Probabilistic Prototype Model for Serendipitous Property Mining

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 663–673, Osaka, Japan, December 11-17 2016. Probabilistic Prototype Model for Serendipitous Property Mining Taesung Lee∗ IBM T. J. Watson Research Center taesung.lee@ibm.com Seung-won Hwang Yonsei University seungwonh@yonsei.ac.kr Zhongyuan Wang† Microsoft Research wzhy@outlook.com Abstract Besides providing the relevant information, amusing users has been an important role of the web. Many web sites provide serendipitous (unexpected but relevant) information to draw user traffic. In this paper, we study the representative scenario of mining an amusing quiz. An existing approach leverages a knowledge base to mine an unexpected property then find quiz questions on such property, based on prototype theory in cognitive science. However, existing deterministic model is vulnerable to noise in the knowledge base. Therefore, we instead propose to leverage probabilistic approach to build a prototype that can overcome noise. Our extensive empirical study shows that our approach not only significantly outperforms baselines by 0.06 in accuracy, and 0.11 in serendipity but also shows higher relevance than the traditional relevance-pursuing baseline using TF-IDF. 1 Introduction Unlike the traditional purpose of the web providing relevant information or answers to user questions, conversely, recent web services ask users unexpected trivia questions to amuse them. Bing provides a set of interesting quizzes with an image of the day on the front page. Figure 1 describes an amusing quiz question generation on a long ‘neck’ of giraffe, which we use as a motivating scenario throughout this paper. Inspired by Figure 1, we study the problem of finding a “serendipitous” property a such as ‘neck’ for any given entity e. Table 1 categorizes existing automatic quiz generation efforts pursuing relevance and unexpectedness, respectively. Inspired by Jeopardy!, Seyler et al. (2015) focus on relevance to the domain and a certain difficulty level. Inspired by Bing questions, Lee et al. (2016) seek unexpected entity-property pair (e, a). In clear contrast, we complement the solution space by pursuing the intersection of finding unexpected but still relevant properties, which is often named as serendipitous (or Case B). As the existing determin- istic model (DM, Lee et al. (2016)) fails to distinguish Case B and C, we propose a new probabilistic model (PM). • DM: Unexpectedness of ‘neck’ can be found by building a deterministic prototype using the average of (normalized) property frequencies of all MAMMALS. As the frequency of ‘neck’ for a giraffe is far higher than the average, this can be found. • PM: DM is effective when the underlying data (i.e., knowledge base) to derive probabilities are not noisy, but an automatically harvested knowledge base inevitably contains noise. For example, Probase, mining textual patterns such as “<property> of <category>,” may contain ‘some part’ as a property of a giraffe, which can be recognized as a desirable unexpected property by DM. We propose a probabilistic model that overcomes this problem. ∗ This work was done at Yonsei University and Microsoft Research. † Zhongyuan Wang is now with Facebook Inc. This work is licenced under a Creative …

Content truncated to 500 words