File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/98/w98-1406_metho.xml
Size: 17,745 bytes
Last Modified: 2025-10-06 14:15:14
<?xml version="1.0" standalone="yes"?> <Paper uid="W98-1406"> <Title>De-Constraining Text Generation</Title> <Section position="3" start_page="0" end_page="49" type="metho"> <SectionTitle> 2 Ontology-Based Modularization </SectionTitle> <Paragraph position="0"> In contrast to modularization by tasks such as discourse structuring, clause structuring and lexical choice, the Mikrokosmos project (http://crl.nmsu.edu/Research/Projects/mikro/index.html) attempts to modularize on the ontological and linguistic data that serves as inputs to the text generation process, that is, based on the types of inputs we expect, not on the types of processing we need to perform. A typical semantic representation that serves as the input to the generation process is shown in Figure 1. This semantic input was produced by the Mikrokosmos analyzer from an input Spanish text.</Paragraph> <Paragraph position="1"> The generation lexicon in our approach is essentially the same as the analysis lexicon, but with a different indexing scheme: on ontological concepts instead of NL lexical units 1. (\[Stede1996\] is an example of another generator with a comparable lexicon structure, although our work is richer, including collocational constraints, for example). The generation lexicon contains information (such as, for instance, semantics-to-syntax dependency mappings) that drives the generation process, with the help of several dedicated microtheories that deal with issues such as focus and reference (values of which are among the elements of our input representations). The Mikrokosmos Spanish core lexicon is complete with 7000 word senses defined; the English core lexicon is still under development with a projected size of over 10,000 word senses. Both of these core lexicons can be expanded with lexical rules to contain around 30,000 entries (\[Viegas et al.1996\]).</Paragraph> <Paragraph position="2"> Lexicon entries in both analysis and generation can be thought of as &quot;objects&quot; or &quot;modules&quot; corresponding to each unit in the input. Such a module has the task of realizing the associated unit, while communicating with other objects around it, if necessary (similar to \[De Smedt1990\]).</Paragraph> <Paragraph position="4"> Each module can be involved in carrying out several of the *tasks like those listed by Wanner and Hovy. For * instance, modules for specific *events or properties are used in setting up clause and sentence structures as well as lexical choice, as will be shown below. Interactions and constraints * flow freely, with the control mechanism dynamically tracking the connections 2. One outcome of this division of labor between declarative data and the control architecture is that the bulk of knowledge processing resides in the lexicon, indexed for both analysis and generation. This has greatly simplified knowledge acquisition in general \[Nirenburg et a1.1996\] and made it easier to * adapt analysis knowledge sources to generation as well as to convert knowledge sources acquired for one language to use with texts in another.</Paragraph> <Paragraph position="5"> Below we sketch out how this organization works. We begin by describing the * main types of lexicon entries with the goal of demonstrating how each performs various generation *tasks. We then take a look at the different types of constraints associated with each kind of entry.</Paragraph> <Paragraph position="7"/> <Section position="1" start_page="49" end_page="49" type="sub_section"> <SectionTitle> 2.1 Types of Lexicon Entries </SectionTitle> <Paragraph position="0"> The main types of lexicon entries correspond to the ontological categories of OBJECTS, EVENTS and PROPERTIES (for simplicity, we will avoid discussion of synonyms and stylistic variations): * Objects. In English, Objects are typically realized by nouns, although the actual mapping might be rather complex \[Beale and Viegas1996\]. In general, object generation lexicon entries can have one-to-one mappings between concept and lexical unit, or can contain additional semantic restrictions, both of which are illustrated in Figure 2. The use of collocational information is described below. * * , Events. * Events, as shown pictorially in Figure 3, can be realized as verbs (&quot;divided&quot;) or nouns (&quot;division&quot;) in English.* Furthermore, the lexicon entries for events typically determine the structure * of the nascent clause by mapping the expected case roles into elements of the verb subcategorization frame (or other structures). Rules for planning passives and relative clauses, for instance, are also available. These rules can be used to fulfill different grammatical and clause combination requirements as described below. Conceptually, all the entries produced by these rules can be thought of as being resident in the lexicon. Practically speaking, many of them can be produced on the fly automatically, reducing the strain on knowledge acquisition.</Paragraph> <Paragraph position="1"> degkilt VAtZa comi~si~ bJt IBM.</Paragraph> <Paragraph position="2"> PP-~eUam~t (opo, omm~ .</Paragraph> <Paragraph position="3"> soot &quot;by&quot; * degbJ~t VAILI diwLcdon of,Apple Into ...</Paragraph> <Paragraph position="4"> IBM, who dlvdded Apple gngo ... 1 J sg~ companies, which IBM divided --. \] J MkpplC/. whkh was divided Into ... \[ J</Paragraph> <Paragraph position="6"> .... p...v, .h~,h has ,~.k \] the coml~tny , whose *took I * Figure 4: PROPERTY Lexicon Entries - A Simplified View typically are consumed by the event entry, except in the case of some nominalizations. DISCOUI~SE-</Paragraph> </Section> </Section> <Section position="4" start_page="49" end_page="53" type="metho"> <SectionTitle> * RELATIONS contribute to setting up sentence boundaries, sentence ordering and pronominalization. </SectionTitle> <Paragraph position="0"/> <Paragraph position="2"> The above generation lexicon entries are the primary knowledge sources used in the generation process 4. Five different types of constraints are automatically generated which constrain the combi: nations of entries allowed to globally realize a semantic input. The Mikrokosmos control mechanism efficiently processes constraints to produce optimal global answers.</Paragraph> <Paragraph position="3"> Binding Constraints. One of the primary advantages of input-based modularization is that the individual knowledge sources (lexicon entries) can be grounded in the input they expect to be matched against. For instance, in Figure 3, the semantic* input expected shows three variables, corresponding*to the three case roles normally associated with a DIVIDE event. The process of linking these variables to the actual semantic structures for a particular input is known as binding.</Paragraph> <Paragraph position="5"> For the input shown in Figure 1, VAR2 will be bound to CONGLOMERATE-32 and VAR3 will be bound to CORPORATION-34.* * Notice that, for this example, no AGENT exists for the DIVIDE-31 event, so that VAR1 will be left unbound. Binding constraints will simply eliminate any syntactic choices that contain non-optional unbound variables. In this case, it will rule out the first syntactic realization for DIVIDE shown in Figure 3.</Paragraph> <Paragraph position="6"> The grounding of the input afforded by the binding process also allows us to simplify the other types of constraints described below. Each of these types of constraints ~ automatically processed in our system, in task-based systems typically require complex rules to be acquired manually. Grammatical Constraints. An example of a grammatical constraint is shown in Figure 5. A lexicon entry can specify grammatical constraints on the realization of any of the variables in it. One *possible syntactic realization for ASSERTIVE-ACT is shown. It requires its VAR2 to be realized as a clause. This particular entry allows the system to produce &quot;John said that Bill went to the store&quot; but not &quot;John said that Bill.&quot; A comparison with Figure 1 shows that the binding process will link VAR2 of the ASSERTIVE-ACT entry to DIVIDE-31. In effect, the*resulting constraint will eliminate any realization for DIVIDE-31 (in Figure 3) that does not produce a full clause at the top-level, through nominalization and relativization. It should be stressed that this filtering occurs only in conjunction with the given *realization of ASSERTIVE-ACT; there may be other realizations that would go fine with, for example, a nominalized realization of DIVIDE.</Paragraph> <Paragraph position="7"> CoUocational Constraints. Figure 6 illustrates the familiar notion of collocational constraints.</Paragraph> <Paragraph position="8"> Again, the fact that the lexicon entryis grounded in the input allows a simple representation of collocations. In this case, the different realizations of LOCATION usually correspond to the semantic type Of the object. Collocations :can be used to override the default. The co-occurrence zone of the STOCK-MARKET entry simply states that if it is used as the range of a LOCATION</Paragraph> <Paragraph position="10"> relation, then the LOCATION relation should be introduced with &quot;on.&quot; This produces an English collocation such as &quot;the stock is sold on the stock market&quot; as opposed to the less natural &quot;... sold at the stock market.&quot; Notice that no additional work on collocations needs to be performed beyond the declarative knowledge encoding. The constraint-based control architecture will identify and assign preferences to collocations.</Paragraph> <Paragraph position="11"> Clause Combination Constraints. Various kinds of constraints arise when clauses are combined :to form complex sentences. The strategies for clause combination come from three sources: * * Directly from a lexicon entry associated with an input. For example, a discourse relation such as CONCESSION might directly set up the syntax to produce a sentence structure such as &quot;Although Jim admired her reasoning, he rejected her thesis.&quot; * Verbs which take complement clauses as arguments also set up complex sentence structures and impose grammatical constraints (if present) on the individual clause realizations: &quot;John said that he went to the store&quot; or &quot;John likes to play baseball.&quot; * Indirectly, from a language-specific source of clause combination techniques (such as relative clause formation or coordination in English).</Paragraph> <Paragraph position="12"> These three sources correspond to the three input situations depicted in Figure 7. The first two have explicit relations linking two EVENTs. The first (the non-case-role relation) will have a * corresponding lexicon entry which directly sets up the sentence structure, along with specific constraints on the individual clauses. The second possibility typically occurs with EVENTs that take complement clauses as case-role arguments. The lexicon entries for these usually will specify the complex clause structure needed. The third situation has no explicit connection in the input; therefore, some sort of language-specific combination strategy must be used to fill the same task.</Paragraph> <Paragraph position="13"> Even though the latter case appears to be a situation that requires a task-oriented procedure, in reality it is as easy to use general purpose structure constraints along with a declarative representation of possible transformations available. Assuming, for the sake of illustration, that due to some external reason a single sentence realization of two clauses is preferred 5, a general purpose structural constraint prevents two clauses from embedding a single referent into distinct syntactic structures. For instance, 1 and 2 below are grammatical, but 3 is not, because both the clauses try to use &quot;conglomerate&quot; as their subject.</Paragraph> <Paragraph position="14"> 1. The conglomerate, whose stock is sold on the stock market, was divided into nine corporations.</Paragraph> <Paragraph position="15"> SConstraints which might produce such a preference can come from a variety of Sources; a common one is the realizations of discourse relations.</Paragraph> <Paragraph position="16"> 2. The conglomerate , which was divided into nine corporations, is sold on the stock market.</Paragraph> <Paragraph position="17"> 3. *The conglomerate was divided into nine corporations is sold on the stock market.</Paragraph> <Paragraph position="18"> The general purpose constraint will automatically prevent such a realization and trigger the consideration of subordinate clause transformations.</Paragraph> <Paragraph position="19"> In addition, the examples of clause combination given above and in Figure 7 all contain e~amples of coreference across clause boundaries. Although coreference realization has its own microtheory that is triggered by instances of coreference in the *text, clause combination techniques may interact with it. For instance, the lexicon entry for a RELATION might specify that a pronoun be used in the second clause.</Paragraph> <Paragraph position="20"> The important thing to note for this presentation is that these types of constraint are either directly found in the lexicon or are produced automatically by the planner. Special situations such as coreference can be easily identified because the lexicon entries ar e grounded in their inputs. This method appears to be much simpler than those needed by task-based generators.</Paragraph> <Paragraph position="21"> Semantic Matching Constraints. Matching constraints take into account the fact that, first of all, certain lexicon entries may match multiple elements of the input structure and, secondly, that the matches that do occur may be imperfect or incomplete.</Paragraph> <Paragraph position="22"> In general, the semantic matcher keepstrack of which lexicon entries cover which parts of the input, which require other plans to be used with it, and which have some sort of semantic mismatch with the input. The following sums up the types of mismatches that might be present, each of which receives a different penalty (penalties are tracked by the control mechanism and help determine which combination of realizations is optimal): tant thing to note here is that input-based modularization in our knowledge sources enables this type of constraint to be tracked automatically. In combination with the other constraints described above, we Can avoid the complex mechanisms needed by task-based generators for interacting realizations of input semantics. :</Paragraph> </Section> <Section position="5" start_page="53" end_page="55" type="metho"> <SectionTitle> 3 Efficient Constraint-based Processing * </SectionTitle> <Paragraph position="0"> The Mikrokosmo s project utilizes an efficient, constraint:directed control architecture called Hunter-Gatherer (HG). \[Beale et a1.1996\] overviews how it enables semantic analysis to be performed in near linear-time. Its use in generation is quite similar. \[Beale1997\] describes HG in detail.</Paragraph> <Paragraph position="1"> Consider Figure 8, a representation of the constraint interactions present in a section of Figure 1. Each label, such as DIVIDE, is realizable by the set of choices specified in the lexicon. Each</Paragraph> <Paragraph position="3"> solid line represents an instance of one of the above constraint types. For example, DIVIDE and ORG-INVOLVED-IN are connected because of the structural constraint described above (they both cannot set up a structure which nests the realization of CONGLOMERATE=32 into different subject positions).</Paragraph> <Paragraph position="4"> The key to the efficient constraint-based planner Hunter-Gatherer is its ability to identify constraints and partition the overall problem into relatively independent subproblems. These subproblems are tackled independently and the results are combined using solution synthesis techniques. This &quot;divide-and-conquer&quot; methodology substantially reduces the *number of combinations that have to be tested, while * guaranteeing an Optimal answer. For example, in Figure 8, if we assume that each node had 5 possible choices (a conservative assumption), there would be 51deg, or almost 10 million combinations of choices to examine. Using the partitions shown in dotted lines, however, HG only examines 1200 combinations, In general, HG is able to process semantic analysis and generation problems for natural language in near linear-time \[Beale et a1.1996\]. = While a detailed explanation of Hunter-Gatherer is beyond the scope of this paper, * it is fairly easy to explain the source of its power. Consider Figure 9, a single subproblem from Figure 8.</Paragraph> <Paragraph position="5"> The key thing to note is *that, of the three nodes, BUY, LOCATION and STOCK-MARKET, only BUY is connected by constraints to entities outside the subproblem. This tells us that by looking only at this subproblem we will not be able to determine*the optimal global choice for BUY, *since there are constraints we cannot take into account. What we can do, howeve r , is, for each possible choice for BUY, pick the choices for LOCATION and STOCK-MARKET that optimize it. Later, when we combine the results of this subproblem with other subproblems and thus determine which choice for BUY is optimal, we will already have determined the choices for LOCATION and STOCK-MARKET that go best with it.</Paragraph> <Paragraph position="6"> The following sums up the advantages Hunter-Gatherer has for text generation:</Paragraph> </Section> class="xml-element"></Paper>