TECHNICAL REPORT Avoid Ambiguity ! ( If You Can )

When people create linguistic expressions, they should avoid ambiguity. Current evidence on this is mixed. In two experiments, subjects read sentences including passive relative clauses, which can be written in full or reduced form (The team (that was) defeated in the Super Bowl vowed revenge); when reduced (without the that was), ambiguity is a threat. Subjects were told about the optional material, and instructed to include or omit it to make the sentence “easier to understand.” One experiment manipulated past participle ambiguity (The team defeated... is ambiguous, whereas The team beaten... is not). Another experiment also manipulated plausibility (The winning team defeated... is more ambiguous relative to The losing team defeated...). Past-participle ambiguity failed to influence whether subjects wrote full embedded clauses, but plausibility tended to. A third experiment verified that the ambiguous fragments are consistent with main-verb interpretations. Thus, when instructed to edit sentences to make them easy to understand, subjects avoid ambiguity not based on morphological ambiguity, but (perhaps) only as conditioned upon pragmatic, real-world knowledge.highlight certain common, but problematic, methodological practices in patient research, and alternative approaches are suggested. It’s right there in the Gricean maxim of manner (Grice, 1975): Avoid ambiguity. The reason for this directive is obvious: An ambiguous linguistic expression is difficult to understand, and so to produce one is to be uncooperative (unless of course the aim is to flout the maxim of manner, but that’s a different matter entirely). Therefore, speakers -or, more relevantly in the present context, writers -should avoid ambiguity when they can. The “avoid ambiguity” component of the Gricean maxim of manner is obviously important when full ambiguity is at stake. Full ambiguity arises when a complete linguistic expression can be interpreted in more than one way. The recommendation, “I can’t say enough good things about this candidate,” is fully ambiguous. (Is the candidate stronger than words can express? Or is the candidate not worthy of positive description?) However, temporary ambiguity too is likely to be problematic for linguistic communication. Temporary ambiguity arises when part-way through, a linguistic expression can be interpreted in more than one way, but by the end of the expression, the inappropriate interpretation becomes impossible. Temporary ambiguities are called garden paths when, at the point of ambiguity, linguistic or contextual factors encourage the ultimately incorrect interpretation. “The team defeated in the Super Bowl vowed revenge next season” is a garden path, because after reading “The team defeated...,” the main-verb interpretation (whereby defeat is a main verb and so the team did the defeating) is especially likely; however, by the end of the expression, this main verb interpretation proves incorrect, because defeat is actually a passive verb in a relative clause (implying that the team got defeated). Garden paths pose demonstrable problems for readers (as shown by the last few decades of research on language comprehension), and so, by the logic described above, writers should avoid them too. (In fact, once the wider contexts of utterances in natural language use are taken into account, this distinction between full and temporary ambiguity is likely to collapse. “I can’t say enough good things” is likely unambiguous in its larger context, though its tolerance of an inappropriate [here opposite] interpretation might pose temporary comprehension difficulties.) So, do speakers avoid full ambiguities or garden paths? Many investigations have explored this question. Two early (and conflicting) observations relevant to this point come from Elsness (1984). He looked at naturally occurring text from the Brown corpus (a variety of written forms, including books, periodicals, government documents, etc.), investigating “that” omission in fully ambiguous sentences like “The newspaper reported (that) yesterday (that) the politician had lied” (without one of the “thats,” it is uncertain whether the reporting or the lying happened yesterday). He found that sentences with such medial adverbials (“yesterday”) indeed included “thats” more often than other similar sentences, suggesting a disambiguating function for CRL Technical Reports, Vol. 18 No. 2, December 2006 4 the “that.” (Note, however, that it may be the presence of the adverbial per se, rather than its ambiguity, that may compel “that” inclusion.) In the same report, Elsness also assessed sentences like “The newspaper reported (that) you had lied.” Such sentences can include a temporary ambiguity, because upon its comprehension, “you” might either be the direct object of the preceding main verb, or the subject of an upcoming embedded-clause verb. Including the optional “that” eliminates this ambiguity. However, the ambiguity only arises for case-ambiguous pronouns (or noun phrases in general) like “you” or “it”; for a case-unambiguous pronouns like “she” or “he,” the sentence is never ambiguous (“The newspaper reported she...” can only be interpreted with “she” as the subject of an upcoming embedded verb). Elsness found no tendency to use “thats” more often in sentences with ambiguous pronouns, in which the “that” can disambiguate, than in sentences with unambiguous pronouns, in which the “that” need not disambiguate. This suggests that the optional “that” does not serve a disambiguating function. Another observation from a corpus analysis comes from Temperley (2003). He also examined the Brown corpus, and suggested that optional “thats” were used to avoid temporary ambiguities (which were likely garden paths as well) that arise with relative clause structures like “the lawyer companies like...” (because “lawyer companies” could be a compound noun). A number of laboratory studies have also investigated ambiguity avoidance. Ferreira and Dell (2000) explored a contrast similar to Elsness’s (ambiguous versus unambiguous pronouns), by having subjects say sentences from memory. Like Elsness, we found no tendency for speakers to use “thats” to avoid garden paths. Similarly, Arnold, Wasow, Asudeh, and Alrenga (2004) had subjects say sentences like “The judge sent the letter to the president to the members of the congressional subcommittee,” which include a temporary ambiguity (momentarily, “to the president” sounds like where the letter is sent -as if it attaches ‘high’ to the verb phrase -rather than what kind of letter was sent -attaching ‘low’ to the noun phrase). The ambiguity can be avoided by phrasing the sentence as, “The judge sent the members of the congressional subcommittee the letter to the president.” When speakers rephrased paraphrases of sentences like these with the instruction to create maximally understandable new sentences, no tendency for them to use the unambiguous rather than the ambiguous forms of these sentences was observed. On the other hand, Haywood, Pickering, and Branigan (2005) found that when speakers took turns instructing each other to carry out actions with utterances like “put the penguin (that’s) in the cup (that’s) on the plate,” they used “that’s” more in ambiguous circumstances (when either instruction was possible) than in unambiguous circumstances (when only one instruction was possible). (However, see Kraljic & Brennan, 2005 described below.) Ferreira, Slevc, and Rogers (2005) looked at ambiguity avoidance by investigating speakers’ object descriptions in a communication task. The objects were made ambiguous either by including in the same displays other different objects that happened to have the same name (e.g., a flyingmammal bat was to be described in displays also including a baseball bat), or by including objects of the same type that differed in some describable way (e.g., a flying-mammal bat accompanied by a larger flying-mammal bat). Results showed that this latter nonlinguistic form of ambiguity was easily avoided (the ambiguity is nonlinguistic because it arises from the conceptual similarity between objects of the same type, not directly because they happen to be described with the same words). However, the former linguistic ambiguity was avoided very weakly, and only under specific circumstances (the ambiguity is linguistic because the only reason for the ambiguity is that the otherwise distinct objects happen to have the same name). These results suggest that speakers strive to avoid ambiguity in general, as shown by their avoidance of nonlinguistic ambiguity, but they nonetheless have greater difficulty avoiding linguistic ambiguity, and they seem to do so only under specific circumstances. (Note that the ambiguities under scrutiny throughout this paper are linguistic in nature.) A number of reports have looked at whether speakers alter the prosody of sentences to avoid ambiguity, and most have found this not to be the case. Kraljic and Brennan (2005) had speakers describe displays to listeners, and found that they did not use different prosody for sentences like “put the dog in the basket on the star” when describing ambiguous displays (when either instruction was possible) versus unambiguous displays (when only one instruction was possible). (And, unlike Haywood et al. described above, they did not find that speakers used additional lexical material more often in ambiguous rather than unambiguous situations.) Schafer, Speer, Warren, and White (2000) also found that speakers failed to use disambiguating prosody specifically in CRL Technical Reports, Vol. 18 No. 2, December 2006 5 ambiguous situations in a scripted game-playing tasks. Allbritton, McKoon, and Ratcliff (1996) had subjects read aloud ambiguous sentences of a number of different types, and found that they used disambiguating prosody only when explicitly told about the possible ambiguity of the sentences. Similarly, Snedeker and Trueswell (2003) found that when speakers gave instructions to act on displays to one another (e.g., “tap the frog with the flower”), they used disambiguating prosody, but only if they were explicitly aware of the ambiguity. In short, whereas most evidence suggests that speakers do not systematically avoid ambiguity (Allbritton et al., 1996; Arnold et al., 2004; Elsness, 1984; Ferreira & Dell, 2000; Ferreira et al., 2005; Kraljic & Brennan, 2005; Schafer et al., 2000), some evidence suggests that speakers sometimes avoid ambiguity, at least under certain circumstances (Allbritton et al., 1996; Elsness, 1984; Ferreira et al., 2005; Haywood et al., 2005; Snedeker & Trueswell, 2003; Temperley, 2003). Specific reasons for this mixed picture are not yet known, but it is likely that the heterogeneity of the tasks, the ambiguities, and the assessed linguistic properties are relevant. The different investigations above explored different production modalities (edited written text vs. spoken utterances), different naturalistic conditions (production from memory vs. different kinds of monologue and dialogue tasks), different forms of ambiguity (garden paths, full ambiguities, prepositional-phrase attachment ambiguities, mainverb vs. relative clause ambiguities, subject vs. object ambiguities, noun-noun compound vs. noun-relative clause ambiguities, etc.), different ambiguityavoidance devices (mention of optional words, use of alternative word orders, prosody), and whether speakers are aware of ambiguities or not, just to name a few factors. Each of these differences might correspond to specific conditions under which speakers avoid ambiguity, or might introduce confounding factors that could lead to the appearance of avoidance or non-avoidance. The objective of the present experiments was to determine the limiting conditions on ambiguity avoidance. Specifically, if we simply ask subjects to avoid ambiguity (in lay terms, by asking them to write sentences that are “easier to understand”), what kinds of ambiguity do they avoid and what kinds do they not avoid? Thus, the question is not whether speakers naturally avoid ambiguity when speaking or writing, but rather, whether speakers can avoid ambiguity at all. To do this, subjects were shown sentences, instructed as to how they could change those sentences (which will be the ambiguity-avoidance device that will be assessed), and asked to make that change (or not) so as to make the resulting sentence “easier to understand.” The sentences were manipulated for whether they would be ambiguous even without the disambiguating device. If speakers avoid garden paths, they should use disambiguating devices more in sentences that need those disambiguating devices in ones that would otherwise be ambiguous -than they do in sentences that do not need those disambiguating devices. For example, take the sentence “The team defeated in the Super Bowl vowed revenge the next season.” As noted, after “defeated,” the sentence contains a garden path, because “defeated” is easily interpreted as a main verb when really it is a passive verb in a relative clause. One reason for the garden path is that the passive relative clause in the sentence (“defeated in the Super Bowl”) is reduced -the sentence does not include an optional relative pronoun (“who” or “that”) and auxiliary verb (“was,” “were,” etc.). If it did -“The team that was defeated in the Super Bowl vowed revenge the next season” -the garden path is eliminated. So, in the experiments, subjects were alerted to the option of using this optional material: Please read each of the following sentences. Each of them can contain two optional words, like that was, who were, that is, and so forth. In some of them, we have inserted those optional words into the sentence, but the sentence would be easier to understand if they weren’t there. For other sentences, we have left the optional words out, but the sentence would be easier to understand with those optional words back in. (Note that subjects were instructed that sentences would be easier to understand without the optional words so they wouldn’t adopt a strategy of always including them.) Then, subjects were instructed: Look at each sentence, and rewrite it with or without the that was (or that were, or whatever), depending on which would be easier to understand. If the sentence would be easier to understand with the that was, make sure that it’s there, but if it would be easier to understand without the that was, make sure that it’s not. CRL Technical Reports, Vol. 18 No. 2, December 2006 6 In short, these instructions practically beg subjects to avoid ambiguity by inserting (or omitting) specific optional material in written sentences. Below, two experiments are reported that were conducted with this “copy editor” task (between which a comprehension norming study is reported). In the first experiment, subjects were shown two kinds of sentences (half of each containing the optional material). In one, the form of the critical past-participle verb was ambiguous, such as the above “defeated.” In the other, the form of the critical past-participle verb was unambiguous. For example, take the (reduced) sentence, “The team beaten in the Super Bowl vowed revenge the next season.” The verb “beaten” has an unambiguous “-en” pastparticiple suffix rather than an ambiguous “-ed” pastparticiple suffix. Therefore, the sentence is unambiguous even when it is reduced: “the team beaten” cannot be taken such that “beaten” is a main verb (i.e., the team could not have beaten anything; they must have been beaten). (For evidence that readers find reduced relative-clause sentences with unambiguous past participles easier to understand, see Trueswell, Tanenhaus, & Kello, 1993.) If speakers are at least in principle sensitive to this ambiguity-relevant property (i.e., past-participle ambiguity), subjects in this copy-editor task should be more likely to write full relative clauses into sentences that have ambiguous past-participle verbs (“The team that was defeated...”) than into sentences that have unambiguous past-participle verbs (“The team that was beaten...”).