1. Introduction

1.2. Present study

In the present study, we are interested to find out if coarticulation of handshape occurs in conversational signing in Sign Language of the Netherlands, similar to what has been shown for ASL. Such coarticulation effects can inform our phonological modeling of handshape. More specifically, we examined whether effects of surrounding handshape are present for the degree of flexion as well as the degree of abduction of the thumb in target signs that have similar handshapes with all fingers selected (see Figure 1), none of which have the features ‘aperture’ or ‘spreading’ (typically referred to as the B handshape, because of the form of this letter in the Dutch and American hand alphabets). The term ‘spreading’ is used in a dual sense in the sign language literature: When it comes to finger configuration, the term refers to the abduction of the thumb, index, ring, and pinky finger at the metacarpophalangeal joints, away from the middle finger. In addition, ‘spreading’ is used in the more common temporal sense of feature spreading over neighboring segments. In this paper, we will only use spreading to refer to the finger configuration. We examined anticipation effects and perseveration effects of the position of the thumb. In addition, we investigated whether spreading of the fingers predicts the position of the thumb at the target sign and to what extent coarticulation of the position of the thumb between surrounding signs and target sign is mediated by the degree of spreading of all fingers.

Figure 1

Three variants of a flat handshape with all fingers extended.

The position of the thumb is thought to play no phonological role in the B handshape in NGT (Crasborn, 2001; van der Kooij, 2002), even though in early descriptions of the language a distinction was made between flat handshapes with the thumb extended (B0, see Figure 1b) and the thumb unflexed, parallel to the other fingers (B1, see Figure 1c; KOMVA, 1986). The thumb is one of the fingers that can realize the selected finger specification ‘all’ in these flat handshapes (Crasborn & van der Kooij, 2003). Its most clear articulation would be the B1 handshape, where the thumb is maximally parallel to the other fingers, maximizing the size of the plane formed by the palm and fingers. There are three phonological features that affect thumb position in other handshapes. First of all, the B handshape contrasts phonologically with the 5 handshape, which likewise has all fingers selected, but in addition has the feature ‘spread fingers.’ As this feature pertains to the relation between the fingers, the thumb and index finger spread (i.e., thumb extended to some degree) just as the other fingers spread (index, ring, and pinky finger). Thus, the thumb is extended in so-called 5 handshapes, where all fingers are extended and away from the midline of the hand (roughly the middle finger). This handshape contrasts with the number ‘4’ in NGT and in many other signed languages. In this latter sign, the four fingers are extended, iconically representing four items, and the thumb is out of the way. The optimal phonetic contrast between the four selected fingers and the thumb would be if the thumb were opposed and touching the palm.

In addition to the feature ‘spread fingers,’ a second relevant phonological property of handshapes that affects the thumb is ‘aperture.’ Aperture specifies the (open or closed) relationship of the thumb to the fingers when the thumb is opposed to the palm. This occurs when the thumb and fingers move (opening and closing with respect to each other), but the closed aperture relation also occurs without movement.

Finally, in contrast to the selection of all fingers in the B handshape, the thumb can be the only selected finger in a handshape, the other fingers being closed to a fist while the thumb is fully extended (the A handshape) or it can be one of the selected fingers in a handshape, with other fingers also being fully extended, such as in the NGT fingerspelling handshapes Y (pinky and thumb), W (middle finger, index finger, and thumb), and L handshapes (index finger and thumb).

In the three cases above, the thumb is a selected finger, whether alone or with other fingers, and as such it can be further specified with a configuration feature. Unselected fingers in the phonological model of van der Kooij (2002) and Crasborn (2001), on which we build here (as opposed to that of Brentari, 1998), cannot have a configuration feature. In the B handshape that is the focus of the present study, the thumb is not selected, and thus cannot carry any further configuration features (such as ‘curved’). This restriction applies to the four fingers just as much as to the thumb.

Variation in the phonetic position of the thumb has been studied before by Battison et al. (1975), looking at ASL. They tried to create a set of rules predicting the position of the thumb in various handshapes, depending on the phonological context within the sign (such as the location or the movement). They elicited judgments from 39 ASL signers on the acceptability (yes or no) of thumb extension in 11 signs with either an ASL ‘G’ handshape (extended index finger) or ‘H’ handshape (extended index and middle finger, non-spread). Thumb extension was found to occur more for H than for G handshapes, more when the location is on the face, and less for signs that have forearm rotation as a lexical movement. A later study on ASL (Kettrick, 1983) argued that it is most likely that the position of the thumb in handshapes in which it is not selected is dependent on the neighboring signs. It is this hypothesis that we seek to test for NGT.

3. Results

We first consider the two sub-questions that are needed in order to answer the main question. The first sub-question is: Which positions of the thumb can be identified for the target sign and for the surrounding signs and what is the distribution pattern of the various positions? The distribution of frequencies of thumb positions across all target signs and their surrounding signs is given in Figure 3; quantification of the distribution of thumb positions across the 14 specific target signs in citation form is given in Table 2.

Figure 3

Frequency of occurrence of various thumb positions at the averaged preceding signs, target signs, and following signs.

Both the figure and the table indicate that there is variation in thumb positions, both in the target signs and in the preceding and subsequent signs. Table 2 further shows that some target signs (e.g., CHILD) are more variable in terms of thumb position than others (e.g., YOUNG), and that not all possible thumb positions were observed in each single target sign. We return to this point in the discussion.

The second sub-question is: How often is the B handshape of the target sign articulated with some or a large degree of spreading of the fingers? This is reported in Table 3, which provides an overview of the various degrees of spreading at each of the 14 target signs.

As can be seen in this table, some signs show a highly homogeneous set of articulations with respect to spreading, where none (e.g., for HUMAN-BEING) or only a few (e.g., for RECENTLY, TO-CALL, and OH-I-SEE) of the sign tokens showed spreading. Other signs showed a more heterogeneous set of articulations. In the discussion, we will return to this issue, when we attempt to explain these variable findings. Now we move on to the main research question, i.e., does the position of the thumb (in particular the degree of flexion and the degree of abduction of the thumb) at a preceding and/or a subsequent sign predict the position of the thumb in a target sign with a B handshape? To answer this question, we performed two mixed-effects regression analyses—one with respect to ‘Flexion’ and one with respect to ‘Abduction.’ A major advantage of mixed-effects analysis is that it can take into account the role of coincidental variation in the data caused by specific participants and/or specific individual signs (e.g., Baayen et al., 2008). By including these ‘random’ variables in the statistical model, and distinguishing them from the central variables of interest (i.e., the fixed variables), mixed-effects modeling makes it possible to test the effects of fixed variables, while taking into account the fact the effect of individual participants and items on the observed variance. Another advantage of mixed-effects modeling is that it is less dependent on well-balanced, factorial designs than ANOVAs; mixed-effects models are well capable of analyzing relatively unbalanced data, such as data from naturalistic corpora (Baayen et al., 2008).

The analyses were performed using the lme4-package (Bates & Maechler, 2010) in R 2.11.1 (R Development Core Team, 2010). The dependent variable in the first analysis was the degree of flexion in the target sign. As can be seen in Table 2, this variable had values ranging from –2 to 2 and was therefore treated as a continuous variable, on which we performed a linear mixed-effects analysis. The independent variables (i.e., predictors) in this analysis were (1) the degree of flexion in the sign that preceded the target sign, (2) the degree of flexion in the sign that followed the target sign, and (3) the degree of spreading of the other fingers in the target sign. The dependent variable in the second analysis was the degree of abduction in the target sign. Like flexion, this variable could in theory range from –2 to 2 (see Table 1), but the vast majority of signs either had the value of 0 or 1 (see Table 2: Only 7 out of 728 observations had an abduction value that was not 0 or 1, which is less than 1% of the data). For this reason, we treated the variable as a dichotomous variable in which only those instances with values 0 or 1 were included. We performed a logistic mixed-effects regression analysis on this data. The independent variables (i.e., predictors) in this second analysis were (1) the degree of abduction in the sign that preceded the target sign, (2) the degree of abduction in the sign that followed the target sign, and (3) the degree of spreading of the other fingers in the target sign.

We always began our analyses with a full model, including two-way and three-way interactions between the independent variables. We then tested whether the fit of the full model compared to simpler versions of the model in which non-significant interactions between predictors were not included, by using a likelihood ratio test that examines whether the log-likelihood of one model versus the other differs significantly from zero (cf., Baayen et al., 2008). If the fit of a simpler version of the model was not significantly different from the fit of the more complicated model, then the simpler model was considered a more optimal reflection of the data. With respect to the random effects, we always included by-item and by-participant random intercepts; for the inclusion of random slopes we used the same procedure as with the fixed effects, namely by starting with the maximal random slope model (following Barr et al., 2013), and then testing whether the fit of the model would be statistically different with simpler versions of the random slope model.¹ By testing the role of random effects in this way, it is possible to gain insight into the role of by-item and by-participant variation with respect our variables of interest.

The analyses that we report below provide the results of the optimal model in terms of fixed and random effects. The mixed-effects models are summarized in tables that report the influence of each predictor variable by giving its parameter estimate (B), the standard error of the parameter estimate (SE B), its t-value or z-value (t-value in the linear regression on flexion; z-value in the logistic regression on abduction), and its p-value.

3.1. Results: Flexion

A summary of the mixed-effects analysis on flexion that best fits the data is given in Table 4. The analysis yielded significant main effects for flexion in the preceding sign and flexion in the following sign, as well as a significant interaction between flexion in the preceding sign and flexion in the following sign. The main effects indicate that a higher degree of flexion in the preceding and/or following sign leads to a higher degree of flexion in the target sign (see Figure 4). The interaction effect is illustrated in Figure 5, and shows that the effect of flexion in the following sign is stronger with lower degrees of flexion in the preceding sign (and vice versa: The effect of flexion in the preceding sign on flexion in the target sign is stronger with lower degrees of flexion in the following sign). The analysis also yielded a significant main effect of spreading. This effect indicates that the degree of flexion in the target sign is higher with higher degrees of spreading of the other fingers in the target sign (see Figure 6). There were no significant interaction effects of spreading with the other predictors included in the model.

Table 4

Summary of the optimal linear mixed-effects regression analysis for variables predicting the degree of flexion in the target sign.

Predictor	Estimate	SE	z-value	p-value
(Intercept)	1.194	0.103	11.487	< .001
Flexion preceding sign	0.198	0.037	5.321	< .001
Flexion following sign	0.121	0.022	5.416	< .001
Spreading other fingers in target sign	0.171	0.049	3.458	< .001
Flexion preceding sign × Flexion following sign	–0.044	0.016	–2.606	.009

• Note. N = 728 (61 signers, 14 signs), Log-likelihood of model: –819.5. Standard deviations of random effect terms were: 0.261 for by-participants random intercepts, 0.324 for by-items random intercepts, 0.100 for by-items random slopes for Flexion preceding sign, and 0.691 for residual error.

Figure 4

Graphical depiction of the main effects of flexion of the preceding sign (left panel) and flexion of the following sign (right panel) on flexion in the target sign.

Figure 5

Graphical depiction of the interaction effect between flexion in the preceding sign and flexion in the following sign on flexion in the target sign.

Figure 6

Graphical depiction of the main effect of spreading of the other fingers in the target sign on flexion in the target sign.

With respect to the random effect structure, the optimal model included only by-items random slopes for Flexion of the preceding sign. This indicates that there is some variation between items with respect to the strength of the effect of Flexion of the preceding sign on Flexion in the target sign. This could be traced back to the descriptive result reported in Table 2 that some target signs turned out to show more variation in terms of thumb position than others. Thus, some target signs were more flexible in terms of thumb position, and therefore more likely to be influenced by the thumb position of the preceding sign. The fact that exclusion of other random slopes did not lead to a decrease of the fit of the model compared to the maximal model indicates that the strength and directionality of all other fixed effects were roughly the same across participants and items.

3.2. Results: Abduction

A summary of the mixed-effects analysis on abduction that best fits the data is given in Table 5. The analysis yielded significant main effects for abduction in the preceding sign and abduction in the following sign. The main effects indicate that a higher degree of abduction in the preceding and/or following sign leads to a higher degree of abduction in the target sign (see Figure 7). In contrast with the results on flexion, there was no significant interaction effect between these main effects. Similar to the results on flexion, the analysis on abduction yielded a significant main effect of spreading, indicating that the degree of abduction in the target sign is higher with higher degrees of spreading of the other fingers in the target sign (see Figure 8). There were no significant interaction effects of spreading with the other predictors included in the model.

Table 5

Summary of the optimal mixed-effects logistic regression analysis for variables predicting the degree of abduction in the target sign.

Predictor	Estimate	SE	z-value	p-value
(Intercept)	–3.013	0.380	–7.926	< .001
Abduction preceding sign	1.464	0.226	6.472	< .001
Abduction following sign	0.601	0.214	2.804	.005
Spreading other fingers in target sign	0.351	0.174	2.020	.043

• Note. N = 728 (61 signers, 14 signs), Log-likelihood of model: –311.6. Standard deviations of random effect terms were: 0.467 for by-participants random intercepts, 1.002 for by-items random intercepts.

Figure 7

Graphical depiction of the main effects of abduction of the preceding sign (left panel) and abduction of the following sign (right panel) on abduction in the target sign.

Figure 8

Graphical depiction of the main effect of spreading of the other fingers in the target sign on flexion in the target sign.

With respect to the random effect structure, the optimal model only contained random intercepts. This indicates that the strength and directionality of all fixed effects were roughly the same across participants and items.

All in all, the results indicate that both flexion and abduction in the target sign are influenced by perseveration processes (preceding sign) and anticipatory processes (following sign). Interestingly, with respect to flexion, these perseveration and anticipation processes turn out to mutually influence each other (interaction effect). In addition to these perseveration and anticipation aspects of thumb position, the effects of spreading indicate that fingers other than the thumb can also influence flexion and abduction in the target sign.

4. Discussion

The main question in the present study was whether the position of the thumb (in particular the degree of flexion and the degree of abduction of the thumb) at a preceding or a subsequent sign predicted the position of the thumb at a target sign, when the articulation of the thumb is thought to play no phonological role in the target sign but does in either the preceding or following signs. We assumed that underspecification of features would lead to increased susceptibility to coarticulation (Keating, 1988), even though in the present corpus study we were not able to evaluate a detailed phonetic hypothesis about the precise shape of the trajectory of thumb articulation between specified targets (Keating, 1988; Boyce et al., 1991; Browman & Goldstein, 1992). Secondly, we examined whether phonetic spreading of the fingers at the target sign (which was assumed not to have the phonological feature spreading) played a role for the anticipation and perseveration effects of the position of the thumb.

We found evidence for coarticulation of handshape on the part of deaf signers of NGT, which replicates and expands on the general finding of coarticulatory effects in handshapes by Cheek (2001), Mauk (2003), and Wilcox (1992), who all looked at ASL. The realization of the preceding and following signs influenced the realization of the B handshape in target signs in NGT. This turned out to be true for the degree of flexion of the thumb depending on the degree of flexion of the preceding or following signs, as well as for the degree of abduction of the thumb, which similarly depended on the degree of abduction of the preceding or following signs. Moreover, the effect of flexion in the following sign was stronger with lower degrees of flexion in the preceding sign (and vice versa: The effect of flexion in the preceding sign on flexion in the target sign was stronger with lower degrees of flexion in the following sign).

We tested whether the extent of coarticulation was due to the degree of realization of spreading of the fingers. Results showed that spreading of the fingers indeed predicted the degree of flexion and abduction at the target sign. The data demonstrate that the degree of coarticulation of the flexion and abduction of the thumb was mediated by the degree of spreading of the fingers in the target sign. In other words, the degree of flexion as well as the degree of abduction in the target sign was higher with higher degrees of spreading of the other fingers in the target sign. Importantly, the effects of spreading did not interact with the coarticulation effects from the preceding and following signs. This means that the coarticulation effects were independent from the effects of spreading.

Interestingly, the different signs showed rather dissimilar patterns of spreading (see Table 3). Whereas some (B handshape) signs were hardly ever realized with spread fingers, other (B handshape) signs showed highly diverse realizations of the degree of spreading, even though none of the selected signs were assumed to have the phonological specification ‘spreading of the fingers.’ This result may partly be due to physiological constraints. For instance, when the fingers are flexed at the metacarpophalangeal (MCP) joints, very little spreading can be articulated due to the shape of the joint surfaces (cf. Ann, 2006: 80 for discussion). Another conclusion related to spreading is based on our finding that for none of the signs did we find a strong tendency for maximally spread fingers. We interpret this as evidence for the sign types we selected to have flat hands without the phonological feature ‘spreading.’

In a similar vein, the results indicated that the variation of positions of the thumb was not equal in the different (B handshape) target signs (see Table 2). While some signs were principally realized by a maximally extended thumb, other signs showed limited extension of the thumb. The same variation in realizations was found for the degree of abduction of the thumb. Again, this may partly be due to physiological constraints to articulate specific positions in certain signs (see also Russell et al., 2010). While the short hand muscles and the tendons of the long thumb muscles that allow movement at the three thumb joints are all independent of those of the other fingers (Kapandji, 1981), it could be that there are complex interactions with the degree of flexion at the wrist that would merit further investigation.

An additional part of the explanation of the variation of thumb positions at the targets may be of a phonological nature. It could be the case that certain combinations of phonological features (within the main parameters handshape, orientation, location, and movement) lead to a preference for specific realizations in a given phonetic dimension. This is what Battison et al. (1975) attempted to find out for some ASL signs, although they considered the ‘±extended thumb’ to be a binary phonological variable, rather than a phonetic continuum. This idea was taken up by van der Kooij (2002). She argued that NGT signs that are produced with lateral movements (in which the ulnar or radial sides of the hand either face the end of the lexical movement or contact the body location, if any) seem to be more prone to show an extended thumb position.

When we zoom in on our individual items, signs that have the relative orientation specification ‘ulnar side of the hand,’ and moreover do not have a location specification at the body, do indeed seem to show the most consistent pattern with respect to extension of the thumb. Take for instance the signs ALREADY and NOT, which have the relative orientation specification ‘ulnar side of the hand’; these signs are in many instances realized with maximally extended thumbs. The signs OH-I-SEE and HUMAN-BEING also have the relative orientation ‘ulnar’ and at the same time touch the location ‘torso’ with the fingertips. Similar to signs that had relative orientation ‘ulnar’ and had no location specification, those signs show a strong tendency to flex the thumb. The sign RECENTLY, which is located close to the signs OH-I-SEE and HUMAN-BEING, does not have the relative orientation specification ‘ulnar’ since the fingertips do not move down across the torso, as is the case in the previous two signs. However, some other signs that do not have the relative orientation specification ‘ulnar’ also showed a strong tendency to extend the thumb, such as the signs for YOUNG, SUPPOSE-THAT and TO-ASK. Overall, those patterns seem to suggest that signs with relative orientation specification ‘ulnar’ have a strong tendency to show extended thumbs when articulated in neutral space. However, this also holds for some signs that have different relative orientation specifications. Thus, the relative orientation seems to be influential, but it is not the sole factor determining thumb extension. More work is needed to investigate combinations of phonological specifications of different types.

This interaction between relative orientation and handshape features was also argued for by Crasborn and van der Kooij (2003), looking at finger configuration (bending), location, and orientation. While the specific nature of the articulators is different between sign and speech in many respects, the question of whether there are coarticulatory interactions between phonological elements of a different nature is not. In spoken languages for instance, while velar and laryngeal features would appear to be clearly dissociated, some have argued that vocalic nasality arises after laryngeal consonants in certain languages (‘rhinoglottophilia,’ Matisoff, 1975). Further investigation of the interactions between phonological features of different articulatory domains may prove especially fruitful for sign languages, since the articulators involved would at first sight appear to be more closely interlinked in sign than in speech: Flexion of the fingers has an impact on the location of the finger tips, to name but one example. Whether or not this has an impact on other phonological features or only on the articulatory gestures needed to optimally realize a phonological feature in that context is something that future research could investigate. In other words, it appears to be an interesting question whether the difference in articulators between the spoken and the signed modality only has an impact on the nature of the phonological features, their overall organizational principles being the same across modalities (as argued by Sandler, 1989, van der Hulst, 1993, and Brentari, 1998, among others), or whether the articulatory characteristics of sign also lead to substantially more interactions between features.

In previous studies on ASL, speed of articulation turned out to be an additional strongly influencing factor for the extent of coarticulation. In the present study, we did not experimentally control the signs; signers were recorded at their own natural signing speed. Nevertheless, we still found a clear pattern of coarticulation, both for flexion and abduction of the thumb. Thus, assuming that we can generalize across sign languages, the previous findings of coarticulation appear not to be limited to the elicitation of brief isolated sentences at a relatively high speed, but are generalizable to a more naturally (and thus variably) paced signing speed as shown in the conversations analyzed for the present study. The current study expanded on experimental work as well as the corpus study by Russell et al. (2010), by showing a unique and intricate role for spreading of the fingers in relation to coarticulation of the position of the thumb.

5. Conclusion

The current findings of signing in relatively free conversation show that the position of the thumb at the B handshapes is strongly influenced by the position of the thumb at the surrounding signs as well as by the degree of spreading of all fingers. We conclude from this that the active articulator used in signs is not always the whole hand, in the sense that the position of all fingers is fully specified. This is implicit in the concept of ‘selected fingers’ that is agreed upon by all phonological models, but few studies have actually empirically tested the predicted variability of the realization of the ‘unselected fingers.’ Probably for reasons of convenience, most authors have continued to use the term ‘handshape,’ but we argue that this term is in fact deceptive. By extension, visual representations of the manual articulators as whole hands in nearly all dictionaries and text books are also deceptive: They display phonetic forms that may be composed of phonologically driven material and phonetically variable material. Similarly, an audio example of a single vowel in a spoken language is not necessarily or fully informative about its phonological form. Thus, our study provides phonetic evidence for many phonological analyses of handshape in the past twenty years, which claimed that in many signs, the thumb is not an active articulator, and only (some of) the other fingers are phonologically specified (Sandler, 1989; Brentari, 1998; van der Kooij, 2002; Crasborn, 2003; Brentari & Eccarius, 2010). As earlier phonological models have implicitly (Brentari, 1998) or explicitly (Eccarius & Brentari, 2007) aimed to describe universal characteristics of sign language phonology, pertaining to multiple if not all sign languages, we likewise would like to suggest that phonetic evidence for different sign languages of the type contributed by the present study for NGT could inform our view on what are universal phonological properties of the thumb in sign language.

The present findings of coarticulation of handshape, combined with some suggested explanations of the degree of phonological variation in the different signs, seems to suggest that phonetic as well as phonological aspects contribute to the extent of coarticulation. Handshape as the description of ‘what is going on beyond the wrist’ appears to be a difficult construct, given that the articulation of specific handshapes is strongly related to the creation of a fluent stream of signs. Thus far, no study had established a clear link between the degree of spreading of the fingers and the flexible use of the thumb. Based on the findings in our present study, it is foreseeable that not only the thumb in the B handshape is predictable from various phonetic-phonological factors, but also other fingers and more hand configurations.

We see four ways in which the present study could be improved upon in future work. First of all, by taking into account the further phonological specification of the signs in the immediate environment, the interaction of other phonological features with the coarticulation process we observed can be established. The orientation of the selected fingers in relation to the location specification of the sign might restrict or facilitate thumb extension. As the two context signs SCHOOL and GOOD in Figure 2 illustrate, signs can look really differently in terms of their location and orientation, but also for instance in whether or not the thumb contacts any of the other fingers (called ‘aperture’ in the sign language literature). Secondly, improved training of the persons doing the transcription would lead to higher-quality data and contribute to our insight in thumb coarticulation. Thirdly, in addition to improved training of the annotators, it may also be possible to actually take inter-annotator variability into account by having multiple annotators code all data and include by-annotator random intercepts and/or slopes in the regression models. This was not possible in our study, as there was only one annotator who coded the entire dataset, and a subset of the data was recoded to obtain a view on the reliability of the annotations. Finally, avoiding the use of transcription by using kinematic measurements of the thumb rather than existing corpus data would be an attractive alternative. However, given the complexity of the thumb joints, this may well prove difficult.

Competing Interests

The authors have no competing interests to declare.