<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1404">
  <Title>I I I I References</Title>
  <Section position="1" start_page="0" end_page="0" type="metho">
    <SectionTitle>
AN ARCHITECTURE FOR OPPORTUNISTIC TEXT GENERATION
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="2" start_page="0" end_page="28" type="metho">
    <SectionTitle>
Abstract
</SectionTitle>
    <Paragraph position="0"> We describe the architecture of the ILEX system, * which supports opportunistic text generation. In * web-based text generation, the SYstem cannot plan the entire multi-page discourse because the user's browsing path is unpredictable. For this reason, * the system must be ready opportunistically to take * advantage of whatever path the user chooses. We describe both the nature of opportunism in ILEX's museum domain, and then show how ILEX has been designed to function in this environment. The architecture presented addresses opportunism in both content determination and sentenceplanning.</Paragraph>
    <Paragraph position="1"> 1 Exploiting opportunities in text generation * Many models of text generation make use of standard patterns (whether expressed as schemas (e.g.</Paragraph>
    <Paragraph position="2"> \[McKeown 85\]) or plan operators (e.g. \[Moore and Paris 93\])) to break down communicative goals in such a way as to produce extended texts. Such models are making two basic assumptions:  1. Text generation is goal directed, in the sense that spans and subspans of text are designed to achieve unitary communicative goals \[Grosz and Sidner 86\].</Paragraph>
    <Paragraph position="3"> 2. Although the details Of the structUre of a text may have to be tuned to particulars of the  communicative situation, generally the structure is determined by the goals and their decomposition. That is, a generator *needs strategies for decomposing the achievement of complex * goals into sequences of utterances, rather than ways of combining sequences of utterances into more complex structures. Generation is &amp;quot;top-down&amp;quot;, rather than&amp;quot;bottom-up&amp;quot; \[Marcu 97\]. Our belief is that there is an important class of NLG problems for which these basic assumptions* are not helpful. These problems all involve situations where semi-fixed explanation strategies are less useful than the ability to exploit opportunities. WordNet gives the following definition of 0pportunity': Opportunity: &amp;quot;A possibility due to a favorable combination of circumstances&amp;quot; Because * opportunities involve *combinations of circumstances, they are often unexpected and hard to predict. It may be too expensive or impossible to have complete knowledge about them. Top-down generation strategies may not be able *to exploit opportunities (except at the cost of looking for all opportunities at all* points) because it is difficult to associate classes of opportunities with fixed stages in the explanation *process.</Paragraph>
    <Paragraph position="4"> We are investigating opportunistic text generation in the Intelligent Labelling Explorer (ILEX) project, which seeks automatically to generate a sequence of commentaries for items in an electronic  catalogue (or museum gallery) in such a way as to reflect the interest of the user and also to further certain educational (O r other)aims. The current domain of the system is the 20th Century Jewellery Exhibit in the Royal Museum of Scotland but ILEX is* designed to work with any domain where object descriptions are required. In ILEX, the system has an agenda of communicative., goals to achieve, which reflect the goals of the curators. The user has the freedom to look at any object in the gallery at any time. The system produces a description of each object asked for by the user, such that each description contributes as best it can to the system's goals and the sequenc e of descriptions fits together into a coherent whole. The result is a variety of mixed-initiative dialogue, in which the user is in control of the high-level communicative goal (what gets described) but the system is in control of how the goal is realised (how the chosen object is described).</Paragraph>
    <Paragraph position="5"> In such a dynamically unfolding environment, it is not possible to predict all possible paths through the interaction. The system must thus be ready to exploit opportunities in order to achieve its goals. In ILEX, the user's arbitrary choice represents a horizon beyond which is it not practical to predict. Each generated page may be the lastone to be generated and therefore has to be planned to achieve as much as possible on its own. Moreover, almost any part of the generated text can be optimised to exploit the arbitrary Situation that the user has got themself into.</Paragraph>
  </Section>
  <Section position="3" start_page="28" end_page="30" type="metho">
    <SectionTitle>
2 Opportunities: evidence and models
2.1 Evidence: the goals of a museum curator
</SectionTitle>
    <Paragraph position="0"> A museum curator seeks to achieve general educational goals through the description of a set of carefully selected objects. In general, the goals are to convey important generalisations (e.g.</Paragraph>
    <Paragraph position="1"> &amp;quot;Organic jewellery tends to have natural themes&amp;quot;) and to dispel important misconceptions (e.g.</Paragraph>
    <Paragraph position="2"> &amp;quot;Jewellery tends to be made of expensive materials&amp;quot;).* These important points have to be brought in appropriately during the description of the exhibits which are selected by the visitor.</Paragraph>
    <Paragraph position="3"> * In order to see how a human being tackles such complex goals, we performed a &amp;quot;CuratOr of Oz&amp;quot; experiment, * in which we chose an arbitrary sequence of exhibits:in the 20th Century Jewellery gallery of the National Museum of Scotland and asked the curator to give Us a commentary. The curator intro~luced general points/themes suggested by the Objects, moving from the objects to the general issues surrounding them, using the objects merely as an excuse to introduce these topics, for instance as in the following (&amp;quot;V&amp;quot; indicates the visitor and &amp;quot;C&amp;quot; the curator): V: &amp;quot;There's a set of three objects here?' C: &amp;quot;What these symbolise for me are the preoccupations of the 1980's with .... &amp;quot; She reinforced points from the past, exploiting an excuse to come back to an important point that has already been made and show its relevance in a new situation: _~ V: C: She also &amp;quot;This one here...&amp;quot; &amp;quot;Yes, you've made a link with the first piece that we looked at, which is the idea of a jewel which is also a work of art and a sculpture...&amp;quot; made links to previous items, thereby improving the continuity of the discourse: C: &amp;quot;... and it was work like this which directly inspired work like the Roger Morris brooch on the stand which we looked at earlier.&amp;quot; : 29 All of these can be regarded as ways of exploiting opportunities offered by the situation. There is nothing like a conventional schema structure to the descriptions produced. The approach looks a lot more like puttingtogether arbitrary pieces of interesting material subject to only very loose retrictions. This may not be the best way to produce a carefully-argued'written text, and clearly the result is not always fluent according to stringent criteria. In some but not all--respects, it resembles the Unplanned discourses discussed by \[Ochs 79\]. Furthermore, in the interactive and*relatively informal setting of a museum tour, it works.</Paragraph>
    <Paragraph position="4"> We thus decided that ILEX should have a whole set of goals about things tosay. These are linked into a single metalevel goal, which is something like &amp;quot;to achieve as many of the individual * goals as possible, within the space available, in the context of a globally coherent discourse which maintains the reader's interest&amp;quot;.</Paragraph>
    <Section position="1" start_page="28" end_page="30" type="sub_section">
      <SectionTitle>
2.2 Models: planning for opportunities *
</SectionTitle>
      <Paragraph position="0"> We discussed above why t0p-down planning seems an unnatural basis for formulating an NLG model that can exploit opportunities. In contrast, ILEX is inspired loosely by ideas from opportunistic planning \[Hayes-Roth and Hayes-Roth 79, Pryor 96\]. Key elements of this are:  vehicle is given orders to deliver various objects to various building sites, and needs to locate these objects at other sites. The system is opportunistic in that while the truck is working on one goal, it is always read Y to switch to another if an *object on its find-and-deliver list turns up. For example, if thetruck stops at one place to pick up a hammer, it may notice a saw, which is also on its list, * and thus pick it up and proceed to its delivery point.</Paragraph>
      <Paragraph position="1"> Pryor's planning occurs within a limited horizon--the robot only has certain knowledge in regards to the immediate location, and outside of that, the world is uncertain (objects are sometimes randomly moved between sites in the world). ILEX inhabits a world analogous in certain respects to PARETO's: .each*page is a site on the map, and itis up to us to find opportunities for realising our goals at each site. However, while in the truck world the system is in control of motion to the next site, in the museum, it is the user who chooses the next page. Conversely, while objects outside the truck's immediate vicinity may move autonomously, for ILEX, facts and their values do not change.</Paragraph>
      <Paragraph position="2"> Opportunisticplanning has similarities with a number of other approaches to planning. It * shares with incremental planning (used in NLG by \[Cawsey 92\]) the idea of starting to execute a plan before the plan is complete, andbeing prepared to repair the partial plan in the light of feedback. It shares with reactive planning the idea of being directed as much by the characteristics of the state of the world at execution time as by the pursuit of preconceived goals. However, unlike pure reactive planning it does acknowledge the need for explicit plans to be manipulated and it</Paragraph>
      <Paragraph position="4"> differs from many models of incremental planning in the extent to which the original plan can be diverted to exploit the characteristics of the world at execution time.</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="30" end_page="34" type="metho">
    <SectionTitle>
3 The ILEX architecture *
</SectionTitle>
    <Paragraph position="0"> To show how ILEX supports opportunistic * text generation, we will here outline the parts of the system and the operation of its text planning. Basically the ILEX task agenda at each point consists of the facts that the system knows which have not yet been conveyed to the user. Each of these 'tasks' has an opportunity value (its educational value, assumed interest to the reader and contribution to coherence). At each point of the discourse, we 'perform tasks' (include facts) which provide the highest opportunity gain.</Paragraph>
    <Section position="1" start_page="30" end_page="32" type="sub_section">
      <SectionTitle>
3.1 The Content Potential
</SectionTitle>
      <Paragraph position="0"> The facts of our knowledge base are interconnected in various ways, and to facilitate content selection and structuring, we organise the facts into a content potential - a graph of facts interconnected in terms of thematic and rhetorical relations. The content potential is an intermed.iary stage between the knowledge base and text, motivated in a similar way to DRSs \[Kamp 81\] by the desire explicitly to represent the selection of possible knowledge structures that can be reflected linguistically. ;As Figure 1 shows, the content potential forms a three-tiered structure of entities, * facts and relations. There are links between items in adjoining tiers, but no links within a tier or between entities and relations. We now discuss the three tiers in turn.</Paragraph>
      <Paragraph position="1">  Entities are the participants in facts (things and qualities in terms of Penman's Upper Model). Entities may be of two kinds: specific entities - such as an individual jewel or person; and gener/c entities - an entity representing some class of entities, such as Scottish jewellers, or art-deco brooches. Generic entities are treated essentially in the same way as specific entities in the content potential, for purposes such as the tracking of focus, anaphor generation, and so on.</Paragraph>
      <Paragraph position="2">  Facts represent the relations between entities, in both events (e.g., X made Y), and states (e.g., X owns Y). In ILEX, we have assumed that all facts are binary (simple relations between two entities), e.g., made-by(J-9999, K+-ng01) represents the fact that the designer King made item  J-9999. The binary assumption simplifies our architecture, allowing quicker text generation. At a later stage, we may allow more complex fact-representation. Complex sentences can be formed through aggregating together these binary facts. Each fact has the following fields: 2 Pred : The name of the Predicate connecting the two entities.</Paragraph>
      <Paragraph position="3"> Argl : The entity in the relationship which the fact is primarily about. For instance, &amp;quot;J-999 was designed by *Jessie King&amp;quot; is primarily about J-999, not about King.</Paragraph>
      <Paragraph position="4"> Arg2 : The other entity in the relationship. This is sometimes another thing (such as &amp;quot;Jessie King&amp;quot;) and Sometimes a quality.</Paragraph>
      <Paragraph position="5"> Various other fields exist which detail the polarity, defeasibility, interest, importance and assimilation Of the *fact. Facts representing general principles or negations of general misconceptions are expressed using generi c entities and can be included in a text just like any other facts. * 3.1.3 Relations * * Relation nodes represent relations between facts. Although based on conceptual relations, they  qualify as rhetorical in that only the Subset of relations that could explicitly be conveyed is in-Cluded in the content potential. Relations include Example, Concession, Amplification, Similarity, Contrast, &amp;quot;In that&amp;quot;, &amp;quot;In other words&amp;quot;, Specification, Whereas and While. Each relation has a nucleus and satellite (as in RST) as well usa set of precondition facts, which must be assimilated before the relation can be. There are no relations between relation-nodes in the content Potential at present. Relation-nodes only link fact-nodes.</Paragraph>
      <Paragraph position="6"> Relations in the content potential present a uniform interface as nodes connected to facts in the graph but we do not have a uniform theory of all the relations. Figure 2 shows a small subgraph of the content-potential , showing two Concession relations between facts.</Paragraph>
      <Paragraph position="7"> Most of the content potential is precompiled, though relevant negations and comparisons depend on the set of entities already encountered and have to be computed on demand, causing the addition of various consequent facts and relations.</Paragraph>
      <Paragraph position="9"/>
    </Section>
    <Section position="2" start_page="32" end_page="33" type="sub_section">
      <SectionTitle>
3.2 Content Determination
</SectionTitle>
      <Paragraph position="0"> ILEX plans a single page of text, describing a single entity, at a time. The content potential represents the information we can express, and the interconnectivity of information. When we receive the resquest for an entity description, the planner sets that entity as the global focus of the current page. Opportunistic planning then commences: The facts directly Connected to that entity represent opportunities: the system can coherently include these facts in the text. *If any of these facts are actually selected, then new opportunities are created in two ways: Entity-based moves: From the fact, we go to the argument which we didn't enter the fact from. We then select anew fact reachable from this node. See Figure 3. If we followed the Arg2 role of a fact, then we are in a sense selecting a new focus (local focus). The facts we generate about this entity should have the new entity as the * focus. Thus in the example, King becomes the Theme of the second sentence. Sentences introduced using entity-based moves can be realised using an Elaboration relation to the starting fact.</Paragraph>
      <Paragraph position="1"> An entity:based move from an individual entity to its generic class *entity can be made once the appropriate &amp;quot;isa&amp;quot; fact has been selected: This item is an organic jewel.</Paragraph>
      <Paragraph position="2"> Organic jewels tend to be ...</Paragraph>
      <Paragraph position="3"> * Relation-based moves: from the initial fact, we follow a relation-node to some new fact. The new fact will be realised textually as a satellite to the original fact's nucleus. The type Of the relation-node will determine the rhetorical relation of the link. See Figure 4.</Paragraph>
      <Paragraph position="4"> Once we select a new fact in either of the ways described above, the new fact may act as the * starting point for new opportunistic expansion. Alternatively, we may decide to backtrack to some earlier point, effecting a focus pop in Grosz and Sidner's \[Grosz and Sidner 86\] terms.</Paragraph>
      <Paragraph position="5"> The selection of which opportunity to explore is determined by a *number of heuristic factors. Firstly, facts are weighted according to the chain of relations back to the focus of the page \[O'Donnell 97\]. This is a way of preventing lengthly digressions from the supposed topic of the text. Secondly, each fact is associated with numbers which represent the opportunity 'value' of the fact. The opportunities are of two kinds: Interest. the estimated value of the fact to the user, e.g. being made of plastic or paper are more interesting (to the user), because they are unusual in jewellery. Canned anecdotes about a piece of jewellerY may also have high interest values.</Paragraph>
      <Paragraph position="6">  Importance. * the value of the fact as regards the system's educational agenda, e.g., the system * considers it important to educate on stylistic development, so facts about styles are rated *highly.</Paragraph>
      <Paragraph position="7"> These values are moderated by a third fact annotation: Assimilation. the degree to which the fact is assumed known to the user, either from general knowledge, or through prior mentions in the web interaction (these values change dynamically). null The three values interest, importance and (1 - assimilation) are multiplied together to calculate the local score of each fact. The overall opportunity value of a fact is the product of the local score of the fact, the overall opportunity value of the parent (the node through which it was reached) and a weight for the relation between them. It is the overall opportunity values that axe used to select which textual opportunities to follow. We have no special theory about where interest and importance come from, though the above examples suggest that there may be domain- and user-type-specific rules that can be used to derive some of them.</Paragraph>
      <Paragraph position="8"> In Summary, content-determination in ILEX is seen as the task of optimising the selection of opportunities that are offered by the topic of the text, subject to not moving too far from that topic. The result of content-determination is a connected subgraph of the content potential (Figure 5). The use of interest and importance in ILEX is analogous to theuse of &amp;quot;salience&amp;quot; in \[McDonald and Conklin 82\]. Because the process is seen as a graph traversal problem, there are als0 similarities with work on generating text from semantic networks \[Simmons and Slocum 72, Sibun 92\]. In a sense, our work aims to combine the best of both.</Paragraph>
    </Section>
    <Section position="3" start_page="33" end_page="34" type="sub_section">
      <SectionTitle>
3.3 Text Planning
</SectionTitle>
      <Paragraph position="0"> Although the process of content determination has worked through a number of moves that may be made in the generated text, the result is not the kind of tree structure that one needs for realisation and also has been influenced only by local considerations of coherence. Text planning therefore requires the following two steps: I. Extend the subgraph to a complete subgraph that includes all the relations linking the selected fact nodes.</Paragraph>
      <Paragraph position="1">  2. Produce from this an &amp;quot;optimal&amp;quot; selection of relations, so as to give rise to an RST* tree * including all the selected facts~  The idea of combining a set of facts together into an &amp;quot;optimal&amp;quot; text is compatible with \[Hovy 90\] and the earlier work of \[Mann and Moore 81\]. Again this involves exploiting opportunities. For in*stance, in order to avoid an awkward focus shift at some point, one might attempt to include a selected fact about a new entity immediately after another one that mentions the same entity. Other text * planning operations that are opportunistic in nature include aggregation \[Dalianis and Hovy 96\] and redundancy suppression \[McDonald 92\], though we will not consider these here. The second step described above is exactly that described by \[Marcu 97\]. That is, one is given a set of facts all of which should be included in a text and a set of relations between facts, some of which can be included in the text. The task is to produce a legal RS tree using the facts and some relations (or the &amp;quot;best&amp;quot; such tree). Marcu's approach first of all attempts to find the best ordering 0f the facts. For every relation that could be indicated, constraints are generated saying what the order of the two facts involved should be and that the facts should be adjacent. The Constraints are weighted accord!ng to attributes of rhetorical relations that have been determined empirically. A standard constraint satisfaction algorithm is used to find the linear sequence such that the total weight of the satisfied constraints is maximal. Once the sequence of facts is known, a general algo~thm is used to construct all possible RS trees based on those facts.</Paragraph>
      <Paragraph position="2"> We could use Marcu's methods directly, but are exploring more widely because we would like to take into account a wider range of preference criteria, develop algorithms that treat entity-based elaborations rather differently from other rhetorical relations \[Oberlander et al 98\] and investigate heuristic approaches that wilt scale up better. We are currently experimenting with *three different algorithms for building an RST tree. These are all opportunistic in nature, rather than being strongly goal-directed or schema-based:  1. The RS tree (realised depth-first) is built to directly reflect the tree of nodes explored (breadth=first) in the content potential.</Paragraph>
      <Paragraph position="3"> 2. The best trees up to a fixed depth using relational moves are constructed; these are &amp;quot;glued ''* together using entity based moves \[Oberlander et al 98\].</Paragraph>
      <Paragraph position="4"> 3. A genetic algorithm is used to search for a legal tree that is of as high quality as possible  \[Mellish et al 98\].</Paragraph>
      <Paragraph position="5"> The current version of ILEX, which is being prepared for evaluation, uses the second of these algorithms and generates context-dependent descriptions for 32 different items of modern jewellery</Paragraph>
      <Paragraph position="7"> This jewel is a bracelet and is in the Organic style. It draws On natural themes for inspiration, in that it is a remarkably fluid piece. Indeed Organic style jewels usually draw on natural themes for inspiration; for instance this jewel is inspired by forms found in natural wood, in that it has a bracelet with a twig-like appearance. It resembles the Arts and Crafts style necklace, in that like the *necklace it is made from silver metal. However this jewel differs* from the necklace, in that it was made by Gerda Flockinger, whereas the necklace was made by Arthur and Georg~e Gaskin ....</Paragraph>
      <Paragraph position="8"> Organic style jewels differ from Art Deco style jewels, in that they are usually * made up of asymmetrical shapes, whereas Art Deco style jewels usually use geometric forms.</Paragraph>
      <Paragraph position="9"> Other jewels in the Organic style include...</Paragraph>
      <Paragraph position="10">  (the non-demo Version deals with 120). *Descriptions of different lengths can be obtained (for the evaluation, the system generates on demand 4 or more pages of about 10 clauses each for each item). Figure 6 Shows part of a relatively long description generated.</Paragraph>
    </Section>
    <Section position="4" start_page="34" end_page="34" type="sub_section">
      <SectionTitle>
3.4 ILEX and Opportunistic Planning
</SectionTitle>
      <Paragraph position="0"> Withthis description of ILEX in mind, we can explore the analogy with PARETO in more detail.</Paragraph>
      <Paragraph position="1"> Where PARETO embarks on the execution of a sketchy plan *to start moving around the truck * world, ILEX embarks on a graph traversal, starting out from the topic entity and guided by the desire not to digress excessively~ Thecontent potential offers options to ILEX in a similar way to PARETO's world. In PARETO, reference features indicate possible opportunities; in ILEX this role is played by the interest and importance annotations. Deeper analysis is require d by PARETO before seizing an opportunity; this is *probably analogous to the way that ILEX attempts to find th e globally best way of incoporating material into the RST tree.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>