In human-human conversation, people frequently ask questions about a person with whom to talk. Since such questions also asked in human-agent conversations, previous research developed a Person DataBase (PDB), which consists of questionanswer pairs evoked by a pre-defined persona to answer user's questions. PDB contains static information including name, favorites, and experiences. Therefore, PDB cannot answer questions about events that occurred after it was built. It means that this approach does not focus on answering questions about more recent things (recent personal questions), e.g., Have you seen any movies lately? In contrast, since recent questions are frequently asked in a casual conversation, conversational agents are required to answer recent questions for maintaining a conversation. In this paper, we collect event data that consist of a large number of experiences and behaviors in daily lives, which enables to answer recent questions. We analyze them and show that our data is effective for answering recent questions.
Questions about a conversational partner are called “personal question,” which are an essential factor for expressing interest in conversational partners. Such questions frequently occur in casual human-human conversations. Nishimura et al. showed that such questions occurred in both human-human and human-agent conversations [Nisimura et al., 2003]. Adequately answering them is an essential factor in the development of conversational agents [Sugiyama et al., 2017].
To answer personal questions, previous works developed Person DataBase (PDB), which consists of question-answer pairs evoked by a pre-defined persona [Batacharia et al., 1999; Sugiyama et al., 2014]. Although their approach covers a wide variety of personal questions, developing a highquality PDB is too expensive. The cost problem makes it difficult to update constantly; consequently, PDB usually contains only static information that rarely changes over time. Therefore, conversational agents using PDB cannot answer questions about recent events such as What did you have for dinner yesterday?. Also, it is easy to imagine that immutable responses to recent personal questions make conversational agents unnatural; therefore, conversational agents have to spend different days that like people spend different days, and it is more natural to return different answers to recent personal questions. To solve this problem, preparing other kinds of data which expresses recent experiences helps conversational agents to answer such questions about recent things (recent personal questions).
One simple idea is to collect data that express such recent events as a diary that is updated by the user. Previous work on response generation leveraged diaries or microblogs as a corpus that includes people’s recent personal information [Li et al., 2016]. Even though this approach seems reasonable, handcrafted-data-driven approach such as PDB has practical advantages in controllability and reliability. In this paper, we collected event data from participants who take part in shortand long-term periods. This collected data is hand-crafted, high-quality and easy to update (adding new day’s data). We clarified the potential of event data to answer questions about recent behaviors/experiences in casual conversations through analysis. 2
As mentioned in the introduction, PDB is the most closely related research to answer user questions. Batacharia et al. developed PDB about Catherine, a 26-year-old female living in New York City [Batacharia et al., 1999]. To cover more questions and with different personas, Sugiyama et al., developed a PDB with six personalities such as a 20-year-old female, a 50-year-old male, and robots [Sugiyama et al., 2014]. Both PDBs contain only static information; therefore, they cannot answer recent questions. If we want to answer recent questions by PDB, we have to update PDB’s contents constantly; however, updating PDBs constantly causes too expensive costs. The difficulty of updating PDB is the relationship of questions and answers (QA); for example, when the content of a base QA changes (e.g., Question:Do you have any pets?, Answer:Yes, I have a dog. change to new Answer:No, I don’t have.), related contents of QAs should be changed depending on a changed content of base QA (e.g., Question:Do you have a dog?, Answer:Yes, I have a dog. should be changed to new Answer:No, I don’t have.). PDB has many complicated relations of QAs, it makes updating PDB difficult and expensive. A PDB’s merit, which is found in handcrafted-data-driven methods, is the ability to generate answers based on facts and consistency from the data. Such handcrafted-data-driven approaches answer questions with consistent replies and without a lie. The consistency of the responses based on facts has the potential for improving the performance of conversational agents.
Although there are many studies on conversational agent’s response generation [Ritter et al., 2011; Inaba and Takahashi, 2016], few studies focus on the consistency depending on an agent’s personality. Persona-based conversation models treat personality as speaker-embedding to increase the sentence quality [Li et al., 2016]. This model is the state-of-theart model to generate conversational agent’s responses using an embedding vector that expresses agent’s personality. This approach has potential to answer recent personal questions; however, it indicates two critical problems. One is that this approach cannot promise to answer without lies; this problem is strongly related with research of PDB. Hand-crafted database approach such as PDB can answer responses that reflected a right personality unless it gets wrong matching of questions. In contrast, neural network based approaches often answer questions with response sentences that do not exist in training data; since such models are optimized only for maximizing the naturalness of response sentences. Even though this approach has the potential to answer recent personal questions, it can offer no guarantee that the answers exist in training data.
Another problem is that this model does not consider the past consistency depended on the day, time, and past events. When we asked a question such as What did you eat last night? to conversational agents, this approach always replies the same response such as I ate ramen. This QA pair is natural when we check only this one pair; however, eating the same food every dinner is too unnatural in the daily life of conversational agents. Therefore, to establish a long-term conversation with human-agent and to make conversational agents more natural, we must solve this invariance problem of responses. Using information of date and time to train speakerembedding vector, it may help to solve this problem. However, we can imagine easily that this model requires much training data which is insufficient with the amount we have now.
In this paper, we created event data for answering recent personal questions in casual conversations. This approach by the created event data is identical with PDB as the handcrafted-data-driven approach and is essential to verify answering based on facts, and we use it as the first step to develop a function that answers questions about recent experiences based on facts. 3
To answer recent personal questions that ask about recent experiences and behaviors, we collect the consistent data from humans as events that express experiences and behaviors. This event data has to be collected from participants with lowcosts; because we need to update it constantly. Besides, we have to collect data from various participants because we do not know what kind of persona influences events.
We recruited 62 Japanese-speaking participants of roughly equal numbers of both genders whose ages ranged from 10s to 60s and collected daily experiences and behaviors as event data. They wrote down 20 events every day and at least two events every four hours. We collect event name, reasons, time, and impressions for each event; because these aspects are asked in casual conversation frequently. Participants were indicated not to write any descriptions including privacy. Such diary-like method to write down like a diary is low-cost compare than the PDB’s collecting method. Specifically, we prepare an Excel file and ask participants to write four aspects such as name, reasons, time, and impressions, for each column. The format of this Excel file is simple; one line is for one event, one sheet is for one day, one file is for one participant.
An event includes four aspects: 1. Event name: What is happened? What did you do? 2. Event reasons: Why did it happen? What did you do? 3. Event time: Selected from the following four-hour time blocks: 0:00-4:00, 4:00-8:00, 8:00-12:00, 12:00-16:00, 16:00-20:00, or 20:00-24:00.
4. Event impressions: How did you feel? For the aspects of reasons and impressions, participants can write more than one sentence with a space between phrases. We define two groups for collecting data. One is the longterm group which takes data with many days from a few participants, and this facilitates the comparison between participants. Another one is the short-term group which takes data with few days from a lot of participants; this is necessary to collect various event data. Five participants wrote 20 events per day for 30 days (long-term group), and 57 participants wrote 20 events per day for seven days (short-term group); finally, we collected a total of 10,980 events. Table 1 shows examples of events collected from a participant who belongs to the short-term group. The example shows that we obtain a variety of events even if the only one participant wrote. 4
We analyze next two viewpoints to show that our collected data helps to answer recent personal questions that related to personality and date. First, the tendency of events was varying among participants; it shows that we have to reflect participant’s characteristics to answer recent personal questions. Second, the tendency of events was varying according to a day of the week; it shows that we have to reflect a day of the week and update event data constantly.
To analyze the tendency of events, we categorized the collected event data since they have slightly different event names, with which we cannot count the occurrence of each event. For example, we wish to handle two events such as Went to school’ and Went to high school as the same event. To collect such similar events as the same event, we perform the word-based hierarchical clustering using word2vec that trained from Wikipedia data.
Next, we highlight the difference between event’s tendencies among participants and days. We calculate frequency distributions of events for each participant and each day, and compare a JS divergence of these frequency distributions. This comparison clarifies two relationships of event tendencies: Distributions of event frequency depend on each participant and Participants have different distributions of event frequency depending on each day. 4.1
We performed hierarchical clustering to find similar events in the collected data [Larsen and Aone, 1999]. This clustering is both analysis and a necessary procedure to compare events among participants or days by collecting clusters. Before clustering, we trained word2vec [Mikolov et al., 2013] from Wikipedia articles; word2vec is useful to convert event names to a word embedding. In clustering, we tokenize event names by mecab [Kudo, 2006], and restore tokenized words to original forms. Next, we calculate vectors by adding together word2vec of each tokenized words, and cluster these calculated vectors with Ward’s method [Szekely and Rizzo, 2005]. Figure 1 shows a dendrogram and a heatmap of each vector. To confirm the difference of clustering results by the number of clusters, we respectively show the hierarchical clustering results of ten clusters and 30 clusters. Table 2 and Table 3 are ten and 30 lists of events. Event names are the nearest event to the center of each collected cluster, and cluster sizes are numbers of events included in each cluster.
From Table 2, we obtained common events which seem to happen to anyone such as Got up, Took a meal, Drank drink, Took a bath, Ate lunch and more. In contrast, from Table 3, we obtained detailed events which seem to happen to specific personas such as Look SNS by PC, Played a game, Watched a video and more. Such events which indicates participant’s characteristic are important to highlight the difference between participants; therefore we use the 30 lists of events as clustering result to compare events between participants and days in following analysis sections.
Note that we defined size of clusters based on a few preliminary experiments. Proposing the clustering method that determines a size of clusters based on clusters variances or entropy has a potential to improve clustering performance; therefore, we will tackle defining a better model to handle event data in future work.
From Figure 1, we can find a few particularly bright clusters that include very similar events. In contrast, some clusters with less brightness include various events that are not so similar.
Since this hierarchical clustering successfully makes clusters, we can benefit by using clustering results for data analyses. This clustering method that considers word meaning as word2vec, could make clusters which gathered almost same meaning events. 4.2
First, we analyzed events among participants. To highlight the differences between participants, we calculate the distribution on clusters of every participant. To calculate it, we used the 30 clusters in Table 3. We compare these cluster distributions between each participants using JS divergence.
The averaged JS divergence of every participant was 0.39. The minimum JS divergence is 0.063, and the maximum JS divergence is 0.77, these scores were found among participants who are the short-term group. Averaged JS divergence is not close to 0; it means that distributions of event frequency
E1 E2 E3 E4 E5 E6 E7 E8 E9 E10 E11 E12 E13 E14 E15 E16 E17 E18 E19 E20 E21 E22 E23 E24 E25 E26 E27 E28 E29 E30
To analyze details of event tendencies, we show counts of event cluster assignment in each participant who is a longterm group, in Table 4. Most participants have different distributions of events, but E4, E6, E7, E9( and E5), E13, E14, E15, E16, E20, E22, E23, E24, E26, and E27 occurred more than once in all participants of the long-term group. E4, E6, E16, E22, and E24 are clusters containing mainly housework such as washing, cleaning, cooking, shopping and more. E7, E13, E14, E15, and E20 are clusters that indicate physiological desires such as eating, drinking, sleeping, taking a bath and more. E9 (and E5) is the cluster that indicates mainly working, E23 indicates reading, and E26 indicate talking. The last E27 is a cluster included various events such as child rearing, one’s hobby, and school life. Such basic events that are related to living were observed in almost participants. In contrast, we obtained that events which relate to entertainment such as Played a game were observed in a specific participant such as P3.
These results show that participants have different event tendencies. This indicates that we should collect data which depends on each persona to answer recent personal questions. Second, we analyzed events among days. Like Section 4.2, we counted the clusters of every participant with 30 clusters in Table 3. We compare cluster counts of all participants in a total at a day of the week. We show JS divergences between days of the week in Table 5.
We focus on JS divergences between weekdays and weekends. These high JS divergences (We showed it as bold in Table 5) show the difference between weekdays and weekends; furthermore, the small JS divergence scores are concentrated in between weekdays and between weekends. This result shows that participants spend different life between weekdays and weekends. Such result that we can imagine easily lets us reconfirm the importance to answer depending on a day. Therefore, we need data which depend on each day to answer questions that ask about events. 5
In this section, we discuss the potential to answer recent personal questions by our collected data. Our discussion follows the “comparison with the conversation corpus” in Sugiyama et al. [Sugiyama et al., 2014], whose PDB covers 41.3% of questions in real conversations and explains why other questions were excluded. The top reason that questions were excluded is “limited by specific words, date, or time” such as What did you eat for lunch today? or Where did you go this summer vacation?, such questions are about 71.2% of a whole of excluded questions. We mainly focus on these excluded questions and show case studies which can answer by our collected data. Tackling to answer such questions helps to solve future works of the previous research.
First of all, we collect 286 questions that are the same as excluded questions by Sugiyama et al. [Sugiyama et al., 2014], and extract 204 questions that were excluded by “limited by specific words, date, or time.” In previous works, they said that these questions are difficult to maintain consistency with 5W1H answers in particular. We focus on these questions and find questions that can answer questions if we use event data. We show examples of such question which can answer based on an event in Table 6.
From Table 6, some questions which ask about speaker’s recent behaviors can answer by our collected data. For example, we can answer a question such as What did you eat for lunch today?, an answer is Yes, I ate a curry and rice by using an event such as Ate a curry and rice. In this manner, we can make an answer utterance that based on an event matched with a question. These results show us the possibility to answer a part of questions that were unsolved future works of PDB with our collected data.
We can also answer questions that ask opinions. Such questions frequently occurred after disclosure or an answer that replied to first questions. To answer with opinions, we use aspects of event impressions. We show examples of questions which ask opinions about events in Table 7. Specifically, when a conversational agent say I watched a movie. as disclosure, and the user asks Do you like it? that asks conversational agent’s opinion, a conversational agent can answer It is fun. by using an aspect of event impression from our collected data. In question-answering based on the conventional PDB, we cannot handle such kind of questions which continued the same topic as the previous turn. Answering questions about the details of the same one event, it shows the potential which improves the question-answering function to talk deeper.
However, we obtain some questions that we could not answer by our collected data; there are Questions that ask about agent’s past custom and Questions that ask about agent’s future. To answer questions that ask about agent’s future, we have to prepare the other data such as plans made by agents. These plans may need the approach such as the belief-desireintention model that is different from our event data. To answer questions that ask about agent’s past custom, we need data which indicates habitual events and experiences. Such data seem closely related to our event data, because habitual events and experiences may be made by the accumulation of recent events. We clarify the relationship between past custom and events, and will propose a method that generates past custom based on accumulated recent events in future work.
From analyses and case studies, we showed the potential of answering for recent personal questions that cannot be answered by the previous PDB. Our collected data helps to answer not only asking events but also asking opinions. However, we obtain some problems that remain about questions which ask about past custom and future. In future work, we tackle to answer questions that ask past custom such as habitual events using our event data. Furthermore, clarifying volumes and frequency to collect enough event data; these are How many events do we need in one day?, How many times do we ask to write per one day?, and How many days do we ask to write events?. Besides the data collection, to develop conversational agents that answer recent personal questions using event data, we have to propose a method that finds events which match with user’s recent personal questions. 6
In this paper, we collect 10,980 events which express recent experiences and behaviors to help conversational agents answer questions about recent experiences. First of all, we analyze collected data to highlight the tendencies of events based on each participant and each weekday, and we show the necessity of our event data that make conversational agents more natural. Our analysis shows that event data reflect participant’s characteristics and dependencies on weekdays, and we show two knowledge about tendencies of events. One, event tendencies are depending on each participant; we Question Did you eat for lunch today? What did you eat for lunch today? Did you play video games lately? What did you play video games lately? What kind of games did you play lately? Did you watch a movie? Where did you go out somewhere recently? Where did you go out somewhere recently? Answer Yes, I ate.
I ate a curry and rice.
Yes, I played video games.
I played smartphone games.
Smartphone games.
Yes, I watch a movie.
I went to the nearby French restaurant.
I went to the spa.
Event Ate a lunch Ate a curry and rice Played video games Playing smartphone games Playing smartphone games Watched a movie Went to the neighboring French restaurant Going to the spa should collect event data which depends on each conversational agent’s persona. Another one, event tendencies are depending on each weekday; we should collect event data which depends on each day to make conversational agent’s answers more natural.
In the discussion, we followed the previous works and obtained case studies that can answer by our collected event data. Our event data helps to answer recent personal questions such as What did you eat for lunch today? that asks about doing conversational agent’s events; therefore, results show potential to achieve our first purpose that answers a part of questions that cannot be answered by the previous PDB. Furthermore, aspects of event impressions help to answer questions that ask opinions such as Do you like it?. This continuous question-answering shows the potential which improves the question-answering function to talk deeper.
In future work, we clarify volume and frequency to collect enough event data, and develop conversational agents that answer recent personal questions by collected event data. ACM SIGKDD international conference on Knowledge discovery and data mining, pages 16–22. ACM, 1999.