the element coordinate setting unit 23a sets virtual coordinates of each of the elements constituting the natural text to be analyzed by compressing the distributed representation obtained in step S202 to a predetermined dimension by principal component analysis (step S203). When the virtual coordinates are assigned to the virtual particles, the process proceeds to step S204.

[0063] In step S204, energy indicating interactions between virtual particles when each of a plurality of elements is regarded as a virtual particle is calculated for the entire natural text to be analyzed. The inter-element energy calculation unit 24a regards a plurality of elements as a plurality of virtual particles placed at virtual coordinates corresponding to each element, and calculates energy indicating interactions between virtual particles corresponding to the plurality of elements by applying a potential function to the plurality of virtual particles and performing force field calculation. The method of calculating the energy can be substantially the same as that described in the first embodiment, including the method of selecting the potential function and the method of determining the parameters, and therefore the description will be omitted. After that, the process proceeds to step S205.

[0064] In step S205, an evaluation value of the entire analysis target natural text is calculated based on the potential energy of the analysis target natural text. The evaluation value calculation unit 28 calculates an evaluation value of the analysis target natural text based on the energy indicating the interaction between virtual particles corresponding to a plurality of elements. The method of calculating the evaluation value is roughly the same as that described in the first embodiment, and therefore the description will be omitted. After that, the processing shown in this flowchart ends.

[0065] According to the system of the present embodiment described above, it is possible to quantitatively evaluate natural language (advertising copy in the example shown in the present embodiment). More specifically, according to the system of the present embodiment described above, it is possible to evaluate the stability of natural language by evaluating the interaction / compatibility between multiple elements that make up the natural language.

[0066] <Third embodiment> In the first embodiment described above, an example was described in which the patent claim is divided into multiple stages of sentence units and elements, the energy is calculated, and an evaluation value is obtained. However, as in the second embodiment described above, the patent claim may also be evaluated based on only one stage of energy calculation. Here, when the patent claim is divided into multiple morphemes and the evaluation value is obtained in only one stage of calculating the energy, the content of the process is roughly the same as in the second embodiment described above, so the description is omitted. Below, the content of the process when the patent claim is divided into multiple sentence units and the evaluation value is obtained in only one stage of calculating the energy is described. Note that the hardware configuration of the system according to this embodiment is roughly the same as in the first embodiment, so the description is omitted.

[0067] 14 is a diagram showing an outline of the functional configuration of an information processing device 1c according to this embodiment. In the information processing device 1c, a program recorded in the storage device 14 is read into the RAM 13 and executed by the CPU 11, and each piece of hardware provided in the information processing device 1c is controlled, so that the information processing device 1c functions as an information processing device including a unit dividing section 21, a unit coordinate setting section 23b, an inter-unit energy calculation section 24b, and an evaluation value calculation section 28. Note that in this embodiment and other embodiments described later, each function provided in the information processing device 1c is executed by the CPU 11, which is a general-purpose processor, but some or all of these functions may be executed by one or more dedicated processors.

[0068] An information processing device 1c according to this embodiment differs from the information processing device 1 according to the first embodiment in that the element identifying unit 22, the element coordinate setting unit 23a, and the inter-element energy calculating unit 24a in the first embodiment are omitted.

[0069] 15 is a flowchart showing the flow of the natural language analysis process according to this embodiment. The process shown in this flowchart is executed when a user inputs a natural language sentence to be evaluated (a natural language sentence to be analyzed) into the information processing device 1c and then issues an instruction to start the natural language analysis process.

[0070] In step S301, the natural sentence to be analyzed is divided into one or more sentence units. The unit division unit 21 divides the natural sentence to be analyzed into a plurality of sentence units. The specific method for dividing the natural sentence to be analyzed into a plurality of sentence units is roughly the same as in the first embodiment, and therefore the explanation is omitted. After that, the process proceeds to step S302.

[0071] In steps S302 and S303, virtual coordinates are assigned to each of the multiple sentence units when they are regarded as virtual particles. The unit coordinate setting unit 23b inputs each of the multiple sentence units constituting the target natural sentence to be analyzed to an encoder to obtain an N-dimensional distributed representation output from the model (step S302). As in the example in the first embodiment, a trained natural language processing model can be used for the encoder. Then, the unit coordinate setting unit 23b sets virtual coordinates of each sentence unit by compressing the distributed representation obtained in step S302 for each of the multiple sentence units constituting the target natural sentence to be analyzed to a predetermined dimension by principal component analysis (step S303). When the virtual coordinates are assigned to the virtual particles, the process proceeds to step S304.

[0072] In step S304, energy indicating interactions between virtual particles when each of a plurality of sentence units is regarded as a virtual particle is calculated for the entire natural sentence to be analyzed. The unit energy calculation unit 24b regards a plurality of sentence units as a plurality of virtual particles placed at virtual coordinates corresponding to each sentence unit, and calculates energy indicating interactions between virtual particles corresponding to the plurality of sentence units by applying a potential function to the plurality of virtual particles and performing force field calculation. The method of calculating the energy can be substantially the same as that described in the first embodiment, including the method of selecting the potential function and the method of determining the parameters, and therefore the description will be omitted. After that, the process proceeds to step S305.

[0073] In step S305, an evaluation value of the entire natural text to be analyzed is calculated based on the potential energy of the natural text to be analyzed. The evaluation value calculation unit 28 calculates an evaluation value of the natural text to be analyzed based on the energy indicating the interaction between virtual particles corresponding to a plurality of text units. The method of calculating the evaluation value is roughly the same as that described in the first embodiment, and therefore the description will be omitted. After that, the processing shown in this flowchart ends.

[0074] <Other variations> In the above-described embodiment, the energy of the system and the evaluation value corresponding to the energy are calculated by applying a potential function to a plurality of virtual particles and performing a force field calculation, but the energy or evaluation value may be estimated by a machine learning model instead of calculating using a potential function. That is, if the potential energy or evaluation value of the system (for example, if the natural text to be analyzed is a patent claim, the transfer price of the patent may be used as the evaluation value) is given, the machine learning model is made to learn the correspondence between the set of positional relationships of the virtual particle group (vector group) and the energy or evaluation value, and the positional relationships of the virtual particle group are input to the trained machine learning model, so that an estimated value of the potential energy or evaluation value of the system can be obtained as an output.

[0075] When a machine learning model is used instead of a potential function, in the natural language analysis process, virtual coordinates are assigned to the virtual particles, and then the positional relationship of the virtual particle group based on the assigned virtual coordinates is input to the machine learning model. When potential energy is obtained as an output from the machine learning model, the evaluation value is calculated based on the obtained potential energy. When an evaluation value is directly obtained as an output from the machine learning model, calculation of the evaluation value based on the energy is omitted. In addition, although an example in which the positional relationship of the virtual particle group is used as an input to the machine learning model has been described above, a feature amount described based on a symmetry function related to the distance distribution and angle distribution around the particles extracted from the positional relationship of the virtual particle group may be used as an input to the machine learning model.

[0076] 16 is a flowchart showing the flow of the natural language analysis process according to this variation. The process shown in this flowchart is executed when a user inputs a natural language sentence to be evaluated (a natural language sentence to be analyzed) into the information processing device and then issues an instruction to start the natural language analysis process.

[0077] The process from step S501 to step S503 is similar to the process from step S201 to step S203 described with reference to Fig. 13, and therefore the description will be omitted. After that, the process proceeds to step S504.

[0078] In steps S504 and S505, an evaluation value of the natural sentence to be analyzed is obtained based on the positional relationships created for the virtual particles. The information processing device calculates the positional relationships (for example, may be expressed by vectors) of the virtual particles based on the virtual coordinates set in step S503, and creates a set of positional relationships (for example, a vector group) for the combination of the virtual particles (step S504). Then, the evaluation value calculation unit 28 inputs the set of positional relationships of the virtual particles created in step S504 into a trained machine learning model, thereby obtaining an evaluation value of the natural sentence to be analyzed as an output from the machine learning (step S505). Note that, although an example of directly obtaining an evaluation value as an output from the machine learning model has been described here, a method of obtaining potential energy of the system from the machine learning model and calculating an evaluation value based on the obtained potential energy may be adopted. In this case, the calculation method of the evaluation value based on the energy is as described in the first embodiment. Thereafter, the process shown in this flowchart ends.

[0079] In this flowchart, a variation in which part of the processing in the second embodiment is replaced with a machine learning model is specifically described, but a variation in which a machine learning model is used instead of a potential function can be applied to any of the embodiments described above. [Explanation of symbols]

[0080] 1, 1b, 1c Information processing device

Claims

1. a coordinate setting means for setting virtual coordinates of each of a plurality of elements constituting the natural text to be analyzed based on a distributed representation of the element; an energy calculation means for calculating an energy indicating an interaction between the virtual particles by regarding the plurality of elements as a plurality of virtual particles placed at the virtual coordinates corresponding to each element and performing a force field calculation by applying a potential function to the plurality of virtual particles; an evaluation value calculation means for calculating an evaluation value indicating an interaction between the plurality of elements constituting the natural text to be analyzed based on the energies calculated for the plurality of virtual particles; A natural language processing system that utilizes force field calculations.

2. an element specifying means for extracting a plurality of morphemes constituting the analysis target natural sentence by performing a morpheme analysis of the analysis target natural sentence, and specifying the plurality of morphemes as the plurality of elements; A natural language processing system utilizing the force field calculation according to claim 1.

3. the energy calculation means calculates the energy by using the potential function to which a parameter set based on an attribute of the element corresponding to the virtual particle is applied. A natural language processing system utilizing the force field calculation according to claim 1.

4. A unit dividing means for dividing a natural text to be analyzed into a plurality of text units; an element coordinate setting means for setting virtual coordinates of a virtual particle when the element is regarded as a virtual particle for each of a plurality of elements constituting each of the plurality of sentence units, based on a distributed representation of the element corresponding to the virtual particle; an inter-element energy calculation means for calculating, for each of the plurality of sentence units, energy indicating an interaction between the virtual particles corresponding to the plurality of elements by regarding the plurality of elements constituting the target sentence unit as a plurality of virtual particles placed at the virtual coordinates corresponding to each element and performing a force field calculation by applying a potential function to the plurality of virtual particles; a unit coordinate setting means for setting virtual coordinates of each of the plurality of sentence units constituting the natural sentence to be analyzed, the virtual coordinates of the sentence unit being based on the virtual coordinates corresponding to the plurality of elements constituting the sentence unit; a unit-to-unit energy calculation means for calculating an energy indicating an interaction between the virtual particles corresponding to the plurality of sentence units by regarding the plurality of sentence units as a plurality of virtual particles placed at the virtual coordinates corresponding to each sentence unit and performing a force field calculation by applying a potential function to the plurality of virtual particles; an evaluation value calculation means for calculating an evaluation value of the natural text to be analyzed based on energies indicating interactions between the virtual particles corresponding to the plurality of elements and energies indicating interactions between the virtual particles corresponding to the plurality of sentence units; A natural language processing system that utilizes force field calculations.

5. The method further comprises an element specifying means for extracting a plurality of morphemes constituting the analysis target natural sentence by performing a morpheme analysis of the analysis target natural sentence, and specifying the plurality of morphemes as the plurality of elements. A natural language processing system utilizing the force field calculation according to claim 4.

6. the unit energy calculation means calculates the energy by using the potential function to which a parameter set based on an attribute of the sentence unit corresponding to the virtual particle is applied; A natural language processing system utilizing the force field calculation according to claim 4.

7. On the computer, a coordinate setting step of setting virtual coordinates of each of a plurality of elements constituting the natural text to be analyzed based on a distributed representation of the element; an energy calculation step of calculating energy indicating an interaction between the virtual particles by regarding the plurality of elements as a plurality of virtual particles placed at the virtual coordinates corresponding to each element and applying a potential function to the plurality of virtual particles to perform a force field calculation; an evaluation value calculation step of calculating an evaluation value indicating an interaction between the plurality of elements constituting the natural text to be analyzed based on the energies calculated for the plurality of virtual particles; A natural language processing method for executing the above.

8. On the computer, A unit division step of dividing a natural text to be analyzed into a plurality of text units; an element coordinate setting step of setting virtual coordinates of a virtual particle when each of a plurality of elements constituting each of the plurality of sentence units is regarded as a virtual particle based on a distributed representation of the element corresponding to the virtual particle; an inter-element energy calculation step of regarding each of the plurality of sentence units as a plurality of virtual particles placed at the virtual coordinates corresponding to each element, and calculating an energy indicating an interaction between the virtual particles corresponding to the plurality of elements by applying a potential function to the plurality of virtual particles and performing a force field calculation; a unit coordinate setting step of setting virtual coordinates of each of the plurality of sentence units constituting the natural sentence to be analyzed, the virtual coordinates of the sentence unit being based on the virtual coordinates corresponding to the plurality of elements constituting the sentence unit; a unit energy calculation step of calculating an energy indicating an interaction between the virtual particles corresponding to the plurality of sentence units by regarding the plurality of sentence units as a plurality of virtual particles placed at the virtual coordinates corresponding to each sentence unit and applying a potential function to the plurality of virtual particles to perform a force field calculation; an evaluation value calculation step of calculating an evaluation value of the natural sentence to be analyzed based on energies indicating interactions between the virtual particles corresponding to the plurality of elements and energies indicating interactions between the virtual particles corresponding to the plurality of sentence units; A natural language processing method for executing the above.

9. Computer, a coordinate setting means for setting virtual coordinates of each of a plurality of elements constituting the natural text to be analyzed based on a distributed representation of the element; an energy calculation means for calculating an energy indicating an interaction between the virtual particles by regarding the plurality of elements as a plurality of virtual particles placed at the virtual coordinates corresponding to each element and performing a force field calculation by applying a potential function to the plurality of virtual particles; an evaluation value calculation means for calculating an evaluation value indicating an interaction between the plurality of elements constituting the natural text to be analyzed based on the energies calculated for the plurality of virtual particles; A program that functions as a

10. Computer, A unit dividing means for dividing a natural text to be analyzed into a plurality of text units; an element coordinate setting means for setting virtual coordinates of a virtual particle when the element is regarded as a virtual particle for each of a plurality of elements constituting each of the plurality of sentence units, based on a distributed representation of the element corresponding to the virtual particle; an inter-element energy calculation means for calculating, for each of the plurality of sentence units, energy indicating an interaction between the virtual particles corresponding to the plurality of elements by regarding the plurality of elements constituting the target sentence unit as a plurality of virtual particles placed at the virtual coordinates corresponding to each element and performing a force field calculation by applying a potential function to the plurality of virtual particles; a unit coordinate setting means for setting virtual coordinates of each of the plurality of sentence units constituting the natural sentence to be analyzed, the virtual coordinates of the sentence unit being based on the virtual coordinates corresponding to the plurality of elements constituting the sentence unit; a unit-to-unit energy calculation means for calculating an energy indicating an interaction between the virtual particles corresponding to the plurality of sentence units by regarding the plurality of sentence units as a plurality of virtual particles placed at the virtual coordinates corresponding to each sentence unit and performing a force field calculation by applying a potential function to the plurality of virtual particles; an evaluation value calculation means for calculating an evaluation value of the natural text to be analyzed based on energies indicating interactions between the virtual particles corresponding to the plurality of elements and energies indicating interactions between the virtual particles corresponding to the plurality of sentence units; A program that functions as a

Images