<?xml version="1.0" standalone="yes"?>
<Paper uid="W94-0324">
  <Title>Generating Cooperative System Responses in Information Retrieval Dialogues</Title>
  <Section position="4" start_page="211" end_page="214" type="metho">
    <SectionTitle>
3 The Corinna System
</SectionTitle>
    <Paragraph position="0"> In order to create an information system based on the integrated model of COR and RST, several aspects had to be considered which go beyond the previously outlined theoretical framework.</Paragraph>
    <Paragraph position="1"> The models COR and RST are of descriptive nature; they are not sufficient to actually construct a dialogue between user and system. To do this, a dialogue manager also needs further semantic information which constitutes the propositional content of the dialogue act to be generated. This semantic knowledge depends on and is provided by the genre (information seeking), the application area (information retrieval) and the actual domain.</Paragraph>
    <Paragraph position="2"> As domain we chose the content of the CORDIS databases, which represent information about EC-funded research projects, participating organizations and strategic funding programs. The databases are offered online by ECHO, a database host of the Commission of the European Community. They have been transformed to and integrated into a relational database system which is publicly available. For the implementation of Corinna a way of incorporating this domain knowledge into the dialogue management and the generation of the dialogue acts had to be found.</Paragraph>
    <Paragraph position="3"> Natural language was decided to be the main mode of interaction in the Corinna system, because it achieves a high level of expressive power. The user is presented a menu of possible dialogue acts (according to the current dialogue state) among which she may choose the most appropriate one. The PENMAN generation system was used to produce both the system's and the user's dialogue contributions. An additional modality of deictic gestures was introduced to allow convenient interaction with the system when the type of user contribution cannot be generated, for example when the basic search term and method is selected from a visual presentation.</Paragraph>
    <Section position="1" start_page="211" end_page="212" type="sub_section">
      <SectionTitle>
3.1 Discourse and Knowledge Structures
3.1.1 Concepts and relations
</SectionTitle>
      <Paragraph position="0"> COR focuses on the dialogue pragmatics and omits semantic aspects in order to be domain-independent. As described above, RST can be used to introduce an additional level of coherence description between dialogue contributions. Additionally, some of the RST relations can be specialized in such a way that they represent relations between concepts that constitute the propositional content of dialogue acts.</Paragraph>
      <Paragraph position="1"> The propositional level was modeled and implemented in a completely object-oriented way. Concepts of all levels of knowledge (dependent as well as independent from the domain) were classified as classes, instances and attributes. The object-oriented approach makes use of the  CLASS-INSTANCE, and!INSTANCE-ATTRIBUTE.</Paragraph>
      <Paragraph position="2"> The ideational branch of Maier and Hovy's (1991) metafunctionally motivated taxonomy of RST relations made it particularly easy to augment the hierarchy, achieving a powerful model that uses relations as a unified way to structure knowledge.</Paragraph>
      <Paragraph position="3"> 3.1.2 Developing the dialogue Within the process of information seeking, the situations of the search - specifically failure or partial success - direct the dialogue in certain:ways. Three main decision criteria have been incorporate~d in the Corinna system: 1) the requested data must be available in the database, 2) the data must be presentable in such a way that it is of use for the information seeker, and 3) the data has to be relevant with respect to the expressed information need. Depending on these factors, a follow-up dialogue act can be selected. For example, an OFFER is only possible when availability and presentability of the data can be granted. If the data is relevant, a follow-up ACCEPT is possible. If, on the other hand, the information need is not matched by the offered information, the OFFER will have to be answered (i.e., evaluated) by a REJECT-OFFER.</Paragraph>
      <Paragraph position="4"> Currently, the system is mainly able to judge availability and presentability of data. It is able to reason about the user's information need only in a very limited way.</Paragraph>
      <Paragraph position="5"> However, this does not lead to a significant drawback, since the user is always given the choice between several alternatives, as for example between an ACCEPT and a REJECT-OFFER.</Paragraph>
      <Paragraph position="6"> Taking into account what has been said above about the integration of COR and RST, the process for the generation of new dialogue contributions also includes the of the sample dialogue treatment of RST relations. Given the propositional content of a dialogue act, the determination of a follow-up act is influenced by the RST relation which holds between the two acts. For example, if a relation of the type rneta-inforrna~ion (see section 2.2.4) has been chosen where contextual information has to be supplied, an appropriate proposition expressing the state or direction of the database search has to be composed. By this way sequences of coherent propositions for dialogue acts are produced.</Paragraph>
      <Paragraph position="7">  Typically, an information retrieval system has its main challenge in the treatment of very large amounts of data. The CORDIS databases contain information about some 14,000 projects and 10,000 organizations, persons and other entities.</Paragraph>
      <Paragraph position="8"> The object-oriented structure of Corinna enabled us to implement retrieval methods for several types of data storage - internally stored discourse knowledge and database contents represented externally - still having the benefit of a uniform programming interface. Thus, the dialogue manager is transparently accessing two distinct types of data, simplifying the design and implementation of the system significantly.</Paragraph>
    </Section>
    <Section position="2" start_page="212" end_page="213" type="sub_section">
      <SectionTitle>
3.2 Realizing Dialogue Acts
</SectionTitle>
      <Paragraph position="0"> So far the focus was on the underlying models of a dialogue. This is motivated by the fact that the described mechanisms for performing a dialogue based on COR and RST apply to any modality, not necessarily natural language. For the Corinna prototype we chose the linguistic form as interaction mode because it incorporates a high level of expressive power needed for Information Retrieval tasks. Also, it has significant advantages over other interaction modes, as far as the clarification of misunder- null ' 7th International Generation Workshop * Kennebunkport, Maine * June 21-24, 1994 standing and dialogue failure are concerned. Information types required in such dialogue states cannot be presented raphically - language is the preferable interaction mode f. also Dilley et al., 1992).</Paragraph>
      <Paragraph position="1"> The realization of dialogue contributions is divided into two major steps. First, the internal representation of a dialogue act is transferred into a high-level logical form that is capable of dealing with language-oriented attributes of an utterance, out of which English text is then automatically generated in the second step. For this task we used the PENMAN generator (PENMAN (1989))with the Sentence Plan Language SPL (Kasper and Whitney (1989)) as intermediate representation.</Paragraph>
      <Paragraph position="2">  We want to focus here on three main parameters influencing the process of realizing a dialogue act: illocntion, proposition and RST relation.</Paragraph>
      <Paragraph position="3"> It has been pointed out earlier that the propositional content and illocutionary point of dialogue acts are managed independently. During the design of Corinna, it became evident that also the realization process can take place quite independently, as far as the creation of the SPL plan is concerned. The flexibility needed for natural language utterances is achieved by combining various intermediate logical repreaentations of realizations for propositions and illocutions. Several typical linguistic attributes for the illocutionary acts in the COlt model can be assumed in order to allow management in the context of a limited computer program. For example, many forms of rtE-QUEST incorporate the use of imperative forms like &amp;quot;tell me&amp;quot;, &amp;quot;show me&amp;quot;, etc.</Paragraph>
      <Paragraph position="4"> In cases where contextual information is to be given for an assertive dialogue act the ltST relation can make significant contributions towards a coherent realization of the two utterances involved. For example, relations like PUrr-POSE or INSTRUMENT can be signaled by markers such as &amp;quot;in order to&amp;quot;, &amp;quot;for&amp;quot;, etc.</Paragraph>
      <Paragraph position="5"> Taking together 1) the fragmentary realization plan of the proposition on a purely assertive level, 2) the generated fragments for an illocutionary act of COlt, and 3) the RST relation that connects this act to the previous act, a suitable SPL statement for PENMAN can be composed.</Paragraph>
      <Paragraph position="6"> The final result is an English sentence, which is passed to the user interface.</Paragraph>
      <Paragraph position="7">  The following simplified example illustrates the various steps !n the planning of a contribution. We assume the scenario as given in figure 6 where the system has just asked for a SUPERTERM of &amp;quot;NL Generation&amp;quot;. In this example the user wants to know the reason of the request and the system gives the corresponding answer. The SPL statement for this utterance is given in figure 8. Three distinct types of information were used to compute this sentence plan: (1) The RST relation CAUSE was selected, since the user asked the question &amp;quot;Why? . (2) The domain of this relation is the verbal action that was performed by the system previously in the dialogue: The REQUEST for the more general term. (3) The motivation for making the utterance is the situation in the database - the search term is unknown. This fact is represented by a propositional item describing the state of the search. Note that the system's response is a dialogue act of type INFORM. As it has been stated earlier, the acts of type INFORM become the nucleus of the whole dialogue, which means that the (sub)dialogue takes on the role of an assertion after its completion. In case of a subdialogue, this assertion then carries the accompanying contextual information for another act one dialogue level up - in our case this is the REQUEST of the system, asking for a replacement term, and the contextual information is the CAUSE of that question. The RST relation plays then an important role in creating the coherence between the original question of the system and the contextual information, which is the reason, why it asked this question.</Paragraph>
    </Section>
    <Section position="3" start_page="213" end_page="214" type="sub_section">
      <SectionTitle>
3.3 Interaction with the system
</SectionTitle>
      <Paragraph position="0"> It has been pointed out previously that we aimed at having a mainly language-driven interface. Additionally to the generation of the system's utterances we decided to also generate the dialogue contributions of the user. That is, there is no Natural Language Understanding module in Corinna; instead, in each dialogue situation where it is the turn of the user, Corinna generates several suggested utterances, from which the user can then choose the one that suits her best in the context of the current dialogue situation. We believe that this does not pose a severe limitation, the dialogue model based on COlt-RST, as well as the semantic knowledge about the search in the database allow predictions about how the dialogue might be continued.</Paragraph>
      <Paragraph position="1"> Figure 9 shows a sample interaction. The screen is divided into two halves. In the upper one, the instantiated dialogue history is displayed and updated after each contribution. The lower half of the interface is reserved for presentation of the database queries and their results. In the dialogue state as represented by the snapshot, the system has just provided the table of results. The user then has the choice to select one out of three possible acts: BE-CONTENTED (&amp;quot;OK&amp;quot;), CONTINUE (&amp;quot;This is not sufficient&amp;quot;) and BE-DISCONTENTED (&amp;quot;Let's stop this dialogue&amp;quot;).</Paragraph>
      <Paragraph position="2"> Note that, depending on the situation of the dialogue, a participant - system or user - may utter two acts in a row, as opposed to simple system-user alternations. For example, after the user is contented with the answer of the system about why it asked for a term replacement,</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="214" end_page="214" type="metho">
    <SectionTitle>
AND 12:00:O0:000AM 12:00:00:O00AM
INTER~OGATION OF
KNOWLEDGE BASES
USING NATURAL
LANGUAGE TEXT
</SectionTitle>
    <Paragraph position="0"> (&amp;quot;Okay&amp;quot;), she then makes a contribution to the dialogue one level up - levels are represented by indentation of contributions - which was a term the system had asked for (&amp;quot;Natural Language&amp;quot;). A similar phenomenon can be observed in the last two utterances of the system. To decide whether the user is allbwed to utter two acts in a row, the system applies a set of (relatively simple) heuristics; they are giving the user preference in guiding the dialogue.</Paragraph>
  </Section>
  <Section position="6" start_page="214" end_page="214" type="metho">
    <SectionTitle>
4 Summary
</SectionTitle>
    <Paragraph position="0"> The Corinna system iis the result of an integration of two models: The CQR model and RST. Compared to the single theories the integrated model has a number of significant advantages: (1) The scope of RST could be extended from monologues to (information seeking) dialogues - without changing major assumptions of the original theory; (2) RST adds an additional level of coherence to the existing COR model. While COP,. focuses on illocutions, dialogue acts and their sequential order, PaST provides means to describe the semantics of the links between the single acts; (3) The description of dialogues in terms of COR-RST gives rise to a systematic approach to dynamically generate dialogue contributions for either of the participants. Corinna, as a prototypical implementation of the combined theories, uses PENMAN to generate natural language utterances. It accesses a database typical for the information retrieval task (CORDIS).</Paragraph>
    <Paragraph position="1"> Of course there are many possibilities for improvement.</Paragraph>
    <Paragraph position="2"> First, the flexibility and power of the generated utterances can be further increased. The plans for the dialogue contributions could be further parameterized, also the knowledge base covering the current situation of the search in the database needs additional concepts. On the side of Information Retrieval, more tactics could be implemented making the search more efficient.</Paragraph>
    <Paragraph position="3"> We think that the outlined theoretical framework is powerful enough to go beyond the state of the system as it is implemented so far, and we will continue this research especially in the context of concrete application areas.</Paragraph>
  </Section>
  <Section position="7" start_page="214" end_page="215" type="metho">
    <SectionTitle>
Acknowledgements
</SectionTitle>
    <Paragraph position="0"> We would like to thank Stefan Sitter for support and inspiring discussions. We also gratefully acknowledge the comments of three anonymous reviewers who made many helpful suggestions. Part of the work of the second author was funded by the German Ministry for Research and Technology (BMFT) under contract 01 IV 101 k/1 (V RBMOBI0.</Paragraph>
  </Section>
class="xml-element"></Paper>