Q20: Rinna Riddles Your Mind by Asking 20 Questions
暂无分享,去创建一个
In this paper, we introduce a novel puzzle riddling framework for an emotional chatbot, Rinna (Wu et al. 2016). First, the user is asked to think one topic word or a role (or, named entities, such as a famous personal name) and that topic word is unknown to Rinna. Then, Rinna tries to guess (or, mind-reading) user’s topic word by asking a sequence of questions, such as “Is your role a male or a female?”. Based on user’s responses such as “yes”, “no”, or “not sure”, the candidate roles are ranked from the Rinna side. Sequentially, the selection of the next question is based on user’s answer to the former question and the existing prior ranking of the questions as well as mutex relations among questions. With no more than 20 questions, Rinna is supposed to tackle user’s mind by delivering the correct answer. If Rinna’s answer is confirmed by the user, then that gaming session is recorded for a future usage of data mining (such as updating the reference answers of one role’s questions, in case that the candidate role’s questions are answered differently with the reference answer). On the other hand, if Rinna could not provide a correct answer (i.e., all Rinna’s answers were wrong), then Rinna asks the user to provide the correct answer and that session will be recorded as well for (1) new role appending and (2) existing role updating to Rinna. We argue that this Q20 gaming benefits the following applications: (1) user profile construction, means that user’s answer in mind reflects his/her interested topic, and (2) product recommendation, means that the roles can be replaced from famous person’s names to partner companies’ product names that allow Rinna to guess by following a similar Q20 workflow. It is not trivial for constructing such a Q20 framework, due to the following difficulties: (1) a database of roles, or topic words of various domains (such as persons, foods, animals, places) should be constructed from Rinna’s side; (2) respectively for each of the roles and topic words, a list of candidate questions together with their reference answers should be collected and constructed in a pre-determined form; (3) use users’ historical answer recoding to update the reference answers to target roles’ (such as the answer of one famous
[1] Hugo Zaragoza,et al. The Probabilistic Relevance Framework: BM25 and Beyond , 2009, Found. Trends Inf. Retr..
[2] Jeffrey Dean,et al. Distributed Representations of Words and Phrases and their Compositionality , 2013, NIPS.
[3] Richard S. Sutton,et al. Reinforcement Learning: An Introduction , 1998, IEEE Trans. Neural Networks.
[4] Nello Cristianini,et al. Classification using String Kernels , 2000 .