<?xml version="1.0" standalone="yes"?>
<Paper uid="P01-1065">
  <Title>A Generic Approach to Parallel Chart Parsing with an Application to LinGO</Title>
  <Section position="6" start_page="0" end_page="0" type="evalu">
    <SectionTitle>
5 Related Work
</SectionTitle>
    <Paragraph position="0"> Parallel parsing for NLP has been researched extensively. For example, Thompson (1994) presented some implementations of parallel chart parsers. Nijholt (1994) gives a more theoretical overview of parallel chart parsers. A survey of parallel processing in NLP is given by Adriaens and Hahn (1994).</Paragraph>
    <Paragraph position="1"> Nevertheless, many of the presented solutions either did not yield acceptable speedup or were very speci c to one application. Recently, several NLP systems have been parallelized successfully. Pontelli et al. (1998) show how two existing NLP applications were successfully parallelized using the parallel Prolog environment ACE. The disadvantage of this approach, though, is that it can only be applied to parsers developed in Prolog.</Paragraph>
    <Paragraph position="2"> Manousopoulou et al. (1997) discuss a parallel parser generator based on the Eu-PAGE system. This solution exploits coarse-grained parallelism of the kind that is unusable for many parsing applications, including our own (see also G orz et. al. (1996)).</Paragraph>
    <Paragraph position="3"> Nurkkala et al. (1994) presented a parallel parser for the UPenn TAG grammar, implemented on the nCUBE. Although their best results were obtained with random grammars, speedups for the English grammar were also considerable.</Paragraph>
    <Paragraph position="4"> Yoshida et. al. (Yoshida et al., 1999) presented a 2-phase parallel FB-LTAG parser, where the operations on feature structures are all performed in the second phase. The speedup ranged up to 8.8 for 20 processors, Parallelism is mainly thwarted by a lack of parallelism in the rst phase.</Paragraph>
    <Paragraph position="5"> Finally, Ninomiya et al. (2001) developed an agent-based parallel parser that achieves speedups of up to 13.2. It is implemented in ABCL/f and LiLFeS. They also provide a generic solution that could be applied to many parsers. The main di erence with our system is the distribution of work. This system uses a tabular chart like distribution of matches and a randomized distribution of uni cation tasks. Experiments we conducted show that the choice of distribution scheme can have a signi cant in uence on the cache utilization.</Paragraph>
    <Paragraph position="6"> It should be mentioned, though, that it is in general hard to compare the performance of systems when di erent grammars are used.</Paragraph>
    <Paragraph position="7"> On the scheduling side, our approach shows close resemblance to the Cilk-5 system (Frigo et al., 1998). It implements work stealing using similar techniques. An important difference, though, is that our scheduler was designed for chart parsers and tabular algorithms in general. These types of applications fall outside the class of applications that Cilk is capable of handling e ciently.</Paragraph>
  </Section>
class="xml-element"></Paper>