Interpersonal Relationship Labels for the CALLHOME Corpus

Interpersonal Relationship Labels for the CALLHOME Corpus Denys Katerenchuk1, David Guy Brizan1,2, Andrew Rosenberg3 1CUNY Graduate Center 365 Fifth Avenue, New York, NY 10016 2University of San Francisco 101 Howard Street, San Francisco, CA 94105 3IBM Research AI 1099 Kitchawan Rd., Yorktown Heights, NY, 10598 1dkaterenchuk@gradcenter.cuny.edu, 2dgbrizan@usfca.edu, 3amrosenb@us.ibm.com Abstract The way we speak to our friends, colleagues, or partners is different in both the explicit context, what we say, and the implicit, how we say it. Understanding these differences is important because it provides additional information that can be used in natural language processing tasks. For example, knowing the relationship between interlocutors can help to narrow the range of topics and improve automatic speech recognition system results. Unfortunately, the lack of corpora makes exploration of this problem intractable. In this work, we release a set of interpersonal relationship labels between conversation participants for the CALLHOME English corpus. We make the labels freely available for download on our website and hope that this effort can further boost research in this area. Keywords: corpus annotation, interlocutor relationship labels, CALLHOME corpus 1. Introduction Communication is one of the most invaluable tools humans have. It enables us to understand each other, share ideas, coordinate work, and build relationships. Through speech, we carry explicit information in terms of words, as well as implicit information that is usually expressed by an acoustic signal of the human voice. For example, when you are on the phone with a friend, it is often clear when the person is distressed, even when they are trying to conceal or deny this fact. Word choice and speech intonation are formed on the conscious and unconscious levels. Unlike the conscious level, we have little control over the unconscious. Understanding the relation between explicit and implicit information during conversations can provide additional information about the speakers. We have all been in a park or a restaurant and overheard or observed strangers interacting with each other. Have you ever tried to guess the relationships between them? Were they two old friends, cousins, or maybe a couple? On what distinct characteristics did you base your assumption? Was it the body language, voice, or word choice? It may seem like an easy task for humans to identify the relationships. However, this is a very difficult task for computers. People are interacting with cellphones, smart TVs, and computers on a daily basis using voice-based interfaces. However, these interactions can be harmed by misun- derstandings. One reason for the occurrence of the misunderstandings is that these systems rely on automated speech recognition (ASR) systems, which, despite showing strong improvements in performance, are far from perfect. One way to improve an ASR system’s performance is to take into account not only spoken words, but also speaker and domain information. For example, information about speaker’s gender, age, or voice characteristics can be used to reduce recognition word error rate (Saon et al., 2013; Serizel and Giuliani, 2017). The knowledge of a discussion topic, in a similar fashion, can aid the …

Content truncated to 500 words