TextTiling parameters

Primary system parameter experimentation involved varying pseudosentence length and block size. The initial values are shown in table 4.1:

Table 4.1: Experiment 1

Pseudosentence length	20
Block size	6

These values were chosen as a starting point, as the original TextTiling system decided on these values as best for expository text [6]. It was theorised that these settings would be suboptimal for spoken dialogue.

It was noted that more data points were acquired when the pseudosentence size was reduced. Given the data sparsity observed in some evaluation dialogues (which were shorter than the magazine article examined by Hearst) which would not have allowed many troughs to appear due to a lack of data points, an experiment was performed using a pseudosentence of length 1 (table 4.2). To avoid altering more than one parameter, the effective total block length (in words) was kept the same, by changing block_size to 120. These single-word-pseudosentence experiments were time intensive, and only a limited number were performed.

Table 4.2: Experiment 2

Pseudosentence length	1
Block size	120

The experiments in tables 4.3 and 4.4 were then performed, modifying Hearst's settings more conservatively:

Table 4.3: Experiment 3

Pseudosentence length	40
Block size	3

Table 4.4: Experiment 4

Pseudosentence length	10
Block size	12

While still maintaining an overall block-size of 120 words (as with experiments 1 and 2), experiments 3 and 4 respectively double and halve the pseudosentence length, effectively modifying the step-size and overlap size of the rolling window.

Because experiment 2 (table 4.2) showed interesting results, a further two experiments were performed modifying the size of the window, shown in tables 4.5 and 4.6.

Table 4.5: Experiment 5

Pseudosentence length	1
Block size	240

Table 4.6: Experiment 6

Pseudosentence length	1
Block size	60

This modification has the result of changing the total size of the area of the dialogue currently under consideration by the system, without affecting the step size or the overlap size. With a larger window, more words form the two blocks under comparison, increasing the sample size over which the word frequency table is constructed. Therefore, random frequency fluctuation would be expected to be reduced by a larger window, while resolution, as well as the ability to perform comparisons near the beginning and end of the document due to the requirement to have a whole block to the left and right of each measurement point, might suffer.