A ministry of (not-so-silly) walks: Investigating classifier handshapes for animate referents in DGS

2.2.1 Syntactic factors

One principal difference between the ‘1’ and ‘GC’ classifiers on the one hand, and the ‘2’ classifier on the other, is that the ‘2’ classifier has both whole entity and body-part classifier uses. The extended index and middle fingers can either represent a two-legged entity as a whole, or they can represent its legs only. Kimmelman et al. (2020) suggest for Russian SL (RSL) that internal movement of the index and middle fingers during the production of a ‘2’ classifier (predicate) indicates that its subject is agentive, while a lack of hand-internal movement indicates that the subject is a Theme (Benedicto and Brentari 2004, for ASL). Body-part classifiers tend to encode agentive, external arguments, while whole entity classifiers tend to encode thematic, internal arguments (Benedicto and Brentari 2004, Kimmelman et al. 2020, Gökgöz 2024). Translated into grammatical relations, an intransitive motion or location predicate could thus have either an unergative subject, i.e., an external, agentive single argument, or it could have an unaccusative subject, i.e., an internal, thematic single argument. Hence, while the ‘1’ and ‘GC’ classifiers in DGS are predicted to encode internal, thematic arguments (unaccusative subjects), the ‘2’ classifier may encode an external, agentive argument (unergative subject) if it has hand-internal movement.^[5] We will test this assumption by operationalizing agentivity and unaccusativity/unergativity via the notion of control. If an entity has control over the depicted movement, we assume they are agentive and form unergative subjects (e.g., climbing up a windowsill); if they do not have control over the movement, we assume them to be thematic and form unaccusative subjects (e.g., being catapulted into the air).

2.2.2 Semantic factors

Classifier predicates primarily describe events of motion, positioning of an entity in space, or their existence (Zwitserlood 2012). All three classifiers under investigation can express directed motion events (e.g., Zwitserlood 2003 for NGT, Engberg-Pedersen 1993 for ‘1’ and ‘2’ in DTS). Concerning positioning, NGT has been claimed to frequently use the ‘1’ classifier to localize, i.e., anchor an entity in the signing space (Zwitserlood 2003, 133), while in DTS, signers are reluctant to use ‘1’ with stative verbs. Instead, the ‘2’ classifier has been observed to contrast different positions of an entity well (e.g., sitting, lying on one’s back vs on one’s side, standing on one’s head, Engberg-Pedersen 1993, 248). One factor we investigate for DGS is thus whether classifier choice is associated with a particular event type: motion vs location of an entity. Within motion events, we further examine whether the shape of the movement path or its direction has an effect on classifier choice. For instance, movement toward the signer is preferentially expressed via the ‘1’ classifier in NGT and DTS (Zwitserlood 2003, 133, Engberg-Pedersen 1993, 248), while some DTS signers do not accept the ‘2’ classifier to depict approaching a viewpoint holder (often the signer). The ‘GC’ classifier in NGT is hypothesized to occur when the trajectory of a movement is emphasized over any particular properties of the moving entity (Zwitserlood 2003, 134).

In addition to the path and direction of a motion event, classifier predicates can also encode the manner of a movement. The concept ‘manner’ is relatively broadly defined; Geuder and Weisgerber (2008) identify the following sub-components of manner based on Levin’s (1993) list of verb types:

• Internal movement, i.e., movement of an entity around its own axis (spin, swing, rock, oscillate, and roll);

• Path shape (zigzag);

• Type of surface contact (slide, bounce);

• Step pattern (run, limp, crawl, gallop);

• Speed (dash, hurry).

Some manner components relate more to the path (e.g., its shape) or to movement dynamics (e.g., speed), while others concern the moving object itself, e.g., which extremities are used to propel the entity forward, or how they make contact with the ground. One might hypothesize that manners which emphasize the movement of the legs are preferably encoded by the ‘2’ classifier in DGS, an observation frequently echoed in the sign language literature (e.g., Benedicto and Brentari 2004 for ASL; Engberg-Pedersen 1993 for DTS, Zwitserlood 2003 for NGT, or Kimmelman et al. 2020 for RSL). In contrast, manners highlighting the dynamics of a movement event may preferentially be encoded by ‘1’ or ‘GC’; in fact, NGT seems to prefer ‘1’ for fast movements (Zwitserlood 2003, 133). In this article, we will look at manner by coding each scene from the Canary Row cartoons with respect to the subcomponents of manner depicted therein.

2.2.3 Discourse-pragmatic factors

This section discusses three factors that relate to discourse-level phenomena: the discourse status of a referent, whether the non-dominant hand represents a second discourse referent that the dominant hand interacts with, and the presence of constructed action (CA). We list CA among the discourse-level phenomena here because it frequently occurs in narrative sequences and relates to different signing perspectives.

3.2.1 Exclusion criteria

We annotated only classifier predicates containing the ‘1’, ‘2’, or ‘GC’ classifier and their respective variants whenever they were used to refer to anthropomorphic referents, i.e., humans and human-like creatures. This means that classifier predicates depicting the movements of the monkey in episode four were not annotated. It also meant excluding classifier predicates featuring a different handshape, e.g., (Section 2.1). Finally, it means excluding both lexical signs and CA used to describe animate movement, location, or manipulation. Some lexical signs were easy to identify as such (e.g., take-off, run), but in other cases, the distinction was more difficult. Here, we list the criteria we used to distinguish between lexical signs and classifier predicates.

• Annotation as $PROD in the DGS Corpus: The DGS Corpus annotates classifier predicates as $PROD or ‘productive’ signs and we adopt their classification (this criterion applies only to the three DGS Corpus re-tellings).

• Depictive movement: Signs with a ‘1’, ‘2’, or ‘GC’ classifier representing the actions of an anthropomorphized or human referent are considered classifier predicates (even without a $PROD gloss) if their movement can be interpreted depictively. Take, for instance, the depictions of ‘falling’ in Figure 3. The lexical sign fallen in 3a employs the 2-handshape representing the legs of a falling human being and has an outward path movement with a wrist flip. It does not mark the start- or endpoint of a real movement event, nor its trajectory. The token in 3c, on the other hand, marks the starting point of the cat’s fall on the non-dominant hand as the open window, which is consistent with the referent’s location established in the previous frame (shown in 3b). Further, it has a larger path movement and additional wrist movement, which depict more aspects of the dynamics of the movement event. Hence, tokens like 3c but not like 3a were annotated as classifier predicates in our dataset.

Figure 3

(a) Lexical sign fallen ‘to fall’; (b) referent ‘cat’ set up slightly to the left in signing space; (c) classifier predicate congruent with the lexical verb fallen (images from Konrad et al. 2020).

3.2.2 Phonological variants

Given our primary focus on classifier choice for anthropomorphic entities, we first annotated the respective handshape on a handshape tier. To distinguish between ‘1’ and ‘GC’ classifiers, which share a handshape, an extended index finger was categorized as ‘1’ wherever the orientation of the handshape was compatible with the position of the depicted character (typically upright, but during fast movement sometimes almost horizontal). If the index finger was oriented palm-down or with the fingertips pointing down, the token was categorized as ‘GC’. During a character’s upward movement, ‘1’ and ‘GC’ are impossible to distinguish; in those cases, we categorized tokens as ‘1’ because it is the more specific entity classifier. Note that there were also a few tokens of the ‘2’ classifier whose orientation did not match that of the character. These were all classified as uses of ‘2’.

Of the three classifiers under investigation, the 1- and 2-classifiers in our dataset show some form variation, which we annotated in the ELAN tier handshape. Both classifiers have variants with extended vs slightly bent knuckles (bent variants are represented in Figure 4a, b, and d). The 2-classifier furthermore appears in a variant where the index and middle finger contact each other. We annotated this variant as ‘N’ for ‘non-spreading’ and illustrate the extended variant in 4c and the bent variant in 4d. Further, both ‘2’ and ‘N’ can move the index and middle fingers independently (annotated as 2m or Nm). Finally, finger bending and moving can occur simultaneously, which is notated as 2bm or Nbm.

Figure 4

Handshape variants of classifiers ‘1’ and ‘2’ (Image (a) is from the VIDI Sign Space project; (b–d) are taken from Konrad et al. 2020).

3.2.3 Syntactic, semantic, and pragmatic factors annotated

Given the exploratory nature of this investigation, we annotated 15 different factors to obtain a broader overview of which aspects potentially influence classifier choice. The factors are listed in Table 1. Note that this investigation focuses on linguistic factors constraining the use of the different classifiers only. Sociolinguistic factors such as gender, region, age, or age of acquisition of the signers were not taken into consideration.

The entity referenced by the classifier is annotated in the entity tier. Possible values are Sylvester (the cat), Tweety bird, the monkey, Granny, or the receptionist.

In the event tier, we annotated whether the predicate depicts the motion of the referent or its location in space. For movements, there are more detailed annotations in the following tiers. The shape of the movement, for example straight, arc-shaped, or complex, is annotated in the movement tier. The direction then indicates where the hand moves along the transverse, vertical, and sagittal axes. Given the discussion in Section 2.2.2, we also annotated movement toward or away from another character (including the signer).

In the congruence tier, we noted whether a classifier predicate has the same shape as a lexical verb (congruent) or not. For example, if two people walk toward each other in a straight line and meet, this event will likely be represented by the hands with extended index fingers meeting in front of the torso. This form is identical to that of the lexical verb treffen ‘meet’ and thus congruent with it, but it also depicts the event iconically, as a classifier predicate does. Such cases are annotated as ‘congruent’ with a lexical sign. At 79.3%, most of the classifier predicates we analyzed were depictive (490/618 tokens), and only 128 (20.7%) were annotated as congruent. Since Pearson’s chi-square test revealed that congruence does not have a significant association with classifier handshape choice (p = 0.2784), we assume that congruent forms behave like classifier predicates for all intents and purposes and include them in the analysis.

For motion predicates only, we annotated whether the classifier predicate describes the path of the motion (e.g., when the cat moves from A to B), or the manner of the motion (e.g., when the cat jumps) on the tier manner. Manner and path can also be combined. Whether a motion predicate depicted a manner of motion was determined on the basis of several factors. On the one hand, variants of the ‘2’ classifier that featured the index and middle fingers extending and flexing were considered to depict a walking or running manner. On the other hand, wrist flexion and extension in the ‘2’ classifier and shoulder flexion and extension in the ‘1’ classifier were interpreted as a bouncing or hopping manner. Following Levin (1993), ‘fall’ is not characterized as a manner verb, while ‘jump’ and ‘zigzag’ are.

The last tier pertaining only to motion events is the controlled tier, where we differentiate between a controlled motion, such as walking, and an uncontrolled motion, such as falling down, being catapulted upward, or crashing into a wall. Importantly, we focus only on the action itself and not on the context of the cartoon. This means that an incident where Sylvester runs on an electric cord and Tweety and Granny are behind chasing him, although the situation as a whole is not under Sylvester’s control, his running action is still considered a controlled motion.

The discourse tier records the discourse status of a referent as introduced for the first time (usually at the start of the re-telling of an episode), as reintroduced after at least one clause (i.e., predicate) featuring a different discourse referent, or as maintained. If the preceding clause was about the same referent as the current one, we count the referent as maintained.

If the classifier predicate on the one hand interacts with another referent represented on the second hand, we annotated this in the interaction tier. Whether the second referent was inanimate, animate, and/or human was specified in the second-hand tier. Inanimate referents are, for example, Tweety bird’s cage, a pipe, or a catapult. We did not annotate interactions with default surfaces on which someone moves, e.g., the street or the pavement. An animate referent would be the monkey, whereas human-like referents are the cat, the bird, the grandmother, and the receptionist.

We also annotate if a classifier predicate is accompanied by CA to depict the emotions or bodily actions of a referent in the CA tier. A related tier, perspective, recorded whether an event is portrayed from an observer’s perspective, from a character’s perspective, or from both. Observer perspective was annotated if a referent is only represented via a classifier on the hand, while character perspective was annotated if the signer uses CA and/or a handling classifier on the non-dominant hand to portray the same character that is represented as ‘1’, ‘2’, or ‘GC’ on the dominant hand. The value ‘both’ was chosen if the signer represents one entity via ‘1’, ‘2’, or ‘GC’ on the hand and another entity on the signer’s body (via CA and/or a handling classifier).

We also sorted the classifier predicates into the episodes or scenes they appeared in and the lexical meaning they convey. Finally, accompanying mouthing or mouth gestures were recorded in the mouth movements tier. They were used to aid in the distinction between classifier predicates and lexical verbs, as the latter tend to be accompanied by mouthing, while the former tend not to be.

4.4.1 Motion or location

We wanted to know whether there is a connection between the choice of the ‘1’- vs ‘2’-classifier handshape with the type of event that is depicted: the motion of an entity or its positioning/location in space. Since locations can never be depicted by ‘GC’, we excluded it from the analysis. From a total of 527 tokens, the majority of classifiers depicted motion events (n = 479 or 91%), while 48 (9%) depicted the location of the entity (one token was unclear). A chi-square test revealed a significant association between event type and classifier choice (χ(1) = 8.05, p = 0.0046). The standard residuals show that while there is no significant difference in handshape choice in motion events, events of positioning an entity in space are significantly less often than expected represented by the ‘1’ classifier (z = −2.53).

4.4.2 Movement shape and direction

We checked whether there was an association between the types of movement paths an entity performed as well as the direction of movement, and the choice of classifier. Given a lack of pre-specified categories for movement type, this factor had a wider range of annotation values (straight, arc-shaped, circular, zigzag, spiral, hopping, and complex) and cannot be analyzed statistically. Still, looking at the absolute numbers for each annotation value, we note that ‘1’ encodes mostly straight and arc-shaped movements, as well as zigzag (12/29 tokens) movements. ‘1’ was not used for spiral, hopping, or complex movements. ‘GC’ does not encode spiral or hopping movements but can sometimes encode complex movements combining more than one movement type (3/11 tokens). No clear (dis-)preference for a movement type is observable for ‘2’.

Turning to the direction of movement, we looked at both absolute direction and direction relative to another character in the event. This character could either be represented by the body (2 tokens), or by the other hand (25 tokens). For absolute directions, we coded movement along the transverse (left or right), the vertical (up or down), and the sagittal (toward or away from body) axes. An association with classifier choice was found (χ(4) = 27.20, p < 0.001) that was driven by the distribution of the ‘GC’ classifier: It occurs more often than expected in sagittal movements (z = 2.93), all of which move away from the body (36/36 tokens), and it occurs less frequently than expected with vertical movements (z = −3.40).

Turning to relative movement, we observed an association between relative movement (absent, present and toward a character, present and away from a character) and classifier choice (Fisher’s exact test, p < 0.001). This association was mostly driven by the ‘1’ classifier occurring significantly more often when encoding movement toward another character (z = 4.03) and a trend against using either ‘2’ (z = −1.51) or ‘GC’ (z = −1.51) for this purpose. More generally, the ‘1’ classifier is used somewhat more often when describing movement relative to another character; it also encodes movement away from a character tendentially more often than expected (z = 1.89).

4.4.3 Manner and path

To show whether a given classifier is preferentially used to encode the path of a movement, the manner in which it is performed, or both, we coded each classifier predicate depicting a motion event (n = 555) for the presence of manner and/or path. Only six classifier predicates did not encode a path (five of them using the ‘2’-classifier), while 414 (74%) encoded only path, and 135 (24%) combined path and manner. In the following, we thus focus on comparing classifier predicates that do express a manner with those that do not. The association between the presence/absence of manner encoding and the type of classifier selected was significant (χ(2) = 80.90, p < 0.001). Based on an inspection of the standardized residuals, the ‘2’-classifier occurs significantly more often than expected (z = 4.65) when manner is depicted, while the ‘1’ and the ‘GC’ classifiers occur significantly more often than expected (z = 2.38 and z = 2.74) when only path is encoded.

4.4.4 Influence of different manner components in the depicted event

Each re-told scene from the Canary Row cartoons was coded for the presence of one or more of Geuder and Weisgerber’s (2008) manner components: type of surface contact, step pattern, path shape, and speed. Given the nature of our dataset, we also checked whether the salient use of extremities other than the feet (wings, front paws) had an impact on classifier choice. In the following subsections, we show that neither manner of surface contact nor step pattern was associated with a particular classifier handshape, whereas path shape, speed, and the salient use of front paws and wings impacted classifier choice.

4.5.1 Discourse status

About 77.2% of the classifier predicates had a maintained discourse status (477/618 tokens); 126 times the entity was reintroduced (20.4%) and 14 times the entity was first introduced at the beginning of the narrative (2.3%). One token was unclear. Keeping in mind the small number of occurrences for discourse-new referents, Fisher’s exact test nonetheless attested a significant association between discourse status and the handshape of the classifier (p = 0.0016). The standardized residuals reveal that if an entity is first introduced into the discourse, signers use ‘1’ more often than expected (z = 3.29).

4.5.2 Presence of CA

Since sign languages allow the simultaneous encoding of one entity via more than one articulator (body partitioning, see e.g., Dudis 2004), a sizeable number of tokens in the Canary Row re-tellings involve a character being represented at the same time on the hand via ‘1’, ‘2’, or ‘GC’, and on the signer’s face and body via CA. The question arises whether a given classifier exhibits a preference for occurring with or without CA. Looking only at cases where the same character is represented via a classifier and CA, we find an association between the presence of CA and classifier choice (χ(2) = 7.66, p = 0.022). The association seems to be driven primarily by the ‘GC’ classifier, which tends to occur less often than expected with CA (z = −1.64). The association persists if we take into consideration all uses of the classifier and CA, including those where hand and body represent different characters (χ(2) = 10.24, p = 0.0006). Here, ‘GC’ is significantly less frequently used with CA (z = −2.28).

4.5.3 Interaction with other entities

Sylvester, Tweety & co. were depicted as interacting with another animate (33 tokens) or inanimate entity (273 tokens) in about 50% of all their occurrences. Inanimate entities depicted on the non-dominant hand included the source or goal of movement as well as non-default surfaces such as a windowsill or a plank of wood. There was a significant association between interaction (absence, interaction with animate entity, interaction with inanimate entity) and classifier choice (Fisher’s exact test, p < 0.001). The ‘1’ classifier was used significantly less frequently than expected with inanimate objects depicted on the non-dominant hand (z = −3.02). A trend was also observed to use the ‘GC’ classifier more often than expected with inanimates (z = 1.87), which may be due to the frequency of a predicate congruent with the lexical verb enter (illustrated in Figure 5a). Interactions with other animate referents were less frequent, in most cases the second referent was depicted on the non-dominant hand, but in four tokens, it was mapped onto the signer’s body (as shown in Figure 5b). An inspection of the standardized residuals shows that ‘1’ occurs significantly more often than expected in depictions of interactions between two animates (z = 4.30), while ‘2’ tendentially occurs less often than expected (z = −1.72).

Figure 5

(a) Sign enter in DGS; (b) interaction with a second animate referent mapped onto the signer’s body: Sylvester (‘1’ classifier) is chased and beaten by Granny (CA) (both images are taken from Konrad et al. 2020).

5.4.1 Syntactic factors

Syntactic factors were shown to influence classifier choice in DGS. The type of thematic role assigned to the argument of a movement verb impacted the choice between ‘1’, 2’, and ‘GC’, but not always in the way we predicted. Our study confirms findings from other sign languages (Benedicto and Brentari 2004, Kimmelman et al. 2020, Gökgöz 2024) suggesting that body part classifiers align with agentive, external arguments. Whenever ‘2’ is used as a body part classifier in our DGS data, i.e., when it has internal movement, it shows a clear preference for encoding arguments whose referents perform a controlled movement and are, hence, agentive. However, we cannot confirm proposals that whole entity classifiers generally show a preference for encoding internal arguments. In the DGS data, ‘1’ shows a dispreference for encoding entities involved in uncontrolled movements (i.e., internal arguments), and ‘GC’ showed no preference either way. Only ‘2’ tendentially occurred more often with internal arguments when used as a whole entity classifier (i.e., without internal movement).

5.4.2 Semantic factors

We showed that classifier choice was associated with event type in that the ‘GC’ classifier only occurred in motion events and the ‘1’ classifier was dispreferred in positioning events, leaving the ‘2’ classifier for events of locating or positioning an entity in space. This is in line with previous findings on DTS (Engberg-Pedersen 1993) but differs from NGT (Zwitserlood 2003), where the ‘1’ classifier occurs with stative localizing verbs more often. Within movement events, we showed the direction of movement relative to another entity to affect classifier choice. Similar to DTS and NGT, the ‘1’ classifier was preferred for movements toward the signer, but we observed a more general tendency for ‘1’ to be used for movement in relation to another animate entity (mostly toward them). We did not replicate DTS signers’ dispreference for showing movement toward the signer with the ‘2’ classifier, but we did observe a more general dispreference for depicting interaction between animate entities using the ‘2’ classifier.

Concerning manner of movement, DGS aligns with many other sign languages in preferentially encoding different manners of movement (in addition to a path) by the ‘2’ classifier. However, it turned out useful to break down manner into Geuder and Weisgerber’s (2008) more nuanced sub-components. We showed that what is intuitively considered ‘manner’, i.e., types of surface contact and step patterns, had no impact on classifier choice, while movement dynamics and typical vs atypical manners of (human) motion did have an impact. Only when manners not involving salient bipedal motion were depicted (flying, crawling, and climbing) did signers prefer ‘1’ or ‘GC’. Likewise, movement speed was found to affect classifier choice, in that ‘1’ was slightly dispreferred for fast movements. Note that this finding differs from NGT, where ‘1’ is preferred for fast movements (Zwitserlood 2003). In sum, at least for classifier constructions in DGS, it is helpful to consider the different components of the rather heterogeneous concept ‘manner’ separately, in turn validating Geuder and Weisgerber’s distinctions.

When only paths were encoded, ‘GC’ and ‘1’ were observed more frequently than expected given their absolute frequency in the dataset (the same was proposed for ‘GC’ in NGT by Zwitserlood 2003).

5.4.3 Discourse-pragmatic factors

While discourse status had some impact on classifier choice, the findings did not align with our predictions: New referents were introduced into the discourse with either of the three classifiers, but preferentially with ‘1’. The ‘1’ classifier is less phonologically complex than the ‘2’-classifier and should thus be preferred for discourse referents with high cognitive accessibility (i.e., maintained referents), but that is not the case. Given the double duty of ‘2’ as a whole entity and a body part classifier, we propose that ‘1’ is nonetheless the more likely choice for discourse-new referents as it presents a more holistic introduction of the referent, not focusing on its legs. For the same reason, our prediction that ‘GC’ should be preferred for new discourse referents was also not borne out; ‘GC’ might easily encode less individuated referents, but since it does not depict any properties of a referent, it is unlikely to be able to introduce a new referent into the discourse at all (with one exception).

Observations from DTS had led us to predict that interactions with another entity should influence classifier choice, and we did find that whenever another animate entity is present in the discourse (encoded on the signer’s body or second hand), DGS signers prefer ‘1’ and disprefer ‘2’. On the other hand, the presence of CA had no impact on the choice between ‘1’ and ‘2’; each occurred in accordance with their absolute frequencies when signers partitioned off a part of their body to represent an entity on their hand and (another or the same one) on their torso/face. However, we did find the ‘GC’ classifier to be dispreferred when CA is present. This seems odd at first, given that there’s a clear division of labor between ‘GC’ representing a movement trajectory and CA individuating the moving entity. If we assume, however, that a classifier needs to introduce enough properties of an entity to anchor a referent in the discourse and make them available for anaphoric reference, and that CA constitutes a form of anaphoric reference, then perhaps it is not surprising that GC should tend not to occur with CA. This behavior would align with ‘GC’ not introducing new discourse referents.