CLAUSE-LEVEL RELATIONSHIP ANALYSIS IN THE TANKA SYSTEM

CLAUSE-LEVEL RELATIONSHIP ANALYSISIN THE TANKA SYSTEM

KEN BARKER

Department of Computer Science, University of OttawaOttawa, Ontario, Canada K1N 6N5

email: [email protected]: 613-564-5420 fax: 613-564-9486

Abstract

Knowledge acquisition from text is often attempted in the presence of largeamounts of pre-coded domain knowledge. Seeding a system with suchknowledge is often a huge knowledge acquisition effort in itself. Breakingthis circle requires a text processing system that relies on little a priorisemantic information. In the absence of such knowledge, processing mustrely on available information, such as surface syntax. Semantic analysis willconsist of determining the semantic relationships between the varioussurface-syntactic constituents in the sentence.

This report describes the design and implementation of a system thatinteractively interprets semantic relationships between clauses. The Clause-Level Relationship Analyzer inspects the parse trees produced by a domain-independent parser for syntactic features that can be used to determine whatkinds of semantic relationships exist between clauses. The system isdesigned to require as little a priori semantic knowledge as possible. Thisdocument offers a set of semantic labels appropriate to syntacticallyconnected clauses and a description of the theory behind assigning theselabels to particular inputs. It also presents elements of the implementation ofthe Clause-Level Relationship Analyzer and looks at its performance.

1 Introduction

This report describes the design and implementation of a semantic analysis module to dealwith semantic relationships between clauses. This module comprises one part of theinteractive semantic analyzer within the TANKA system (Text ANalysis for KnowledgeAcquisition).

The objective of the TANKA project is to build a conceptual network model of a technicaldomain by processing English text that describes the domain. A guiding principle of theproject is to use as little a priori semantic knowledge as possible. In the absence of suchknowledge, processing must be driven by the syntax and assisted by a user. A researchgoal, therefore, is to determine how much of the meaning of a text can be uncovered fromits syntax and how unambiguous the derived meaning will be.

TANKA’s text analysis system starts with an unsimplified text and applies syntacticknowledge and surface-level lexical knowledge to build a conceptual model of the domain.The system consists of three phases: syntactic analysis, semantic analysis and conceptualnetwork building.

CLR Analysis - 2

TANKA has recently been adopted as the linguistic processing module for a MachineLearning from text project. See Delannoy et al. (1993) and Delisle et al. (1994) for adescription of the MaLTe project (Machine Learning from Text).

This document begins with an overview of the main components of TANKA. Section 2describes theoretical issues both in the construction of a set of semantic relationships at theclause level and in the design of an analysis module for these relationships. Section 3presents implementation details of the Clause-Level Relationship Analyzer and section 4addresses limitations of the design and the Analyzer.

1 . 1 Syntactic Analysis in TANKA: DIPETT

Syntactic analysis is performed by DIPETT (Domain Independent Parser of English TechnicalTexts). DIPETT is a broad-coverage DCG parser whose rules are based primarily on Quirket. al. (1985). The parser produces a single initial parse which can be subsequentlycorrected (if necessary) by an interactive reattachment module. The reattachment moduleallows a user to make slight changes to the parse tree by, for example, reattaching aprepositional phrase to a noun instead of to the verb. Work on this module is currently inprogress. For details on DIPETT, see Copeck et al. (1992) and Delisle (1994).

1 . 2 Semantic Analysis in TANKA: HAIKU

The semantic analysis module, called HAIKU, consists of three levels of semanticprocessing: Clause-Level Relationship Analysis, Case Analysis and Noun-ModifierRelationship Analysis. The semantic relationships occurring at each level are listed inTable 1.

Clause-Level Relationships (CLRs) are the semantic relationships between acts,events or states represented syntactically by syndetically connected finite clauses. CLRAnalysis attempts to assign a semantic label to the relationship expressed in the connectionof these clauses. The performance goal of this module is to automatically find the best CLRthat represents the input sentence, thereby easing the burden on the user who is expected toaccept or reject suggestions made by the CLR Analyzer.

Cases represent semantic relationships between the main verb within a clause and itssyntactic arguments (subject, objects, prepositional phrases and adverbials). For eachclause in the input, HAIKU’s Case Analyzer interactively assigns semantic labels to thesyntactic arguments based on syntactic and lexical clues. The Case Analyzer learns to makebetter assignments by calculating the “closeness” of the current input to patterns memorizedfrom previous processing. The Case Analyzer begins by finding in the input the syntacticelements known to mark Cases (prepositions, adverbials and positional noun phrases). Theuser assigns Case labels to the constituents marked by these Case Markers based onsuggestions made by the Case Analyzer. The Case Analyzer consults a list of CaseMarker→Case mappings as well as dictionaries of previous Case assignments in order tomake appropriate suggestions. The Cases identified by Case Analysis will eventually beused to construct links between activities and objects in the growing conceptual network.For details on the Case Analyzer, see Delisle et. al. (1993). For details on the Case Systemused by the Case Analyzer, see Barker et. al. (1993).

Noun-Modifier Relationships (NMRs) represent semantic relationships between thehead noun of a noun phrase and its modifiers (adjectives, nouns, relative clauses, etc.).The relationships identified between a noun and its modifiers will eventually be used toconstruct links between objects and properties and between objects and other objects in thegrowing conceptual network. Research on NMRs and NMR Analysis is in progress.

CLR Analysis - 3

CLRs CasesCausal Participant

Causation AgentEnablement BeneficiaryEntailment ExperiencerPrevention InstrumentDetraction Object

Temporal RecipientTemporal Co-occurrence CausalityTemporal Precedence Cause

Conjunctive EffectConjunction OppositionDisjunction Purpose

TimeFrequencyTime_at

NMRs1 Time_fromAgentive Time_toAppositive Time_throughBenefactive SpaceCausative DirectionClassificatory Location_atCompositional Location_fromDestination Location_toInstrumental Location_throughLocative OrientationObjective QualityPossessive AccompanimentProductive ContentPurposive ExclusionSource MannerTemporal MaterialTopic Measure

Order

Table 1: Semantic Relationships

The division of semantic relationships into these three levels follows from their syntacticanalysis at different levels, but we recognize the fact that a single semantic concept may berealized syntactically at any one of the levels. Consider the following two examples:

A man was murdered yesterday with a handgun because a jealous husband returnedhome early.

the murder of a man yesterday by a handgun because of the early return of a jealoushusband

1 The list of NMRs is a preliminary attempt to enumerate semantic relationships between a noun and itsmodifiers based primarily on Levi (1978). It has not yet been carefully scrutinized or validated.

CLR Analysis - 4

Syntactically, the first example contains one CLR (between the two clauses connected by‘because’), six Cases (corresponding to ‘a man’, ‘yesterday’, ‘a handgun’, ‘a jealoushusband’, ‘home and ‘early’) and one NMR (marked by ‘jealous’). The second examplehas four top-level NMRs (marked by ‘of a man’, ‘yesterday’, ‘with a handgun’ and‘because of the early return of a jealous husband’), two level-two NMRs (marked by‘early’ and ‘of a jealous husband’) and one level-three NMR (marked by ‘jealous’). Thesimilarity of the second example to the first may be due (in part) to the use of a noun todenote an activity instead of an object2. Because the noun denotes an activity, its modifiersare more likely to fill Case-like roles than the roles associated with regular object-denotingnouns. In order to accurately analyze such inputs, it may be desirable to allow theassignment of Case roles to noun modifiers. This idea will be investigated further in theresearch on Noun-Modifier Relationships.

It can argued that every distinct surface representation has a unique meaning. The argumentclaims that the intent of the writer/speaker dictates the particular surface structure for an ideathat may have more than one syntactic realization. This choice of one surface representationover another is sufficient to differentiate between all possible representations. However,such fine distinctions cannot be represented in a broad-scope knowledge base efficiently.Therefore, we accept different syntactic representations as potentially representing a single(if somewhat ambiguous) deep structure.

The structures resulting from semantic analysis will be used by the Network FragmentBuilder to construct clusters of concepts. These clusters (conceptual network fragments)will be added to a growing network intended to be a representation of the domain. Detailsof the Conceptual Network formalism can be found in Yang & Szpakowicz (1991a, 1991b,1994).

2 Theory

2 . 1 A CLR System

HAIKU’s CLR System consists of a small set of general semantic relationships (seeTable 1) and a list of the lexical items that mark them (coordinators, correlatives,subordinators, etc.). In this document, these lexical items are referred to as CLR Markers(in a semantic context) or clausal connectives (in a syntactic context). The CLR setitself arose from an exhaustive study of the connectives and was designed to cover therelationships marked by their various meanings. The CLR Analyzer has been tested withseveral larger sample texts and the CLR set has so far proven to be adequate based onexperiments such as the one described in section 3.3.

The CLR set has also been compared to other lists of relationships between clauses (seesection 2.3 on Related Work). However, most of these lists were developed in the area ofDiscourse Analysis. These “Discourse Relations” between clauses (or larger text units) dealnot only with the relationships between the events or acts represented by clauses, but alsowith the rhetorical functions of the clauses themselves. This distinction has been variouslyreferred to as “Subject Matter” vs. “Presentational” or “Semantic” vs. “Pragmatic”. Our

2 Quirk et. al. (1985) refer to such nouns as deverbal nouns. The problem with attempting to treatdeverbal nouns differently than regular nouns is in identifying which nouns are deverbal in the absenceof lexical semantic information.

CLR Analysis - 5

CLRs only deal with the relationships between the events or acts denoted by the clauses(i.e., the “Subject Matter” or “Semantic” relationships).

2 . 1 . 1 CLR Markers

CLR Analysis is centred around the CLR Marker Dictionary which contains a single entryfor each CLR Marker. Each Marker entry enumerates the CLRs that can be associated withthe Marker. These CLR Marker→CLR mappings were determined by studying the varioususages of each connective to learn what semantic relationships they represent.

The CLR set was constructed by introducing a new label for each relationship indicated bya particular meaning of a connective. The set was then edited to be specific enough tomaintain coverage over all of the connective meanings but general enough that each CLRwas well represented in the Dictionary: a CLR marked by a only single meaning of a singleconnective was a candidate for merging with another CLR. The process was iterative:elements of the Marker list sometimes suggested a weakness in the CLR set, while amaturing CLR set often uncovered overlooked Marker usages. For example, a particularusage of a connective having no obvious mapping to a CLR indicated a possibly incompleteCLR set while a single connective usage seeming to map to more than one CLR indicated apossible redundancy in the CLR set. The mapping of a Marker to a particular CLR mightalso indicate a potential mapping from the same Marker to a related CLR. The current stateof the CLR Marker list is presented in Appendix I.

Inspection of the connective usages also uncovered a result useful to CLR Analysis: theidentification of a consistent “direction” for each connective. For most of our CLRs, theorder of the CLR arguments is relevant. The exceptions in our set of CLRs areConjunction, Disjunction and Temporal Co-occurrence. For example, for Causation, oneclause is the antecedent and the other is the consequent. With some connectives, the clauseintroduced by the connective is the antecedent (e.g. ‘The file will print if the programworks’) whereas with other connectives the clause introduced by the connective is theconsequent (e.g. ‘The program works so the file printed’). This correspondence betweenthe syntactic arguments to the connective and the semantic arguments to the CLR isconsistent among the usages of the connective. Each connective’s direction is also stored inthe CLR Marker Dictionary so that CLR Analysis can automatically determine the correctordering of CLR arguments, given the order of the syntactic arguments.

2 . 1 . 2 CLR Glossary

This section describes the meaning of each Clause-Level Relationship. It is meant (in part)to guide TANKA users when asked by the system to approve the program’s suggestion of aCLR for a given input. Each description gives the name of the link that will be used in theNetwork Fragment Builder between network nodes corresponding to CLR arguments.Also accompanying each description is an example sentence that illustrates the CLR. Theabbreviations can be typed by the user when assigning a CLR.

The descriptions of the CLRs appear in three groups according to the type of relationshipthey represent: Causal, Temporal and Conjunctive. This division corresponds to a kind ofranking of CLRs: The Causal CLRs imply a temporal ordering (a cause temporallyprecedes its effect) and also imply a Conjunctive relationship (Prevention impliesDisjunction; the others imply Conjunction). Similarly, the Temporal CLRs implyConjunction. In the absence of any other information, this ranking could be used as adefault to prefer one CLR over another. For example, by default Temporal Precedence maybe preferred over Causation, since Causation implies Temporal Precedence. However, the

CLR Analysis - 6

system can afford to be ambitious and prefer the “stronger” Causation, since the user canalways reject the suggested assignment.

For the remainder of this document, CLRs are treated as binary relationships. The CausalCLRs are all binary. The Temporal and Conjunctive CLRs can have more than twoarguments. However, the only connectives that can mark these n-ary CLRs (n > 2) are thecoordinators ‘and’ and ‘or’. Of these two, ‘or’ unambiguously marks Disjunction, while‘and’ can mark Conjunction, Temporal Co-occurrence or Temporal Precedence. TheCausal CLRs present more difficult CLR disambiguation problems. Therefore, the CLRAnalysis mechanisms presented in the rest of this document apply to binary CLRs.

In the following definitions, A1 and A2 refer to acts or states denoted by clauses. Theycorrespond to the “first” and “second” CLR (semantic) arguments. Note that A1, the “first”semantic argument, often corresponds to the “second” syntactic argument. In the rest ofthis document, reference to the first and second CLR arguments will correspond to A1 andA2 respectively in the CLR definitions. For Causal CLRs, A2 will often be referred to asthe resultant.

In the examples, A2 appears boldface (where the distinction between A1 and A2 isrelevant) and the connective (CLR Marker) is capitalized.

Causal

Causation (caus)

The Causation relationship will be represented in a network by a link causes(A1,A2) whereA1 makes A2 occur or exist. A1 is sufficient to cause A2 and the occurrence or existence ofA1 is required.

The file printed BECAUSE the program issued a print command.

Enablement (enab)

The Enablement relationship will be represented by a link enables(A1,A2) where A1 makesA2 possible. A1 is necessary to enable A2 but is not sufficient.

The printer can print IF the paper tray contains paper.

Entailment (entl)

A link entails(A1,A2) will represent an Entailment relationship. If A1 exists or occurs thenA2 must also exist or occur. Unlike Causation, however, A1 is not known to exist oroccur.

The printer will print IF a print command is issued.

Prevention (prev)

The Prevention relationship will be represented by a link prevents(A1,A2) where A1 ismeant to keep A2 from occurring or existing. A1 is sufficient to prevent A2.

The files were not copied SINCE the hard disk crashed.

CLR Analysis - 7

Detraction (detr)

The Detraction relationship will be represented by a link detracts_from(A1,A2). A1 detractsfrom A2 but is insufficient to prevent A2 from occurring or existing.

Although the server was very busy, the program ran.

Temporal

Temporal Co-occurrence (ctmp)

Temporal Co-occurrence will be represented by a link cooccurs_with(A1,A2) where A1and A2 occur or exist at the same time.

One job can run in the background WHILE another job runs in the foreground.

Temporal Precedence (prec)

A link temporally_precedes(A1,A2) will represent the relationship in which A1 occurs orexists (or begins to occur or exist) before A2.

The file printed BEFORE I changed the toner cartridge.

Conjunctive

Conjunction (conj)

A Conjunction relationship exists between two acts or states about which no more can besaid than that they both occur or exist. This relationship will be represented by a linkin_conjunction_with(A1,A2).

The computer runs applications AND the printer prints documents.

Disjunction (disj)

A Disjunction relationship exists between two acts or states about which no more can besaid than that one or both occur or exist. This relationship will be represented by a linkin_disjunction_with(A1,A2).

The program may terminate OR it may not.

2 . 2 CLR Analysis

CLR Analysis begins by consulting the CLR Marker Dictionary to determine which CLRsare marked by the connective in the current input. Although the CLR set is quite small tobegin with, the Marker Dictionary lists only those CLRs marked by each connective. Thesystem will then choose one of the CLR candidates for the given connective as the mostappropriate.

Since one of the fundamental tenets of TANKA is to avoid the use of a priori semanticknowledge, the most promising approach for choosing between candidate CLRs is to findsyntactic information in the input that can be used for disambiguation. There are two mainsources of such syntactic information: the clausal connective and the verb phrase features ofeach of the clausal arguments.

CLR Analysis - 8

2 . 2 . 1 Connective Polarity and Argument Polarity

The polarity of the verb phrase (positive or negative) can be used to distinguish betweenpositive (Causation, Enablement, Entailment) and negative (Detraction, Prevention) CausalCLRs. For example, the difference between Entailment and Prevention is often the fact thatthe resultant for Entailment is positive whereas the resultant for Prevention is negative:

The document will not print IF the paper tray is empty. (prev)

The document will print IF the printer is ready. (entl)

Note that in the final CLR representation, the negation in the second clause should bedeleted. Consider the following example:

The machine will not work IF the power is off. (1)

The desired CLR representation for this sentence is

'the power is off' <prevents> 'the machine will work' (2)

In general, for a sentence of the form

A entails (not B)

the desired CLR interpretation is

A prevents B

where the polarity of the CLR’s second argument has been reversed. We will refer to thisrule as the Polarity-Reversal Rule for negative Causal CLRs (Prevention and Detraction).

However, with certain connectives the role of verb phrase polarity has the opposite effect indetermining the CLR:

The machine will work UNLESS the power is off.

This sentence has the same desired CLR representation (2) as sentence (1). However, forthis sentence, the polarity of the second CLR argument should not be reversed, but shouldbe left as it is in the original sentence.

Conversely, if the CLR is positive (Causation, Enablement or Entailment), the second CLRargument’s polarity should be reversed when marked by ‘unless’:

He will not attend UNLESS he finishes his paper.

This sentence has the desired representation:

'he finishes his paper' <entails> 'he will attend'

These examples suggest that the nature of the connective as either “positive” or inherently“negative” should be taken into consideration when performing CLR Analysis. From thispoint on, a connective will be referred to as either a positive connective or a negativeconnective whenever the distinction is relevant. The “positivity” or “negativity” of a

CLR Analysis - 9

connective will be referred to as the connective polarity3. Note that both positive andnegative connectives can mark both positive and negative Causal CLRs.

So we now have the following Polarity-Reversal Rule:

For Causal CLRs: reverse the polarity of the second CLR argument in the final CLRrepresentation if either

a) the CLR is a negative CLR marked by a positive connective, orb) the CLR is a positive CLR marked by a negative connective.

Other connectives from the CLR Marker dictionary that exhibit the behaviour of negativeconnectives include: ‘although’, ‘but’, ‘either-or’, ‘except’, ‘lest’, ‘notwithstanding’,‘only’, ‘or’, ‘save that’, ‘saving that’, ‘though’, ‘till’, ‘unless’, ‘until’, ‘yet’.

The problem with using argument and connective polarity to distinguish between positiveand negative CLRs is that the polarity of a clause is often implicit in one of its elements.Compare ‘the program will not succeed’ (explicit verb phrase negation) versus ‘theprogram will fail’ (implicitly negative verb phrase) or ‘the program will experience failure’(implicit negation in the complement). Only in the first example can polarity can be used toassist CLR Analysis, since the other two examples would require the a priori lexicalsemantic information that “fail” and “failure” are somehow implicitly negative4.

2 . 2 . 2 Argument Tense and Modality

In a surface-syntactic analysis, tense and modality are often ambiguous. For example, theauxiliary ‘could’ is sometimes used as the past tense of the modal of Ability ‘can’ (as in ‘Icould stand on my head when I was seven’), and is sometimes used as a conditionalauxiliary (as in ‘I could do that If I wanted to’). Similar tense/modality ambiguities existwith ‘may’, ‘will’ and others5. These ambiguities suggest that tense and modality beconsidered together.

The tenses and modalities of the clauses in a sentence often hold clues to the semanticrelationships between clauses. The difference between many of the Causal CLRs is thedegree of certainty of the resultant. For example, the difference between Entailment andEnablement is the degree of certainty that the resultant may occur, given the precedent:Entailment implies a high degree of certainty that the resultant will occur; Enablement, amuch weaker certainty.

3 This terminology is not meant to suggest that the connective itself is inherently positive or negative.Rather, it reflects the connective’s relationship to verb phrase polarity in determining whether the CLRis positive or negative.

4 In practice, for inputs with implicitly negative Resultants, A prevents B will be interpreted as A entails(implicitly negative) B. For example, the sentence ‘The job failed to terminate BECAUSE the machinedied’ will be interpreted as 'the machine died' <entails> 'the job failed to terminate' ,which is valid, if not as elegant as 'the machine died' <prevents> 'the job terminates' .

5 Lyons (1977) describes in detail the overlap between futurity and modality, citing the fact that in manylanguages, future tenses developed from forms denoting uncertainty.

CLR Analysis - 10

Stronger Modals Weaker Modals<no modal> cancannot coulddare not maymust mightneed need notshall oughtwill shouldwould

Table 2: Strength of the Modals

Although much of the research into modals has identified hierarchies of modal “strength”6,the definition of our Causal relationships as sufficient or insufficient suggests that a binarydivision of modals into “stronger modals” and “weaker modals” is appropriate fordisambiguating the CLRs.

Table 2 shows the modals (and marginal modals7) recognized by DIPETT, divided intostronger and weaker modals. Some notes on the division are justified:

• Since modals represent some degree of uncertainty, a clause with no modalauxiliary will always be considered stronger than a clause with a weaker modal.

• Several researchers (see, for example, Hermerén, 1978) have noted that manymodals in negative verb sequences signal an opposite semantic modality to that expressedby the same modal in positive verb sequences. For example, the modal ‘can’ representsPossibility when appearing in positive verb sequences (as in ‘the program can print’) butrepresents Necessity in negative verb sequences (‘the program cannot print’). On ourstronger/weaker scale, ‘can’ would be a weaker modal while ‘cannot’ (or ‘can’t’) would bea stronger modal. Modals exhibiting this behaviour in declarative texts appear in both theirpositive and negative forms in the table.

• The modal ‘dare’ only occurs in modern declarative English in negation (seePalmer, 1979, p. 27). The negated ‘dare’ (as in ‘I dare not go’) is a stronger modal (assupported by the paraphrase ‘It is certain that I will not go’). Any remaining forms of‘dare’ in the positive, such as ‘I dare say’ have become idiomatic and would occurextremely rarely (if ever) in declarative technical texts.

• ‘May’ only becomes a stronger modal when appearing in a negative verb phrasewith a second person subject (contrast ‘you may not leave’ with ‘the program may notprint’). The sentence with the second person subject is unambiguous between a stronger

6 Many researchers, including Palmer (1979), have identified the modals as falling into two broadcategories: those denoting some degree of Necessity (stronger modals) and those denoting some degreeof Possibility (weaker modals). Hermerén (1978) goes further to suggest that the modals representdistinct degrees of modality. He offers the ranking Certainty⇒Prediction⇒Probability⇒Possibility, forexample.

7 The marginal modals in the table are ‘dare’, ‘need’ and ‘ought’. These modals are considered similar tothe central modal auxiliaries in function. However, they usually appear in patterns where they moreclosely resemble main verbs than auxiliaries (see Quirk, et al., 1985:3.40-43).

CLR Analysis - 11

sense of permission and a weaker sense of intention because language users do not tend tomake assertions about an addressee’s intent, unambiguously making ‘may not’ with asecond person subject a stronger modal. However, since the third person subject is morecommon in declarative technical texts, ‘may not’ does not appear with the stronger modalsin the table.

• Similar to ‘dare’, ‘need’ usually only occurs in modern declarative English innegation. The negated ‘need’ (as in ‘I need not go’) is a weaker modal (as supported by theparaphrase ‘It is uncertain that I will (not) go’). However, in the rare situation where needoccurs in the positive (as in ‘Only graduates need apply’) it is a stronger modal.

• The modal ‘should’ can be interpreted as the past tense of ‘shall’, suggesting that itis a stronger modal (as in ‘I felt sure that we should meet again’ (from Quirk et al.,1985:4.58). However, such a reading is extremely rare in modern English and its use isrestricted to indirect speech, which has little relevance to our work. We can restrictourselves to the more common, tentative (weaker modal) reading of ‘should’.

2 . 2 . 3 Choosing Between Competing CLRs

This section addresses all of the possible pairs of CLRs8. For each pair, the syntacticfeatures of CLR arguments that may be useful for distinguishing between the CLRs arelisted. These distinguishing features were used to build a set of “preference rules” tochoose one of the CLRs over the other, given a particular input with particular syntacticfeatures. Using these rules, the CLR Analyzer suggests a CLR assignment to the user, whocan accept or reject the suggestion.

These features were derived by examining each of the connectives in the CLR MarkerDictionary to find pairs of potentially competing CLRs—i.e. CLRs that can be marked bythe same connective. For each pair, several example sentences were constructed differing inmodality = {stronger modal, weaker modal, no modal} and polarity = {positive, negative}for each CLR argument. In many cases, only one value for each feature was possible forone of the arguments without changing the meaning of the example sufficiently to be betterdescribed by a different CLR. For example, in the sentence

IF the software is working, the printer can print. (enab)

if the modality of the second argument is changed to a stronger modal, such as ‘will’, theexample is better described as Entailment.

For CLRs that seemed potentially ambiguous but that had no common connective, the CLRMarker Dictionary was examined closely to see if it might be incomplete. Several suchinstances were found.

For some pairs of CLRs, the syntactic features of the arguments are not consistentlydifferent enough to allow disambiguation. These ambiguous pairs have an effect on theoutcome of CLR competitions, described in section 3.2.2. Moreover, many of the pairshave no known connectives in common (in the Marker dictionary). If the user ever assignsone CLR from such a pair to a connective that marks the other CLR, the heuristics will not

8 For the nine CLRs in our set, there are 9

2

= 36 such pairs.

CLR Analysis - 12

be able to disambiguate between them in future processing. In many of these cases,however, it is unlikely that such an assignment will ever be made. For example, it isunlikely that a user would ever assign the Conjunction CLR to the connective ‘or’ (whichwould cause ambiguity between Conjunction and Disjunction in future processing).

As usual, where there is a distinction between the first and second CLR arguments(corresponding to A1 and A2 in section 2.1.2), the second CLR argument is shownboldface.

Causation vs. Enablement

For Causation, the occurrence of the resultant is required. Logically, Causation isimplication where the antecedent is known to have occurred. For this reason, modals areunlikely to appear in either clause (since modals, by definition, express a degree ofuncertainty; see Quirk et al., 1985:4.49).

For Enablement, the occurrence of the resultant is uncertain, since the enabling clause isinsufficient to guarantee its occurrence. The resultant clause is likely to contain modals—likely weaker modals.

The program stopped BECAUSE the machine did not have enough memory.(caus)

The program should run BECAUSE the machine has enough memory.(enab)

Causation vs. Entailment

Again, the occurrence of the resultant is required for Causation. For Entailment, theresultant is contingent on the truth of the proposition represented by the first CLRargument. Modals (in particular stronger modals) may be common in the resultant forEntailment, but should not appear if the relationship is Causation.

The program failed, THEREFORE the system crashed. (caus)

The program failed, THEREFORE the system will crash. (entl)

Causation vs. Prevention

When Causation and Prevention are marked by the same connective, the resultant clause forCausation has a positive polarity whereas for Prevention the polarity of the resultant isnegative. Note that only condition a) of the Polarity Reversal Rule will apply sinceCausation is never marked by a negative connective.

We adjourned the meeting AS it was getting late. (caus)

We could not continue the meeting AS it was getting too late. (prev)

Causation vs. Detraction

The distinction between Causation and Detraction is the same as that between Causationand Enablement. Namely, for Detraction, weaker modals are expected in the resultantclause while no modals are expected for Causation. The polarity of the resultant may alsobe used as a distinguishing feature (as with Causation vs. Prevention). Although a negativeresultant is not a required feature for Detraction, it is unlikely to appear with Causation.

CLR Analysis - 13

The program worked BECAUSE there was sufficient memory. (caus)

The program may not work BECAUSE memory is low. (detr)

Causation vs. Conjunction

None of the connectives in the CLR Marker Dictionary mark both Causation andConjunction. Therefore, it is unlikely that the CLR Analyzer will ever need to distinguishbetween them.

Causation vs. Disjunction

None of the connectives in the CLR Marker Dictionary mark both Causation andDisjunction. Therefore, it is unlikely that the CLR Analyzer will ever need to distinguishbetween them.

Causation vs. Temporal Precedence

In general, it is difficult to distinguish between the Causal and Temporal (or Causal andConjunctive) CLRs when marked by the same connective:

I cried SINCE she left. (caus)

I watched television SINCE she left. (prec)

It is possible that the presence of modals might signal a preference for the Causal CLRs.However, as claimed above, the Causation CLR is distinguished by an absence of modals.Moreover, the ambiguity between ‘will’ used as a modal and ‘will’ as a tense indicatormakes a distinction even more difficult (since it is likely that ‘will’ as a tense indicator willoccur frequently with Temporal CLRs).

Lacking a compelling theoretical basis for distinguishing between these types of CLRs, itmay be useful to consider a practical basis. Any module that makes use of CLR Analyzeroutput will probably be able to infer more from Causal CLRs than from Temporal orConjunctive CLRs9. This suggests a performance preference for choosing Causal CLRs.Of course, the user will always have the option of rejecting any suggestions made by theCLR Analyzer.

Causation vs. Temporal Co-occurrence

See Causation vs. Temporal Precedence for a discussion on choosing between Causal andTemporal CLRs.

Enablement vs. Entailment

The resultant of Enablement tends to have weaker modals whereas the Entailment resultantcontains stronger modals.

9 For example, a module attempting to learn from CLR structures would be able to construct a rule fromCausal CLRs. Given a Conjunctive CLR, it may only be able to assert the CLR arguments as looselyrelated facts.

CLR Analysis - 14

IF the power is on, the computer can work. (enab)

IF the program responds, the computer must be working. (entl)

Enablement vs. Prevention

These two CLRs differ both in strength of modal in the resultant (Enablement—weakermodal, Prevention—stronger modal) and in resultant/connective polarity.

IF the power is on, the printer can be used. (enab)

IF the power is off, the printer will not print. (prev)

Enablement vs. Detraction

It is very difficult to distinguish between Enablement and Detraction.

IF the display is broken, the printer can still print files. (detr)

IF the power is on, the printer can quickly print files. (enab)

By definition, Enablement is an insufficient positive Causal CLR and Detraction is aninsufficient negative Causal CLR. Both CLRs are likely to have weaker modals in theresultant. If the CLR arguments have different polarities with a positive connective or thesame polarities with a negative connective, Detraction should be chosen. Otherwise, wecannot disambiguate between the two.

Enablement vs. Conjunction

None of the connectives in the CLR Marker Dictionary mark both Enablement andConjunction. Therefore, it is unlikely that the CLR Analyzer will ever need to distinguishbetween them.

Enablement vs. Disjunction

None of the connectives in the CLR Marker Dictionary mark both Enablement andDisjunction. Therefore, it is unlikely that the CLR Analyzer will ever need to distinguishbetween them.

Enablement vs. Temporal Precedence


Enablement vs. Temporal Co-occurrence


Entailment vs. Prevention

Although Entailment and Prevention are both sufficient Causal CLRs, they differ inpolarity. With a positive connective Entailment should have a positive resultant andPrevention should have a negative resultant. If the connective is negative, Entailment

CLR Analysis - 15

should have a negative resultant and Prevention should have a positive resultant. ThePolarity Reversal Rule applies in all cases.

IF the printer has paper, the file will print. (entl—pos. conn.)

IF the printer is out of paper, the file will not print. (prev—pos. conn.)

UNLESS the printer has paper, the file will not print. (entl—neg. conn.)

UNLESS the printer is out of paper, the file will print. (prev—neg. conn.)

After application of the Polarity Reversal Rule, the desired CLR interpretation for the firstand third examples will be

'the printer has paper' <entails> 'the file will print'

The interpretation for the second and fourth examples will be

'the printer is out of paper' <prevents> 'the file will print'

Entailment vs. Detraction

The Entailment CLR tends to have stronger modals in the resultant while the DetractionCLR tends to have weaker modals.

IF the system is unresponsive, the computer may still be working. (detr)

IF the system responds, the computer must be functioning. (entl)

Entailment vs. Conjunction

None of the connectives in the CLR Marker Dictionary mark both Entailment andConjunction. Therefore, it is unlikely that the CLR Analyzer will ever need to distinguishbetween them.

Entailment vs. Disjunction

None of the connectives in the CLR Marker Dictionary mark both Entailment andDisjunction. Therefore, it is unlikely that the CLR Analyzer will ever need to distinguishbetween them.

Entailment vs. Temporal Precedence


Entailment vs. Temporal Co-occurrence


CLR Analysis - 16

Prevention vs. Detraction

The resultant of Prevention must be negative (although possibly implicitly negative) whenmarked by a positive connective; the resultant of Detraction need not be negative. If theresultant is negative with a positive connective (or positive with a negative connective) or ifthere is any implicit negation in the arguments, it will probably not be possible toautomatically disambiguate between these two CLRs.

IF the system is hung, the computer will not work. (prev)

IF the system is unresponsive, the computer will still work. (detr)

Prevention vs. Conjunction

None of the connectives in the CLR Marker Dictionary mark both Prevention andConjunction. Therefore, it is unlikely that the CLR Analyzer will ever need to distinguishbetween them.

Prevention vs. Disjunction

The only connectives that mark both Prevention and Disjunction are ‘or’ and ‘either-or’.Since there are no readily available sufficient syntactic distinctions between their respectivearguments, a pragmatic choice should be made. Since Disjunction is only marked by or and‘either-or’ (whereas Prevention is marked by several other markers), Disjunction will besuggested to the user. In fact, the Prevention marked by ‘or’ and ‘either-or’ is a veryspecific (and possibly infrequent) usage.

The program must terminate OR the system will crash. (prev)

The program will terminate OR it will fail to terminate. (disj)

Prevention vs. Temporal Precedence

None of the connectives in the CLR Marker Dictionary mark both Prevention and TemporalPrecedence. Therefore, it is unlikely that the CLR Analyzer will ever need to distinguishbetween them.

Prevention vs. Temporal Co-occurrence


Detraction vs. Conjunction

See Causation vs. Temporal Precedence for a discussion on choosing between Causal andConjunctive CLRs.

Detraction vs. Disjunction

None of the connectives in the CLR Marker Dictionary mark both Detraction andDisjunction. Therefore, it is unlikely that the CLR Analyzer will ever need to distinguishbetween them.

CLR Analysis - 17

Detraction vs. Temporal Precedence

None of the connectives in the CLR Marker Dictionary mark both Detraction and TemporalPrecedence. Therefore, it is unlikely that the CLR Analyzer will ever need to distinguishbetween them.

Detraction vs. Temporal Co-occurrence

The only case where Detraction and Temporal Co-occurrence compete is a weak example ofDetraction. Temporal Co-occurrence should be preferred (overriding the general rule inCausation vs. Temporal Precedence for preferring Causal CLRs over Temporal CLRs).

The program suggests one interpretation WHILE the user prefers another.(detr)

The backup continues WHILE normal processing takes place. (ctmp)

Conjunction vs. Disjunction

None of the connectives in the CLR Marker Dictionary mark both Conjunction andDisjunction. Therefore, it is unlikely that the CLR Analyzer will ever need to distinguishbetween them.

Conjunction vs. Temporal Precedence

The only case where Conjunction and Temporal Precedence compete is a weak example ofTemporal Precedence. Conjunction should be preferred.

The program ran AND the printer printed. (conj)

The program computed the value AND it terminated. (prec)

Conjunction vs. Temporal Co-occurrence

There is little syntactic information to distinguish between Conjunction and Temporal Co-occurrence. Since Conjunction is relatively semantically vacuous, Temporal Co-occurrencewill be the preferred suggestion.

The program ran AND the printer printed. (conj)

The program ran AND the user waited. (ctmp)

Disjunction vs. Temporal Precedence

None of the connectives in the CLR Marker Dictionary mark both Disjunction andTemporal Precedence. Therefore, it is unlikely that the CLR Analyzer will ever need todistinguish between them.

Disjunction vs. Temporal Co-occurrence

None of the connectives in the CLR Marker Dictionary mark both Disjunction andTemporal Co-occurrence. Therefore, it is unlikely that the CLR Analyzer will ever need todistinguish between them.

CLR Analysis - 18

Temporal Precedence vs. Temporal Co-occurrence

None of the connectives in the CLR Marker Dictionary mark both Temporal Precedenceand Temporal Co-occurrence. Therefore, it is unlikely that the CLR Analyzer will ever needto distinguish between them.

2 . 3 Related Work

The following review of related work includes many older, more seminal references in thefields of Semantics and Discourse Analysis, as well as more recent proposals in theseareas.

2 . 3 . 1 CLRs

Several authors have proposed sets of relationships between clausal (and larger) textelements. However, as mentioned in section 2.1, most of these sets have been presented inthe context of Discourse Analysis. The resulting sets of “Discourse Relations” cover both“Subject Matter” and “Presentational” relations (Mann & Thompson, 1988). Thisdistinction has also been investigated by van Dijk (1977).

The goal of the TANKA project is to model the domain knowledge in a text about thedomain. The Pragmatic relations in a text do not represent domain knowledge. Rather, theypromote text coherence. These relations should not be represented by links in theConceptual Network and therefore, are not included in our CLR set.

One of the earlier attempts in Computational Linguistics to enumerate a set of semanticrelationships between propositions was an extension to Schank’s Conceptual DependencyTheory (Schank, 1975). Although the original theory only defined primitive acts and therelationships between acts and their participants, a later refinement to the theory (Schank &Abelson, 1977) introduced Conceptual Relations (including Enable, Result, Reason, andInitiate) to capture the semantic relationships between acts.

Halliday & Hasan (1976) present several Conjunctive Relations divided into two “planes”of relations: “External” and “Internal”. This division corresponds roughly to Mann &Thompson’s Subject Matter vs. Presentational. The Conjunctive Relations are furtherdivided into Additive, Adversative, Causal and Temporal relations. There is a closecorrespondence of the External relations within these categories to our Conjunctive,negative Causal, positive Causal and Temporal CLRs.

Van Dijk (1977) gives a thorough treatment of “Semantic” vs. “Pragmatic” considerationsin the study of discourse. The types of “Semantic Connections” include Conjunction,Disjunction, Conditionals (including causal relations) and Contrastives.

Hobbs (1983) presents a taxonomy of Coherence Relations that includes Enablement,Cause and Contrast (similar to our Detraction) along with several Pragmatic relations.

Bäcklund (1984) identifies six types of clauses (Temporal, Concessive, Conditional,Comparative, Conditional-Concessive and Locative) defined by their semantic functionwithin the discourse. For each type a set of connectives that typically introduce clauses ofthat type are enumerated. For example, Temporal clauses are introduced by the connectives‘when’, ‘whenever’, ‘after’, ‘as long as’, etc. The Temporal functions are further dividedinto temporal relations and non-temporal relations introduced by the temporal connectives.

CLR Analysis - 19

Mann & Thompson’s Rhetorical Structure Theory offers a list of Rhetorical Relations(RST) that includes relations roughly corresponding to all of our CLRs (Mann &Thompson 1986a, 1986b, 1988). However, as with Hobbs’ Coherence Relations, Mann &Thompson’s list is aimed at capturing both Subject-Matter and Presentational relations.Several authors have proposed sets of relations based on RST (including Hovy (1993),Lascarides et al. (1992) and Sanders et al. (1992) among others).

Schiffrin (1987) presents a list of Discourse Relations based on a study of the lexical items(Discourse Markers) that signal them. Although many of the Markers are unique toconversational speech (such as ‘oh’, ‘well’, ‘y’know’, etc.), the resulting set of Relationshas much in common with our CLRs: Conjunctive Relations (Conjunctive, Coordinative,Continuative, Contrastive, Disjunctive), Causal Relations (Cause-Result, Warrant-Inference, Motive-Action) and Temporal Relations (Reference Time, Event Time,Discourse Time).

Dahlgren (1988) summarizes the work of a number of researchers (including Cohen, Fox,Grosz & Sidner, Hirst, Hobbs, Lockman & Klappholz, Litman & Allen, Mann &Thompson, Polanyi & Scha and Reichman). The summary results in a list of twentyCoherence Relations: Sequence, Reported Event, Enablement, Cause, Goal, Parallel,Contrast, Evidence, Elaboration, Generalization, Restatement, Qualification, Evaluation,Description, Situation-Activity, Situation-Time, Situation-Place, Import, UnbiasedComment and Biased Comment. Many of these relations (such as Evaluation and BiasedComment) are unique to Discourse Analysis and therefore have no corresponding CLRs inour set.

Knott & Dale (1993) describe a method of motivating a set of Rhetorical Relations (basedon the set of relations defined in Mann & Thompson’s RST) through inspection of a largecorpus. The process involved scanning a corpus for “cue phrases” (defined as “phraseswhose function it is to link spans of discourse together”). These cue phrases (similar to ourlist of CLR Markers) were then used as a basis for classification of relations.

Kehler (1993) hints at a set of Coherence Relationships that includes Contrast,Comparison, Result, etc. In his frame representation for a clause, there is also reference totense, polarity and modality. However, the Relationships have not yet been fullyenumerated and the potential uses of tense, polarity and modality have not been investigated(Kehler, 1993: personal communication).

2 . 3 . 2 Tense and Modality

Heremerén (1978) presents a list of several semantic modalities and the particular modalauxiliaries that mark them. The semantic modalities are arranged in hierarchies indicating“strength”. Those modalities higher up in the hierarchy logically imply the modalities belowthem. For example, Certainty of a proposition implies Prediction of that proposition;Prediction of the proposition implies Probability; Probability implies Possibility, etc.

Palmer (1979) considers modality only as marked by the English modals. He identifies two“degrees” of modality: Necessity and Possibility, which roughly correspond to themodalities expressed by our stronger and weaker modals (respectively).

Coates (1983) lists the semantic modalities expressed by each of the English modals. Thesemodalities express modal “strength” in the labels she gives them (e.g. Strong Obligationvs. Weak Obligation; Confident Inference vs. Tentative Inference; etc.). The mappingbetween the modals and these modalities implies a strong/weak classification of the modalsthemselves.

CLR Analysis - 20

Quirk et al. (1985) give a similar mapping between modal auxiliaries and semanticmodalities. That work also makes note of how tense affects modality. For example, thepast tense of modals often express a hypothetical reading of the modalities.

3 Practice

While section 2 dealt with theoretical aspects, this section will deal with the issues specificto implementing HAIKU’s CLR Analyzer. section 3.1 shows the existing DIPETT andHAIKU structures relevant (as input) to CLR Analysis as well as CLR output structures.section 3.2 describes the actual analysis of CLRs as implemented in the CLR Analyzer andsection 3.3 presents results of an experiment using the Analyzer.

3 . 1 Structures

3 . 1 . 1 Syntactic Input

The input to CLR Analysis consists of several structures prepared by the HAIKU shell forthe purposes of semantic analysis. These structures include the Input String, the ParseTree, the Logical Clausal Structure and a list of parse tree fragments corresponding toeach clause in the input. For CLR Analysis, the relevant structures are the Logical ClausalStructure (which represents the syntactic organization of clauses in the sentence) and theclausal parse tree fragments. Table 3 shows the three basic variations of the format of theLogical Clausal Structure corresponding to different syntactic configurations of connectedclauses.

• coordination or subordination[ argument1, connective, argument2 ]

• correlative coordination[ [ coord1, argument1 ], [ coordinator2, argument2 ] ]

• subordinator/conjunct correlation[ coordinator, [ argument1, subordinator, argument2 ] ]

Table 3: Valid Logical Clausal Structures

The first format is the most common: a connective linking two arguments. The connectivecan be either a coordinate conjunction (such as ‘and’) or a subordinator (such as ‘until’).Each argument can be either a pointer to the corresponding clausal parse tree fragment or anembedded Logical Clausal Structure. An example sentence producing this type of LogicalClausal Structure might be

The program will print UNTIL the printer runs out of paper.

The second format corresponds to the linking of two clauses by a correlative (‘either-or’)10.Again, each argument can refer to a clause or an embedded Logical Clausal Structure.

EITHER the system crashed OR somebody killed the job.

10 Although ‘both-and’ and ‘neither-nor’ are also valid correlatives, they cannot be used to correlate wholeclauses (see Quirk, et al., 1985:13.35,37).

CLR Analysis - 21

The third format corresponds to the linking of two clauses by a two-part subordinator (seeQuirk et al., 1985:8.145).

IF the power is on THEN the display should be working.

3 . 1 . 2 CLR Output Structures

Following CLR Analysis, the structures prepared by the HAIKU shell are supplementedwith the Clause-Level Relationships structure:

[clr_structure(<clr_label>,[<arg1>,<arg2>,...])]

In the structure, <clr_label> is the CLR assigned to the input and <arg1> and <arg2>are either pointers to the appropriate clauses in the list of clausal parse tree fragments orembedded CLR structures. The argument list accommodates more than two arguments forthe case of n-ary CLRs (see section 2.1.2). CLR output for some example sentences ispresented in Appendix II.

3 . 2 Algorithms

3 . 2 . 1 Basic CLR Analysis

The CLR Analyzer consists of two main modules. The CLR Driver interprets the inputfrom the HAIKU shell, determines the correct output representation and invokes the CLRAssignment module. This Driver handles the various syntactic formats of connectedclauses, clause nesting as well as user interaction and CLR structure building.

For nested structures, analysis begins with the innermost nested pairs (or sequences) ofclauses and proceeds to the outermost relationships. Consider the following example:

The printer can print if the program issues the print command before the jobterminates.

For this sentence the HAIKU shell will produce the following Logical Clausal Structure:

[*statement1*,'if',[*statement2*,'before',*statement3*]]

where *statement1* refers to the parse tree fragment corresponding to ‘The printer canprint’, *statement2* refers to the fragment for ‘the program issues the print command’and *statement3* refers to ‘the job terminates’. The CLR Driver will first invoke theAssignment module to label the relationship between *statement2* and *statement3* .This relationships constitutes a composite statement S). Next, it will invoke the Assignmentmodule to assign a label to the relationship between *statement1* and S.

The CLR Assignment module chooses the best CLR to suggest to the user for verification.The CLR Marker Dictionary is consulted for the CLRs marked by the given connective.From these CLR candidates, if this module manages to determine a single best CLR for therelationship, that CLR will be submitted to the user for acceptance or rejection. If it cannotdecisively disambiguate between the valid CLRs, the user will be asked to select the mostappropriate CLR from the list of candidates for this input. In either case, the user can accepta suggested CLR or reject the suggestions and enter a CLR of his own choosing.

For the nested example sentence above, the resulting CLR structure might be:

CLR Analysis - 22

[clr_structure(enab, [clr_structure(prec,[*statement2*,*statement3*])], *statement1*)]

3 . 2 . 2 CLR Competitions

Choosing a single CLR from the set of valid CLRs for an input is based on the preferencerules presented in section 2.2.3. Each of those rules is a heuristic for choosing between agiven pair of candidate CLRs. However, if there are more than two candidates, theheuristics cannot be applied directly: they must be applied to individual pairs within the setof candidates. The simplest approach would be to apply the heuristics to the first two CLRsto determine the preferred of those two, then apply the heuristics to the third CLR and the“winner” of the first competition, etc.

For example, suppose there are four candidate CLRs (A, B, C, D) for some input.Suppose further that the heuristics provide the following preferences for this same input:

Prefer OverB AB DC BD C

Using the simple approach described above, the system would select D as the mostappropriate CLR for the input. However, if the selection algorithm worked backwardsthrough the list starting with the last two candidates, it would select B as the mostappropriate CLR. This deviant behaviour is obviously undesirable.

A better approach (the one implemented in the CLR Analyzer) would be to holdcompetitions between all pairs of candidate CLRs and choose the CLR with the bestpreference record.

Each candidate CLR competes against all other candidates. Each time a candidate ispreferred over another candidate, the “winner” collects two points, while the “loser”collects no points. If the heuristics are unable to prefer one candidate over another, thematch is declared a tie and each candidate receives a single point. Once all matches havebeen played, each competitor’s points are examined to determine if there is a single CLRthat was preferred over all other CLRs. If there is such a victor, it is presented to the useras the single best suggestion for the current input. If no single CLR won all its matches, theuser is presented with all competitors with at least one point, starting with the CLR with themost points. A competitor failing to gain a single point represents an unlikely choice for aCLR for the input (since all other candidates were preferred over it). CLRs with zero pointsare never presented to the user.

The example above would produce the following standings:

CLR Wins Losses Ties PointsA 0 1 2 2B 2 1 0 4C 1 1 1 3D 1 1 1 3

CLR Analysis - 23

3 . 3 Experimental Results

To check the accuracy of the preference rules and the validity of the “competition” model, atest was conducted with 100 sentences from the Ontario Building Code (Ontario Ministryof Housing, 1991). The Ontario Building Code is the legal document setting out theregulations for the design and construction of buildings in Ontario; it contains more than400,000 words. For the test, sentences were chosen containing clauses connected byconnectives known to mark two or more CLRs (from the CLR Marker Dictionary). Apartfrom this restriction, the sentences were chosen randomly from throughout the entire text.For each sentence, the CLR Analyzer held competitions to determine the most appropriateCLR based on syntactic features of the clauses (described in section 3.1 and 3.2) and thepreference rules (described in section 3.3). The results are summarized in Table 4.

Single CLR Chosen as Winner Single CLR with Most Points Multiple CLRs with Most Points

Right Wrong Right Wrong (ambiguous)

83 2 11 0 4

Table 4: CLR Competition Test Results

These results suggest that:

• the CLR set is adequate for the relationships in the given sentences;

• even if there is no single CLR that was preferred over all others, it is useful topresent the CLR with the most points to the user first (all 11 of thesesuggestions were correct for this test);

• the preference rules and CLR Marker Dictionary are accurate for this text (giventhe success rate of (83+11)/100);

• the syntactic features contain enough information for the system to choose asingle CLR (it was unable to suggest a single CLR in only 4 cases);

• the user is still required to correct wrong suggestions (columns 2 & 4) and todisambiguate when the system is unable to do so (column 5).

4 Limitations and Future Work

4 . 1 Embedded CLRs

All of the techniques described in section 2 for CLR Analysis were for disambiguatingbetween candidate CLRs based on the syntactic features of the CLR arguments. However,as shown in section 3.2.1, the CLR arguments may be embedded CLR structures. Theseembedded CLR structures have no distinct modality or polarity. The system is thereforeunable to suggest a single CLR or a subset of the candidates. The user is required to choosea CLR from the complete list of valid CLRs for a given input.

One partial solution to this problem would be to attempt a CLR competition with theembedded CLR argument. Many of the preference rules check the features of only one ofthe CLR arguments (often the resultant). If the input contains one embedded CLR structure

CLR Analysis - 24

and one clausal argument, a CLR competition may still be possible. Those heuristicsrequiring syntactic features of the embedded CLR argument would fail, resulting in a tie,but those heuristics only requiring features of the clausal argument would usually be able tochoose a winner.

4 . 2 Argument Ordering

Section 2.1.1 described a method for automatically determining the correspondencebetween syntactic arguments and semantic arguments. For some connectives, the firstsyntactic argument corresponds to the first semantic argument while for other connectives,the order of arguments is reversed. The CLR heuristics depend on the order of thearguments. If the assumption regarding the ordering is incorrect, the heuristics are likely toproduce meaningless results. Of course, the user is always given the option of reversingthe arguments and assigning the correct CLR.

4 . 3 Implicit Polarity

Many of the heuristics make use of the polarity of the CLR arguments. As noted in section2.2.1, however, an argument’s polarity is often implicit in the semantics of the clause’sverb or complement. Since the system does not have any lexical semantic knowledge, itcannot accurately determine the polarity of an implicitly negative clause.

4 . 4 Non-Auxiliary Modality

The modality of a clause is not always conveyed by a modal auxiliary. Adverbs (such as‘possibly’ and ‘certainly’) and modal paraphrases (such as ‘it is possible that...’ and ‘it iscertain that...’) are often used to express modality. Again, in the absence of lexicalsemantics, the CLR Analyzer is unable to interpret these forms as expressing modality.

However, it might be tractable to pre-code this lexical semantic information, since the set of“implicitly modal” words is probably quite small and closed. The same may be true forimplicitly negative words. A close dictionary inspection would be necessary to confirmthese suspicions.

5 Conclusion

This report has described a system that interprets the semantic relationships between actsrepresented by syndetically connected clauses. The use of syntactic features (tense,modality and polarity) together with knowledge of the usages of connectives enables theCLR Analyzer to be selective about which CLRs to suggest to the user for verification.Often, the system will propose a single CLR for a given input.

The promising results of early tests on the CLR Analyzer suggest that it has a valuablecontribution to make in automating semantic analysis. The system’s independence of mucha priori semantic domain knowledge makes it particularly well suited to the task ofKnowledge Acquisition from large, real-world texts.

CLR Analysis - 25

Acknowledgments

This work is supported by the Natural Sciences and Engineering Research Council ofCanada. Several people have contributed to the work described in this report.

Members of the TANKA team did important early work on semantic analysis in HAIKU andprovided invaluable direction and support for the current work: Terry Copeck, SylvainDelisle, Stan Szpakowicz. The MaLTe team continues to offer a stimulating forum forpresenting research as well as providing a practical focus: Jean-François Delannoy, CaoFeng, Louis Massey, Stan Matwin, Messaouda Ouerd, Riverson Rios, Stan Szpakowicz.

References

BÄCKLUND , INGEGERD (1984). Conjunction-Headed Abbreviated Clauses in English.Stockhom: Almqvist & Wiksell, International.

BARKER, KEN, TERRY COPECK, SYLVAIN DELISLE & STAN SZPAKOWICZ (1993). “AnEmpirically Grounded Case System”. submitted to the International Journal ofLexicography; an earlier version is available as TR-93-08, Department of ComputerScience, University of Ottawa.

COATES, JENNIFER (1983). The Semantics of the Modal Auxiliaries. London: CroomHelm.

COPECK, TERRY, SYLVAIN DELISLE, & STAN SZPAKOWICZ (1992). “Parsing and CaseAnalysis in TANKA ”. Proceedings of COLING-92. Nantes, France, 1008-1012.

DAHLGREN, KATHLEEN (1988). Naïve Semantics for Natural Language Understanding.Boston: Kluwer Academic Publishers.

DELANNOY, JEAN-FRANÇOIS, CAO FENG, STAN MATWIN & STAN SZPAKOWICZ (1993).“Knowledge Extraction from Text: Machine Learning for Text-to-Rule Translation”,Proceedings of the Machine Learning and Text Analysis Workshop, ECML-93, Vienna,1-7.

DELISLE, SYLVAIN (1994). “Text Processing without A-Priori Domain Knowledge: Semi-Automatic Linguistic Analysis for Incremental Knowledge Acquisition”, Ph.D. thesis, TR-94-02, Department of Computer Science, University of Ottawa.

DELISLE, SYLVAIN , KEN BARKER, TERRY COPECK & STAN SZPAKOWICZ (1993).“Interactive Semantic Analysis of Technical Texts”. submitted to ComputationalIntelligence (in revision); an earlier version is available as TR-93-13, Department ofComputer Science, University of Ottawa.

DELISLE, SYLVAIN , KEN BARKER, JEAN-FRANÇOIS DELANNOY, STAN MATWIN & STANSZPAKOWICZ (1994). “From Text to Horn Clauses: Combining Linguistic Analysis andMachine Learning”. to appear in Canadian AI-94.

HALLIDAY , M.A.K. & RUQAIYA HASAN (1976). Cohesion in English. London:Longman.

CLR Analysis - 26

HERMERÉN, LARS (1978). On Modality in English: A Study of the Semantics of theModals. Lund: CWK Gleerup.

HOBBS, JERRY (1983). “Why is Discourse Coherent?”. Fritz Neubauer (ed.). Coherencein Natural Language Texts. Hamburg: Buske.

HOVY, EDUARD H. (1993). “Automated Discourse Generation Using Discourse StructureRelations”. Fernando C.N. Pereira & Barbara J. Grosz (eds.). Artificial Intelligence 63:1-2(October, 1993), 341-385.

K EHLER, ANDREW (1993). “The Effect of Establishing Coherence in Ellipsis andAnaphora Resolution”. Proceedings of the 31st Annual Meeting of the Association forComputational Linguistics. Columbus, Ohio, 62-69.

KNOTT, ALISTAIR & ROBERT DALE (1993). “Using Linguistic Phenomena to Motivate aSet of Rhetorical Relations”. Human Communication Research Centre, University ofEdinburgh.

L ASCARIDES, ALEX , NICHOLAS A SHER & JON OBERLANDER (1992). “InferringDiscourse Relations in Context”. Proceedings of the 30th Annual Meeting of theAssociation for Computational Linguistics. Newark, Delaware, 1-8.

LEVI, JUDITH N. (1978). The Syntax and Semantics of Complex Nominals. New York:Academic Press.

LYONS, JOHN (1977). Semantics. Cambridge: Cambridge University Press.

M ANN , WILLIAM C. & SANDRA A. THOMPSON (1986). “Relational Propositions inDiscourse”. Discourse Processes 9:1 (1986), 57-90.

MANN, WILLIAM C. & SANDRA A. THOMPSON (1986). “Rhetorical Structure Theory:Description and Construction of Text Structures”. Gerard Kempen (ed.). Natural LanguageGeneration: New Results in Artificial Intelligence, Psychology and Linguistics. Dordrecht:Martinus Nijhoff Publishers.

MANN, WILLIAM C. & SANDRA A. THOMPSON (1988). “Rhetorical Structure Theory:Toward a functional theory of text organization”. Text 8:3 (1988), 243-281.

ONTARIO M INISTRY OF HOUSING (1991). “The Ontario Building Code”. OntarioRegulation 413/90. Queen’s Printer for Ontario.

PALMER, F. R. (1979). Modality and the English Modals. London: Longman.

QUIRK, RANDOLPH, SIDNEY GREENBAUM, GEOFFREY LEECH & JAN SVARTVIK (1985).A Comprehensive Grammar of the English Language. London: Longman.

SANDERS, TED J.M., WILBERT P.M. SPOOREN & L EO G.M. NOORDMAN (1992).“Toward a Taxonomy of Coherence Relations”. Discourse Processes 15 (1992), 1-35.

SCHANK, ROGER C. (1975). Conceptual Information Processing. Amsterdam: North-Holland Publishing Company.

SCHANK, ROGER C. & R.P. ABELSON (1977). Scripts, Plans, Goals and Understanding.Hillsdale, NJ: Erlbaum.

CLR Analysis - 27

SCHIFFRIN, DEBORAH (1987). Discourse Markers. Cambridge: Cambridge UniversityPress.

VAN D IJK, TEUN A. (1977). Text and Context. Explorations in the Semantics andPragmatics of Discourse. London: Longman.

YANG, LEIXUAN & STAN SZPAKOWICZ (1991a). “Inheritance in Conceptual Networks”.Proceedings of the Sixth International Symposium on Methodologies for IntelligentSystems. Charlotte, NC, 191-202.

YANG, LEIXUAN & STAN SZPAKOWICZ (1991b). “Planning in Conceptual Networks”. F.Dehne, F. Fiala and W.W. Koczkodaj (eds.). Advances in Computing and Information -ICCI ’91. Lecture Notes in Computer Science 497, Springer-Verlag, 669-671.

YANG, LEIXUAN & STAN SZPAKOWICZ (1994). “Path-Finding in Networks” ICCI-94. (toappear).

CLR Analysis - 28

APPENDIX I - THE CLR MARKER DICTIONARY

This appendix lists the Clause-Level Relationship Markers and the CLRs they mark. EachMarker entry includes the Marker, the polarity of the Marker (see section 2.2.1) and a listof CLRs. For each CLR the direction of the CLR arguments relative to the direction of thesyntactic arguments is given along with an example sentence.

Marker ArgumentMarker Polarity CLR Direction Example

after pos prec reverse We came AFTER THEY LEFT.enab reverse The printer can print AFTER THE PAPER TRAY HAS BEEN FILLED

also pos conj same He was mean, ALSO he was ugly.

although neg detr reverse He wouldn't come ALTHOUGH I ASKED HIM.

and pos conj same He ate AND she drank.prec same SHE READ FOR AN HOUR AND she went to bed.

as pos ctmp same We were leaving AS they arrived.caus reverse AS IT WAS GETTING LATE, we adjourned.enab reverse The printer can print AS IT HAS POWER.entl reverse AS THE FILE WAS PRINTED, the job must have run.prev reverse AS THE JOB FAILED, the file was not printed.

because pos caus reverse The program died BECAUSE THE MACHINE HAD NO MORE MEMORY.entl reverse The job must have been killed BECAUSE THE LOG SHOWS IT.enab reverse The printer should work BECAUSE THE POWER IS ON.prev reverse The file did not print BECAUSE THE JOB FAILED.

before pos prec same YOU'D BETTER BE GONE BEFORE I come.enab same YOU MUST TURN ON THE COMPUTER BEFORE you can use it.

but neg prev reverse We would attend BUT WE HAVE NO MONEY.enab reverse The printer would not have printed BUT THE POWER WAS ON.detr same WE HAVE NO MONEY BUT we will attend.

considering pos detr reverse CONSIDERING SHE ARRIVED ONLY RECENTLY she is well adjusted.prev reverse CONSIDERING I HAVE NO MONEY I will be unable to attend.

'either-or' neg disj same EITHER I will do homework OR I will watch television.prev same EITHER give yourself up OR I will shoot.

except neg enab reverse The file would not have printed EXCEPT THE PROGRAM SUCCEEDED.prev reverse We would attend EXCEPT WE HAVE NO MONEY.

even_if pos detr reverse We will not attend EVEN IF WE ARE INVITED.

CLR Analysis - 29

for pos caus reverse The file did not print FOR THE JOB FAILED.enab reverse The printer can print FOR IT HAS POWER.entl reverse The file must have printed FOR THE JOB SUCCEEDED.prev reverse The file must not have printed FOR THE JOB FAILED.

if pos prev reverse IF THE POWER IS OFF, the printer won't work.enab reverse IF THE POWER IS ON, the printer can be used.entl reverse IF THE LIGHT IS ON, he must be home.

'if-then' pos prev same IF THE POWER IS OFF THEN the printer won't work.enab same IF THE POWER IS ON THEN the printer can be used.entl same IF THE LIGHT IS ON THEN he must be home.

lest neg prev same MAKE A NOTE LEST you forget.

notwithstanding neg detr reverse He wouldn't come NOTWITHSTANDING I ASKED HIM.

once pos prec reverse ONCE HE HAS COME you can go.enab reverse ONCE THE POWER IS ON the printer can be used.

only neg prev reverse We would attend ONLY WE HAVE NO MONEY.

or neg disj same I will do homework OR I will watch television.prev same Give yourself up OR I will shoot.

provided pos entl reverse You will graduate PROVIDED YOU PASS YOUR COURSES.

provided_that pos entl reverse You will graduate PROVIDED THAT YOU PASS YOUR COURSES.

providing pos entl reverse You will graduate PROVIDING YOU PASS YOUR COURSES.

providing_that pos entl reverse You will graduate PROVIDING THAT YOU PASS YOUR COURSES.

rather pos detr reverse He didn't do it, RATHER HE SAID WHO DID.

save_that neg detr reverse We would be ruined SAVE THAT THE TORNADO MISSED US.

saving_that neg detr reverse We would be ruined SAVING THAT THE TORNADO MISSED US.

seeing_that pos detr reverse SEEING THAT SHE ARRIVED ONLY RECENTLY she is well adjusted.enab reverse The printer should work SEEING THAT THE POWER IS ON.

since pos prec reverse I've been unhappy SINCE THEY LEFT.caus reverse SINCE THE PROGRAM SUCCEEDED, the file was printed.prev reverse SINCE THE PROGRAM FAILED, the file was not printed.entl reverse SINCE THE PROGRAM IS CORRECT, the file must print.enab reverse SINCE HE HAS THE MONEY, he may attend.

so pos caus same HE CAME SO I left.enab same THE POWER IS ON SO the printer can print.entl same THE PROGRAM IS CORRECT SO it must terminate.prev same THE PROGRAM IS BUGGY SO the file will not print.

that pos caus same SHE DID SO WELL THAT I had to congratulate her.

CLR Analysis - 30

then pos prec same SHE READ FOR AN HOUR THEN she went to bed.

therefore pos caus same I THINK THEREFORE I am.enab same I AM HUMAN THEREFORE I can think.entl same THERE IS SMOKE THEREFORE there must be fire.

though neg detr reverse He wouldn't come THOUGH I ASKED HIM.

till neg prec same SHE'LL STAY AWAKE TILL she finishes the book.enab reverse The printer won't work TILL THE POWER IS ON.

unless neg prev reverse It will happen UNLESS I OPPOSE IT.entl reverse He will not be home UNLESS THE LIGHT IS ON.

until neg prec reverse SHE'LL STAY AWAKE UNTIL she finishes the book.enab reverse The printer won't work UNTIL THE POWER IS ON.

when pos ctmp same I always call WHEN the clock strikes twleve.caus reverse He is upset WHEN HIS TEAM LOSES.enab reverse We can use the printer WHEN IT IS TURNED ON.

whenever pos ctmp same I always call WHENEVER the clock strikes twleve.caus reverse He is upset WHENEVER HIS TEAM LOSES.enab reverse You can use the printer WHENEVER IT IS TURNED ON.

whereas pos detr reverse He wants three WHEREAS SHE WANTS FOUR.caus reverse WHEREAS ALL ARE AGREED it shall be law.

whereupon pos ctmp same Six birds flew WHEREUPON the sun shone.enab same A WINDOW OPENED WHEREUPON it rained in.

whether pos detr reverse WHETHER SHE IS YOUNG OR NOT she is still very healthy.

while pos detr reverse He wants three WHILE SHE WANTS FOUR.ctmp same He always comes WHILE I'm away.

yet neg conj same The economy is bad YET we have our health.detr reverse We will attend YET WE HAVE NO MONEY.

CLR Analysis - 31

APPENDIX II - SAMPLE CLR ANALYSIS SESSION

Example 1:

Enter sentence:

The printer can print because the power is on.

Unit #1: the printer can print because the power is on .

First parse for input sentence:...

HAIKU: initiating semantic analysis...

LOGICAL CLAUSAL STRUCTURE

top_structure

[*statement1*,subordinator(because),*statement2*]

SYMBOLIC A-LISTS

clauses

[*statement1*,statement(...)]


INPUT STRING([the,printer,can,print,because,the,power,is,on,.])

HAIKU: Clause-Level Relationship Analysis of current input ...

> There is a Clause-Level Relationship marked by 'because':

'the printer can print'

'because'

'the power is on'

CLR competition between caus and entl... entl wins.

CLR competition between caus and enab... enab wins.

CLR competition between caus and prev... caus wins.

CLR competition between entl and enab... enab wins.

CLR competition between entl and prev... entl wins.

CLR competition between enab and prev... enab wins.

> The CLR Analyzer's best suggestion(s) for this input:

Enablement (enab) (1)

> Please enter a number between 1 and 1

or enter a valid CLR label for this relationship (CR to abort): 1

> Your CLR assignment will be stored as:

'the power is on'

<enables>


> Do you accept this assignment

(enter r to reverse the arguments, a to abort) [Y/n/r/a]? y

CLAUSE-LEVEL RELATIONSHIPS

clr_structure(enab,[*statement12*,*statement11*])

CLR Analysis - 32

Example 2:

Enter sentence:

The printer must have printed because the program issued a "print" command.

Unit #2: the printer must have printed because the program issued a " print " command .




top_structure

[*statement13*,*coord1*,*statement14*]

SYMBOLIC A-LISTS

coordinators

[*coord1*,coord(because,nil)]

clauses



INPUT STRING([the,printer,must,have,printed,because,the,program,issued,a,",print,",command,.])



'the printer must have printed'

'because'

'the program issued a " print " command'

CLR competition between caus and entl... entl wins.

CLR competition between caus and enab... enab wins.

CLR competition between caus and prev... caus wins.

CLR competition between entl and enab... entl wins.

CLR competition between entl and prev... entl wins.

CLR competition between enab and prev... tie.


Entailment (entl) (1)





<entails>

'the printer must have printed'




clr_structure(entl,[*statement14*,*statement13*])

CLR Analysis - 33

Example 3:

Enter sentence:

The printer didn't print because it was out of paper.

Unit #4: the printer didn't print because it was out of paper .




top_structure

[*statement15*,subordinator(because),*statement16*]

SYMBOLIC A-LISTS

clauses



INPUT STRING([the,printer,didn't,print,because,it,was,out,of,paper,.])



'the printer did not print'

'because'

'the printer was out_of paper'

CLR competition between caus and entl... caus wins.

CLR competition between caus and enab... caus wins.

CLR competition between caus and prev... prev wins.

CLR competition between entl and enab... entl wins.

CLR competition between entl and prev... prev wins.

CLR competition between enab and prev... prev wins.


Prevention (prev) (1)




'the printer was out_of paper'

<prevents>

'the printer did print'




clr_structure(prev,*statement16*,*statement15*)

CLR Analysis - 34

Example 4:

Enter sentence:

The printer can print because the power was on before the program issued a "print"

command.

Unit #9: the printer can print because the power was on before the program issued a " print "

command .




top_structure

[*statement30*,*coord3*,[*statement31*,subordinator([before,tokens([before])]),*statement32*]]

SYMBOLIC A-LISTS

coordinators

[*coord3*,coord(because,nil)]

clauses




INPUT STRING([the,printer,can,print,because,the,power,was,on,before,the,program,issued,a,",print,",

command,.])


> There is a Clause-Level Relationship marked by 'before':

'the power was on'

'before'


CLR competition between prec and enab... prec wins.


Temporal-Precedence (prec) (1)




'the power was on'

<temporally_precedes>




CLR Analysis - 35



'because'

'the power was on before the program issued a " print " command'


Causation (caus) (1),

Entailment (entl) (2),

Enablement (enab) (3),

Prevention (prev) (4)




'the power was on before the program issued a " print " command'

<enables>





clr_structure(enab,[clr_structure(prec,[*statement31*,*statement32*]),*statement30*])

CLAUSE-LEVEL RELATIONSHIP ANALYSIS IN THE TANKA SYSTEM

Documents

Transcript of CLAUSE-LEVEL RELATIONSHIP ANALYSIS IN THE TANKA SYSTEM