<?xml version="1.0" standalone="yes"?>
<Paper uid="P03-2038">
  <Title>An Intelligent Procedure Assistant Built Using REGULUS 2 and ALTERF</Title>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 REGULUS 2
</SectionTitle>
    <Paragraph position="0"> REGULUS 2 (Rayner et al., 2003; Regulus, 2003) is an Open Source environment that supports efficient compilation of typed unification grammars into speech recognisers. The basic intent is to provide a set of tools to support rapid prototyping of spoken dialogue applications in situations where little or no corpus data exists. The environment has already been used to build over half a dozen applications with vocabularies of between 100 and 500 words.</Paragraph>
    <Paragraph position="1"> The core functionality provided by the REGULUS 2 environment is compilation of typed unification grammars into annotated context-free grammar language models expressed in Nuance Grammar Specification Language (GSL) notation (Nuance, 2003). GSL language models can be converted into runnable speech recognisers by invoking the Nuance Toolkit compiler utility, so the net result is the ability to compile a unification grammar into a speech recogniser.</Paragraph>
    <Paragraph position="2"> Experience with grammar-based spoken dialogue systems shows that there is usually a substantial overlap between the structures of grammars for different domains. This is hardly surprising, since they all ultimately have to model general facts about the linguistic structure of English and other natural languages. It is consequently natural to consider strategies which attempt to exploit the overlap between domains by building a single, general grammar valid for a wide variety of applications. A grammar of this kind will probably offer more coverage (and hence lower accuracy) than is desirable for any given specific application. It is however feasible to address the problem using corpus-based techniques which extract a specialised version of the original general grammar.</Paragraph>
    <Paragraph position="3"> REGULUS implements a version of the grammar specialisation scheme which extends the Explanation Based Learning method described in (Rayner et al., 2002). There is a general unification grammar, loosely based on the Core Language Engine grammar for English (Pulman, 1992), which has been developed over the course of about ten individual projects. The semantic representations produced by the grammar are in a simplified version of the Core Language Engine's Quasi Logical Form notation (van Eijck and Moore, 1992).</Paragraph>
    <Paragraph position="4"> A grammar built on top of the general grammar is transformed into a specialised Nuance grammar in the following processing stages:  1. The training corpus is converted into a &amp;quot;treebank&amp;quot; of parsed representations. This is done using a left-corner parser representation of the grammar.</Paragraph>
    <Paragraph position="5"> 2. The treebank is used to produce a specialised grammar in REGULUS format, using the EBL algorithm (van Harmelen and Bundy, 1988; Rayner, 1988).</Paragraph>
    <Paragraph position="6"> 3. The final specialised grammar is compiled into a Nuance GSL grammar.</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 ALTERF
</SectionTitle>
    <Paragraph position="0"> ALTERF (Rayner and Hockey, 2003) is another Open Source toolkit, whose purpose is to allow a clean combination of rule-based and corpus-driven processing in the semantic interpretation phase. There is typically no corpus data available at the start of a project, but considerable amounts at the end: the intention behind ALTERF is to allow us to shift smoothly from an initial version of the system which is entirely rule-based, to a final version which is largely data-driven.</Paragraph>
    <Paragraph position="1"> ALTERF characterises semantic analysis as a task slightly extending the &amp;quot;decision-list&amp;quot; classification algorithm (Yarowsky, 1994; Carter, 2000). We start with a set of semantic atoms, each representing a primitive domain concept, and define a semantic representation to be a non-empty set of semantic atoms. For example, in the procedure assistant domain we represent the utterances please speak up show me the sample syringe set an alarm for five minutes from now no i said go to the next step respectively as fincrease volumeg fshow, sample syringeg fset alarm, 5, minutesg fcorrection, next stepg where increase volume, show, sample syringe, set alarm, 5, minutes, correction and next step are semantic atoms. As well as specifying the permitted semantic atoms themselves, we also define a target model which for each atom specifies the other atoms with which it may legitimately combine. Thus here, for example, correction may legitimately combine with any atom, but minutes may only combine with correction,set alarm or a number.1.</Paragraph>
    <Paragraph position="2"> Training data consists of a set of utterances, in either text or speech form, each tagged with its intended semantic representation. We define a set of feature extraction rules, each of which associates an utterance with zero or more features. Feature extraction rules can carry out any type of processing. In particular, they may involve performing speech recognition on speech data, parsing on text data, application of hand-coded rules to the results of parsing, or some combination of these. Statistics are then compiled to estimate the probability p(a j f) of each semantic atom a given each separate feature f, using the standard formula p(a j f) = (Naf + 1)=(Nf + 2) where Nf is the number of occurrences in the training data of utterances with feature f, and N af is the number of occurrences of utterances with both feature f and semantic atom a.</Paragraph>
    <Paragraph position="3"> The decoding process follows (Yarowsky, 1994) in assuming complete dependence between the features. Note that this is in sharp contrast with the Naive Bayes classifier (Duda et al., 2000), which assumes complete independence. Of course, neither assumption can be true in practice; however, as argued in (Carter, 2000), there are good reasons for preferring the dependence alternative as the better option in a situation where there are many features extracted in ways that are likely to overlap.</Paragraph>
    <Paragraph position="4"> We are given an utterance u, to which we wish to assign a representation R(u) consisting of a set of semantic atoms, together with a target model comprising a set of rules defining which sets of seman1The current system post-processes Alterf semantic atom lists to represent domain dependancies between semantic atoms more directly before passing on the result. e.g. (correction, set alarm, 5, minutes) is repackaged as (correction(set alarm(time(0,5)))) tic atoms are consistent. The decoding process proceeds as follows:  1. Initialise R(u) to the empty set.</Paragraph>
    <Paragraph position="5"> 2. Use the feature extraction rules and the statistics compiled during training to find the set of all triples hf; a; pi where f is a feature associated with u, a is a semantic atom, and p is the probability p(a j f) estimated by the training process.</Paragraph>
    <Paragraph position="6"> 3. Order the set of triples by the value of p, with the largest probabilities first. Call the ordered set T .</Paragraph>
    <Paragraph position="7"> 4. Remove the highest-ranked triple hf; a; pi from  T . Add a to R(u) iff the following conditions are fulfilled: p pt for some pre-specified threshold value pt.</Paragraph>
    <Paragraph position="8"> Addition of a to R(u) results in a set which is consistent with the target model.</Paragraph>
    <Paragraph position="9"> 5. Repeat step (4) until T is empty.</Paragraph>
    <Paragraph position="10"> Intuitively, the process is very simple. We just walk down the list of possible semantic atoms, starting with the most probable ones, and add them to the semantic representation we are building up when this does not conflict with the consistency rules in the target model. We stop when the atoms suggested are too improbable, that is, they have probabilies below a cut-off threshold.</Paragraph>
  </Section>
class="xml-element"></Paper>