JP2022187821A - Interactive intention information extraction program, apparatus, and method - Google Patents

Abstract

To provide an interactive intention information extraction program, apparatus, and method capable of extracting intention information of a user from dialogue with the user in a chat format.SOLUTION: A program implemented in a computer causes the computer to operate as: text input means 11 for interactively capturing a free answer from a user; intention information extraction means 12 for extracting intention information expressing user's intention contained in the user answer using a comparative expression database 21; dialogue control means 13 for controlling progress of the dialogue with the user by referring to a classification in the comparative expression database 21 to which the extracted intention information belongs and a dialogue scenario database 23; and log analysis means 14 for analyzing and aggregating dialogue history.SELECTED DRAWING: Figure 4

Description

The present invention relates to an interactive intention information extraction program, device and method capable of extracting intention information of a user from conversations with the user in a chat format.

Multistage questions are frequently asked in various questionnaires and market surveys, but these have conventionally been done manually. As a result, it was time-consuming and costly, and complicated processing was required, from writing questions to questions to subjects, etc., to tabulating the obtained data. By the way, with the progress of information processing technology in recent years, various AI (artificial intelligence) technologies have been introduced into various fields of society, and AI has come to be used in questionnaires and the like. rice field. For example, Patent Literature 1 discloses a technique aimed at extracting appropriate information from various conversations with a user.

JP 2009-193533 A

Although Patent Literature 1 attempts to save labor by using AI, it does not describe the details of Japanese language analysis, which is indispensable for processing and is related to the success or failure of the system. Therefore, even if yes/no type or multistage questions using a menu format can be realized, it is difficult to realize a question type that allows the user to freely answer.
In order to solve this problem, by using Japanese language processing technology that uses structural comparison of sentences, multi-step processing is easy to understand and manage for users without using keyword matching that is prone to mistakes or complicated menu formats. It is the subject of the present invention to implement the query.
This multistage question is expected to be used in various fields, one of which is laddering. Laddering is a marketing research method that clarifies the utility and value of products by asking survey respondents a series of questions about products, brands, etc., like climbing a ladder. That is.
In the present invention, this laddering survey is carried out using AI, and not only the characteristics and elements of the target product etc. are extracted from the free answers from the survey subjects, but also how they are It also aims to clarify whether they are connected and create value.

In order to solve the above problems, the interactive intention information extraction program of the present invention includes:
A program that extracts intention information through multi-stage questions to a user,
the computer,
a text input means for interactively inputting a user's free answer;
intention information extracting means for extracting intention information expressing the user's intention contained in the user's answer using a comparative expression database;
It is characterized by referring to the classification in the comparative expression database to which the extracted intention information belongs and the dialogue scenario database to operate as dialogue control means for controlling the progress of the dialogue with the user.
Furthermore, it is also preferable to operate as log analysis means for analyzing and aggregating dialogue histories.
The user's response sentence is uttered in Japanese, and may be a sentence (single sentence, compound sentence, compound sentence, or multiple sentence), a word, or a combination of an adjective and a noun.

Here, the intention information is information included in the user's utterance that the listener of the user's utterance wishes to extract for a certain purpose. For example, when I asked him about beer, he said, "I drink wine, beer, or anything else. When I eat lasagna, I drink wine. The best way to drink beer is with yakitori on a hot day." Suppose you answered If the listener is conducting market research on beer, the information "Yakitori and beer on a hot day" is of interest to the listener, and they simply ignore the story about wine. In this way, information that the listener wants to extract from the utterance content is intention information. The speaker intends to talk about his favorite drinks, wine and beer, but the only mention of beer that is of interest to the listener matches his intentions. In this way, the intention information of the present invention is to be interpreted from the standpoint of the listener, and is the general ``intention'' of the utterer, such as ``thinking. The meaning is that there is a gap.
For the theme of market research on beer, a large number of expressions related to beer must be registered in the comparative expression database. This expression (either a sentence or a word) is called a "factor" in the present invention.

The comparative expression database has regions of interest that are classified into one or more arbitrary numbers for each purpose of dialogue, and sentences to be compared are registered in each region of interest, and all or part of the user's answer sentence and either is matched with the comparison target sentence, it is preferable to regard the comparison target sentence as the intention information.
A "region of interest" corresponds to a "layer" in the following embodiments, and a "comparison target sentence" corresponds to a "factor".
The sentences to be compared are also written in Japanese, and may be sentences (single sentences, compound sentences, compound sentences, or multiple sentences), words, or combinations of adjectives and nouns.

A dialog scenario database preferably stores rules that define the progression of the dialog and the individual drill-down questions that make up the multi-stage questions to the user.
As a basic form of progress of the dialogue, starting from the primary response to the primary question issued to the user, multi-stage questions are executed as many as the number of intention information extracted from the primary response, deep-dive questions that compose multi-stage questions It is preferable to extract intent information from the responses to

Preferably, the comparative expression database has an area of interest to which a comparison sentence that indicates that the user is indifferent belongs, and the dialogue control means terminates the dialogue when the comparison sentence belonging to this is extracted.

The present invention can be realized by installing an interactive intention information extraction program in a computer and operating it as an interactive intention information extraction apparatus, or as an operation method of an interactive intention information extraction apparatus.

According to the present invention, AI can automatically extract intention information from dialogue with the user. you can get an answer. In addition, although it was not easy to tally the freely input sentence type questionnaires manually, the present invention can automatically perform this by AI and express the analysis results in easy-to-read charts.
Furthermore, simply by replacing the comparative expression database, it opens up a way to enable dialogue on themes in various fields.

It is a figure explaining the process outline|summary of this Embodiment. It is an intention information extraction state transition diagram of the embodiment of the present invention. It is a figure which shows the comparison object sentence (factor) registered into the factor list of embodiment of this invention. 1 is a functional block diagram of an intention information extraction device according to an embodiment of the present invention; FIG. FIG. 4 is a diagram showing the structure of a factor list according to an embodiment of the present invention; FIG. 4 illustrates an in-depth question decision rule of an embodiment of the present invention; It is a figure explaining the basic ladder movement state (up, down, slide) of embodiment of this invention. It is a figure explaining the operation example of the laddering of embodiment of this invention. It is a figure explaining the factor extraction process of embodiment of this invention. FIG. 4 is an overall processing flow diagram of the embodiment of the present invention; It is a processing flow diagram of the primary response phase of the embodiment of this invention. It is a processing flow diagram of the laddering phase of the embodiment of the present invention. It is a figure explaining the factor restriction|limiting of the transition destination layer of embodiment of this invention. FIG. 4 is a chain diagram of extracted factors of an embodiment of the present invention; 4 is a context map of extracted factors of an embodiment of the invention; FIG. 4 is a diagram clustering and visualizing the extracted factors of the embodiment of the present invention; It is the figure which analyzed and visualized the extracted factor from the ladder|ladder connection rate and the appearance number of a factor of embodiment of this invention. FIG. 3 is a visualization of the extracted factors of the embodiment of the present invention compared between different brands; FIG. 10 is a diagram normalizing and visualizing the transition of the user's extraction factor of an embodiment of the present invention; It is a figure explaining the example which utilizes this invention to fields other than marketing, such as healthcare.

An embodiment of the present invention will be described below with reference to the drawings. In the following description, the comparative expression database will be called a "factor list", and the sentences to be compared, that is, the words and sentences registered in the factor list will be called "factors". A "user" is a speaker, but it can also mean a "subject" or a "research subject". In addition, although the entity that asks the user questions and obtains answers is called the "system," it is sometimes called the "listener" or the "reader."

≪1. Overview of this system≫
First, with reference to FIG. 1, an outline of a processing system (hereinafter referred to as "this system") using an interactive intention information extraction device (hereinafter referred to as "this device") 1 implementing an interactive intention information extraction program will be described. explain. [1] When a user-input sentence in Japanese is input to the device 1, [2] this input sentence and each factor in a factor list uploaded from the outside are read, and the factor and If a similar expression is found, this factor is extracted as intention information. Since this system is characterized by asking multistage questions to grasp the user's intention, [3] what kind of in-depth question to ask or whether to terminate the dialogue Controls the dialogue and asks the user a new question or send a closing message. A history of interactions such as questions and answers to users and extracted factors are stored in a log file. Then, [4] refer to the log file, aggregate and analyze the data, and output it visually in an easy-to-understand manner.

An overview of this system from the viewpoint of data transition is shown in FIG.
That is, an input sentence, which is an answer to one question sentence, and a layer are associated with each other. The factor list is grouped into several groups from a predetermined point of view, and each group is called a layer. In the example of FIG. 3, the factor list has three layers: Layer A, which collects factors from the viewpoint of "attributes"; Layer B, which collects factors from the viewpoint of "functions"; and Layer C, which collects factors from the viewpoint of "emotions". consists of one layer. (Layer Z will be discussed later.)
Each layer has one or more intention information, ie, an expression expressing the user's intention. For example, Layer A (attribute) has intention information such as "product package" and "product logo", Layer B (function) has intention information such as "delicious" and "delicious", and Layer C (emotion) has It has intention information such as "cheerful" and "fun", which is the "factor" in this system.
In other words, a layer is a group associated with a user's intention information.
The layer is determined by the factors contained in the user input sentence. For example, when the factor "delicious" is extracted, it is determined that the user's answer belongs to layer B (function), and the layer and question content to be asked next are determined according to the in-depth question determination rule. Usually, the layer to be questioned is the layer C or the layer A adjacent to the layer B to which the extraction factor "delicious" belongs. This is ladder-up or ladder-down, which will be described later.

The outline of the processing and data transition of this system has been described above.
Next, the configuration and operation of this system will be described in detail with reference to the drawings.

≪2. Functional block configuration of this system≫
FIG. 4 is a functional block diagram of the device 1. As shown in FIG.
This device 1 has a text input unit 11 , an intention information extraction unit 12 , a dialogue control unit 13 and a log analysis unit 14 . These functional components can be realized by executing a computer program installed in the apparatus 1. FIG.

The text input unit 11 inputs text in Japanese sent from the user U. FIG. The user U may be a mobile terminal or a personal computer used by the user. Any input format may be used, such as character input from a keyboard or touch panel, voice input, or the like. For voice input, the device 1 must be equipped with a voice recognition unit (not shown). Also, it is generally considered that questions to the user U are voiced, and a voice synthesizer (not shown) is also required.
The input text may be one or more sentences, one word, or a list of two or more words. Also, the sentence may be a simple sentence, a compound sentence, a compound sentence, or an overlapping sentence.

The intention information extracting unit 12 includes a factor extracting unit 121 that externally refers to a factor list 21, which is a comparative expression database, to extract factors that establish matching from the input text, and an extraction factor that saves factors each time they are extracted. It consists of a file 122 and an extraction factor filtering section 123 that appropriately selects from extracted factors. It should be noted that "matching is established" does not mean that each word matches, but rather that it substantially matches. For example, if the input text is "I'm going to buy a beer for my friend" and the factor is "I'm going to buy a beer for my friend", matching is established. Also, if the input text is "like beer" and the factor is "like beer", matching is established.
Information such as the selected extraction factor and layer is input to the dialog control unit 13 .

The dialog control unit 13 acquires a layer from the extracted factors, and based on the acquired information and the dialog history information up to now in the history file 22, specifies the next layer to transition to, and judges the end of laddering. a laddering processing unit 131, information from the laddering processing unit 131 and the dialogue scenario DB 23, a deep question identification unit 132 that identifies the next question to the user U, and a next question sentence to the user U or It consists of a user output unit 133 that transmits a message sentence to the user U.

The log analysis unit 14 aggregates and analyzes the data stored in the history file 22, outputs it in the form of a table or diagram on the screen, prints it, or transmits it to an external computer via a communication line. do.

Next, the factor list 21 and the dialogue scenario DB 23 storing data necessary for executing this system will be described.

The factor list 21 needs to be replaced in view of the nature of the information that the system wants to obtain from the user U, so it is uploaded from the outside each time the system starts processing. Since this factor list 21 is created by an external system different from this system, a description of how to create it will be omitted. As already mentioned, the factor list 21 is a set of layers consisting of one or more factors. Here, the data structure of one layer will be explained using FIG.

As in FIG. 5, similar factors are grouped under the same factor ID. A label is attached to each factor ID. A group (=layer) is a set of these superordinates. The index in the figure is for distinguishing the factors with the same factor ID. Each factor is represented by text in Japanese.
The factor label "CM" includes the factor "CM" and the factor "CM is interesting". In this case, "CM is interesting" includes two independent words "CM" and "interesting", and "CM" consists of one independent word.
Since one factor list 21 can include a plurality of layers, the layers A, B, and C illustrated in FIG. 3 constitute one factor list 21 .
In the factor list 21, as shown in FIG. 3, it is preferable to register a layer Z, which is a collection of factors representing the intention of "indifference".

The dialogue scenario DB 23 stores rules necessary for dialogue control.
One of them is the in-depth question determination rule illustrated in FIG. 6(1).
In-depth question decision rules describe the current layer, the direction of laddering from the previous layer, the case of question selection, specific question examples, and so on.
In Figure 6 (1), the line marked with (*1) is at the start stage where the layer has not yet been decided, and asks "What kind of image do you have about ▲▲▲?" Indicates that the first question (primary question) is prepared.
The line with (*2) indicates that you have transitioned to layer A (attribute) by ladder down, and the first question to obtain the factor related to layer A (attribute) is "What is the factor that made you feel that way?" ?” is provided.
The line with (*3) was expected to transition to layer A (attribute) by ladder down and to acquire factors related to layer A (attribute), but other layers (from B or It shows that the factors of C) have been obtained, and that questions with the same content (but with different expressions) are prepared for obtaining factors related to layer A (attributes).
The line with (*4) indicates that it slid from Layer A (attribute) to Layer A (attribute), and the first question to obtain the factor related to Layer A (attribute) is ``Do you know anything else? Do you have anything to say?” indicates that it is available.

In the in-depth question determination rule, as shown in FIG. 6(2), there is an unintended answer to the question, that is, when the factor corresponding to the factor list 21 is not found, or when there is an answer expressing indifference. We have also prepared questions for you to ask. As with the factor list 21, the in-depth question determination rule needs to be replaced in view of the nature of the information that the system wants to obtain from the user U, so it is uploaded from the outside each time the system is started.

Laddering, that is, the orientation of a series of dialogues for extracting factors from each layer, is an important piece of information for identifying in-depth questions.
In order to obtain intention information from the user appropriately and efficiently, the transition between layers considered desirable is assumed on the system side. The order of "Layer A (attribute)"→"Layer B (function)"→"Layer C (emotion)" shown in FIG. 7 is desirable, and the transition in this direction is called ladder-up. Conversely, the transition in the order of "Layer C (emotion)" → "Layer B (function)" → "Layer A (attribute)" is called a ladder down. However, when one laddering is completed and the next laddering is started, a transition to the same layer is allowed, which is called ladder slide (lateral movement).
Since these ladder-up, ladder-down, and ladder slides are very important concepts in dialogue control processing, which will be explained later, the main points will be explained here.

Laddering is done according to the following rules: This rule should also be stored in the dialogue scenario DB 23 .
★Rules for first laddering★
- Ladder starts from the start up (down).
- Up (down) from the layer by the factor acquired in the primary response (hereinafter referred to as "primary response factor" to distinguish from the "factor" acquired during laddering) until there is no further layer.
・When there is no place to go up (down), search for an empty layer and go down (up).
★Laddering 2nd rule★
・When there are multiple primary response factors, the layer by the primary response factor is increased (or decreased) until there is no further layer. In other words, it is the same as the first laddering.
- If there is only one primary answer factor, return to the first layer (the layer that first acquired the factor) and slide.
In other words, the question is asked to extract the factors of the same layer.
★ Ladder ring afterwards ★
・Try ruddering as much as possible, but if ruddering has been completed twice at the time of failure, do not slide and finish.

A specific operation example according to this laddering rule is as shown in FIG.
When laddering is established only from the ladder-up, as shown in the left column of FIG. 8, primary questions are acquired from the in-depth question determination rule, primary answers are obtained from the user, and laddering is started. For example, in response to the primary question (see row (*1) in Figure 6), "What kind of image do you have of ○○ Beer Co., Ltd.'s 'Beer Biyori'?" Assume that the primary response is "I always buy OO beer. It's my favorite maker."
From this primary response, the factor "favorite manufacturer" belonging to layer A (attribute) (see (*a) in FIG. 3) was obtained as the primary response factor. Since it is Layer B that ladders up from Layer A, the first question associated with Layer B (function), "What good can you get from that?" ). Suppose that this question is posed to the user, and the answer is "Because it is a product of XX beer, it should be delicious." From this answer, the factor "delicious" (see (*b) in Fig. 3) of Layer B (function) was acquired, so the first question corresponding to Layer C (emotion), which is subsequently laddered up from Layer B Acquire "How does that make you feel?" (see row (*6) in FIG. 6). Assume that this question is posed to the user and that the user receives the answer, "I feel positive that I will do my best tomorrow." From this answer, I was able to acquire the Leia C (emotion) factor (see (*c) in Fig. 3) that "I can feel positive", so one laddering was successful.
Thus, laddering usually starts up from the start. Factors are acquired in the order of layer A, layer B, and layer C, and when there is no layer ahead of layer C, there is no layer for which factors have not been extracted, so one laddering is completed.

The middle row in FIG. 8 shows the case where ladder-up and ladder-down are mixed in one laddering. Once the primary answer factor of layer B (function) is obtained from the answer to the primary question, the first question corresponding to layer C (emotion) is obtained. If this question is asked to the user and the factor of layer C (emotion) is obtained, there is no layer to upload next. When there are no more upload destinations, it searches for layers whose factors have not yet been extracted and downs them. Since no factor is extracted from layer A (attribute), the question of layer A (attribute) is obtained. This is the ladder down from Layer C to Layer A. If the question of layer A (attribute) is posed to the user and the factor of layer A (attribute) can be acquired, the laddering for one time is successful.

When laddering is established only from ladder-down, as shown in the right column of FIG. 8, the idea is the same except that the order of ladder-up (left column of FIG. 8) and layer is reversed.

The block configuration of the device 1, the factor list 21 and the dialogue scenario DB 23 have been described above.
Next, regarding the operation of the device 1,
(1) factor extraction, (2) dialog control including identification of in-depth questions, and (3) log data analysis will be focused on.

≪3. Operation of this system≫
<<3.1 Factor Extraction>>
Processing by the factor extraction unit 121 will be described with reference to FIG.
The user's input sentence is run through a syntactic analyzer, and based on the parsed result, it is converted to facilitate the matching process. For example, if the original input text is a compound sentence or compound sentence, it is divided into a plurality of simple sentences. At that time, ``He went to a convenience store and bought a can of beer.'' is changed to ``He went to a convenience store.'' Add the subject "wa". This is an example of complementary and redundant structures.

Factors belonging to each layer of the factor list 21 are also passed through a syntactic analyzer, and for example, the factor "I drank a lot of new beer on holiday" is decomposed as follows.
to drink
<------- many <<modify predicate>>
<--Holiday Complement
<--beer
<-- New Release 《Modify Complement》
In this way, adjuncts (particles, auxiliary verbs) are excluded. Conjunctions may be omitted. As a result of parsing, the factors are simplified compared to the user's input sentence, which is made redundant, in order to perform matching accurately.

The basis of the factor extraction process is Japanese language processing that can process any Japanese sentence using structural comparison of sentences. Since there is an existing technology for this Japanese language processing, this technology can be used. For example, the technology disclosed in Japanese Patent Application No. 2021-79401 filed on May 8, 2009 is one example.

Since the extracted factors may have overlapping intentions or may be included as part of other factors, the extracted factor filtering unit 123 performs the following processing.
First, if multiple different factors containing the same intent information are detected, we want to eliminate duplication. Various methods are conceivable for this purpose, but for example, a factor with strong restrictions is selected. Specifically, it is preferable to count the number of independent words included in one factor and select the factor with the largest number. Alternatively, the factor with the most characters may be selected. The point is, whatever the method, the appropriate one should be selected from multiple factors.
If multiple factors with the same information are detected, select the first one found. For example, when the factor "Showa" is detected three times in total in one input sentence as "Showa", "Showa", and "Showa", only the first "Showa" is left.

Factors selected by filtering are output to the laddering processing unit 131 and also stored in the history file 22 . If it is an answer to the first interview question, the history file 22 registers the user's identification information and, if necessary, personal information. If voice input is used, the user's voice may be used to automatically determine gender and approximate age. As for the extraction factor, the serial number of the answered question, the layer to which the factor belongs, the factor ID, etc. are recorded. These provide essential information for identifying the next question.

<<3.2 Dialogue control>>
First, the process from start to finish of interaction processing with the user will be reviewed according to FIG.
The entire interaction process is roughly divided into two phases, Phase 1 and Phase 2.
Phase 1 corresponds to the interview, which is the beginning of the laddering investigation, and the interview starts from the chat screen being displayed. A question first displayed on the chat screen is called a primary question, and in phase 1, primary answer factors are extracted from the user's answers to the primary question (step F1). In other words, the primary question is a question to the user for the purpose of obtaining primary answer factors for laddering.
The subsequent Phase 2 is the laddering process starting from the primary response factors extracted here, and Phase 1 is distinguished from the Phase 2 laddering. The questions asked to the user in phase 2 are called "deep questions" and are distinguished from the above "primary questions".

Phase 1 will be described with reference to the processing flow diagram of FIG.
A response to a primary question issued to a user at the beginning of an interview is called a primary response. A primary response factor is extracted from this primary response (step S11). There may be multiple primary response factors extracted. If there is an extraction factor (Yes in step S12), it is stored as a primary response factor (step S13). Suppose that two primary response factors, factor A (attribute) and factor B (function), are extracted. In this case, laddering is started in the order of factor A and factor B.
The layer to which the extracted primary response factor belongs is obtained (step S14). Here, since laddering starts from the primary answer factor A, layer A (attribute) is acquired.
Determine that the next transition destination of layer A (attribute) is layer B (function) by ladder-up, acquire a question that can extract the factor of layer B (function) from the in-depth question determination rule, and use the user terminal is displayed on the chat screen (step S15). Phase 2 begins by asking this drill-down question to the user (see step F2 in FIG. 10).

If the number of primary answer factors is 0 in step S12 (No in step S12), the primary question is issued again, but if the number of primary questions that have already been issued is 3 or more (Yes in step S16), the interview ends. (Step S17). The multi-step question ends without transitioning to phase 2, that is, without executing laddering even once.
If the number of primary questions is less than 3 (No in step S16), the primary questions are reacquired and displayed on the chat screen (step S18).
Based on the result of step S14, the first laddering is executed assuming that the factor of layer A (attribute) has been acquired, and when this is completed, the laddering of factor B (function) follows (second laddering). If three or more primary response factors are extracted in step S12, laddering is performed for the number of primary response factors. If a plurality of primary response factors can be obtained in Phase 1 of FIG. These are called the first laddering, the second laddering (and the third laddering...).

In phase 2 of FIG. 10, the first laddering is performed using the primary response factors. Phase 2 laddering is repeated for the number of primary response factors, but if only one primary response factor is obtained, this primary response factor is used for traversing. In other words, in the second laddering when only one primary response factor could be obtained, lateral movement is performed in order to reuse the same primary response factor as in the first laddering.

Returning to FIG. 10 again, phase 2 will be described.
Assume that the primary response factor of layer A (attribute) is extracted from the primary response at the start of the interview.
In-depth questions are acquired by referring to the dialogue scenario DB 23 (step F2), and factors belonging to layer B (function) are extracted from the answers returned from the user (step F3), or layer C (emotion) is extracted (step F4), it means that the ladder-up question was successful. On the other hand, when neither the factor of layer B (function) nor the factor of layer C (emotion) can be extracted, that is, when "there is no factor", it is determined as a failure, and the ladder-up question is posed again. This is repeated a maximum of three times (step F5), and when the factor cannot be extracted, this round of laddering is terminated.
Note that "no factor" refers to, for example, a case where a question was asked in anticipation of an answer about a layer (attribute), but the extracted factor was in a different layer (function or emotion). The question is repeated three times until the answer to the attribute) is obtained. However, this number can be changed.

When the factor of layer B (function) can be extracted in step F3, the ladder-up question is obtained according to the in-depth question determination rule (step F6), and the factor belonging to layer C (emotion) is included in the answer returned from the user. is extracted (step F7), it means that the ladder-up question was successful. On the other hand, when "there is no factor", the question is repeated a maximum of three times (step F8), and when the factor cannot be extracted, this round of laddering ends.
When the factor of Layer C (emotion) can be extracted in step F4, the ladder-down question is obtained according to the in-depth question determination rule (step F9), and the factor belonging to Layer B (function) is included in the answer returned from the user. is extracted (step F10), it means that the ladder-down question was successful. On the other hand, when "there is no factor", the question is repeated a maximum of three times (step F11), and when the factor cannot be extracted, this round of laddering ends.

When step F7 is reached, primary response factors have been extracted for layer A (attribute), and factors have been extracted during laddering from both layer B (function) and layer C (emotion). In this way, factors have been extracted from all layers, and this round of laddering ends (step F12).
Also when step F10 is reached, factors have already been extracted from all layers, layer A (attribute), layer C (emotion), and layer B (function), and this round of laddering ends (step F12).
Here, when the number of primary answer factors obtained in the primary answer (F1) is two or more and laddering corresponding to each primary answer factor is completed (step F13), the series of dialogue processing is terminated.

On the other hand, when the number of times of laddering is less than 2 (step F14) and the primary answer factors obtained in phase 1 (step F1) are less than 2 (step F15), the ladder slide question is obtained from the in-depth question determination rule. (step F16), and the second laddering is started (step F17). Laddering is performed at least twice, but since the second primary response factor has not been acquired in phase 1, laddering is performed from the slide of the first primary response factor.
When the number of times of laddering is less than 2 (step F14) and two or more primary response factors are obtained in phase 1 (step F1), the second laddering is started (step F18).
If a factor belonging to layer Z (indifference) is extracted (step F19), a series of interactive processing is terminated. Laddering terminates with laddering failure if no extraction factors or indifferent responses are followed. In this way, there are ways to end laddering that are never completed.

Next, according to the processing flow of FIG. 12, Phase 2 of FIG. 10 will be described in detail.
Factors are extracted from the user's answers to in-depth questions (step S101). If there is an extraction factor (Yes in step S102), it is checked whether the factor belongs to the expected layer (step S103). If the extraction factor is in the expected layer (Yes in step S104), the next transition destination layer is obtained (step S105). If all layers contain factors, the transition destination layer no longer exists. S107).
In step S107, the next in-depth question is determined by referring to the in-depth question determination rule. For example, if the expected layer is (attribute) and the next transition destination is (function), from FIG. A probing question is identified, "What good comes from that?"

If there is no extraction factor in step S102, or if there is no expected layer in step S104, the process proceeds to step S115. S116). The total number of in-depth questions in one layer is 3, and if the factor cannot be acquired even after exceeding this, the interview ends. When the number of questions is less than 3 (No in step S115), the in-depth question is acquired again (step S117).
In the case of "no extraction factor" (No in step S102), an in-depth question may be "Could you tell me more specifically?"
This "no extraction factor" occurs when the questioner's intention and the user's answer do not match. If the questioner's intended answer is still not obtained, the interview ends. (Step S116).

When no transition destination layer remains in step S106, that is, when factors are entered in all layers, this round of laddering is terminated (step S108). If there is a primary response factor for which laddering has not yet been performed, it is acquired (step S109).
If there is an unexecuted primary response factor (Yes in step S110), a new round of laddering, that is, phase 2 in FIG. 10 is started from the beginning (step S111). If there is no unexecuted primary answer factor (No in step S110), the history file 23 is referenced to obtain the number of times of laddering that has been executed (step S112). If the number of laddering completion times is less than 2 (Yes in step S113), slide questions are acquired (step S114). If the laddering completion count is two or more (No in step S113), the interview ends (step S116).

By the way, if the transition destination layer can be acquired in step S104 of FIG. 12, it is preferable to limit the factor of the planned transition destination layer by the factor extracted from the current layer.
For example, the (attribute) factor, the (function) factor, and the (emotion) factor enclosed by ellipses in FIG. 13 are related in meaning. That is, when a factor different from this is selected at the destination of laddering, it is treated as a failure, so that more accurate laddering can be expected.
From the viewpoint of the structure of the factor list 21, it is preferable to tag the factor ID of the factor extracted in the current layer, tag the factor ID of the planned transition destination layer, and link these tags. Determination as to whether the tag is linked is performed in step S104 of FIG.
Specifically, referring to the factor list 21 in FIG. 13, when the user intends the factor "I can drink in cans. I can drink in bottles." A more natural response can be achieved if the intention is "easily drinkable." However, if the factor "delicious" is intended for layer B (function), it cannot be said to be a function of this attribute, and it cannot be said to be a natural response.
In this case, the interaction control unit 13 repeats questions with different expressions within a set number of times as a control when an expected answer cannot be obtained, which makes it easier to obtain a natural answer.

The interaction control unit 13 controls overall interaction with the user. The number of times of laddering and the number of re-questions within the same ladder may be set as parameters in the storage unit (not shown) of the device 1 . A primary question, an in-depth question, a dialog end message, and the like displayed on the user's chat screen are sent to the user U via the user output unit 133 .
The dialogue control unit 13 stores a series of dialogue contents such as questions to the user, answers from the user, and layers and factors extracted from the answers in the history file 22 . In conventional questionnaire surveys, etc., questionnaire forms are collected and input into the system by keyboard input or OCR. However, in this system, it is automatically incorporated into the system as the dialogue with the user progresses.

≪3.3 Log analysis≫
The log analysis unit 14 refers to the history file 22 and visualizes the chain of factors. FIG. 14 shows a display example.
The mechanism for visualizing the chain of factors (elements) is as follows. (In FIG. 14, a new layer D (life) is provided for convenience of explanation of visualization.)
Let An, Fn, En, and Ln be the factors detected in conversations with 10 users User1 to User10 in laddering, which has attributes, functions, emotions, and lives as layers, and the number of detected factors is represented by the size of the circle. do. In FIG. 14, the factors are arranged from the top in the order of the most frequent factors in the layer.
If the chain between each factor is graphically illustrated as shown in FIG. 14, the following can be immediately understood. For example, it can be seen that the most important factor in layer D (life) is L2, followed by L3. In addition, it can be seen that users who place importance on these factors place importance on factors E8 and E10 in Leia C (emotion). Similarly, the chain of factors of layer B (function) and layer A (attribute) can also be confirmed.

In FIG. 14, the numbers of extracted factors are represented in descending order by the size of the circle, but a table may be used to represent the appearance ranking. Frequent factors can be analyzed with any expression method.
Furthermore, by multiplying not only by ladder hierarchy, but also by demographics (demographic attributes), it is possible to aggregate in various ways.

From the viewpoint of the chain of factors, the context around the factors may be visualized as shown in FIG. For example, the factor "the taste of sake is strong" is decomposed into the words "sake", "taste", and "strong" by morphological analysis, but these words co-occur in this factor. Called. In FIG. 15, factors having at least one co-occurring word in common are assumed to be related and mapped.

By performing cluster analysis of appearing factors, the factors may be grouped based on similarities and visualized. The dendrogram of FIG. 16(1) and the mapping diagram of FIG. 16(2) are examples of visualization.

Suggestions for marketing and promotion can be derived by analyzing the extracted factors from two axes, such as the rate of connection to a specific ladder and the number of appearances of the factor itself.
FIG. 17 shows (1) a visualization of analysis results and (2) a marketing interpretation example.

14 to 17 above visualize the factors obtained for the same product. It is also useful for marketing to summarize and visualize the factors obtained from surveys targeting multiple products (different brands of the same company, brands of competitors, etc.). FIG. 18(1) compares and graphs the number of factors that appeared between different brands. The number of factors may be the sum of all extracted factors or the total number for each specific factor.
Also, as shown in FIG. 18(2), the factor diversity indices may be compared and graphed. For example, when a conspicuous number of factors related to the popularity of commercials, such as "the commercial has an impact" or "the commercial is interesting," are extracted, and only a few other factors are extracted, the diversity index is a low value. take. On the other hand, when various factors such as CM, appearance, and taste are evenly extracted, the diversity index becomes a high value.

FIG. 19 shows an example of normalizing and visualizing transitions for specific users from the history saved in the log.
When User6 transitions A9→F8→L5→E8 and laddering succeeds, the order of transitions is normalized by ladder-up, so the chain is represented by the dashed circles and arrows in the figure. .
By the way, the order in which factors are obtained differs depending on ladder-up, ladder-down, and the like. However, the acquisition order for each laddering does not matter in the aggregation process, and a series of values is used as the user's answer. This is because if the order of factor extraction for each individual user is taken into account in the aggregation and analysis, it becomes complicated, and it becomes difficult to grasp the overall trend.

In addition to the above, it is conceivable to analyze and visualize the aggregated results of dialogue histories from various perspectives.

This completes the description of the embodiment of the present invention.
However, the above embodiments are merely examples. Therefore, the structure of the database such as the above factor list, the contents of the data stored, the processing flow, the method of displaying the aggregated results, etc. are merely examples.
For example, in the above-described embodiment, the standard laddering is to start from the attribute layer and dig deep to the emotional layer, but in actual research, the number of layers varies. In addition, although the order of Leia attribute→Layer function→Layer emotion is called a ladder-up, this order is an artificial arrangement and may be flexibly changed. The name of the layer (attribute, function, etc.) is also not limited.
Although both input sentences and factors were written in Japanese, it is not impossible to apply them to other languages.

In short, it is important that AI automatically conducts dialogue with the user without human intervention, and that information in the area of interest for the listener can be picked up from the dialogue.
In addition, questionnaires and market research have always required a great deal of effort to aggregate and analyze the collected data, but it is also important for AI to do this instead.
The present invention having such features can be applied not only to laddering as in the above embodiment, but also to many other fields.
Some of the application examples are as follows.
・Draw a customer journey ・Collect word-of-mouth by digging deep into reviews ・Mental care (relieving anxiety by digging deep into anxiety)
・Brain dump tool ・Coaching chatbot ・Automatic navigation ・Interactive recommendation service ・Content development for diagnosis, fortune-telling, matching, etc. A service that satisfies both the implementing entity and the user can be realized.

FIG. 20 is a diagram explaining application of the present invention to mental care.
In specific fields such as health care and psychology, the present invention can be utilized as a tool for analyzing and diagnosing responses from persons in whom specific factors have been detected in collaboration with experts in the field.
In the example shown in FIG. 20, factors with a negative tendency such as "not interested" or "don't care" were detected from some of the respondents. When referring to the history files of these respondents, we find that "I don't feel like drinking alcohol because the taste is mild", "I can't remember the name because it's too long", "The scent is strong and gives me a headache", etc. Suppose you know that you have answered By having industrial physicians and clinical psychologists analyze and diagnose such respondents from a professional perspective, it is possible to take an appropriate approach to the respondents.
In this case, analysis and diagnosis are performed by human experts, but the present invention, that is, AI, is in charge of the preliminary work of extracting respondents who are likely to have problems.

Regarding the convenience of being able to automatically process free input sentences, I would like to add an example of a multiple-choice (including two-choice) questionnaire. Suppose that there is a question "Please select one of '1. very good, 2. good, 3. normal, 4. bad, 5. very bad' regarding today's lecture". If "today's lecture" had both good and bad points, it would be hard to choose. Also, whether to select "1. Very good" or "2. Good" is subjective to the respondent, and there is no clear standard. Furthermore, there may be non-response if there are no suitable options.
The present invention overcomes the limitations of such multiple-choice questionnaires.

As a system that can accurately and labor-savingly perform interactive processing in Japanese, the present invention has the potential to be utilized in many fields. It can be used not only for consumer preference surveys, questionnaires after various events and lectures, but also for checking proficiency at cram schools.

1: Interactive intention information extraction device 12: Intention information extraction unit 13: Dialogue control unit 14: Log analysis unit 21: Comparative expression database (factor list)
22: History file 23: Dialogue scenario database (laddering rule, in-depth question determination rule, etc.)

Claims (8)

A program that extracts intention information through multi-stage questions to a user,
the computer,
a text input means for interactively inputting a user's free answer;
intention information extracting means for extracting intention information expressing the user's intention contained in the user's answer using a comparative expression database;
An interactive intention information extractor, characterized in that it operates as dialogue control means for controlling the progress of dialogue with a user by referring to the classification in the comparative expression database to which the extracted intention information belongs and the dialogue scenario database. program. said computer,
2. The interactive intention information extracting program according to claim 1, further operated as log analysis means for analyzing and aggregating dialogue histories. The comparative expression database has regions of interest classified into an arbitrary number of one or more for each purpose of dialogue, and sentences to be compared are registered in each region of interest,
2. The interactive intention information extraction according to claim 1, wherein when all or part of the user's answer sentence is matched with any of the comparison sentences, the comparison sentence is regarded as the intention information. program. 2. The interactive intent information extraction according to claim 1, wherein said dialog scenario database stores rules that define the progress of the dialog and individual in-depth questions constituting multistage questions to the user. program. The basic mode of progress of the dialogue is to start with a primary answer to a primary question issued to the user, execute multi-step questions for the number of intention information extracted from the primary answer, and deep-dive questions that compose multi-step questions. 5. The interactive intention information extraction program according to claim 4, wherein the intention information is extracted from the answer to the . The comparative expression database has a region of interest to which a comparison target sentence that indicates that the user is indifferent,
4. The interactive intention information extraction program according to claim 3, wherein said dialogue control means terminates the dialogue when a comparison target sentence belonging to this is extracted. A device for extracting intention information through multistage questions to a user,
a text input means for interactively inputting a user's free answer;
intention information extracting means for extracting intention information expressing the user's intention contained in the user's answer using a comparative expression database;
An interactive intention information extraction comprising: a classification in the comparative expression database to which the extracted intention information belongs; and a dialogue control means for controlling progress of dialogue with a user by referring to a dialogue scenario database. Device. A method for extracting intention information through multi-step questioning of a user, comprising:
the computer
a step of interactively capturing a free answer text by the user;
a step of extracting intention information expressing the user's intention contained in the user's answer using a comparative expression database;
An interactive intention information extraction method, comprising: classifying the extracted intention information in the comparative expression database; and referring to a dialogue scenario database to control progress of dialogue with a user. .

Images