<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-1430">
  <Title>From Context to Sentence Form</Title>
  <Section position="3" start_page="225" end_page="227" type="metho">
    <SectionTitle>
2 Approach: froln context to
</SectionTitle>
    <Paragraph position="0"> sentence form via attention focus Our goal is to obtain a less intrusive device through context sensitivity of the spoken output. The presupposition is that utterances anchored to the hearer's multidimensional context will require less cognitive effort to attend to. Our strategy is based on the discourse pragmatic account for grammatical differences between utterances expressing the same propositional content. Figure 2 shows how we envision the connection between various disciplines relating context to utterance form.</Paragraph>
    <Paragraph position="1"> Context is considered to activate particular mental representations in the heater's mind (modeled as her attention space). In order to be communicative, the speaker hypothesises about this attention space and structures his utterance accordingly. Informalion Structure Theory accounts for this adaptation process. We use our earlier conlext model and developed a strategy for determining topic and focus based on the analysis of COMFIIS&amp;quot; discourse wag-</Paragraph>
    <Section position="1" start_page="225" end_page="226" type="sub_section">
      <SectionTitle>
2.1 Context modeling
</SectionTitle>
      <Paragraph position="0"> Context perception and adaptation are important in research on wearable technologies. Nomadic Radio (Sawhney and Schmandt, 1999) adapts information delivery (news, email, phone-calls) to the user's physical environment through varying output sound technology. Remembrance agents act like memory extensions by pointing to related information in appropriate contexts (De Vaul et al., 2000). Neither use linguistic form variation. Our I3 sister project.</Paragraph>
      <Paragraph position="1"> HIPS (Benelli et al., 1999) does and focuses on the interaction between physical and hypertext navigation for a wearable museum guide. Schmidt and colleagues provide a platform for inferring tile relationship between low-level contextual data and application dependent concepts (Schmidt et al., 1999).</Paragraph>
      <Paragraph position="2"> When dealing with content delivery to human users.</Paragraph>
      <Paragraph position="3"> the use and interpretation of symbolic data in combination with quantitative data remains an important issue on the research agenda. Our context model is a first proposal in that direction.</Paragraph>
      <Paragraph position="4"> When focusing on lexical phenonaena like d,qctie expressions (this afternoon, here) and anaphora (she, the same topic) or the inclusion of appositions related to the bearer's profile (one of your favourite topics), .we proposed a three-dimensional context model (see figure 3) in order to generate truly context sensitive expressions. Objects mentioned to the user are recorded in a discourse model, her location in space and time is monitored via beacons. The Information \[,aver provides user profile information (in terms of persons and topics of interest). Entil ies in the NLG input structure are annotated with con-I ext ual informal ion of t hose different perspect ires.  We will use the same multi-dimensional context * model for building an .attention space..model of .the hearer. Only for the physical context, we need additional reasoning on the time and location indexes in terms of the activity of the user (cfr. 2.3). Indeed, knowing which kind of activity the user is involved in at each moment (i.e. the ontology instances involved in that activity) we hypothesise on which person and keyword the user's attention is focused on.</Paragraph>
    </Section>
    <Section position="2" start_page="226" end_page="226" type="sub_section">
      <SectionTitle>
2.2 Attention focus and Information
</SectionTitle>
      <Paragraph position="0"> Structure Theory ........ .</Paragraph>
      <Paragraph position="1"> Other researchers have investigated attention focus in larger spans of discourse .(McCoy and Cheng, 1991; Grosz and Sidner, 1986) and in dialogue (Jokinen et al., 1998). Corpus analysis (Rats, 1996) confirms the existence of a mechanism called topic, through which interlocutors strive at discourse coherence to reduce the cognitive effort of the hearer. The terminology used in the different frameworks is confusing, even contradictory (Bosch and van der Sandt, 1999). Information Structure Theory (Lambrecht, 1994) accounts for exactly those phenomena we are interested in: grammatical differences between allo-sentences (expressing the same semantic content). Lambrecht considers information structure as an integral part of the grammar of natural languages. After determining what to say, a speaker structures this information in terms of his own presupposition of the hearer's attention state.</Paragraph>
      <Paragraph position="2"> Identifiability (whether a shared representation exists in both interlocutors' minds) and activation status (how much a known representation is at the forefront of the bearer's mind (Chafe, 1987)) determine pragmatic role assignments. Topic, the role of aboutness is attributed to a discourse referent that is identifiable and more or less active. Focus is the unpredictable part of tile utterance. Whereas all utterances have a focus (in order to be comnmnicative), some may be topic-less. Lambrecht distinguishes 3 types of sentence constructions (according to whether the predicate, the argument or the whole sentence is in focus 3) and demonstrates through granamatical analysis, that tile first construction is tile most natural one. Languages use different gramnlatical markers to realise information structure and there is no one-to-one correspondence between grammatical markers (e.g. definiteness, pronominalization, accentuation) and topic or  3Examples taken from .(Lambrecht, 1994): ((S.\taLL CaPS indicate prosodic accent) (a) predicate focus: what did the children do? The children went to SCHOOL.</Paragraph>
      <Paragraph position="3"> (b) argument focus: who went to school? The CHILDREN went to school.</Paragraph>
      <Paragraph position="4"> (c) sentence focus (topic-less): what happened't I'he (HIL-DRKN wellt to SLItOOL.</Paragraph>
      <Paragraph position="5"> focus. In English, topic is preferably realised as an ..... trnaccented.~pronounC/- while., focus elements,,usually carry prosodic accent 4.</Paragraph>
    </Section>
    <Section position="3" start_page="226" end_page="227" type="sub_section">
      <SectionTitle>
2.3 COMRIS discourse pragmatics
</SectionTitle>
      <Paragraph position="0"> There is no content-based discourse planning in COMRIS. The propositional content of parrot messages is provided by agents that represent particular user interests in the virtual world. A mechanism of competition for attention determines whether a message will actually be pushed to the user. As a '..consequences, ~the sentences to-be generated-are topicless: each message conveys only new information, as if answering the hypothetical question: 'what is appropriate for me to do now?'. Thus they bare the danger of coming 'out of the blue', as in the following  sequence: o &amp;quot;There will be an interesting presentation by Amanda Huggenkiss about 'knowledge systems and AI'.&amp;quot; (propagandist message) o &amp;quot;Enric Plaza proposes to meet you to discuss about 'machine learning'.&amp;quot; (appointment proposal) null o &amp;quot;Josep Arcos, who shares your interest in 'agents', is currently in your neighbourhood.&amp;quot; (proximity alert) o -&amp;quot; Please note you have to give a presentation on 'NLG and context' within 5 minutes.&amp;quot; (commitment reminder).</Paragraph>
      <Paragraph position="1">  The intuition that such a sequence is not ideal from the communicative point of view, confirms our interpretation of information structure theory in view of communicative effect. Whereas topic expression creates discourse continuity (i.e. links tile message to tile context in a broad sense: an active mental representation), topic-less sentences can be assumed to require a higher cognitive effort from tile hearer. Therefore our communicative strategy for COM RIS will be to look for a topic candidate within a given propositional content. To be communicatively more effective, we try to somehow link a message to the attent.ion space of the user-hearer.</Paragraph>
      <Paragraph position="2"> Obviously, the bearer's mind is a black box and all we can do is hypo/hesise about the activation of mental representations by contextual factors. In line with our previous work, we argue that the 3 dimensions of the user's context (linguistic, physical and profile) should be taken into account. Given the COMRIS ontology, the attention state model can be represented as a simple tree structure (see examples in section 3): each utterance conveys information 4This is a simplification of I,ambrecht 'sanalysis. Our point is that less prosodic accents reduce the cognitive effort of the hearer, which is our goal. Combined with the choice of sentence structure, it constitutes our strategy for reduced obtrusi v{~lless.</Paragraph>
      <Paragraph position="3">  about, an event characterised by a key-word (-list), involving a person, and possibly a time/location specification. Thus we will search in the hearer's discourse and physical context which are the activated instances of the concepts event, person, key-word and time/location. To find out which instances are contributed by the physical context, we hypothesise about the user's current activity by comparing her physical position with the conference programme or her agenda. For instance, if we know that the user is attending a particular presentation, we can query the conferenceprogram for the speaker and the keywords of that presentation. Alternatively, if tim user's physical location confirms that she attends an appointment, her agenda will reveal the name of the person she's meeting and maybe some topics of discussion. Any of these instances may also carry context annotation w.r.t, the user's interest profile. Section 3 explains this further through scenarios.</Paragraph>
    </Section>
    <Section position="4" start_page="227" end_page="227" type="sub_section">
      <SectionTitle>
2.4 NLG strategy
</SectionTitle>
      <Paragraph position="0"> Assignment of topic and focus follows from our application of Information Structure Theory to tile discourse pragmatic situation in COMRIS. Our search for a topic candidate in the NLG input structure considers time pressure first, then the activation of entities via discourse or activity and finally the hearer's interest profile, as detailed in the following rules: 1. (physical context first) If the NLG input structure contains a time expression that is annotated as being very close to the current point in time (physical context value), then let the time expression be the topic, opening the sentence and carrying a prosodic accent. The sentence structure is predicate focus.</Paragraph>
      <Paragraph position="1"> e.g. Please note that, within FIVE MINUTES, you have to give a presentation on 'NATURAL</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="227" end_page="227" type="metho">
    <SectionTitle>
LANGUAGE GENERATION AND CONTEXT'.
</SectionTitle>
    <Paragraph position="0"> 2. (topzc candidate in attention space) If one of the entities of the input structure is also present in the (hearer's) attention space map, let it be the topic, realised as an unaccented pronoun (preferred topic marking) in case it occurred in tile immediate linguistic context a or as a leftdislocated constituent in case it was present in the physical context.</Paragraph>
    <Paragraph position="1"> e.g. She will give a presentation on KNOWL-</Paragraph>
  </Section>
  <Section position="5" start_page="227" end_page="227" type="metho">
    <SectionTitle>
EDGE SYSTEMS and AI.
</SectionTitle>
    <Paragraph position="0"> AMANDA HUGGENKISS, she will give a presentation on KNOWLEDGE SYSTEMS and AI.</Paragraph>
    <Paragraph position="1"> 3. (profile conte~'l also matters) If none of tile above situations occur, verity' whether any of the entities of the inpul stru'lure has a high -:'immediat,'ly preceding message, m,t tmJre than X ag~,. wh,,re X i~ a tim,&amp;quot; Ihteshoht profile value (indicating tile hearer's interest in * that keyword, or person), ff the physical.context also allows topic shift, use an argument focus structure (after introducing the new topic): e.g. Someone interested in 'AGENTS' is in your neighbourhood. It's JOSEP ARCOS 6.</Paragraph>
    <Paragraph position="2"> 4. Else (default) use a sentence focus structure.</Paragraph>
    <Paragraph position="3"> e.g. PENN SILL proposes to MEET you to talk about 'NLG and CONTEXT.</Paragraph>
    <Paragraph position="4"> The scenarios below will further concretise tile relationship between context, attention space and topic-focus assignment, but tile above examples already illustrate our main point. The first 3 rules are aimed at linking an element of the propositional content to the user's attention focus, in virtue of tile preceding discourse, the physical context or her interest, profile. Topic expression often leads to de-accentuation. In other words, rule 4 applies when there is no way to anchor the utterance to the user's context and requires to accent every information entity. Empirical experiments will have to verify the hypothesis that the non-default sentence constructions are perceived as less intrusive.</Paragraph>
  </Section>
class="xml-element"></Paper>