The contribution of context and contour to perceived belief in polar questions

Highlights

• We measure perceived belief of ¡H*L% and L*HL% in Puerto Rican Spanish in various contexts.

• Mild positive bias contexts did not significantly affect perceived belief for ¡H*L% and L*HL%.

• Strong positive bias and mismatch bias significantly affected perceived belief for ¡H*L% and L*HL%.

• We argue that ¡H*L% in PRS conveys one layer of meaning (polar question marking).

• We argue that L*HL% in PRS conveys two layers of meaning (polar question marking + disbelief).

Abstract

In the past decades, we have advanced significantly in our knowledge of intonational meaning, but few studies have tested experimentally the way in which discourse contexts affect intonational meaning. In this work we were specifically interested in how listeners use both intonation and discourse context to infer information about speaker belief states. We examined the effect of five bias types on two intonation contours used for polar questions (PQs) in Puerto Rican Spanish (PRS). The bias types consisted of unbiased, mild positive bias, strong positive bias and mismatch bias contexts. The intonation contours had been previously claimed to differ in the belief state information they convey – ¡H*L% is known to mark utterances as PQs without encoding specific belief state on the part of the speaker, while L*HL% is known to convey a state of disbelief on the part of the speaker (Armstrong, 2010). We hypothesized that the lack of belief and the presence of disbelief for these contours, respectively, would be perceived by listeners when these contours were heard in an unbiased context. We also predicted that listeners would rely on contextual bias more for the ¡H*L% contour than the L*HL% contour, and that the disbelief meaning would persist regardless of discourse context. Perceived belief scores were analyzed, and results show that different bias types affected perceived belief scores in different ways. Mild positive bias did not seem to affect perceived belief for the two contours, while strong positive bias and mismatch bias did. Since L*HL% conventionally conveys disbelief, a reversal effect was shown when it was heard in the strong belief context. Participants’ comments indicate that in such cases, an ironic interpretation of the contour is available. These results, in addition to the comments provided by participants, show that perceived belief will depend both on the type of contextual bias, as well as the type of information conveyed intonationally. This work provides more evidence for the dynamic relationship between specific context types and intonation contours that differ in terms of the amount and type of meaning they convey.

Introduction

Over the last forty years, studies in intonational meaning have made clear that speaker attitude and belief states can be encoded intonationally. A classic claim comes from the seminal work of Pierrehumbert and Hirschberg (1990), who propose that intonation contours are chosen by speakers (1) to create a relationship between propositional content and previous/upcoming utterances and (2) to create a relationship between the propositional content of an utterance and the beliefs of the speaker and the hearer. This paper is concerned with the latter function of intonation. Following Pierrehumbert and Hirschberg's observations, we expect that from the hearer's perspective, intonation is used to glean information about speaker belief states. The role of intonation in encoding belief is especially relevant when considering polar questions (PQs), an area of the grammar where we can find varying levels of epistemic asymmetry in terms of speaker commitment to propositional content, what the speaker knows, what the hearer knows, the attitudes the speaker may have about propositional content, and so forth (Enfield et al., 2013). For instance, when the speaker produces a true out-of-the-blue PQ, s/he may have no assumptions about whether the answer is yes or no. Enfield et al. claim that for PQs, speakers generally have a low commitment to the truth of a proposition, while the hearer is assumed to have a higher commitment (since the speaker directs his/her question to the hearer). However, we know that often times, speakers have evidence that leads them to draw conclusions about the truth value of propositions, and therefore PQs are frequently not so out-of-the-blue. Speakers are able to encode their own beliefs and attitudes about propositions, and can use the grammar to make more fine-grained distinctions within a ‘gradient possibility space’ (Enfield et al., 2013:194). Enfield et al. discuss the semantics of sentence-final particles and their use for ‘tilting the epistemic gradient’. For instance in Lao, a tonal language spoken in Thailand, the sentence-final particles bòò3, vaa3, tii4 and nòq1 are used for PQs. The first means ‘I want to know if p is the case’, the second means ‘I want to know if p is the case; I’d say it is, based on current evidence’, the third means ‘Maybe ∼p is the case, I don’t know; I’d say p is the case, based on independent evidence’, and the last means ‘I’m saying p is the case; I’d think you say this as well’ (p. 204). The semantics of these particles are then modulated by pragmatic context. The underpinnings of sentence-final particles mirror quite closely the function of intonation in PQs. For instance, Vanrell et al. (2013) show that through intonation, speakers can encode their greater degree of commitment to propositional content in PQs. Vanrell et al. (2014) showed that when Mallorcan Catalan speakers have physically perceived evidence for a proposition, they are likely to mark PQs with the particle que and the L+H* L% nuclear configuration. We thus expect that the meaning of intonation contours will be modulated by pragmatic context much like sentence-final particles are. One example of the pragmatics of PQ intonation can be found considering the high-rise pitch contour (H* H−H%) in American English, which may be used for PQs, but has other, more context-specific meanings. Pierrehumbert and Hirschberg (1990) give the example of the utterance My name is Mark Liberman produced with the H* H−H% contour. As mentioned, this contour is used for PQs in American English. However, in the context given by Pierrehumbert and Hirschberg, the speaker is not questioning whether his name is Mark Liberman, which would be the case if H* H−H% were being used as a PQ. Rather, in a given context, where the speaker has approached the receptionist window at a doctor's office, the tune can be used to convey something like ‘My name is Mark Liberman and are you expecting me, or am I in the right place?’. The speaker is able to convey information (he asserts his identity), but the contour allows him convey additional information, giving the utterance another dimension of meaning in context. This same contour H* H−H% is analyzed by Hirschberg and Ward (1995) as being used by the speaker to convey that propositional content should be added to the mutual belief space but at the same time allows the speaker to question whether the hearer can relate this propositional content to their own private beliefs. Based on this analysis of H* H−H%, they propose that “the tonal features associated with phrases should better be seen as relating the interpretation of phrases in some way to speaker or hearer's private beliefs” (411). Later, Escandell-Vidal's (1998) analysis of interrogatives in Peninsular Spanish assumes that speakers make intonational choices based on their assessment of their beliefs, as well as the beliefs of others. Her proposal for “attributed interrogatives” is based on the speaker's attribution of propositional content as the belief of a speaker vs. the belief of someone else (often times the hearer). She finds that in Peninsular Spanish, the low-rise conveys only one level of meaning (PQ-marking), while the high rise and the rise-fall, in addition to marking a sentence as a PQ, mark the proposition as the belief of some individual. These “attributed interrogatives”, according to Escandell-Vidal's account, convey two levels of meaning.

If speakers convey information about speaker and hearer beliefs intonationally, then we also expect that this information is accessible to the hearer. After all, why would the speaker go through the trouble of marking such information intonationally? To this end, Gravano and his colleagues (2008) used experimental methods to test the contribution of intonation to listeners’ assessment of speaker certainty. In this study, they showed that just as the epistemic modal would conveyed more certainty than the verb is, the downstepped contour in American English conveyed more certainty than the declarative or yes-no question contours. Their findings confirm that like modal verbs, listeners depend on intonation in assessing the degree of speaker certainty. Within the domain of PQs, few studies have explored the effect of context and intonation contour on perceived belief experimentally, especially within the realm of PQs. Gunlogson's (2003) pioneering work on PQ intonation in American English made clear the crucial role of discourse context for the interpretation of PQ intonation. However, experimental research examining how intonation and context conspire to lead a listener to a given meaning is rather limited, and we find even less work relating these factors to the perception of speaker belief states. Nevertheless, we review some of the relevant literature here. The L*+H L−H% contour was investigated experimentally by Ward and Hirschberg, 1988, Ward and Hirschberg, 1992 to understand which phonetic details were necessary to distinguish between uncertainty vs. incredulity meanings, both of which are available for L*+H L−H% in American English. These authors found that while listeners were influenced by other spectral characteristics, they were most influenced by pitch range. However, the authors examined ambiguous contexts in which both the uncertainty and incredulity interpretations were available, and not how context might influence the listeners’ choice of uncertainty vs. incredulity, but rather a range of phonetic details.¹ Later Nilsenová (2006) investigated perceived bias based on intonation contour in American English questions. Listeners were presented with positive PQs produced with any of eight intonation contours (H* H−H%, L* H−H%, L* L−H%, L* H−L%, L* L−L%, H* L−H%, H* L−L%, H* H−L%). Using a 5-pt Likert scale, listeners had to make a decision about the answer the speaker expected: the speaker definitely expects ‘no’; probably expects ‘no; [has] no expectations; probably expects ‘yes’; definitely expects ‘yes’. Expectations of a negative answer were found to be related to L* L−H%, while expectations of a positive answer tended to be associated with high phrase and high boundary tones. Listeners tended to perceive the speaker as expecting a positive answer with low-rises and high-rises. One limitation in this design that the author points out is that these questions were presented to listeners decontextualized, and thus it is not clear how the tunes would have been perceived in context. Speech corpora have also been used in experimental work on perceived certainty. Lai (2010) investigated the effect of context type, intonation and cue words on perceived certainty. Using the Switchboard I Release 2 corpus of telephone conversational speech, Lai examined the effect of different types of contexts (factual, evaluative, attributed, inferred), prosodic contours (rises vs. falls) and cue words (e.g. really, well, okay, sure, yeah). Participants were presented with written conversations and audio response (from the corpus, but later synthesized for the experiment) which they then rated using a 7-pt. Likert scale. Participants gave ratings based on expectedness, credibility and evidence. Lai found that rising intonation did not affect perceived credibility of the speaker, but did lead participants to rate speakers as sounding more uncertain. She also found that ratings varied depending on the specific cue word used. For instance, the cue word yeah was found to be highly susceptible to sounding uncertain with rising intonation, but not the cue word well. Hara et al. (2013) used naturalness ratings to explore the role of prosody in negative polar questions in Mandarin Chinese as well as Japanese. They found that in Mandarin, sentence-final stress in PQs conveys an epistemic conflict between speaker and addressee. In Japanese, on the other hand, deaccenting in negative polar questions can be used when the speaker seeks agreement. They characterize both non-canonical uses as instances of prosody being used to convey a meta-question about the discourse. Interestingly, both of the prosodic strategies investigated by the authors are more prototypically known in the literature as prosodic correlates of focus or information structure, but the authors show that listeners associate them with belief and discourse evidence. Most recently, Arvaniti et al. (2014) explored the pragmatics of two types of wh-question intonation in Greek: LH* L−L% and L*H L−¡H*% tunes (labels are from the GRToBI system). With the goal of showing evidence for a phonological difference between the two contours, they used two experimental designs: one where participants listened to one contour or the other with different contexts (default, information-seeking wh-contexts vs. rhetorical wh-contexts), and another in which participants listened to a question produced with either contour and picked an appropriate response. Response patterns for both experiments clearly differed depending on the intonation pattern, allowing the authors to argue for a phonological distinction between the two tunes.

Puerto Rican Spanish (henceforth PRS) is a Caribbean variety of Spanish known to exhibit variation in the contours possible to convey polar questions. In the Spanish Tones and Breaks Indices System (Sp_ToBI) (Armstrong, 2010, Hualde and Prieto, 2015) this variety of Spanish has an intonational inventory comprised of (L*, H*, ¡H*,² L+H*, L+<H*, L*+H) and five boundary tones (L%, !H%, H%, LH% and HL%). The ¡ and ! diacritics indicate phonological upstep and downstep, respectively. Like other languages (D’Imperio, 2002 for Neopolitan Italian; Grabe et al., 2004 for varieties of British English; Lucente, 2012 for Brazilian Portuguese; Vanrell et al., 2013 for Majorcan Catalan; Henriksen et al., 2015 for Manchego Spanish), an utterance-final fall is used to mark polar questions in PRS. Speakers produce an extra-high tone in the nuclear stressed syllable, followed by a fall to a low boundary. The schematic in Fig. 1 illustrates the ¡H*L% nuclear configuration³ used for polar questions in PRS.

This differs from the type of fall found for declaratives, for example, which are also characterized by final falls. However, declarative falls are characterized by a fall through the nuclear stressed syllable, labelled H+L* L% in the Sp_ToBI system. As can be seen in Fig. 2, the pretonic syllable is high, followed by an F0 fall through the nuclear accented syllable to a low boundary tone.

Armstrong (in press) has also described a contour used for PQs that express ‘disbelief’ in PRS. This contour is described as a low, flat tone in the nuclear tonic syllable (L*), followed by a bitonal boundary tone HL% in the post-tonic syllables. Fig. 3 illustrates this contour.

In production work using the Discourse Completion Task (Prieto, 2001, Nurani, 2009, Prieto and Roseano, 2010) paradigm, it has been shown that in PRS, the ¡H*L% contour in the Sp_ToBI system (Armstrong, 2010, Hualde and Prieto, 2015) is used as a general PQ marker, (sentence-level marking, but no belief marking) while the L*HL% contour is used to grammatically encode the speaker's state of incredulity or disbelief with respect to p(roposition) (Armstrong, in press). While many varieties of Spanish convey incredulity using the pitch shape of a general PQ-marking intonation contour in addition to expanded pitch range or non-modal voice quality, PRS is unique in that a specific intonation contour, L*HL%, has been found to be used rather reliably to express incredulity or disbelief. As far as we know, the relationship between the L*HL% tune and incredulity/disbelief meaning has not been shown for other Spanish dialects (though it is true that not all Spanish dialects have intonational descriptions). Intonational encoding of speaker disbelief, or incredulity, has been documented for other Romance languages such as Valencian and Central Catalan (Crespo-Sendra, 2011, Crespo-Sendra et al., 2013), Buenos Aires Spanish (Ar Lee et al., 2010), Brazilian Portuguese (Truckenbrodt et al., 2009, Frota et al., 2015), Friulian (Roseano et al., 2015), and French (Michelas et al., 2013). The PQs in Fig. 1, Fig. 3 would differ then, in the amount and type of information they convey, as shown in Table 1a, Table 1b.

We can refer to L*HL% questions as multidimensional, i.e. conveying two “layers” of meaning in that they seem to convey both information about sentence type (i.e. PQ-marking) and belief state (i.e. conveying the speaker's state of disbelief). For the case of ¡H*L%, it is possible that the speaker's state of “unknowing” would be inferred, by nature of the utterance being marked as a question.

We do not restrict the idea of layered intonational meaning to PQs, since we assume that such strategies would also available for other sentence types. If it is the case that the default PQ-marking contour ¡H*L% in PRS does not encode any specific belief information/speaker bias apart from the sense of “wondering” or “unknowing” associated with questions in general, then it follows that this contour would be felicitous in unbiased contexts, as in (1):

Here José has never met Felicia and therefore has no way of knowing whether or not she has children. The default contour, but not the disbelief contour, would be felicitous here. A context where the disbelief contour might arise is shown in (2):

In (2) Chris reacts to the presupposition in Rebecca's prior utterance, that she has a child. Since Chris has known Rebecca for a decade and she has never mentioned having children, the idea that Rebecca has children does not match with Chris’ expectations about the world, resulting in his state of incredulity. The PRS L*HL% intonation would allow a speaker in this context to both question the propositional content and express this state of disbelief at the same time. Thus, the contour carries two layers of meaning: sentence-level (PQ marking, or perhaps more minimally, response-eliciting) and belief state (disbelief marking). There are no restrictions in terms of how the information is activated for L*HL% questions – p may be activated linguistically (as in (2)) or extra-linguistically (as in (3)):

Since a goat is not a typical gift for city-dwellers nor is it a pet typically kept by New York City apartment-dwellers, Jessica conveys her state of disbelief that her husband got her a goat. However, the proposition is activated extra-linguistically – it is something Jessica has perceived through sight.

Given the ways in which the PRS PQ contours have been claimed to be pragmatically restricted (Armstrong, in press), we would expect that when listeners hear PQs produced with ¡H*L% they should not perceive any specific belief on the part of the speaker (other than a state of unknowing), but should perceive disbelief when they hear PQ questions produced with L*HL%. But to what extent does discourse context play a role? The meanings proposed above were based on production experiments and in context, and therefore the first goal of this work was to confirm whether the belief states that listeners perceived in a perception experiment could confirm earlier claims about the meaning of the two contours described above. Below we outline a rating experiment designed to examine the meanings of ¡H*L% and L*HL%, in addition to how pragmatic context might modulate their meanings. We reasoned that if the default contour ¡H*L% does not convey belief state information, then participants should perceive neutral belief states when the contour is heard in an unbiased context. Disbelief, on the other hand, should be perceived even in an unbiased context for utterances produced with L*HL%, if disbelief meaning is encoded through the contour. We also hypothesized that perceived belief states would be more affected by context when the default contour is heard than when the disbelief contour is heard, since there would be no encoded belief state information competing with contextual bias, i.e. since no belief is encoded through the default contour, it would be easier to base perceived belief purely on context. However, since disbelief intonation (L*HL%) in PRS encodes two levels of meaning, we expect context to affect its meaning differently from the default contour. In order to understand how context might affect belief states that are encoded intonationally, we used an experimental methodology that used contexts differing in degree of bias. We qualified the degree of bias based on the type of evidence for a given proposition available to the speaker in the discourse context. The contexts we chose were by no means exhaustive. We pitted one unbiased context against four different biased contexts. Our predictions and hypotheses are tested in the experiment outlined in section 2.