File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/90/c90-3051_metho.xml
Size: 20,141 bytes
Last Modified: 2025-10-06 14:12:29
<?xml version="1.0" standalone="yes"?> <Paper uid="C90-3051"> <Title>Incremental Parsing and Reason Maintenance</Title> <Section position="2" start_page="0" end_page="0" type="metho"> <SectionTitle> 2 3TMS-Style Approaches </SectionTitle> <Paragraph position="0"/> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 2.1 Overview </SectionTitle> <Paragraph position="0"> An early example of adopting reason maintenance in parsing (more precisely, story analysis) is the system RESUND (O'Rorke 1983). O'Rorke considers the problem of correcting inferences 4 that conflict with subsequent information in the story. He combines the story processor with a JTMS, using dependency-directed backtracking to determine incompatible assumptions, and a collection of preference heuristics to choose from among candidate solutions. 5 3Sense-semantic interpretation; possibly also contextual interpretation.</Paragraph> <Paragraph position="1"> 4 An inference here corresponds to a selection (assumption) of a schema, a script-like knowledge structure used for deriving information not explicitly mentioned in the text. ~Other work, ttmugh not adopting downright JTMSs, makes use of somewhat similar techniques for the purpose of recovering from erroneous inferences. For example, Jacobs (1988), in dealing with the problem of &quot;concretion&quot; -- developing a most specific or metaphorical interpretation -- lets his system, TRUMP, treat each interpretation as an assumption.</Paragraph> <Paragraph position="2"> If an assumed interpretation results in a contradiction, dependencies are used to discard (chains of) assumptions that conflict with the preferred interpretation. Conflicting information is simply thrown away, ~hus in a sense making the system even More recently, Zernik and Brown (1988) suggest a coupling of a DCG parser with a JTMS, both of which are embedded in a Prolog system. The DCG is extended with default rules to enable nonmonotonic reasoning. These defaults are used to guide the entire parsing and interpretation process. The division of labour is as follows: Given a new piece of input, the parser outputs an (extended) analysis in the form of a dependency network. This, in turn, is fed to the JTMS, resulting in an updated in~out labelling of the network corresponding to the currently believed interpretation. The purpose of the JTMS is to obtain a system which avoids (chronological) backtracking and instead handles inconsistent information (ambiguities) by choosing to believe a different portion of the previous inferences. For example, in parsing a sentence like &quot;The child sold by his parents was found alive&quot; (cf. Zernik and Brown 1988:802), the system initially assumes &quot;the child&quot; to be the agent since the &quot;default voice&quot; is considered to be &quot;active&quot;. When later the word &quot;by&quot; is parsed, a nonmonotonic supporter of &quot;active&quot; becomes in, thus making &quot;active&quot; become out and &quot;the child&quot; be considered direct object.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 2.2 Problems </SectionTitle> <Paragraph position="0"> This section dis'cusses some general problems of JTMSs and JTMS-style approaches to incrementM parsing with special reference to the framework of Zernik and Brown (1988) whose notion of parsing comes closest to the one considered here.</Paragraph> <Paragraph position="1"> Perhaps the most important characteristic of a JTMS is that it insists on global consistency; in other words, it is limited to one single solution (one context) at a time: &quot;At each point \[presumably after each new word\], the parser must deposit a hypothesis based on a partial set of clues, a hypothesis which might later be retracted.&quot; (Zernik and Brown 1986:802.) Unfortunately, in a domain like natural-language analysis where local ambiguity constantly plagues the parser with inconsistent information, this becomes problematic. First of all, when the set of assumptions admits multiple solutions, these cannot be compared: since a JTMS only allows one context, there is simply no way to examine two sets of beliefs simultaneously in order to gauge their relative strengths.</Paragraph> <Paragraph position="2"> Furthermore, upon each incremental change (i.e., parsing of a new word), new JTMS labellings have more insistent on consistency than a JTMS. Story processors like ARTHUR (Granger 1980), FAUSTUS (Norvig 1983), and ATLAST (Eiselt 1987) keep track of successive (candidate) inferences and reconsider rejected ones when faced with conflicting information. It could finally be mentioned that in the postscript to his book, Mellish (1985:114) suggests a combination of chart parsing and JTMS, but does not further develop tiffs.</Paragraph> <Paragraph position="3"> 288 2 to be comlmted tbr the network. If, in the com'se of this, a contradiction arises - tile probability of which increases with the size of the grammar, i.e., with the number of (l)otentially competing) default rules dependency-directed backtracking h~s to be invoked to identify the sources of the contradiction and resolve the conllict. This requires extensive search and often results in new contradictions. Until all conflicts are resolved, the status of some nodes may have changed between in and out several times. 'l?hus, &quot;tile machinery is cumbersome&quot; (de Kleer 1986:139)fi A further problern is that the irfl'erence engine works on only one part of the search space at a time. For example, if a word has two senses, only one of them will be worked on by the parser. But all that is known in such a ca,se is that both senses cannot be part of the same final solution; it may still be important to draw int~rences Dora them independently. For further discussion of these and other problems in connection with JTMSs, see de Kleer (1986:138 ft.).</Paragraph> </Section> </Section> <Section position="3" start_page="0" end_page="0" type="metho"> <SectionTitle> 3 ATMS-Style Approaches </SectionTitle> <Paragraph position="0"/> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 3.1 Overview </SectionTitle> <Paragraph position="0"> Charniak and Gohlman (1988) make use of an NI)MS for keeping track of alternatives arising in sensesemantic and co,ltexl;ual interpretation, for example, wil.h respect to word sense, case, and noun-phrase reference. Each alIernative is treated as an ATMS assun~l)tion and is fnr~hermore assigned a pro/mbility. By comparing segs of assumpl.ioi~s underlying various potential interpreta.lions, the system can choose the &quot;best&quot; alternative, i.e., the one with the highest probability.</Paragraph> <Paragraph position="1"> Nagao (1989) provides an approach where ditferent assumptions about sentence interpretations constitute different &quot;worlds&quot; of formulae in a way which resembles an ATMS representation. A characteristic of these fra.meworks is that they only handle semantic-interpretation alternatiw',s and do not (attempt to) integrate this with parsing.</Paragraph> <Paragraph position="2"> A different kind of Ni'MS--style approach, and one that is grounded in parsing, can be obtained by extending a chart parser with dependencies; more specifically, by recording for each edge its immediate and ultimate source edges (Wirdn 1989). The next section develops this.</Paragraph> <Paragraph position="3"> 6IncidentMly, the sole explicit example provkted by Zernik and Brown (eited above) ouly involves one default rule. Since the original as well as the revised interpretation represent coherent, sets of justified beliefs, both c~m be arrived at by straightforward (re)labelling, titus avoiding the more cumbm~ some process of dependency-directed backtracking -- in fact, ~ernik and Brown do not mention dependency-directed back-tracking at all. They also do no |state any systematic preference policy, so it is not clear how they generally gauge the relative strengths of incompa/.ible assumptions when trying to resolve contradictions.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 3.2 An ATMS-Style Chart Parser </SectionTitle> <Paragraph position="0"> tIow can a chart parser extended with edge depen-.</Paragraph> <Paragraph position="1"> dencies be viewed as consisting of an inference engine alld a (simplified) ATMS? This section develops an infonnal answer to this question.</Paragraph> <Paragraph position="2"> 'llb begin with, an ATMS, just like a chart parser, can be seen as a tool for organizing efficient search through a space of alternatives by providing a &quot;cache&quot; of partial results and by sharing these results across different branches of the search space. Both the (I);-usic) A'I?MS and a chart parser are inonotonic in the sense of only providing monotonic derivability (and, in case of the ATMS, monotonic justifications).</Paragraph> <Paragraph position="3"> Both frameworks are incremental, performing piecemeal updates in response to a constant stream of new assumptions, nodes, and justifications. Furfllermore, the order in which updates are made does not affect tile final outcome of the process. In particular, the following correspondences hold: (r) A chart edge corresponds to an NI'MS 7~ode.</Paragraph> <Paragraph position="4"> . A preterminal (lexical) edge corresponds to an assumption node)' (r) The immediate source information of an edge corresponds t.o a justification.</Paragraph> <Paragraph position="5"> * Information about the set (of sets) of ultimate source edges of an edge corresponds to its ATMS label.</Paragraph> <Paragraph position="6"> . The chart corresponds to an ATMS network.</Paragraph> <Paragraph position="7"> . An analysis (or interpretation) of a phr~e, sentence, etc. corresponds to an ATMS context, i.e., the theory of an environment, where the latter is a set of assumptions.</Paragraph> <Paragraph position="8"> (r) The (standard) chart-parsing algorithm corre null sponds to the inference engine.</Paragraph> <Paragraph position="9"> More precisely, information about so,tee edges (i.e., justifications) can be derived as follows: An edge formed through a combination depends on the active--inactive edge pair that generated it. An edge formed through a prediction depends on the (one) edge that triggered it. (Alternatively, predicted edges can be left out of the dependency trails altogether: since they only represent inferential hypotheses and do not carry any analysis structure, they could be seen ~s belonging with the inference engine rather than with the ATMS. On this view, a prediction has neither dependants nor sources, s) A scanned edge does not depend upon any other 7Thus, each word sense corresponds to an assumption. In a system whidl haaldles noisy input (e.g., ill-formed or spoken input), one might instead let hypothesized word forms correspond to assmnptions.</Paragraph> <Paragraph position="10"> 8This would also soNe the. problenl with top-down parsing pointed out in Wirdn (1989:245 f.). --- Note that, if we want to introduce gm'bage collect.ion of useless predictions, it would still be necessary to Imep a record of their dependencies.</Paragraph> </Section> </Section> <Section position="4" start_page="0" end_page="0" type="metho"> <SectionTitle> 3 289 </SectionTitle> <Paragraph position="0"> edge (but on an instance of a word or a lexicalized phrase).</Paragraph> <Paragraph position="1"> Labels are likely to be simple in an ATMS-style chart parser. Normally, each edge has a unique, stable set of sources ~- zero, one, or two edges which are determined once and for all when the edge is created) A potential exception to this uniqueness of source is the case of the parser attempting to regenerate an (existing) edge, something which is prohibited by a redundancy test. This attempted regeneration actually corresponds to introducing an additional justification for the edge. Allowing this would require a more elaborate machinery for computing ATMS labellings in accordance with de Kleer (1986).</Paragraph> <Paragraph position="2"> Whether this capability is needed or not would have to be decided with respect to the demands of the particular application of the ATMS-style parser (of.</Paragraph> <Paragraph position="3"> section 4).</Paragraph> <Paragraph position="4"> It could finally be noted that a chart parser, ~s normally conceived of, does not record &quot;nogoods&quot; (inconsistent combinations of assumptions) aus does the ATMS. Instead, the chart parser by itself ensures that inconsistent combinations of edges do not get further worked on (through predictions, the agenda, etc.).</Paragraph> </Section> <Section position="5" start_page="0" end_page="0" type="metho"> <SectionTitle> 4 Applications </SectionTitle> <Paragraph position="0"/> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 4.1 Incremental Parsing and Interpretation </SectionTitle> <Paragraph position="0"> A strong definition of incremental parsing would require that text can be added or deleted in a piecemeal fashion and, furthermore, that the system behaves monotonically in processing such updates; i.e., as words are added/deleted, the set of possible analyses increases/decreases monotonically. An attractive property which follows from this is that the amount of processing of an update is roughly proportional to the size of the update (eft Wirdn 19.89:2,i2, Earley and Caizergues 1.972:1040). An ordinary chart parser as well as an ATMS-style chart parser as put forward above are incremental in this sense, whereas the nonlnonotonic model of Zernik attd Brown (1988) attd, say, an ATN parser are not. In the latter frameworks, a previously determined analysis can be modified by a subsequent step in the parsing process, tdeg As for interpretation with respect to a model or a context, in order for a system to be incremental in the above sense, it should be compositional (and again monotonic) such that the interpretation 9In bidirectionM chart paa'sing (Satta and Stock 1989) an edge might have three solwces.</Paragraph> <Paragraph position="1"> ldegOne might ask which granunatical formalisms enable incremental processing. Cai, egorial gratmnoa- in its various incantations is an obvious alternative (for example, Ades and Steedman 1982). Unification-based grammar formalisms (Shieber 1986) provide another alternative given that the ratification component is capable of incremental processing (Bresnml and Kaplan 1982:xliv ft., Stee~:hnan 1985).</Paragraph> <Paragraph position="2"> of a phrase is a flmction of the interpretations of its syntactic constituents and their associated contexts. One computationally-oriented model which flflfils this requirement, and which is indeed taylored for incremental interpretation, is that of Haddock (1987, 1988, 199(I). Haddock's model, which can be seen as a continuation and refinement of Mellish (1985), incrementally interprets singular noun-phrases which refer to known contextual entities.</Paragraph> <Paragraph position="3"> Translated into a chart-parsing framework, upon scanning a word, the corresponding set of predicates and potential referents (obtained from the context) is associated with the new edge, and upon combining edges, a constraint-satisfaction algorithm is run to narrow down the set of possible referents.</Paragraph> <Paragraph position="4"> The possibility of using dependencies also to incrementally handle deletions of words (assumptions) is investigated in Wirdn (1989). Actually, a machinery is developed to handle arbitrary sgutactic changes , which can be thought of as edit operations (insertion, deletion, replacement), and which can be fl'eely combined and applied to arbitrary portions of the text -- for example, input can be entered in any order.</Paragraph> <Paragraph position="5"> Edge dependencies are used to propagate the effects of a change precisely to those edges that are affected by the change. II This has potential computational applications in interactive na.turaldanguage systems such as language-sensitive text editing. In psycholinguistic terms, an edit operation might correspond to correction of a misread or misheard passage.</Paragraph> <Paragraph position="6"> Since IIaddock's model for incremental interpretation is not in any way limited to left-to-right incrementality, it is possible to adopt it also within the system of Wir~;.n (1989). Furthermore, it is possible to conceive of a semantic analogue to this processing of syntactic changes. Consider a dynamic context, for example a database representing the real-time state of some world. By maintaining dependencies between entities and relations in the contextual model and their counterparts in the linguistic analysis, a machinery for incrementally reevaluating previously made interpretations with respect to the changing context could be attained.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 4.2 Comparison of Interpretations </SectionTitle> <Paragraph position="0"> The chart allows comparison of (competing) interpretations ill the sense that any analyses can be simultaneously examined. WhaZ one cannot do in ordinary chart parsing is to ~k which particular edges (assumptions, etc.) underlie a given analysis since edges are not labelled with justifications or assumptions (and all information from lower-level edges may not have been percolated upwards). Put differently, the chart cannot &quot;explain&quot; its analyses. Of course, extending a chart parser to record dependencies is a simple thing. The point is that, in doing so, one has in effect obtained a simple KI'MS-style problem 11 This could also be achieved in a JTMS-style parser.</Paragraph> <Paragraph position="1"> 290 4 solver. Charniak and Goldman (1988) provide an example of how ATMS techniques could be used for comparisons (eft section a.1).</Paragraph> </Section> <Section position="3" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 4.3 Revision of Interpretations </SectionTitle> <Paragraph position="0"> The basic ATMS does not provide nonmonotonic justifications and hence no machinery for actually revising a previously held interpretation through default reasoning, etc. (Similar effects are instead achieved by maintaining multil)le contexts and by switching between these.) However, certain semantic and pragmatic phenomena, like anaphoric reference, ,~eem to require a capability tbr default reasoning; processing such phenomena by gradually developing all possible analyses wouht lead to combina~ torial explosion (Asher 1984). '\]~hus, although non-monotonic devices destroy the strong notion of incrementality discussed above (the effects of an update cannot hr general be kept local, the amount of compntation needed is not bounded by the size of the update, etc.), they are sometimes needed. A recent example of this is Pollack and Pereira (1988) who use a strict compositional semantics but. with non-monotonic interpretation rules; other examples are Asher (1984), Dunin-Keplicz (1984), and Appelt and Konolige (1988).</Paragraph> <Paragraph position="1"> Dre,~sler (1989) shows how to extend the basic ArMS with a new type of node which, in particular, allows the encoding of nonmonotonic justifications and default rules. Given the relationship between chart parsing and ATMS, it seems like there should be a way of translating such a default machinery to an ATMS-style parsing framework. Furthermore, given a framework which integrates parsing and interpretation by performing interpretation on-line to the parser, it might be advantageous to allow encoding of default reasoning at the level of the parser, as indeed Zernik artd Brown do, but to use it in a nmch more restricted way.</Paragraph> </Section> </Section> class="xml-element"></Paper>