<?xml version="1.0" standalone="yes"?>
<Paper uid="J04-3001">
  <Title>c(c) 2004 Association for Computational Linguistics Sample Selection for Statistical Parsing</Title>
  <Section position="7" start_page="270" end_page="272" type="evalu">
    <SectionTitle>
5. Related Work
</SectionTitle>
    <Paragraph position="0"> Sample selection benefits problems in which the cost of acquiring raw data is cheap but the cost of annotating them is high, as is certainly the case for many supervised learning tasks in natural language processing. In addition to PP-attachment, as discussed in this article, sample selection has been successfully applied to other classification  Computational Linguistics Volume 30, Number 3 applications. Some examples include text categorization (Lewis and Catlett 1994), base noun phrase chunking (Ngai and Yarowsky 2000), part-of-speech tagging (Engelson Dagan 1996), spelling confusion set disambiguation (Banko and Brill 2001), and word sense disambiguation (Fujii et al. 1998).</Paragraph>
    <Paragraph position="1"> More challenging are learning problems whose objective is not classification, but generation of complex structures. One example in this direction is applying sample selection to semantic parsing (Thompson, Califf, and Mooney 1999), in which sentences are paired with their semantic representation using a deterministic shift-reduce parser. A recent effort that focuses on statistical syntactic parsing is the work by Tang, Lou, and Roukos (2002). Their results suggest that the number of training examples can be further reduced by using a hybrid evaluation function that combines a hypothesisperformance-based metric such as tree entropy (&amp;quot;word entropy&amp;quot; in their terminology) with a problem-space-based metric such as sentence clusters.</Paragraph>
    <Paragraph position="2"> Aside from active learning, researchers have applied other learning techniques to combat the annotation bottleneck problem in parsing. For example, Henderson and Brill (2002) consider the case in which acquiring additional human-annotated training data is not possible. They show that parser performance can be improved by using boosting and bagging techniques with multiple parsers. This approach assumes that there are enough existing labeled data to train the individual parsers. Another technique for making better use of unlabeled data is cotraining (Blum and Mitchell 1998), in which two sufficiently different learners help each other learn by labeling training data for one another. The work of Sarkar (2001) and Steedman, Osborne, et al. (2003) suggests that co-training can be helpful for statistical parsing. Pierce and Cardie (2001) have shown, in the context of base noun identification, that combining sample selection and cotraining can be an effective learning framework for large-scale training. Similar approaches are being explored for parsing (Steedman, Hwa, et al.</Paragraph>
    <Paragraph position="3"> 2003; Hwa et al. 2003).</Paragraph>
  </Section>
class="xml-element"></Paper>