Presented at the International Conference on Applied Human Factors and Ergonomics (AHFE), Las Vegas, NV (July 2015).

Business, news, financial, and intelligence analysts collect vast amounts of information to compose, reason about, and refine hypotheses that are often communicated as narratives. Narratives are a natural and prevalent form of communication that provide organizing principles to guide the development of hypotheses for events across domains. Narrative structure can promote analyst understanding, hypothesis testing, and general reasoning about large collections of data. However, narrative-based approaches for collecting and reasoning about data available from modern sources (e.g., blogs, social media) must address issues of data overload and data quality to be effective. For example, while analysts may be interested in performing forensic analysis of events in foreign countries, it can be difficult to identify objective, truthful information from a large collection of data from open sources. Analysts must identify conflicting and corroborating information to develop those hypotheses and generate conclusions. While narrative-based approaches offer a natural framework to capture this reasoning and communicate findings to others, the mostly manual processes of collecting new data, refining hypotheses, and maintaining conflicting and corroborating information places a heavy burden on analysts. To address this issue, we designed and prototyped a set of visualization methods that convey key information elements (e.g., actors, motivations, relationships), meta-information (e.g., certainty, recency, source), and relationships across information elements to help analysts create narrative-based hypotheses. These visualizations facilitate the development of coherent and cohesive narratives while attempting to account for analysts’ bias.

In this paper, we describe key elements for narrative-centric reasoning over large data sets and application towards hypothesis generation. We also present examples of our approach to support narrative-based hypothesis generation within a visualization framework. Finally, we describe preliminary evaluation efforts taken with representative users to assess the usability and effectiveness of these approaches. We anticipate these methods will reduce the cognitive workload of analysis and improve the quality of narrative-based hypotheses generated, leading to more useful and correct hypotheses in less time.