Some Results

This section presents a number of preliminary results to indicate how well the system currently performs. I used a corpus of 500 word-graphs, output of a preliminary version of the speech recognizer, and typical of the intended application. The size of these word-graphs is given in table 1

Table 1: The size of the test set of 500 word graphs

number of transitions	number of graphs
0-5	250
6-10	103
11-15	50
16-20	26
21-30	23
31-40	16
41-50	8
51-75	15
> 75	9

Currently the grammar assigns the following number of readings to each of these word-graphs (table 2). It can be concluded from this table that in about 20% of the cases the grammar cannot assign a full parse to the input.

Table 2: Number of readings assigned to the word-graphs by the grammar

number of readings	number of graphs
0	104
1	261
$\geq$ 2	135
2	76
3	18
4	20
5	3
6	4
8	4
9	1
10	1
11	1
14	1
16	2
18	1
24	1
48	1
84	1

The cputime and memory requirements of three different parsers are listed in the tables 3 and 4.

Table 3: Cpu-time (in milliseconds) needed to analyse the test set of 500 word-graphs. Timings include the robustness component. Timings were obtained on a fast HP-UX 735. This machine is about three times faster than a Pentium 75MHz. The first parser is the head-corner parser as described in this paper. The second parser is an inactive bottom-up chart parser with packing. The third parser is a bottom-up Earley parser with packing.

parser	total (msec)	average	maximum
head-corner	50983	101	2820
inactive-chart	348867	697	37761
bottom-up Earley	443364	886	51460

Table 4: Memory requirements (in kilo-bytes) of the parsers. The figures listed indicate the size of the parse-tables.

parser	total (Kbytes)	average	maximum
head-corner	7280	14	299
inactive-chart	34025	68	1912
bottom-up Earley	84295	169	5403

The final table lists the penalty score (as explained above) for the set of 500 word-graphs ( 5).

Table 5: Penalty scores for the set of 500 word-graphs. The best and worst possibilities are defined by the best path through the word-graph (which ignores bigrams) and the best possible path through this word-graph. Using bigrams currenlty works only slightly better than the NLP-component including the robustness component.

	total score	number of hits
speech (no bigram) best path	771	277
speech (no bigram) best possible	170	402
speech (with bigram)	508	353
nlp (robustness, no bigram, speech for skips)	532	349
nlp (robustness, no bigram, none for skips)	557	341
nlp (no robustness, no bigram, speech for skips)	663	315
nlp (no robustness, no bigram, none for skips)	869	288