JP2023073104A - Matching device, matching method, program, and recording medium - Google Patents

Abstract

To improve accuracy in a sequential pattern matching technique.SOLUTION: A matching device (1) comprises: an acquisition section (11) that acquires object information to be matched; a first matching section (12) that performs pattern matching for the object information by executing a search, which is aimed at each of a plurality of pieces of partial information included in the object information and in which the number of search paths is limited to a predetermined number or smaller number, using a state model corresponding to a regular expression; a second matching section (15) that performs pattern matching for the object information by executing a beam search aimed at each of the plurality of pieces of partial information included in the object information without using the state model corresponding to the regular expression; and a selection section (16) that outputs a matching result by the first matching section or a matching result by the second matching section.SELECTED DRAWING: Figure 11

Description

The present invention relates to a matching device, a matching method, a program, and a recording medium that perform matching with target information.

Conventionally, in pattern matching techniques such as OCR (Optical Character Recognition), a search algorithm is also used to sequentially identify a plurality of characters. On the other hand, beam search is known as an example of such a search algorithm (for example, Non-Patent Document 1). In the beam search, only a predetermined number of candidates with the highest probability are left in each search step, and the search step is repeated.

“Towards End-To-End Speech Recognition with Recurrent Neural Networks,” Alex Graves, Navdeep Jaitly, Proceedings of the 31st International Conference on Machine Learning, PMLR 32(2); 1764-1772, 2014. “Regular expressions for decoding of neural network outputs,” Tobias Strauss et.al, arXiv:1509.04438v2, 2016.

By using beam search in pattern matching technology such as OCR, it is expected to improve accuracy compared to search algorithms such as greedy search. It is possible to shorten the search time compared to the conventional method.

However, even if the beam search is used, there is still room for improvement in terms of accuracy deterioration due to misrecognition of characters and the like. In this regard, Non-Patent Document 2 proposes a beam search using a nondeterministic finite automaton (NFA) corresponding to regular expressions. However, if there is no a priori information about the type of information to be recognized, even if the method of Non-Patent Document 2 is used, it is difficult to improve accuracy. There is

In order to solve the above problems, a matching device according to an aspect of the present invention uses an acquisition unit that acquires target information to be matched, and a state model that corresponds to a certain regular expression. a first matching unit that performs pattern matching on the target information by executing a search that targets each of a plurality of included partial information and limits the number of search paths to a predetermined number or less; A beam search targeting each of a plurality of partial information included in the target information without using a state model corresponding to a regular expression, the beam search limiting the number of search paths to a predetermined number. a second matching unit that performs pattern matching with respect to the target information; and a selection unit that selects according to the probability information of the matching result from the second matching unit.

In order to solve the above problems, a matching method according to an aspect of the present invention includes an acquisition step of acquiring target information to be matched; a first matching step of performing pattern matching on the target information by executing a search targeting each of a plurality of included partial information, the search limiting the number of search paths to a predetermined number or less; A beam search targeting each of a plurality of partial information included in the target information without using a state model corresponding to a regular expression, the beam search limiting the number of search paths to a predetermined number. a second matching step that performs pattern matching on the target information, and either the matching result of the first matching step or the matching result of the second matching step as the matching result of the first matching step and a selection step of selecting according to the probability information of the matching result of the second matching step.

According to one aspect of the present invention, improved accuracy can be obtained even in the absence of a priori information about what type of information is to be recognized.

1 is a block diagram showing the configuration of a matching device according to Embodiment 1 of the present invention; FIG. FIG. 4 is a diagram showing an example of a regular expression and an example of a state model referred to by the matching device according to Embodiment 1 of the present invention; FIG. 4 is a flow chart showing the flow of matching processing executed by the matching device according to the first embodiment of the present invention; FIG. 5 is a diagram showing an example of a plurality of matching candidates and probability information acquired by the matching unit according to the first embodiment of the present invention; FIG. 10 is a diagram illustrating possible transitions from the current state searched using each matching candidate in a beam search according to a comparative example; 6 is a diagram showing a search path corresponding to FIG. 5 superimposed on the prediction result shown in FIG. 4; FIG. FIG. 5 is a diagram showing an example of a state model used for pattern matching performed by the matching unit according to the first embodiment of the present invention; FIG. 10 is a diagram illustrating possible transitions from the current state searched using each matching candidate in the search for pattern matching processing by the matching unit according to the first embodiment of the present invention; FIG. 9 is a diagram showing a search path corresponding to FIG. 8 superimposed on the prediction result shown in FIG. 4; FIG. 5 is a diagram showing an output result example by the matching device according to the first embodiment of the present invention, together with a comparative example; FIG. 7 is a block diagram showing the configuration of a matching device according to Embodiment 2 of the present invention; FIG. 10 is a diagram showing an example of matching result output by the matching device according to the second embodiment of the present invention; FIG. 10 is a diagram showing another example of matching result output by the matching device according to the second embodiment of the present invention; It is a block diagram which shows the hardware structural example of the matching apparatus which concerns on Embodiment 1 of this invention.

[Embodiment 1]
An embodiment of the present invention will be described in detail below. The matching device according to the present embodiment performs matching processing on target information such as image information, and outputs matching results. As an example, the matching device according to this embodiment can be used to identify a character string included in image data captured by an imaging device, but this is not a limitation of this embodiment.

(matching device)
The configuration of the matching device will be specifically described with reference to FIG. FIG. 1 is a block diagram showing the configuration of a matching device 1 according to this embodiment. As shown in FIG. 1, the matching device 1 includes a control unit 10, a storage unit 20, a communication unit 30, and an input/output unit 40.

(storage unit)
First, various data stored in the storage unit 20 will be described with reference to FIG. As shown in FIG. 1, the storage unit 20 contains
・Regular Expression REGEX, and
・State Model SM corresponding to the regular expression
is stored. Here, the regular expression REGEX may be obtained from outside the matching device 1 via the communication unit 30 and the input/output unit 40, which will be described later, or may be obtained via the communication unit 30 and the input/output unit 40. It may be generated by the generating unit 14, which will be described later, based on the acquired user's instruction.

A regular expression REGEX includes at least one or more literal characters and/or one or more meta-characters. Here, literal characters refer to explicit numbers, characters, strings, etc. that need not be represented by identifiers. On the other hand, meta-characters are represented by identifiers and have various meanings. As an example, meta-characters include meta-characters that represent location (position), meta-characters that represent characters and groups of characters, meta-characters that represent escapes, meta-characters that represent repetition, and meta-characters that represent grouping. Although there are metacharacters and the like to express, this is not intended to limit the present embodiment.

The upper part of FIG. 2 shows an example of the regular expression REGEX stored in the storage unit 20. As shown in FIG. The regular expression REGEX shown in Figure 2 is
The literal letters 'a' and 'b'
・Meta character "()" that means grouping
・The meta character ``|'', which means an alternative choice
・The meta character “{}” that means the specified number of repetitions and the literal character “2” that indicates the specified number of times
・A meta character “*” that means any repetition of 0 or more times
and as a whole,
- It expresses 0 or more repetitions of (one a or two b).

On the other hand, the state model SM is a model including a plurality of states and transitions between states, and is also called a state machine. The state model SM includes, for example, nodes representing states and directed links representing transitions between states. The state model SM is referred to in matching processing by the matching unit 12, which will be described later.

Any regular expression can be represented as a state model. In other words, whether or not a certain target character string matches a certain regular expression (whether or not the certain target character string can be expressed as the certain regular expression) is determined by the corresponding regular expression. It can be determined using a state model that

An example of a state model is a Finite Automaton, but this is not a limitation of this embodiment. Also, as an example of a finite automaton,
・Deterministic Finite Automaton
・Non-deterministic Finite Automaton
However, this does not limit the present embodiment. A finite automaton can be represented, by way of example, by a computer program containing one or more states and rules describing one or more transitions.

The state model SM may be obtained from the outside of the matching device 1 via the communication unit 30 and the input/output unit 40, which will be described later, or may be obtained from the user may be generated by the generation unit 14, which will be described later, based on the instruction of .

As an example, the state model SM according to this embodiment is generated in association with the regular expression described above. For example, prior to matching processing by the matching unit 12, which will be described later, the generation unit 14 can generate in advance the state model SM corresponding to the regular expression REGEX used in the matching processing, and store it in the storage unit 20. .

The lower part of FIG. 2 shows an example of the state model SM stored in the storage unit 20. As shown in FIG. The state model SM shown in the lower part of FIG. 2 is a state model expressing an example of the regular expression REGEX shown in the upper part of FIG.

The state model SM illustrated in FIG.
・Nodes “1” and “2”
A link from node '1' to node '2', a link from node '2' to node '1', and a link from node '1' to its own node '1'. For the state model SM, the target character string is input to node "1", and if the first character string is "a", it stays at node "1", and the first character string is "b". If so, transition to node "2". If the first character in the string is anything other than "a" or "b", the string is rejected. Also, in the state of node “2”, if the second character string is “b”, the state is changed to node “1”.

In this way, when the target character string satisfies the regular expression REGEX in the upper part of FIG. Become. This may also be expressed as that the subject character string has been accepted by the state model SM.

On the other hand, when the target character string does not satisfy the regular expression REGEX in the upper part of FIG. 2 (the target character string cannot be expressed by the regular expression REGEX), the character string is in a state other than node "1". This is sometimes expressed as that the target character string is rejected by the state model SM.

Returning to the description of other data stored in the storage unit 20 . As shown in FIG. 1, the storage unit 20 stores image data IMG. The image data IMG is, for example, image data including one or more characters in the angle of view. The image data IMG may be acquired from the outside of the matching device 1 via the communication unit 30 or the input/output unit 40, or may be captured by a camera (imaging device) (not shown) included in the matching device 1. can be anything.

In addition, as shown in FIG. 1, the storage unit 20 also stores an inference model PM. More specifically, the storage unit 20 stores various data defining the inference model PM, and the later-described matching unit 12 refers to these data to execute the inference model PM. Although the specific configuration of the inference model PM does not limit this embodiment, as an example, the inference model PM can be configured using a CNN (Convolutional Neural Network), which limits this embodiment. not something to do. The inference model PM may be configured by combining a CNN and an RNN (Recurrent Neural Network), or may use another DNN (Deep Neural Network).

Further, as shown in FIG. 1, the storage unit 20 also stores a pattern matching result PMR indicating the result of matching processing by the matching unit 12 . When the image data IMG is image data including one or more characters in the angle of view, the pattern matching result PMR indicating the result of pattern matching by the matching unit 12 for the image data IMG is, for example, a string. . The pattern matching result PMR is output outside the matching device 1 via the communication unit 30 and the input/output unit 40 .

(communication department)
The communication unit 30 communicates with devices external to the matching device 1 . The communication unit 30 transmits data supplied from the control unit 10 to an external device, and supplies data received from an external device to the control unit 10 .

(input/output unit)
The input/output unit 40 receives data from outside the matching device 1 and outputs data to the outside of the matching device 1 . As an example, the input/output unit 40 includes a camera or scanner, and supplies an image acquired by the camera or scanner to the control unit 10 . Here, the image may include the image data IMG described above. As another example, the input/output unit 40 includes input devices such as a keyboard and a touch pad, and receives data input via these input devices. Then, it supplies the received data to the control unit 10 . Further, as an example, the input/output unit 40 includes a display panel, and displays data supplied from the control unit 10 via the display panel. The data displayed by the display panel may include the matching result by the matching unit 12 .

(control part)
Next, the configuration of the control unit 10 included in the matching device 1 will be described with reference to FIG. As shown in FIG. 1, the control unit 10 includes an acquisition unit 11, a matching unit 12, a selection unit 13, and a generation unit .

(acquisition part)
The acquisition unit 11 acquires target information to be matched. Here, the target information is, for example, image information including one or more characters or patterns in the angle of view. As an example, the target information may include one image including one or more characters or patterns in the angle of view, or may include multiple images including one or more characters or patterns in the angle of view. The image data IMG described above is an example of target information. However, these examples do not limit this embodiment.

(Matching part)
The matching unit 12 executes pattern matching processing on target information. More specifically, the matching unit 12 uses a state model corresponding to a certain regular expression to search each of a plurality of pieces of partial information included in the image data IMG, comprising: Pattern matching is performed on the target information by executing a search limited to a predetermined number or less. The matching section 12 is sometimes called a first matching section. In addition, in this embodiment, the search route is also simply referred to as a route.

Here, the matching unit 12 can use, as an example, the state model SM corresponding to the regular expression REGEX described above.

Also, the plurality of pieces of partial information included in the image data IMG are, for example, partial images obtained by dividing the image data IMG into a plurality of pieces. For example, the matching unit 12 may divide the image data IMG into partial images each having a predetermined pixel width, and predict a matching candidate for each partial image. Here, the predetermined pixel width is 4 pixels as an example, but this does not limit the present embodiment. The predetermined pixel width may be, for example, 6 pixels, 8 pixels, or another pixel width. Also, the prediction of the matching candidate for each partial image by the matching unit 12 can be executed by the inference model PM described above, as an example.

More specifically, in the matching unit 12, the inference model PM predicts a character included in the image data IMG for each partial image of a predetermined pixel width included in the image data IMG, and outputs the prediction result. The plurality of matching candidates represented is output along with probability information for each matching candidate.

In addition, as described above, the matching unit 12 executes a search targeting each of the plurality of partial images included in the image data IMG, in which the number of search paths is limited to a predetermined number or less. do. Here, each time the inference model PM outputs a matching candidate for a partial image,
(1) selecting K matching candidates in descending order of probability indicated by the probability information; (2) generating one or more search paths by adding the selected plurality of matching candidates to the previous search paths;
(3) Among one or more generated search paths, a process of leaving the top N-th search paths and excluding other paths is performed.

Here, in addition to the processes (1) to (3) above, the matching unit 12 applies a selection process with reference to the state model SM to eliminate matching candidates that do not match the state model SM. . As an example, the matching unit 12 applies selection processing with reference to the state model SM to a plurality of matching candidates output by the inference model PM before the processing of (1) above. Perform a filtering process that eliminates matching candidates that do not match. Then, the matching unit 12 performs the processes (2) and (3) above on the matching candidates after the filtering process.

In this way, the matching unit 12 applies the selection process with reference to the state model SM to the plurality of matching candidates described above, thereby eliminating matching candidates that do not match the state model SM and increasing the number of transition paths. is limited to a predetermined beam width or less, the process proceeds to matching processing for the next partial image. The matching unit 12 executes beam search using the state model by performing such processing. The matching unit 12 performs pattern matching on the image data IMG by executing beam search using such a state model.

The matching unit 12 outputs the result of pattern matching as a pattern matching result PMR, and the output pattern matching result PMR is stored in the storage unit 20 as an example.

(Selection part)
The selection unit 13 (state model selection unit 13 ) selects a state model SM to be used for pattern matching by the matching unit 12 from the state models SM stored in the storage unit 20 . As an example, the selection unit 13 selects a state model to be used for pattern matching from a plurality of state models individually corresponding to individual regular expressions based on the content type information received by the input/output unit 40 .

If the matching device 1 uses only one specific state model SM, the selector 13 is not essential. As an example, the multiple state models SM include
・State model that supports regular expressions suitable for pattern matching on dates (including year, month, day, etc.) ・Supports regular expressions suitable for pattern matching on addresses (including prefectures, cities, towns, addresses, etc.)・State model that supports regular expressions suitable for pattern matching on phone numbers ・State model that supports regular expressions suitable for pattern matching on e-mail addresses ・Financial values (price, various indices, etc.) It can be prepared in advance according to the type of matching target, such as a state model corresponding to a regular expression suitable for pattern matching.

(generator)
The generation unit 14 generates the state model SM corresponding to the regular expression REGEX stored in the storage unit 20 from the regular expression REGEX. The generator 14 can generate the state model SM as a finite automaton in the form of a computer program, for example. However, this does not limit the present embodiment, and the state model SM may be a set of rules and parameters and may be generated in a form that can be referred to by a computer.

For example, the generating unit 14 generates a non-deterministic finite automaton corresponding to the regular expression REGEX stored in the storage unit 20, and generates a deterministic finite automaton based on the non-deterministic finite automaton. may be generated.

A deterministic finite automaton is preferably used rather than a non-deterministic finite automaton as the state model SM used for pattern matching performed by the matching unit 12 described above. This is because deterministic finite automata can process acceptance or rejection faster than non-deterministic finite automata.

Note that the state model SM may be configured to be created by a human and stored in the storage unit 20 . In such a configuration, the generator 14 is not essential.

(Flow of pattern matching processing)
Next, the flow of pattern matching processing by the matching device 1 will be described with reference to FIG. FIG. 3 is a flow chart showing the flow of pattern matching processing by the matching device 1. As shown in FIG.

(Step S101)
First, in step S101, the acquisition unit 11 acquires target information. As described above, for example, the acquiring unit 11 acquires the image data IMG as the target information.

(Step S102)
Step S102 is the beginning of a processing loop for a plurality of pieces of partial information included in target information. As described above, as an example, the partial information is a partial image with a predetermined pixel width included in the image data IMG. A loop related to partial information is expressed by a loop variable n (n is a natural number from 0 to N-1 (N is the total number of partial images included in the image data IMG)), as an example.

(Step S103)
Subsequently, in step S103, the matching unit 12 predicts a character included in the image data IMG by the inference model PM referring to the partial information In, and selects a plurality of matching candidates representing the prediction result as each matching candidate. Acquired with the probability information of

(Step S104)
Subsequently, in step S104, the matching unit 12 obtains possible transitions from the current state using each matching candidate obtained in step S103. Here, the current state refers to the state determined by the processing up to the loop variable n-1, and as an example, refers to the current state in the state model SM. , refers to possible transitions from the current state in the state model SM. Alternatively, one or a plurality of paths remaining as selectable states in a transition diagram representing transitions corresponding to each process up to the loop variable n-1 can also be referred to as the current state.

(Step S105)
Subsequently, in step S105, the matching unit 12 checks whether each of the possible transitions obtained in step S104 matches the state model SM. In other words, each matching candidate is checked for consistency with the state model SM. where if said possible transition (or its corresponding candidate match) is consistent with the state model SM, then said transition (or its corresponding candidate match) is accepted by the state model SM; rejected by

(Step S106)
Subsequently, in step S106, the matching unit 12 eliminates transitions (or matching candidates corresponding thereto) that do not match the state model SM, in other words, transitions that are rejected by the state model SM (or matching candidates corresponding thereto). , is not considered in subsequent processing.

(Step S107)
Subsequently, in step S107, the matching unit 12 selects K matching candidates from the matching candidates that were not eliminated in step S106. As an example, matching candidates up to the Kth with the highest probability are selected.

(Step S108)
Subsequently, in step S108, the matching unit 12 updates the route using the matching candidate selected in step S107.

(Step S109)
Subsequently, in step S109, the matching unit 12 leaves the routes up to the N-th route selected in step S108, and excludes the other routes. Excluded paths are not considered in further processing.

(Step S110)
Subsequently, in step S110, the matching unit 12 increments the loop variable n to n+1.

(Step S111)
Step S111 is the end of the processing loop for a plurality of pieces of partial information included in the target information.

(Step S112)
After completing the above-described processing for all of the plurality of pieces of partial information included in the target information, in step S112, the matching unit 12 outputs the matching result corresponding to the route with the highest probability.

As described above, the pattern matching processing by the matching device 1 is performed as matching processing for certain partial information included in the target information.
a process of acquiring one or more matching candidates for the certain partial information and probability information of each matching candidate (step S103);
a first selection process (step S106) for excluding matching candidates that do not match the state model from among the one or more matching candidates;
and a second selection process (step S107) for selecting matching candidates equal to or less than the predetermined number from the matching candidates after the first selection process by referring to the probability information.

(Example of state transition by processing of matching unit 12)
An example of state transition by the processing of the matching unit 12 will be described below. In the following description, the case where the character string included in the image data captured by the imaging device is "1st year February 8" will be described as an example. In other words, in step S101 described above, a case where the acquisition unit 11 acquires the image data IMG in which the character string "February 8, 1st year" is imaged as the target information will be described as an example.

In the above example, the multiple pieces of partial information In (n=0 to 5) included in the target information are "1", "year", "2", "month", "8", and "day". , six partial images corresponding to .

As described above, in step S103, the matching unit 12 predicts characters included in the image data IMG by the inference model PM that refers to the partial information In, and selects a plurality of matching candidates representing the prediction results from each Acquired together with probability information of matching candidates. As a result, for example, the information shown in FIG. 4 is acquired.

FIG. 4 is a diagram showing an example of a plurality of matching candidates and probability information acquired by the matching unit 12. As shown in FIG.

In the example of FIG. 4, the partial information I0 to I5 are represented as Slice0 to Slice5. Also, Characters “1”, “2”, “3”, . . . “Month”, “Day”, .

In FIG. 4, the columns corresponding to each Slice and the rows corresponding to each Characters indicate the probability information of each matching candidate as the prediction result for each piece of partial information. Here, for each piece of partial information, the top three probability information items of each matching candidate are selected in descending order of value. In FIG. 4, the top three pieces of probability information are shown with different hatching.

Next, examples of search paths and matching results obtained by the processing of the matching unit 12 will be described more specifically.

(Processing according to comparative example)
First, for comparison, the transition by the conventional beam search will be described as a comparative example.

FIG. 5 is a diagram illustrating possible transitions from the current state searched using each matching candidate in the conventional beam search. In this example, the number N of beams to maintain is two and the number K of matching characters considered for transition destinations is three.

That is, in the beam search for each of the partial information I0 to the partial information I5, matching candidates up to the third with the highest probability are selected, and the beam search path is updated using the selected matching candidates. At this time, among the paths of the beam search, the paths up to the second highest are left, the other paths are excluded, and the excluded paths are not considered in the subsequent processing.

In this case, the state transitions with the beam search for each of the partial information I0 to I5. Therefore, in FIG. 5, Step 0 indicates the state corresponding to the partial information I0, Step 1 indicates the state corresponding to the partial information I1, Step 2 indicates the state corresponding to the partial information I2, and so on. A state corresponding to the partial information I5 is shown in Step5.

Also, in FIG. 5, the Considered characters are matching candidates up to K-th with high probability. ing.

Furthermore, in FIG. 5, the paths up to the top two are represented by Beam 1 and Beam 2. In FIG. Matching candidates corresponding to the current state of path Beam 1 are indicated by Beam 1 characters, and matching candidates corresponding to the current state of path Beam 2 are indicated by Beam 2 characters.

In FIG. 5, the matching candidate probability information indicated by Beam 1 characters is indicated by Beam 1 probability, and the matching candidate probability information indicated by Beam 2 characters is indicated by Beam 1 probability.

Further details will be described with reference to FIG. 6 in conjunction with FIG. FIG. 6 is a diagram showing search paths corresponding to FIG. 5 superimposed on the prediction result shown in FIG. In FIG. 6, the path Beam 1 and path Beam 2 up to the second highest are indicated by straight arrows and dashed-dotted line arrows, respectively.

Here, path Beam 1 and path Beam 2 are also referred to as Best beam and 2nd best beam, respectively. Also, in FIG. 6, excluded pathways are called Considered expansion and indicated by dashed arrows.

In FIG. 6, in Step 0 (Slice 0), the probability information is referenced, "/" is the best beam, and "1" (number 1) is the 2nd best beam. On the other hand, "l" (alphabet l) has the lowest probability information value among the three matching candidates, so it is considered expansion and is eliminated.

In Step 1 (Slice 1), beam search is executed with each of "/" and "1" as the current state. Considered characters in Step1 (Slice1) are "year", "month", and "day", so the path from "/" to "year", the path from "/" to "month", the path from "/" to " There can be a path to day, a path from '1' to 'year', a path from '1' to 'month', and a path from '1' to 'day'.

The probability corresponding to the route from "/" to "year" is 0.32 (=0.4×0.8), the probability corresponding to the route from "/" to "month" is 0.04 (=0.4×0.1), "/ ” to “日” is 0.02 (=0.4×0.05). The probability corresponding to the route from "1" to "year" is 0.24 (=0.3×0.8), and the probability corresponding to the route from "1" to "month" is 0.03 (=0.3×0.1 ), and the probability corresponding to the path from “1” to “moon” is 0.015 (=0.3×0.05).

Among the six paths described above, the top two paths are the path from "/" to "year" and the path from "1" to "month". Therefore, the best beam in Step1 (Slice1) is the path from "/" to "year", the 2nd best beam is the path from "1" to "year", and other paths are excluded. .

Similarly, in Step 2 (Slice 2), the path of '/', 'year', and '2' is the best beam, and the path of '1', 'year', and '2' is the 2nd best beam.

By repeating such processing up to Step 5 (Slice 5), the best beam and the 2nd best beam as the prediction results of the target information "February 8, 1st year" are obtained. In this example, the best beam is the path of '/', 'year', '2', 'month', '8', '8', and the 2nd best beam is the path of '/', 'year', ' 2", "month", "8", "day". Here, the probability of the best beam (Beam 1 probability) is 0.09216, and the probability of the 2nd best beam (Beam 2 probability) is 0.073728.

(Example of processing according to the present embodiment)
Next, a more specific description will be given of the state transition by the pattern matching processing by the matching device 1 of the present embodiment.

FIG. 7 is a diagram showing an example of the state model SM used for pattern matching performed by the matching unit 12. As shown in FIG. FIG. 7 is a state model corresponding to regular expressions suitable for pattern matching on dates (including year, month, day, etc.). The state model shown in FIG. 7 is selected by the selection unit 13 from among a plurality of state models, as an example.

Referring to the state model of FIG. 7, the matching unit 12
In Step 0 (transition from node 0 to node 1), each matching candidate is matched with the regular expression "[0-9]", and matching candidates matching the regular expression "[0-9]" are left. Exclude matching candidates that do not match the regular expression "[0-9],
In Step 1 (transition from node 1 to node 2), the consistency between each matching candidate and the regular expression "year" is determined, and matching candidates that match the regular expression "year" are left, and the regular expression "year" Filter out candidate matches that do not match the
・・・
In Step 5 (transition from node 5 to node 6), the consistency between each matching candidate and the regular expression "day" is determined, and matching candidates that match the regular expression "day" are left, and the regular expression "day" filter out candidate matches that do not match the
process.

FIG. 8 is a diagram for explaining possible transitions from the current state searched using each matching candidate in the search for pattern matching processing by the matching device 1 of this embodiment. In this example, the number N of beams to maintain is two and the number K of matching characters considered for transition destinations is three.

That is, as in the case of FIG. 5, in the search for each of the partial information I0 to I5, matching candidates up to the third with the highest probability are selected, and the search path is updated using the selected matching candidates. be. At this time, among the search routes, the routes up to the second highest are retained, the other routes are excluded, and the excluded routes are not considered in subsequent processing.

In FIG. 8, selection of matching candidates is different from that in FIG. 5 because the state model of FIG. 7 is applied.

That is, in Step 0 (Slice 0), the regular expression "[0-9]" is applied, so there are three matching candidates for Considered characters: "1", "2", and "3". ” is excluded from matching candidates.

Further details will be described with reference to FIG. 9 in conjunction with FIG. FIG. 9 is a diagram showing search paths corresponding to FIG. 8 superimposed on the prediction result shown in FIG. In FIG. 6, the path Beam 1 and path Beam 2 up to the second highest are indicated by straight arrows and dashed-dotted line arrows, respectively.

Here, path Beam 1 and path Beam 2 are also referred to as Best beam and 2nd best beam, respectively. Also, in FIG. 9, the excluded pathways are called Considered expansion and are indicated by dashed arrows.

In the examples of FIGS. 8 and 9, in Step 1 (Slice 1), a search is executed with each of "1" and "2" as the current state. By applying the state model in Fig. 7, the Considered characters in Step1 (Slice1) are only "Year", so the path from "1" to "Year" and the path from "2" to "Year" there can be a route. Therefore, the best beam in Step 1 (Slice 1) is the path from "1" to "year", and the 2nd best beam is the path from "2" to "year".

Similarly, in Step 2 (Slice 2), the path of "1", "year" and "2" is the best beam, and the path of "2", "year" and "2" is the 2nd best beam.

By repeating such processing up to Step 5 (Slice 5), the best beam and the 2nd best beam as the prediction results of the target information "February 8, 1st year" are obtained. In this example, the Best beam will be the path of '1', 'Year', '2', 'Month', '8', 'Day', and the 2nd best beam will be '2', 'Year', ' 2", "month", "8", "day". Here, the probability of the best beam (Beam 1 probability) is 0.055296, and the probability of the 2nd best beam (Beam 2 probability) is 0.

According to the beam search according to the comparative example described with reference to FIGS. day" are obtained, both of which are incorrect prediction results.

On the other hand, according to the search of this embodiment described with reference to FIGS. However, since the probability of "February 8, 2nd year" is 0, only "February 8th, 1st year" is actually obtained.

As described above, according to the present embodiment, since the state model corresponding to the regular expression is used in the pattern matching performed by the matching unit 12, more accurate prediction results can be obtained.

(Example of output result by matching device 1)
FIG. 10 is a diagram showing an output result example by the matching device 1 according to the present embodiment together with a comparative example. The example shown in FIG. 10 shows the matching result when the matching device 1 acquires the image data 101 as the target information.

In the example shown in FIG. 10, the input/output unit 40 accepts "date" as the content type information, and the selection unit 13 selects multiple regular expressions individually corresponding to individual regular expressions based on the content type information. state model corresponding to a regular expression suitable for pattern matching on dates (including year, month, day, etc.).

In the example shown in FIG. 10, the matching unit 12 uses the state model to search each of a plurality of pieces of partial information included in the target information as described above. Pattern matching is performed on the target information by performing searches limited to no more than the number of . Then, the input/output unit 40 displays the matching result 103 by the matching unit 12, as shown in FIG.

On the other hand, matching results 102A and 102B indicate matching results according to comparative examples. A matching result 102A shows an example of the result of executing a beam search without using the state model.

More specifically, the matching result 102A has the following character types among the outputs of the inference model PM:
A process that leaves only numbers, kanji, and punctuation, in other words, sets the probability information for outputs other than numbers, kanji, and punctuation to 0. After executing the processing, the result of executing the beam search without using the state model is shown.

Matching result 102B shows another example of the result of performing beam search without using the state model. More specifically, the matching result 102B has the following character types among the outputs of the inference model PM:
After executing the process of leaving only numbers and kanji, in other words, the process of setting the probability information of outputs other than numbers and kanji to 0, the state model is used. It shows the result of performing a beam search that does not have

As shown in FIG. 10, a search that does not use a state model outputs "/" instead of "1" as shown in matching result 102A, or "8" instead of "day" as shown in matching result 102B. " is output, and it is not possible to output an appropriate matching result.

On the other hand, as shown in FIG. 10, the matching result 103, which is the search result using the state model, is an appropriate matching result.

As described above, according to the matching device 1 according to the first embodiment, it is possible to improve accuracy in the sequential pattern matching process.

[Embodiment 2]
A second embodiment of the present invention will be described below. The same reference numerals are given to the same configurations as those described in the first embodiment, and the description thereof will be omitted as appropriate.

A matching device 1A according to the present embodiment, like the matching device 1 according to the first embodiment, performs matching processing on target information such as image information, and outputs a matching result.

FIG. 11 is a block diagram showing the configuration of a matching device 1A according to this embodiment. As shown in FIG. 11, a control unit 10, a storage unit 20, a communication unit 30, and an input/output unit 40 are provided.

Various data stored in the storage unit 20 are substantially the same as those in the storage unit 20 included in the matching device 1 according to the first embodiment. However, as an example, only one state model SM is stored in the storage unit 20 according to the present embodiment. The state model SM according to this embodiment uses, as an example, a more general expression than the first embodiment as the regular expression REGEX.

Further, as shown in FIG. 11, the control unit 10 of the matching device 1A according to the present embodiment includes an acquisition unit 11, a first matching unit 12, a second matching unit 15, a matching result selection unit 16, and a matching result output unit. 17.

The first matching unit 12 has the same configuration as the matching unit 12 according to the first embodiment. Therefore, various processes of the matching unit 12 described in the first embodiment are also applied to this embodiment. On the other hand, the second matching unit 15 performs beam search without using the state model. In other words, the second matching unit 15 performs beam search according to the comparative example described in the first embodiment. More specifically, the second matching unit 15 performs a beam search for each of a plurality of pieces of partial information included in the target information without using a state model corresponding to a regular expression. is limited to a predetermined number, pattern matching is performed on the target information.

The matching result selection unit 16 selects either the matching result from the first matching unit 12 or the matching result from the second matching unit 15 as probability information of the matching result from the first matching unit 12 and the matching result from the second matching unit 15. Select according to probability information. Then, the matching result output unit 17 outputs the selection result by the matching result selection unit 16 . The matching result output by the matching result output unit 17 is displayed on a display panel included in the input/output unit 40 as an example.

In other words, the matching method by the matching device 1A is
an acquisition step of acquiring target information to be matched;
By executing a search targeting each of a plurality of partial information contained in the target information using a state model corresponding to a certain regular expression, the search limiting the number of search paths to a predetermined number or less , a first matching step of performing pattern matching on the target information;
A beam search targeting each of a plurality of partial information included in the target information without using a state model corresponding to a regular expression, the beam search limiting the number of search paths to a predetermined number. A second matching step of performing pattern matching on the target information by
Either the matching result of the first matching step or the matching result of the second matching step is combined with probability information of the matching result of the first matching step and probability information of the matching result of the second matching step. and a selection step of selecting according to .

FIG. 12 is a diagram showing an example of matching result output by the matching device 1A according to the present embodiment. The example shown in FIG. 12 shows the matching result when the matching device 1A acquires the image data 101 as the target information.

In the example shown in FIG. 12, the first matching unit 12 uses the general state model described above to search each of a plurality of pieces of partial information included in the target information. Pattern matching is performed on the target information by performing a search limited to the number of , and a matching result 103 is obtained.

On the other hand, the second matching unit 15 performs a beam search targeting each of a plurality of pieces of partial information included in the target information without using a state model corresponding to a regular expression. Pattern matching is performed on the target information by performing a beam search restricted to , and a matching result 102 is obtained.

In the example shown in FIG. 12, the probability (beam probability) of the matching result 103 obtained by the first matching unit 12 is 0.98, which can be said to be a sufficiently large value. In other words, assuming that the preset threshold value is 0.8 as an example, the probability of a value greater than the threshold value is obtained.

On the other hand, in the example shown in FIG. 12, the probability (Beam probability) of the matching result 102 obtained by the second matching unit 15 is 0.99, which can be said to be a sufficiently large value. In other words, assuming that the preset threshold value is 0.8 as an example, the probability of a value greater than the threshold value is obtained.

Thus, when both the matching result probability information obtained by the first matching unit 12 and the matching result probability information obtained by the second matching unit 15 indicate high probability, in other words, When the probability information of the matching result is both larger than the predetermined threshold, the matching result output unit 16 outputs the matching result obtained by the first matching unit 12 and the matching result obtained by the second matching unit 15, 12 outputs only the matching results obtained. Alternatively, the matching result obtained by the first matching unit 12 is preferentially displayed.

On the other hand, FIG. 13 is a diagram showing another example of matching result output by the matching device 1A according to the present embodiment. The example shown in FIG. 13 shows the matching result when the matching device 1A acquires the image data 101 as the target information.

In the example shown in FIG. 13, the probability (beam probability) of the matching result 103 obtained by the first matching unit 12 is 0.02, which can be said to be a sufficiently small value. In other words, assuming that the preset threshold value is 0.8 as an example, the probability of a value smaller than the threshold value is obtained.

On the other hand, in the example shown in FIG. 13, the probability (beam probability) of the matching result 102 obtained by the second matching unit 15 is 0.96, which can be said to be a sufficiently large value. In other words, assuming that the preset threshold value is 0.8 as an example, the probability of a value greater than the threshold value is obtained.

Thus, when the probability information of the matching result obtained by the first matching unit 12 is smaller than the probability information of the matching result obtained by the second matching unit 15, in other words, the matching obtained by the first matching unit 12 If the probability information of the result is smaller than the predetermined threshold and the probability information of the matching result obtained by the second matching unit 15 is larger than the predetermined threshold, the matching result output unit 16 outputs the matching obtained by the first matching unit 12. Among the results and the matching results obtained by the second matching unit 15, only the matching results obtained by the second matching unit 15 are output. Alternatively, the matching result obtained by the second matching unit 15 is preferentially displayed.

According to the matching device 1A configured as described above, more appropriate matching results can be presented to the user. Further, according to the matching device 1A that performs the above process, it is possible to improve accuracy without a priori information regarding what type of information the target information (image data) is.
(Specific processing example)
A specific processing example by the matching result selection unit 16 may be expressed as follows. That is, the matching result selection unit 16
A value indicating a ratio obtained by dividing the probability value indicated by the probability information of the matching result by the first matching unit 12 and the probability value indicated by the probability information of the matching result by the second matching unit 15 is compared with the threshold. Then, the matching result selection unit 16
If the value indicating the ratio is equal to or greater than the threshold, selecting the matching result by the first matching unit 12;
If the value indicating the ratio is less than the threshold, the matching result by the second matching unit 15 is selected. Here, the threshold can be set as appropriate. As an example, the threshold may be any value between 0.9 and 1.0, or may be another value.

As an example, the example of FIG. 12 will be described from the viewpoint of this processing example. 2 by dividing by 0.99, which is the beam probability of the matching result 102 obtained by the matching unit 15, the ratio
0.989
Calculate Here, the threshold value is set to 0.9.

The matching result selection unit 16 compares the ratio value of 0.989 with the threshold value of 0.9. Since the value of the ratio is greater than or equal to the threshold, the matching result selection unit 16 selects the matching result 103 by the first matching unit 12 .

On the other hand, in the example of FIG. 13 , the matching result selection unit 16 sets the beam probability of 0.02 for the matching result 103 obtained by the first matching unit 12 to 0.02 for the matching result 102 obtained by the second matching unit 15 . By dividing by 0.96, the Beam probability of 0.021, the ratio value
Calculate Here, the threshold is set to 0.9 as in the case of the example of FIG.

The matching result selection unit 16 compares the ratio value of 0.021 with the threshold value of 0.9. Since the value of the ratio is less than the threshold, the matching result selection unit 16 selects the matching result 102 from the second matching unit 15 .

Note that the matching result selection unit 16 may set the threshold to a value corresponding to at least one of the matching result by the first matching unit 12 and the matching result by the second matching unit 12 . As an example, the threshold is set to a value corresponding to at least one of the probability value indicated by the probability information of the matching result by the first matching unit 12 and the probability value indicated by the probability information of the matching result by the second matching unit 15. may

(Hardware configuration example of matching device)
FIG. 14 is a block diagram illustrating the physical configuration of a computer used as matching devices 1 and 1A. The information processing apparatus 1 can be configured by a computer having a bus 210, a processor 201, a main memory 202, an auxiliary memory 203, a communication interface 204, and an input/output interface 205, as shown in FIG. . Processor 201 , main memory 202 , auxiliary memory 203 , communication interface 204 and input/output interface 205 are interconnected via bus 210 . An input device 206 and an output device 207 are connected to the input/output interface 205 .

As the processor 201, for example, a CPU (Central Processing Unit), a microprocessor, a digital signal processor, a microcontroller, or a combination thereof is used.

As the main memory 202, for example, a semiconductor RAM (random access memory) or the like is used.

As the auxiliary memory 203, for example, flash memory, HDD (Hard Disk Drive), SSD (Solid State Drive), or a combination of these is used. Auxiliary memory 203 stores a program for causing processor 201 to execute the operations of matching devices 1 and 1A described above. The processor 201 expands the program stored in the auxiliary memory 203 onto the main memory 202 and executes each instruction included in the expanded program.

A communication interface 204 is an interface that connects to the network N. FIG.

As the input/output interface 205, for example, a USB interface, a short-range communication interface such as infrared rays or Bluetooth (registered trademark), or a combination thereof is used.

As the input device 206, for example, a keyboard, mouse, touch pad, microphone, or a combination thereof is used. A display, a printer, a speaker, or a combination thereof is used as the output device 207, for example.

In this example, processor 201 is an example of a hardware element that implements control unit 10 . Also, the main memory 202 and the auxiliary memory 03 are examples of hardware elements that implement the storage unit 20 . Also, the communication interface 204 is an example of a hardware element that implements the communication unit 30 . Also, the input/output interface 205, the input device 206, and the output device 207 are examples of hardware elements that implement the input/output unit 40 described above.

[Example of realization by software]
The functions of the information processing apparatuses 1 and 1A are programs for causing a computer to function as the apparatus, and are realized by a program for causing the computer to function as each control block (especially each part included in the control unit 10) of the apparatus. can do.

In this case, the apparatus comprises a computer having at least one control device (eg processor) and at least one storage device (eg memory) as hardware for executing the program. Each function described in each of the above embodiments is realized by executing the above program using the control device and the storage device.

The program may be recorded on one or more computer-readable recording media, not temporary. The recording medium may or may not be included in the device. In the latter case, the program may be supplied to the device via any transmission medium, wired or wireless.

Also, part or all of the functions of the above control blocks can be realized by logic circuits. For example, integrated circuits in which logic circuits functioning as the control blocks described above are formed are also included in the scope of the present invention.

Some or all of the functions of each control block may be operated by the control device, or may be operated by another device (for example, an edge computer or a cloud server). There may be.

〔summary〕
Some or all of the above-described embodiments may also be described as follows. However, the present invention is not limited to the embodiments described below.

(Appendix 1)
A search targeting each of a plurality of partial information included in the target information using an acquisition unit that acquires target information to be matched and a state model corresponding to a certain regular expression, wherein: a first matching unit that performs pattern matching on the target information by executing a search that limits the number of the target information to a predetermined number or less.

(Appendix 2)
The search executed by the first matching unit acquires one or more matching candidates for the certain partial information and probability information of each matching candidate as matching processing for certain partial information included in the target information. a first selection process for excluding matching candidates that do not match the state model from among the one or more matching candidates; referring to the probability information, from the matching candidates after the first selection process; , and a second selection process of selecting matching candidates equal to or less than the predetermined number.

(Appendix 3)
3. A matching apparatus according to any one of claims 1 or 2, wherein the state model is a deterministic finite automaton.

(Appendix 4)
4. The method according to any one of appendices 1 to 3, further comprising a selection unit that selects a state model to be used for the pattern matching from a plurality of state models individually corresponding to each regular expression based on content type information. matching device.

(Appendix 5)
5. The matching device according to any one of appendices 1 to 4, further comprising a generator that generates the state model.

(Appendix 6)
6. A matching device according to any one of appendices 1 to 5, wherein the target information includes one or more images.

(Appendix 7)
7. The matching device according to any one of appendices 1 to 6, further comprising an output unit that outputs a matching result obtained by the first matching unit.

(Appendix 8)
A beam search targeting each of a plurality of partial information included in the target information without using a state model corresponding to a regular expression, the beam search limiting the number of search paths to a predetermined number. further comprising a second matching unit that performs pattern matching on the target information,
One of the matching result by the first matching unit and the matching result by the second matching unit is combined with probability information on the matching result by the first matching unit and probability information on the matching result by the second matching unit. 7. A matching device according to any one of claims 1 to 6, comprising an output unit for outputting in response to .

(Appendix 9)
An acquisition step of acquiring target information to be matched, and a search targeting each of a plurality of partial information included in the target information using a state model corresponding to a certain regular expression, wherein: and performing pattern matching on said target information by performing a search limited to a predetermined number or less.

(Appendix 10)
A program for causing a computer to function as the matching device according to appendix 1, the program for causing the computer to function as the acquisition unit and the matching unit.

(Appendix 11)
A computer-readable recording medium recording the program according to appendix 10.

The present invention is not limited to the above-described embodiments, but can be modified in various ways within the scope of the claims, and can be obtained by appropriately combining technical means disclosed in different embodiments. is also included in the technical scope of the present invention.

1, 1A matching device 10 control unit 11 acquisition unit 12 matching unit (first matching unit)
13 selection unit (state model selection unit)
14 generation unit 15 second matching unit 16 matching result selection unit 17 matching result output unit 20 storage unit 30 communication unit 40 input/output unit

(Matching part)
The matching unit 12 executes pattern matching processing on target information. More specifically, the matching unit 12 searches each of a plurality of pieces of partial information included in the image data IMG using a state model corresponding to a certain regular expression. Pattern matching is performed on the target information by performing searches limited to no more than a number. The matching section 12 is sometimes called a first matching section. In addition, in this embodiment, the search route is also simply referred to as a route.

Referring to the state model of FIG. 7, the matching unit 12
In Step 0 (transition from node 0 to node 1), the consistency between each matching candidate and the regular expression "[0-9]" is determined , and matching candidates that match the regular expression "[0-9]" are left. , exclude matching candidates that do not match the regular expression "[0-9],
In Step 1 (transition from node 1 to node 2), the consistency between each matching candidate and the regular expression "year" is determined, and matching candidates that match the regular expression "year" are left, and the regular expression "year" Filter out candidate matches that do not match the
・・・
In Step 5 (transition from node 5 to node 6), the consistency between each matching candidate and the regular expression "day" is determined, and matching candidates that match the regular expression "day" are left, and the regular expression "day" filter out candidate matches that do not match the
process.

Claims (12)

an acquisition unit that acquires target information to be matched;
By executing a search targeting each of a plurality of partial information contained in the target information using a state model corresponding to a certain regular expression, the search limiting the number of search paths to a predetermined number or less , a first matching unit that performs pattern matching on the target information;
A beam search targeting each of a plurality of partial information included in the target information without using a state model corresponding to a regular expression, the beam search limiting the number of search paths to a predetermined number. A second matching unit that performs pattern matching on the target information by
One of the matching result by the first matching unit and the matching result by the second matching unit is combined with probability information on the matching result by the first matching unit and probability information on the matching result by the second matching unit. and a matching result selector that selects according to. The search performed by the first matching unit includes:
As a matching process for certain partial information included in the target information,
a process of acquiring one or more matching candidates for the certain partial information and probability information of each matching candidate;
a first selection process for excluding matching candidates that do not match the state model from among the one or more matching candidates;
a second selection process of selecting matching candidates equal to or less than the predetermined number from the matching candidates after the first selection process, with reference to the probability information;
2. The matching device of claim 1, comprising: 3. A matching device according to claim 1 or 2, wherein said state model is a deterministic finite automaton. 4. The method according to any one of claims 1 to 3, further comprising a state model selection unit that selects a state model used for pattern matching from a plurality of state models individually corresponding to individual regular expressions based on content type information. A matching device according to paragraph. 5. The matching device according to any one of claims 1 to 4, further comprising a generator that generates the state model. 6. A matching device according to any one of claims 1 to 5, wherein said target information includes one or more images. 7. The matching device according to any one of claims 1 to 6, further comprising an output section for outputting the selection result of the matching result selection section. The matching result selection unit
A probability value indicated by the probability information of the matching result by the first matching unit and a value indicating a ratio obtained by dividing by the probability value indicated by the probability information of the matching result by the second matching unit are compared with a threshold value. death,
If the value indicating the ratio is equal to or greater than the threshold, selecting the matching result by the first matching unit;
8. The matching device according to any one of claims 1 to 7, wherein if the value indicating the ratio is less than the threshold value, the matching result by the second matching section is selected. 9. The matching device according to claim 8, wherein the matching result selection unit sets the threshold to a value corresponding to at least one of the matching result by the first matching unit and the matching result by the second matching unit. an acquisition step of acquiring target information to be matched;
By executing a search targeting each of a plurality of partial information contained in the target information using a state model corresponding to a certain regular expression, the search limiting the number of search paths to a predetermined number or less , a first matching step of performing pattern matching on the target information;
A beam search targeting each of a plurality of partial information included in the target information without using a state model corresponding to a regular expression, the beam search limiting the number of search paths to a predetermined number. A second matching step of performing pattern matching on the target information by
Either the matching result of the first matching step or the matching result of the second matching step is combined with probability information of the matching result of the first matching step and probability information of the matching result of the second matching step. a selection step of selecting according to the matching method. A program for causing a computer to function as the matching device according to claim 1, the program for causing the computer to function as the acquisition unit, the first matching unit, and the second matching unit. A computer-readable recording medium on which the program according to claim 11 is recorded.

Images