Integration of soft data for information fusion: Pitfalls, challenges and trends

This paper explores the question of soft data integration for information fusion. The topic is of particular interest in military and civilian fusion applications making use of both sensor and human-based information. Soft data has much to offer for such applications, where subtle connections are difficult to observe with physical sensors. However, the purpose of this analysis is to broaden our comprehension of soft data as a key issue for future fusion approaches without making any assumption on the interplay between models processing hard and soft data. From a series of efforts conducted in various research projects, we have identified the main pitfalls and challenges of soft data processing. Among them, ambiguity in natural language, uncertainty in observation and characterization of human sources necessitate original processing methods. Additionally, the paper illustrates two complementary trends developed for soft data integration. The first one relies on the use of BML, an unambiguous controlled language developed to be understood by both operators and automated systems. The second one is using previously acquired domain semantics to perform analysis of soft data at different conceptual levels. Lessons learned from practical use of those techniques allow us to build up a personal perception of the current situation in the field of soft data integration for information fusion.