Manual wh-signs and English wh-mouthings differentiate BSL content and polar questions

4.1 BSL corpus data and study participants

Our analysis is based on approximately 5.5 h of dyadic pre-arranged free conversation between 40 deaf signers documented in the BSL Corpus (Schembri et al. 2013).^[2] Participants were balanced for gender (20 women, 20 men) and four regions in England (10 for each of the four cities Birmingham, Bristol, London, Manchester). Participants were on average 47.8 years old at the time of recording (SD = 17.8 years, 17–77 years). Most participants are right-hand-dominant signers (n = 37). Only 3 signers are left-hand-dominant. There are 24 participants (60 %) who come from hearing families, while 16 come from families with at least one deaf parent (40 %). Those from deaf families learned BSL from birth, while most of those from hearing families reported learning BSL before age seven. Two signers reported learning BSL between ages eight and twelve (LN25 and BL22). The majority of participants identified as White (n = 36, 90 %), with 3 participants identifying as Black and 1 participant identifying as South Asian. We also considered if participants had experience teaching BSL at the time of filming: 10 out of the 40 signers had experience teaching BSL (25 %), as opposed to 12 who had no experience teaching BSL (30 %). We have no information about whether any of the remaining 18 signers did or did not teach BSL (45 %). A total of 726 tokens of questions were annotated in this dataset. As we are using conversational data, tokens of questions are unequally distributed across participants, ranging from a minimum of 5 tokens per participant to a maximum of 44 tokens per participant (M = 18.2).

4.2 Representation in the BSL corpus and current study

Before proceeding further, we want to acknowledge some issues with the representation of deaf signers in the BSL Corpus as a whole, and the current study in particular. It is important to note that a corpus can rarely, if ever, claim to be entirely balanced and representative: it can only be described in terms of how it is balanced and representative (Gries 2009). The BSL Corpus was initially designed to document eight major regional urban varieties of BSL as used by deaf people who learned the language from their parents and/or peers as young children growing up in the UK, i.e., cohorts of deaf signers traditionally described as “native and near-native” signers (Schembri et al. 2013). Thus regional representation was limited in each country, and data from only four regions in England form the focus of this study. The 10 % figure mentioned above representing Black, South Asian, and other minoritised ethnic deaf people within this cohort was based on 2001 Census of England and Wales data used to plan participant demographics for the BSL Corpus in 2008 (Office for National Statistics 2006). The main focus was varieties of BSL used by deaf signers who experienced relatively uninterrupted signed language acquisition during early childhood, and who therefore represent language conventions transmitted across multiple generations of signing peoples (see Johnston 2006). In addition, the categories provided for participants to choose on the background questionnaire for gender were only male and female.

This obviously does not capture the full range of signing peoples who use BSL in the UK, including regions outside the eight urban cities studied, those who are non-binary, hearing people who acquired BSL from birth or in early childhood, deafblind signers who use tactile BSL, or deaf or hearing people who learned BSL during or after transition to adulthood (the latter sometimes referred to as “new signers”; De Meulder 2019). Minority ethnic representation among signers in the BSL Corpus is also lower than intended, mainly because deaf signers from Black or South Asian backgrounds could not be recruited in some locations. This at least partly reflects the fact that no deaf fieldworkers from minoritised ethnic backgrounds were employed during the data collection phase of the BSL Corpus project. It is clear that the 10 % figure itself was too low in 2008 and should have been closer to the 20 % identified in the 2011 Census (Oxford Policy Management 2015). Of course, it is also misguided to rely on percentage representation figures in the context of corpus data sampling. It is not possible, for example, that data from one deaf Black signer can be representative of all deaf people from Afro-Caribbean or African backgrounds who grew up using BSL in the UK.

4.3 Annotation of BSL corpus data

Data consist of one ELAN file for each signer in the dyadic conversation. The ELAN software enables precise time alignment of annotations with the corresponding video sources on multiple user-specifiable tiers (Crasborn and Sloetjes 2008). Approximately 500 manual signs in each file had previously been tokenised and assigned ID-glosses, along with free translations into English (Fenlon et al. 2015; see Johnston 2010 for more information about ID-glossing). The same files (n = 40) were enriched with annotations for the current study. In the first instance, we created clause and other relevant annotations for approximately 500 manual signs in each file. Each file therefore contains whatever number of clause annotations are necessary to parse the initial 500 tokens of manual signs.

The shortest clip is 03:03 min long; the longest clip is 20:31 min long. The mean duration across video clips is 08:09 min. These durations vary due to the conversational nature of the discourse. Signers were encouraged to discuss whatever topic they wanted to discuss, for however long they wanted, up to around 30 min. Signers covered a broad range of topics in their conversations, including their experiences growing up deaf, their relationships with parents and teachers, their working life, deaf clubs and other organisations, deaf sports, childhood and travel memories, and other events that were current during 2008–2010, when the BSL Corpus was documented. Examples include attending the 2005 Melbourne Deaflympics (a major event in modern deaf history), the social effects of the 2007–2008 global financial crisis (referred to as “credit crunch” in the UK), and the 2008 BBC drama House of Saddam (then available with closed captions on BBC iPlayer).

BSL questions were first identified and annotated following the fundamental unit of conversation structure: the adjacency pair. The first part of an adjacency pair is typically an initiated action, such as a question, and the response is the second part, a related action that in this case is an answer following the initial turn (Sacks et al. 1974). For example, an answer to a question could be displayed as a simple nod responding to a polar question. Adjacency pairs such as question-answer, greeting-greeting, or offer-acceptance have certain characteristics: they present two ordered turns (e.g., a question followed by an answer), displayed by different participants, and are adjacently placed (i.e., one after the other), belonging to a pair-type relationship (Schegloff 2007: 13). In this study, we also identified whether a question was directed to the other participant, or whether it featured in enactments where the same participant enacts a question and provides the answer as part of a narrative. This distinction helps to differentiate between two different types of conversational units (see Section 5.1 below for more on this subset of enacted questions).

Annotation of the BSL Corpus data analysed here was based on guidelines developed for the Auslan Corpus (Johnston 2010) and documented in the Auslan Corpus Annotation Guidelines (Johnston 2019). Questions were coded and analysed in terms of their semantic function: whether the question is a polar question or a content question. We then coded how each question was signalled: whether manual signs and/or non-manual forms were used, and if there was one form of signalling or multiple forms. We also coded aspects of syntactic structure, such as whether the polar or content question contained manual question signs, where these question signs were located within the question utterances, the scope of any non-manual forms used to signal the question, and some pragmatic functions. For example, we coded whether the question was initiated at the start of a turn or during the continuation of a turn, and whether the question occurred within enacted discourse (cf. reported speech, constructed action, role shift, personal transfer). We did this to check whether the fundamental pragmatics of turn-taking might have an effect on question signalling. Finally, we considered whether any socio-demographic factors influence individual signers’ expression of these question types in BSL, such as their age, gender, region, family background (e.g., parents deaf or hearing), reported age of BSL acquisition, and BSL teaching experience.

Each ELAN file was viewed multiple times by three annotators, including Hodge and Manrique, and the research assistant Maxwell Barber (a deaf lifelong signer of BSL). All instances of questions were identified and details about each token were annotated further on additional tiers. Table 1 describes the function of the ELAN tiers used to enrich the BSL question data with information relating to our research questions. The main unit of analysis is the Question tier, which was used to identify and annotate tokens of questions in the first instance, while additional information about each token was annotated on other tiers.

Tokens where the annotator was not confident about the identification of the question were tagged with a question mark. This annotation process ensured that all instances of questions evidenced within the conversations in the study dataset were identified and included, and none were cherry-picked or arbitrarily excluded, in keeping with the variationist principle of accountability (e.g., Tagliamonte 2006). It also means that in our statistical analysis, we could focus on those tokens that are confidently identified, all of which have been revised by more than one annotator, and some of which have also been checked by two other annotators (see Section 4.4 below). As all tokens of questions were produced spontaneously during these pre-arranged free conversations, the total number of tokens identified provides a fairly accurate indication of the proportion of questions within the conversations documented in the BSL Corpus.

4.4 Assessing stability of annotations in the BSL corpus

It is important to check if other annotators agree or disagree with the annotations created to investigate questions in BSL. This is done as an assessment of how annotations stabilise over time to represent a broad consensus, more so than a form of ‘validation’ (see Hodge and Crasborn 2022; Johnston 2019). A detailed coding schema to annotate and quantify the frequency of use of multiple manual and non-manual components of questions was developed as per Table 1. Questions were initially identified by Hodge, Barber, and Manrique. All question tokens were then reviewed, revised and coded by the three annotators. Any tokens where annotators disagreed or were uncertain about the analysis were annotated as uncertain with a question mark in the Question tier. The final dataset was then analysed by the third author, Winter, and Hodge using the statistical methods described in Section 4.5.

One outcome of the annotation and revision process was greater confidence in our overall analysis of polar and content questions, and fewer uncertain cases in the final results presented here. Henceforth, when counts of uncertain cases are very low for particular categories in the sections below, they are omitted from the data visualisations for ease of presentation. Another outcome was greater attention to other-initiated repair-type questions, some of which have not typically been discussed in the signed language literature and were overlooked during earlier revisions (see Manrique 2016, 2017]). This revision process highlights the benefits of uniting researchers with expertise in different yet complementary cognitive-functional frameworks within a single corpus study.

4.5 Statistical analysis

All analyses were conducted with R version 3.6.1 (R Core Team 2019). The ‘tidyverse’ package version 1.3.0 (Wickham et al. 2019) was used for data carpentry and plotting. The ‘ranger’ package version 0.12.1 (Wright and Ziegler 2017) was used for random forest analysis. The ‘tuneRanger’ package version 0.5 was used to tune random forest parameters (Probst et al. 2019). The ‘FactoMineR’ package version 2.4 was used to perform Multiple Correspondence Analysis (Lê et al. 2008). The ‘factoextra’ package version 1.0.7 (Kassambara and Mundt 2017) was used for interpreting and visualising the results of the Multiple Correspondence Analysis. Scripts, data, tables and figures are available in the following Open Science Framework repository: https://osf.io/mjwug/. The basic annotations for each file (ID-glosses and translations) are available to registered researchers via the CAVA repository for the BSL Corpus: https://bslcorpusproject.org/cava/. Other annotations can be made available on request.

4.6 Examples from the BSL corpus dataset

(2)

MC21 asks MC22 if she is taking a towel on their beach holiday (polar question with non-manual forms only, i.e., eyebrows raised and head nod movement).

(3)

LN06 checks with LN05 where he flew to on his holiday (content question with non-manual forms only, i.e., eyebrows down, pressed lips, and no wh-sign, showing an exception to the tendency for content questions to have wh-signs and English mouthing).

(4)

BM08 checks with BM07 who he was texting (content question with manual and non-manual forms, i.e., eyebrows down, head forward, manual wh-question form and English wh-mouthing, showing a typical content question).

(5)

LN24 checks that LN23 is not driving home from the party (polar question with non-manual forms only, i.e., eyebrows raised at the end, and note that headshake and brows down at the start is likely related to overall negation).

(6)

LN04 asks LN03 if she went to a deaf school (polar question with manual and non-manual forms, in this case the manual forms are double pointing).

5.1 Syntactic context of questions in this dataset

We first provide an overview of the syntactic context in which these questions occurred. Most questions were produced as single clauses (546 tokens, 75 %). The remaining question tokens occurred within complex clause utterances (82 tokens, 11 %), within utterances shared between two signers (38 tokens, 5 %), or as interactional fragments (31 tokens, 4 %). Some question tokens (29 tokens, 4 %) occurred within clause utterances that could not be confidently identified and analysed. In the following analyses, we exclude questions occurring within enacted utterances or constructed actions (n = 103; 14 % of total question tokens) because these questions do not request a response from the other participant. Instead, both question and answer are enacted by the same participant as part of a narrative, i.e., they involve the same signer re-enacting two or more interlocutors asking and answering questions within a period of enactment. The dataset analysed subsequently therefore comprises a total of 623 question tokens (86 % of the total tokens annotated).

5.2 Semantic type of questions

Figure 1 shows that most questions in the revised dataset of 623 question tokens were polar questions (416 tokens, 67 %), and fewer tokens were content questions (n = 140, 22 %). There were only 45 tokens that were of type ‘other’ (7 %), and 22 that were disjunctive ‘alternative’ questions (4 %), which are questions such as Are you coming today or tomorrow? that often present two options (see Bolinger 1978; Koshik 2005). The ‘other’ type is a catch-all category for anything that does not fit the types polar, content, or alternative, such as rhetorical questions or outlouds, where signers express thoughts such as Where did I put my key? that are not meant to be responded to by another person. As there were far fewer tokens of ‘other’ and ‘alternative’ question types observed in the corpus data, our analysis focuses on polar and content questions only (556 tokens, i.e., 89 % of the revised dataset of 623 question tokens). In other words, there are 556 tokens of polar and content questions in the final dataset analysed below. This is in line with our research questions, which are focused on the difference between polar and content questions.

Figure 1:

Proportion of questions according to semantic type (n = 623).

5.3 Manual and non-manual signalling of polar and content questions

Figure 2(a) shows that the overwhelming majority of both polar questions (n = 409, 98 %) and content questions (n = 135, 96 %) were produced with some form of non-manual question signalling coded in this study. This includes English mouthing (e.g., who, what, when), mouth gestures (e.g., downturned mouth), eyebrow movements and head movements. Only very few polar questions (n = 7, 2 %) or content questions (n = 5, 4 %) were produced without any of these non-manual signals. Figure 2(b) reveals that there were more differences between polar and content questions with respect to manual signalling. For polar questions, manual signalling was absent in the majority of cases (n = 357, 86 %). A few polar questions were signalled using one manual question sign (n = 44, 11 %); even fewer were signalled using multiple question signs (n = 15, 4 %; see Example [5]). Content questions differed from polar questions in having more tokens produced with a single manual question sign (n = 76, 54 %) or multiple signs (n = 30, 21 %). Only 34 content questions had no manual question sign (24 %). What Figure 2(b) does not show is that there were also differences in the types of question signs. While the manual question signs for content questions included signs such as who, what, where, when and how, this was not the case with polar questions, which included PALM-UP and some kinds of pointing.

Figure 2:

Semantic types broken up by (a) manual and (b) non-manual signalling of questions according to semantic type.

5.4 Use of English question mouthings and mouth gestures

Figure 3(a) shows that signers frequently incorporated English question mouthings like who, what, where and when into their content questions, yet it was rare for signers to use English question mouthings in polar questions. For polar questions, question mouthings were absent for the overwhelming majority of cases (n = 403, 97 %); a single question mouthing was present in only 13 cases (3 %),^[3] and there was not a single question that included multiple question mouthings. For content questions, question mouthings were absent for only 35 tokens (25 %); all other tokens either had single question mouthings (n = 87, 62 %), or multiple mouthings (n = 18, 13 %). Figure 3(b) shows that mouth gestures were rarely involved in signalling questions in this dataset: the majority of polar questions (n = 401, 96 %) and content questions (n = 130, 93 %) were characterised by an absence of mouth gestures relating to question signalling.

Figure 3:

Distribution of (a) English mouthings and (b) mouth gestures used to signal questions.

5.5 Use of eye, eyebrow and head movements

For eye and eyebrow movements, our annotation system allowed for multiple eye or eyebrow movements to be annotated within a given question token. For example, we annotated if a question was signalled with an upward eyebrow movement followed by a downward movement, and/or whether the signer was squinting. There were a total of 410 question tokens (74 %) that occurred with a single eye and eyebrow movement. A much smaller number (n = 49, 9 %) involved multiple eyebrow movements. Ninety question tokens (16 %) were not accompanied by any discernible eyebrow movements. Seven tokens (1 %) were excluded from further analysis because the eyes were not visible enough to determine if an eyebrow movement occurred. As can be seen in Figure 4a, whether there were single, multiple, or no eyebrow movements was largely the same for polar questions and content questions. The majority of polar questions (n = 303, 74 %) and content questions (n = 107, 78 %) involved single eyebrow movements; very few polar questions (n = 66; 16 %) and content questions (n = 24, 18 %) had no eyebrow movements; and there was a slightly higher proportion of multiple eyebrow movements for polar questions (n = 43, 10 %) than content questions (n = 6, 4 %).

Figure 4:

Distribution of (a) eyebrow movements and (b) type of eye and eyebrow movements for question tokens with only one eyebrow movement.

In considering what type of eyebrow movements were involved, we focus on the question tokens with only one eye or eyebrow movement, and only on the three most dominant types “up”, “down” and “squint” (see Figure 4b, n = 393). Content questions have proportionally more downward eyebrow movements (n = 61, 58 %) compared to polar questions (n = 75, 26 %). In contrast, polar questions showed proportionally more upward eyebrow movements (n = 166, 58 %) compared to content questions (n = 37, 35 %). There was overall a low number of eye squints, but these occurred proportionally slightly more often with polar questions (n = 46, 16 %) than content questions (n = 8, 8 %).

We used the same approach for considering head movements. Most question tokens involved single head movements, with similar proportions for polar questions (n = 241, 58 %) and content questions (n = 77, 55 %). There was a slightly higher proportion of polar questions with multiple head movements (n = 50, 12 %) compared to content questions (n = 4, 3 %). There were proportionally slightly more content questions (n = 58, 42 %) than polar questions (n = 125, 30 %) with no discernible head movement whatsoever.

In considering what type of head movements were involved, we focus on those questions with a single head movement. Since there were overall extremely low total counts of ‘up’ (n = 10) and ‘backward’ head movements (n = 8), we will not discuss these two types of movements with respect to their distribution in polar and content questions, and the totals in Figure 5b reflect this. Figure 5 shows the relative proportions for all other types of head movements in polar and content questions (n = 300). There were proportionally slightly more ‘sideward’ head movements for content questions (n = 13, 17 %) than polar questions (n = 21, 9 %). Polar questions in turn had proportionally many more nods (n = 122, 51 %) than content questions (n = 22, 29 %). The number of head shakes was similar for polar questions (n = 32, 13 %) and content questions (n = 11, 14 %). There were proportionally slightly fewer ‘forward’ head movements for polar questions (n = 36, 15 %) than content questions (n = 18, 23 %). Finally, the proportion of ‘down’ head movements was relatively similar for polar questions (n = 17, 7 %) and content questions (n = 8, 11 %). Taken together, this analysis of eyebrow movements and head movements suggests that polar and content questions do not differ that much from each other in terms of whether they occur with head or eyebrow movements. Rather, the type of movement differs much more strongly.

Figure 5:

Distribution of (a) number of head movements and (b) type of head movements for question tokens with only one head movement (‘up’ and ‘backward’ not shown due to few cases).

5.6 Use of multiple strategies for signalling polar and content questions

So far we have analysed various strategies for signalling polar and content questions in isolation. Now we consider signalling with multiple strategies in combination, such as whether an eyebrow movement was used together with a manual wh-sign to signal a content question. Figure 6 shows how many question tokens had zero, one, two or more distinct signalling strategies. The maximum that can be obtained was five strategies, i.e., a question token signalled using at least one manual question sign and any head movement, any eyebrow movement, any question-related English mouthing, and/or any question-related mouth gesture. As can be seen, most question tokens involved at least two strategies (404 tokens, 73 %). Comparably fewer question tokens relied on one encoding strategy (124 tokens, 22 %). Even fewer question tokens relied on none of the strategies quantified here (28 tokens, 5 %). These were tokens that may have been signalled using strategies that were not coded in this study, such as the use of sign repetitions to seek clarification; manual or non-manual sign holds at the end of a question; and freeze looks (see Floyd et al. 2016; Manrique and Enfield 2015).

Figure 6:

Number of strategies used to signal polar and content questions.

There was a clear difference between polar and content questions with respect to the number of strategies involved in question signalling. Nearly all content questions were signalled with two or more strategies (n = 130, 93 %), and only very few with just one strategy (n = 8, 6 %). Conversely, polar questions had a lower proportion of questions with multiple strategies (n = 274, 66 %), and correspondingly more questions signalled with just one strategy (n = 116, 28 %). The proportion of cases with zero strategies was low for both polar questions (n = 26, 6 %) and content questions (n = 2, 1 %). More detail about the combinations of strategies used to signal polar and content questions is described in Section 5.10.

5.7 Position of manual question signs in polar and content questions

Figure 7 shows the distribution of manual question signs in terms of their position within the polar or content question. It focusses only on those question tokens that include one or more manual question signs (n = 163, 106 content questions, 59 polar questions). For both types of questions, the most frequent distribution for the position of manual forms was the category ‘other’. This includes, for example, tokens of single-sign utterances, where the only sign is a manual question sign, as well as utterances with multiple question signs that are not a reprise, i.e., two unique question signs. These tend to be questions with pointing signs at the start and end of the question, and questions that include manual question signs and PALM-UP – in other words, questions with manual forms that do not fit any of the other categories discussed in the literature and also annotated for this study. Beyond the category ‘other’, the most common pattern for both content and polar questions is for a manual question sign to occur post-predicate (non-reprise and clause-final), then pre-predicate. Finally, the least common positions for manual signs are reprise (non-clause-final), reprise (clause-final) and post-predicate (non-reprise).

Figure 7:

Position of manual question signs within polar and content questions.

In order to determine if BSL Corpus data can shed light on the wh-movement debate described in Section 2, we looked at the 27 tokens of pre-predicate question signs in more detail. Most of these question signs were adjuncts (signs such as where, when, why and how) rather than arguments. We did find four tokens of question signs who or what in object position, such as Example (4).^[4] These four tokens do suggest that question signs in object-initial position are not ungrammatical as claimed by Neidle et al. (2000). However, none of these four tokens had a clear overtly expressed subject and thus none were in the order of OSV. It may be that BSL Corpus data does not directly address the issue of object-initial orders involving question signs because these orders are not observed in the BSL Corpus dataset described here.

It is more important to interpret results about the position of manual question signs in light of how signers also use English mouthing to signal content and polar questions. For example, consider a question that contains an utterance-initial mouthed English wh-question form and a post-predicate manual question sign, as in Example (7). This is different to a question that contains a sole post-predicate manual question sign without English mouthing, as in Example (8). The former contains more conventionalised encoding regardless of syntactic position of the manual sign. These kinds of possibilities are frequent in BSL conversations. However, they were not recognised in earlier investigations of question signalling in signed languages. As the BSL Corpus data shows, both manual signs and English mouthing are important for investigations of questions in BSL. Thus, studies on the position of manual question signs need to also concurrently consider data on the occurrence and positions of English mouthing forms that signal questions.

(7)

A BSL signer demonstrates a BSL content question with wh-mouthing.

(8)

A BSL signer demonstrates a BSL content question without wh-mouthing.

5.8 Pragmatic context of polar and content questions

Several pragmatic factors may influence how BSL signers signal questions, including what was signed before it was produced and whether the question was requested by another person. In terms of the non-manual strategies used to signal questions, following Deuchar (1984) and noted in Section 2, we wanted to check if it may just be necessary for the polar or content question to look different from whatever the signer was doing with their face and body before asking the question. In other words, whether visible perceptual contrasts are important to question signalling, in addition to or instead of specific manual and/or non-manual forms for signalling questions. We were also curious if there was any effect on whether the question requested a response during a conversation from the other signer, e.g., if questions that did not elicit a response have something noteworthy about them. Therefore we annotated the Question tier to investigate whether signers’ choices of manual and non-manual signalling may be influenced by what they were doing before and during their question utterance.

The Question tier included codes for whether a question was asked within a continuation of the signer’s discourse, i.e., the signer was already signing and then asked a question, or if it occurred as a self-initiated turn, i.e., the signer was not signing, but instead interrupted the other signer or otherwise initiated their own turn by asking a question. Figure 8a shows that the majority of polar and content questions were asked by the signer who was already holding the floor at the time of asking a question (311 tokens, 56 %). Slightly fewer questions were asked by a signer interrupting or otherwise initiating a turn (242 tokens, 44 %). A small proportion of questions were not certainly identified as either continuing or initiating, or as questions at all (3 tokens, 1 %). Content questions were proportionally slightly more likely to be initiating (n = 76, 54 %) than polar questions (n = 166, 40 %). Polar questions were slightly more likely to be continuing (n = 249, 60 %) than content questions (n = 62, 44 %).

Figure 8:

(a) Proportion of questions that were initiating or continuing, (b) proportion of questions that requested a response (yes/no).

We also coded whether questions function as adjacency pairs requesting a response from the other signer, in order to filter out those questions that did not demand a response from the recipient during a conversation. Figure 8b shows that most questions did request a response (488 tokens, 88 %), while a few questions did not request any response (68 tokens, 12 %). Most of the latter were rhetorical questions, outlouds or enacted questions that did not seek a response. The majority of polar questions (n = 370, 89 %) and content questions (n = 118, 84 %) requested a response.

5.9 Question signalling as a function of sociolinguistic factors

We also explored the extent to which strategies for signalling both polar and content questions differ as a function of five sociolinguistic variables: whether the signer grew up in a deaf or hearing family; whether the signer teaches BSL or not; and their age, gender, and region (Birmingham, Manchester, London, Bristol). We separated signers into a ‘younger’ and ‘older’ category based on a median split. For each of the sociolinguistic variables, we first established that each of the groups used a similar proportion of polar and content questions. This tells us that any differences in strategies between these groups may be due to different preferences, rather than resulting from using more or less content or polar questions within the corpus. The analyses of sociolinguistic factors reported in this section combine polar and content questions as otherwise numbers are too low when broken up by question type and the sociolinguistic factors. In other words, the following analysis looks at how question signalling in general differs as a function of sociolinguistic factors, regardless of the polar/content distinction.

We explored the data with Pearson standardised residuals to check if particular form features are over- or under-represented for any of the groups mentioned above, focusing on those results for which residuals are larger than |2| as a heuristic threshold to gain insight into this data. There were no such ‘significant’ over-or under-representations as a function of BSL teaching, deafness in the family, age, or gender. However, there was noteworthy regional variation, with signers from Birmingham and Manchester being more likely to not use manual question signs (83 % and 77 %) in comparison to Bristol (54 %) and London (62 %) signers. Conversely, Bristol and London signers were more likely to use either single or multiple manual question signs. Table 2 shows the total counts of manual question signalling, collapsed across the polar/content distinction, with standardised residuals in brackets.

Another aspect of question signalling that showed noteworthy regional variation relates to head movements. Table 3 shows the proportion of single, multiple and no head movements during question signalling (regardless of polar/content distinction), with standardised residuals in brackets highlighting patterns that deviate strongly from expectation (>|2|). A noteworthy pattern in this table is that Manchester signers showed a higher proportion of questions with no head movement (43 %) than Bristol (18 %) and London (27 %) signers. Proportionally speaking, Manchester signers performed fewer single head movements for questions (49 %) than signers from Birmingham (53 %), London (62 %) and Bristol (69 %).

Next, we zoom into what specific head movements the single head movements represent, looking at the same five most common categories – down, forward, sideward movements, or headshakes and nods – that we also analysed in Section 5.5. This revealed that Birmingham signers were somewhat more likely to perform sidewards head movements (standardised residual = +2.6), and correspondingly less likely to perform forward movements (–2.0). In terms of descriptive percentages, 20 % of single head movements for Birmingham signers were sidewards, as opposed to only 12 % for Bristol, 8 % for Manchester, and 6 % for London. Regarding forward head movements, Bristol (21 %), Manchester (21 %) and London (19 %) signers showed very similar proportions, compared to signers from Birmingham (10 %). Signers from Manchester showed fewer downwards movements (–2.1 standardised residual, 3 % of single head movements), compared to signers from Bristol (7 %), London (13 %) and Birmingham (7 %). Thus, there were regional differences both with respect to whether head movements were used (Manchester signers show fewer head movements), as well as with respect to which specific head movements featured in question signalling (Birmingham more sideward, fewer forward head movements; Manchester fewer downward movements).

In terms of whether questions occurred together with eye movements, signers from Manchester had the most questions without any eye movements (23 %), followed by signers from Bristol (17 %), London (16 %) and Birmingham (10 %) (see Table 4). In contrast, Birmingham signers showed the largest proportion of single eye movements (83 %), followed by signers from London (73 %), Bristol (72 %) and Manchester (70 %). There were no stark differences in the use of multiple eye movements, which showed similarly low proportions for all four groups.

Focusing on singular eye movements that were either “up”, “down” or “squint” (the three most common categories, see Section 5.5), Table 5 shows that Birmingham signers used squints proportionally more often (54 %) than up (21 %) or down (25 %) eye movements. Manchester showed a similar pattern, with a higher proportion of squints (56 %) and correspondingly fewer up (16 %) or down (28 %) movements. London signers also used squints frequently in question signalling (55 %), but they showed a much stronger preference for downwards eye movements (43 %) compared to the other regions, and correspondingly very few up movements (3 %). Bristol patterned similar to London, but with even more down movements (56 %) and even fewer squints (41 %), and similarly low numbers of up movements (4 %). Overall, these results suggest that Birmingham and Manchester signers pattern more closely together with respect to manual signs (Table 2), as well as with respect to eyebrow movements. This mirrors similar sociolinguistic patterns for clause negation in BSL, suggesting a possible difference between northern and southern regions in the UK (Hodge et al. under review).

5.10 Multiple correspondence analysis of all strategies for signalling polar and content questions

So far, the only analysis that has considered multiple strategies for signalling questions has lumped all encoding strategies together into a simple count variable (see Section 5.6). This revealed the dominant pattern is for signers to use multiple strategies with 73 % of cases, with the remaining questions relying on a single strategy for signalling a question, or a strategy that was not quantified in this study. In this section, we consider if and how specific strategies co-occur within polar and content questions.

We used Multiple Correspondence Analysis (MCA) to assess these co-occurrences. This method can be used to investigate the structure of associations between all the categorical variables together. MCA is a dimension reduction technique similar to Principal Component Analysis (PCA), but for categorical data. It allows us to reduce a large number of categorical variables to a smaller number of dimensions based on whether those categorical variables are associated with each other. In order to do this, all strategies relating to how questions were signalled (e.g., head movements, manual question signs) were split into a binary (yes/no) predictor variable. This meant that some of the more complex variables were broken up into their individual components. For example, the head movement variable was broken up into a variable ‘HeadForward’ coding for whether the head moved forward or not, another variable ‘HeadNod’ whether there was a nod or not, and so on. Consequently, the form features annotated in the BSL Corpus dataset were broken up into 16 different binary variables. All 16 variables were input into the Multiple Correspondence Analysis (see Figure 9).

Figure 9:

Multiple correspondence analysis of all categorical variables.

Figure 9 shows how particular question tokens are situated in the two-dimensional space spanned by the first two dimensions resulting from the Multiple Correspondence Analysis (dimension 1 captures 11 % of the variation in strategies; dimension 2 captures 10 %).^[5] Each dimension will be explained in turn. Looking at the contributions to the first dimension showed that the presence of a single mouthing and the presence of a single manual sign latched strongly onto this dimension; they contribute 20 % and 15 % of the definition of this variable, respectively. Moving from left to right along the x-axis of Figure 9, we can see that dimension 1 separates polar questions (fewer single mouthings and fewer single manual signs) from content questions (more single mouthings and more single manual signs), although there is also overlap. It should be emphasised that the MCA does not “know” about the polar/content type distinction, and only investigates which strategies are statistically associated with which other strategies. Dimension 1 shows that mouthings and manual signs strongly go together, and our visualisation with separate colours for polar and content questions shows that this happens to also strongly differentiate the two question types.

The second dimension is most strongly defined by the absence of eye or eyebrow movements (23 % of the definition of the dimension), as well as by the absence of head movements (14 %). This shows that questions that have no eye or eyebrow movements also tend to have no head movements (high values on dimension 2). However, when we colour each data point separately for question type, we can see that dimension 2 does not allow us to distinguish easily between polar and content questions. Altogether, this analysis suggests that of all the form features that distinguish polar and content questions, the presence or absence of mouthings and manual signs (dimension 1) is most important, and these two variables are associated with each other. Head and eye or eyebrow movements are also associated with each other, but in comparison less distinctive with respect to the polar/content question distinction.

5.11 Random forest analysis of all strategies for signalling polar and content questions

We then used all variables as predictors to assess whether specific strategies allow predicting whether a question is a polar question or content question. This analysis also allows us to do two things: first, we can assess how successful question type can be predicted from form features alone; second, we can assess which strategies are most predictive of question type. For this analysis we used random forests, a machine learning algorithm that performs well in situations when there are many potential predictors to consider and these are potentially collinear, such as when question tokens with manual signs are more likely to have English mouthings. The random forest is trained on predicting the polar vs. content question distinction from all 16 different binary form feature variables. Table 6 cross-tabulates whether the token question was a polar or content question (rows) against the predicted class (columns). Overall, accuracy was very high with 91.4 % accuracy in predicting the polar/content distinction from presence of strategies alone. The out-of-bag prediction error is 8.6 %, suggesting that this analysis would generalise well to unseen data.^[6]

Figure 10 shows the variable importances for the top ten most predictive strategies. The absolute values of this variable importance measure are not interpretable, i.e., this statistic only yields insights into the relative importance of different strategies with respect to each other. There was a clear pattern where the presence/absence of mouthing and the presence/absence of manual signalling were vastly more predictive of the polar/content question distinction than any other variables. This suggests that these two form features are most important for the difference between polar and content questions. In fact, if one uses mouthing as sole criterion for whether a question is polar or content, one would be correct 91.4 % of the time in classifying a question as one of these two types. Using manual question signs as a sole criterion would lead to a correct classification 83.3 % of the time for this dataset.

Figure 10:

Relative variable importances for the top ten most predictive predictors in the random forest analysis.

The random forest analysis corroborates the picture painted by the MCA, even though both approaches are conceptually very different. The MCA looks at correlations between strategies; the random forest analysis tries to predict question type from strategies. Together however, both analyses show that content questions are associated with English mouthings and manual signs, in contrast to polar questions. It also shows that in comparison to the overwhelming importance of manual signs and mouthings, non-manual eye, eyebrow and/or head movements are relatively much less important in distinguishing content and polar questions. This is not to say that there are not clear differences between content and polar questions for these other variables, as revealed by our descriptive analysis; however, when everything is taken together, highly conventionalised mouthings and manual signs dominate the polar/content question type distinction.

6.1 Overview of findings

In this study, we used descriptive and exploratory statistical methods to investigate how BSL signers signal polar and content questions in conversation data in the BSL Corpus. Based on a subset of 623 tokens of polar and content questions analysed here, we found that polar questions are more frequent in dyadic conversations than content questions, and that almost all questions involve non-manual signalling, including eye and eyebrow movements, head movements, and English wh-mouthings. A major finding is that multiple strategies are most often used to co-signal question types, which means that any theory or analysis where the differentiation of question types primarily relies on, or overly emphasises just one of these strategies – as is built into the manual versus non-manual marking paradigm – is missing a lot.

The main differences between polar and content questions relate to signers’ use of manual wh-question forms and mouthing of English wh-question forms. Content questions were more likely to involve single or multiple manual wh-forms. The most common position of these wh-signs within content utterances can be described as ‘other’, which ranges from questions consisting of one manual wh-form (i.e., a single-sign utterance), to questions with both manual and mouthed English wh-question forms, to questions with multiple manual wh-forms in more than one syntactic position in a multi-sign utterance, either with or without mouthed English wh-forms. Next most common was for manual wh-question forms to occur in post-predicate positions, then pre-predicate. Other positions (e.g., reprise) were also observed. If polar questions did include manual question forms, these were not wh-forms, but instead other manual strategies not analysed here (see Section 5.6). Content questions were also more likely to include single or multiple English wh-mouthings, while these were largely absent for polar questions.

Taken together, the data on manual question and English mouthing wh-forms present a more complex picture of question signalling than has been previously reported for BSL or other signed languages, especially for the signalling of content questions. These data indicate that theoretical investigations of wh-sign positions, such as the debate about wh-movement under generative analyses, may be oversimplified by focusing solely on manual question signs and non-manual brow, eye and head movements. Instead, this data from the BSL Corpus demonstrates how claims regarding the position of manual question wh-forms should also take into account the presence or absence of mouthed English wh-forms for signed languages with English as the majority ambient language. Likely signed languages with other ambient spoken languages should consider this too. Most importantly, analysis of corpus data illustrates just how problematic it can be to base theoretical analyses on a few judgements of isolated, contrived examples.

Following typological investigations of questions in signed languages, we also investigated signers’ use of non-manual signals such as eyebrow and head movements. We found it is most common for questions to be signalled using just one type of eyebrow and/or head movement, regardless of whether it is a polar or content question. With respect to the type of movements, content questions were more likely to have downward eyebrow movements, whereas polar questions more likely to have upward eyebrow movements. If there was no discernible head movement, these were most likely content questions. Nods were more likely to be used in polar questions, as were eye squints, which were infrequent overall. Questions with multiple eyebrow and/or head movements tended to be polar questions. Yet all types of eyebrow and head movements were identified in both polar and content questions, and none of these strategies differentiated between polar and content questions as strongly as manual and mouthed English wh-forms did.

We also considered how these different manual and non-manual strategies are combined within question tokens. Most question tokens involved at least two strategies; fewer relied on only one strategy. Nearly all content questions were signalled with two or more strategies, and only a very few with just one strategy. Conversely, polar questions had a lower proportion of questions with multiple strategies, and correspondingly more questions signalled with just one strategy. It is interesting to observe that the less frequent question type (content questions) is signalled with more strategies than the more frequent type (polar questions), as discussed in Section 5.6. This is in line with what has been observed for spoken languages, where frequent forms generally tend to be shorter, or less marked (Piantadosi et al. 2011; Zipf 1949). The proportion of cases with zero strategies (i.e., strategies observed in the data but not coded for this study) was low for both polar and content questions.

Overall, analysis of manual and non-manual question signalling in BSL suggests that this traditional binary distinction is too simplistic. The data presented here instead suggest that it is the presence or absence of conventionalised wh-forms that most strongly differentiates content from polar questions in BSL, and that these can be signed and/or mouthed. While mouthing played a large role in the BSL Corpus with especially content questions, mouth gestures did not. It is possible that the mouth gestures coded in the data have other more primary functions beyond question signalling, perhaps relating to stance, expressions of uncertainty, and/or disapproval.

6.2 Other factors influencing question signalling in BSL

The majority of the questions analysed here were produced from one of the parties during dyadic conversations and requested some type of response from the other signer. We checked if the interactional context of questions might have an effect on how questions were signalled, such as whether the question was asked by someone holding or interrupting a turn. Most questions were asked by the signer who was holding the floor, and fewer were asked by the signer interrupting or otherwise initiating a turn. Content questions were slightly more likely to be used to initiate a turn than polar questions, whereas polar questions were more likely to be used in continuing a turn. This indicates something of the conversational nature of the interaction and how different types of questions may be driven by the demands of solving problems of understanding within the interaction.

When one of the parties during an interaction does not understand or does not see what was signed before, they might initiate repair to fix the problem using a question-like form. These questions index a prior moment in a given conversation, and function as a way to confirm or check understanding (e.g., Do you mean X?, Tonight?). On the other hand, questions that initiate a new topic or ask for new information via standard content questions do not deal with problems of seeing or understanding when they are not related to what was signed before by the other signer. Note that wh-question signs are often used to initiate repairs during conversation, but in relation to what was said previously, such as asking What? Who? When? (see Manrique 2016, 2017] for more details).

Some of the manual question forms included in the current study, such as PALM-UP and some kinds of pointing signs, are different from the traditional wh-question forms typically considered in the signed language literature, i.e., manual signs meaning who, what, when, etc. These may have been identified as having a question function in the context of other signals that we did not code for in the current study such as manual holds or freeze looks (Floyd et al. 2016; Manrique and Enfield 2015). This would make sense given that PALM-UP and pointing signs are known to regulate interaction, in addition to holds and freeze looks (Cooperrider et al. 2018; Ferrara 2020). In future research, it would be useful to also code for holds and freeze looks to empirically determine their role in question signalling in BSL (see Manrique 2016, 2017] for LSA). During conversational interactions, polar questions and content questions are also often used to check, confirm and repair problems of understanding (Dingemanse et al. 2015). The previous literature on questions in signed languages has largely ignored repair-type questions, and addressee signing in general. The results of this study and other work on other-initiated repair perspectives in signed language interaction all highlight the importance of distinguishing the pragmatic functions or everyday use of questions in face-to-face conversation (Byun et al. 2020; Manrique 2017; Skedsmo 2021).

A major benefit of corpus-based investigations is the possibility to check the role of sociolinguistic factors in various identified patterns of use. The corpus data analysed here suggests that Birmingham and Manchester signers pattern more closely with each other than with London and Bristol signers, with respect to their use of manual wh-forms and eyebrow movements. This north/south England divide mirrors sociolinguistic patterns for how clause negation is signalled by signers from these regions (see Hodge et al. under review) and north/south England distinctions found in mouthing in BSL (Proctor and Cormier 2023). However, with the possible exception of region, there were no strong patterns that could be described by any of the sociolinguistic variables accounted for in this study. We therefore conclude that most of the patterns for signalling questions may instead be driven by interactional demands such as repair described above rather than any sociolinguistic factors analysed here. A clear example of this is the increased use of mouthing and other non-manual signs, such as head movements and more marked use of eyebrows, in repair practices (see Manrique 2016, 2017], for more details). This is attributed to the distinct interactional demands associated with repair. In conversation analysis, repair refers to the processes by which participants address problems in speaking, hearing (signing and seeing) and understanding that arise during interaction (Manrique 2017). Participants often need to clarify or emphasise the entire message or part of it to ensure mutual understanding.

Another benefit is the possibility to conduct exploratory analyses on all data points annotated. We used MCA to identify any strong associations between the different strategies used to signal polar and content questions. MCA suggests that single manual and single mouthed English wh-forms are strongly associated and together account for a significant amount of the difference between content and polar questions. It also suggests an association between the presence/absence of eyebrow and head movements, but these associations do not strongly differentiate between content and polar questions. It is enough to say that both are present (or absent) in both types of questions.

We also used random forest analyses to assess whether individual strategies may predict whether a question is a polar or content question. The presence/absence of manual wh-forms and mouthed English wh-forms was strongly predictive of content questions compared to any other variables annotated and analysed here. In other words, these two forms of signalling are most important for differentiating between polar and content questions in BSL. Content questions are strongly signalled using manual and mouthed wh-forms, in contrast to polar questions, which are not. The random forest analysis also predicts that eyebrow and head movements do not distinguish between polar and content questions. Instead, signers’ use of highly conventionalised manual and mouthed wh-question forms are what makes the difference.

Some of the findings from this study align with what has been claimed in earlier studies of polar and content questions in signed languages. For example, where there was a single eyebrow movement, we found a preference for content questions to use furrowed brows and for a preference for polar questions to use raised brows (Zeshan 2004, 2006]). Yet we also found brow and head movements to be plurifunctional across question types, similar to more recent studies such as Lajolinna (2021) and Puupponen (2019). More specifically for BSL, we find consistencies with Deuchar (1984) and Sutton-Spence and Woll (1999), who observed that while raised and furrowed eyebrows are typical of polar and content questions in BSL respectively, these can be overridden by other factors. Additionally Deuchar noted that a contrast with non-manual movements occurring just before the question may be more important than specific eyebrow positions. The BSL Corpus data appear to be consistent with these observations.

The main differences we find in the current study compared to earlier studies stem from using an interactional dataset, identifying not only grammatical functions, but also pragmatic functions that reflect language practices in conversation. Regarding the grammatical aspects of polar and content questions, we considered English mouthing in addition to manual question signs, in distinguishing content from polar questions. In relation to pragmatic functions, this study identifies multiple social actions such as repair practices, and whether a given question is answered or not, among other interactional aspects. This highlights the need to consider a wider range of factors in the study of questions in signed languages beyond those features already known to contribute to question signalling.

Overall, the patterns of question signalling we find in BSL mirror patterns identified for many spoken languages that use conventional wh-question forms (whether lexical items or mouthing) to differentiate content from polar questions. In terms of facial expressions and head movements, the fact that multiple eyebrow and head movements (rather than single, particular movements) distinguish polar from content questions suggests that these features are complex and plurifunctional. This may also relate to brow and head movement used by both signers and non-signers when asking questions (Campbell et al. 1999). This study has also shown that to have a better understanding of questions it is crucial to look at questions within an interactional context and not as isolated occurrences limiting the functions of question formats to confirming or asking for new information. As it has been shown in previous studies (e.g., Manrique 2016), problems of understanding are common in everyday interaction and prompt the use of question-like formats that play a different role in communication (e.g., Huh? What? Could you repeat that?).

While this study focused on the syntactic and semantic aspects of question signalling in BSL, we also considered the pragmatic context, specifically whether a given question initiated an adjacency pair by requesting a response from another interlocutor. This has been relevant for considering differences between ‘proper’ questions in conversation in spoken language literature, based on conversational structure units as question-answer sequences (Sacks et al. 1974). For example, whether questions request information or request confirmation from another person during interaction is important, compared to rhetorical questions or question-answer sequences that are re-enacted within a depiction of a previous interaction that do not seek a response from the current interlocutor.

Although such questions in the BSL Corpus dataset were not analysed here, we can confirm there were no syntactic or semantic differences to the findings outlined above, i.e., the same patterns described above for polar and content questions are also present in rhetorical and enacted questions. However, from a conversation analytic approach, these types of questions represent distinct practices that influence the unfolding of interaction and the roles of the participants (see Heritage 2010; Stivers and Enfield 2010). As Schegloff (2007) and Heritage (2012) have highlighted, the context and interactional purposes of questions significantly impact their role and function within conversation. Understanding the nuances and variations in question practices is crucial for a comprehensive analysis of conversational dynamics and the diverse functions that questions can serve (Clayman and Heritage 2002). This will form the basis for a later study building on the current dataset.

6.3 Limitations

This study has two main limitations. The first limitation relates to the type of data collected, which involved dyadic interactions filmed in a language lab environment. Future studies should continue contributing and building on our understanding of everyday interaction in signed languages by looking at both dyadic and multiparty interactions and across different social contexts, including those where participants engage in other social actions while interacting in more natural settings, such as eating together, working together, etc. (see e.g. Danet et al. 2023; Manrique 2016; Shaw 2019). The second limitation relates to the focus on quantitative methods of analysis. This leaves some open questions unanswered, which might benefit from complementary qualitative approaches such as Conversation Analysis and Interactional Linguistics. These approaches would be useful for clarifying further the quantitative findings described here, bringing new insights into the discussion and understanding of questions in signed languages more generally.

These two limitations are related. For example, this study found that polar and content questions in BSL may be signalled with all types of eye, eyebrow and head movements. However, we do not yet know if these different non-manual movements might be essential for different types of conversational functions. Examples might include standard questions when asking someone else for new or more information versus initiating repair on someone else’s prior turn due to problems of seeing or understanding (Manrique 2016, 2017]). Each function employs different conversational practices that might provide a better understanding of the use of non-manual movements in relation to the use of question-like formats, compared to syntactic or other factors that have been prioritised in the signed language literature so far (see Manrique 2016, 2017]).