A Context-Aware NLP Approach For Noteworthiness Detection in Cellphone Conversations

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 25–36, Dublin, Ireland, August 23-29 2014. A Context-Aware NLP Approach For Noteworthiness Detection in Cellphone Conversations Francesca Bonin ∗ Trinity College Dublin Dublin, Ireland boninf@tcd.ie Jose San Pedro Telefonica Research Barcelona, Spain jspw@tid.es Nuria Oliver Telefonica Research Barcelona, Spain nuriao@tid.es Abstract This papers presents a context-aware NLP approach to automatically detect noteworthy infor- mation in spontaneous mobile phone conversations. The proposed method uses a supervised modeling strategy which considers both features from the content of the conversation as well as contextual information from the call. We empirically analyze the predictive performance of features of different nature on a corpus of mobile phone conversations. The results of this study reveal that the context of the conversation plays a crucial role on boosting the predictive perfor- mance of the model. 1 Introduction More than 6 billion people worldwide use their cellphones daily for a variety of purposes: contacting colleagues, relatives or friends, doing business, getting help in emergency situations, etc. Previous work (Carrascal et al., 2012) has shown that almost 40% of users frequently feel the need to recall bits of information from their phone conversations and that 27% of the users consider the recall task to be difficult, mainly because taking notes during a mobile phone call is not always possible (e.g. hands not free, lack of time or devices for note-taking). In a related user study, Cycyl et al. reveal that users are often engaged in concurrent tasks during mobile phone conversations (e.g. walking, jogging, driving, cooking, etc), which makes taking notes an unfeasible task (Cycil et al., 2013). In this setting, information extraction techniques could be applied to automatically detect noteworthy information from mobile phone conversations. Related studies have focused on detecting noteworthiness from meeting transcripts (Banerjee and Rudnicky, 2009). However, very little work has been done to date to identify this kind of information in other types of human communication, such as spontaneous phone conversations. In this paper, we present a data-driven information extraction approach aimed at automatically detect- ing fragments of phone conversations worth annotating for future recall, i.e. noteworthy. These call notes could then be presented to the users to enable fast browsing of their conversation history, and leveraged to design efficient information interaction techniques for supporting smart user interfaces. Given the particular characteristics of mobile phone calls, detecting noteworthiness in them is chal- lenging at many levels. First, the audio is captured in a natural environment rather than in controlled settings, which results in noisy signals, and consequently in noisy transcriptions. Second, the conversa- tions are highly fragmented due to their spontaneous nature. Finally, at a conceptual level, judging which pieces of information are noteworthy is a very subjective task, as emerged in (Banerjee and Rudnicky, 2009), who investigated the feasibility of the task by conducting a Wizard of Oz-based user study. Our noteworthiness modeling approach considers a supervised learning paradigm which takes into account two types of information: (1) Contextual …

Content truncated to 500 words