<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-1020">
  <Title>User-Friendly Text Prediction for Translators</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> The idea of using text prediction as a tool for translators was first introduced by Church and Hovy as one of many possible applications for &amp;quot;crummy&amp;quot; machine translation technology (Church and Hovy, 1993). Text prediction can be seen as a form of interactive MT that is well suited to skilled translators. Compared to the traditional form of IMT based on Kay's original work (Kay, 1973)--in which the user's role is to help disambiguate the source text-prediction is less obtrusive and more natural, allowing the translator to focus on and directly control the contents of the target text. Predictions can benefit a translator in several ways: by accelerating typing, by suggesting translations, and by serving as an implicit check against errors.</Paragraph>
    <Paragraph position="1"> The first implementation of a predictive tool for translators was described in (Foster et al., 1997), in the form of a simple word-completion system based on statistical models. Various enhancements to this were carried out as part of the TransType project (Langlais et al., 2000), including the addition of a realistic user interface, better models, and the capability of predicting multi-word lexical units. In the final TransType prototype for English to French translation, the translator is presented with a short pop-up menu of predictions after each character typed.</Paragraph>
    <Paragraph position="2"> These may be incorporated into the text with a special command or rejected by continuing to type normally. null Although TransType is capable of correctly anticipating over 70% of the characters in a freely-typed translation (within the domain of its training corpus), this does not mean that users can translate in 70% less time when using the tool. In fact, in a trial with skilled translators, the users' rate of text production declined by an average of 17% as a result of using TransType (Langlais et al., 2002). There are two main reasons for this. First, it takes time to read the system's proposals, so that in cases where they are wrong or too short, the net effect will be to slow the translator down. Second, translators do not always act &amp;quot;rationally&amp;quot; when confronted with a proposal; that is, they do not always accept correct proposals and they occasionally accept incorrect ones.</Paragraph>
    <Paragraph position="3"> Many of the former cases correspond to translators simply ignoring proposals altogether, which is understandable behaviour given the first point.</Paragraph>
    <Paragraph position="4"> Association for Computational Linguistics.</Paragraph>
    <Paragraph position="5"> Language Processing (EMNLP), Philadelphia, July 2002, pp. 148-155. Proceedings of the Conference on Empirical Methods in Natural This paper describes a new approach to text prediction intended to address these problems. The main idea is to make predictions that maximize the expected benefit to the user in each context, rather than systematically proposing a fixed amount of text after each character typed. The expected benefit is estimated from two components: a statistical translation model that gives the probability that a candidate prediction will be correct or incorrect, and a user model that determines the benefit to the translator in either case. The user model takes into account the cost of reading a proposal, as well as the random nature of the decision to accept it or not.</Paragraph>
    <Paragraph position="6"> This approach can be characterized as making fewer but better predictions: in general, predictions will be longer in contexts where the translation model is confident, shorter where it is less so, and absent in contexts where it is very uncertain.</Paragraph>
    <Paragraph position="7"> Other novel aspects of the work we describe here are the use of a more accurate statistical translation model than has previously been employed for text prediction, and the use of a decoder to generate predictions of arbitrary length, rather than just single words or lexicalized units as in the TransType prototype. The translation model is based on the maximum entropy principle and is designed specifically for this application.</Paragraph>
    <Paragraph position="8"> To evaluate our approach to prediction, we simulated the actions of a translator over a large corpus of previously-translated text. The result is an increase of over 10% in translator productivity when using the predictive tool. This is a considerable improvement over the -17% observed in the TransType trials.</Paragraph>
  </Section>
class="xml-element"></Paper>