JP2024078016A - Work support system, work support device, and work support method - Google Patents

Abstract

To properly set an important work showing work details.SOLUTION: A work support system 30 has: a past processing part 10 which learns an evaluation model 17, which evaluates an important word included in manual data 36 from a sentence structure extracted from the manual data 36, so as to output a word included in a past manual 11 as an important word included in the past manual 11 to the evaluation model 17; and a current processing part 20 which inputs the sentence structure extracted from the current manual data 36 to the evaluation model 17 to register the important word output from the evaluation model 17 in an important word dictionary 34 as a keyword for confirming whether the important word is uttered in inspection work based upon the current manual data 36.SELECTED DRAWING: Figure 8

Description

The present invention relates to a work support system, a work support device, and a work support method.

Speech recognition systems are becoming more and more popular as a means of inputting character data. Since workers at work sites have their hands full with tools, it is convenient to use voice input for inputting character data.
For example, Patent Document 1 describes a report creation system that consolidates the results of maintenance and inspection for each work item of an object using information based on voice input by a worker to a mobile terminal.

JP 2015-75792 A

In some cases, work such as inspections at work sites is carried out by a work team consisting of one supervisor and one or more workers. The supervisor and workers in the work team communicate with each other by voice via a communication device such as a walkie-talkie. Below, an example of the outline of the procedure for an inspection work, which is one example of work, is shown.
(Step 1) The person in charge of the inspection will have a "manual" that describes the work procedures for each inspection item. The manual may be in the form of electronic data that can be referenced on a screen.
(Step 2) The person instructing reads out the inspection items in the manual to the worker, thereby communicating the inspection instructions to the worker.
(Step 3) The worker repeats the inspection items that were conveyed to the person instructing that the inspection instructions have been received, and then performs the inspection work in accordance with the inspection instructions.
(Step 4) The instructor checks whether the worker has performed the inspection correctly by checking the contents of the repetition. If an inspection item was not repetitive, there is a risk of a mistake in the work, so the instructor checks it again with the worker.

Here, in order to reduce the burden on the instructor in (Step 4), we consider introducing a voice recognition system to mechanically check whether the contents of the instructor and the worker's recitation are correct. However, if the contents of the manual were read out verbatim, it would take a long time to communicate intentions, and it would be a heavy burden on both the instructor and the worker. Therefore, the worker tends to communicate important words (hereinafter, important words) extracted from the manual to avoid misunderstanding the contents of each inspection item, and omit other unimportant words in the manual.
Therefore, in a voice recognition system to be introduced into inspection work, a key point to be checked mechanically is whether or not important words were correctly uttered by the instructor and the worker.

Thanks to advances in speech recognition systems, the accuracy of text conversion, which converts spoken voice data into text data, continues to improve. On the other hand, in terms of whether the work content of an inspection item is correctly communicated from the supervisor to the worker, how key words that characterize the inspection item are set has a significant impact on the accuracy of the communication check. Note that communication check accuracy refers to the accuracy of confirming that the work content has been correctly communicated, in other words, the accuracy of the system that checks communication using key words.

For example, suppose that the inspection manual states an inspection item such as "Periodic inspection: pull out the cable from cable terminal Z1 in terminal block number Y1 in panel number X1."
Assume that the person giving the instruction utters, "Start regular inspection. Pull out the cable from cable terminal Z1 of panel number X1."
Assume that the worker repeats, "Pull out the cable from cable terminal Z1."
In this case, the instructor and the worker have the same understanding of the task, "pulling out the cable from cable terminal Z1." However, the worker does not repeat the instructor's "panel number X1," and may mistakenly perform the task for "panel number X2." Furthermore, the worker does not say "terminal block number Y1," and may mistakenly perform the task for "terminal block number Y2."
On the other hand, the worker did not repeat the commander's statement, "Start regular inspection," but this discrepancy did not affect the communication of the work content in any way.

In this way, simply comparing what the instructor said with what the worker said and detecting whether they match or mismatch may not always be enough to determine whether the inspection items listed in the instructions have been correctly communicated from the instructor to the worker. In the above example, it is desirable to set four important words: "panel number X1," "terminal block number Y1," "cable terminal Z1," and "pull out the cable." If these four important words are covered by the speech of the instructor and the worker, it can be assumed that the work content has been correctly communicated. Also, the more important words there are, the better, and the word "regular inspection" may be excluded from the list of important words.

However, conventional techniques do not provide a mechanism for supporting the setting of important words from the viewpoint of how to set important words to ensure that the work content is correctly communicated. For example, Patent Document 1 describes that the contents of the inspector's speech are summarized based on predetermined keywords indicating the work inspection location, predetermined keywords indicating the equipment to be inspected, and predetermined keywords indicating the inspection overview. However, it does not describe how to set the "predetermined keywords."
Although the above is an example of inspection, the above problems can occur during general work.

Therefore, the main objective of this invention is to appropriately set important words that indicate the work content.

In order to solve the above problems, the work support system of the present invention has the following features.
The present invention provides an important word extraction device that trains an evaluation model that evaluates important words included in a manual from a sentence structure extracted from the manual so as to output words included in a past manual as important words included in the past manual;
and a work support device which inputs sentence structures extracted from the current manual into the evaluation model, and registers important words output from the evaluation model in an important word dictionary as keywords to be used in work based on the current manual.
Other means will be described later.

According to the present invention, it is possible to appropriately set important words that indicate the work content.

FIG. 2 is a schematic diagram of an inspection operation according to the present embodiment. 1 is a configuration diagram of an inspection support device according to an embodiment of the present invention. FIG. 2 is a hardware configuration diagram of the inspection support device according to the present embodiment. 4 is a flowchart showing an overview of processing by a dictionary generating unit according to the present embodiment. 11 is a table showing repetition confirmation results extracted from collected data relating to the present embodiment. 1 is an explanatory diagram showing two types of sentence structures extracted from instruction manual data according to the present embodiment; 7 is a table showing a result of dependency analysis of FIG. 6 according to the present embodiment; FIG. 2 is a configuration diagram showing details of a first example of a dictionary generating unit according to the present embodiment. FIG. 9 is an explanatory diagram showing details of the evaluation model of FIG. 8 according to the present embodiment. FIG. 13 is a configuration diagram showing details of a second example of a dictionary generating unit according to the present embodiment. FIG. 11 is an explanatory diagram showing details of the evaluation model of FIG. 10 according to the present embodiment. FIG. 13 is a configuration diagram showing details of a third example of a dictionary generating unit according to the present embodiment. FIG. 13 is a configuration diagram showing details of a fourth example of a dictionary generating unit according to the present embodiment.

This embodiment will be described with reference to the drawings.

FIG. 1 is a schematic diagram of the inspection work.
In the following, in this specification, an inspection of equipment is illustrated as an example of work, but the present invention can also be applied to any work site where work is done in a team, such as a disaster site. In this team, one inspection instructor 1 instructs one or more inspection workers 2 to carry out the inspection work, and the inspection workers 2 repeat important words previously spoken by the inspection instructor 1 in response to the utterance, thereby carrying out the inspection work in the following procedure.
(Step 1) Inspection instructor 1 checks the "instruction manual" that describes the work procedures for each inspection item on a tablet terminal (inspection terminal 3 in Figure 2) held by inspection instructor 1.
(Step 2) The inspection instructor 1 communicates the inspection instructions to the inspection worker 2 by reading out the inspection items in the manual to the inspection worker 2.
(Step 3) The inspection worker 2 repeats the transmitted inspection items to notify the inspection instructor 1 that the inspection instructions have been received. Then, the inspection worker 2 carries out the inspection work in accordance with the inspection instructions.
(Step 4) The work support system 30 in FIG. 2 performs voice recognition on the spoken content 102 of (Step 2) and (Step 3), and feeds back the repetition confirmation result obtained from the result to the inspection instructor 1 by voice or screen display.

The procedure data 101 is data describing the work procedures for the inspection items. The procedure data 101 in Fig. 1 describes the inspection items as text, but may be in other data formats such as a checklist or a checklist.
The utterance contents 102 show the utterances of (step 2) and (step 3) between the inspection instructor 1 and the inspection worker 2 in accordance with the procedure data 101. For example, in response to the inspection instructor 1's "Please lift the cable," the inspection worker 2 repeats, "Lift it, right?". This shows that the inspection work action "lift" has been communicated from the inspection instructor 1 to the inspection worker 2.

The confirmation screen 103 is a display screen of the repetition confirmation result recognized by the work support system 30 from the spoken content 102 in (step 4), and is displayed on a tablet terminal (the inspection terminal 3 in FIG. 2) held by the inspection instructor 1. The confirmation screen 103 has a procedure column 103A and a warning column 103B.
The result of the repetition check is reflected in the instruction manual column 103A. For example, the instruction manual column 103A displays important words (such as lift and 1-2345A) written in the instruction manual data 101 underlined, and highlights important words that were repeated in the utterance content 102 by encircling them. The important words are recited in the order in which they appear in the instruction manual data 101, and if there are multiple inspection operators 2, at least one of them needs to repeat them.
The attention-calling column 103B displays a message calling attention to an important word (here, "B67") that is underlined in the procedure column 103A but not repeated.
This confirmation screen 103 allows the inspection instructor 1 to confirm the inspection items that were not repeated (inspection items with discrepancies) and to confirm again with the inspection worker 2, thereby preventing human error.

2 is a configuration diagram of the work system 100. In the work system 100, a work support system 30 and an inspection terminal 3 are connected via a network.
The work support system 30 has an inspection result notifying unit 31, a voice recognition unit 32, a dictionary generating unit 33, and a procedure notifying unit 35. The work support system 30 stores an important word dictionary 34 and procedure data 36 in a storage unit.
The procedure data 36 is data in which the work procedures for each inspection item such as the procedure data 101 in FIG. 1 are described.
The important word dictionary 34 contains important words extracted from the procedure data 36. Important words are important words that affect the accuracy of the check of the communication of the work contents, and as shown in the confirmation screen 103 of FIG. 1, important words that are not repeated are considered to mean that the work contents have not been communicated.
The dictionary generating unit 33 generates the important word dictionary 34 from the procedure data 36 (details are shown in FIG. 8, etc.). As described later, the dictionary generating unit 33 learns the contents of utterances repeated in past inspection work, and can extract important words with high accuracy while reducing the time and effort required for the manager to manually extract the important word dictionary 34.

The procedure notification unit 35 notifies the inspection terminal 3 of the procedure data 101 in FIG. 1 and the like (step 1).
The voice recognition unit 32 converts the spoken voices (such as the spoken content 102 in FIG. 1 ) acquired by a transceiver or the like from the inspection instructor 1 and the inspection worker 2 into text by voice recognition, and obtains a repetition confirmation result from the result (step 4). The voice recognition unit 32 then compares the important words registered in the important word dictionary 34 with the words repeated as the repetition confirmation result, determines the inspection result based on whether they match, and transmits the inspection result to the inspection result notification unit 31.
The inspection result notification unit 31 notifies the inspection terminal 3 of the inspection result transmitted from the voice recognition unit 32 via the confirmation screen 103 in Figure 1 or the like, and also outputs the contents of the warning column 103B in Figure 1 by voice to a transceiver held by the inspection instructor 1.

FIG. 3 is a diagram showing the hardware configuration of the work support system 30. As shown in FIG.
Each device of the task support system 30 (such as the key word extraction device and task support device described below) is configured as a computer 900 having a CPU 901 , RAM 902 , ROM 903 , HDD 904 , communication I/F 905 , input/output I/F 906 , and media I/F 907 .
The communication I/F 905 is connected to an external communication device 915. The input/output I/F 906 is connected to an input/output device 916. The media I/F 907 reads and writes data from a recording medium 917. Furthermore, the CPU 901 executes a program (also called an application or its abbreviation, an app) loaded into the RAM 902, thereby improving and controlling each processing unit. This program can also be distributed via a communication line, or recorded on a recording medium 917 such as a CD-ROM and distributed.
The dictionary generating unit 33 may be configured as a separate device outside the task support system 30 .

The following four examples will be described below as details of the dictionary generating unit 33.
(First Example) A method for creating an evaluation model 17 of words appearing in the instruction manual data 36 based on a sentence structure (details in FIG. 6) of the word sequence relationship extracted from the instruction manual data 36 and the content of utterances during past inspections (FIGS. 8 and 9).
(Second Example) A method for creating an evaluation model 17 of words appearing in the instruction manual data 36 based on the sentence structure of the word sequence relationship extracted from the instruction manual data 36 and the content of utterances during past inspections and the current exercise (rehearsal) (FIGS. 10 and 11).
(Third Example) A method for creating an evaluation model 17 of words appearing in the instruction manual data 36 based on a sentence structure of dependency relationships between words extracted from the instruction manual data 36 (details in FIG. 6) and the content of utterances during past inspections (FIG. 12).
(Fourth Example) A method for creating an evaluation model 17 of words appearing in the instruction manual data 36 based on the sentence structure of the dependency relationships between words extracted from the instruction manual data 36 and the content of utterances during past inspections and the current exercise (FIG. 13).

FIG. 4 is a flowchart showing an outline of the process of the dictionary generating unit 33.
The dictionary generating unit 33 collects data such as utterance contents from past inspection work based on past procedure data (S11). The dictionary generating unit 33 creates a past evaluation model from the data collected in S11 (S12). The processes of S11 and S12 are executed in (First Example) to (Fourth Example).
The dictionary generating unit 33 collects (S13) data such as utterance contents from the current inspection exercise work based on the current procedure data 36. This process of S13 is omitted in the (first example) and (third example).
The dictionary generation unit 33 generates the current important word dictionary 34 by inputting the current procedure data 36 to the past evaluation model of S12 (S14). In the second and fourth examples in which S13 is executed, the repetition confirmation result extracted from the speech content collected in S13 is also input to the past evaluation model of S12.
The dictionary generation unit 33 supports the current inspection by providing the important word dictionary 34 of S14 to the voice recognition unit 32 (S15), as described in Fig. 2. In response to the support of S15, the work support system 30 checks whether or not important words registered in the important word dictionary 34 have been uttered in the work based on the current procedure data 36, as described in Fig. 2, and issues a warning about important words that have not been uttered (attention call column 103B in Fig. 1).

FIG. 5 is a table showing the replay confirmation results extracted from the data collected in S11 and S13.
The repetition confirmation result is associated with the "number of repetitions" repeated by the inspection worker 2 for each word appearing in the instruction manual data 36, and the "repetition interval" indicating the number of utterances from when the inspection instructor 1 spoke until the inspection worker 2 repetitions it. For example, in FIG. 1, the inspection instructor 1 says "Please lift the cable," and the inspection worker 2 repeats "Lift, right?" in the speech bubble immediately after. Therefore, the repetition interval = 1 (responding immediately after the call). Also, there are a total of two places where the inspection instructor 2 says "lift," so the number of repetitions = 2.

FIG. 6 is an explanatory diagram showing two types of sentence structures extracted from the instruction manual data 36.
The procedure data 101 shown in Fig. 1 describes the work procedures for the inspection items as sentences. For two types of sentence structures described below, morphological analysis, which divides the sentence into words and breaks it down, is performed as preprocessing for both.
The word sequence relationship 111 is data in which word strings after morphological analysis are connected (by arrows in the drawing) in the order in which they appear in the instruction manual data 101. The word sequence relationship 111 is extracted from the instruction manual data 101 by applying, for example, a machine learning-based model.

The dependency analysis result 112 is data in which the dependency structure between parts of speech such as subject, predicate, object, and modifier that make up the sentence is output for a word string after morphological analysis. First, the inspection term "lift" registered in a dictionary of inspection words prepared in advance is extracted as the verb of the sentence. Next, words closely related to the verb "lift" that have a dependency structure with the verb "lift", such as the object of the verb "lift" (= cable, terminal) and the demonstrative "1-2345A, B67, CD890A12, terminal T1", are also extracted.

FIG. 7 is a table showing the dependency analysis result 112 of FIG.
This table associates, for each word after morphological analysis, information indicating whether the word corresponds to an inspection term and information indicating whether the word corresponds to various dependency structures (subject, object, modifier, clause, etc.) with the verb of the inspection term. "Yes" in the table indicates that the word corresponds.

FIG. 8 is a block diagram showing the details of a first example of the dictionary generating unit 33. As shown in FIG.
The dictionary generation unit 33 has a past processing unit (key word extraction device) 10A that processes data of past inspection work (S11, S12 in FIG. 4), and a current processing unit (work support device) 20A that processes data of a current inspection work (S13 to S15 in FIG. 4).
The past processing section 10A has a past speech recognition section 12, a word sequence analysis section 13A, and a learning section 15, and stores the past instructions 11, sequence words 14A, past repetition words 16, and an evaluation model 17 in the memory section.
The past procedure manual 11 is sample data used in past inspection work as the procedure manual data 36 in Fig. 2, and may be one or more (n past inspection works). Note that the work site, the inspection items performed at the work site, and the members of the team inspecting the inspection items may be the same or different between the past inspection work and the current inspection work.

The past voice recognition unit 12, as the voice recognition unit 32 in Fig. 2, obtains a repetition confirmation result from past speech based on the past manual 11 acquired from the inspection instructor 1 and the inspection worker 2. The past voice recognition unit 12 stores words uttered (repetition) by both the inspection instructor 1 and the inspection worker 2 as past repetition words 16. Note that the past repetition words 16 may store only some parts of speech (nouns, verbs, adjectives, etc.). In addition, statistical information for each word (number of repetitions, repetition intervals, etc.) may be added to the past repetition words 16 as shown in Fig. 5. Furthermore, the past repetition words 16 may be manually input instead of being extracted by voice recognition.
As explained in the word sequence relationship 111 of FIG. 6, the word sequence analysis unit 13A extracts the word sequence relationship appearing in the past manual 11 as a sentence structure, and stores the extraction result as sequence word 14A.

The learning unit 15 performs machine learning to generate an evaluation model 17 from the sequence words 14A and the previously recited words 16. The learning unit 15 then stores the evaluation model 17 learned from the past data as an evaluation model 25 of the current processing unit 20A, thereby causing the evaluation model 25 to be used in the inference process of the current data.
The evaluation model 17 is a model that has learned the relationship between the sentence structure of the instruction manual data 36 and the importance of each word that constitutes the sentence of the instruction manual data 36. The importance is the degree to which a word is likely to be selected as an important word, and can take a value ranging from 0 to 1, for example. A value of 0 indicates that the word is definitely not important, and a value of 1 indicates that the word is definitely important.
The order of the words constituting the sentences in the instruction manual data 36 often affects the importance of each word. For example, the importance of the same word "cable" will be high if it is followed immediately by a verb that means inspection, such as "lift." On the other hand, if it is followed immediately by a verb that is unrelated to inspection, such as "sell," the importance of "cable" will be low.

Here, the learning unit 15 uses the previously recited words 16 as input data for the evaluation model 17, so that the recited confirmation result is reflected in the importance of each word, as exemplified below.
The importance of words included in the previously recited words 16 (words with a repetition count of 1 or more in the table of FIG. 5) is increased above the importance of words not included in the previously recited words 16.
The more frequently a word is repeated in the table of FIG. 5, the greater the importance of the word is increased.
The shorter the repetition interval of a word in the table of FIG. 5, the greater the importance of the word is increased.

The current processing unit 20A has a word sequence analysis unit 22A and a recognition unit 24, and stores in the memory unit the current instruction manual data 36, sequence words 23A, the evaluation model 25, past repetition words 16, and the current important word dictionary 34. The word sequence analysis unit 22A, like the word sequence analysis unit 13A, extracts the sequence relationship of words appearing in the current instruction manual data 36 as a sentence structure, and stores the extraction result as sequence words 23A.
The identification unit 24 functions as a machine learning inference unit corresponding to the machine learning learning unit 15. In other words, the identification unit 24 inputs the sequence of words 23A to the evaluation model 25, and causes the evaluation model 25 to output the importance of each word of the sequence of words 23A. Then, the identification unit 24 extracts important words from the sequence of words 23A using a judgment formula such as "if importance>0.80, it is considered to be an important word" based on the importance of each word, and sets the extraction result as the current important word dictionary 34.

FIG. 9 is an explanatory diagram showing details of the evaluation model 17 in FIG.
8. A sequence of words 201 is input data to the evaluation model 17, which corresponds to the sequence of words 14A in Fig. 8. A vector conversion 202 is a processing unit that converts the sequence of words 201 into a word vector 203, and corresponds to the input layer of the evaluation model 17.
The word vector 203 is data in which each word in the sequence of words 201 is expressed by a plurality of vector component values. The first word in the sequence of words 201, "below," is converted into the first word vector 203 "x1," and the second word in the sequence of words 201, "cable," is converted into the second word vector 203 "x2." Note that the seventh word vector 203 "x7," which corresponds to the seventh word, "(2-921A)," is an unknown word because all vector component values are "1," and unknown words are likely to be technical terms and therefore likely to be important words.

The intermediate layer 204 is a processing unit that outputs the estimation result 205 of each word by inputting the word vector 203 of each word in the order of the sequence words 201, and corresponds to the intermediate layer of the evaluation model 17. In FIG. 9, a time-series DNN (Deep Neural Network) model such as an RNN (Recurrent Neural Network) or an LSTM (Long short-term memory) is used as the evaluation model 17. Therefore, the evaluation functions h1, h2, ..., h4, ..., h7 are connected one by one to the evaluation functions in the subsequent stage. The first evaluation function h1 of the intermediate layer 204 inputs the first word "below" of the sequence words 201, and the second evaluation function h1 of the intermediate layer 204 inputs the second word "cable" of the sequence words 201.

The estimation result 205 indicates a combination of the estimation result vector (y1, y2, ..., y7) and the likelihood (range of value 0 to value 1) for each word in the sequence of words 201 as the evaluation result of the intermediate layer 204, and corresponds to the input layer of the evaluation model 17.
The inference result vector is a vector that stores two values, one above and one below, where the upper value indicates that it has been classified as an "important word" by being "valued 1," and the lower value indicates that it has been classified as a "not important word" by being "valued 1."
The likelihood is a value indicating the likelihood of a word becoming an important word, and the larger the value, the more likely the word is to become an important word.

The important words 206 is a list of words corresponding to the output data of the evaluation model 17. For example, the identification unit 24 adds words classified as "important words" in the estimation result vector from the estimation result 205 and having a likelihood equal to or greater than a predetermined value (0.80) to the important words 206 as importance=likelihood. On the other hand, for words classified as "important words" in the estimation result vector, regardless of the likelihood value, the identification unit 24 excludes them from the important words 206 by setting importance=0.
The learning unit 15 inputs the word sequence 14A in Fig. 8 to the evaluation model 17 as the word sequence 201 in Fig. 9, and learns the evaluation model 17 so that the evaluation model 17 outputs the previously recited words 16 as important words 206. In other words, the previously recited words 16 are utilized for the machine learning of the learning unit 15 as teaching material data indicating the correct answers of the important words 206 output by the evaluation model 17.

The first example of the dictionary generating unit 33 described above with reference to FIGS.
The evaluation model 17 is a model that receives time-series input data.
The past processing unit 10A and the present processing unit 20A extract the sequence relationship of the words extracted from the manual data 36 as a sentence structure, and input the extracted sequence relationship of the words to the evaluation model 17 as time-series input data.

FIG. 10 is a block diagram showing the details of a second example of the dictionary generating unit 33. As shown in FIG.
The dictionary generating unit 33 has a past processing unit 10B that processes data on past inspection work, and a current processing unit 20B that processes data on a current inspection work.
The past processing unit 10B is the same as the past processing unit 10A in Fig. 8. The current processing unit 20B in the second example in Fig. 10 is obtained by adding a practice speech recognition unit 26B and a practice repetition word 27B to the current processing unit 20A in Fig. 8.

The practice speech recognition unit 26B obtains practice repetition words 27B from speech based on the current procedure data 36 acquired from the inspection instructor 1 and the inspection worker 2, in the same manner as the past speech recognition unit 12, for the practice of the current inspection work. The practice repetition words 27B indicate words that have been repetitively read, like the past repetition words 16, but the time of repetition is changed from the past to the present (this time). The practice repetition words 27B may be manually input instead of being extracted by speech recognition.
The identification unit 24 inputs the practice repetition words 27B in addition to the sequence words 23A to the evaluation model 25, thereby extracting important words from the sequence words 23A, and sets the extraction result as the current important word dictionary 34. In this way, the practice repetition words 27B are taken into consideration in determining important words, thereby improving the detection accuracy of important words.

The second example of the dictionary generating unit 33 described above with reference to FIG. 10 has the following features.
The current processing unit 20B adds words contained in the current instruction manual data 36 that were commonly uttered by multiple persons during an exercise based on the current instruction manual data 36, which is performed prior to the inspection work based on the current instruction manual data 36, to the input data of the evaluation model 17.

FIG. 11 is an explanatory diagram showing details of the evaluation model 17 in FIG.
The evaluation model 17 in Fig. 11 is basically the same as the evaluation model 17 in Fig. 9, but a practice word repetition flag 203B1 is added to the end of each word in the word vector 203B in Fig. 11. The practice word repetition flag 203B1 indicates whether each word is included in the practice words repetition 27B in Fig. 10 (value = 1) or not (value = 0). By expanding the practice word repetition flag 203B1, the intermediate layer 204 can output the estimation result 205 of each word taking into account the practice words repetition 27B.

FIG. 12 is a block diagram showing the details of the third example of the dictionary generating unit 33. As shown in FIG.
The dictionary generating unit 33 has a past processing unit 10C that processes data on past inspection work, and a current processing unit 20C that processes data on a current inspection work.
The third example in Fig. 12 is obtained by replacing the method of extracting sentence structures from the instruction manual data 36 in the first example in Fig. 8 with the method of the dependency analysis unit 13C, instead of the method of the word sequence analysis unit 13A. Therefore, in the past processing unit 10C in Fig. 12, the word sequence analysis unit 13A and the sequence words 14A are deleted from the past processing unit 10A in Fig. 8, and instead, the dependency analysis unit 13C, the dependency words 14C, and the inspection word dictionary 18C are added.
The inspection word dictionary 18C and the inspection word dictionary 21C are dictionaries in which verbs that are likely to be uttered during inspection, such as "lift," are registered in advance as inspection terms.

As shown in the dependency analysis result 112 of Fig. 6, the dependency analysis unit 13C focuses on the inspection term in the inspection word dictionary 18C from the sentence of the past manual 11, and extracts other words having a dependency relationship with the inspection term as dependency words 14C. As shown in Fig. 7, in addition to the inspection term, the dependency words 14C also include information indicating the type of dependency relationship (subject, object, ...) for each other word having a dependency relationship with the inspection term (hereinafter, "feature amount" of the word).
The evaluation model 17 in Fig. 12 is a model that learns the relationship between the sentence structure of the instruction manual data 36 and the importance of each word constituting the sentence of the instruction manual data 36, similar to the evaluation model 17 in Fig. 8. In the evaluation model 17 in Fig. 8, the sequence words 14A extracted by the word sequence analysis unit 13A are input as the sentence structure of the instruction manual data 36. On the other hand, in the evaluation model 17 in Fig. 12, the dependency words 14C extracted by the dependency analysis unit 13C and their feature amounts are input as the sentence structure of the instruction manual data 36.

The evaluation model 17 in the third and fourth examples is realized using any of the following.
- A time-series DNN model such as LSTM shown in Figure 9, used in the first and second examples.
・Model using neural networks, support vector machines, regression analysis, etc.
An evaluation function that inputs the information in the table in Figure 7 (whether the word is an inspection term, whether the word is the subject of an inspection term, etc.) as features for each word and calculates the importance of each word from those features.
This evaluation function is, for example, a full combination of features extracted from each word. The learning unit 15 inputs the dependent words 14C to the evaluation function of the evaluation model 17, and learns (updates) the contents of the evaluation function so that the evaluation function outputs the past repetition words 16 as important words 206.

The initial state of the evaluation model 17 that has not yet been learned by the learning unit 15 is created, for example, by any of the following methods.
(Method 1) An evaluation model 17 is created in which only the inspection terms in the inspection word dictionary 18C are output as important words.
(Method 2) An evaluation model 17 is created in which, in addition to the inspection terms in the inspection word dictionary 18C, words that are in dependency relationships (subject, object, ...) with the inspection terms written in the sample instructions are also output as important words.
(Method 3) An evaluation model 17 is created in which, in addition to the inspection terms in the inspection word dictionary 18C, words that are in dependency relationships with the inspection terms in the utterance contents extracted in the exercise are also output as important words.

12, the word alignment analysis unit 22A and aligned words 23A are deleted from the current processing unit 20C in FIG. 8, and instead, a check word dictionary 21C, a dependency analysis unit 22C, and a dependency word 23C are added.
The dependency analysis unit 22C, like the dependency analysis unit 13C, focuses on the inspection terms in the inspection word dictionary 21C from the sentences of the current instruction manual data 36, and extracts other words having a dependency relationship with the inspection terms as dependency words 23C. Then, the identification unit 24 inputs the dependency words 23C to the evaluation model 25 (time-series DNN model or evaluation function), and registers important words output from the evaluation model 25 in the current important word dictionary 34.

The third example of the dictionary generating unit 33 described above with reference to FIGS. 11 and 12 has the following features.
The evaluation model 17 is a model that receives input data of features indicating dependency relationships between words.
The past processing unit 10C and the present processing unit 20C extract features indicating dependency relationships between words for inspection terms indicating inspection actions extracted from the instruction manual data 36 as sentence structures, and input the extracted features to the evaluation model 17 as input data.
The current processing unit 20C may register a verb from among the important words registered in the important word dictionary 34 as a check term for the next time.

FIG. 13 is a block diagram showing the details of the fourth example of the dictionary generating unit 33. As shown in FIG.
The dictionary generating unit 33 has a past processing unit 10D that processes data on past inspection work, and a current processing unit 20D that processes data on a current inspection work.
The past processing unit 10D is the same as the past processing unit 10C in Fig. 12. The current processing unit 20D in the fourth example in Fig. 13 is obtained by adding a practice speech recognition unit 26D and a practice repetition word 27D to the current processing unit 20C in Fig. 12. The practice speech recognition unit 26D and the practice repetition word 27D are the same as the practice speech recognition unit 26B and the practice repetition word 27B in Fig. 10.

The exercise voice recognition unit 26D, similar to the past voice recognition unit 12, obtains exercise repetition words 27D from the spoken voice based on the current instruction manual data 36 obtained from the inspection instructor 1 and the inspection worker 2, for the current inspection work exercise.
The evaluation model 25 is a model that, for example, in addition to the feature quantities for each word shown in the evaluation model 17 of Figure 12 (such as whether the word is a check term), also uses statistical information on the practice repetition words 27D (such as the number of repetitions and the repetition interval in Figure 5) as input data for the evaluation function.
The identification unit 24 inputs the practice repetition words 27D in addition to the sequence words 23A to the evaluation model 25, thereby extracting important words from the sequence words 23A, and sets the extraction result as the current important word dictionary 34. In this way, the practice repetition words 27D are taken into consideration in determining important words, thereby improving the detection accuracy of important words.
Furthermore, the identification unit 24 may add verbs from the extraction results of the important words to the inspection word dictionary 21C as inspection terms to be used next time. This makes it possible to increase the number of important words to be extracted next time.

The fourth example of the dictionary generating unit 33 described above with reference to FIG. 13 has the following features.
The evaluation model 17 is a model that receives input data of features indicating statistical information about words commonly uttered by a plurality of people, in addition to features indicating dependency relationships between words.
The current processing unit 20D adds statistical information about words commonly spoken by multiple people during an exercise based on the current instruction manual data 36, which is performed prior to the inspection work based on the current instruction manual data 36, to the input data of the evaluation model 17.

The dictionary creation unit 33 of the present embodiment described above has past processing units 10A to 10D and current processing units 20A to 20D.
The past processing units 10A to 10D train the evaluation model 17, which evaluates important words included in the instruction manual data 36 from sentence structures extracted from the instruction manual data 36, so as to output words included in the past instruction manuals 11 that are commonly uttered by a plurality of persons as important words included in the past instruction manuals 11.
The current processing units 20A to 20D input the sentence structures extracted from the current instruction manual data 36 to the evaluation model 17, and thereby register the important words output from the evaluation model 17 in the important word dictionary 34 as keywords for confirming whether or not the important words were uttered during the inspection work based on the current instruction manual data 36 (a feature common to the first to fourth examples).
As a result, important words that have a proven track record in past inspection work as past repetition words 16 are also extracted into the current important word dictionary 34, so that important words that indicate the work content can be appropriately set. Furthermore, by extracting important words based on words of sentence structures extracted from the past manuals 11, words that are not used in the inspection work can be appropriately excluded from the important word dictionary 34 as noise.

Furthermore, in the first and second examples of the dictionary generating unit 33, the order of words is extracted as a sentence structure. This allows the sentence structure to be extracted with a small amount of calculation. On the other hand, in the third and fourth examples of the dictionary generating unit 33, the dependency relationship between words is extracted as a sentence structure. This allows the sentence structure to be extracted with high accuracy, going as far as the meaning of the sentence.
In the second and fourth examples of the dictionary generating unit 33, the contents of utterances during not only past inspections but also the current exercise (rehearsal) are reflected in the evaluation model 17. As a result, the same manual is used for both the exercise and the actual inspection as the current manual data 36, so that a highly accurate important word dictionary 34 can be prepared before the actual inspection.

Furthermore, the present invention is not limited to the above-described embodiments, and various other applications and modifications are possible without departing from the gist of the present invention as described in the claims. For example, the above-described embodiments are detailed and specific descriptions of the configuration of the work support system 30 in order to explain the present invention in an easy-to-understand manner, and are not necessarily limited to those including all of the components described. It is also possible to replace part of the configuration of one embodiment with a component of another embodiment. It is also possible to add a component of another embodiment to the configuration of one embodiment. It is also possible to add, replace, or delete other components from part of the configuration of each embodiment.

In addition, the above-mentioned configurations, functions, processing units, etc. may be realized in part or in whole by hardware, for example, by designing them as an integrated circuit, etc. As the hardware, a broad processor device such as an FPGA (Field Programmable Gate Array) or an ASIC (Application Specific Integrated Circuit) may be used.
Furthermore, each component of the task support system 30 according to the above-described embodiment may be implemented in any hardware as long as the respective hardware can transmit and receive information to each other via a network. Furthermore, the processing executed by a certain processing unit may be realized by a single piece of hardware, or may be realized by distributed processing by multiple pieces of hardware.

1 Inspection instructor 2 Inspection worker 3 Inspection terminal 10A, 10B, 10C, 10D Past processing section (important word extraction device)
11 Past Guidelines (Past Guidelines)
12 Past speech recognition unit 13A Word sequence analysis unit 14A Sequence word 13C Dependency analysis unit 14C Dependency word 15 Learning unit 16 Past repetition word 17 Evaluation model 18C Check word dictionary 20A, 20B, 20C, 20D Current processing unit (work support device)
21C Inspection word dictionary 22A Word sequence analysis unit 23A Sequence word 22C Dependency analysis unit 23C Dependency word 24 Identification unit 25 Evaluation model 26B, 26D Exercise voice recognition unit 27B, 27D Exercise repetition word 30 Work support system 31 Inspection result notification unit 32 Voice recognition unit 33 Dictionary generation unit 34 Important word dictionary 35 Manual notification unit 36 Manual data (current manual)
100 Work System

In order to solve the above problems, the work support system of the present invention has the following features.
The present invention provides an important word extraction device that trains an evaluation model that evaluates important words included in a manual from a sentence structure extracted from the manual so as to output words used in past work based on past manuals as important words included in the past manuals;
and a work support device which inputs sentence structures extracted from the current manual into the evaluation model, and registers important words output from the evaluation model in an important word dictionary as keywords to be used in work based on the current manual.
Other means will be described later.

Claims (11)

an important word extraction device for training an evaluation model that evaluates important words included in a manual from a sentence structure extracted from the manual so as to output words included in a past manual as important words included in the past manual;
and a work support device that inputs sentence structures extracted from the current manual into the evaluation model, and registers important words output from the evaluation model in an important word dictionary as keywords to be used in work based on the current manual. The evaluation model is a model that accepts time-series input data,
The work support system according to claim 1, wherein the important word extraction device and the work support device extract a relationship between words extracted from a manual as a sentence structure, and input the extracted relationship between words as time-series input data to the evaluation model. The work support system according to claim 2, wherein the work support device adds, to input data of the evaluation model, words included in the current manual and uttered in common by a plurality of persons during an exercise based on the current manual, which is performed before the work based on the current manual. the evaluation model is a model that receives input data of features indicating dependency relationships between words,
The work support system according to claim 1, wherein the important word extraction device and the work support device extract, as a sentence structure, features indicating dependency relationships between words for work terms indicating the actions of the work extracted from the manual, and input the extracted features as input data to the evaluation model. The task support system according to claim 4 , wherein the task support device registers a verb from among the important words registered in the important word dictionary as the task term for the next task. the evaluation model is a model that receives input data of features indicating statistical information about words commonly uttered by a plurality of persons in addition to features indicating dependency relationships between words,
The work support system according to claim 4, wherein the work support device adds statistical information about words commonly uttered by the plurality of persons during an exercise based on the current manual, which is performed prior to the work based on the current manual, to input data of the evaluation model. The work support system according to claim 1, wherein the important word extraction device outputs, from among words contained in past manuals, words contained in the past manuals that have been commonly uttered by a plurality of people as important words contained in the past manuals. The plurality of people refers to a team in which one work instructor instructs one or more workers to perform work,
The work support system according to claim 7 , wherein the words commonly uttered by the plurality of persons are words uttered first by the work instructor and repeated by the worker in response to the utterance. 2. The work support system according to claim 1, wherein the work support device checks whether or not an important word registered in the important word dictionary has been uttered in the work based on the current manual, and issues a warning about an important word that has not been uttered. A work support device comprising: a current processing unit that inputs a sentence structure extracted from a current manual into an evaluation model that evaluates important words contained in a manual from a sentence structure extracted from the manual, and registers important words output from the evaluation model in an important word dictionary as keywords to be used in work based on the current manual. The work assistance system includes an important word extraction device and a work assistance device,
the important word extraction device learns an evaluation model that evaluates important words included in a manual from a sentence structure extracted from the manual so as to output words included in a past manual as important words included in the past manual,
The work support method is characterized in that the work support device inputs sentence structures extracted from the current manual into the evaluation model, and registers important words output from the evaluation model in an important word dictionary as keywords to be used in the work based on the current manual.

Images