Measuring Semantic Relations between Human Activities

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 664–673, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Measuring Semantic Relations between Human Activities Steven R. Wilson and Rada Mihalcea University of Michigan {steverw|mihalcea}@umich.edu Abstract The things people do in their daily lives can provide valuable insights into their personality, values, and interests. Unstruc- tured text data on social media platforms are rich in behavioral content, and auto- mated systems can be deployed to learn about human activity on a broad scale if these systems are able to reason about the content of interest. In order to aid in the evaluation of such systems, we introduce a new phrase-level semantic textual sim- ilarity dataset comprised of human activ- ity phrases, providing a testbed for auto- mated systems that analyze relationships between phrasal descriptions of people’s actions. Our set of 1,000 pairs of activi- ties is annotated by human judges across four relational dimensions including simi- larity, relatedness, motivational alignment, and perceived actor congruence. We eval- uate a set of strong baselines for the task of generating scores that correlate highly with human ratings, and we introduce sev- eral new approaches to the phrase-level similarity task in the domain of human ac- tivities. 1 Introduction Our everyday behaviors say a lot about who we are. The things we do are related to our personality (Ajzen, 1987), values (Rokeach, 1973), interests (Goecks and Shavlik, 2000), and what we are go- ing to do next (Ouellette and Wood, 1998). While we cannot always directly observe what people are doing on a day-to-day basis, we have access to a large number of unstructured text sources that describe real-world human activity, such as news outlets and social media sites. Fiction and non- fiction writings often revolve around the things that people do, and even encyclopedic texts can be rich in descriptions of human activities. Al- though many common sources of text contain hu- man activities, reasoning about these activities and their relationships to one another is not a trivial task. Descriptions of human actions are fraught with ambiguity, subjectivity, and there are multi- tudinous lexically distinct ways to express highly similar events. If we want to gain useful insights from these data, it should be beneficial to develop effective systems that can successfully represent, compare, and ultimately understand human activ- ity phrases. In this paper, we consider the task of automati- cally determining the strength of a relationship be- tween two human activities,1 which can be helpful in reasoning about texts rich with activity-based content. The relationship between activities might be similarity in a strict sense, such as watching a film and seeing a movie, or a more general related- ness, such as the relationship between turn on an oven and bake a pie. Another way to categorize a pair of activities is by the degree to which they are typically done with a similar motivation, like eat- ing dinner with family and visiting relatives. Or, in order to …

Content truncated to 500 words