Parameter tuning

Using the Honours Advising dialogue due to its interesting features (the existence of several convincing `hits', and some human-annotated topic changes not even hinted at by the TextTiling metric), and using as a starting point Hearst's default settings [6] of pseudosentence length=20, block size=6, experiments were performed to determine the optimal settings for these two parameters. Maintaining the total window size as an approximate constant, the pseudosentence length was first halved, as seen in figure 5.16. Then the same run was performed with the pseudosentence length doubled from its original size of 20, as seen in figure 5.17. While interesting, these results seemed to be universally worse than Hearst's original settings, and so it was decided that the domain in this respect had no effect on parameter requirements.

Figure 5.16: Honours Advising dialogue: $ps=10$ $bs=12$

Figure 5.17: Honours Advising dialogue: $ps=40$ $bs=3$