The Fingerprint of Human Referring Expressions and their Surface Realization with Graph Transducers (IS-FP, IS-GT, IS-FP-GT)}

The algorithm IS-FP takes up the idea from the IS-FBN algorithm developed for the shared task 2007. Both algorithms learn the individual attribute selection style for each human that provided referring expressions to the corpus. The IS-FP algorithm was developed with two additional goals (1) to improve the indentification time that was poor for the FBN algorithm and (2) to push the dice score even higher. In order to generate a word string for the selected attributes, we build based on individual preferences a surface syntactic dependency tree as input. We derive the individual preferences from the training set. Finally, a graph transducer maps the input strucutre to a deep morphologic structure. 1 IS-FP: Generating Referring Expression with a Human Imprint A review of the referring expressions shows that humans prefer frequently distinct attributes and attribute combination such as in the following examples. grey desk (30t1), a red chair (30t2), red sofa (30t3), blue chair (30t5), a small green desk (30t6) the one in the top left corner (31t1), the one to the left in the middle row (31t2), the bottom right most one (31t3), the blue chair at the bottom center (31t5), etc. The first individual (#30) seems to prefer colour and size while the second one (#31) seems to prefer the relative position (to the left) and places (the top left corner). Because of the review, we checked, if the Incremental Algorithm (Dale and Reiter, 1995) using the order for the attributes due to the frequency calculated for each individual can outperform the algorithm using the order for the attributes due to the frequency of the complete training set. This was the case. Table 1 shows the results. Using the individual attribute order, the IA performed as good as the FBN algorithm, cf. Table 1. Algorithm Furniture People Avg. IA (complete) 0.796 0.710 0.753 IA (individual) 0.835 0.736 0.7855 FBN 0.810 0.762 0.786 Table 1: Incremental algorithm and FBN However, the FBN algorithm generates all possible referring expressions and selects based on the dice metric the most similar expressions of the same human. Since there is usually a set of equal good referring expressions, it is possible to select a referring expression among these results due to another metric. That this would improve the results shows the experiment to selected among these results the referring expression that is closest to the correct result. The outcome was that the FBN algorithm has still about 9% room for improvements. The following sections investigates possibilities to use this chance. 1.1 Identification Time An important metric is the identification time that is the time which is used by a human to identify an entity due to a given referring expression. The identification time is very loosely related with the number of minimal referring expressions and therefore likely with the length of a referring expression. The best identification times had a system with 74% minimal referring expressions and the second and third best