Utterance selection device, utterance selection method, and utterance selection program

Abstract

An utterance selection device according to the present invention creates preference information indicating a user's preference for each topic by means of conversation with the user, and stores the preference information in a preference information DB. Next, the utterance selection device searches the preference information DB for preference information about a topic having at least a predetermined similarity to a topic of the utterance indicated in an utterance flow for a task. Also, the utterance selection device estimates the user's preference for the topic indicated in the searched preference information as the user's preference for the utterance indicated in the utterance flow for the task. Next, the utterance selection device outputs the utterances indicated in the utterance flow, while excluding those for which the user's preference is low.

Description

Utterance selection device, utterance selection method, and utterance selection program

[0001] The present invention relates to an utterance selection device, an utterance selection method, and an utterance selection program.

[0002] Dialogue systems are required to encourage users to take specific actions through dialogue with them (behavior induction). For example, user preferences are thought to have a significant influence on actions such as purchasing a product or selecting a tourist destination. Therefore, in order to induce user behavior, it is important to collect user preferences through dialogue with the user and effectively utilize the collected preferences. In addition, for efficient behavior induction, it is also important to improve the user's satisfaction with the dialogue itself.

[0003] One example of a technique for improving user satisfaction with the dialogue itself is a rule-based dialogue system. This dialogue system branches the content of the utterance to the user depending on the content of the response to the user's question and the user's emotion at the time of the response, thereby improving user satisfaction with the dialogue.

[0004] Yuki Kubo, et al., “Team OS's System for Dialogue Robot Competition 2022”, arXiv:2210.09928v1 [cs.HC] 18 Oct 2022, [online], [Retrieved September 16, 2023], Internet <URL: https: / / arxiv.org / abs / 2210.09928>

[0005] However, conventional dialogue systems do not use collected user preference information to select topics during dialogue, so they cannot effectively utilize the collected user preference information. As a result, they cannot efficiently induce user behavior. Furthermore, rule-based dialogue systems have the problem of high costs required to build rules to determine what to say to the user.

[0006] Therefore, an object of the present invention is to realize a dialogue with a user that induces the user to take action efficiently and at low cost.

[0007] In order to solve the above-mentioned problems, the present invention is characterized by comprising a memory unit that stores preference information indicating a user's level of favorability for each topic that may be covered in a conversation with the user; an estimation unit that searches the memory unit for preference information for topics whose similarity to the topic is equal to or greater than a predetermined value for each topic of utterances shown in a speech flow to the user, and estimates the user's level of favorability for each utterance shown in the speech flow based on the user's level of favorability for the topics shown in the searched preference information; and a selection unit that excludes utterances shown in the speech flow for topics whose estimated user's level of favorability is less than a predetermined value and outputs the utterances.

[0008] According to the present invention, a dialogue with a user that induces the user to take action can be realized efficiently and at low cost.

[0009] FIG. 1 is a diagram illustrating an example of the operation of an utterance selection device. FIG. 2 is a diagram illustrating an overview of the utterance selection device. FIG. 3 is a diagram illustrating an example of the configuration of the utterance selection device. FIG. 4 is a flowchart illustrating an example of a processing procedure executed by the utterance selection device. FIG. 5 is a diagram illustrating a dialogue in a chat phase in an experiment. FIG. 6 is a diagram illustrating a baseline dialogue in the experiment. FIG. 7 is a diagram illustrating a dialogue with a user using utterances selected from a task utterance flow. FIG. 8 is a diagram illustrating a computer that executes an utterance selection program.

[0010] Hereinafter, a mode (embodiment) for carrying out the present invention will be described with reference to the drawings. The present invention is not limited to this embodiment. The utterance selection device of this embodiment selects an utterance to be output to a user from, for example, utterances prepared in advance (for example, utterance flows for tasks). For example, the utterance selection device selects an utterance that is effective for assisting the user in selecting a product or suggesting a new action to the user.

[0011] [Overview] First, an overview of the utterance selection device will be described. The utterance selection device determines, based on user preference information, which utterances shown in the task utterance flow are on topics the user does not like. The utterance selection device then excludes utterances on topics the user does not like from the utterances shown in the task utterance flow and outputs them. This allows the utterance selection device to utilize the user preference information and output utterances that efficiently induce the user to take action.

[0012] An example of the operation of the utterance selection device will be described with reference to Fig. 1. First, in a "chat phase" equivalent to an icebreaker, the utterance selection device collects information indicating users' preferences for topics (preference information) and stores it in a preference information DB (database).

[0013] Next, in the "task phase," the utterance selection device selects an utterance based on the user's preferences. For example, the utterance selection device searches the preference information DB for preference information on a topic similar to the topic of the utterance indicated in the task utterance flow. Then, the utterance selection device determines whether or not the topic indicated in the task utterance flow can be uttered based on the user's preference (likeability) for the topic indicated in the searched preference information.

[0014] For example, as shown in Fig. 1, the utterance selection device collects user preferences for topics in the "chat phase." For example, the utterance selection device outputs questions to the user that allow a "like / dislike" response for each topic, and if the utterance selection device receives a response from the user that includes the intention of "like," it analyzes that the user likes that topic.

[0015] For example, the utterance selection device outputs the question "Do you like ramen?" to the user, and if the response "Yes, I like it" is received, the device analyzes that the user likes (positively) topics related to "ramen." The "like / dislike" analysis uses, for example, BERT sentiment analysis. The utterance selection device stores information (e.g., (ramen, positive)) that includes the topic and the analysis result of the user's preference for that topic as preference information in the preference information DB.

[0016] Next, in the "task phase," the utterance selection device acquires preference information (e.g., (ramen, positive), (music, negative)) from the preference information DB and the topic of the utterance in the task utterance flow (e.g., "Chinese food").

[0017] The utterance selection device then vectorizes the topics "Chinese," "ramen," "music," etc. Next, if the word with the greatest cosine similarity to the vectorized "Chinese" is "ramen," the utterance selection device searches the preference information DB for "ramen." Next, the utterance selection device extracts the emotion "positive" associated with "ramen" from the search results (the searched preference information for "ramen"). The utterance selection device then outputs information associating the emotion "positive" with the topic "Chinese." Note that if the utterance selection device is unable to search the preference information DB for a topic similar to the topic "Chinese" (for example, if no preference information for a topic with a cosine similarity to the topic "Chinese" that is equal to or exceeds a predetermined threshold), it outputs specific information such as "neutral." For example, the utterance selection device outputs information associating the emotion "neutral" with the topic "Chinese."

[0018] The utterance selection device then determines whether or not to utter the topic based on the output information. For example, if the user's feelings toward the topic "Chinese" are "positive or neutral," the utterance selection device outputs the utterance "Do you like Chinese food?" in the task utterance flow. On the other hand, if the user's feelings toward the topic "Chinese" are "negative," the utterance selection device skips the utterance of the topic "Chinese."

[0019] This allows the utterance selection device to output utterances from the task utterance flow, excluding utterances on topics that are estimated to be disliked by the user. As a result, the utterance selection device can efficiently select utterances that are appropriate for inducing user behavior, for example.

[0020] Furthermore, by executing the above process, the utterance selection device can reduce the number of times the user utters "I don't like it" during the dialogue, thereby improving the user's satisfaction with the dialogue as a whole. As a result, it is possible to realize a dialogue that is more likely to induce user behavior.

[0021] Furthermore, the utterance selection device uses the similarity between keywords to search for similar topics, which eliminates the need to prepare data describing the connections between topics in advance when selecting utterances, thereby reducing the cost of building a dialogue system.

[0022] The above process is outlined in Figure 2. In the chat phase, the utterance selection device acquires the user's response to each topic by asking the user questions, etc., and then analyzes the user's preference for each topic using sentiment analysis such as BERT. The utterance selection device then stores information (preference information) that associates the topics with the user's preference for each topic in a preference information DB.

[0023] After that, when the utterance selection device acquires the topic of the utterance (topic B) indicated in the task utterance flow, it estimates the user's feelings toward topic B. For example, the utterance selection device searches the preference information stored in the preference information DB for preference information about topic A (topic A, favorability rating A) that has the highest similarity to topic B (preference information estimation). The utterance selection device then estimates that the favorability rating (favorability rating B) of topic B indicated in the searched preference information is the same as the favorability rating A of topic A. The utterance selection device then uses topic B and favorability rating B to determine whether or not topic B can be uttered. Note that if the utterance selection device cannot find preference information about a topic similar to topic B after searching the preference information DB, it assigns a predefined value (e.g., a value indicating a specific emotion, such as "neutral") as the favorability rating of topic B.

[0024] For example, if the popularity rating of topic B is "positive" or "neutral," the utterance selection device determines that the utterance of topic B is acceptable and outputs it. On the other hand, if the popularity rating of topic B is "negative," the utterance selection device determines that the utterance of topic B is unacceptable and does not output it.

[0025] [Configuration Example] Next, a configuration example of the utterance selection device 10 will be described with reference to Fig. 3. The utterance selection device 10 includes, for example, an input / output unit 11, a storage unit 12, and a control unit 13.

[0026] The input / output unit 11 is an interface that controls the input and output of various data. The input / output unit 11 receives inputs such as user utterances, task utterance flows, etc. The input / output unit 11 also outputs utterances selected by the control unit 13.

[0027] The storage unit 12 stores data, programs, etc. that are referenced when the control unit 13 executes various processes. The storage unit 12 is realized by a semiconductor memory element such as a random access memory (RAM) or a flash memory, or a storage device such as a hard disk or an optical disk. For example, the storage unit 12 stores a preference information DB, a speech flow for a task, etc.

[0028] The preference information DB stores information (preference information) that associates topics with the user's level of preference for the topics (e.g., favorability rating). The preference information is created by, for example, the analysis unit 132.

[0029] The task utterance flow is information that shows a set of a series of utterances (utterance contents) to be output to the user and the topic of those utterances (see System 1 to System 4 in Figure 5). This task flow is designed in advance by an administrator or the like in the form of "utterance contents, topic" for each turn of utterance.

[0030] The control unit 13 controls the entire utterance selection device 10. The functions of the control unit 13 are realized, for example, by a CPU (Central Processing Unit) executing a program stored in the storage unit 12.

[0031] The control unit 13 includes, for example, a dialogue unit 131 , an analysis unit 132 , an estimation unit 133 , a selection unit 134 , and an output processing unit 135 .

[0032] The dialogue unit 131 dialogues with the user to collect preference information of the user. For example, the dialogue unit 131 outputs a question to the user to which the user can respond with likes / dislikes, and acquires the response from the user. For example, the dialogue unit 131 outputs a question to the user such as "Do you often eat ramen?" and acquires a response such as "I like it" from the user.

[0033] The analysis unit 132 analyzes the user's preference for the topic indicated in the question based on the question to the user and the response thereto obtained by the dialogue unit 131 .

[0034] For example, the analysis unit 132 uses a large-scale language processing model such as BERT to analyze the user's preference (for example, "like (positive) / neutral / dislike (negative)") for a topic indicated in a question (for example, "ramen"). If the user's utterance is x and the result (output) of the sentiment analysis by BERT is y, then the result is as shown in the following formula (1).

[0035] y = BERT(x)...Equation (1)

[0036] The analysis unit 132 then stores information (preference information) that associates the topic with the analysis result of the user's level of favorability for the topic in the preference information DB. For example, the analysis unit 132 stores information that associates "like (positive) / neutral / dislike (negative)" as the user's level of favorability for the topic in the preference information DB as preference information.

[0037] For each topic that appears in the task-use utterance flow, the estimation unit 133 searches the preference information DB for preference information on a topic whose similarity to the topic is equal to or greater than a predetermined value. If the estimation unit 133 cannot find preference information on a topic whose similarity to the topic is equal to or greater than a predetermined value from the preference information DB, the estimation unit 133 sets the user's preference level for the topic to a specific preference level, such as "neutral."

[0038] For example, the estimation unit 133 converts (vectorizes) each topic of preference information stored in the preference information DB and each topic of utterances indicated in the task utterance flow into an embedded representation, and calculates the cosine similarity between the embedded representations. Next, the utterance selection device searches the preference information DB for the preference information of the word (topic) that has the highest cosine similarity with the topic indicated in the task utterance flow.

[0039] Then, the estimation unit 133 estimates the user's preference level for the topic indicated in the searched preference information as the user's preference level for the topic indicated in the task-use utterance flow. Note that if the estimation unit 133 cannot find preference information for a topic whose similarity to the topic is equal to or greater than a predetermined value from the preference information DB, the estimation unit 133 estimates the user's preference level for the topic as, for example, "neutral."

[0040] The selection unit 134 excludes and outputs, from the utterances shown in the task utterance flow, utterances on topics for which the user's favorability rating estimated by the estimation unit 133 is less than a predetermined value. For example, from the utterances shown in the utterance flow, the selection unit 134 excludes and outputs, from the utterances shown in the utterance flow, utterances on topics for which the user's favorability rating is estimated to be "negative."

[0041] The output processing unit 135 outputs to the user the utterance output by the selection unit 134. For example, the output processing unit 135 converts the utterance output by the selection unit 134 into voice data and outputs it.

[0042] According to the utterance selection device 10, it is possible to exclude and output utterances on topics that are estimated to be disliked by the user from among the utterances shown in the task utterance flow.

[0043] [Example of Processing Procedure] Next, an example of processing procedure executed by the utterance selection device 10 will be described with reference to Fig. 4. First, the utterance selection device 10 executes processes in the chat phase (S1, S2), and then executes processes in the task phase (S3, S4, S5).

[0044] [Chat Phase] For example, the dialogue unit 131 engages in a dialogue (chat) with the user regarding preference information. The analysis unit 132 then analyzes the user's level of favorability for each topic based on questions to the user and their responses obtained through the dialogue (S1), and stores information indicating the user's level of favorability for each topic (preference information) in the preference information DB (S2).

[0045] [Task Phase] Next, the utterance selection device 10 executes processing in the task phase. For example, for each topic of utterance indicated in the task utterance flow, the estimation unit 133 searches the preference information DB for preference information related to a topic similar to the topic (S3).

[0046] Thereafter, the estimation unit 133 estimates the likeability of each utterance in the task-use utterance flow based on the user's likeability indicated in the preference information searched for in S3 (S4). For example, the estimation unit 133 searches for preference information on topic A that is similar to topic B of the utterances indicated in the task-use utterance flow. If the user's likeability indicated in the searched preference information on topic A is likeability A, the estimation unit 133 estimates likeability A as the likeability of topic B. If the estimation unit 133 cannot find preference information on a topic similar to topic B in the preference information DB, the estimation unit 133 estimates a specific emotion, such as neutral, as the likeability of topic B.

[0047] After S4, the selection unit 134 excludes utterances on topics with low favorability ratings (e.g., those with a "negative" favorability rating) from the utterances in the task flow and outputs the utterances (S5). After that, the output processing unit 135 outputs the utterances output by the selection unit 134 to the user.

[0048] By performing the above process, the utterance selection device 10 can output utterances from among the utterances shown in the task utterance flow, excluding utterances on topics that are estimated to be disliked by the user.

[0049] [Other Embodiments] Note that, although the preference information in the preference information DB is created by, for example, the analysis unit 132 in the above description, the present invention is not limited to this. For example, the preference information in the preference information DB may be created by conducting a survey or asking questions to users in advance.

[0050] [Experimental Results] An evaluation experiment was conducted on the utterance selection device 10 of this embodiment. The topics covered in this experiment are as follows:

[0051] ・Chat phase: History, work, music, ramen ・Task phase: Temples, studying, live music, Chinese food

[0052] The tokenizer for user input is the one published at the following URL.

[0053] https: / / huggingface.co / cl-tohoku / bert-base-japanese-whole-word-masking

[0054] The BERT model used is the one published at the following URL:

[0055] https: / / huggingface.co / koheiduck / bert-japanese-finetuned-sentiment

[0056] The Word2Vec model used to convert topics into vectors is available at the following URL:

[0057] https: / / aial.shiroyagi.co.jp / 2017 / 02 / japanese-word2vec-model-builder /

[0058] The experimental results will be explained using Figures 5 to 7. In Figures 5 to 7, lines beginning with "System" and a number, such as "System1" and "System2," indicate that the utterance is from the system and which turn it is. Lines beginning with "User" and a number, such as "User1" and "User2," indicate that the utterance is from a user and which turn it is.

[0059] In the "chat phase," the utterance selection device 10 conducted a dialogue with the user as shown in Fig. 5. The underlined parts in Fig. 5 are parts where the user responded positively.

[0060] Fig. 6 is a diagram showing the result of a dialogue with a user using the task-specific utterance flow as is (baseline dialogue). Fig. 7 is a diagram showing the result of a dialogue with a user by the utterance selection device 10 using utterances selected from the task-specific utterance flow. The utterance selection device 10 searched for preference information using similar topics as "history" and "temples," "work" and "study," "ramen" and "Chinese food," and "music" and "live music."

[0061] The baseline shown in Figure 6 includes questions about topics that are presumed to be disliked by the user. On the other hand, the utterances shown in Figure 7 do not include questions about topics that are presumed to be disliked by the user. Therefore, it was confirmed that the utterance selection device 10 makes it possible to collect information about things that the user likes in a shorter number of turns.

[0062] [Application Examples] The utterance selection device 10 can be applied to various situations. For example, consider a case where a user selects one product from multiple products (product A, product B, ...). In this case, by applying the utterance selection device 10, it is possible to output utterances that enable the user to select a product that matches the user's preferences, for example, based on the user's preference information.

[0063] Also, consider a case where the utterance selection device 10 is given an instruction to recommend product A to a user. In this case, by applying the utterance selection device 10, it is possible to output utterances for efficiently collecting information for the user to select product A based on, for example, user preference information.

[0064] For example, consider a case where the utterance selection device 10 has a dialogue with a user for about 10 turns to collect the user's preferences. In this case, the number of turns for each phase is assigned as follows:

[0065] ・Chat phase: 5 turns ・Task phase: 5 turns

[0066] In this case, according to the above experimental results, the utterance selection device 10 can perform three additional turns of dialogue in the task phase by excluding topics to which the user has responded "disliked" from the task utterance flow. As a result, the utterance selection device 10 can conduct a more in-depth dialogue with the user based on the user's preference information collected up to that point. Furthermore, the utterance selection device 10 prevents the user from speaking about topics to which the user has responded negatively, which is expected to improve satisfaction with the dialogue.

[0067] Furthermore, the utterance selection device 10 can reduce the cost of constructing an utterance flow for a user.

[0068] For example, consider a case where the utterance selection device 10 engages in a dialogue with a user to recommend a tourist spot. Note that the topics for recommending a tourist spot vary depending on the tourist spot. Therefore, for example, it is assumed that the utterance selection device 10 actively outputs utterances on topics related to history when recommending Kyoto, and actively outputs utterances on topics related to food when recommending Hokkaido. It is also assumed that the utterance selection device 10 provides fixed topics during the chat phase, independent of the tourist spot being recommended.

[0069] In this case, the dialogue carried out by the utterance selection device 10 follows the flow below: Chat phase: dialogue on a fixed topic is carried out Task phase (dialogue for recommending tourist spots): utterances on topics related to tourist spots and for encouraging user actions are selected

[0070] In this case, for example, if the entire dialogue flow were written according to conditional branching, as in the prior art, the cost of constructing the dialogue would be enormous. However, by applying a method for inferring similar topics, as in the utterance selection device 10, it is possible to realize dialogue for recommending tourist spots with limited construction costs.

[0071] [System Configuration, etc.] The components of each unit shown in the figure are conceptual functional units and do not necessarily have to be physically configured as shown. In other words, the specific form of distribution and integration of each device is not limited to that shown, and all or part of them can be functionally or physically distributed and integrated in any unit depending on various loads, usage conditions, etc. Furthermore, all or any part of the processing functions performed by each device can be realized by a CPU and a program executed by the CPU, or can be realized as hardware using wired logic.

[0072] Furthermore, among the processes described in the above embodiments, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically using a known method.In addition, the information including the processing procedures, control procedures, specific names, various data and parameters shown in the above documents and drawings can be changed as desired unless otherwise specified.

[0073] [Program] The utterance selection device 10 can be implemented by installing a program (utterance selection program) as package software or online software on a desired computer. For example, by executing the program on an information processing device, the information processing device can function as the utterance selection device 10. The information processing device referred to here includes mobile communication terminals such as smartphones, mobile phones, and PHS (Personal Handyphone Systems), as well as terminals such as PDAs (Personal Digital Assistants).

[0074] 8 is a diagram showing an example of a computer that executes an utterance selection program. The computer 1000 includes, for example, a memory 1010 and a CPU 1020. The computer 1000 also includes a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. These components are connected by a bus 1080.

[0075] The memory 1010 includes a read-only memory (ROM) 1011 and a random access memory (RAM) 1012. The ROM 1011 stores a boot program such as a basic input / output system (BIOS). The hard disk drive interface 1030 is connected to a hard disk drive 1090. The disk drive interface 1040 is connected to a disk drive 1100. A removable storage medium such as a magnetic disk or optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to a mouse 1110 and a keyboard 1120, for example. The video adapter 1060 is connected to a display 1130, for example.

[0076] The hard disk drive 1090 stores, for example, an OS 1091, an application program 1092, a program module 1093, and program data 1094. That is, the programs that define the processes executed by the utterance selection device 10 are implemented as program modules 1093 in which computer-executable codes are written. The program modules 1093 are stored, for example, in the hard disk drive 1090. For example, the program modules 1093 for executing processes similar to those of the functional configuration of the utterance selection device 10 are stored in the hard disk drive 1090. The hard disk drive 1090 may be replaced with an SSD (Solid State Drive).

[0077] Data used in the processing of the above-described embodiment is stored as program data 1094, for example, in the memory 1010 or the hard disk drive 1090. The CPU 1020 then reads the program module 1093 or the program data 1094 stored in the memory 1010 or the hard disk drive 1090 into the RAM 1012 as necessary and executes them.

[0078] The program module 1093 and program data 1094 may not necessarily be stored in the hard disk drive 1090, but may also be stored in a removable storage medium and read by the CPU 1020 via the disk drive 1100 or the like. Alternatively, the program module 1093 and program data 1094 may be stored in another computer connected via a network (such as a local area network (LAN) or a wide area network (WAN)). The program module 1093 and program data 1094 may then be read by the CPU 1020 from the other computer via the network interface 1070.

[0079] REFERENCE SIGNS LIST 10 Utterance selection device 11 Input / output unit 12 Storage unit 13 Control unit 131 Dialogue unit 132 Analysis unit 133 Estimation unit 134 Selection unit 135 Output processing unit

Claims

1. A speech selection device comprising: a memory unit that stores preference information indicating a user's level of preference for each topic that may be covered in a conversation with the user; an estimation unit that searches from the memory unit for preference information for topics whose similarity to the topic is greater than or equal to a predetermined value for each topic of utterances shown in a speech flow to the user, and estimates the user's level of preference for each utterance shown in the speech flow based on the user's level of preference for the topics shown in the searched preference information; and a selection unit that excludes and outputs utterances shown in the speech flow for topics whose estimated user's level of preference is less than a predetermined value.

2. The utterance selection device according to claim 1, characterized in that the estimation unit converts the topics appearing in the utterance flow and the topics indicated in the preference information into embedded expressions, respectively, and searches the memory unit for preference information of topics whose similarity to the topic in question is equal to or greater than a predetermined value, using the similarity between the embedded expressions.

3. A speech selection method executed by an utterance selection device, comprising: a step of searching for preference information of a topic having a similarity to the topic shown in an utterance flow to a user from a memory unit that stores preference information indicating the user's preference level for each topic that may be dealt with in a dialogue with the user, and estimating the user's preference level for each utterance shown in the utterance flow based on the user's preference level for the topics indicated in the searched preference information; and a step of excluding and outputting, from the utterances shown in the utterance flow, utterances on topics whose estimated user's preference level is less than a predetermined value.

4. A speech selection program for causing a computer to execute the steps of: searching for preference information of topics having a similarity to a predetermined value or more from a memory unit that stores preference information indicating the user's level of preference for each topic that may be covered in a dialogue with the user, for each topic of utterances shown in a speech flow to a user, and estimating the user's level of preference for each utterance shown in the speech flow based on the user's level of preference for the topics indicated in the searched preference information; and excluding and outputting utterances shown in the speech flow for topics whose estimated user's level of preference is less than a predetermined value.

Images