On the optimal decision rule for sequential interactive structured prediction

Interactive structured prediction (ISP) is an emerging framework for structured prediction (SP) where the user and the system collaborate to produce a high quality output. Typically, search algorithms applied to ISP problems have been based on the algorithms for fully-automatic SP systems. However, the decision rule applied should not be considered as optimal since the goal in ISP is to reduce human effort instead of output errors. In this work, we present some insight into the theory of the sequential ISP search problem. First, it is formulated as a decision theory problem from which a general analytical formulation of the optimal decision rule is derived. Then, it is compared with the standard formulation to establish under what conditions the standard algorithm should perform similarly to the optimal decision rule. Finally, a general and practical implementation is given and evaluated against three classical ISP problems: interactive machine translation, interactive handwritten text recognition, and interactive speech recognition.

Graphical abstractFigure optionsDownload full-size imageDownload high-quality image (171 K)Download as PowerPoint slideHighlights► We present an optimum algorithm for sequential interactive structured prediction. ► A general analytical formulation of the optimum search algorithm is derived. ► The classical algorithm can be seen as a maximum approximation to our algorithm. ► A general and practical implementation of the optimum algorithm is given. ► Our approach performs better although the classical algorithm is a good approx.