<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-2302">
  <Title>Stochastic Language Generation in a Dialogue System: Toward a Domain Independent Generator</Title>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 Stochastic Generation (HALogen)
</SectionTitle>
    <Paragraph position="0"> We used the HALogen framework (Langkilde-Geary, 2002) for our surface generation. HALogen was originally created for a domain within MT and is a sentence planner and a surface realizer. Analysis and MT applications can be found in (Langkilde and Knight, 1998; Knight and Langkilde, 2000).</Paragraph>
    <Paragraph position="1"> HALogen accepts a feature-value structure ranging from high-level semantics to shallow syntax. Figure 1 shows a mixture of both as an example. Given this input, generation is a two step process. First, the input form is converted into a word forest (a more efficient representation of a word lattice) as described in (Langkilde-Geary, 2002). Second, the language model chooses the most probable path through the forest as the output sentence. null  moved the ambulance.</Paragraph>
    <Paragraph position="2"> The word forest is created by a series of grammar rules that are designed to over-generate for a given representation. As figure 1 shows, there is a lot of syntactic information missing. The rules are not concerned with generating only syntactically correct possibilities, but to generate all possibilities under every input that is not specified (our example does not provide a determiner for ambulance, so the grammar would produce the definite and indefinite versions). Once the forest is created, the language model chooses the best path(s) through the forest.</Paragraph>
    <Paragraph position="3"> We modified HALogen's grammar to fit the needs of a dialogue system while maintaining the same set of roles and syntactic arguments recognized by the grammar. The</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
TRIPS Logical Form uses many more roles than HALo-
</SectionTitle>
    <Paragraph position="0"> gen recognizes, but we converted them to the smaller set. By using HALogen's set of roles, we can be assured that our grammar is domain independent from TRIPS.</Paragraph>
    <Paragraph position="1"> We did, however, expand the grammar within its current roles. For instance, we found the theme role to be insufficient and changed the grammar to generate more syntactic constructs (for example, we generate the theme in both the object and subject positions). We also expanded the production rules for interrogatives and imperatives, both of which were sparsely used/tested because of HALogen's original use in MT domains.</Paragraph>
    <Paragraph position="2"> HALogen is able to expand WordNet word classes into their lexical items, but due to the difficulty of mapping the TRIPS word classes to WordNet, our input terms to HALogen are the desired lexical items instead of word classes as shown in figure 1. Future work includes linking the grammar to the TRIPS word classes instead of WordNet.</Paragraph>
  </Section>
  <Section position="6" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 The Dialogue System
</SectionTitle>
    <Paragraph position="0"> We developed our approach within TRIPS, a collaborative planning assistant that interacts with a human user mainly through natural language dialogue, but also through graphical displays. The system supports many domains involving planning scenarios, such as a 911 disaster rescue assistant and a medical adviser. TRIPS per- null (Dzikovska et al., 2003)) forms advanced reasoning and NLP tasks including, but not limited to, interpretation in context, discovering user intentions, planning, and dialogue management. Language generation has largely been ignored in the system until recently. As with many dialogue systems, it has simply been a means to show results in the above areas through a language back-end. Recently, Stent (Stent, 1999) did extensive work on dialogue management through rule-based generation (Allen et al., 2001).</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.1 Logical Form of Meaning
</SectionTitle>
      <Paragraph position="0"> There are two meaning representations in TRIPS. The first is a domain independent representation called the logical form (LF). The second is a domain dependent knowledge representation (KR). The effort toward creating the domain independent LF is part of an overall goal of creating a dialogue system that is easily portable to new domains. A domain-specific representation is always needed for reasoning, and mapping rules are created to map the LF into the KR for each domain. These rules are easier to create than a new logical representation for each domain.</Paragraph>
      <Paragraph position="1"> Dzikovska, Swift and Allen (Dzikovska et al., 2003) have built a parser that parses speech utterances into this domain-independent LF. The LF is very important to this paper. One of the biggest problems that any surface generation approach faces is that it takes a lot of work to generate sentences for one domain. Moving to a new domain usually involves duplicating much of this work. However, if we create a surface generator that uses the LF as input, we have created a surface generator that is able to generate English in more than one specific domain.</Paragraph>
      <Paragraph position="2"> The LF ontology consists of a single-inheritance hierarchy of frame-like LF types that classify entities according to their semantics and argument structure. Every LF type can have a set of thematic arguments with selectional restrictions. The ontology is explicitly designed to capture the semantic differences that can affect sen- null The parser uses the LF type definitions to build a general semantic representation of the input. This is a flat and unscoped representation of the semantics of the sentence that serves as input to the TRIPS discourse interpretation modules (which perform reference resolution, disambiguation, intention recognition to produce the final intended meaning). Figure 3 gives an example of the LF representation of the sentence, he took an aspirin. It can be read as follows: A speech act of type SA TELL occurred with content being V11, which is a proposition of type LF CONSUME (more specifically &amp;quot;take&amp;quot;), with AGENT V123 and THEME V433. V123 is pronominal form of type LF PERSON and pro-type HE, and V433 is an indefinitely specified object that is of type LF DRUG (more specifically &amp;quot;aspirin&amp;quot;).</Paragraph>
      <Paragraph position="3"> The LF representation serves as the input to our surface generation grammar after a small conversion. If natural human quality dialogue can be produced from this LF, not only has a domain independent generator been created, but also a generator that shares ontologies and lexicons with the parser.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.2 Integrating HALogen into TRIPS
</SectionTitle>
      <Paragraph position="0"> The task of converting our independent Logical Form (LF) into HALogen's Abstract Meaning Representation was relatively straightforward. Several rules were created to change LF specific roles into the smaller set of roles that the surface generation grammar recognizes. LF roles such as COGNIZER and ENTITY are converted to AGENT and THEME respectively. Verb properties represented by TMA are converted into the appropriate syntactic roles of TENSE, MODALITY, AUXILLARY, etc. The LF type triple is reduced to just the lexical item and appropriate determiners are attached when the LF provides enough information to warrant it. It is best  This resulting AMR is the input to HALogen where it is converted into a word forest using our modified dialogue-based HALogen grammar. Finally, the language model chooses the best output.</Paragraph>
      <Paragraph position="1"> The above conversion applies to declarative, imperative and interrogative speech acts. These are translated and generated by the method in section 3. We also take a similar approach to Stent's previous work (Stent, 1999) that generated grounding and turn-taking acts using a template-based method. These usually short utterances do not require complex surface generation and are left to templates for proper production.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>