<?xml version="1.0" standalone="yes"?>
<Paper uid="N04-4010">
  <Title>Using N-best Lists for Named Entity Recognition from Chinese Speech</Title>
  <Section position="3" start_page="0" end_page="1" type="intro">
    <SectionTitle>
1. Introduction
</SectionTitle>
    <Paragraph position="0"> Named Entity Recognition (NER) is the first step for many tasks in the fields of natural language processing and information retrieval. It is a designated task in a number of conferences, including the Message Understanding Conference (MUC), the Information Retrieval and Extraction Conference (IREX), the Conferences on Natural Language Learning (CoNLL) and the recent Automatic Content Extraction Conference (ACE).</Paragraph>
    <Paragraph position="1"> There has been a considerable amount of work on English NER yielding good performance (Tjong Kim Sang et al. 2002, 2003; Cucerzan &amp; Yarowsky 1999; Wu et al. 2003). However, Chinese NER is more difficult, especially on speech output, due to two reasons. First, Chinese has a large number of homonyms and the vocabulary used in Chinese person names is an open set so more characters/words are unseen in the training data. Second, there is no standard definition of Chinese words. Word segmentation errors made by recognizers may lead to NER errors. Previous work on Chinese textual NER includes Jing et al. (2003) and Sun et al. (2003) but there has been no published work on NER in spoken Chinese.</Paragraph>
    <Paragraph position="2"> Named Entity Recognition for speech is more difficult than for text, since the most reliable features for textual NER (punctuation, capitalization, and syntactic patterns) are often not available in speech output. NER on automatically recognized broadcast news was first conducted by MITRE in 1997, and was subsequently added to Hub-4 evaluation as a task. Palmer et al.</Paragraph>
    <Paragraph position="3"> (1999) used error modeling, and Horlock &amp; King (2003) proposed discriminative training to handle NER errors; both used a hidden Markov model (HMM). Miller et al.</Paragraph>
    <Paragraph position="4"> (1999) also reported results in English speech NER using an HMM model. In a NIST 1999 evaluation, it was found that NER errors on speech arise from a combination of ASR errors and errors of the underlying NER system.</Paragraph>
    <Paragraph position="5"> In this work, we investigate whether the NIST finding holds for Chinese speech NER as well. We present the first known result for recognizing named entities in realistic large-vocabulary spoken Chinese. We propose to use the best-known model for Chinese textual NER-a maximum entropy model--on Chinese speech NER.</Paragraph>
    <Paragraph position="6"> We also propose using re-segmentation and post-classification to improve this model. Finally, we propose to integrate the ASR and NER components to optimize NER performance by making use of the n-best  ASR output.</Paragraph>
    <Paragraph position="7"> 2. A Spoken Chinese NER Model</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.1 LVCSR output
</SectionTitle>
      <Paragraph position="0"> We use the ASR output from BBN's Byblos system on broadcast news data from the Xinhua News Agency, which has 1046 sentences. This system has a character error rate of 7%. We had manually annotated them with named entities as an evaluation set according to the PFR corpus annotation guideline (PFR 2001).</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="1" type="sub_section">
      <SectionTitle>
2.2 A maximum-entropy NER model with post-
</SectionTitle>
      <Paragraph position="0"> classification To establish a baseline spoken Chinese NER model, we selected a maximum entropy (MaxEnt) approach since this is currently the single most accurate approach known for recognizing named entities in text (Tjong Kim Sang et al., 2002, 2003, Jing et al., 2003)  . In the CoNLL 2003 NER evaluation, 5 out of 16 systems use MaxEnt models and the top 3 results for English and top</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>