File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/00/j00-2001_metho.xml
Size: 77,186 bytes
Last Modified: 2025-10-06 14:07:13
<?xml version="1.0" standalone="yes"?> <Paper uid="J00-2001"> <Title>Integrating Text Planning and Linguistic Choice Without Abandoning Modularity: The IGEN Generator</Title> <Section position="5" start_page="110" end_page="110" type="metho"> <SectionTitle> Rubinoff The IGEN Generator </SectionTitle> <Paragraph position="0"> expressed as an adjective. This is because the only available adjective is the somewhat awkward water-going, and it turns out that in the sentences TEXT happens to generate, WATER only appears in an adjective context in cases where it's not important to say it. Thus, this strategic decision (that WATER can be omitted) is encoded permanently into the tactical component; there is no way for the strategic component to control this decision.</Paragraph> <Paragraph position="1"> In MUMBLE (Meteer et al. 1987), in contrast, the interactions have pushed linguistic information into the text planner. For example, MUMBLE can take the already constructed phrase Fluffy, Floyd's dog, buries bones and modify it with the information that this was reported by Helga to produce Helga reported that Fluffy, Floyd's dog, buries bones. But in order to do this, it has to mark the information about Helga with the new-main-clause feature. So the planner has to know what clauses are, know that the earlier information was turned into a clause, and know that making Helga reports a new main clause is a useful thing to do. Many of the linguistic decisions are thus actually being made by the planner.</Paragraph> <Section position="1" start_page="110" end_page="110" type="sub_section"> <SectionTitle> 2.2 Facing the Dilemma </SectionTitle> <Paragraph position="0"> We are now faced with a dilemma. On the one hand, the separation of planning and linguistic realization into distinct components seems natural and useful. On the other hand, it precludes making decisions involving interactions between text planning and linguistic issues.</Paragraph> <Paragraph position="1"> One possible response would be to abandon the separation; the generator could be a single component that handles all of the work. This approach has occasionally been taken, as in Kantrowitz and Bates (1992) and Danlos (1987) 4 and, at least implicitly, in Paris and Scott (1994) and Delin et al. (1994); however, under this approach, all of the flexibility and simplicity of modular design is lost.</Paragraph> <Paragraph position="2"> The opposite approach is to simply ignore the limitations of a modular design and proceed as if there need be no interactions between the components. Whatever problems result will be handled as best they can, on a case-by-case basis. This approach is the one taken (implicitly or explicitly) in the majority of generators. In fact, Reiter has even argued in favor of this approach, claiming that the interactions are sufficiently minor to be ignored (or at least handled on an ad hoc basis) (Reiter 1994). While this certainly has appeal as a design methodology, it seems reckless to assume that problems will never appear. Certainly an approach to generation that does handle these interactions would be an improvement, as long as it didn't require abandoning modularity. There have in fact been attempts to develop modified modular designs that allow generators to handle interactions between the components. These include devices such as interleaving the components (McDonald 1983; Appelt 1983), backtracking on failure (Appelt 1985; Nogier 1989), allowing the linguistic component to interrogate the planner (Mann 1983; Sondheimer and Nebel 1986), and Hovy's notion of restrictive (i.e., bottom-up) planning (Hovy 1988a, 1988c). All of these approaches, though, require that potential interactions be determined either by the tactical component or by the system designer in advance. The text planning component still has no way to detect and respond to unanticipated interactions on its own initiative. 5</Paragraph> </Section> </Section> <Section position="6" start_page="110" end_page="114" type="metho"> <SectionTitle> 4 Danlos still has a separate low-level &quot;syntactic&quot; component, but essentially all of the generator's decisions are made by the strategic component. 5 In fact, adding additional components may make the problem even worse, as decisions may then be spread across three or more separate components. It might seem that a component placed between the </SectionTitle> <Paragraph position="0"> planner and the realizer could handle interactions between the levels, but this would simply recreate the original problem within the &quot;middle&quot; component; coping with the interactions would still require dealing with the whole range of issues that come up in both planning and realization.</Paragraph> <Paragraph position="1"> Computational Linguistics Volume 26, Number 2 The IGEN generator resolves the modularity dilemma by means of a new approach that preserves complete modularity but allows the generator to handle interactions between the components. The key to IGEN's approach is that the linguistic component provides the planner with an abstract description of the effects of using particular linguistic structures. This abstract description is encoded in annotations that the linguistic component attaches to the structures it builds. In contrast with previous feedback-based approaches, in which the feedback only occurs when the linguistic component needs additional help from the planner, IGEN annotates all of its linguistic structures. This allows the planner to detect situations where linguistic choices cause problems at the planning level; the linguistic component need not be aware of them.</Paragraph> <Paragraph position="2"> Conversely, the planner can build communicative plans without concern for whether the linguistic component can actually express the elements of the plan; any such problems will be detected by the linguistic component and reflected in the annotations.</Paragraph> <Paragraph position="3"> The annotations allow a cycle of negotiation between the planner and the linguistic component; each component works entirely within its own level, drawing only on the results of the work at the other level.</Paragraph> <Paragraph position="4"> 3. IGEN: A New Model for Generation The IGEN generator overcomes the limitations of the modular approach, without giving up its advantages, by the use of annotations that provide feedback from the linguistic component to the planner. These annotations abstract away from the details of the linguistic expressions, allowing the planner to detect interactions between decisions at the linguistic and planning levels without having to know anything about the knowledge or reasoning used by the linguistic component. The planner can then respond to these interactions by choosing among the expressions proposed by the linguistic component, using the annotations to evaluate how particular linguistic choices affect the successful achievement of the goal(s) driving the generator. This approach allows IGEN's components to interact while preserving the strict separation between their knowledge and processing.</Paragraph> <Paragraph position="5"> The annotations are predicates that apply to the linguistic expression they are attached to, indicating some effect of or property of the expression. Most of them also take an explicit argument indicating a semantic expression whose relation to the option is indicated by the annotation. The types of annotations used by IGEN are shown in Figure 1. They fall into the following general categories: 6 Meaning-based These relate the meaning of a linguistic option to the information it is supposed to express, both in terms of how explicitly and how completely the information is expressed.</Paragraph> <Paragraph position="6"> Contextual These indicate dependencies and effects on the presence and status of elements in the discourse context, including not only availability for anaphoric reference but also effects on how various elements in the discourse context will be perceived.</Paragraph> <Paragraph position="7"> Pragmatic These indicate various pragmatic and stylistic features of the option. 7 6 There are also some annotation types used to keep track of IGEN's internal processing; these are not shown here. 7 Note that IGEN currently only uses one annotation of this type: concise-construction, used to mark constructions that are particularly concise.</Paragraph> <Paragraph position="8"> Annotation types.</Paragraph> <Paragraph position="9"> The IGEN generator uses these annotations to allow for interactions between its various components, shown in Figure 2. IGEN works as follows: The communicative planner responds to goals given to the generator by building an appropriate plan, which will contain a number of communicative actions, i.e., actions that communicate information, and places requests for ways to express that information into the workspace. The linguistic component then responds by producing a series of annotated linguistic options for each of the planner's requests. These are also placed in the workspace, where the planner can use the annotations to evaluate and rank the options and also, when appropriate, to modify the plan based on what the linguistic component is able to produce.</Paragraph> <Paragraph position="10"> This process doesn't actually produce any output, though, since there is no way to determine when it is &quot;finished.&quot; The linguistic component can always come up with yet another way to express something, and the planner can always do more reasoning about the interactions between some expression and the planner's specific and general goals. There must therefore be another component, here called the utterer, that is responsible for assembling utterances from the options preferred by the planner and shipping them off to be spoken or written. The utterer's job is to balance the planner's preferences against time pressure, producing output when whatever deficiencies the options have are outweighed by the consequences of remaining silent. The various components thus work together in parallel, incrementally building, improving, and producing utterances. Each component of the generator handles a particular aspect of the generation task, involving different kinds of knowledge and different constraints on the generator's work.</Paragraph> <Paragraph position="11"> IGEN in some ways resembles a blackboard architecture (Nii 1986b, 1986a); like a blackboard, its workspace contains several possible utterances that are described at different levels of structure and are built up incrementally by various independent components. On the other hand, it doesn't have the strict hierarchical structure that usually characterizes blackboard systems; each request from the planner may correspond to several linguistic options, and the linguistic options may handle (parts of) more than one request. Also, blackboard systems usually have sequential scheduling of actions, whereas IGEN's planner, linguistic component, and utterer all run in parallel. Furthermore, the relevant issue here for IGEN is whether there are real interactions between the modules, which a blackboard-style approach to generation doesn't Architecture of the IGEN generator.</Paragraph> <Paragraph position="12"> guarantee. Blackboard-based generation can still be structured in a way that makes such interactions impossible, either explicitly, as in DIOGENES (Nirenburg et al. 1988), which uses several blackboards successively, or implicitly, as in GLINDA, in which successive stages of rules depend on the results of previous stages for their triggers. 8</Paragraph> <Section position="1" start_page="113" end_page="114" type="sub_section"> <SectionTitle> 3.1 The Communicative Planner </SectionTitle> <Paragraph position="0"> This is a special-purpose planner that accepts goals from the overall system and plans out ways to achieve them via communication. These can be &quot;communicative&quot; goals such as &quot;transmit this information to the user&quot; (or more precisely &quot;get the user to believe this information&quot;), but they can also be social interaction goals such as &quot;establish a deferential (or collegial, or superior) attitude toward the user&quot; or cooperative behavior goals such as &quot;be straightforward&quot; or &quot;transmit as little information as possible&quot; 8 As can be seen from the diagram on page 24 of Kantrowitz and Bates (1992) and the accompanying discussion.</Paragraph> </Section> </Section> <Section position="7" start_page="114" end_page="118" type="metho"> <SectionTitle> Rubinoff The IGEN Generator </SectionTitle> <Paragraph position="0"> or even more general goals such as &quot;get the user to take his umbrella&quot; or &quot;convince the user to improve his diet.&quot; The communicative planner draws on a set of specialized plans that capture how communication acts can achieve various goals. It draws on knowledge of the effects of communication to plan out how to achieve various kinds of goals by expressing various kinds of information. Since it has to evaluate how well particular linguistic expressions further those goals, the planner needs to build an explicit representation of its plan(s) that indicates how particular actions contribute to their achievement. The planner depends on the linguistic component to tell it how to carry out these actions, i.e., how to express particular information in language.</Paragraph> <Paragraph position="1"> While the planner is specialized for planning communicative actions, it can in principle draw on anything in the system's general body of knowledge and belief that might support the goals it is given. Thus (as we shall see) in the second example in Section 4.3, the planner decides to mention the rain not because it bears any special relationship to anything in the input goal, but rather because mentioning it is part of a plan that achieves the goal. Even in cases where the planner's goal is simply to transmit some specific information, it can include any additional information that will help that transmission to succeed. The planner's role is not just to organize information, or even to collect information on some topic and organize it, but rather to identify a set of communicative acts that will achieve some goal.</Paragraph> <Paragraph position="2"> 3.1.1 Building an Initial Plan. The communicative planner starts off by building an initial plan to achieve the goals it is given. The form of that plan is important. It cannot be just a (partially or totally) ordered set of information to go into the text, as is the case in many generators. Since IGEN's planner needs to evaluate how well linguistic options achieve the plan's goals, the plan must record what each piece of information in the plan is intended to accomplish and how it supports the purpose of the larger text containing it. Sometimes the planner's goal may simply be to convey certain information to the user, but there may also be more indirect goals that conveyance is intended to achieve. Thus the plan must indicate in detail why the information is being expressed. An example of the type of plan the planner must build is shown in Figure 3 (Section 4.1 discusses the use of this plan in more detail). This model of text plans meshes well with the model of discourse structure developed by Grosz and Sidner (Grosz and Sidner 1985, 1986), in which the purpose of each discourse segment is an important part of the structure. 9 IGEN constructs its plans using a hierarchical planning algorithm (Nilsson 1980).</Paragraph> <Paragraph position="3"> The planner first checks all of its top-level plans to see which have effects that match the goal. Each matching plan's preconditions are checked; if they are currently (believed to be) true, the planner then attempts to find all instantiations of the plan's body. 1deg The body of a plan can be an action or sequence of actions, a goal or sequence 9 Moore and Paris also note that &quot;a generation system must maintain the kinds of information outlined by Grosz and Sidner&quot; (Moore and Paris 1989, 203). Their planner uses plan structures similar to IGEN's, except that the plan operators they use are generally instantiations of rhetorical relations drawn from Rhetorical Structure Theory (Mann and Thompson 1987). In IGEN, the plans can involve any goals or actions that could be achieved via communication.</Paragraph> <Paragraph position="4"> Hovy has described another text planner that builds similar plans (Hovy 1988b). This system, however, starts with a list of information to be expressed and merely arranges it into a coherent pattern; it is thus not a planner in the sense used here (as Hovy makes clear). 10 Since text planning was not the primary focus of this work, IGEN is designed to simply assume that any false preconditions are unattainable. IGEN's planner divides the requirements of a plan into two parts: the preconditions, which are not planned for, and those in the plan body, which are. This has no Support (2) by performing: Utter(\[first day's information\]) Support (2) by performing: Utter(\[second day's information\]) Support (2) by performing: Utter(\[last day's information\]) Figure 3 A sample initial plan.</Paragraph> <Paragraph position="5"> of goals, a subplan, or a goal decomposition. 11 To instantiate the body, the planner accordingly collects the action(s), recursively plans for the goal(s), instantiates the subplan, or carries out the goal decomposition. The net result of this process is a list of possible complete plans that achieve the original goal.</Paragraph> <Paragraph position="6"> up an initial plan of what it wants to say, it puts requests for the individual bits of information contained in the plan into the workspace for the linguistic component to respond to. The planner, however, cannot assume that its requests will be followed precisely; the linguistic component may not be able to find a way of completely satisfying them. The planner must therefore examine and evaluate the responses that the linguistic component puts into the workspace (as described below in Section 3.2).</Paragraph> <Paragraph position="7"> These responses consist of linguistic structures together with the annotations that provide descriptions of how closely the option carries out the request(s) it responds to, as well as any other information or effects arising from the use of the option. Drawing on these annotations, the planner evaluates how well the various options fit into the plan and assigns them appropriate ratings along a scale from VERY-HIGH to VERY-L0W.</Paragraph> <Paragraph position="8"> The evaluation algorithm is divided into two phases (as shown in Figure 4), which work roughly as follows: The first phase looks at how well the option matches the request, checking first to see whether the option misses any requested information; if there is no missing information, the rating is based on annotations indicating the relationship between the option's meaning and the request. The second phase then adjusts the option's rating based on how any missing or added information interacts with the plan structure. This ability to adjust preferences among linguistic options based on planning issues allows IGEN to overcome the usual limitations of a modular generator design. Without the annotations (or some similar feedback mechanism), the planner would have to supply the linguistic component with any information that might be relevant to choosing among the options. This could potentially include practical effects, since the preconditions, by and large, are types of goals that IGEN doesn't have any plans for; any attempt to plan for them would immediately fail. 11 A goal decomposition is a way to decompose a goal into subgoals that can be planned for independently. For example, the utter-sequence plan works by decomposing the information to be uttered according to some relevant ordering.</Paragraph> <Paragraph position="9"> Rubinoff The IGEN Generator EVALUATE(Option, Request) Phase 1: 1. Let Covered-Pieces = UNION(Request, arguments of any COVERS-OTHER-ENTRY annotations).</Paragraph> <Paragraph position="10"> 2. If Option has a MISSING-INFO annotation whose argument is a member of Covered-Pieces, or has more than one MISSING-INFO annotation, evaluation is : VERY-LOW.</Paragraph> <Paragraph position="11"> 3. Else if Option has any MISSING-INFO annotations, evaluation is :LOW.</Paragraph> <Paragraph position="12"> 4. Else if Option has any INDIRECTLY-SUGGESTS annotations, evaluation is : MEDIUM.</Paragraph> <Paragraph position="13"> 5. Else if Option has any MA~S-IMPLICIT annotations, evaluation is : HIGH.</Paragraph> <Paragraph position="14"> 6. Else evaluation is :VERY-HIGH (because Option must have a MAKES-EXPLICIT annotation).</Paragraph> <Paragraph position="15"> Phase 2: 1. Let Adjustment = 0.</Paragraph> <Paragraph position="16"> 2. For each MISSING-INFO annotation attached to Option: Subtract 1/2 from Adjustment if its argument is an item in the plan (1 if it's a critical item).</Paragraph> <Paragraph position="17"> 3. For eachACTIVATES-IN-CONTEXT or EXTRA-INFO annotation attached to Option: * If its argument is a critical item in the plan, add I to Adjustment.</Paragraph> <Paragraph position="18"> * Else if its argument supports or conflicts with a critical item in the plan or strengthens or weakens a goal or precondition of the plan or the head of the request, add or subtract 1 from Adjustment accordingly.</Paragraph> <Paragraph position="19"> * Else if its argument is a sequence, Request is a member of another sequence that is a critical item in the plan, and the two sequences are along the same scale but are disjoint or only overlap slightly, subtract 1 from Adjustment. For example, the sequences \[1, 2, 3, 4\] and \[4, 5, 6, 7\] would trigger this adjustment, but the sequences \[1, 2, 3, 4\] and \[2, 3, 4, 5\] wouldn't.</Paragraph> <Paragraph position="20"> * Else if its argument is an element in the plan, add 1/2 to Adjustment.</Paragraph> <Paragraph position="21"> 4. Adjust Option's Phase I evaluation up or down Adjustment levels (ignoring 1/2s), e.g. an Adjustment of -2 or - 21/2 would change : HIGH to : LOW, and add an ADJUSTED-UP or ADJUSTED-DOWN annotation to Option as appropriate.</Paragraph> <Paragraph position="22"> Note: a &quot;critical&quot; item in a plan is one that plays a central role in the structure of the plan, e.g. the basis of a goal decomposition or an action, goal or plan sequence. Figure 4 The linguistic option evaluation algorithm.</Paragraph> <Paragraph position="23"> anything in the plan, so the linguistic component would have to understand and reason about every aspect of the plan, defeating the whole purpose of a modular architecture.</Paragraph> <Paragraph position="24"> In addition to evaluating options, the planner looks for opportunities to revise the plan based on the work of the linguistic component. For example, the same linguistic structure may appear as an option for several different parts of the plan, or an option for one part of the plan may also express information contained in (or related to) Computational Linguistics Volume 26, Number 2 another part of the plan. The planner can reorganize the plan to take advantage of these kinds of situations. 12 Thus the communicative plan can be modified based on linguistic information, again without violating the strict modularity of the generator.</Paragraph> <Section position="1" start_page="117" end_page="118" type="sub_section"> <SectionTitle> 3.2 The Linguistic Component </SectionTitle> <Paragraph position="0"> The linguistic component is the part of IGEN that produces the actual bits of language that make up the utterance. It acts like a &quot;consultant&quot; to the communicative planner, providing it with linguistic expressions that attempt to capture part or all of the information the planner wants to express. 13 The linguistic component monitors the workspace, looking for requests from the planner, to which it responds by suggesting linguistic expressions that capture some or all of the information in the request. These expressions are placed in the workspace as options for the planner to evaluate; the linguistic component continues suggesting options for a request as long as that request remains in the workspace. 14 The linguistic options are pieces of surface structure; they can be at any level of grammatical structure and need only be partially specified. Thus an option could be a full sentence, phrase, or word, or a particular clause structure (e.g., a topicalization or an/t-cleft) with no further detail filled in, or a noun phrase with a determiner but with the head noun left unspecified. The options are represented as feature sets indicating the various syntactic and lexical properties of the option. 15 For example, the following feature set represents the partial phrase &quot;It will be ADJP&quot; in which the predicate adjective has not yet been specified:</Paragraph> <Paragraph position="2"> The FRED role in this phrase has not been filled yet, so its value is ADJF, indicating that it must be filled by an adjective phrase.</Paragraph> <Paragraph position="3"> The linguistic component is also responsible for annotating the options it puts into the workspace. The annotations are derived primarily from three sources: a comparison of the structure and position of the option's meaning and the request in IGEN's semantic network, the induced perspective shift of the option's meaning, and special annotations associated directly with the option. 16 The annotation algorithm works roughly 12 Another possibility would be for the planner to revise the plan when it detects problems with the options. For example, in a tutoring situation, if the linguistic component only returned options that involve concepts the student doesn't yet understand or know about, the planner might decide to revise the plan (perhaps to simplify what it's trying to say). IGEN doesn't implement this kind of plan revision, although the annotations provide the planner with the information necessary to do so.</Paragraph> <Paragraph position="4"> 13 IGEN's design allows for multiple linguistic components, each providing a different kind of linguistic knowledge, although only one was actually used; see Rubinoff (1992) for details.</Paragraph> <Paragraph position="5"> 14 As currently implemented, the linguistic component finds options by scanning through IGEN's semantic network, looking for items that are part of the meaning of some expression (Rubinoff 1992). The particular method used to produce the options, though, is incidental to the the overall functioning of the generator; it has no effect on the annotation or evaluation of the options, as the annotations depend only on the relationship between the option's meaning and the request.</Paragraph> <Paragraph position="6"> 15 There is also a MAP feature used to connect unfilled syntactic arguments with the corresponding semantic objects.</Paragraph> <Paragraph position="7"> 16 The induced perspective shift of the option's meaning is computed using a model of perspective built into IGEN's knowledge representation formalism; see Rubinoff (1992) for details.</Paragraph> </Section> </Section> <Section position="8" start_page="118" end_page="120" type="metho"> <SectionTitle> Rubinoff The IGEN Generator </SectionTitle> <Paragraph position="0"> as follows: First the linguistic component adds an annotation that indicates the kind of connection (if any) between the option's meaning and the request; this produces either a MAKES-EXPLICIT, MAKES-IMPLICIT, INDIRECTLY-SUGGESTS, or MISSING-INFO annotation. Next it adds EXTRA-INFO and MISSING-INFO annotations to indicate the differences between (the meaning of) the option and the request, and COVERS-OTHER-ENTRY annotations for any other requests that are included in the option's meaning. Finally, it adds ACTIVATES-IN-CONTEXT annotations to indicate any associated perspective shifts.</Paragraph> <Paragraph position="1"> For an example, consider the situation described in Section 2.1 where the generator is trying to express the information that John shot and killed someone. 17 Among the annotated options produced to express #<SHOOT-AND-KILL (JOHN, BILL)> are:</Paragraph> <Paragraph position="3"> In each of these cases, there is a MAKES-EXPLICIT annotation for the meaning of the word, a MAKES-IMPLICIT annotation for the requested information, and a MISSINGINF0 annotation for the property of the request that is not covered by the option. (Of course, MAKES-EXPLICIT annotations on the options that produce with a gun and dead indicate how the missing information can be added to the utterance.) The full annotation algorithm is as follows: 18 meaning and collect them in Annotations.</Paragraph> <Paragraph position="4"> If Option's meaning contains Request and nothing else, return Annotations. If Request is an instance of an element in Option's meaning and they have the same label, add a MAKES-EXPLICIT annotation for Request to Annotations.</Paragraph> <Paragraph position="5"> Else if Request is an instance, subconcept, or subrange of an element in Option's meaning or a position within (or identical to) an element in Option's meaning, then add a MAKES-IMPLICIT annotation for Request, MISSING-INFO annotations for any links connected to Request that don't match links connected to the element in Option's meaning, and EXTRA-INFO annotations for any links connected to the element in Option's meaning that don't match links connected to Request.</Paragraph> <Paragraph position="6"> Else if an element in Option's meaning is an instance or subconcept of Request, add a MAKES-IMPLICIT annotation for Request and MISSING-INFO annotations as described in step (4) to Annotations.</Paragraph> <Paragraph position="7"> instances or subconcepts of the same element) or an &quot;uncle&quot; of Request (i.e., the sibling of an element of which Request is an instance or subconcept), add a MAKES-IMPLICIT annotation for Request and MISSING-INFO and EXTRA-INFO annotations as described in step (4) to Annotations.</Paragraph> <Paragraph position="8"> Else if there is any other link between Request and an element in Option's meaning, add an INDIRECTLY-SUGGESTS annotation for Request to Annotations. Else add a MISSING-INF0 annotation for Request to Annotations.</Paragraph> <Paragraph position="9"> If Option's meaning contains a node or link that is another part of the planner's request, then add a COVERS-0THER-ENTRY annotation for the workspace entry containing the request to Annotations, and repeat steps (3) through (8) for that request.</Paragraph> <Paragraph position="10"> Add ACTIVATES-IN-CONTEXT annotations for every element that would be shifted into perspective by an element in Option's meaning to Annotations.</Paragraph> <Paragraph position="11"> Add any special annotations associated with Option to Annotations.</Paragraph> <Section position="1" start_page="119" end_page="120" type="sub_section"> <SectionTitle> Return Annotations. 3.3 The Utterer </SectionTitle> <Paragraph position="0"> The utterer is responsible for the final assembly and output of the utterance. Its main concern is balancing the desire for the most appropriate possible utterance against time pressure. Since the planner and the linguistic component build up and refine the utterance incrementally, they can continue working indefinitely; there's never a point where they can declare the utterance &quot;finished&quot; and quit working. On the other hand, IGEN can't go too long without uttering something or it will lose its turn, because the user will either utter (i.e., type) something else or get tired of waiting and leave (or quit the program). So the utterer must strike a balance between the advantage of improvements to the utterance that may yet be found and the disadvantage of continued delay.</Paragraph> <Paragraph position="1"> The actual work of the utterer is fairly straightforward. It constantly monitors the workspace, checking the available options against a time-sensitive acceptability level.</Paragraph> <Paragraph position="2"> Whenever the best available utterance meets or exceeds the minimum acceptable evaluation, the utterer proceeds to output the utterance. The criterion of acceptability is determined via a set of delay thresholds that control how long to wait before lowering the minimum acceptable rating. For example, the default threshold settings are \[0,1,2,12,16\], meaning: if the time since the last utterance is more than 0, accept an option rated VERY-HIGH or better; if the time since the last utterance is more than 1, accept an option rated HIGH or better; and so on through all the ratings. 19 If the time since the last utterance is more than 16, then any option will be accepted; this essentially means that the generator would rather say something completely stupid than remain silent.</Paragraph> <Paragraph position="3"> If any of the candidate utterances' evaluations meets or exceeds the current minimum acceptable rating, the utterer chooses the best candidate utterance and outputs it, removing the workspace entries it derives from. If there are two or more candidate utterances with the highest rating, the utterer breaks the tie by comparing the ratings 19 IGEN measures time in terms of cycles through its process scheduler, an arbitrary but workable approach that allows for consistent timing behavior.</Paragraph> <Paragraph position="4"> Rubinoff lqle IGEN Generator of the individual options they are composed of. If there is still a tie, the utterer chooses, in order of preference, the candidate utterance that:</Paragraph> <Paragraph position="6"> violates fewer constraints. 2deg was the only one whose rating was adjusted up by the planner, since this means there was something specific about the option that fit well into the plan.</Paragraph> <Paragraph position="7"> is more concise, indicated by the presence of a CONCISE-CONSTRUCTION annotation. 21 covers more entries in the workspace.</Paragraph> <Paragraph position="8"> covers an earlier entry in the workspace.</Paragraph> <Paragraph position="9"> If none of these criteria distinguish the candidate utterances, the utterer simply picks one at random.</Paragraph> </Section> </Section> <Section position="9" start_page="120" end_page="123" type="metho"> <SectionTitle> 4. Examples </SectionTitle> <Paragraph position="0"> The workings of the various components and representational machinery described above can be seen more clearly by looking at a few examples of IGEN at work. These examples demonstrate how IGEN can handle decisions that involve various kinds of interactions between linguistic and planning issues, despite the strict separation of the planning and linguistic components of the generator. In addition, the examples show how IGEN is sensitive to time pressure, improving its output when given more time to work and degrading gracefully when forced to work faster.</Paragraph> <Section position="1" start_page="120" end_page="123" type="sub_section"> <SectionTitle> 4.1 Supporting the Plan Structure </SectionTitle> <Paragraph position="0"> In this example, the generator is given the goal #<KNOW(HEARER,WEATHER-INFD)>, where #<WEATHER-INF0> is an unordered collection of facts about the weather over a three-day span starting on the current day. The final output in response to this goal is: (4) It is warm on today. It will be warm on Monday. It will be cool on Tuesday.</Paragraph> <Paragraph position="1"> Since the input goal simply involves transmitting a fixed set of information to the hearer, the construction of this response is fairly straightforward. Nevertheless, even in this simple example there are a few places where purely linguistic preferences are overridden by the planner in order to support the text plan structure. municative planner, which constructs an initial plan. The planner starts by applying the utter-sequence plan, which involves decomposing the goal based on a relevant sequence. Since the members of #<WEATI-IER-INF0~ all have connections to the sequence of days in the month, the planner is able to decompose the original goal into a sequence 20 The only possible constraint in the current implementation is a missing (obligatory) feature value; all other grammatical and conceptual constraints are strictly enforced. 21 This is actually controlled by a parameter (*conciseness-preference*) that determines whether the utterer will prefer concise or verbose options or neither. In theory, tlnis parameter should be set by the planner based on its knowledge of the situation; in practice, it is set by hand to prefer concise options. Requests to express the information about each day's weather are then passed on to the linguistic component.</Paragraph> <Paragraph position="2"> In particular, the planner carries out step (4) of the plan by placing in the workspace the request:</Paragraph> <Paragraph position="4"> A separate workspace entry is created for each node and link in this request, 22 with the entry for the top-level #<0VER-TIME-SPAN> link marked as their &quot;head&quot; entry.</Paragraph> <Paragraph position="5"> begins to produce and annotate linguistic options for each of these entries. It also runs the first phase of the evaluation algorithm (described in Figure 4) to produce an initial evaluation. 23 For #<TEMPERATURE>, the linguistic component finds the following options: 24</Paragraph> <Paragraph position="7"> The linguistic component draws on the portion of IGEN's knowledge shown in Figure 5 to annotate this option. Since the option directly expresses the concept #<WEATHER>, the linguistic component attaches a MAKES-EXPLICIT annotation to the option. Since #<TEMPERATURE> has a #<PART> link connecting it to #<WEATHER>, it attaches an INDIRECTLY-SUGGESTS annotation for #<TEMPERATURE>.</Paragraph> <Paragraph position="8"> 22 This doesn't limit the linguistic component to working on only one node or link at a time; options can cover or depend on other entries, as, for example, the future tense option in Section 4.1.2. 23 This should really be done by the planner, but the first phase doesn't actually depend on any information about the plan, so it can safely be run by the linguistic component without compromising the separation of the components. The second phase, which does depend on the plan, will be run by the planner.</Paragraph> <Paragraph position="9"> 24 It actually produces several other options that are immediately ruled out as being completely inappropriate, e.g. the word temperature as an option for #<SEPT-26>, or because they require arguments that aren't present in the request, e.g. the verb be (which needs two arguments) as an option for #<TEMPERATURE>. Throughout these examples, such options will simply be ignored in the interests of brevity.</Paragraph> <Paragraph position="10"> Since the usage of #<w: IT> here is an idiomatic construction, it has some special handling in the lexicon to indicate that its meaning depends on what it is being used to describe. In particular, #<w: IT> here is taken to mean #<TEMPERATURE>. The linguistic component thus produces a MAKES-EXPLICIT annotation for #<TEMPERATURE>, as it does for the option using the word #<w : TEMPERATURE>. #<w:IT> is also marked as producing a CONCISE-CONSTRUCTION annotation, so the linguistic component attaches one to the option.</Paragraph> <Paragraph position="11"> Computational Linguistics Volume 26, Number 2 As these options are produced, they are evaluated based on their annotations according to the algorithm described in Figure 4. In the first phase, #<w : WEATHER> is rated MEDIUM, reflecting the fact that the option doesn't really express the request, but does suggest it indirectly by mentioning a closely related concept. #<w : TEMPERATURE> is rated VERY-HIGH, reflecting the option's exact match with the request, as is #<w: IT>, which also expresses #<TEMPERATURE> (albeit idiomatically). The second phase of the evaluation algorithm makes no changes to the ratings, since the options' annotations don't indicate anything that would interact with the plan, leaving #<w: IT> and #<w : TEMPERATURE> as the preferred options. 25 For #<WITHIN-RANGE>, the only acceptable option the linguistic component finds is:</Paragraph> </Section> </Section> <Section position="10" start_page="123" end_page="124" type="metho"> <SectionTitle> (SUBJ HEAD PRED) </SectionTitle> <Paragraph position="0"> The SUBJ and PRED features indicate #<w:BE>'s syntactic roles; the value of these features indicates the syntactic constraints on the structures that may fill the roles.</Paragraph> <Paragraph position="1"> In addition, the structure has a MAP feature indicating which workspace entries can provide options to fill the roles.</Paragraph> <Paragraph position="2"> Since #<WITHIN-RANGE> is a subconcept of #<BE-LOCATED> (one meaning of #<w:BE> in IGEN's lexicon), the linguistic component produces (MAKES-EXPLICIT #<BE-LOCATED>) and (MAKES-IMPLICIT #<WITHIN-RANGE>) annotations for the option. These annotations lead the first phase of the evaluation algorithm to rate the option as HIGH; the second phase leaves this evaluation unadjusted. (Of course, since there is only one acceptable option here, its rating doesn't really matter.) For #<56DEG-Z5DEG-F>, the options produced are (omitting the details of the structures to show only the annotations): #<w:WARM>:</Paragraph> <Paragraph position="4"> These annotations are based on the network fragment shown in Figure 5. #<w:WARM> is rated HIGH, because it does a reasonable job of describing the requested temperature range} 6 The #<w : TEMPERATURE> option is not as good, but it at least gives a sense of what is being talked about, so it gets the next lower rating of MEDIUM.</Paragraph> <Paragraph position="5"> The entries we have seen so far have not involved any interactions between the</Paragraph> </Section> <Section position="11" start_page="124" end_page="126" type="metho"> <SectionTitle> Rubinoff The IGEN Generator </SectionTitle> <Paragraph position="0"> options and the plan. Our first instance of such interactions arises in the options for</Paragraph> <Paragraph position="2"> The first phase evaluations suggest that #<w: OVER> is the preferred option. Here, though, the second phase of the evaluation algorithm modifies the ratings based on interactions with the plan. The ACTIVATES-IN-CONTEXT annotations for #<w:0VER> and #<w:0N> indicate that these options affect how the system (and the hearer) perceive the time span being talked about. Since the #<SEQUENCE> of days being talked about is the organizing principle of the plan, it is a &quot;critical&quot; piece of the plan, as are the #<MEMBER> links connecting it to the pieces of the original goal. The ACTIVATES-IN-CONTEXT annotation for #<w:0VER> indicates that it encourages the hearer to think of #<SEPT-26> as a #<LINEAR-SPAN>. This conflicts with the (critical) #<MEMBER> link between #<SEPT-26> and the #<SEQUENCE>, because it describes the day as a span of time rather than as a discrete point in a sequence. The second phase algorithm thus downgrades #<w:0VER>'s rating from VERY-HIGH to HIGH. Conversely, #<w:0N> encourages the hearer to think of #<SEPT-26> as a #<LINEAR-POS>, i.e., a discrete position within a linear span, supporting the critical #<MEMBER> link. \]&quot;he rating for #<w: ON> is therefore upgraded from HIGH to VERY-HIGH, and it becomes the preferred option. 28 Note here how the use of annotations allows the planner to remain ignorant of purely linguistic issues. The planner has no way to distinguish between an option realizing #<0VER-TIME-SPAN> as a lexical item and one realizing it as the value of some feature of another option. Its decision is based solely on the contextual effects of the available options. The planner can only make distinctions based on the effects of the options, not on the forms of the options.</Paragraph> <Paragraph position="3"> The final piece of the request is #<SEPT-26>, for which the linguistic component produces the following options:</Paragraph> <Paragraph position="5"> The first phase of the evaluation algorithm rates both of these options as HIGH, since they match the planner's request equally well. The second phase, though, distinguishes between them on the basis of the particular time sequences they activate. The sequence #<WEEK> activated by the option #<w :MONDAY> is significantly different from the #<SEQUENCE> of three days around which the plan is organized, so #<w : MONDAY> is downgraded to #<MEDIUM>. The #<FEW-DAYS> sequence activated by the option #<w : TOMORROW>, on the other hand, largely overlaps with the plan's #<SEQUENCE>, so no adjustment is made. As a result, #<w:TOMORR0W> becomes the preferred option.</Paragraph> <Paragraph position="6"> 4.1.3 Assembling the Utterance. As these options are produced and evaluated, the utterer watches the workspace, trying to assemble them into an acceptable utterance.</Paragraph> <Paragraph position="7"> Whenever an option appears that meets the current acceptability threshold, the utterer utters (i.e., prints) it. The final result is: (5) It will be warm on Monday.</Paragraph> <Paragraph position="8"> Note that this output does not simply use the highest-rated option from each entry in the workspace, because that would involve the ungrammatical phrase on tomorrow. In addition to responding to time pressure, the utterer also enforces syntactic constraints; it chooses the best possible combination of options that is syntactically acceptable. 29 4.1.4 Summary. This simple example illustrates how the annotations allow IGEN to maintain the traditional division into two independent components while still handling interactions between decisions at the planning and linguistic levels. The planner never has to deal with linguistic structures, and the linguistic component never has to deal with plans or information structures. For example, the planner can't distinguish between lexical items and feature values; conversely the linguistic component has no access to the goals driving the planner or the plans it builds. Nevertheless, when there are decisions that depend on both informational and linguistic structures, the annotations allow IGEN to handle the interactions between the different levels. The effects of these interactions are relatively minor in this example, but subsequent examples will show how they can be more dramatic.</Paragraph> <Section position="1" start_page="125" end_page="126" type="sub_section"> <SectionTitle> 4.2 Revising the Plan </SectionTitle> <Paragraph position="0"> In addition to merely selecting among options, IGEN's planner can revise the plan in response to them. Consider a slight modification to the previous example, in which IGEN is given the same goal except that the temperature on the third day falls within 29 This is an oversimplification, as can be seen from the output in (4) in Section 4.1, where IGEN produces the phrase on today. What's really going on is that IGEN's lexicon has two entries for each of today and tomorrow; one as a noun phrase and one as a prepositional phrase. As it happens, the PP entry turns up first. In the case here, the option for Monday is produced and selected before the NP option for tomorrow turns up. In the first sentence of (4), there is no such option, so the linguistic component has time to produce the NP today option.</Paragraph> </Section> </Section> <Section position="12" start_page="126" end_page="128" type="metho"> <SectionTitle> Rubinoff The IGEN Generator </SectionTitle> <Paragraph position="0"> the range considered &quot;warm.&quot; IGEN proceeds to build a plan that is identical to the previous one (except for the difference in the information to be expressed). As before, the planner places requests in the workspace and the linguistic component starts to construct options. At this point, IGEN is well on its way to generating text similar to (4): (6) It is warm on today. It will be warm on Monday. It will be warm on Tuesday.</Paragraph> <Paragraph position="1"> However, once the option for warm is produced, something new happens. The planner detects that there are several parallel structures in the workspace that can use the same linguistic structure (warm) to realize different concepts (the different temperature ranges) in corresponding positions within each structure. This provides the opportunity to combine the parallel structures into a single conjoined structure, which the planner proceeds to do. All of the workspace entries for the parallel structures are removed, and the planner enters requests for a new conjoined structure: 3deg</Paragraph> <Paragraph position="3"> IGEN then proceeds as in the previous example, ultimately generating: (7) It will be warm on today, Monday, and Tuesday.</Paragraph> <Paragraph position="4"> Note that the planner couldn't have created this conjoined structure initially because it didn't know that all three days' temperatures could be realized the same way. To the planner, each day has a different temperature range. This example appears exactly the same as the previous one. Once the linguistic component indicates that all three temperature ranges can be realized the same way, though, the planner can then determine that this results in a parallelism that can be reduced to a conjoined structure. The annotations allow the planner to modify its plan, based on the results of the realization process, despite not having any access to purely linguistic knowledge or processing.</Paragraph> <Section position="1" start_page="126" end_page="128" type="sub_section"> <SectionTitle> 4.3 Varying the Utterance in Response to the Plan </SectionTitle> <Paragraph position="0"> In the previous examples, we have seen how IGEN can use the feedback from the linguistic component to improve the organization of the text it generates. The annotations also allow IGEN to tailor the generated text so as to better achieve the goals behind the utterance. Consider IGEN's response to the goal &quot;make the user happy&quot;: 31 (a) Goal: Downplay unpleasant information 30 The temperature range used in the conjoined structure is chosen randomly from among the structures being combined. This is safe to do because it is certain that all of the temperature ranges can be realized the same way; that's what triggered the combination in the first place. 31 The construction of the plans in this section is admittedly ad hoc; IGEN certainly doesn't have a complete model of the psychological factors and issues they involve. What's important here, though, is not the construction of the plan but rather how the plan and the annotations are used to produce the kind of subtle effects that can't be produced by the linguistic component or the planner alone.</Paragraph> <Paragraph position="1"> example. Some of the choices, however, are the result of interactions of a sort beyond what was seen there.</Paragraph> <Paragraph position="2"> Consider the options for #<CONSIST-0F (PRECIPITATION,RAIN) > in step (3) of the plan; these include: 32</Paragraph> <Paragraph position="4"> The first phase of the evaluation algorithm rates the word #<w:RAINS> highest, since it captures the intended meaning precisely. #<w: DRIZZLES> and #<w: POURS> are the next best options, since they capture the basic meaning, and are equally appropriate, since the linguistic component can't evaluate the significance of the extra information they present.</Paragraph> <Paragraph position="5"> The second phase, however, determines that the EXTRA-INF0 expressed by the option #<w:DRIZZLES> implies another part of the information requested by the planner (the #<MINIMAL-SIGNIFICANCE> relation). Its rating is therefore increased to VERY-HIGH. Conversely, the rating for #<w:POURS> is lowered to MEDIUM, since it 32 To simplify the discussion, I have included the first-phase evaluations of the options here.</Paragraph> </Section> </Section> <Section position="13" start_page="128" end_page="130" type="metho"> <SectionTitle> Rubinoff The IGEN Generator </SectionTitle> <Paragraph position="0"> contradicts the #<MINIMAL-SIGNIFICANCE> relation. This leaves both other options, #<w:RAINS> and #<w:DRIZZLES>, rated VERY-HIGH; the utterer will prefer to use #<w:DRIZZLES>, though, since there was something specific about the option that the planner approved of.</Paragraph> <Paragraph position="1"> The decision here to say #<w : DRIZZLES> instead of #<w:RAINS> could not have been made by either the planner of the linguistic component alone. The decision depends on both the planner's knowledge that the concept #<DRIZZLE> supports another piece of the plan and the linguistic component's knowledge that there is a way to express #<DRIZZLE> without making explicit its distortion of the actual information that it's raining. It's only the interaction between the two components provided by the annotations that makes saying #<w : DRIZZLES> instead of #<w:RAINS> possible.</Paragraph> <Paragraph position="2"> The point of using the annotations, though, was not just to handle these kinds of interactions, but more importantly to do so without compromising the principled separation of the two components. Among other things, this makes it possible for IGEN to express the same information in a different manner, based solely on differences in the plan. If the purely linguistic issues are the same, the linguistic component will do exactly the same work, even though the differences in the output may depend in part on that work.</Paragraph> <Paragraph position="3"> We can see this by comparing the previous example with IGEN's response to the goal #<TAKE(HEARER, UMBRELLA)>, for which the planner constructs the following plan: i.</Paragraph> <Paragraph position="4"> Note that the information to be expressed here in step (4) is also expressed in step (3) of the previous example. Since the linguistic component doesn't have access to the plan, it necessarily produces the same options as before. The planner, though, evaluates these options differently, since it is using them to carry out a different plan. The three options #<w: RAINS>, #<w: POURS>, and #<w: DRIZZLES> are produced and annotated as before. Their evaluations are different, though. Step (3) here has a precondition that #<CONSIST-0F(PRECIPITATION,RAIN) > be #<DANGEROUS> in order to achieve its intended goal. The extra information expressed by #<w:POURS> supports this precondition, so #<w:POURS> is upgraded to VERY-HIGH. On the other hand, #<w:DRIZZLES> undercuts the precondition and is therefore downgraded to MEDIUM. The preferred option here is therefore #<w:POURS>, in contrast to the previous example, and the final result is then: (9) It is pouring.</Paragraph> <Paragraph position="5"> Computational Linguistics Volume 26, Number 2 The final output differs even though the linguistic component does exactly the same processing in these two examples. From the linguistic component's point of view, the examples are exactly the same; the same request is placed in the workspace, so the same options are produced. The difference lies solely in the plan structure, not in the content of the explicit message to be realized, so only the planner behaves differently. The linguistic component's design and functioning are completely indifferent to the issues that concern the planner.</Paragraph> <Section position="1" start_page="129" end_page="130" type="sub_section"> <SectionTitle> 4.4 The Range of Variation </SectionTitle> <Paragraph position="0"> We can see similar variation in the handling of #<N0-PRECIPITATION>, which is realized in three different ways in three different examples because of the different roles it plays in the three plans. In Section 4.3, it was realized as clear; the two examples that follow illustrate how IGEN can select alternative realizations.</Paragraph> <Paragraph position="1"> Consider first what happens when the generator is given the goal &quot;get the user to conserve water.&quot; In response, it builds up a plan to do so by talking about the drought: This plan is similar to the one built in Section 4.3 for the goal #<TAKE(HEARER,UM-BRELLA) >; they both involve motivating the hearer to take some action by explaining some danger that the action will provide protection from. The main difference here is that rather than simply telling the user that there is a serious drought, the planner draws on a more specific (ad hoc) plan that involves expressing more detailed information.</Paragraph> <Paragraph position="2"> The final result for this example is: (10) The drought is continuing. It is dry today. It will be dry again tomorrow. Note that IGEN uses #<w:DRY> here to express #<N0-PRECIPITATION>, in contrast with the first example in Section 4.3, where it was expressed by #<w : CLEAR>. This happens because #<w: DRY> has an (ACTIVATES-IN-CONTEXT #<DROUGHT>) annotation, and activating the concept #<DROUGHT> supports the sense of danger that is a precondition of step (3) of the plan. #<w:DRY>'s evaluation is thus upgraded, making</Paragraph> </Section> </Section> <Section position="14" start_page="130" end_page="132" type="metho"> <SectionTitle> Rubinoff The IGEN Generator </SectionTitle> <Paragraph position="0"> it the preferred option. In the example in Section 4.3, this activation doesn't interact with the plan, so #<w:CLEAR> remains the preferred option.</Paragraph> <Paragraph position="1"> Yet another realization is selected when IGEN is given the goal of convincing the user to go to the park, resulting in the following plan: The only difference from previous examples lies in the choice of #<w: SUNNY> to express the concept #<N0-PRECIPITATION> rather than using either #<w:CLEAR> or #<w:DRY>. This choice arises because of an ACTIVATES-IN-CONTEXT annotation indicating that the use of #<w:SUNNY> activates the concept #<SUNSHINE>, which strengthens the enjoyability precondition in step (3) of the plan. It is this sort of interaction between linguistic choices and the plan structure that allows IGEN to express #<N0-PRECIPITATION> three different ways in three different examples.</Paragraph> <Paragraph position="2"> The point here is not just that IGEN can produce different lexical realizations for a particular concept. If that were the only goal, we could dispense with the feedback mechanism and simply design some sort of discrimination network (or similar device) to test various features of the information being expressed. The planner could supply whatever information is needed to drive the network. 33 Something like this approach is in fact used in some systems (e.g., Elhadad and Robin 1992; PenMan 1989; Hovy 1988a).</Paragraph> <Paragraph position="3"> The problem with the discrimination network approach is that it can't handle the range of examples shown here without violating the modularity of the generator. The examples shown here have involved linguistic choices that depend on the plan structure, on other actions in the plan, and on preconditions of the plan. Furthermore, the plan dependencies can involve a nontrivial amount of reasoning (e.g., the example immediately above, which involves the connection between sunshine and enjoying 33 Of course, that still wouldn't allow for examples like the one in Section 4.2, where the planner modifies the plan in response to the work of the linguistic component.</Paragraph> <Paragraph position="4"> Computational Linguistics Volume 26, Number 2 being outdoors). Encoding these decisions in a discrimination net (or any other kind of processing) within the linguistic component would mean giving that component access to all of the plan and the reasoning used to construct it. Much of the planner's knowledge and processing would have to be duplicated in the linguistic component. IGEN, in contrast, handles these examples while maintaining strict modularity.</Paragraph> <Section position="1" start_page="131" end_page="131" type="sub_section"> <SectionTitle> 4.5 The Effects of Time Pressure </SectionTitle> <Paragraph position="0"> The discussion of the examples in Section 4.3 glosses over an important detail: IGEN will only generate the output shown there if the time pressure is eased by increasing the delay thresholds. In particular, the choice of #<w :DRIZZLES> and #<w: POURS> over #<w:RAINS> in those examples is made with the thresholds set to force the utterer to always wait at least five time units before uttering the next piece of text. This is necessary because the linguistic component produces the option for #<w:RAINS> sooner than the options for #<w: POURS> and #<w :DRIZZLES>.</Paragraph> <Paragraph position="1"> If the delay thresholds are left at the default settings, the utterer will accept an option that is rated VERY-HIGH without any delay. Thus when the #<w:RAINS> option appears, the utterer immediately accepts and utters it. While # <w:DRTZZLES > and #<w: POURS > would be better options in these examples, the utterer never gets a chance to see them. Because of time pressure, the decision of how to realize the planner's request is made before the linguistic component has a chance to find #<w : DRIZZLES>.</Paragraph> <Paragraph position="2"> There's nothing wrong with this result; it is raining is a perfectly good way to express #<CONSIST-0F(PRECIPITATION,RAIN)>. It's just that if IGEN is given more time to work, it can produce a better utterance. The utterer's job is to balance the quality of the utterance against time pressure; in this example, changing the intensity of the time pressure induces a corresponding change in the refinement of the utterance.</Paragraph> <Paragraph position="3"> Of course, if the time pressure is set sufficiently high (e.g., to simulate a situation where the the hearer is about to walk away), IGEN will sometimes produce an utterance that has some slight problems, because speed is more important than perfection. Consider what happens when the example in Section 4.1 is run with the delay thresholds set to accept anything that's not terrible (i.e., any option rated MEDIUM or higher) right away. The planning and initial production and evaluation of options proceeds as with the default thresholds. The handling of #<0VER-TIME-SPAN> works out differently, though. The option for #<w:0VER> is produced by the linguistic component several time units before the option for #<w:0N>. With the default thresholds, that's not a problem, because the utterer can wait long enough for #<w: 0N> to be produced.</Paragraph> <Paragraph position="4"> With the increased time pressure, though, the utterer is forced to produce the utterance before the option #<w:0N> is available. The resulting utterance (for the second day's information) is it will be warm over Monday.</Paragraph> <Paragraph position="5"> This result is not awful; the sentence still manages to express the intended information. It just doesn't fit quite as well into the intended overall structure of the text. So the main consequence here of increasing the time pressure is to generate a text that is slightly less coherent and fluent. In this case, IGEN responds to increasing time pressure with a slight degradation in the generated text.</Paragraph> </Section> <Section position="2" start_page="131" end_page="132" type="sub_section"> <SectionTitle> 4.6 The Effects of Vocabulary Limitations </SectionTitle> <Paragraph position="0"> Since IGEN is designed to always find the best utterance possible in the available time, without assuming any notion of a &quot;correct&quot; utterance, it can produce reasonable output even when generating in domains where it has limited linguistic resources. This can be seen by looking at the result of selectively disabling elements of its &quot;vocabulary,&quot; i.e., the lexical and syntactic resources available to the linguistic component. The result</Paragraph> </Section> </Section> <Section position="15" start_page="132" end_page="134" type="metho"> <SectionTitle> Rubinoff The IGEN Generator </SectionTitle> <Paragraph position="0"> is a gradual, graceful degradation of the generated text as the generator's linguistic resources become poorer.</Paragraph> <Paragraph position="1"> Consider again IGEN's response (under relatively mild time pressure) when given the goal #<TAKE(HEARER,UMBRELLA)>. The resulting plan, as we have seen, requires expressing the information #<CONSIST-0F (PRECIPITATION,RAIN) >, which is accomplished by uttering it is pouring. If the verb #<w: POUR> is removed from the generator's vocabulary, though, this utterance is not possible; instead the output becomes it is raining. Note that this is the same output IGEN produces if the time pressure is increased; the consequence of limiting either time or vocabulary is the same.</Paragraph> <Paragraph position="2"> If the verbs #<w:RAINS> and #<w:DRIZZLES> are removed as well, the generator is left with no single option that can express all of the requested information. Instead, the utterer resorts to assembling individual options for the various pieces of the request, producing the sentence the precipitation consists of rain. 34 While this utterance does adequately express the requested information, it is fairly awkward. Still, it's not bad, given that the two best options are unavailable.</Paragraph> <Paragraph position="3"> Beyond this point, further vocabulary removal starts to make things much worse.</Paragraph> <Paragraph position="4"> Removing the phrase #<w: CONSIST OF> leads to the utterance of it will be clear. This utterance is terrible; it completely misses the intended information. That's not surprising; after all, the options that should be used have all been removed from IGEN's vocabulary. On the other hand, it's not completely bizarre; it's still talking about the weather, and in fact it's even still describing the precipitation. That is, IGEN still has a vague sense of what it's trying to say; it's not simply producing an utterance at random. This can be seen again by removing the phrase #<w: BE CLEAR>; the resulting utterance is then it snowed. IGEN is still trying to get as close as it can to the intended information, even though it can't really get it right.</Paragraph> <Paragraph position="5"> Eventually, of course, removing linguistic resources will make the generator collapse; if every word or phrase relating to precipitation of any sort is removed, then IGEN produces it is medium, which, though grammatical, is gibberish. If the linguistic component has absolutely no way to produce options related to what the planner wants to express, then there's really nothing IGEN can do. What's significant here, though, is that as the generator's linguistic resources are gradually impoverished, the generator's output degrades gracefully. Rather than immediately collapse when its preferred options are removed, IGEN continues to produce the best possible utterance it can build with its remaining linguistic resources. 3s 4.7 Generating in a Different Language As a test of IGEN's modularity, a simple French linguistic component was developed and tested on the examples in Sections 4.1 and 4.3. Switching to a different language provides an extreme example of how the linguistic component can be varied without affecting the planner. The change primarily involved defining new lexicon entries for the French words, phrases, and feature values needed to talk about the weather. In addition, the routine that handles verb inflection had to be replaced, since French verb endings are different from English ones. Other than these changes, though, no mod34 The actual output is precipitation consists of rain, but this is just due to a limitation of the implemented grammar. Specifically, there's no requirement that noun phrases have determiners, so the linguistic component never bothers to suggest one for the NP headed by #<w:PRECIPITATIOR>.</Paragraph> <Paragraph position="6"> 35 Of course, in many situations it's best to put a lower bound on the quality of IGEN's output; this can be done simply by setting the appropriate delay thresholds to infinity. It would also be possible to set an overall time limit on IGEN's processing, after which an uncompleted plan would be assumed to have failed. IGEN could then modify the plan or take other appropriate corrective action.</Paragraph> <Paragraph position="7"> Computational Linguistics Volume 26, Number 2 ification to IGEN was necessary. The linguistic component was otherwise unaltered, and the planner and utterer required no changes at all. 36 The main difference between the ways English and French describe the weather lies in the &quot;weather it&quot; construction. English allows weather descriptions that use it together with any of a class of verbs of being, seeming, or becoming. Thus the following sentences are all possible in English: (12) It is warm.</Paragraph> <Paragraph position="8"> (13) It seems warm.</Paragraph> <Paragraph position="9"> (14) It feels warm.</Paragraph> <Paragraph position="10"> (15) It's becoming warm.</Paragraph> <Paragraph position="11"> (16) It's getting warm.</Paragraph> <Paragraph position="12"> In these sentences, the meaning of it seems to be the weather or the temperature; thus in any of these sentences it could be replaced by the weather or the temperature without changing the meaning or acceptability. This phenomenon is modeled in IGEN by having a special word #<w:IT> whose meaning can range over the various items constituting the weather; this word is constrained to appear only as the subject of an appropriate verb. Thus, as we saw in Section 4.1, #<w: IT> is represented as meaning #<TEMPERATURE> and annotated appropriately.</Paragraph> <Paragraph position="13"> A literal translation of any of the sentences in (12) to (16) into French, though, would be ungrammatical. The corresponding French construction uses the verb faire (literally 'to make' or 'to do'), as in (17): It is warm.</Paragraph> <Paragraph position="14"> *Le temps fait chaud.</Paragraph> <Paragraph position="15"> The weather makes/does warm The weather is warm.</Paragraph> <Paragraph position="16"> *La temp4rature fait chaud.</Paragraph> <Paragraph position="17"> The temperature makes/does warm The temperature is warm.</Paragraph> <Paragraph position="18"> *I1 semble chaud.</Paragraph> <Paragraph position="19"> It seems warm.</Paragraph> <Paragraph position="20"> I1 semble faire chaud.</Paragraph> <Paragraph position="21"> It seems to make/do warm.</Paragraph> <Paragraph position="22"> It seems warm.</Paragraph> <Paragraph position="23"> Unlike it in (12) to (16), the il here is not referring to the weather or the temperature. Rather it is an expletive, i.e., a dummy subject required by the grammar of the language, similar to the it in the English it seems John left. This can be seen by comparison 36 It was necessary to add a few more elements to the semantic network (which is used by both components) to capture the meaning of the French word sur, which does not correspond exactly to the meaning of any English word, as discussed below. These new elements were used only by the linguistic component, not by the planner.</Paragraph> </Section> <Section position="16" start_page="134" end_page="135" type="metho"> <SectionTitle> Rubinoff The IGEN Generator </SectionTitle> <Paragraph position="0"> with (18) and (19), in which replacing il with le temps (&quot;the weather&quot;) or la temperature (&quot;the temperature&quot;) leads to an ungrammatical sentence. (20) and (21) demonstrate that the crucial element in the French construction is the verb faire; other verbs can be used only if they combine with faire. This contrasts with the English construction, in which it is the crucial element and can be used with a range of verbs. The French construction is thus represented in IGEN as a version of the verb faire that takes an expletive il as its subject and whose meaning ranges over specifications of the various elements of the weather.</Paragraph> <Paragraph position="1"> A second difference between the two languages shows up in Section 4.1 in the options for #<0VER-TIME-SPAN>. In English, IGEN has to choose between using on or over to express this link. In French, though, these concepts are expressed by the same word (sur). 37 Furthermore, sur can't be used for temporal expressions, only for spatial expressions. So the options for #<0VER-TIME-SPAN> work out quite differently in French. French in general uses ~ (&quot;at&quot;) rather than sur for temporal expressions; however ~ cannot be used with expressions describing days. Instead, these temporal modifiers are expressed using bare NP's with no preposition. This construction is modeled in IGEN by a null preposition whose object must be a definite NP.</Paragraph> <Paragraph position="2"> Given all of this, IGEN produces the following results in French for the example in Section 4.1: (22) I1 fait chaud aujourd'hui. I1 fera chaud lundi.</Paragraph> <Paragraph position="3"> it does warm today it will do warm Monday It is warm today. It will be warm Monday.</Paragraph> <Paragraph position="4"> I1 fera frais mardi.</Paragraph> <Paragraph position="5"> it will do cool Tuesday It will be cool Tuesday.</Paragraph> <Paragraph position="6"> This is equivalent to the English result except for expressing the temporal modifier as a noun phrase rather than as a prepositional phrase headed by on, as discussed above.</Paragraph> <Paragraph position="7"> Similarly, the variant in Section 4.2 produces: (23) I1 fera chaud aujourd'hui lundi et mardi.</Paragraph> <Paragraph position="8"> it will do warm today Monday and Tuesday It will be warm today Monday and Tuesday.</Paragraph> <Paragraph position="9"> For the goal of making the user happy, IGEN produces the French output: (24) I1 bruine seulement. I1 fait chaud. Le temps sera clair bientdt. it drizzles only it does warm the weather will be clear soon It's only drizzling. It's warm. It will be clear soon.</Paragraph> <Paragraph position="10"> This is exactly the same as the English output except for the use of the French &quot;weather faire&quot; construction or the explicit use of le temps (&quot;the weather') rather than the English &quot;weather it&quot; construction.</Paragraph> <Paragraph position="11"> 37 Note that the point here is not that sur is the (unique) translation of on and over into French; IGEN is doing generation, not machine translation. It's just that sur is the French word that most naturally expresses the particular concepts IGEN is using. Computational Linguistics Volume 26, Number 2 The goal of convincing the user to take his umbrella, though, produces the output: (25) I1 pleut.</Paragraph> <Paragraph position="12"> it rains It's raining.</Paragraph> <Paragraph position="13"> because French doesn't have a verb corresponding to to pour. Since no option equivalent to the one for #<w: POURS> is proposed, the option for #<w: PLEUT> is the preferred option, just as #<w:RAINS> is the preferred option in English when lack of time or deliberate removal prevents the linguistic component from finding #<w : POUR$>. The point here is not simply that IGEN can handle these changes; after all, one of the primary motivations for dividing the generator into separate planning and linguistic components is to allow the components to be modified independently. The point is that IGEN allows the linguistic component to be modified so dramatically without limiting its ability to handle interactions between linguistic and planning issues. In fact, IGEN can handle these interactions equally with either version of the linguistic component without having to make any corresponding changes to the planner.</Paragraph> </Section> class="xml-element"></Paper>