File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/evalu/05/w05-1523_evalu.xml
Size: 2,557 bytes
Last Modified: 2025-10-06 13:59:34
<?xml version="1.0" standalone="yes"?> <Paper uid="W05-1523"> <Title>Parsing Generalized ID/LP Grammars</Title> <Section position="5" start_page="192" end_page="192" type="evalu"> <SectionTitle> 4 Evaluation </SectionTitle> <Paragraph position="0"> To evaluate the effectiveness of the GIDLP formalism, a moderate-scale grammar of German was obtained from Professor Martin Volk (Stockholm University). This grammar combines ID/LP rules with PS rules, as argued for in (Volk 1996), and uses a flat structure to encode the flexibility of the German Mittelfeld. As an example, the rule for ditransitive verbs is given in (1).</Paragraph> <Paragraph position="2"> This grammar can be mechanically translated into the GIDLP format, as each of Volk's PS rules corresponds to a GIDLP rule. This translation establishes an 'initial' GIDLP grammar.2 The grammar was then optimized in two successive steps to take advantage of the GIDLP formalism. First, a 'medial' grammar was created in which word order domains were introduced only when necessary. (In the ID/LP formalism, every local tree is an order domain.) Second, a 'final' grammar was created by reordering the RHS order of each rule so as to put the most discriminatory RHS element first - generally the finite verb. To compare these three grammars, a testsuite of 150 sentences was constructed. The sentences were generally chosen to equally cover the sentence types recognized by the grammar. The results from parsing this testsuite with each grammar are summarized in Figure 3, which shows the average number of chart insertion attempts at each sentence length.</Paragraph> <Paragraph position="3"> (Chart insertion attempts have traditionally been used as an overall metric for parsing efficiency, as parse time tends to be dominated by the time taken searching the chart for blocking edges.) Overall, the final grammar shows a clear improvement over the medial and initial grammars.</Paragraph> <Paragraph position="4"> 2As Volk's parser is not available, the relative performance of the GIDLP parser on the initial grammar and of Volk's parser on his grammar cannot be determined. Thus Volk's grammar is only used as a basis for the three GIDLP grammars described here.</Paragraph> <Paragraph position="5"> mar sees a decrease of 69.2% in the number of chart insertion attempts compared to the initial grammar.</Paragraph> <Paragraph position="6"> Thus the expressive capabilities provided by the GIDLP formalism lead directly to improvements in parsing efficiency.</Paragraph> </Section> class="xml-element"></Paper>