JP7199067B2 - Reasoning device, reasoning method and program - Google Patents

Description

The present invention relates to an inference device, an inference method, and a program.

In Patent Document 1, "The position information detection sensor 9 included in the terminal 3 may be a GPS ... measures the current position of the vehicle and outputs this current position information" (paragraph 0086), " The information sent from the terminal 3 to the server 20 is described as location information such as "0 prefecture x city △ town XXX" (paragraph 0109).
[Prior art documents]
[Patent Literature]
[Patent Document 1] Patent No. 6095835

In one aspect of the invention, a reasoning apparatus is provided. The reasoning apparatus may comprise an acquisition unit for acquiring geographic data relating to an area containing a particular point. The inference device acquires a first language expression related to a specific point and an acquisition unit based on learning results of a correspondence relationship between the first language expression related to the arbitrary point and geographic data related to the area including the arbitrary point and the arbitrary point. An inference unit may be provided for inferring a particular point from the provided geographic data.

The inference unit includes a plurality of linguistic feature amounts specified from a first language expression regarding an arbitrary point, and a plurality of map images or feature images identified from a plurality of map images or feature images included in geographic data regarding an area including the arbitrary point. Correspondence relationships between image feature amounts and arbitrary points may be learned in advance. The inference unit may specify a plurality of image feature amounts from a plurality of map images or feature images included in the geographic data acquired by the acquisition unit. The inference unit may identify a plurality of linguistic features from the first linguistic expression regarding the specific point. The inference unit may infer a specific spot from a plurality of image feature quantities and a plurality of language feature quantities relating to the specific spot.

The inference unit generates a plurality of linguistic feature values and a plurality of image feature values related to the arbitrary point, step by step from a first area including the arbitrary point to a second area smaller than the first area and including the arbitrary point. , and an arbitrary point may be learned. The inference unit generates a plurality of image feature values and a plurality of language feature values related to the specific point, step by step from a third area including the specific point to a fourth area including the specific point that is smaller than the third area. A particular point may be inferred from

The inference unit may generate an image of a specific point by adding together geographical likelihoods indicated by a plurality of image feature amounts respectively corresponding to a plurality of linguistic feature amounts related to the specific point.

The inference device further includes a calculation unit that determines a point with the highest geographical likelihood as the specific point in the image of the specific point generated by the inference unit, and calculates the latitude and longitude of the determined point. may

When the point specified by the latitude and longitude calculated by the calculation unit is not located on the road, the inference device converts the latitude and longitude calculated by the calculation unit to the specified point on the road closest to the specified point. A correction unit may be further provided for correcting the latitude and longitude of the point closest to the point where the measurement is performed.

The acquisition unit acquires open data of an area including the specific point based on the second language expression when a second language expression related to the specific point is input, and obtains landmarks, roads, and intersections from the open data. A plurality of map images or feature images classified by at least one of them may be generated.

The acquisition unit may acquire a natural language expression regarding the specific point, convert the natural language expression into a first language expression that is a vector expression regarding the specific point, and output the first language expression to the inference unit.

In one aspect of the invention, an inference method is provided. The inference method may comprise obtaining geographic data for an area containing a particular point. The inference method is based on the learning result of the correspondence relationship between the first language expression regarding the arbitrary point and the geographic data regarding the area including the arbitrary point and the arbitrary point, and the first language expression regarding the specific point and the acquisition stage. An inference stage may be provided for inferring a particular point from the geographic data obtained.

In one aspect of the invention, a program is provided. The program may cause the computer to perform an acquisition procedure to acquire geographic data for the area containing the specified point. The program provides a computer with a first language expression of a specific point and an acquisition procedure based on learning results of a correspondence relationship between a first language expression of an arbitrary point and geographic data of an area containing the arbitrary point and the arbitrary point. An inference procedure may be executed to infer a specific point from the geographic data obtained in .

It should be noted that the above summary of the invention does not list all the necessary features of the invention. Subcombinations of these feature groups can also be inventions.

1 is a schematic diagram of a system 10 according to one embodiment; FIG. 1 is a block diagram of a reasoning apparatus 100 according to one embodiment; FIG. FIG. 4 is a flow diagram of an inference method according to one embodiment; 6 is an example of an area 60 that includes a particular point where user 20 is located, according to one embodiment. FIG. 4 is a diagram for explaining an example of a method of generating a map image or a feature image from open data according to one embodiment; FIG. 4 is a diagram for explaining an example of a method of generating a map image or a feature image from open data according to one embodiment; 4 is an example of multiple map or feature images classified into a landmark layer and a road network layer, respectively, according to one embodiment. FIG. 4 is a diagram illustrating an example of a method for generating an image of a specific point with a neural network structure, according to one embodiment; FIG. 10 is a diagram illustrating an example of a method for summing geographical likelihoods, according to one embodiment; FIG. 5 is a diagram illustrating another example of a method for generating an image of a specific point with a neural network structure, according to one embodiment; FIG. 4 is a flow diagram of a learning method according to one embodiment; FIG. 4 is a diagram showing an example of map images or feature images used in the learning method according to one embodiment; 12 illustrates an example computer 1200 in which aspects of the present invention may be implemented in whole or in part; FIG.

Hereinafter, the present invention will be described through embodiments of the invention, but the following embodiments do not limit the invention according to the claims. Also, not all combinations of features described in the embodiments are essential for the solution of the invention. In addition, in the drawings, the same or similar parts may be denoted by the same reference numerals to omit redundant description.

FIG. 1 is a schematic diagram of a system 10 according to one embodiment. The system 10 includes a communication terminal 30 and an inference device 100 that communicate with each other via a communication network 50 . The system 10 is a system for inferring the geographical location of the user 20 based on natural language expressions written in natural language from the user 20 who owns the communication terminal 30 . It does not matter whether the communication network 50 is wired or wireless.

The communication terminal 30 in this embodiment is, as an example, a communication device such as a smart phone. The communication terminal 30 communicates with the inference device 100 by circuit switching. In the system 10 according to this embodiment, the user 20 makes a call to the inference device 100 from the communication terminal 30 to make a voice call with the inference device 100 .

The inference device 100 in this embodiment identifies the latitude and longitude of the current location of the user 20 based on a natural language expression written in a natural language spoken by the user 20 during voice communication.

More specifically, during a voice call with the user 20, the inference apparatus 100 receives from the user 20 the address of the location where the user 20 is currently located and information about the location, such as landmarks visible to the user 20, The name of the nearest road or intersection is obtained from the user 20. - 特許庁For example, the address may be general address information at the municipality level, or may be specific address information at the street level. The inference device 100 identifies the latitude and longitude of the current location of the user 20 by performing rule-based search and machine learning inference from these natural language expressions.

A rule-based search may include, for example, using a geocoder to determine latitude and longitude from an address. The rule-based search may also include, for example, using Open Street Map (OSM) to acquire point data and line data within a certain area centered on the latitude and longitude. In addition, rule-based search, for example, identifies a position represented by an address, and if the position is not on a road or intersection, the position is replaced with a position on the road or intersection nearest to the position. may include correcting to

Inference by machine learning may be, for example, taking linguistic expressions and geographic data as input and inferring geographic locations indicated in language. Inference apparatus 100 performs rule-based search and machine learning inference even when the above-described natural language expression indicates a location that is vaguely described in natural language, so that the latitude and longitude of the location can be calculated as high as possible. Can be specified with precision.

As an example, the inference apparatus 100 of the present embodiment responds to a telephone call from the user 20 through the communication terminal 30 with automatic voice, thereby eliciting the natural language from the user 20 and writing it in the natural language. to obtain a natural language representation of Additionally or alternatively, the reasoning device 100 may obtain natural language expressions written in the above-described natural language dictated by the operator of the company that manages the reasoning device 100 from the user 20 .

Note that the address of the location where the user 20 is currently located, which the inference device 100 hears from the user 20 during voice communication with the user 20, is an example of a second language expression regarding the address of a specific location. Also, the above names of landmarks, roads, intersections, etc. that the inference device 100 hears from the user 20 during voice communication with the user 20 are examples of natural language expressions relating to specific points.

Communication terminal 30 and reasoning apparatus 100 may communicate using a packet switching system instead of the circuit switching system described above. It may be transmitted to the reasoning device 100 by communication.

FIG. 2 is a block diagram of a reasoning apparatus 100 according to one embodiment. The inference device 100 includes an acquisition unit 101 , an inference unit 103 , a calculation unit 105 , a correction unit 107 and an accident information database 109 .

Acquisition unit 101 communicates with a device external to inference device 100 , such as communication terminal 30 . The acquisition unit 101 may communicate with another communication terminal, for example, a computer of a management company that manages the inference device 100 .

The acquisition unit 101 acquires geographic data regarding an area including a specific point. More specifically, while communicating with the communication terminal 30, the acquisition unit 101 in this embodiment decodes the natural language about the current location of the user 20, which is spoken by the user 20 during the voice call, and obtains the natural language Get representation. The natural language may indicate, for example, the address of the point, or may indicate a specific road, landmark, or the like near OO IC on the highway or near OO station. It should be noted that the natural language expression is an example of a second language expression regarding a specific point.

When acquiring the natural language expression of the point in this way, the acquisition unit 101 in this embodiment acquires the open data of the area including the point where the user 20 is currently located based on the natural language expression. The acquisition unit 101 further generates a plurality of map images or feature images classified by at least one of landmarks, roads, and intersections from the open data.

As a specific example, the acquisition unit 101 in this embodiment converts the address into latitude and longitude using a geocoder, and uses an open street map (OSM) to calculate points within a certain area centered on the latitude and longitude. Get data or line data.

The acquisition unit 101 in the present embodiment further generates the plurality of map images or feature images from the point data and line data using the map drawing library. Note that open data, map images, and feature images are all examples of the geographic data described above.

The acquisition unit 101 outputs the acquired geographic data to the inference unit 103 . More specifically, the acquisition unit 101 in this embodiment outputs a plurality of map images or feature images generated from open data acquired by OSM to the inference unit 103 .

The acquisition unit 101 in the present embodiment further acquires a natural language expression regarding a specific point, and converts the natural language expression into a vector expression regarding the specific point. As a specific example, when the acquisition unit 101 in this embodiment acquires a natural language expression from the natural language about the current location of the user 20 while communicating with the communication terminal 30, Word2Vec is The natural language representation is converted into a vector representation using the parser used. Acquisition unit 101 outputs the vector representation to inference unit 103 . Note that the vector representation is an example of a first language representation of a specific point.

Further, when the acquisition unit 101 in this embodiment receives the latitude and longitude input from the correction unit 107 , the acquisition unit 101 transmits the latitude and longitude to, for example, a computer of a management company that manages the inference device 100 .

The inference unit 103 acquires a first language expression related to a specific point based on the learning result of the correspondence relationship between the first language expression related to the arbitrary point and the geographic data related to the area including the arbitrary point, and the arbitrary point. The particular point is inferred from the geographic data obtained by unit 101 .

More specifically, the inference unit 103 in this embodiment identifies a plurality of linguistic feature amounts from the first linguistic expression related to a specific point, and also identifies a plurality of map maps included in the geographic data acquired by the acquisition unit 101. A plurality of image feature values are specified from an image or a feature image. In this embodiment, the first language expression is the vector expression described above, and the vector expression regarding a specific point is input from the acquisition unit 101 to the inference unit 103 as described above.

The inference unit 103 further includes a plurality of linguistic feature amounts specified from the first language expression regarding an arbitrary point, and a plurality of map images or feature images included in the geographic data regarding the area containing the arbitrary point. A specific point is inferred from a plurality of image feature amounts and a plurality of linguistic feature amounts related to the specific point, based on learning results of correspondence relationships between the plurality of image feature amounts and the arbitrary point.

As a specific example, the inference unit 103 in this embodiment has a deep learning model with a neural network structure including an LSTM (Long Short-Term Memory) network and a CNN (Convolution Neural Network). The inference unit 103 identifies a plurality of linguistic feature amounts from the vector representation using the LSTM network, and identifies a plurality of image feature amounts from the plurality of map images or feature images using the CNN.

In a specific example, the inference unit 103 uses a plurality of linguistic feature amounts specified from a vector representation of an arbitrary point, and a plurality of map images or feature images of the arbitrary point, using the deep learning model described above. Deep learning is performed on the correspondence relationship between the specified plurality of image feature values and the arbitrary point. The corresponding relationship includes the degree of association between the language feature amount and the image feature amount. The inference unit 103 reads, for example, a vector representation and a plurality of map images or feature images relating to an arbitrary point stored in the accident information database 109, identifies the language feature amount and the image feature amount, respectively, and performs the correspondence. Deep learning relationships.

The accident information database 109 stores natural language expressions for arbitrary points and natural language expressions for addresses of arbitrary points instead of vector expressions and multiple map images or feature images for arbitrary points. good too. In this case, the inference unit 103 outputs these natural language expressions read from the accident information database 109 to the acquisition unit 101, thereby obtaining from the acquisition unit 101 a vector expression and a plurality of map images or feature images regarding an arbitrary point. may be entered.

In a specific example, the inference unit 103 further generates an image of a specific point by adding together geographical likelihoods indicated by a plurality of image feature amounts corresponding to a plurality of linguistic feature amounts related to the specific point. do. The geographical likelihood indicated by the image feature amount is, for example, the degree of likelihood of the location represented by the linguistic expression that is the basis of the linguistic feature amount corresponding to each image feature amount, that is, the position where the user 20 is currently located. It is a mapping within an image and can be handled including range, multiple locations, and ambiguity. Generating an image of the specific point corresponds to inferring the specific point. The inference unit 103 outputs the image of the specific spot to the calculation unit 105 .

In this embodiment, the calculation unit 105 determines the point with the highest geographical likelihood as the specific point in the image of the specific point generated by the inference unit 103, and calculates the latitude and longitude of the determined point. calculate. The image is associated with latitude and longitude information of the open data that is the source of the image. Calculation unit 105 outputs the calculated latitude and longitude to correction unit 107 .

In this embodiment, the correction unit 107 determines whether or not the point specified by the latitude and longitude calculated by the calculation unit 105 is located on the road. If the point is not located on the road, the correction unit 107 replaces the latitude and longitude calculated by the calculation unit 105 with the latitude and longitude of the point closest to the specified point on the road closest to the specified point. and corrected for longitude. As an example, the correction unit 107 may acquire from the inference unit 103 via the calculation unit 105 the image feature amount specified from the map image of the road in order to perform these determinations and corrections. Correction unit 107 outputs the latitude and longitude to acquisition unit 101 .

Note that the inference apparatus 100 does not have to include the calculator 105 and the corrector 107 . For example, the inference apparatus 100 may cause an external device to execute each process of the calculation unit 105 and the correction unit 107 described above. An example of an external device may be the computer of the management company mentioned above.

In this embodiment, the accident information database 109 stores each data described above. Each data may be data indicating a location where a vehicle accident occurred in the past. Note that the inference apparatus 100 may include a memory separate from the accident information database 109 that stores programs and the like for controlling the entire inference apparatus 100 , and the memory may include the accident information database 109 . Each component of the inference apparatus 100 reads the program stored in the memory in order to execute each function described above.

Note that the inference device 100 does not have to include the accident information database 109 . For example, the inference device 100 may acquire each of the above-described data from a database of an external device and execute the above-described inference and learning.

FIG. 3 is a flow diagram of an inference method according to one embodiment. As an example, the flow shown in FIG. 3 may be started by the user 20 using the communication terminal 30 to make a voice call to the reasoning device 100, and the reasoning device 100 starting to respond to the user 20 with automatic voice. good. The user 20 speaks the natural language required for the flow according to the automatic voice guidance.

The inference device 100 acquires (1) an address language expression related to the address of a specific point and (2) a point language expression related to the specific point from the natural language uttered by the user 20 (step S101). An example of the address language expression is the language expression “Chuo-ku, Kobe City, Hyogo Prefecture”. An example of a point linguistic expression is the linguistic expression "near the intersection of national highway A and national highway B, near a convenience store in Sannomiya-cho."

The inference device 100 (1) acquires open data of an area containing the address, generates Ni map images or feature images from the open data, and (2) converts the point language expression into a vector expression of length Nv. Convert (step S103).

FIG. 4 is an example of an area 60 that includes a particular point where user 20 is located, according to one embodiment. In FIG. 4, the latitude and longitude of the area 60 are indicated by vectors. The area 60 is, for example, a 5 km square square area centered on the latitude and longitude of the address indicated by the linguistic expression "Chuo-ku, Kobe City, Hyogo Prefecture".

As shown in FIG. 4, an area 60 includes a user 20 operating a communication terminal 30 beside his/her own vehicle, four roads that intersect at four intersections, and three convenience stores located along each road. included. In addition, "Sannomiya-cho" is written on the area 60 only for the purpose of clarifying the description.

The inference apparatus 100 uses a geocoder to identify the latitude and longitude from the address indicated by the linguistic expression “Chuo-ku, Kobe-shi, Hyogo”. In this way, when the address indicated in the address language expression is at the municipal level, the address is not a point but a range. may be specified. Alternatively, the reasoning apparatus 100 may acquire open data based on the range. If the address shown in the address language expression includes a house number and a number, the address is a point, so the latitude and longitude of the point may be specified.

The inference apparatus 100 further determines an area 60 centered on the latitude and longitude, acquires a plurality of point data and line data of the area 60 using OSM, and extracts Ni pieces, for example 150 points, from these open data. Generate individual map and feature images.

FIG. 5 is a diagram for explaining an example of a method for generating a map image or feature image from open data according to one embodiment. FIG. 5 shows, from top to bottom, an OSM point data (landmark) table, a graph of the geographical likelihood of each landmark shown in the table, and a feature image of each landmark based on the graph. indicates

In this table, the name of a broadly classified group called OSM point data (landmarks) is described in the first row, and the type of landmark, the name of each landmark, and the The latitude of the mark and the longitude of each landmark are described. Also, each point data is stored in each cell in the third and subsequent rows of the table.

FIG. 6 is a diagram for explaining an example of a method of generating a map image or feature image from open data according to one embodiment. FIG. 6 shows, from top to bottom, a line data (road) table and a point data (intersection) table of OSM, and a graph of the geographical likelihood of each of the roads and intersections shown in the table, as in FIG. , and a map image of roads and a map image of intersections based on the graph.

In the table shown on the top left side of FIG. 6, the name of a broad group of OSM line data (roads) is described in the first row, and the name of the road and the line data are described in each row after the second row. combinations are stored.

In the table shown on the uppermost right side of FIG. 6, the name of a group of OSM point data (intersections) is listed in the first row, and the name of the intersection and the name of each intersection are written in order from the first column of the second row. The latitude and longitude of each intersection are described. Also, each point data is stored in each cell in the third and subsequent rows of the table.

As illustrated in FIGS. 5 and 6, the reasoning apparatus 100 uses a map drawing library to generate Ni map images and feature images from point data and line data such as landmarks, roads, and intersections in OSM. Generate.

FIG. 7 is an example of multiple map or feature images classified into a landmark layer and a road network layer, respectively, according to one embodiment. The inference device 100 may classify the Ni map images and feature images into the landmark layer and the road network layer, respectively, as shown in FIG.

The landmark layer is a different layer for each landmark, for example, for each type of store, building, landform, landmark, etc., and includes several hundred layers, for example. The inference device 100 classifies, for example, the feature images of the landmarks shown in FIG. 5 into one of the landmark layers.

Also, the road network layer includes sidewalks, roadways, intersections, highways, etc., and includes about 10 layers as an example. The inference device 100 classifies, for example, the map image of roads and intersections shown in FIG. 6 into one of the road network layers.

The reasoning apparatus 100 uses a syntactic analyzer to convert the point language expression "near the intersection of national highway A and national highway B near a convenience store in Sannomiya-cho" to "Sannomiya-cho/no/convenience store/nearby/no/national highway A /and/National road B//intersection/nearby” and embedded in a vector of length Nv. The number Nv of words (tokens) in the vector is variable and includes dependencies.

It should be noted that steps S101 to S103 are an example of acquisition steps for acquiring geographic data relating to an area including a specific point.

Returning to the description of the flow of FIG. 3, the inference device 100 (1) identifies a plurality of image feature amounts from Ni map images or feature images, and (2) identifies a plurality of language features from a vector representation of length Nv. The amount is specified (step S105). The inference device 100 further infers a specific point from the plurality of image feature amounts and the plurality of language feature amounts (step S107). More specifically, the inference apparatus 100 generates an image of a specific spot by summing up geographical likelihoods indicated by multiple image feature quantities corresponding to multiple language feature quantities.

FIG. 8 is a diagram illustrating an example of a method for generating an image of a specific point with a neural network structure, according to one embodiment. The reasoning apparatus 100 has a deep learning model of neural network structure including LSTM network and CNN, as shown in FIG.

The inference apparatus 100 identifies a plurality of linguistic feature values (Mv) from a linguistic expression (Nv), which is a vector expression, by an LSTM network, and from a plurality of map images or feature images (256×256×Ni) by CNN. A plurality of image feature quantities (256×256×Mi) are specified.

The inference apparatus 100 multiplies the i-th layer matrix of the image feature amount by the i-th value of the language feature amount to obtain the i-th layer feature amount, and generates M pieces of this. The inference device 100 further adds up the generated M images to generate an image for estimating a specific point where the user 20 is located by CNN.

When multiplying the i-th layer matrix of the image feature amount and the i-th value of the language feature amount, the inference apparatus 100 regards each value of the language feature amount as a weight for each layer of the image feature amount and multiplies it. may The inference apparatus 100 may obtain the weight of each layer through learning. The inference device 100 may obtain, for example, weights of “Sannomiyacho: convenience store: intersection=3:6:1” as a result of learning.

FIG. 9 is a diagram illustrating an example of a method for summing geographical likelihoods, according to one embodiment. FIG. 9 shows, for the purpose of clarifying the explanation, an image diagram of adding up the geographical likelihood graph of each landmark shown in FIG. 5 and an image diagram of adding up the feature image of each landmark based on the graph. are arranged one above the other. The inference apparatus 100 generates an image of a specific point by adding up the geographical likelihoods indicated by the plurality of image feature quantities corresponding to the plurality of language feature quantities, as shown in the image diagram of FIG.

FIG. 10 is a diagram for explaining another example of a method for generating an image of a specific point with a neural network structure, according to one embodiment. The inference apparatus 100 in this other example embodiment differs from the embodiment shown in FIG. is specified, and a plurality of image feature amounts (256×256×Mi) are specified from a plurality of map images or feature images (256×256×Ni) by CNN.

The reasoning apparatus 100 according to the other example embodiment matches the sizes of the layers of the language feature amount and the layers of the image feature amount, and stacks the layers to obtain 2M layers. The inference device 100 further adds up the 2M layers to generate an image that estimates the specific point where the user 20 is located by CNN.

Note that steps S105 to S107 are based on the learning result of the correspondence relationship between the first language expression regarding the arbitrary point for learning and the geographic data regarding the area including the arbitrary point and the arbitrary point, and the specific point 2 is an example of an inference stage for inferring a particular point from a second linguistic representation of and geographic data acquired in the acquisition stage;

Returning to the description of the flow in FIG. 3, the inference apparatus 100 determines the point with the highest geographical likelihood in the image generated in step S107 as a specific point, and calculates the latitude and longitude of the determined point. (Step S109). The inference apparatus 100 further determines whether the point identified by the calculated latitude and longitude is located on the road (step S111).

When the inference apparatus 100 determines that the point is not located on the road (step S111: YES), the latitude and longitude are the closest road points to the point specified by the latitude and longitude. The latitude and longitude are corrected to those of the nearest point from the point (step S113), the corrected latitude and longitude are sent to, for example, the computer of the management company of the reasoning apparatus 100, and the flow ends.

Reasoning apparatus 100 may utilize OSM to perform the above correction. For example, the reasoning apparatus 100 may retrieve latitude and longitude data of roads and features using OSM, and obtain the latitude and longitude of the nearest position on the road from the latitude and longitude. For example, the inference device 100 may use supplementary information about the location of the accident, such as roads and intersections, stored in the accident information database 109 to perform the above correction.

When the inference apparatus 100 determines that the point is located on the road (step S111: NO), the inference apparatus 100 transmits the calculated latitude and longitude as they are to the computer, for example, and terminates the flow.

FIG. 11 is a flow diagram of a learning method according to one embodiment. Inference apparatus 100 executes the learning method shown in FIG. 11 before executing the inference method shown in FIG. Inference apparatus 100 may additionally execute the learning method shown in FIG. 11 while the inference method shown in FIG. 3 is not being executed.

The inference apparatus 100 determines an area to be learned, acquires open data of the area, and generates Ni map images or feature images from the open data (step S301). The area may be about 5 km square, as shown in FIG. 3, for example. The inference apparatus 100 generates Ni map images or feature images by performing the same processing as (1) in step S103 of the flow in FIG. 3, for example. Note that the inference apparatus 100 repeatedly uses the Ni map images or feature images generated in step S301 as fixed while executing the learning flow shown in FIG.

The reasoning apparatus 100 acquires, as teacher data, (1) a point language expression regarding an arbitrary point within the area, and (2) correct information about the latitude and longitude of the point (step S303). In other words, the reasoning apparatus 100 uses data such as places where accidents have occurred in the past, and sets pairs of the linguistic expression of the accident place and the latitude and longitude of the point as teacher data.

The inference device 100 (1) converts the point language representation into a vector representation of length Nv, and (2) generates a map image of the latitude and longitude of the correct information (step S305). The inference apparatus 100 identifies the vector representation of the length Nv by performing, for example, the same processes as (1) and (2) in steps S103 in the flow of FIG. Generate an image.

The reasoning apparatus 100 receives Ni map images or feature images and a vector representation of length Nv as input, and performs deep learning with a map image based on correct information as an output (step S307), and performs deep learning from steps S303 to S307 until sufficient learning is performed. After repeating (step S309), the flow ends.

In this way, the inference apparatus 100 performs deep learning by fixing the input of Ni map images or feature images, and sequentially providing pairs of linguistic expressions and correct images as inputs and outputs, respectively.

FIG. 12 is a diagram showing an example of a map image or feature image used in the learning method according to one embodiment. FIG. 12 shows an example of a pair of a natural language expression automatically generated by the reasoning apparatus 100 according to the embodiment and a map image or a feature image.

In addition to using actual information such as accident data, the reasoning apparatus 100 uses OSM geographic data and crowdsourcing to compensate for the amount of data, so that pairs such as those shown in FIG. can be automatically generated.

The inference apparatus 100 may automatically generate combinations of various linguistic expressions such as addresses, feature names, and correction information, as shown in FIG. The reasoning device 100 may also generate arbitrary map images or feature images such as single points, multiple points, ranges, etc., as shown in FIG.

When using automatically generated data of natural language expressions, the inference apparatus 100 first receives a map image or a feature image and an automatically generated language expression as input, and performs pre-learning using the automatically generated image feature amount as correct data. conduct. Next, the inference apparatus 100 performs additional learning according to the learning method of FIG. However, the inference apparatus 100 uses the same “map image or feature image” for both pre-learning and additional learning. That is, the reasoning apparatus 100 uses the map image or feature image generated in step S301 of the flow of FIG. 11 for both pre-learning and additional learning.

When using only actual data as shown in FIG. 11, it may be difficult to prepare a necessary and sufficient number of data. Therefore, the inference apparatus 100 may perform pre-learning with a large amount of automatically generated data as shown in FIG. 12 and then perform additional learning with a small amount of actual data to improve accuracy. Inference apparatus 100 may improve the accuracy of the learning model by supplementing the number with automatically generated data when the number of actual data is not sufficient.

Also, the inference apparatus 100 may perform re-learning (fine-tuning). Specifically, the inference device 100 learns a general-purpose base model from data of various areas, and re-learns with data of a specific area based on the base model, thereby achieving high accuracy with a small amount of data. is also possible. Alternatively, the inference apparatus 100 may be applicable to various regions without re-learning as described above by simply replacing the geographic image to be input after learning a general-purpose base model.

As described above, according to the inference apparatus 100 and the inference method according to the above-described multiple embodiments, geographic data relating to an area including a specific point is acquired, and a first language expression relating to an arbitrary point and an area including an arbitrary point A specific point is inferred from the first language representation of the specific point and the acquired geographic data based on the learning result of the correspondence between the geographic data and the arbitrary points.

For example, past accident data stored in an existing accident database owned by an insurance company may not accurately record the address of the accident location. In addition, since the free description of the location of the accident recorded by the person in charge of receiving the accident is stored in the accident database, the contents of the description are often ambiguous. Also, the address of the accident site and the accident location do not exactly match. The address is the location of the building, but most of the accident locations are on roads, so there is a discrepancy between the address of the accident location and the accident location.

According to the inference apparatus 100 and the inference method having the above configuration, for example, information indicating the location of the accident vaguely expressed in natural language is narrowed down to a range such as a town-chome, for example, based on rules, and then machine learning is performed. can be used to estimate the accident location in more detail. For example, according to the inference apparatus 100 and the inference method having the above configuration, by estimating the latitude and longitude from the linguistic expression of the accident location in the accident data recorded in the past, it is possible to determine where accidents frequently occur. It can be used for geographical statistical analysis of accident data, such as analysis, and to create detailed risk maps.

Further, for example, according to the reasoning device 100 and the reasoning method having the above configuration, when receiving a call from a call center of an insurance company or the like, a question for specifying a more specific position can be instructed to assist in specifying the location of the accident. Also, it is possible to improve the accuracy of identifying the accident location.

In the above multiple embodiments, the inference unit 103 additionally gradually shifts from a first area including an arbitrary point to a second area that is smaller than the first area and includes an arbitrary point. You may learn the correspondence of the several language feature-value and several image feature-value regarding, and arbitrary points. In this case, the inference unit 103 sequentially generates a plurality of image feature values and a plurality of A specific point may be inferred from the linguistic feature of

For example, the inference unit 103 may change the size of the area step by step to a range of 2000 km square, a range of 100 km square, and a range of 5 km square, and learn the model in each range. In other words, the inference unit 103 may divide the resolution of the image used in machine learning into a plurality of resolutions and gradually narrow down the range from a wide range to a narrow range. The inference unit 103 may be configured to estimate a detailed position by narrowing down the location gradually from a wide range. In other words, the inference unit 103 may make inferences in order from wide areas and gradually narrow down the area.

In the system 10 according to the above multiple embodiments, the inference device 100 is configured to infer the current location of the user 20 based on the verbal expression from the user 20, which is obtained from the communication terminal 30 via the communication network 50. . Additionally or alternatively, the system 10 may include a reasoning device 100 mounted on the vehicle of the user 20, a reasoning device 100 mounted on the communication terminal 30 of the user 20, or the like.

Further, in the system 10 according to the above multiple embodiments, the inference device 100 is configured to infer the current location of the user 20 . Additionally or alternatively, reasoning device 100 may reason about the intended destination of user 20 . In this case, the inference apparatus 100 additionally uses context information such as the current position and direction of travel of the user 20 to determine the intended destination of the user 20, i.e., the position relative to the current position of the user 20. can be inferred. As an example, the inference device 100 may display the inferred route information to the destination on the communication terminal 30, for example, the car navigation display of the vehicle, in order to inform the user 20 of the inferred route to the destination. As an example, the inference device 100 may input route information to the inferred destination into the vehicle in order to automatically drive the vehicle toward the inferred destination. Note that the inference apparatus 100 may determine the area based on the current position, direction, etc. instead of using the geocoder when inferring the relative position from the current position of the user 20 .

Various embodiments of the invention may be described with reference to flowchart illustrations and block diagrams, where blocks refer to (1) steps in a process in which operations are performed or (2) devices responsible for performing the operations. may represent a section of Certain steps and sections may be implemented by dedicated circuitry, programmable circuitry provided with computer readable instructions stored on a computer readable medium, and/or processor provided with computer readable instructions stored on a computer readable medium. you can Dedicated circuitry may include digital and/or analog hardware circuitry, and may include integrated circuits (ICs) and/or discrete circuitry. Programmable circuits include logic AND, logic OR, logic XOR, logic NAND, logic NOR, and other logic operations, memory elements such as flip-flops, registers, field programmable gate arrays (FPGAs), programmable logic arrays (PLAs), etc. and the like.

Computer-readable media may include any tangible device capable of storing instructions to be executed by a suitable device, such that computer-readable media having instructions stored thereon may be designated in flowcharts or block diagrams. It will comprise an article of manufacture containing instructions that can be executed to create means for performing the operations described above. Examples of computer-readable media may include electronic storage media, magnetic storage media, optical storage media, electromagnetic storage media, semiconductor storage media, and the like. More specific examples of computer readable media include floppy disks, diskettes, hard disks, random access memory (RAM), read only memory (ROM), erasable programmable read only memory (EPROM or flash memory), Electrically Erasable Programmable Read Only Memory (EEPROM), Static Random Access Memory (SRAM), Compact Disc Read Only Memory (CD-ROM), Digital Versatile Disc (DVD), Blu-ray (RTM) Disc, Memory Stick, Integration Circuit cards and the like may be included.

The computer readable instructions may be assembler instructions, Instruction Set Architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state setting data, or instructions such as Smalltalk, JAVA, C++, etc. any source or object code written in any combination of one or more programming languages, including object-oriented programming languages, and conventional procedural programming languages such as the "C" programming language or similar programming languages; may include

Computer readable instructions may be transferred to a processor or programmable circuitry of a general purpose computer, special purpose computer, or other programmable data processing apparatus, either locally or over a wide area network (WAN), such as a local area network (LAN), the Internet, or the like. ) and may be executed to create means for performing the operations specified in the flowcharts or block diagrams. Examples of processors include computer processors, processing units, microprocessors, digital signal processors, controllers, microcontrollers, and the like.

FIG. 13 illustrates an example computer 1200 in which aspects of the invention may be implemented in whole or in part. Programs installed on the computer 1200 cause the computer 1200 to function as one or more "parts" of operations or one or more "parts" of an apparatus according to embodiments of the invention, or to and/or cause computer 1200 to perform processes or steps of processes according to embodiments of the present invention. Such programs may be executed by CPU 1212 to cause computer 1200 to perform certain operations associated with some or all of the blocks in the flowcharts and block diagrams described herein.

Computer 1200 according to this embodiment includes CPU 1212 , RAM 1214 , graphics controller 1216 , and display device 1218 , which are interconnected by host controller 1210 . Computer 1200 also includes input/output units such as communication interface 1222 , hard disk drive 1224 , DVD-ROM drive 1226 , and IC card drive, which are connected to host controller 1210 via input/output controller 1220 . The computer also includes legacy input/output units such as ROM 1230 and keyboard 1242 , which are connected to input/output controller 1220 through input/output chip 1240 .

The CPU 1212 operates according to programs stored in the ROM 1230 and RAM 1214, thereby controlling each unit. Graphics controller 1216 takes image data generated by CPU 1212 into a frame buffer or the like provided in RAM 1214 or into itself, and causes the image data to be displayed on display device 1218 .

Communication interface 1222 communicates with other electronic devices over a network. Hard disk drive 1224 stores programs and data used by CPU 1212 within computer 1200 . DVD-ROM drive 1226 reads programs or data from DVD-ROM 1201 and provides programs or data to hard disk drive 1224 via RAM 1214 . The IC card drive reads programs and data from IC cards and/or writes programs and data to IC cards.

ROM 1230 stores therein programs that depend on the hardware of computer 1200, such as a boot program that is executed by computer 1200 upon activation. Input/output chip 1240 may also connect various input/output units to input/output controller 1220 via parallel ports, serial ports, keyboard ports, mouse ports, and the like.

A program is provided by a computer-readable storage medium such as a DVD-ROM 1201 or an IC card. The program is read from a computer-readable storage medium, installed in hard disk drive 1224 , RAM 1214 , or ROM 1230 , which are also examples of computer-readable storage medium, and executed by CPU 1212 . The information processing described within these programs is read by computer 1200 to provide coordination between the programs and the various types of hardware resources described above. An apparatus or method may be configured by implementing information operations or processing according to the use of computer 1200 .

For example, when communication is performed between the computer 1200 and an external device, the CPU 1212 executes a communication program loaded into the RAM 1214 and sends communication processing to the communication interface 1222 based on the processing described in the communication program. you can command. Under the control of the CPU 1212, the communication interface 1222 reads transmission data stored in a transmission buffer area provided in a recording medium such as the RAM 1214, the hard disk drive 1224, the DVD-ROM 1201, or an IC card. Data is transmitted to the network, or received data received from the network is written in a receive buffer area or the like provided on the recording medium.

In addition, the CPU 1212 causes the RAM 1214 to read all or necessary portions of files or databases stored in external recording media such as a hard disk drive 1224, a DVD-ROM drive 1226 (DVD-ROM 1201), an IC card, etc. Various types of processing may be performed on the data in RAM 1214 . CPU 1212 may then write back the processed data to an external recording medium.

Various types of information, such as various types of programs, data, tables, and databases, may be stored on the recording medium to be processed. CPU 1212 performs various types of operations on data read from RAM 1214, information processing, conditional decisions, conditional branching, unconditional branching, and information retrieval, which are described throughout this disclosure and are specified by instruction sequences of programs. Various types of processing may be performed, including /replace, etc., and the results written back to RAM 1214 . In addition, the CPU 1212 may search for information in a file in a recording medium, a database, or the like. For example, when a plurality of entries each having an attribute value of a first attribute associated with an attribute value of a second attribute are stored in the recording medium, the CPU 1212 selects the first attribute from among the plurality of entries. search for an entry that matches a condition in which the attribute value of the attribute is specified, read the attribute value of the second attribute stored in the entry, and thereby determine the first attribute that satisfies the predetermined condition An attribute value of the associated second attribute may be obtained.

Programs or software modules according to the above description may be stored in a computer readable storage medium on or near computer 1200 . Also, a recording medium such as a hard disk or RAM provided in a server system connected to a dedicated communication network or the Internet can be used as a computer-readable storage medium, whereby the program can be transferred to the computer 1200 via a network. offer.

Although the present invention has been described above using the embodiments, the technical scope of the present invention is not limited to the scope described in the above embodiments. It is obvious to those skilled in the art that various modifications and improvements can be made to the above embodiments. In addition, matters described with respect to a specific embodiment can be applied to other embodiments as long as they are not technically inconsistent. Also, each component may have features similar to other components with the same name but different reference numerals. It is clear from the description of the scope of claims that forms with such modifications or improvements can also be included in the technical scope of the present invention.

For example, a rule-based search acquires geographic data, such as tile coordinates and mesh codes, for an area represented by a range of latitude and longitude, including a specific point, from GPS information in a car navigation system or smartphone. may include That is, as a specific example of acquiring geographic data relating to an area including a specific point, the acquisition unit may acquire geographic data of an area including the specific point from GPS information.

The execution order of each process such as actions, procedures, steps, and stages in the devices, systems, programs, and methods shown in the claims, the specification, and the drawings is particularly "before", "before etc., and it should be noted that they can be implemented in any order unless the output of the previous process is used in the subsequent process. Regarding the operation flow in the claims, the specification, and the drawings, even if the description is made using "first," "next," etc. for the sake of convenience, it means that it is essential to carry out in this order. not a thing

10 system 20 user 30 communication terminal 50 communication network 60 area 100 inference device 101 acquisition unit 103 inference unit 105 calculation unit 107 correction unit 1200 computer 1201 DVD-ROM
1210 host controller 1212 CPU
1214 RAM
1216 graphic controller 1218 display device 1220 input/output controller 1222 communication interface 1224 hard disk drive 1226 DVD-ROM drive 1230 ROM
1240 input/output chip 1242 keyboard

Claims (9)

an acquisition unit that acquires geographic data about an area that includes a specific point;
A first linguistic expression relating to an arbitrary point and a first linguistic expression relating to the specific point and acquired by the acquiring unit based on learning results of a correspondence relationship between the first linguistic expression relating to the arbitrary point and geographic data relating to an area including the arbitrary point and the arbitrary point. an inference unit that infers the specific point from the geographic data obtained ;
The reasoning unit
a plurality of linguistic features specified from the first language expression relating to the arbitrary point, and a plurality of map images or feature images specified from the plurality of map images or feature images included in the geographic data relating to the area containing the arbitrary point Learning the correspondence relationship between the image feature quantity and the arbitrary point,
identifying a plurality of image feature amounts from a plurality of map images or feature images included in the geographic data acquired by the acquisition unit;
Identifying a plurality of linguistic features from the first linguistic expression related to the specific point;
Inferring the specific point from the plurality of image features and the plurality of language features related to the specific point;
reasoning device. The reasoning unit
The plurality of linguistic features and the plurality of images relating to the arbitrary point, step by step from the first area including the arbitrary point to the second area including the arbitrary point smaller than the first area Learning the correspondence relationship between the feature quantity and the arbitrary point,
The plurality of image feature values and the plurality of languages related to the specific point, step by step from a third area including the specific point to a fourth area including the specific point that is smaller than the third area inferring the specific point from the features;
A reasoning apparatus according to claim 1 . The inference unit generates an image of the specific point by adding together the geographical likelihoods indicated by the plurality of image feature amounts respectively corresponding to the plurality of language feature amounts related to the specific point.
3. A reasoning device according to claim 1 or 2 . A calculation unit that determines a point having the highest geographical likelihood as the specific point in the image of the specific point generated by the inference unit, and calculates the latitude and longitude of the determined point. prepare
4. A reasoning apparatus according to claim 3 . When the point specified by the latitude and longitude calculated by the calculating unit is not located on a road, the latitude and longitude calculated by the calculating unit are placed on the road closest to the specified point. In, further comprising a correction unit that corrects the latitude and longitude of the closest point from the identified point,
5. A reasoning apparatus according to claim 4 . The acquisition unit acquires open data of an area including the specific point based on the second language expression when the second language expression regarding the specific point is acquired, and from the open data, landmarks, generating the plurality of map images or feature images classified by at least one of roads and intersections;
A reasoning apparatus according to any one of claims 1 to 5 . The acquisition unit acquires a natural language expression regarding the specific point, converts the natural language expression into the first language expression that is a vector expression regarding the specific point, and outputs the first language expression to the inference unit.
A reasoning apparatus according to any one of claims 1 to 6 . A computer-implemented inference method comprising:
an acquisition step of acquiring geographic data for an area containing a particular point;
A first linguistic expression relating to an arbitrary point and a first linguistic expression relating to the specific point and acquired in the acquiring step based on learning results of a correspondence relationship between the first linguistic expression relating to the arbitrary point and geographic data relating to an area including the arbitrary point and the arbitrary point. an inferring step of inferring the particular point from the geographic data obtained from
The inference step includes:
a plurality of linguistic features specified from the first language expression relating to the arbitrary point, and a plurality of map images or feature images specified from the plurality of map images or feature images included in the geographic data relating to the area containing the arbitrary point learning the correspondence relationship between the image feature amount and the arbitrary point;
identifying a plurality of image feature amounts from a plurality of map images or feature images included in the geographic data acquired in the acquisition step;
identifying a plurality of linguistic features from the first linguistic expression relating to the specific point;
inferring the specific point from the plurality of image features and the plurality of linguistic features relating to the specific point;
including,
reasoning method. to the computer,
an acquisition procedure for acquiring geographic data about an area containing a particular point;
A first linguistic expression relating to an arbitrary point and geographic data relating to an area including the arbitrary point, and a first linguistic expression relating to the specific point and acquired by the acquisition procedure based on learning results of a correspondence relationship between the arbitrary point and the arbitrary point a program for executing an inference procedure for inferring the specific point from the geographic data obtained by
The inference procedure includes:
a plurality of linguistic features specified from the first language expression relating to the arbitrary point, and a plurality of map images or feature images specified from the plurality of map images or feature images included in the geographic data relating to the area containing the arbitrary point learning the correspondence relationship between the image feature amount and the arbitrary point;
identifying a plurality of image feature amounts from a plurality of map images or feature images included in the geographic data acquired in the acquisition step;
identifying a plurality of linguistic features from the first linguistic expression relating to the specific point;
inferring the specific point from the plurality of image features and the plurality of linguistic features relating to the specific point;
including,
program .

Images