<?xml version="1.0" standalone="yes"?>
<Paper uid="C00-2169">
  <Title>Processing Self Corrections in a speech to speech system</Title>
  <Section position="3" start_page="0" end_page="1116" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Spontaneous speech is disfluent. In contrast to read speech the sentences aren't perfectly planned before they are uttered. Speakers often modify their plans while they speak. This results in pauses, word repetitions or changes, word fragments and restarts. Current mltorustic speech understanding systems perform very well in small domains with restricted speech but have great difficulties to deal with such disfluencies. A system that copes with these self corrections (=repairs) must recognize the spoken words and identify the repair to get the intended meaning of an utterance. To characterize a repair it is commonly segmented into the following four parts (el. fig.i):  * reparandum: the &amp;quot;wrong&amp;quot; part of the utterance null * interruption point (IP): marker at the end of the reparandum * editing term: special phrases, which indicate a repair like &amp;quot;well&amp;quot;, &amp;quot;I mean&amp;quot; or filled pauses such as &amp;quot;uhln '~, &amp;quot;uh&amp;quot; * reparans: the correction of the reparandum on Thursday lcannot * no Ican meet &amp;quot;ah afteronc</Paragraph>
    <Paragraph position="2"> Only if reparandum and editing term are known, the utterance can be analyzed in the right way. It remains an open question whether the two terms should be deleted before a semantic analysis as suggested sometimes in the literature 1. If both terms are marked it is a straight-forward preprocessing step to delete reparandum and editing term. In the Verbmobil 2 corpus, a corpus dealing with appointment scheduling a.nd tr~vel planning, nearly 21% of all turns contain at least one repair. As a consequence a speech understanding system thai; cannot handle repairs will lose perforlnance on these turns.</Paragraph>
    <Paragraph position="3"> Even if repairs are defined by syntactic and semantic well-formedness (Levelt, 1983) we observe that most of them are local phenomena..</Paragraph>
    <Paragraph position="4"> At this point we have to differentiate between restarts and other repairs a (modification repairs). Modification repairs have a strong correspondence between reparandum and reparans, 1In most cases a reparaudum could be deleted without any loss of information. But, for exmnple, if it introduces an object which is referred to later, a deletion is not appropriate.</Paragraph>
    <Paragraph position="5"> &gt;l?his work is part of the VERBMOBIL project and was funded by the German Federal Ministry for Research and Technology (BMBF) in the framework of the Verbmobil Project under Grant BMBF 01 IV 701 V0. The responsibility for the contents of this study lies with the authors.</Paragraph>
    <Paragraph position="6"> SOften a third kind of repair is defined: &amp;quot;abridged repairs&amp;quot;. These repairs consist solely of an editing term and are not repairs in our sense.</Paragraph>
    <Paragraph position="7">  whereas restarts a.re less structured. In our believe there is no nted for a. complete syntactic am@sis to detect ~md correct most modification repairs. Thus, in wh~tt follows, we will concentra.te on this ldnd of repa.ir.</Paragraph>
    <Paragraph position="8"> There are two major arguments to process repairs before t)arsing. Primarily spontaneous speech is not always syntactically well-formed even in the absence of sell' corrections. Second (Meta-) rules increase the pa.rsers' search space. This is perhaps acceptable for transliterated speech but not for speech recognizers output like l~ttices because they represent millions of possible spoken utterances. \[n addition, systems whk;h a.re not based on a. deep syntactic and semantic amdysis e.g. statistical dialog act prediction -- require a repa.ir processing step to resolve contr~dictions like the one in tit. 1.</Paragraph>
    <Paragraph position="9"> We propose all algorithm for word lattices th,~t divides repa.ir detection a.nd correction in three steps (of. fig. 2) l&amp;quot;irst, ~r trigger indicates potential 1Ps. Second, a sl;ochasl, ic model tries to lind an appropria.te repair h)r each IP by guessing 1,he mosl; l)robable segmentation, qb accomplish this, repair processing is seen as a statistical machine translation problem where the repa.randum is a transl~tion of the reparans.</Paragraph>
    <Paragraph position="10"> For every repair found, a pa.th representing the spcaker.' intended word sequence is inserted into the la.ttice. In the last step, a lattice parser selects the best pa.th.</Paragraph>
    <Paragraph position="11"> tlll 'llllll'Sday I C/iIIlllt)l IlO \[ CIIII lllCel &amp;quot;ah tiller t)llC</Paragraph>
  </Section>
class="xml-element"></Paper>