<?xml version="1.0" standalone="yes"?>
<Paper uid="P03-1025">
  <Title>Compounding and derivational morphology in a finite-state setting</Title>
  <Section position="9" start_page="0" end_page="0" type="evalu">
    <SectionTitle>
8 Implementation and experiments
</SectionTitle>
    <Paragraph position="0"> We implemented the selective history-based RTNconstruction in Prolog, as a conversion routine that takes as input a definite-clause grammar with compiled-out grounded feature values; it produces as output a Prolog representation of an FSA. The resulting automaton is determinized and minimized, using the FSA library for Prolog by Gertjan van Noord.9 Emphasis was put on identifying the most suitable strategy for dealing with word formation taking into account the relative size of the FSAs generated (other techniques than the selective history strategy were tried out and discarded).</Paragraph>
    <Paragraph position="1"> The algorithm was applied on a sample word formation grammar with 185 compiled-out context-free rules, displaying the principled mechanism of category and other feature selection, but not the full set of distinctions made in the DeKo project. 9 of the rules were compiled from the prefixation rule, and were thus marked as h-rules for the selective method.</Paragraph>
    <Paragraph position="2"> We ran a comparison between a version of the non-selective parameterized RTN-method of (Nederhof 2000) and the selective history method proposed in this paper. An overview of the results is given in fig. 1.10 It should be noted that the optimizations of sec. 7 were applied in both methods (the non-selective method was simulated by mark- null for the selective method is an artefact of the sample grammar. ing all rules as h-rules).</Paragraph>
    <Paragraph position="3"> As the size results show, the non-deterministic FSAs constructed by the selective method are more complex (and hence resource-intensive in minimization) than the ones produced by the &amp;quot;plain&amp;quot; parameterized version. However, the difference in exactness of the approximizations has to be taken into account. As a tentative indication for this, note that the minimized FSA for a35a59a36 a38 in the plain version has only two states; so obviously too many distinctions from the context-free grammar have been lost.</Paragraph>
    <Paragraph position="4"> In the plain version, all word formation operations are treated alike, hence the history list of length one or two is quickly filled up with items that need not be recorded. A comparison of the number of different pairs of categories and history lists used in the construction shows that the selective method is more economical in the use of memory space as the depth parameter grows larger. (For a35 a36a12a1 , the selective method would even have fewer different category/history list pairs than the plain method, since the patterns become repetitive. However, the approximations were impractical for a35 a36a2a1 .) Since the selective method uses non-h-rules only in the determination of legal histories (as discussed in sec. 6), it can actually &amp;quot;see&amp;quot; further back into the history than the length of the history list would suggest.</Paragraph>
    <Paragraph position="5"> What the comparison clearly indicates is that in terms of resource requirements, our selective method with a parameter a35 a25 is much closer to the a35 a25 -version of the plain RTN-method than to the next higher a35 a25 a18a14a13 version. But since the selective method focuses its record-keeping resources on the crucial aspects of the finite-state approximation, it brings about a much higher gain in exactness than just extending the history list by one in the plain method. We also ran the selective method on a more fine-grained morphological grammar with 403 rules (including 12 h-rules). Parameter a35 a36 a38 was applicable, leading to a non-deterministic FSA with 7,345 states, which could be minimized. Parameter a35 a36 a20 led to a non-deterministic FSA with 87,601 states, for which minimization could not be completed due to a memory overflow. It is one goal for future research to identify possible ways of breaking down the approximation construction into smaller subproblems for which minimization can be run separately (even though all categories belong to the same equivalence class of mutually recursive categories).11 Another goal is to experiment with the use of transduction as a means of adding structural markings from which the analysis trees can be reconstructed (to the extent they are not underspecified by the finite-state approach); possible approaches are discussed in Johnson 1996 and Boullier 2003.</Paragraph>
    <Paragraph position="6"> Inspection of the longest few hundred prefixcontaining word forms in a large German newspaper corpus indicates that prefix stacking is rare. (If there are several prefixes in a word form, this tends to arise through compounding.) No instance of stacking of depth 3 was observed. So, the range of phenomena for which the approximation is inexact is of little practical relevance. For a full evaluation of the coverage and exactness of the approach, a comprehensive implementation of the morphological grammar would be required. We ran a preliminary experiment with a small grammar, focusing on the cases that might be problematic: we extracted from the corpus a random sample of 100 word forms containing prefixes. From these 100 forms, we generated about 3700 grammatical and ungrammatical test examples by omission, addition and permutation of stems and affixes. After making sure that the required affixes and stems were included in the lexicon of the grammar, we ran a comparison of exact parsing with the unification-based grammar and the selective history-based RTN-approximation, with parameter a35 a36 a38 (which means that there is a history window of one item). For 97% of the test items, the two methods agreed; 3% of the items were accepted by the approximation method, but not by the full grammar. The approximation does not lose any 11A related possibility pointed out by a reviewer would be to expand features from the original unification-grammar only where necessary (cf. Kiefer and Krieger 2000).</Paragraph>
    <Paragraph position="7"> test items parsed by the full grammar. Some obvious improvements should make it possible soon to run experiments with a larger history window, reaching exactness of the finite-state method for almost all relevant data.</Paragraph>
  </Section>
class="xml-element"></Paper>