Semantic Uses of When ClausesSteven BethardDepartment of Computer ScienceUniversity of Colorado at BoulderBoulder, CO [email protected] 1, 2004AbstractIn English, while when-clauses serve a variety of semantic purposes, it has been claimedthat the ‘canonical’ when-clause serves one of three semantic purposes: to specify the time ofthe head-clause, to specify a time to which the head-clause is related, or to specify a type ofoccasion for which the head-clause describes an instance. This paper uses the Brown, Switch-board and Wall Street Journal corpora to analyze this claim in some detail, examining thefrequency of both temporal and atemporal uses of when-clauses. It concludes that this claimis not unjustified, though strictly ‘canonical’ forms occur between only 57% and 65% of thetime in the corpora studied.1 IntroductionEnglish when-clauses are generally thought of as temporal specifiers, identifying the time a givensituation takes place. The situations taking part in this relation are generally the situation describedin the when-clause and the situation described in the head-clause (in general, the clause that syn-tactically dominates the when-clause). For example,(1) John was in his bedroom when I was sleeping.Here, the sleeping situation in the when-clause is described as being simultaneous with the being-in-bedroom situation described by the head-clause.When-clauses are in fact much more versatile than simple temporal coordination however. Theycan express inclusion between the when-clause and head-clause (in either direction), as in 2 and13. They can express the anteriority or posteriority of the when-clause to the head-clause, as in4 and 5. Additionally, they can describe the starting or ending points of the head-clause situation(called ‘inchoative’ and ‘terminative’ uses, respectively), as in 6 and 7. They can even occasionallyindicate the time at which the speaker evaluates a situation (the ‘focalizing’ point), as in 8.(2) John was in his bedroom when I showed up at the door.(3) I showed up at the door when John was in his bedroom.(4) John woke me up when I went to sleep.(5) I had set my alarm for 6:00am when I went to sleep.(6) When I got into bed, I slept.(7) I was asleep when John woke me up.(8) When the smoke cleared, six people were dead.This brief illustration of some of the temporal uses of when-clauses touches only the surfaceof a rather complex semantic phenomenon. Declerck (1997) explores this area in much greaterdetail, proposing a theory of how tense, aspect and when-clause structure interact to produce thiswide variety of temporal interpretations. His typology divides when-clauses into eight syntactic-semantic classes:1. direct questions2. indirect questions3. relative clauses modifying temporal noun phrases4. nonrestrictive relative clauses without overt antecedents5. free relative noun clauses6. adverbial time clauses7. narrative when-clauses8. atemporal when-clausesOf these when-clause types, Declerck identifies a subtype of adverbial time clauses as the ‘canon-ical’ when-clauses. In this subtype, the when-clause serves one of three semantic purposes: tospecify the time of the head-clause (as in 9), to specify a time to which the head-clause is related(as in 10), or to specify a type of occasion for which the head-clause describes an instance (as in11).(9) John will leave when I arrive.2(10) When I arrive, John will have already left.(11) When John’s bored with a class, he’s uncooperative.Note that the first semantic purpose is the simultaneity use discussed with example 1, and thesecond semantic purpose includes all the other temporal uses of when-clauses discussed with ex-amples 2-8. The third semantic purpose of ‘canonical’ when-clauses, however, is in fact, a com-pletely atemporal use. Declerck’s description of this and a variety of other atemporal uses forwhen-clauses is summarized below.CASE-SPECIFYING when-clauses (12), which we discussed above as the third semantic classof ‘canonical’ when-clauses, specify the type of situation described in the head-clause. When-clauses that establish some sort of contrast with the situation in the head-clause Declerck callsADVERSATIVE when-clauses (13). In the CAUSAL class of when-clauses (14), the when-clausespecifies the cause of the situation in the head-clause. MANNER when-clauses (15) describe themanner or means by which the situation in the head-clause occurs. Declerck’s CONDITIONALwhen-clauses (16) are when-clauses that specify the conditions under which the situation in thehead-clause occurs. Finally, NARRATIVE when-clauses (17) do not modify the head-clause, butinstead introduce the next situation in the narration.(12) When the husband dies first, the widow often stays on her own.(13) The reports put sales at $1 million when they actually exceeded $29.(14) The riots stopped when the police intervened.(15) John provides evidence of this phenomenon when he reports that ...(16) You know I’m compliance itself – when I’m not thwarted.(17) I was sitting quietly in the kitchen when a stranger entered the room.The claim in (Declerck 1997) that ‘canonical’ when-clauses serve semantically as either tempo-ral or CASE-SPECIFYING clauses is taken to suggest that these semantic uses will be most commonin speech and text. To validate this claim, this paper considers an in-depth study of the semanticclasses of various when-clauses in several corpora.32 DataThe data used for this study is taken from three sources: the Brown corpus (Francis and Kucera1964), the Switchboard corpus (Godfrey, Holliman, and McDaniel 1992), and the Wall StreetJournal 1994 corpus. The Brown corpus is a 1 million word corpus representing 15 categoriesof American English text including letters, biographies, fiction, editorials and a variety of othergenres. The Switchboard corpus is a 3 million word corpus providing transcripts of conversationalspeech collected from the long distance telephone conversations of over 500 different speakersin the United States. The Wall Street Journal 1994 corpus is a 9 million word corpus consistingof a variety of financial news reports. These corpora were chosen because the represent threesubstantially different genres of text, and as such, allow the study here to observe whether theDeclerck hypothesis holds equally for all text types.For when-clause data, 100 sentences containing the word when were chosen randomly fromeach corpus. Each of these