<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1093">
  <Title>Automatic Generation of Domain Models for Call Centers from Noisy Transcriptions</Title>
  <Section position="4" start_page="737" end_page="738" type="intro">
    <SectionTitle>
2 Background and Related Work
</SectionTitle>
    <Paragraph position="0"> In this work, we are trying to bridge the gap between a few seemingly unrelated research areas viz. (1) Automatic Speech Recognition(ASR), (2) Text Clustering and Automatic Taxonomy Generation (ATG) and (3) Call Center Analytics. We present some relevant work done in each of these areas.</Paragraph>
    <Paragraph position="1"> Automatic Speech Recognition(ASR): Automatic transcription of telephonic conversations is proven to be more difficult than the transcription of read speech. According to (Padmanabhan et al., 2002), word-error rates are in the range of 78% for read speech whereas for telephonic speech it is more than 30%. This degradation is due to the spontaneity of speech as well as the telephone channel. Most speech recognition systems perform well when trained for a particular accent (Lawson et al., 2003). However, with call centers now being located in different parts of the world, the requirement of handling different accents by the same speech recognition system further increases word error rates.</Paragraph>
    <Paragraph position="2"> Automatic Taxonomy Generation (ATG): In recent years there has been some work relating to mining domain specific documents to build an ontology. Mostly these systems rely on parsing (both shallow and deep) to extract relationships between key concepts within the domain. The ontology is constructed from this by linking the extracted concepts and relations (Jiang and Tan, 2005). However, the documents contain well formed sentences which allow for parsers to be used.</Paragraph>
    <Paragraph position="3"> Call Center Analytics: A lot of work on automatic call type classification for the purpose of categorizing calls (Tang et al., 2003), call routing (Kuo and Lee, 2003; Haffner et al., 2003), obtaining call log summaries (Douglas et al., 2005), agent assisting and monitoring (Mishne et al., 2005) has appeared in the past. In some cases, they have modeled these as text classification problems where topic labels are manually obtained (Tang et al., 2003) and used to put the calls into different buckets. Extraction of key phrases, which can be used as features, from the noisy transcribed calls is an important issue. For manually transcribed calls, which do not have any noise, in (Mishne et al., 2005) a phrase level significance estimate is obtained by combining word level estimates that were computed by comparing the frequency of a word in a domain-specific corpus to its frequency in an open-domain corpus. In (Wright et al., 1997) phrase level significance was obtained for noisy transcribed data where the phrases are clustered and combined into finite state machines. Other approaches use n-gram features with stop word removal and minimum support (Kuo and Lee, 2003; Douglas et al., 2005). In (Bechet et al., 2004) call center dialogs have been clustered to learn about dialog traces that are similar.</Paragraph>
    <Paragraph position="4"> Our Contribution: In the call center scenario, the authors are not aware of any work that deals with automatically generating a taxonomy from transcribed calls. In this paper, we have tried to formalize the essential aspects of a domain model.</Paragraph>
    <Paragraph position="5"> We show an unsupervised method for building a domain model from noisy unlabeled data, which is available in abundance. This hierarchical domain model contains summarized topic specific details for topics of different granularity. We show how such a model can be used for topic identification of unseen calls. We propose two applications for  aiding agents while handling calls and for agent monitoring based on the domain model.</Paragraph>
  </Section>
class="xml-element"></Paper>