<?xml version="1.0" standalone="yes"?>
<Paper uid="I05-6003">
  <Title>A Study of Applying BTM Model on the Chinese Chunk Bracketing</Title>
  <Section position="4" start_page="27" end_page="28" type="concl">
    <SectionTitle>
4 Conclusion and Future Directions
</SectionTitle>
    <Paragraph position="0"> In this paper, we define a word-layer matrix that can be used to translate the CKIP Treebank and the Penn Chinese Treebank into corresponding BTM datasets. By the BTM dataset, we developed a BTM model, adopting two types of conditional probabilities and using full TL POS pattern matching and full TL POS template matching to identify the chunks for each segmented and POS-tagged Chinese sentence.</Paragraph>
    <Paragraph position="1"> Our experiment results show that the BTM model can effectively achieve precision and recall optimization on the matching sets for both perfect input and actual input. The experimental results also demonstrate that:  (1) The BTM threshold value is positively related to the BTM F-measure; (2) The POS layer number is positively related to the BTM F-measure; (3) The F-measure of our BTM model for the  matching set should be not sensitive to two BTM parameters: BTM threshold value and BTM training size; (4) The chunk bracketing of our BTM model on the matching set should be high and stable (or say, robust) against training size, perfect and actual input while POS layer number is t 2 and BTM threshold value is t 0.5; (5) The BTM model can provide a matching set with high and stable performance (more than 95% F-measure) for improving N-gram-like models without trial-and-error, or say, a tuning process. For most statistical language models, such N-gram models, need tuning to improve their performance and large-scale corpus to  overcome corpus sparseness problem (Manning et al., 1999; Gao et al., 2002; Le et al., 2003). Furthermore, it is difficult for them to identify their &amp;quot;matching set&amp;quot; with high and stable performance, whereas our BTM model has the ability to support chunkers and parsers for improving chunking performance. According to the fourth experiment results, when applying a BTM (0.5, 2, 45,000) model on the matching set and a 4-gram model on the non-matching set, the combined system can improve the F-measure of 4-gram model 2.5% for perfect input and 1.0% for actual input. Among the chunking and parsing models, Cascaded Markov Models should be the first one to construct the parse tree layer by layer with each layer's Markov Model. As per (Brants, 1999), each layer's chunk bracketing of Cascaded Markov Models is dependent because the output of a lower layer is passed as input to the next higher layer. On the contrast, our BTM model can independently generate the chunks for top layer without the results of lower layer chunk bracketing; and (6) Since the F-measures of the BTM model for the matching sets of perfect and actual input both are greater than 95%, we believe our BTM model can be used not only to improve the F-measure of existing shallow parsing or chunking systems, but also to help select valuable sentences from the non-matching set for effectively extending the CR of our BTM model.</Paragraph>
    <Paragraph position="2"> In the future, we shall study how to combine our BTM model with more conventional statistical approaches, such as Bayesian Networks, Maximum Entropy and Cascaded Markov Models, etc. Meanwhile, we will also apply our BTM model to the Penn English Treebank as a comparative study.</Paragraph>
  </Section>
class="xml-element"></Paper>