TWI874052B - Life support system, life support device, and life support method - Google Patents

Abstract

[Topic] To support behavioral changes based on the individual lifestyle of the target person at a low cost. [Solution] A life support device (10) comprises: a text processing unit (130) that converts sensor information obtained by a sensor group (40) into text according to a text conversion rule (120) read from a memory unit (12); a word graph generating unit (131) that arranges each word of the text converted by the text processing unit (130) in a word graph (122) that is a vector space of two dimensions or more; and a word graph transition feature extracting unit (132) that outputs a proposal screen that can compare a first word graph generated from words of the text converted from the first sensor information obtained by the sensor group (40) and a second word graph generated from words of the text converted from the second sensor information obtained by the sensor group (40).

Description

Life support system, life support device, and life support method

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

A behavior change support system has been proposed that uses a computer to support parenting to change a child's inappropriate behavior. Patent document 1 describes the essence of determining the response method to be presented to the responder by response method information indicating the response method to the behavior for each combination of the behavior, the opportunity for the behavior, and the result of the behavior.

[Prior Art Literature] [Patent Literature]

[Patent Document 1] Japanese Patent Publication No. 2019-133638

In terms of support for behavior change, in order to establish a living infrastructure that values environmental protection or Well-being, a system that provides information for improving lifestyles to subjects such as caregivers is being examined. In order to propose the lifestyles shown above, detailed and wide-ranging proposals such as which behaviors should be continued and which behaviors should be improved based on the current living habits of the subjects are more effective. In addition, since the subjects have individual lifestyles, personalized suggestions that match their own lifestyles are more convincing than health manuals that are uniform for everyone.

However, the behavior change support system disclosed in Patent Document 1 is limited to supporting the improvement of one inappropriate behavior such as "stealing younger brother's toys", and is not suitable for proposing individual lifestyles. In addition, in the method of Patent Document 1, the rule database such as the response method information for the response method to the inappropriate behavior must be manually input in advance, which costs money in the preparation of behavior change.

Therefore, the main topic of the present invention is to support behavioral changes based on the individual lifestyle of the subject at a low cost.

In order to solve the above-mentioned problems, the life support system of the present invention has the following characteristics.

The present invention is a life support system, which comprises: a sensor group for acquiring sensor information about a subject whose lifestyle is measured; and a life support device for presenting a proposal screen for a proposed lifestyle to the subject based on the acquired sensor information. The life support system is characterized in that the aforementioned life support device comprises: a text processing unit for converting the sensor information acquired by the aforementioned sensor group into text based on a text rule read from a memory unit. ; A word graph generating unit, which arranges each word of the text converted by the text processing unit in a word graph which is a vector space of two dimensions or more; and a word graph comparing unit, which outputs the above-mentioned proposal screen which can compare the first word graph generated by the words of the text converted by the first sensor information obtained by the above-mentioned sensor group and the second word graph generated by the words of the text converted by the second sensor information obtained by the above-mentioned sensor group.

Other methods will be described later.

The present invention can provide support for behavioral changes based on the individual lifestyle of the subject at a low cost.

10: Life support devices

11: Input and output department

12: Memory Department

13: Operation Department

20: Internet

30: Terminal device

31: Prompt Department

32: Input section

40:Sensor group

41A: Camera

41B: Temperature and humidity sensor

42A: Microphone

42B:Millimeter wave sensor

43A: Vital Signs Sensor

43B: Radio wave sensor

44A: Human presence sensor

44B: Depth sensor

45A: Door sensor

45B: Air pressure sensor

46A: Accelerometer

46B: Noise sensor

47A: Gyroscope sensor

47B: Vibration sensor

48A: Illuminance sensor

48B: Taste sensor

49B: Home appliance sensor

100: Life support system

110: Information Input Department

111: Information output department

112: Recommended input department

120: Textual rules

121: Behavior Corpus

122: Word map

123: Attribute information

124: Word category

125: Word dispersion

130: Text Processing Department

131: Word graph generation unit

132: Word-graph transition feature extraction section (word-graph comparison section)

133: Specific information filtering section

134: Word pattern analysis and extraction unit

135: Related word extraction section

136: Suggestions to the editorial department

200: Screen display

201,202,211,212,222:Word chart

201R,202R: Behavior Path

210: Screenshot

213: Conversion function

214: Inverse transformation function

220: Output screen

900: Computer

901:CPU

902: RAM

903:ROM

904:HDD

905: Communication I/F

906: Input/output I/F

907:Media I/F

915: Communication device

916: Input/output device

917: Recording media

1300: Category Classification Department

1301: Time Series Analysis Department

1302: Arithmetic Operation Department

1303: Comparison/Logical Operation Department

1304: Sound recognition unit

1305: Image Recognition Department

T1,T2:Table

[Figure 1] is a diagram showing the structure of the life support system of this implementation form.

[Figure 2] is a diagram showing the structure of the life support device of this embodiment.

[Figure 3] is a flowchart for explaining the text processing of the life support device of this embodiment.

[Figure 4] is a flowchart for explaining the word graph generation process of the life support device executed after Figure 3 of this embodiment.

[Figure 5] is a diagram showing the structure of the life support device of Example 1.

[Figure 6] is a table showing an example of sensor information obtained by the sensor group in Example 1.

[Figure 7] is a table showing an example of textual rules for Example 1.

[Figure 8] is a flowchart showing an example of textual rules for implementation example 1.

[Figure 9] is a flowchart showing an example of textual rules for implementation example 1.

[Figure 10] is a flowchart showing an example of textual rules for implementation example 1.

[Figure 11] is an explanatory diagram showing an example of the behavior corpus (Corpus) of Example 1.

[Figure 12] is a screen shot showing an example of a word graph in Example 1.

[Figure 13] is a screen image in which behavior path information is added to the word image of the screen image of Figure 12 of Example 1.

[Figure 14] is a diagram showing the structure of the life support device of Example 2.

[Figure 15] is a screen diagram showing the conversion function and inverse conversion function between word graphs in Example 2.

[Figure 16] is a diagram showing the structure of the life support device of Example 3.

[Figure 17] is a screen shot of a word graph that collects past opinions in Example 3.

[Figure 18] is a screen shot of a word graph that collects current opinions in Example 3.

[Figure 19] is a diagram showing the structure of the life support device of Example 4.

[Figure 20] is an explanatory diagram showing an example of a behavior corpus of Example 4.

[Figure 21] is a detailed flowchart showing the target language suggestion creation process of Example 4.

[Figure 22] is an example of a word diagram related to Example 4.

[Figure 23] is a screen shot showing an example of an output screen of the suggestion editing section of Example 4.

[Figure 24] is a diagram showing the hardware configuration of each device in the life support system of this embodiment.

FIG1 is a configuration diagram of a life support system 100. The life support system 100 is composed of a life support device 10, a terminal device 30 for information input and output, and a sensor group 40 connected via a network 20. The life support device 10 is equipped with an input and output unit 11, a memory unit 12, and a calculation unit 13.

The input/output unit 11 exchanges input/output data with the terminal device 30 or the sensor group 40. The memory unit 12 stores the pre-stored database and the data extracted by the calculation unit 13 for calculation. The calculation unit 13 executes the functions of the life support device 10.

The terminal device 30 may also be in the form of a PC, a smart phone, a smart watch, a smart speaker, etc., but the type of the terminal device 30 is not limited to these.

The terminal device 30 includes: a prompting unit 31 and an input unit 32.

The prompting unit 31 prompts the service provider or service user the information output by the life support device 10. The input unit 32 inputs the behavior, status, or surrounding environment of the service user by the service provider or service user himself.

The sensor group 40 collects information about the behavior or status of the target person (the service user of the life support device 10 and its related persons) whose lifestyle is measured, or information about the surrounding environment (house, etc.) of the target person. The sensor group 40 includes at least one of the following: camera 41A, microphone 42A, vital sign sensor 43A, human sensor 44A, door sensor 45A, acceleration sensor 46A, gyro sensor 47A, illumination sensor 48A, temperature and humidity sensor 41B, millimeter wave sensor 42B, radio wave sensor 43B, depth sensor 44B, air pressure sensor 45B, noise sensor 46B, vibration sensor 47B, smell sensor 48B, and home appliance sensor 49B.

The vital sign sensor 43A is installed, for example, by an elderly person living in a service-oriented residence (hereinafter referred to as a "residence for the elderly with service-oriented residence"). The human sensor 44A, the door sensor 45A, the illumination sensor 48A, and the home appliance sensor 49B are installed, for example, in the service-oriented residence.

In addition, the sensor group 40 may be partially or entirely included in the terminal device 30. For example, if the terminal device 30 is a smart phone, the information of the acceleration sensor 46A or the gyroscope sensor 47A, GPS, etc. may be output to the life support device 10.

FIG2 is a diagram showing the structure of the life support device 10.

The input/output unit 11 includes: an information input unit 110, an information output unit 111, and a suggestion input unit 112.

The information input unit 110 receives sensor information obtained by the sensor group 40, sensor information loaded in the terminal device 30, or text information input by the service provider or service user through the terminal device 30.

The information output unit 111 outputs information such as the word graph 122 generated by the calculation unit 13, or the extracted word graph transition characteristics (described in detail later in FIG. 4), or the edited suggestions to the terminal device 30.

The suggestion input unit 112 receives input of a suggestion sentence to be edited.

The memory unit 12 stores: textual rules 120, behavior corpus 121, word graph 122, attribute information 123, word norm 124, and word distribution representation 125.

The textual rules 120 store rules for using the sensor information input from the sensor group 40 to treat the behavior or status of the service user and the surrounding environment as text information. For example, it is a list of texts corresponding to the results of sensor information classification, time series analysis, arithmetic operation, comparison operation, logical operation, voice recognition, image recognition, etc.

The behavior corpus 121 stores text information representing the behavior or status of the service user (or its related parties) or text information representing the surrounding environment as a database. Each sentence in the text contains information on "content of behavior (Do)" or "content of status (Be)", and at least one of the information of "when (When)", "where (Where)", "who (Who)", "what to do (What)", and "with whom (With Whom)" that is linked to it.

The word graph 122 is data mapped based on the similarity of the meanings of the words in the usage behavior corpus 121. The meaning of each word is vectorized by a count-based method (a method of reducing the dimension of the co-occurrence matrix of words) or an inference-based method of obtaining the word distribution representation 125.

The arrangement of words in the word graph 122 close to or far from each other is determined by the size of the inner product of the unit vector representation of each word. Thus, words with a large similarity in the meaning used are arranged closer, and words with a small similarity in the meaning used are arranged farther away.

The word graph 122 may include not only the behavior or status of the service user, but also the names of places or rooms in the home or office, furniture or home appliances and their operating status, names or nicknames of the service user or his/her related persons, time periods such as morning or evening, days of the week, etc. In addition, the words in the word graph 122 may also be clustered according to the degree of their similarity.

For example, the places or behaviors, furniture or home appliances, people, time periods, etc. that belong to the clusters containing words such as "dining" or "dining room" can be classified as "dining clusters", and the places or behaviors, furniture or home appliances, people, time periods, etc. that belong to the clusters containing words such as "sleep" or "bedroom" can be classified as "sleep clusters". The number of clusters is determined by considering the Elbow method or the Silhouette method, etc.

Attribute information 123 is labeled data of the attributes of individuals or groups that are the sensing objects, such as the degree of care required by the elderly.

The word norm 124 refers to the length of the word vector derived when the words included in the behavior corpus 121 are vectorized. In the vectorized representation of words, the more specific the context and the more frequently appearing the words, the larger the norm is compared with other words. Therefore, the words with a larger vector norm can be extracted as words that more clearly express the behavior tendency, thinking tendency, habits, etc. of the service user.

The word dispersion representation 125 is generated by converting the words in the behavior corpus 121 into numerical vectors to grasp their meanings using an inference-based method such as Word2Vec.

The calculation unit 13 includes: a text processing unit 130, a word graph generation unit 131, a word graph transition feature extraction unit (word graph comparison unit) 132, a specific information filtering unit 133, a word pattern analysis extraction unit 134, a related word extraction unit 135, and a suggestion editing unit 136.

The text processing unit 130 applies the following example algorithm to the sensor information to convert the sensor information into text.

‧Through the classification algorithm of the class classification unit 1300.

‧The time series analysis algorithm performed by the time series analysis unit 1301.

‧The arithmetic operation algorithm performed by the arithmetic operation unit 1302.

‧Comparison operation algorithm and logic operation algorithm performed by the comparison/logic operation unit 1303.

‧The voice recognition algorithm of the voice recognition unit 1304 is used.

‧The image recognition algorithm provided by the image recognition unit 1305.

In the texting by the texting processing unit 130, the processing result of a single type of sensor information can be used, or two or more types of sensor information can be used. For example, by combining the door sensor of the entrance with the directional human sensor, it is considered to classify/determine whether the text "going out" or "returning home" is generated. In addition, texting can also be performed based on the texting rule 120 that establishes a link with the sensor information or its classification result.

The word graph generation unit 131 vectorizes the words in the behavior corpus 121 by using natural language processing operations such as the aforementioned Word2Vec, and generates a word graph 122 that arranges the vectorized words in a vector space of two dimensions or more. The word graph generation unit 131 can also use all the texts in the behavior corpus 121 to generate the word graph, or can use texts generated in a certain period of time such as the behavior corpus during the last month to generate the word graph.

The word graph generation unit 131 can also generate the word graph 122 in 2-dimension or 3-dimension that is highly readable to humans, or can generate the word graph 122 with a vector representation of 4 or more dimensions. The dimension of the word graph 122 is preferably a dimension that is suitable for calculating the similarity between words arranged in the word graph 122 or for clustering between arranged words. In addition, when the vector representation of words of 4 or more dimensions is arranged in the word graph 122 and visualized, the word graph generation unit 131 can also use t-SNE (t-distributed Stochastic Neighbor Embedding) or PCA (Principal Component Analysis) to reduce the dimension of the vector to 2 or 3, thereby forming a representation that is highly readable to humans.

The word graph change feature extraction unit 132 extracts a conversion function or an adjusted feature quantity for converting a word graph 122 generated by a specific individual or group that changes in time series, or different word graphs 122 generated by different individuals or groups.

The specific information filtering unit 133 filters specific information such as personal information contained in the behavior corpus and generates an anonymous behavior corpus.

The word category analysis and extraction unit 134 analyzes the categories of words included in the behavior corpus as word categories 124.

The related word extraction unit 135 performs morphological analysis on the inputted suggestion and extracts the words constituting the suggestion. Then, the related words of the suggestion to the service user are extracted from the positions on the graph of the words constituting the suggestion sentence. The suggestion editing unit 136 uses the related words extracted by the related word extraction unit 135 to edit the suggestion sentence.

FIG3 is a flowchart illustrating the text processing of the life support device 10.

In S11, the information input unit 110 determines whether the sensor information obtained by the sensor group 40 or the text information input accepted by the terminal device 30 (S12) is information about the behavior or status of the service user or service related person, and information about the surrounding environment.

If text information (S11, text) is input from the terminal device 30, the text processing unit 130 converts the result of the input of the text information (S12) into a behavior corpus 121 as shown in FIG. 11 and stores it in the memory unit 12 (S13).

If sensor information is obtained from the sensor group 40 (S11, sensor), in S20, the information input unit 110 is divided into the following according to the type of the obtained sensor information. The text processing unit 130 converts the obtained sensor information into text by processing each of the divided destinations.

‧If the sensor information is time series data (S20, time series), the classification unit 1300 will classify the value of the timestamp linked to the sensor according to the textualization rule 120, thereby textualizing the sensor information (S21).

‧If the sensor information is a vital sign sensor value, etc. (S20, vital sign, etc.), the vital sign sensor value is converted into text by the time series analysis processing performed by the time series analysis unit 1301, the arithmetic operation processing performed by the arithmetic operation unit 1302, or the logical operation processing performed by the comparison/logical operation unit 1303 (S22, as shown in Figure 8).

‧If the sensor information is a human sensor value, etc. (S20, human sensor, etc.), the comparison/logic operation unit 1303 performs a logical operation on the human sensor value and the values of various sensors (door sensor 45A, life sign sensor 43A, illumination sensor 48A, home appliance sensor 49B), thereby texting the human behavior (S23, as shown in Figures 9 and 10).

‧If the sensor information is sound data (S20, sound), the sound recognition unit 1304 converts the content of the speaker's speech obtained by the microphone 42A into text (S24).

‧If the sensor information is image data (S20, image), the image recognition unit 1305 converts the speaker's non-verbal information (emotions, gestures, etc.) obtained from the image of the camera 41A into text (S25).

The text processing unit 130 converts the text that has been texted in S21 to S25 into a behavior corpus 121 and stores it in the memory unit 12 (S13).

FIG4 is a flowchart illustrating the word graph generation process of the life support device 10 executed after FIG3.

In S31, if the processing of the later-described embodiment 2 (processing of generating a cluster graph) is performed by the settings input in advance (S31, Yes), the specific information filtering unit 133 filters specific information such as personal information from the behavior corpus 121 (S32) to protect privacy. The cluster graph refers to a word graph 122 generated from an unspecified plurality of cluster data. The cluster data is the data of each individual or the data of family members of the same family.

By means of Example 2, a common word graph 122 can be extracted from the data of an unspecified plurality of groups having the same attribute (e.g., the same degree of care required) indicated by the attribute information 123. On the other hand, in Example 11 described later, a word graph 122 is generated from the data of an individual.

In S41, the word graph generation unit 131 generates a word graph 122 from the behavior corpus 121 of S13 or S31. The word graph generation unit 131 performs its function regularly, such as once a week or once every few months, whenever a certain number of new texts are added to the behavior corpus 121.

In S42, the word graph transition feature extraction unit 132 extracts "word graph transition features" that describe the time series transition of the word graph 122 generated in S41. The word graph transition feature refers to a feature that indicates how the words are arranged between the multiple word graphs 122, or whether the clusters that group the words have changed. The transition between the multiple word graphs 122 may refer to the transition over time caused by the same person, or may refer to the transition caused by different people.

In S51, if the processing of the later-described embodiment 3 (processing of collecting discussion opinions) is performed by the settings input in advance (S51, Yes), the word norm analysis and extraction unit 134 analyzes the norm of the word vector for the words included in the behavior corpus 121, and sorts the words according to the size of the norm, thereby extracting keywords (S52).

Next, the word graph generation unit 131 generates a word graph 122 (opinion graph) representing the result of collecting opinions based on the keywords extracted in S52. In this opinion graph, the common identification or majority opinions are reflected in an easy-to-view format by free descriptions of unspecified majority groups.

In S61, if the processing of the later-described embodiment 4 (processing of supplementing the content of the suggestion based on the target word) is performed by the settings input in advance (S61, Yes), the suggestion editing unit 136 creates suggestions based on the similarity between the words in the word graph 122 and the target word (S62, as shown in FIG. 21).

In S43, the word graph transition feature extraction unit 132 arranges and displays the word graph 122 before the transition and the word graph 122 after the transition. The user can grasp the word graph transition feature by comparing the word graphs 122 of both sides. Alternatively, the word graph transition feature extraction unit 132 can compare the word graphs 122 of both sides to mechanically extract the word graph transition feature.

In S44, the terminal device 30 generates a proposal for providing services such as nursing services based on the word graph transition features extracted in S42. The proposal for providing services refers to a proposal for selecting a new service to use or a proposal for adjusting a service that is already in use. The processing of S44 can also be performed by the service user or service provider inputting the proposal content to the terminal device 30, or it can be mechanically generated by a robot or home appliance that provides the service.

[Implementation Example 1]

In Example 1, a case where the life support system 100 is applied to the life habits of an elderly person living in a housing for the elderly with service (hereinafter referred to as a service-oriented housing) is illustrated.

The life support system 100 comprises: a sensor group 40 for acquiring sensor information about a subject whose lifestyle is measured; and a life support device 10 for presenting a screen for proposing a lifestyle to the subject based on the acquired sensor information.

FIG5 is a configuration diagram of the life support device 10 of Example 1. The life support device 10 of FIG5 is a part of the life support device 10 of FIG2 extracted, and specifically has the following components.

‧The input/output unit 11 includes: an information input unit 110 and an information output unit 111.

‧The memory unit 12 includes: textual rules 120, behavior corpus 121, and word graph 122.

‧The calculation unit 13 includes: a text processing unit 130, a word graph generation unit 131, and a word graph transition feature extraction unit 132.

‧The text processing unit 130 includes: a category classification unit 1300, a time series analysis unit 1301, an arithmetic operation unit 1302, and a comparison/logic operation unit 1303.

The life support device 10 in FIG5 mainly has the following components.

‧The text processing unit 130 converts the sensor information obtained by the sensor group 40 into text according to the text rules 120 read from the memory unit 12.

‧The word graph generation unit 131 arranges each word of the text converted by the text processing unit 130 in the word graph 122 which is a vector space of two dimensions or more.

‧The word graph transition feature extraction unit 132 outputs a proposal screen that can compare the first word graph generated by the words of the text after the first sensor information obtained by the sensor group 40 is converted; and the second word graph generated by the words of the text after the second sensor information obtained by the sensor group 40 is converted.

FIG6 is a table T1 showing an example of sensor information obtained by the sensor group 40 of Embodiment 1.

The table T1 stores various sensor information obtained by the sensor group 40, such as acceleration data obtained by the acceleration sensor 46A, at each time stamp. Among them, "Human sensing @ porch = OFF" means that the human sensing sensor 44A that senses people in the porch does not sense a person. In addition, "Human sensing @ living room = ON" means that the human sensing sensor 44A that senses people in the living room senses a person.

Among them, the sensor information in Table T1 of Figure 6 is stored every 1 minute, but the sampling rate is not limited to that.

FIG. 7 is a table T2 showing an example of textual rules 120 used in S21.

Table T2 is the result of classifying the timestamp values into 8 types of time periods every 3 hours as defined by the Weather Bureau, as shown in the "Time" text classification.

In S21, the classification unit 1300 of the text processing unit 130 classifies the value of the timestamp linked to the sensor according to the text rule 120.

FIG8 is a flowchart showing an example of the textual rule 120 used in S22. The flowchart is executed when the heart rate rises. The textual rule 120 is generated by empirical rules such as "most people walk at a speed of more than 2.5 km per hour".

When the walking speed is above 2.5 km/h (S101, Yes), the text processing unit 130 classifies the text as "exercise" (S102). On the other hand, if the text processing unit 130 is (S101, No), the text processing unit 130 classifies the text as "abnormal heart rate" (S103).

By using the textual rule 120 of FIG. 8 , the textual processing unit 130 can derive texts such as "exercise" or "abnormal heart rate" through logical operations of heart rate and acceleration.

Alternatively, the text processing unit 130 may also generate text such as "heart rate rises" or "heart rate decreases" by, for example, analyzing the time series of the heart rate. Alternatively, the text processing unit 130 may also derive text such as "walking" by performing arithmetic operations using an accelerometer or a gyroscope sensor.

FIG9 is a flowchart showing an example of textual rule 120 used in S23. The flowchart is executed when the door open/close sensor value becomes ON.

The comparison/logic operation unit 1303 performs logical operations on the door sensor value and the human sensor value according to the flow chart of Figure 9, thereby converting them into text (S23).

When the human sensor 44A arranged at the entrance of the inner side of the home is on (S111, Yes), the text processing unit 130 classifies the composition as "going out" (S112). On the other hand, when (S111, No) and there is "going out" in the recent behavior corpus 121 (S113, Yes), the text processing unit 130 classifies the composition as "returning home" (S114). On the other hand, if (S113, No) the text processing unit 130 classifies the composition as "visiting" (S115).

FIG. 10 is a flowchart showing an example of textual rule 120 used in S23. The flowchart is executed when the human sensor value becomes ON.

The comparison/logic operation unit 1303 performs logical operations on the human sensor value and the values of various sensors (acceleration sensor 46A, illuminance sensor 48A, home appliance sensor 49B, etc.) according to the flowchart of Figure 10, thereby texting the human behavior (S23).

The text processing unit 130 may also text only the sensor information with the most recent timestamp, or may text the sensor information that is linked to multiple timestamps traced back from the most recent sensor information.

The text processing unit 130 divides the processing according to the location of the person whose human sensor value is ON (S201).

‧When in the living room, the composition book is classified as "relaxation" (S202).

‧When in the dining room, if the person moves near the dining table (acceleration @ dining table is above the critical value) (S211, Yes), the composition is classified as "dining" (S212).

‧When in the bedroom, if the bedroom is dark (illuminance @ bedroom does not reach the critical value) (S221, Yes), the composition is classified as "sleeping" (S222).

‧When in the kitchen, if the kitchen appliances are running (the door of the refrigerator or microwave oven is opened or closed, etc., as detected by the appliance sensor 49B) (S231, Yes), the composition is classified as "cooking" (S232).

FIG11 is an explanatory diagram showing an example of the behavior corpus 121 generated in S13.

The text processing unit 130 generates the behavior corpus 121 of FIG. 11 (S13) of sentences representing human behaviors such as "sleeping in the early morning" by accumulating texts representing behaviors such as "eating" or "sleeping" generated by the text rules 120 of FIG. 7 to FIG. 10 in the following example sequence.

(Sequence 1) Determine the time label (When) by the timestamp.

(Sequence 2) The location label (Where) is determined by the human sensor value.

(Sequence 3) The content of the behavior (Do) is determined by other sensors (acceleration, heart rate, door, illumination, home appliances).

(Sequence 4) Use timestamps to subdivide behavior labels (e.g. subdivide "meal" into "breakfast").

(Sequence 5) Create a sentence from (Sequence 1) to (Sequence 4) that expresses the actions performed by the same person at the same timestamp.

FIG12 is a screen shot showing an example of the word graph 122 (FIG2) extracted in S41.

The word graph generation unit 131 vectorizes the meaning of each word in the text converted by the text processing unit 130 into a vector expression, and calculates the similarity between each word based on the vector expression, so that the words with greater similarity are arranged closer in the word graph 122.

The word graph generation unit 131 generates the word graph 122 shown in FIG. 12 etc. from the behavior corpus 121 shown in FIG. 11 etc. by the following exemplified sequence (S41).

(Sequence 1) The word graph generation unit 131 subdivides the text-based article into word classes by mechanical morphological analysis. For example, the first line of Figure 11 "在早早床妺" is subdivided into the word class "早早", the word class "在", and the word class "睡妺".

(Sequence 2) A machine learning algorithm (Word2Vec) is used to vectorize words of finely differentiated word classes to obtain a distributed representation of words, and the words are plotted on the word graph 122 according to the distributed representation. Next, the word graph generation unit 131 groups (clusters) words according to the arrangement of one or more plotted words and the distance between words on the word graph 122. The distributed representation of words is shown as follows, for example.

‧Word "early morning" = [0,0.3,0.4]

‧Word "sleep" = [0.5, 0, 0.2]

(Sequence 3) The naming of the generated clusters is done by people according to empirical rules. For example, the classification of people's behaviors in the family is determined by classifying them into "bedtime-related", "relaxation-related", "outing-related", and "housework and dining-related". Among them, in the word graph 122 of Figure 12, according to the similarity between words, a total of 4 clusters are generated, including bedtime C (abbreviation of Cluster), relaxation C, outing C, and housework and dining C. However, it is not necessary to create clusters, and the number or types of clusters are not limited to them.

By following the above sequence, even if the sequence is the same, different word graphs 122 are generated because different behavior corpora 121 are input. For example, in FIG. 12 , even if the behavior corpora 121 of the same senior citizen are used, the words arranged or the clusters in which they are clustered are different in the following two word graphs.

‧Word graph 201 (first word graph) generated from sensor information (first sensor information) one week ago. In the first sensor information, the health status is "good" and the activity level is high.

‧Word graph 202 (second word graph) generated from current sensor information (second sensor information). In the second sensor information, the health status is "poor" and the activity level is low.

The word graph transition feature extraction unit 132 generates a screen display 200 in which the word graph 201 and the word graph 202 are arranged for comparison, and displays the screen display 200 on the terminal device 30 (the terminal of the service provider) etc. (S43).

The word graph 201 includes four clusters, namely, bedtime C, relaxation C, going out C, and housework and dining C, and the word graph 202 includes three clusters, namely, bedtime C, relaxation C, and housework and dining C. That is, the user can understand the situation of not going out due to deteriorating health status by comparing the word graph 201 in the screen display 200 with the word graph 202.

The word graph change feature extraction unit 132 extracts the word graph change features of the word graph 201 generated last time and the word graph 202 generated this time. The word graph change feature may also refer to a word added or disappeared from the previous word graph 122, or a cluster added or disappeared from the previous word graph 122. Alternatively, it may refer to a word whose position or cluster to which it belongs has changed from a specific word in the previous word graph 122.

Among them, in FIG12, the words that are surrounded by four corners to form the difference between the word graph 201 and the word graph 202 are used as the word graph transition feature extracted by the word graph transition feature extraction unit 132. For example, the word "bed C" in the word graph 201 when the health status is good does not include the words "abnormal heart rate" or "morning". On the other hand, the word "bed C" in the word graph 202 when the health status is bad includes the words "abnormal heart rate" or "morning".

The word graph transition feature extraction unit 132 in Embodiment 1 extracts the word graph transition features exemplified below. The extracted word graph transition features may also be included in the screen display 200 of S43.

(Feature A) "Out-of-town related cluster" has disappeared

(Feature B) Adding words such as "abnormal heart rate" or "morning" to the "bedtime-related cluster"

(Feature C) Words such as "before noon" or "after noon" are moved from the "going out related cluster" to the "relaxation related cluster"

Thus, the operator of the high-income housing service infers that the health status or activity level of the corresponding elderly person has decreased through the screen display 200 of S43. As a result, the operator of the high-income housing service proposes the following exemplary service provision through the terminal device 30 based on the acquired word graph transition characteristics (S44).

‧For (Feature A), the proposal is to promote walking outside or participating in entertainment in day care services. In addition, the proposal can also be made by words located near going out C.

‧For (Characteristic B) and (Characteristic C), raise the alert level of "health deterioration" and "lifestyle change" for service and housing staff, and propose to increase the frequency of video patrols.

FIG13 is a screen image in which behavior path information is added to the word graph of the screen image of FIG12.

The word graph transition feature extraction unit 132 generates behavior paths 201R and 202R for the first word graph 201 and the second word graph 202 by connecting the words that represent the time periods in the word graph 201 by being close to each other in the word graph 122 and in the order of the passing of the time periods, and displays the generated behavior paths 201R and 202R in the respective word graphs.

The elderly with the behavior paths 201R and 202R can understand that their current living habits (current behavior) are night-time people who stay up late at night and do not go out during the day, and their living habits a week ago (target behavior) were day-time people who went out during the day. Therefore, the elderly can obtain the discovery that they want to change their living habits from night-time people to day-time people from the screen diagram of FIG. 13.

Therefore, the word graph transition feature extraction unit 132 generates a behavior path 201R that is connected in the order of, for example, "before noon" of going out C → "evening" of housework and dining C → "evening" of relaxing C → "late night" of going to bed C. Similarly, the word graph transition feature extraction unit 132 generates a behavior path 202R that is connected in the order of "before noon" of relaxing C → "evening" of housework and dining C → "late night" of going to bed C.

As a result, a behavior path 201R connected in the order of "going out C→ housework dining C→ relaxing C→ going to bed C" is added to the word graph 201. A behavior path 202R connected in the order of "relaxing C→ housework dining C→ going to bed C" is added to the word graph 202.

[Example 2]

In Example 2, instead of generating a word graph 122 for a specific individual as in Example 1, a word graph 122 for a group with a specific attribute (here, the degree of care required) is generated, and the word graph transition characteristics between word graphs 122 with different degrees of care required are extracted. The following is an example of applying the life support system 100 to a group of elderly people living in a high-income housing facility.

FIG14 is a diagram showing the structure of the life support device 10 in Embodiment 2. The life support device 10 in FIG14 is obtained by adding attribute information 123 and a specific information filter 133 to the life support device 10 in FIG5 .

The memory unit 12 additionally stores attribute information 123 for classifying multiple subjects. The word graph generation unit 131 generates a word graph 122 from sensor information about multiple subjects with the same attribute information 123.

That is, the attribute information 123 is referred to in the process of storing the text in the behavior corpus by the text processing unit 130 (S13 in FIG. 3) to allocate the behavior corpus 121 according to the attribute. The attribute information 123 is, for example, information on the care need level of each individual, and the data of individuals with the same care need level are pooled into the same behavior corpus 121.

The specific information filter 133 removes information equivalent to personal information (name, contractor number, ID, etc.) from the behavior corpus 121, thereby creating an anonymous behavior corpus 121 (S32). That is, when the specific information filter 133 generates a word graph 122 from sensor information about multiple subjects, the words equivalent to personal information that can identify the subject are excluded from the word graph 122.

Next, the word graph generation unit 131 generates a word graph 122 (S41) for each different attribute (here, the degree of care required) of the behavior corpus 121 in S32.

FIG. 15 is a screen diagram showing the conversion function 213 and the inverse conversion function 214 between the word graphs 211 and 212.

The word graph transition feature extraction unit 132 extracts conversion rules (conversion function 213 and inverse conversion function 214) for converting word graphs 122 (FIG. 2) of different attribute information 123 to each other.

The word graph transition feature extraction unit 132 applies a conversion rule to one of the first word graph 211 and the second word graph 212 that is equivalent to the same attribute information 123 as the object, thereby generating the other word graph.

On screen 210, a word graph 211 of care need = mild and a word graph 212 of care need = severe are displayed in sequence (S43). Thus, a lifestyle for bringing the group with severe care need closer to the group with mild care need can be proposed. Therefore, the word graph transition feature extraction unit 132 derives a conversion function 213 from the word graph 211 to the word graph 212, and an inverse conversion function 214 from the word graph 212 to the word graph 211.

Specifically, the word graph transition feature extraction unit 132 obtains the conversion function 213 and the inverse conversion function 214 for converting the word graphs of different attributes shown in FIG. 15 based on the different placement tendencies of the common words (words representing time periods, or sleeping, eating, etc.) in the word graphs 122 of different care needs, as the word graph transition feature of S42.

Next, the word graph transition feature extraction unit 132 can display the specific content of the conversion function 213 and the specific content of the inverse conversion function 214 as a warning or suggestion to the elderly themselves or the senior housing service provider in addition to the screen 210. For example, it can display whether there are words such as "morning" in the meal-related cluster, and the difference in the clusters to which "before noon" and "after noon" belong. In this way, it can be recommended to eat breakfast or go out before noon to after noon for the group with a high degree of care needs.

Among them, the conversion function 213 and the inverse conversion function 214 include, for example, the following (conversion B1) to (conversion B3).

(Transition B1) Changes in the distance between words or clusters. For example, in Figure 15, the distance between "morning" and "before noon" is changed between the word graph 122.

(Conversion B2) Changes in the number of clusters.

(Transformation B3) Changes in the words that make up the cluster. For example, in Figure 15, "morning" is changed from housework, dining, C to bedtime, C.

In addition, the conversion function 213 and the inverse conversion function 214 may also be compound functions that perform (conversion B1) to (conversion B3). At this time, the word graph transition feature extraction unit 132 may only adjust the distance between specific words or clusters, or may use the distance between specific words to adjust the distance between all words or clusters.

The conversion function 213 and the inverse conversion function 214 derived as shown above are used, for example, in the following scenario.

An elderly person X with a high care requirement is admitted to a high-end residential care facility. Based on his living habits in the first week of his stay, the word graph generation unit 131 generates a word graph 122 (word graph X1). This word graph X1 visualizes the living habits of a person with a high care requirement. Elderly person X is compared with the word graph 122 (word graph X2) with the goal of improving his living habits to obtain a discovery of improvement in his life. However, since elderly person X has no past data for the first week of his stay, word graph X2 cannot be generated from the personal data of elderly person X.

Therefore, the word graph generation unit 131 applies the inverse conversion function 214 to the word graph X1, thereby creating a word graph X2 that modifies the word graph X1. The word graph X2 is based on the individual word graph X1 of the elderly person X, so it is more convincing to the elderly person X than the general (group) word graph Y1 that is seen as requiring care = mild.

In addition, it is assumed that three months have passed since the elderly X moved in. In this way, since the word graph X3 after three months of moving in can also be created, as described in Example 1, the word graph X1 of the individual in the first week of moving in and the word graph X3 of the individual after three months of moving in can also be arranged and displayed. Here, the word graph transition feature extraction unit 132 regenerates the conversion function X4 of the word graph X1→word graph X3. The word graph transition feature extraction unit 132 can also display the result of comparing the conversion function X4 of the individual with the conversion function 213 (or the inverse conversion function 214) of the group shown in Figure 15. In this way, it is possible to analyze whether the migration status of the elderly X is heading towards the deterioration or improvement of the degree of care required.

In S44, for example, the terminal device 30 of the insurance company periodically determines which attribute of the word graph 122 the elderly who have joined the service are close to, or which attribute they are changing to, and recommends reconsidering the products they have joined. Alternatively, the terminal device 30 can recommend behavioral changes to the subscriber in a way that gradually approaches the lifestyle in the word graph 122 of a mild need for care. As a result, if the word graph 122 of the elderly individual is close to a mild need for care, the insurance premium can be reduced and the added value in the insurance business can be increased.

[Implementation Example 3]

In Example 3, the case where the life support system 100 is applied in the discussion (opinion collection and improvement measures formulation) for the improvement of the working environment in the enterprise union is illustrated. In the enterprise union, opinions from union members are collected regularly through discussions or questionnaires for the improvement of the working environment. However, the collection of various opinions or the extraction of major issues and the formulation of improvement measures are mostly carried out by the union executive committee while performing their daily duties, which may cause a burden on employees. Therefore, in Example 3, a function as a tool to support opinion collection, etc. is provided.

Figure 16 is a diagram showing the structure of the life support device in Example 3.

The life support device 10 of FIG16 is obtained by deleting the category classification unit 1300, the time series analysis unit 1301, the arithmetic operation unit 1302, and the comparison/logic operation unit 1303 from the life support device 10 of FIG5. In addition, the life support device 10 of FIG16 is obtained by adding a voice recognition unit 1304, an image recognition unit 1305, a specific information filtering unit 133, a word mode analysis and extraction unit 134, and a word mode 124 to the life support device 10 of FIG5.

In S11, the information input unit 110 obtains information from the microphone or camera 41A regarding the discussion of union relations. The obtained discussion information is as follows, for example.

‧Information that the voice recognition unit 1304 converts the content of the speaker's speech into text (S24).

‧The image recognition unit 1305 converts the non-verbal information (emotions such as joy and anger) of the speaker's appearance when speaking, which is represented by the camera 41A, into text (S25).

‧The terminal device 30 inputs the text information (S12) of the questionnaire collected in writing or on the Web.

The text processing unit 130 converts the discussion information input in S24, S25, and S12 into a behavior corpus 121 and stores it in the memory unit 12 (S13).

In S52, the word category analysis and extraction unit 134 analyzes the category of words contained in the behavior corpus 121 of S13. Among them, the words that are considered to have more similar context and appear frequently have a relatively large category compared to other words, and are considered to be words that express the common problem consciousness of most union members. Next, the word category analysis and extraction unit 134 is the result of the analysis of the word category 124, sorts the words according to the size of the category, and extracts the upper number words as keywords of the labor environment problem.

As shown above, the text processing unit 130 converts the text into words representing the content of the meeting. Then, the word graph generation unit 131 excludes the words whose vector norms of each word representing the content of the meeting do not reach a predetermined value from the word graph 122.

FIG. 17 is a screen shot of the word graph 221 that collects past opinions in Example 3.

In S53, the word graph generation unit 131 generates a word graph 221 (past opinion graph) that collects past opinions based on the keywords extracted in S52 from past discussions. In the past opinion graph of FIG17, words used in similar contexts or similar topics are arranged close together (clustered), and the problems existing in the labor environment can be classified and their contents extracted.

‧Question 1 C = chores, efficiency, workload, late night work, and overtime are clustered in A: business system/workload issues.

‧Question 2 C = work from home, IT tools, communication, and lack of exercise are clustered together with B: the problem of work from home.

‧Question 3 C = wages, bonuses, basic salary, taxes, and high prices are clustered in C: Questions about wages and living expenses.

‧Question 4 C = Parental leave, maternity leave, staff shortage, and subsidies are clustered in D: Questions about balancing work and childcare.

The word graph generation unit 131 determines the number of clusters by, for example, the Elbow method, and finds the similarity between word vectors by, for example, the k-means method. In this way, the word graph generation unit 131 can derive the similarity of the meanings between words, and cluster words with high similarity of meaning (close to each other on the word graph 221).

FIG. 18 is a screen shot of the word graph 222 that collects current opinions in Example 3.

In S53, the word graph generation unit 131 generates a word graph 222 (current opinion graph) that collects current opinions based on the keywords extracted in S52 from the current discussion.

Next, the word graph change feature extraction unit 132 displays a screen comparing the past opinion graph (FIG. 17) and the current opinion graph (FIG. 18), and extracts the word graph change features between the opinion graphs of both parties. In this way, improved labor environment problems (such as "late night work"), remaining problems (such as "lack of exercise"), and emerging problems (such as "high prices") are extracted.

Thus, in S44, the terminal device 30 of the executive department of the union can formulate improvement measures required as company measures based on the keywords extracted in S52 and the word graph change characteristics extracted in S53.

[Implementation Example 4]

In Example 4, a case where a life support system 100 is applied to a behavior change recommendation service for home or office use is illustrated.

FIG. 19 is a diagram showing the structure of the life support device in Example 4. The life support device 10 in FIG. 19 is obtained by adding a related word extraction unit 135, a suggestion editing unit 136, a suggestion input unit 112, and a word distribution expression 125 to the life support device 10 in FIG. 5.

FIG20 is an explanatory diagram showing an example of the behavior corpus 121 in Embodiment 4.

Compared with the behavior corpus 121 in FIG. 11 , the behavior corpus 121 in FIG. 20 is attached with the subject of each behavior. The category classification unit 1300 performs personal identification when assigning a subject to the text. The method of personal identification can also be face recognition by using the image taken by the camera 41A, or can be a life sign such as heart rate or gait that is more private by comparing it with the image.

In S41, the word graph generation unit 131 generates a word graph 122 based on the word distribution representation 125 etc. which indicates the meaning of each word appearing in the behavior corpus 121. Since a subject is given to the behavior corpus 121, behaviors or places that are highly related to a specific person are arranged in the word graph 122.

FIG21 is a detailed flowchart showing the target language suggestion creation process (S62).

In S621, the suggestion input unit 112 receives input of a clue expression (hereinafter referred to as a clue sentence) serving as a suggestion through the terminal device 30.

The suggestion input unit 112 receives input of clue sentences. The input clue sentences directly instruct the service user to change his/her behavior. For example, the clue sentences may be behaviors related to health or Well-being, such as "Let's sleep", "Let's exercise", "Let's relax", "Let's chat", "Let's take a bath", etc. Alternatively, the clue sentences may be behaviors related to energy saving, such as "Let's turn off unused lighting", "Let's increase the temperature setting of the air conditioner", etc.

Alternatively, in S621, the suggestion input unit 112 may also use clue sentences calculated based on sensor data such as heart rate data measured by an elderly person equipped with a sensor group 40 such as a watch (smart watch) equipped with a vital sign sensor 43A as an input object.

Figure 22 is an example of a word graph 122 of Example 4.

In S622, the related word extraction unit 135 extracts related words of the clue sentence input in S621 from the word graph 122 generated in S41.

The related word extraction unit 135 performs morphological analysis on the clue sentence, identifies the words or similar words (hereinafter referred to as "target words") included in the suggested content of the analysis result on the word graph 122, and extracts the related words. For example, from the clue sentence "Relax", the target word "Relax" is identified.

Next, the related word extraction unit 135 may also generate related words (一郎, 廳, 電視) in the cluster containing the target word as shown in FIG. 22 as the extraction process of the related words for the target word. Alternatively, if no cluster is formed on the word graph 122, the related word extraction unit 135 may also select words arranged around the target word.

In S623, the suggestion editing unit 136 edits the clue sentence mechanically or manually based on the related words extracted in S622 to generate a suggestion sentence.

The suggestion editing unit 136 edits suggestions based on the related words extracted in S621 and the target words of the clue sentence. The suggestion editing unit 136 is a template for suggestion sentences such as "content of behavior (Do)", "content of state (Be)", "when (When)", "where (Where)", "who (Who)", "what (What)", and "with whom (With Whom)", and applies related words to generate suggestion sentences.

Therefore, a word dictionary such as "Living room = a noun indicating a place, which is applied to where~" is prepared in advance. In addition, the suggestion editing unit 136 interpolates the applied word classes with particles, auxiliary verbs, verbs, etc. to generate meaningful suggestion sentences. Among them, the target language of the clue sentence is the candidate of the verb to be supplemented.

FIG. 23 is a screen shot showing an example of an output screen 220 of the suggestion editing unit 136 of Embodiment 4.

In S624, the suggestion editing unit 136 outputs the suggestion sentence edited in S623 to the prompting unit 31 of the terminal device 30.

Among them, Figure 23 is an example of a notification suggestion sentence on the locked screen of a smartphone. Notifications can be output not only on the locked screen but also on a smart watch, and can also be output as voice notifications through a home robot or smart speaker.

As shown above, the related word extraction unit 135 extracts the target word from the expression of the input suggestion clue, searches for the target word from the words in the word graph 122 generated by the word graph generation unit 131, and extracts the related words arranged near the target word in the word graph 122.

Next, the suggestion editing unit 136 generates a suggestion sentence that complements the input suggestion clue based on the target word and related words extracted by the related word extraction unit 135.

FIG24 is a diagram showing the hardware configuration of each device of the life support system 100.

Each device of the life support system 100 (life support device 10, terminal device 30, sensor group 40) is configured as a computer 900 having a CPU 901, a RAM 902, a ROM 903, a HDD 904, a communication I/F 905, an input/output I/F 906, and a 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. In addition, the CPU 901 improves and controls each processing unit by executing a program (also called an application program or simply APP) that has been read into the RAM 902. Then, the program can also be distributed through a communication line or recorded on a recording medium 917 such as a CD-ROM.

Among them, the computer 900 such as the life support device 10 can also concentrate the components shown in Figure 24 on one physical computer. Alternatively, the computer 900 can also be distributed in multiple computing resources such as cloud servers or edge servers, each of which is connected through a network.

In addition, the functions of multiple devices (terminal device 30, life support device 10) can also be centrally configured into one device (for example, in the terminal device 30), or only the computing unit 13 can be configured in the cloud server.

The life support system 100 of each embodiment described above analyzes the information representing the behavior of the service user, the information representing the status, or the textualized data representing the surrounding environment as a word map 122. In this way, words representing the lifestyle of the service user are derived and used in the service content, thereby providing personalized and effective life support according to the lifestyle of the service user or its changes. The words representing the lifestyle in the word map 122 include, for example, words representing behaviors/statuses/times/places/people that are highly related to specific behaviors or states.

On the other hand, in the previous systems that support behavior change, it is not possible to obtain the same effect as the life support system of this embodiment 100 due to the following reasons.

‧The rules for behavior change are uniform for everyone, so there are cases where the content of the prompts deviates from the lifestyle of the service users.

‧When urging people to change their behavior, it is not to draw their attention, but to give direct instructions or orders, so there is a possibility of neglect or rejection.

In addition, the present invention is not limited to the above-mentioned embodiments. As long as it does not deviate from the gist of the present invention described in the scope of the patent application, other various application examples and variant examples can be obtained. For example, the above-mentioned embodiments are for the purpose of explaining the present invention in detail and specifically, and are not necessarily limited to those having all the described components. In addition, part of the components of a certain embodiment can be replaced with components of other embodiments. In addition, components of other embodiments can be added to the components of a certain embodiment. In addition, for part of the components of each embodiment, other components can be added, replaced, or deleted.

In addition, the above-mentioned structures, functions, processing units, etc. may also be partially or entirely implemented in hardware, such as by designing with integrated circuits. In terms of hardware, broad processor components such as FPGA (Field Programmable Gate Array) or ASIC (Application Specific Integrated Circuit) may also be used.

In addition, each component of the life support system 100 of the above-mentioned embodiment can be installed in any hardware if each hardware can transmit and receive information to each other through the network. In addition, the processing performed by a certain processing unit can also be realized by 1 hardware, or by distributed processing using multiple hardware.

10: Life support devices

11: Input and output department

12: Memory Department

13: Operation Department

110: Information Input Department

111: Information output department

112: Recommended input department

120: Textual rules

121: Behavior Corpus

122: Word map

123: Attribute information

124: Word category

125: Word dispersion

130: Text Processing Department

131: Word graph generation unit

132: Word-graph transition feature extraction section (word-graph comparison section)

133: Specific information filtering section

134: Word pattern analysis and extraction unit

135: Related word extraction section

136: Suggestions to the editorial department

1300: Category Classification Department

1301: Time Series Analysis Department

1302: Arithmetic Operation Department

1303: Comparison/Logical Operation Department

1304: Sound recognition unit

1305: Image Recognition Department

Claims (10)

A life support system, which has: a sensor group that obtains sensor information about a subject whose lifestyle is measured; and a life support device that presents a proposal screen for a proposed lifestyle to the subject based on the obtained sensor information, the life support system is characterized as follows: The aforementioned life support device is equipped with: A text processing unit that converts the sensor information obtained by the aforementioned sensor group into text according to the text rules read by the memory unit; A word graph generation unit that arranges each word of the text converted by the aforementioned text processing unit in a word graph that is a vector space of two dimensions or more; and The word graph comparison unit outputs the above-mentioned proposal screen that can compare the first word graph generated by the words of the text converted by the first sensor information obtained by the above-mentioned sensor group and the second word graph generated by the words of the text converted by the second sensor information obtained by the above-mentioned sensor group. A life support system as claimed in claim 1, wherein the word graph generating unit vectorizes the meaning of each word in the text converted by the text processing unit into a vector expression, and calculates the similarity between each word based on the vector expression, so that words with greater similarity are arranged closer to each other in the word graph. A life support system as claimed in claim 1, wherein the word graph comparison unit generates a behavior path for the first word graph and the second word graph, by connecting words that respectively represent time segments in the word graphs by virtue of their close positions in the word graphs and in the order in which the time segments pass, and displays the generated behavior path in each of the word graphs. As in the life support system of claim 1, wherein the memory unit additionally stores attribute information for classifying multiple subjects, and the word graph generation unit generates one word graph from sensor information about multiple subjects with the same attribute information. As in claim 4, the life support device is further provided with: a specific information filter unit, The specific information filter unit is configured to exclude words corresponding to information of an individual who can identify a specific subject from the word graph when generating the word graph from sensor information about a plurality of subjects. A life support system as claimed in claim 4, wherein the word graph comparison unit extracts conversion rules of the word graphs for converting the attribute information different from each other, and applies the conversion rules to the word graph of one of the first word graph and the second word graph that is equivalent to the attribute information identical to the object, thereby generating the word graph of the other one. As in the life support system of claim 1, wherein the text processing unit converts the text into a text containing words representing the content of the meeting, and the word graph generation unit excludes words whose vector norms of each word representing the content of the meeting do not reach a predetermined value from the word graph. A life support system as claimed in claim 1, wherein the life support device further comprises: a related word extraction unit and a suggestion editing unit, The related word extraction unit extracts a target word from an expression that becomes a clue of the input suggestion, searches for the target word in the word graph generated by the word graph generation unit, and extracts related words arranged near the target word in the word graph, The suggestion editing unit generates a suggestion sentence that supplements the expression that becomes a clue of the input suggestion based on the target word and related words extracted by the related word extraction unit. A life support device, which obtains sensor information about a subject whose lifestyle is measured by a sensor group, and presents a proposal screen for proposing a lifestyle to the subject based on the sensor information, characterized by: It has: A text processing unit, which converts the sensor information obtained by the sensor group into text based on the text rules read by the memory unit; A word graph generating unit, which arranges each word of the text converted by the text processing unit in a word graph which is a vector space of two dimensions or more; and The word graph comparison unit outputs the above-mentioned proposal screen that can compare the first word graph generated by the words of the text converted by the first sensor information obtained by the above-mentioned sensor group and the second word graph generated by the words of the text converted by the second sensor information obtained by the above-mentioned sensor group. A life support method, characterized in that: a life support device that obtains sensor information about a subject whose lifestyle is measured by a sensor group, and presents a proposal screen for a proposed lifestyle to the subject based on the sensor information, comprises: a text processing unit, a word graph generation unit, and a word graph comparison unit, The text processing unit converts the sensor information obtained by the sensor group into text based on a text rule read by a memory unit, The word graph generation unit arranges each word of the text converted by the text processing unit in a word graph that is a vector space of two dimensions or more, The word graph comparison unit outputs the proposal screen that can compare the first word graph generated by the words of the text converted from the first sensor information obtained from the sensor group and the second word graph generated by the words of the text converted from the second sensor information obtained from the sensor group.

Images