Information processing system and information processing method

Abstract

[Problem] To make it possible to assist interaction with a user. [Solution] This information processing system is characterized by comprising an analysis unit for analyzing a topic in a session on the basis of communication data during one session, a topic storage unit for storing the topic in association with time information, and a message generation unit for generating a message on the basis of the stored past topic at the start of a new session.

Description

Information processing system and information processing method

[0001] The present invention relates to an information processing system and an information processing method.

[0002] A conversation program using artificial intelligence has been provided (see, for example, Patent Document 1).

[0003] Special Publication No. 2022-531994

[0004] It is difficult to smoothly start a conversation with an unfamiliar user.

[0005] The present invention has been made in view of the above background, and aims to provide a technique that can support dialogue with a user.

[0006] The main invention of the present invention for solving the above problem is an information processing system comprising an analysis unit that analyzes topics within a session based on communication data during the session, a topic memory unit that stores the topics in association with time information, and a message generation unit that generates a message based on the stored past topics at the start of a new session.

[0007] Other problems and solutions disclosed in this application will be made clear in the section on preferred embodiments of the invention and the drawings.

[0008] According to the present invention, it is possible to support dialogue with a user.

[0009] It is a diagram showing an example of the overall configuration of an information processing system. It is a diagram showing an example of the hardware configuration of a management server 2. It is a diagram showing an example of the software configuration of a management server 2. It is a diagram explaining the operation of a management server 2.

[0010] <System Overview> An information processing system according to one embodiment of the present invention will be described below. In the information processing system of this embodiment, it is assumed that an avatar will converse with a user, and the content of the avatar's speech can be generated using, for example, a generation AI. In this embodiment, when the avatar starts a conversation session with a user, the avatar speaks based on topics that have been discussed with the user in the past.

[0011] 1 is a diagram showing an example of the overall configuration of an information processing system. The information processing system of this embodiment is configured to include a management server 2. The management server 2 is communicably connected to a user terminal 1 via a communication network. The communication network is, for example, the Internet, and is constructed using a public telephone line network, a mobile phone line network, a wireless communication path, Ethernet (registered trademark), or the like.

[0012] The user terminal 1 is a computer operated by a user, and may be, for example, a smartphone, a tablet computer, or a personal computer.

[0013] The management server 2 is a computer that analyzes the user's situation and performs processing related to avatars. The management server 2 may be a general-purpose computer such as a workstation or a personal computer, or may be logically realized by cloud computing.

[0014] <Management Server> FIG. 2 is a diagram illustrating an example of the hardware configuration of the management server 2. Note that the illustrated configuration is an example, and other configurations may also be used. The management server 2 includes a CPU 201, memory 202, storage device 203, communication interface 204, input device 205, and output device 206. The storage device 203 stores various data and programs, and is, for example, a hard disk drive, solid state drive, or flash memory. The communication interface 204 is an interface for connecting to a communication network, and is, for example, an adapter for connecting to Ethernet (registered trademark), a modem for connecting to a public telephone network, a wireless communication device for wireless communication, or a USB (Universal Serial Bus) connector or RS232C connector for serial communication. The input device 205 is, for example, a keyboard, mouse, touch panel, button, microphone, or the like for inputting data. The output device 206 is, for example, a display, printer, speaker, or the like for outputting data. Each functional unit of the management server 2 described below is realized by the CPU 201 reading a program stored in the storage device 203 into the memory 202 and executing it, and each storage unit of the management server 2 is realized as part of the storage area provided by the memory 202 and the storage device 203.

[0015] 3 is a diagram illustrating an example of the software configuration of the management server 2. The management server 2 includes a learning model storage unit 231, a topic storage unit 232, a video acquisition unit 211, an analysis unit 212, and a message generation unit 213.

[0016] <Memory Unit> The learning model memory unit 231 stores first and second learning models. The first learning model is for detecting the degree of attribute of a user appearing in a video (hereinafter referred to as attribute degree). The second learning model is for converting audio data into text data. The first and second learning models are created by machine learning. Machine learning can be broadly divided into supervised learning and unsupervised learning. Supervised learning is a method of training a model using input data and corresponding output data (supervisor data). The model parameters are adjusted based on the supervisor data to learn the mapping from input data to output data. On the other hand, unsupervised learning is a method of learning the structure and patterns of input data without using supervisor data, learning the density distribution and feature representation of the input data. Examples of the first and second learning models in this embodiment include neural networks, support vector machines, decision trees, and random forests, which are types of supervised learning. Furthermore, self-organizing maps and k-means methods, which are types of unsupervised learning, can also be used. These machine learning algorithms learn the relationship between input data and output data by weighting and transforming the input data and optimizing the parameters to minimize the error with the output data, thereby learning the mapping from input data to output data.

[0017] The first learning model can be created by machine learning, for example, using features extracted from video and attribute degrees as training data. Features input to the first learning model include, for example, images of the face region extracted from each frame of the video, the positions and sizes of organs such as the eyes, nose, and mouth extracted from the face region, and temporal changes in these organs. On the other hand, attribute degrees output by the first learning model include, for example, the number of blinks, the degree of eye opening, the degree of mouth opening, the position of the eyebrows, the direction of the face, and temporal changes therein. Specifically, the first learning model can input features such as images of the face region and the positions and sizes of organs, and output attribute degrees such as the number of blinks and the degree of eye opening. Attributes can include the number of blinks, eye offset (the angle of the eyes relative to the camera), gaze estimated from the eye offset, and facial expression. Facial expressions include anger, disgust, fear, happiness, sadness, surprise, neutral, negative / positive, etc., and it is possible to infer the number of times these expressions appear in a given period (such as one second), the average or median of their direction, and the degree of emotion (facial expression) such as anger.

[0018] A large-scale dialogue corpus with emotion labels can be used as training data for the machine learning related to the first learning model. For example, text data of various dialogues, such as posts on social media, records of customer support dialogues, and movie lines, can be collected, and emotion labels can be assigned by multiple annotators using crowdsourcing. Each dialogue text is assigned one of predefined emotion categories, such as joy, anger, or sadness. Data representing the degree of each category as a real value between 0 and 1 can also be prepared.

[0019] A model that combines a CNN (convolutional neural network) and an RNN (recurrent neural network) can be used as a machine learning model. A CNN is used to extract local features from text data, while an RNN can be used to perform time series analysis that takes context into account. These networks can be combined in multiple layers, and the final layer can output a probability distribution of emotion categories. Note that other models may also be used.

[0020] The second learning model can be created by machine learning using, for example, features extracted from the speech data itself and / or frequency-converted speech data and text data of the words uttered at that time as training data. A learning model related to general speech recognition can be used as the second learning model.

[0021] The topic storage unit 232 stores topics discussed in a session. A topic may be a word or a sentence. A topic may be data represented as a vector. The topic storage unit 232 stores topics in association with users and time information. The topic storage unit 232 can store users, time information, topics, and emotions among attributes in association with each other. The topic storage unit 232 can store users, time information, topics, emotions among attributes, and the degree of the emotions in association with each other. The topic storage unit 232 can store topics identified by the analysis unit 212 (described later) in association with time information when the conversation took place and information identifying the user. In other words, the topic storage unit 232 can manage the history of topics in past conversations with each user. The topic storage unit 232 can store topic information in the format of, for example, a relational database or a key-value store, and can be configured to enable high-speed search and retrieval as needed.

[0022] <Functional Unit> The video acquisition unit 211 acquires from the user terminal 1 video images obtained by capturing images of the user using a camera provided in the user terminal 1 during an online session. The video images acquired from the user terminal 1 may or may not be configured to be displayed on the screen of the user terminal 1. The video acquisition unit 211 can acquire video images from the user terminal 1, including video images currently being displayed and video images not currently being displayed on the user terminal 1. In this embodiment, the video acquisition unit 211 acquires video images including audio data, but audio data may also be acquired separately from video images from the camera. Audio data can be acquired from the user terminal 1 as audio data obtained by collecting the user's speech using a microphone provided in the user terminal 1.

[0023] The analysis unit 212 analyzes the topic of a session based on communication data during the session. The analysis unit 212 converts the content of utterances contained in the voice data into text data and analyzes the converted text data to identify the topic of the conversation. For example, the analysis unit 212 can input the voice data into a second learning model and segment the text data obtained as the output into words and sentences using language processing techniques such as morphological analysis.

[0024] The analysis unit 212 can identify one or more topics in communication based on the acquired words and sentences. For example, the analysis unit 212 calculates the relevance of the segmented words and sentences to predefined topic categories and identifies the most relevant topic. This topic identification process uses factors such as the frequency of word appearance, co-occurrence, and semantic similarity of sentences. Furthermore, the topic identification process can identify highly relevant topics by calculating the distance between vector data in which words and sentences are embedded and vector data in which predefined topic categories are embedded. The analysis unit 212 may, for example, register a third learning model in the learning model storage unit 231 that identifies topics based on words or vector data in which words are embedded, and the analysis unit 212 may identify topics by providing the acquired words to the third learning model. For example, the analysis unit 212 may provide a prompt to the large-scale language model that includes text data acquired from speech data, words extracted from the text data, and / or vector data in which words are embedded, as well as an instruction to identify a topic, thereby causing the large-scale language model to estimate a topic.

[0025] To extract topics from dialogue text data, the analysis unit 212 can also use an algorithm combining the TF-IDF method and the k-means method. In this case, the dialogue text is first subjected to morphological analysis to extract words, such as nouns and proper nouns, that can be topic candidates. Next, the TF-IDF values ​​of these words are calculated, and words with high values ​​are designated as topic candidates. The TF-IDF value is calculated by multiplying the frequency of a word's appearance in a dialogue by the inverse of its frequency of appearance in all dialogue data. The more characteristic and important a word is to the dialogue, the higher the value. Next, the extracted topic candidate words can be converted into word embedding vectors. A word embedding vector is a vector that represents the meaning of a word as a point in a multidimensional space, with semantically similar words being located close to each other. For example, a word embedding model can be trained in advance using a method such as Word2Vec, and each topic candidate can be vectorized using that model. Finally, the vectorized topic candidate words can be clustered using the k-means method. The k-means method is an algorithm that assumes k cluster centers, assigns each data point to the closest cluster, and updates the cluster centers to minimize the variance within the cluster. This allows semantically similar topic candidate words to be grouped into the same cluster. The label for each cluster can be the word with the highest TF-IDF value among the words in the cluster.

[0026] The analysis unit 212 can analyze the type and / or degree of emotion of the speaker from the user's utterance during a conversation, along with the topic, based on the communication data. The analysis unit 212 can infer the type and / or degree of emotion based on, for example, video images and a first learning model. As types of emotions, for example, six basic emotions, including joy, anger, sadness, fear, surprise, and disgust, can be defined. Furthermore, the degree of emotion can be expressed, for example, by a real number ranging from 0 to 1 for each emotion category. For example, if the degree of "joy" is 0.8, it can be interpreted that the user is feeling a relatively strong emotion of joy.

[0027] Specific emotion analysis techniques can use, for example, both acoustic features extracted from speech data and linguistic features extracted from text data. Examples of acoustic features include pitch, pitch fluctuation, volume, volume fluctuation, and the proportion of high-frequency components in speech. Examples of linguistic features include the frequency of appearance of words expressing emotions, the proportion of positive and negative expressions, and the syntactic structure of sentences. The analysis unit 212 extracts these features and performs machine learning using pre-prepared emotion-labeled training data to construct an emotion classifier. For example, a supervised learning algorithm such as a support vector machine or a neural network can be used to train a model that predicts emotion categories from the features. Alternatively, the predicted probability for each category may be used as the degree of emotion.

[0028] The analysis unit 212 is not limited to analyzing the emotions described above, but can also analyze speaking speed (how fast the speaker speaks), enthusiasm, etc. based on the communication data. It can also detect fever (excitement), the amount of movement of body parts (any part such as eyes, eyelids, fingers, arms, etc.), and the speed of body part movement. All of these detectable feature amounts are considered to be included in the "emotion" in this embodiment.

[0029] To perform emotion analysis that takes context into consideration, the analysis unit 212 may analyze time-series changes in emotions by referring to the dialogue history. For example, emotions can be analyzed for each dialogue segment divided into regular time intervals, and the transition of emotions between segments can be modeled using a Markov model or the like. This allows for a more detailed understanding of dynamic changes in the user's emotional state. The analysis unit 212 can detect, as important time points, points at which emotions change, particularly points at which the trend of time-series changes in emotions changes. For example, points at which the speaking speed suddenly increases or points at which the user smiles can be detected as important time points.

[0030] When starting a new session with a user, the message generation unit 213 generates a message based on past topics stored for the user. The message generation unit 213 can extract topics that correspond to a predetermined period of time from those stored in the topic storage unit 232 and generate a message based on the extracted topic. The message generation unit 213 can generate a message by, for example, providing a large-scale language model with a prompt that includes a topic and an instruction to generate a message to be addressed to the user. The message generation unit 213 can generate sentences using the large-scale language model, for example, using the read topic as a seed. For example, the message generation unit 213 can generate a prompt consisting of words and sentences that represent the topic and input the prompt into the large-scale language model to generate natural sentences that are in line with the topic.

[0031] In this embodiment, an information processing system uses an avatar as an interface when interacting with a user. The avatar is a humanoid character rendered using CG and capable of visually and aurally influencing the user. The avatar's appearance can be generated using 3DCG rendering technology based on predesigned polygon models and textures. For example, multiple avatar models with different genders, age groups, roles, etc. can be prepared, allowing the user to select the most appropriate model based on their attributes and preferences. The avatar's facial expressions, gestures, mouth movements, etc. can be animated in real time in synchronization with the content of the interaction output from the message generation unit 213. The avatar can be rendered on various devices, such as smartphones, tablets, and large displays, depending on the context of the interaction.

[0032] The message generation unit 213 performs processing to have the avatar speak a message (output the message as the avatar's speech). It controls the output of the generated message as the avatar's speech. This makes it possible for the avatar to speak based on past topics, allowing the avatar to start a natural conversation with the user based on past topics, thereby realizing a so-called icebreaker at the start of a session.

[0033] A speech synthesis unit may be provided to convert the text output from the message generation unit 213 into the speech of the avatar. For example, speech synthesis technology using HMM or neural networks can be used to adapt the avatar's tone and voice quality to the user. The avatar can recognize not only the user's speech but also non-verbal information obtained from a camera, such as the user's facial expressions, gaze, and gestures, and behave accordingly. For example, if the user smiles, the avatar can smile back in a similar manner, thereby showing emotional empathy with the user. Furthermore, the avatar can detect the direction of the user's gaze, make eye contact, and enhance the sense of immersion in the conversation.

[0034] The message generation unit 213 can select a topic based on time information and emotion. For example, the message generation unit 213 can select one topic (e.g., randomly) from a predetermined number of topics associated with positive emotions (including happiness, smiles, etc.) starting from the most recent topic. The message generation unit 213 can select a topic based on time information, emotion, and the degree of emotion. For example, the message generation unit 213 can sort topics associated with positive emotions (including happiness, smiles, etc.) in descending order of emotional degree and then in descending order of time information, and select one topic (e.g., randomly) from a predetermined number of topics starting from the top.

[0035] When a new dialogue session with a user is started, the message generation unit 213 can read past topics for that user stored in the topic storage unit 232 and generate a message related to that topic.

[0036] The message generator 213 can use the following indicators as criteria for selecting topics: (1) Temporal proximity: A higher score is assigned to a topic that appeared in a more recent conversation. (2) Topic relevance: A higher score is assigned to a topic that is semantically closer to the content of the current conversation. (3) Topic importance: A higher score is assigned to a topic that appeared repeatedly in past conversations or a topic that the user was emotionally excited about. The message generator 213 calculates a score for each of the retrieved past topics based on the above three indicators, selects the topic with the largest weighted sum of these scores, and generates a message related to that topic. Regarding the importance of a topic, the message generator 213 can, for example, evaluate topics that occurred near (e.g., within a predetermined time before or after) an important point analyzed by the analysis unit 212 as important topics, and select topics based on their importance. However, the message generator 213 can select topics based on their importance and the number of times the topic appeared during the session (e.g., based on a score calculated from the importance and the frequency of the topic's appearance) to prevent topics from being repeated every day.

[0037] The analysis unit 212 then analyzes what the user has said and obtains it as text data, and the message generation unit 213 generates and outputs a message in response to the text data (having the avatar speak), thereby enabling communication between the user and the avatar.

[0038] <Operation> FIG. 4 is a diagram illustrating the operation of the management server 2.

[0039] When the management server 2 starts a session (S301), it retrieves past topics corresponding to the other user from the topic memory unit 232 (S302), generates a message based on the retrieved topic (S303), and has the avatar speak the generated message (S304).

[0040] As described above, according to the information processing system of this embodiment, it is possible to break the ice at the start of a session based on past topics.

[0041] Although the present embodiment has been described above, the above embodiment is intended to facilitate understanding of the present invention and is not intended to limit the present invention. The present invention may be modified or improved without departing from the spirit thereof, and equivalents thereof are also included in the present invention.

[0042] For example, the processing by each of the functional units of the management server 2 described above may be performed by any of the functional units. Also, a different functional unit that performs part of the processing by each of the functional units described above may be added. Also, the functional units of the management server 2 may be distributed across multiple computers.

[0043] Furthermore, the information stored in each storage unit of the management server 2 may be stored in any of the storage units. That is, the information stored in the above-mentioned multiple storage units may be stored in one storage unit, or part of the information stored in one of the above-mentioned storage units may be stored in another storage unit.

[0044] <Modification 1>

[0045] In the first modification, the user's attributes are estimated based on not only the user's emotions but also their age, gender, speaking style, etc., and topics are selected and messages are generated accordingly.

[0046] The analysis unit 212 can estimate the user's age and gender based on acoustic features extracted from the voice data. For example, features that vary depending on age and gender, such as voice pitch, speaking rate, and voice quality, can be input into a machine learning model, and the user's age and gender category can be obtained as the output. Voice data from a large number of speakers and their age and gender labels can be prepared as training data, and a classifier such as an SVM or neural network can be trained.

[0047] The analysis unit 212 analyzes the characteristics of a user's speaking style based on linguistic features extracted from text data. For example, it can define linguistic indicators that reflect the user's attributes, such as vocabulary richness, frequency of use of polite language and abbreviations, and type of first-person pronoun, and record these statistics as a user profile.

[0048] The message generation unit 213 can select a topic based on the user's age, sex, speaking style characteristics, etc. estimated by the analysis unit 212. For example, for a young male user, topics that are popular among men of the same age, such as sports and hobbies, can be preferentially selected, and for a middle-aged or elderly female user, topics that are of high interest to that attribute, such as health and cooking, can be presented.

[0049] The message generator 213 can also take into consideration the attributes of the user when generating a message on the selected topic. For example, for younger users, the message generator 213 can generate a message in a friendly tone, using casual language and a lot of emoticons, while for middle-aged and older users, the message generator 213 can generate a message in a more subdued tone, using polite language.

[0050] The message generator 213 can mimic the characteristics of the user's speaking style to enable more natural communication. For example, it can actively incorporate vocabulary and expressions that the user frequently uses into messages, or adjust the endings of words to match the user's tone of voice.

[0051] <Modification 2>

[0052] In the second modification, not only a conversation with one user but also a group conversation with multiple users is supported.

[0053] The analysis unit 212 can individually analyze the comments of each user participating in a group conversation and extract each user's emotions and topics. For example, the analysis unit 212 can analyze each user's voice data and text data using the same method as in the above embodiment and assign emotion labels and topic labels to each user. This makes it possible to understand the emotional state of each user and the topics they are talking about in a group conversation.

[0054] The analysis unit 212 can analyze the relationships between users' comments. For example, it can determine how other users respond to a comment made by a certain user based on the time interval between responses and the similarity of the content of the responses. This can discover patterns of interaction between users in a group conversation.

[0055] The message generation unit 213 integrates the emotions and topics of each user and the information on interactions between users obtained by the analysis unit 212, and can provide appropriate topics and generate messages while taking into consideration the flow of conversation throughout the group. For example, if other users show interest in a topic presented by a certain user, the message generation unit 213 can stimulate conversation throughout the group by providing information related to the topic or asking questions that delve deeper into the topic. On the other hand, if there is a difference of opinion between users on a certain topic, the message generation unit 213 can encourage constructive discussion by clarifying the points of contention or adding commentary from a third-party perspective.

[0056] The message generation unit 213 can generate a message that deepens mutual understanding between users based on the relationship between the statements made by the users. For example, if another user sympathizes with a statement made by a certain user, the message generation unit 213 can narrow the psychological distance between the users by expressing that sympathy in words. Conversely, if there is a difference of opinion between the users, the message generation unit 213 can resolve the conflict by sending a message that acknowledges the differences in each user's position and seeks a compromise.

[0057] <Modification 3>

[0058] In the third modification, not only past dialogue history but also external information such as the latest news and trends is incorporated, thereby expanding the options for topics to be used in dialogue with the user.

[0059] The management server 2 according to the third modification includes an external information acquisition unit. The external information acquisition unit can periodically collect information about the latest events and trends from news sites, social media, blogs, and the like on the Internet. For example, by using technologies such as RSS feeds, APIs, and web scraping, it can automatically acquire updated information from each site and store the titles, summaries, keywords, and the like in a database.

[0060] The analysis unit 212 can analyze the information collected by the external information acquisition unit and extract topic candidates. For example, as in the above embodiment, trending topics can be discovered by identifying important keywords mentioned in many articles using the TF-IDF method or the like and clustering them.

[0061] The analysis unit 212 analyzes the history of past interactions with the user to estimate the user's interests. For example, it can extract keywords frequently used by the user or topics about which the user has expressed strong feelings, and identify common topics as the user's interests.

[0062] The message generation unit 213 can select topics that match the user's interests from the latest topics extracted by the analysis unit 212 and use them in the dialogue. For example, if it is estimated that the user is interested in sports, the message generation unit 213 can suggest the latest sports news or event information as a topic. Furthermore, if the user is interested in politics or economics, the message generation unit 213 can mention the latest political events or trends in economic indicators.

[0063] The message generator 213 can also suggest new topics to the user based on external information. For example, if an event occurs that is not directly related to the user's interests but is attracting significant social attention, the message generator 213 can broaden the scope of the conversation by asking the user for their opinion on the topic or by providing related background information.

[0064] <Modification 4>

[0065] In variant 4, the system is expanded to a multimodal dialogue system that includes not only text data but also various types of content such as images and videos.

[0066] In a multimodal dialogue system, the management server 2 may further include an image recognition unit and a video analysis unit. The image recognition unit analyzes still images that appear during a dialogue and extracts topics and emotions from their content. For example, it can use a deep learning model such as a convolutional neural network (CNN) to recognize and label objects, scenes, text, and the like that appear in the images. As a result, for example, if a user shares a photo of a travel destination, the image recognition unit 270 can identify tourist attractions and activities that appear in the photo and present them as topic candidates.

[0067] The video analysis unit analyzes video content appearing during a conversation and extracts topics and emotions from the content. For example, the video is divided into frames at regular intervals, and object recognition similar to that performed by the image recognition unit is performed on each frame, while object movement and changes are detected from the differences between frames. Similarly to the above-described embodiment, audio data included in the video is also recognized and converted into text. For example, if a user shares a scene from a favorite movie, the video analysis unit can analyze the characters, background, and dialogue appearing in that scene and suggest related topics.

[0068] The analysis unit 212 can integrate the information extracted by the image recognition unit and the video analysis unit with text data to analyze topics and emotions from a more multifaceted perspective. For example, if a user shares text such as "It was a fun trip" and photos of the travel destination, the analysis unit 212 can extract the emotion "fun" from the text and specific activities at the travel destination from the photos, and combine them to generate a specific topic such as "It was fun experiencing □□ in XX."

[0069] The message generation unit 213 can generate messages related to images and videos based on the topics generated by the analysis unit 212. For example, with respect to a dish appearing in a photo shared by a user, the message generation unit 213 can generate a message introducing the origin and characteristics of the dish, recommended restaurants, etc. Furthermore, with respect to a video shared by a user, the message generation unit 213 can return more specific questions and comments tailored to the content of the video.

[0070] The management server 2 may include an image generation unit to enhance the expressiveness of the avatar. The image generation unit generates an appropriate image in real time in accordance with the message generated by the message generation unit 213. For example, a deep learning model such as a generative adversarial network (GAN) can be used to automatically generate an image corresponding to text. As a result, for example, in response to a message such as "It makes me feel like I'm relaxing at the beach," the image generation unit 290 generates a relaxing image that combines a blue ocean, white sand, palm trees, etc., and presents it to the user along with the message.

[0071] <Modification 5>

[0072] In the fifth modification, a personalized dialogue is carried out by analyzing the dialogue history with the user and learning the user's preferences and habits.

[0073] The management server 2 includes a user profile learning unit. The user profile learning unit accumulates the dialogue history of each user and extracts the characteristics of each user from it. It can analyze topics that users frequently mention, topics that elicit positive responses, and topics that elicit negative responses, and store these in the topic storage unit 232. It can also extract frequently used expressions, phrases, and vocabulary tendencies as characteristics of the user's speaking style, and store these in the speaker information database.

[0074] The message generation unit 213 refers to the user characteristics learned by the user profile learning unit, selects topics suitable for the user, and generates messages. For example, the message generation unit 213 can acquire topics that the user is likely to be interested in and topics that the user should avoid from the topic storage unit 232, and determine the direction of the conversation based on that information. For example, if it is found that a certain user is very interested in history-related topics but tends to have a negative reaction to political topics, the message generation unit 213 can make adjustments such as actively suggesting history-related topics while sparing political topics.

[0075] Furthermore, the message generator 213 can acquire the characteristics of the user's speaking style from the speaker information database and generate messages that imitate them. For example, if it is found that a certain user frequently uses phrases such as "That's right," or "That is," the message generator 213 generates messages that match the user's speaking style, such as "The weather is sunny today, isn't it?" or "You said you went on a trip, where did you go?" This allows the user to communicate with the dialogue agent as if they were talking with a close friend.

[0076] The message generation unit 213 can also estimate changes in the user's interests and life events from past dialogue history and adjust the content of the dialogue accordingly. For example, if a user talks about recently getting married, the message generation unit 213 will weave topics related to marriage into subsequent messages, such as "How is your newlywed life?" and "Do you have any memories of your wedding?". In addition, at the change of seasons, the message generation unit 213 can suggest seasonal topics, such as "It's spring now and it's getting warmer," and "Do you have any plans for summer vacation?"

[0077] The user profile learning unit periodically analyzes the user's dialogue history to detect changes in the user's interests and habits. When a change is detected, the content of the topic storage unit 232 and the speaker information database is updated, so that dialogue can always be based on the latest user profile.

[0078] <Disclosures> The present disclosure also includes the following configurations. [Item 1] An information processing system comprising: an analysis unit that analyzes topics within a session based on communication data during the session; a topic memory unit that stores the topics in association with time information; and a message generation unit that generates a message based on the stored past topics at the start of a new session. [Item 2] The information processing system according to Item 1, wherein the message generation unit extracts topics that correspond to topics within a predetermined period of time in the past from the topics stored in the topic memory unit, and generates the message based on the extracted topic. [Item 3] The information processing system according to Item 1, wherein the message generation unit performs processing to have an avatar speak the message, and the message generation unit causes the avatar to speak based on the past topics. [Item 4] The information processing system according to item 1, wherein the analysis unit analyzes the speaker's emotion along with the topic based on the communication data, the topic memory unit stores the time information, the topic, and the emotion in association with each other, and the message generation unit selects the topic according to the time information and the emotion. [Item 5] The information processing system according to item 1, wherein the analysis unit analyzes the speaker's emotion and the degree of the emotion along with the topic based on the communication data, the topic memory unit stores the time information, the topic, the emotion, and the degree of the emotion in association with each other, and the message generation unit selects the topic according to the time information, the emotion, and the degree of the emotion. [Item 6] An information processing method according to item 1, wherein a computer analyzes the topics within a session based on communication data during the session, stores the topics in association with time information, and generates a message at the start of a new session based on the stored past topics.

[0079] 1 User terminal 2 Management server

Claims

1. An information processing system comprising: an analysis unit that analyzes topics within a session based on communication data during the session; a topic storage unit that stores the topics in association with time information; and a message generation unit that generates a message based on the stored past topics when a new session starts.

2. An information processing system according to claim 1, wherein the message generation unit extracts topics stored in the topic storage unit that correspond to topics within a predetermined period of time in the past, and generates the message based on the extracted topics.

3. An information processing system according to claim 1, wherein the message generation unit performs processing to have an avatar speak the message, and the message generation unit causes the avatar to speak based on the past topics.

4. An information processing system according to claim 1, wherein the analysis unit analyzes the speaker's emotions along with the topic based on the communication data; the topic memory unit stores the time information, the topic, and the emotion in association with each other; and the message generation unit selects the topic according to the time information and the emotion.

5. An information processing system according to claim 1, wherein the analysis unit analyzes the speaker's emotions and the degree of the emotions along with the topic based on the communication data; the topic memory unit stores the time information, the topic, the emotions, and the degree of the emotions in association with each other; and the message generation unit selects the topic according to the time information, the emotions, and the degree of the emotions.

6. An information processing method characterized by the computer analyzing the topics within a session based on communication data during the session, storing the topics in association with time information, and generating a message at the start of a new session based on the stored past topics.

Images