next up previous contents
Next: Evaluation using the constrained Up: Experiments Previous: Experiments   Contents


Evaluation using the unconstrained MERGE

For the unconstrained version (Figure 7.3), positional indexing outperformed the non-indexed parser by an average of 21% (the best improvement being 44%). Although the difference at every sentence between the parsing times for path and positional indexing is rather variable (while the difference between parsing times for positional indexing and for non-indexed parser is almost constant), the improvements in parsing times for path indexing over the non-indexed parser are better than those of positional indexing: an average of 25%, with a maximum of 45%. However, there were no significant variations in the improvements brought by both indexing method as the parsing times increased.

When using the parsing times of the positional indexing parser as a baseline, path indexing performs better than positional indexing with an average of 5% (with a maximum of 24%). These improvements, as well as both indexing methods' improvements over the non-indexed parser, are explained by a reduction in the number of unsuccessful unifications. Some examples illustrating this reduction are presented in Table 7.1.


Table 7.1: The number of successful and failed unifications for the non-indexed and indexed parsers over the unconstrained MERGE grammar, for selected sentences where significant improvements (at least 10%) in parsing times were recorded both between positional indexing and EFD and between path indexing and positional indexing. The sentence numbers are the same as those used in Figure 7.3.
Sentence Successful Failed unifications
number unifications EFD EFD with EFD with
      positional index path index
101 29 217 143 119
851 95 643 384 282
1357 355 2983 2092 1594
1816 557 8626 6475 4676



next up previous contents
Next: Evaluation using the constrained Up: Experiments Previous: Experiments   Contents