<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-2098">
  <Title>Exact Decoding for Jointly Labeling and Chunking Sequences</Title>
  <Section position="3" start_page="0" end_page="763" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> The Viterbi algorithm and the CKY algorithms are two decoding algorithms essential to the area of natural language processing. The former models a linear chain of labels such as part of speech tags, and the latter models a parse tree. Both are used to extract the best prediction from the model (Manning and Schutze, 1999).</Paragraph>
    <Paragraph position="1"> However, some tasks seem to fall between the two, having more than one layer but flatter than the trees created by parsers. For example, in relation extraction, we have entities in one layer and relations between entities as another layer. Another task is shallow parsing. We may want to model part-of-speech tags and noun/verb chunks at the same time, since performing simultaneous labeling may result in increased joint accuracy by sharing information between the two layers of labels.</Paragraph>
    <Paragraph position="2"> To apply the Viterbi decoder to such tasks, we need two models, one for each layer. We must feed the output of one layer to the next layer. In such an approach, errors in earlier processing nearly always accumulate and produce erroneous results at the end.</Paragraph>
    <Paragraph position="3"> If we use CKY, we usually end up flattening the output tree to obtain the desired output. This seems like a round-about way of modeling two layers.</Paragraph>
    <Paragraph position="4"> There are previous attempts at modeling two layer labeling. Dynamic Conditional Random Fields (DCRFs) by (McCallum et al, 2003; Sutton et al, 2004) is one such attempt, however, exact inference is in general intractable for these models and the authors were forced to settle for approximate inference. null Our contribution is a novel model for two layer labeling, for which exact decoding is tractable. Our experiments show that our use of label-chunk structures results in significantly better performance over cascaded CRFs, and that the model is a promising alternative to DCRFs.</Paragraph>
    <Paragraph position="5"> The paper is organaized a follows: In Section 2 and 3, we describe the model and present the decoding algorithm. Section 4 describes the learning methods applicable to our model and the baseline models. In Section 5 and 6, we describe the experiments and the results.</Paragraph>
  </Section>
class="xml-element"></Paper>