Getting attention in different languages: A usage-based approach to parenthetical look in Chinese, Dutch, English, and Italian

2.1 Corpus data

The data for this study comes from the following sources:

1. the Center for Chinese Linguistics Corpus (CCL) and particularly its contemporary language components of general conversation and of institutionalized interviews (see Center for Chinese Linguistics 2020);

2. the Corpus Gesproken Nederlands (CGN), with Dutch speech material from the 1990s and early 2000s, and, more specifically, its Netherlandic Dutch sections of spontaneous conversations and broadcast interviews and discussions (Dutch Language Union 2004);

3. the British National Corpus (BNC), with material from the mid-1970s to the early 1990s, and the spontaneous conversations and interviews and discussions in particular (BNC Consortium 2007);

4. the Perugia Corpus (PEC), a resource of Italian from the 1990s and 2000s, and, more specifically, the components of informal (telephone) conversations and institutionalized interviews (Spina 2014).

The sizes of the subcorpora for each language are given in Table 1.

As the preceding paragraph makes clear, the focus is on spoken language. The rationale for the choice is that parentheticals, especially the ones under examination here, can reasonably be assumed to be more common in and typical of speech than writing (perhaps excluding highly dialogic fiction and plays). Regular conversational data in particular has been selected because it can tell us how look functions in everyday language. The addition of discussions and interviews is motivated by the more institutionalized nature of the interaction in these genres. The more fixed roles of the interlocutors, the more structured distribution of turns and the narrower topic of conversation (e.g. an issue that people have diverging opinions on) may entail a slightly different usage of look. It seems less probable, for example, that the parenthetical is employed in its literal sense, as a directive to look at something present in the situation: the (often neutral) setting of interviews or discussions rarely serves as their subject. Conversely, the use of look in (1) to (3) as what Van Olmen (2010a: 228) labels an “argumentation marker” (see also Aijmer 2018: 183–184), may be more frequent.

It is important to acknowledge that there is variation in the data used for this study. The Italian conversations, for instance, include telephone calls whereas the Chinese ones do not. In the same vein, the English material originates from the mid-1970s to the early 1990s while the Dutch material dates from the 1990s and the early 2000s. Such dissimilarities are almost unavoidable, however, when one requires a sufficiently large amount of comparable data from four different languages and is therefore forced to rely on existing resources. The corpora and components selected here are, in other words, as similar as could practically be achieved within this context and in view of further limitations such as accessibility and searchability. Even so, the variation needs to be borne in mind, of course, when interpreting the results.

The analysis is based on a sample of 200 cases of parenthetical look per language, half of which come from the conversations and the other half from the interviews and discussions. To obtain the samples, we searched the two relevant parts of every corpus separately, by means of the tool specific to the resource or the all-purpose WordSmith 7 software (Scott 2016), for the strings: (i) 你看 nǐkàn and 你们看 nǐmenkàn in Chinese; (ii) kijk in Dutch; (iii) look in English; (iv) guarda, guardi and guardate in Italian. What these forms of look have in common is that:

1. they are the most basic way in which each language can order one or more addressees to look at something, i.e. the imperative;

2. they have the potential to function as parentheticals.

Dutch and English have essentially only one such form, i.e. kijk and look respectively.^[6] In Italian, there are three variants: singular guarda, polite guardi and plural guardate. ^[7] Chinese, lastly, combines 看 kàn ‘look’ with the pronoun 你(们) nǐ(men) ‘you (all)’ (see also Long et al. 2019, who argue that pragmatic 你看 nǐkàn indeed derives from imperative 你看 nǐkàn). In the next step of the extraction, all hits of these strings were put in a random order. From the resulting lists for each component of each corpus, we then manually selected the first 100 parenthetical instances of look, for which (4) can serve as an example.

你看 nǐkàn here is propositionally redundant. It no longer conveys the meaning of visual perception and fulfills the parenthetical function of expressing the speaker’s marked expectation that the addressee will agree with what is being said. Limiting ourselves to parentheticals as defined in Section 1 meant that hits such as (5) and (6) had to be filtered out. In (5), guarda is not used as a propositionally redundant surplus of meaning. It is rather part of the syntax of a longer clause and not some type of aside connecting a sentence in one way or another to the context in which it is uttered. Kijk in (6) does appear to fulfill some sort of pragmatic role, expressing B’s (mild) surprise at the information that A has just given him and acting as a backchanneling device similar to oh ‘oh’ and mm-hu ‘uh-huh’. It is not a parenthetical, however. Similar to guarda in (5), it is not propositionally redundant, constitutes a turn in its own right together with ja ‘yes’ and conveys some kind of directive/expressive illocutionary force as a distinct speech act.

Note that, unlike what (1) to (3) might suggest, parenthetical look in Romance and Germanic languages need not be restricted to a position to the left of the clause. In (7), for instance, look occurs to the right of a syntactically independent clause, which it relates to the situation by urging B to look at the evidence for the claim. Kijk in (8) has a similar function but is found in the middle of a sentence: it is followed by the direct object deze salade ‘this salad’ of the preceding clause fragment ik heb hier nog ‘I still have’.

Another possible analysis of (8) would view ik heb hier nog as an incomplete sentence and kijk as introducing some distinct add-on. We take such examples at face value, however, and thus consider them clause-medial, because deze salade can in principle be regarded as syntactically linked to what precedes kijk and the alternative analysis is hard to prove.

2.2 Analytical framework

The corpus data is analyzed in terms of a variety of parameters, described in Sections 2.2.1–2.2.7 (Section 2.2.8 discusses the analytical procedure), with the aim of identifying in a bottom-up way what we call “illocutional concurrences” (see Tantucci and Wang 2018, 2020a). These are data-driven intersections of form (e.g. kijk), illocution (e.g. the speech act of the clause marked by kijk) and contextual situatedness (e.g. turn or clause position, as well as the genre where the linguistic act is realized). The notion of concurrence thus differs crucially from that of collocation, as it is not limited to the textual co-occurrence of two lexemes. Rather, concurrences underpin behavioral tendencies in which linguistic forms recursively intersect with specific linguistic acts that are distinctively realized under particular contextual conditions. Put simply, while collocates pertain to text as such, concurrences underpin interactional behavior (e.g. factors like illocutionary force, contextual situatedness, position within a turn-at-talk and so on).

3.1 Relative frequency

Figure 1 presents, for each language, the number of instances of parenthetical look per 10,000 words, in both the conversations and the interviews and discussions.

Figure 1:

The relative frequencies of parenthetical look.

Note, however, that these numbers are extrapolations. Conversational kijk can serve as an example. We had to analyze 150 of the 1,297 randomized hits in this corpus component to find the 100th parenthetical attestation. The total number of hits (1,297) was then multiplied by the proportion of 100–150 (0.67), giving us an approximation of 864.67 for the overall absolute frequency of parenthetical kijk. This number, finally, was divided by the size of the subcorpus (1,747,789) and multiplied by 10,000, producing the relative frequency of 4.95. In view of this procedure, some caution about the numbers in Figure 1 is warranted. Another reason^[11] is that the frequency with which individual speakers use look may vary considerably. As not all our corpora have the required information, we are, unfortunately, unable to build inter-speaker variability into our analysis.

To compare the frequencies of look across genres or languages, we use Rayson and Garside’s (2000) log-likelihood calculator.^[12] This test has been developed especially for contrasting absolute frequencies in corpora of possibly dissimilar sizes and produces a G ² value, which is indicative of a significant difference when higher than 3.84 (p < 0.05) and a highly significant one when higher than 10.83 (p < 0.001). When one frequency is compared in this way to more than one other frequency, the standard significance level is Bonferroni-corrected: it is divided by the number of comparisons made, to mitigate the danger of giving too much meaning to one or more of the tests yielding a p-value lower than 0.05 as they could simply result from chance (see Baayen 2008: 114).

A first observation concerns the two genres in the four languages.^[13] In Italian and English, look is substantially more frequent in conversations than in interviews and discussions (16.45 vs. 2.13 cases per 10,000 words and 3.93 vs. 0.61 respectively). In Dutch, however, it occurs significantly more often in interviews and discussions than in conversations (6.59 vs. 4.95) and, in Chinese, there does not appear to be a difference (1.01 vs. 1.12). These results could be interpreted as indicating that guarda and look are more typically colloquial than kijk and nǐkàn. The findings may also be due to the dissimilarities between the subcorpora in the four languages, of course (see Section 2.1). The absence of a difference in Chinese is probably also not unrelated to nǐkàn’s comparative overall infrequency.

A second observation has to do with the rate of occurrence of look across languages.^[14] Guarda and kijk are substantially more frequent in the conversations as well as the interviews and discussions than look and nǐkàn. The relation between Italian and Dutch differs in the two genres, though: in the conversations, guarda (16.45 instances per 10,000 words) occurs significantly more often than kijk (4.95) but, in the interviews and discussions, kijk (6.59) is significantly more frequent than guarda (2.13). The relation between English and Chinese too varies between both genres: in the conversations, look (3.93) has a substantially higher rate of occurrence than nǐkàn (1.12); in the interviews and discussions, nǐkàn (1.01) is substantially more frequent than look (0.61). A range of factors may be at play here, such as the aforementioned differences in the make-up of the subcorpora, the (non-)existence of direct functional competitors to look in the language (see Romero Trillo 1997; Van Olmen 2010b; Aijmer and Elgemark 2013 about look vs. ‘listen’) and the speakers of a language’s partiality to a particular use of the marker in communication (see Section 3.3 on guarda frequently correlating with expressive illocutionary force, a function that is probably fully exploited by speakers of Italian in spontaneous conversation but is less compatible with interviews and discussions).

3.2 Conditional importance of variables

To assess the cross-linguistic variation of look, we first computed the conditional importance of all the parameters/variables that could potentially predict language diversity. To do this, we fitted a random forest model, which is a multivariate learning method for classification resulting from an n number of individual decision trees (see Ho 1995 among others). Decision trees, in turn, are multifactorial models of decisions and consequences, in which the outcome variable depends on a set of statistical “decisions” based on levels of significance that lead to a hierarchical classification. The simulation of a large number of trees allowed us to compare (n = 1,000 in the present analysis), for the outcome variable of language, the weight of each predictor. We fitted language as the outcome variable, while our predictors were: (i) the speech act of the sentence marked by look (SA1); (ii) that of the subsequent one (SA2); (iii) whether look would be used as part of a quotation; (iv) its position within the clause; (v) its position within the turn at talk; (vi) whether it expresses a literal meaning of visual perception or not; and (vii) whether it would include a vocative. The actual model output, in Figure 2, was obtained with the “cforest” function of the R “party” package (see Hothorn et al. 2006).

Figure 2:

Random forest of the conditional importance of variables.

The cut-off value for Figure 2’s interpretation is the absolute importance weight of the variable with the lowest score (see Levshina 2015: 298), which always varies depending on the nature of the data. In this case, the conditional weight of each parameter ranges from around 0 (vocative) to a highest score approaching 0.7 (SA1).^[15]

From the dot-plot in Figure 2, we can clearly gather that it is the illocutionary dimension that primarily predicts the cross-linguistic variation in look.^[16] In particular, the speech act that concurs with the use of the marker (SA1) is the strongest predictor, followed by the parameter of whether look occurs in reported speech or not (quotation) and then by the speech act that sequentially follows a prior occurrence of look (SA2). Other parameters that are at play include turn position and contextual features hinging on the genre in which look appears, followed by less impactful effects such as clause position, perceptual (as opposed to pragmatic) meaning and, finally, the presence/absence of vocatives (this last variable sets the lowest value for the weighted prediction of language diversity). This hierarchy indicates that the most prominent predictor of language diversity when attention-getting with look is at play is the illocutionary force. In view of this fact, the next section is dedicated specifically to the cross-linguistic comparison of speech acts where look is used.

3.3 Speech act

As shown in Figure 2, speech acts are the most important predictor for assessing the cross-linguistic variation of look. However, all we can gather from the model above is that different types of illocutionary force are important to predict how languages differ from one another. To see exactly how they do this, we fitted a network model, so that we could capture the dynamic transition from one speech act (SA1) to another (SA2), when attention-getting is encoded via look as a parenthetical. Figure 3 thus enables us to assess the impact and the directionality of the speech acts associated with look – asser(tives), dir(ectives), comm(issives), expr(essives), decl(arations), eval(uatives) – in Ch(inese), Du(tch), En(glish) and It(alian). The plot is a directed graph obtained with the RStudio ‘Igraph’ package for network analysis (Csardi 2013).

Figure 3:

Speech acts and their sequences concurring with look.

The speech act circles are to be interpreted as follows: the bigger the circle (also termed “edge”), the higher the frequency of its speech act with look in the language. The sequential relationships between the first and second speech acts are captured by the arrows (also termed “vertices”), the thickness of the lines and the distance between the circles. More precisely, thicker lines and closer distances indicate stronger directional relationships from the one speech act to the other.

A first observation concerns assertives. It is clear from Figure 3 that, cross-linguistically, they have the strongest association with look. This is indicated by the size of the assertive circles, which is distinctively bigger than the other speech acts in all languages. This fact may be attributable to an overall prevalence of assertives in conversation, but it is also worth mentioning that this speech act type is already favored by look in its literal sense. Telling someone parenthetically to look at something naturally combines with a description of the object or event highlighted – like the state in (7), repeated as (27), for instance. The high frequency of assertives, with pragmatic cases of the marker as well, may therefore be taken as a sign of persistence of the original meaning/context of look used parenthetically.

This persistence is especially evident for Chinese and even more so for Dutch, as their assertive circles are considerably bigger than the English and Italian ones. In other words, the reason why the variable of speech act scores so high in Figure 2’s random forest analysis is because it sets the languages under investigation apart from each other: if a speech act is not assertive, chances are low that the language is Chinese or Dutch. Additionally, we can see in Figure 3 how, in these two languages, a first assertive exhibits a strong tendency to lead to a second one, as shown by the very thick arrows making a U-turn from “asser” to “asser” and as exemplified by (28). In this example, the second statement can be said to qualify the first one: war is linked to power but power is relative.

In other words, in both Chinese and Dutch, look seems to be adopted mainly to draw attention to one or more factual states of affairs, rather than to expressions of emotion, orders, questions or subjective evaluations. In the case of Chinese, this may also underpin rapport management and harmonious exchange among interlocutors (e.g. Spencer-Oatey 2005), which are distinctively prominent categories in Chinese commissives (Tantucci and Wang 2018, 2020a). In particular, information transmission in Chinese is often characterized by overt markers of expected agreement (cf. the use of 吧 ba as a PM in Tantucci 2017b). The pragmatic function of 你看 nǐkàn can be seen as an another example of this feature of Chinese, as it is added as an extra-propositional surplus of meaning, distinctively aimed at soliciting the interlocutor’s agreement upon the speaker’s assertion, paraphrasable in the form of ‘as you will agree, P’.

Compared to Dutch in particular, Italian and English have a much more even and diversified distribution of speech acts in Figure 3. What is characteristic of Italian, as well as Chinese (but more marginal in the two Germanic languages), is the concurrence of look with expressives: the expressive circles in Figure 3 are relatively small for Italian and Chinese but nearly non-existent for Dutch and English. In Italian, as (29) illustrates, the parenthetical is distinctively employed as an expression of phatic disclosure, with which the speaker gets the addressee’s attention to mark their attempt to share their emotions. In fact, under the assumption that such speech acts are more typical of conversations than of interviews (perhaps especially in Italian), it may be this association of guarda with expressives that accounts for its much higher frequency in conversations.

The use of look with expressives in Chinese is similar. The latter, in addition to expressing the act of sharing emotions from the speaker in the direction of the hearer, it additionally implies some sort of expected sympathy in return, e.g. from the hearer to the speaker (Tantucci and Wang 2020b), as in (30). Emotional disclosure here is therefore inherently geared towards rapport enhancement and bidirectional harmonious interaction.

Note also that such links with expressives are probably not completely unexcepted. look in its original meaning is already often used by speakers to point out some visible object or event that they simultaneously wish to convey their feelings about. It is no coincidence that English, for instance, has a semi-formulaic phrase look at that! expressing wonder (Van Olmen 2011: 409–410; see also Sánchez López 2017 on mirative mira; Badan 2021 on guarda as a marker of surprise).

There are two features typical of English. The first one is the tendency to employ look alongside directive speech acts, as in (31), where the mother draws on look not only to get her child’s attention but also to strengthen the command (together with the additions of the imperative subject you and the threat I won’t tell you again). As their very small directive circles in Figure 3 show, such usage is virtually absent in Dutch and Chinese and rather insignificant in Italian too. A potential reason for this fact is that, already in its literal sense, parenthetical look does not combine well with directives. More specifically, it is hard – though, admittedly, not impossible – to see how/why an appeal to look at something would be directly adjacent to yet another appeal to do something else. This original incongruity would then persist in Chinese, Dutch and Italian, even when look is pragmatic. In English, by contrast, it would have been lost and look would arguably be more bleached – i.e. more of its original function has disappeared – compared to literal ‘look!’ and to 你看 nǐkàn, kijk and guarda.^[17]

A second peculiarity of English is that the speech acts concurring with look are often used in isolation. In Figure 3, the language has the largest “absent” circle cross-linguistically (followed by Italian), which entails that a first act does not culminate in any second act (see Section 2.2.6). This tendency of English is also clear from the relatively thick arrows going from “asser” and “dir” to “absent”, while, in the other languages, the arrows pointing at the latter circle are generally very thin. An indicative example of the phenomenon is (32).

As reported speech is analyzed as a separate turn here (see Section 2.2.3), the assertive starting with and the information is not considered a speech act that follows the quoted directive introduced by look and the example needs to be classified as having no second speech act. Crucially, it is actually such cases, discussed in more detail in the next section, that account for the large “absent” circle for English.

3.4 Quotation, genre and vocative

As per Figure 2, the parameter that plays the second most important role in predicting language diversity is quotation, i.e. whether or not look serves to mark – and, more specifically, introduce some previous first- or third-person direct speech, like in (28) and (13) respectively. To fully evaluate this parameter’s impact, we fitted a mixed effects multinomial logistic regression (see Baayen 2008; Jaeger 2008). Mixed effects logistic regression helps to model binary outcome variables with two types of factors that are mixed in this kind of analysis: fixed factors and random factors. Random factors are used to systematically exclude variation that can be deemed random or unpredictable and thus only indirectly affect the response variable. Separating the effects of random factors allows the analyst to assess more reliably the effects of the remaining fixed factors, meaning that results can be generalized beyond the data set at issue with higher confidence.

After searching for possible interactions and fitting the best predictors through backward stepwise selection (see Levshina 2015: 266), we fitted a Glmer model with: (i) the presence of quotation as the outcome variable; (ii) genre, language and the presence of vocatives as fixed effects; and (iii) the interaction ID (i.e. the tag referring to the speech event) in which look was used as a random effect. The present model is the one that performed best for the cross-linguistic prediction of quotatives’ use. The results are presented in Table 2.

We can tell from Table 2 that the presence of quotations concurring with look is significantly correlated with interviews and discussions (at the reference level), in contrast to spontaneous conversations (β = −1.11, Z = 3.39, p = 3.46e−07). This fact may suggest that interviews and discussions are characterized by a markedly narrative stance and a comparatively more prominent orientation towards entertainment or information transmission that would be relevant to a wider audience. The exchange in (33) is a good illustration: the interviewer A asks a short question and the interviewee B gives a 400-word answer that recounts some past event and includes two quotations (one of which is presented here).

This tendency may arguably be connected with the extended intersubjective awareness (see Tantucci 2013, 2021) that some audience will be watching or hearing what is being said by someone who is being interviewed. In fact, providing anecdotal evidence is an important strategy in a diverse range of interviews (e.g. Moore and Stilgoe 2009), which can either apply to the perlocutionary effects of informing or entertaining. Quite differently, in the contextual setting of general conversation, a larger audience of viewers/hearers is not a constitutive element of the speech event. This fact entails the use of linguistic strategies that are less centered on extended intersubjectivity and the awareness of a generic social persona, but rather comparatively more focused on what is at issue for a (or some) specific interlocutor(s).

Table 2 also reveals clear cross-linguistic variation. To be precise (with genre and use of vocatives being controlled for) English is the language where look appears most frequently in quotations, differing most significantly in this respect from its Germanic cousin Dutch (β = −3.33, Z = −7.5, p = 1.30e−13). Chinese (β = −1.93, Z = −6.11, p = 9.77e−10) and Italian (β = −0.94, Z = −3.98, p = 7.06e−05) negatively correlate with quotations too, in contrast to English (the size of the difference is smallest with Italian, though). These results support previous observations about the exceptional status of English compared to Dutch in terms of look in quotations (e.g. Van Olmen 2010a), which can clearly be extended to other languages. In fact, in interviews and discussions, look is more common in reported than in regular speech (two thirds of the cases).^[18] In the conversations, this context of usage accounts for one fifth of the attestations.

A question that one could raise now is whether English look has developed into a quotative marker. The answer is probably negative, as there exist – at least to our knowledge – no languages with quotatives that have developed out of look. This view does not mean, however, that look should necessarily be regarded as an “actual” part of the quotation. As Tannen (1986: 324) among others argues, much of reported speech is “constructed” by the reporting speaker – rather than an exact repetition of the original utterances – to add a sense of immediacy and engage the addressee in its interpretation. In constructed dialogues, the reporting speaker also animates (Goffman 1981) the reported speaker’s utterance, demonstrating not only what the latter said but also how they said it (Clark and Gerrig 1990). Pragmatic markers are obviously very convenient tools to achieve those aims. What may make look in English particularly useful is its comparatively strong sense of commitment: what follows look is an act that the speaker really believes in (e.g. Saxton 1992: 55–58; Van Olmen 2010b: 80). When a reporting speaker starts a quotation with this marker, they signal to their addressee (and not necessarily the original speaker to the original addressee) that the ensuing utterance is a/the crucial part of their story, which is perhaps also why it is reported directly in the first place. In (32), for instance, the speaker wants to make the general point that it is just easier to ask realtors to send you their information via the post. To do so, he talks about one of his own experiences and, importantly, his actual request is a quotation and is highlighted by look. The frequency of look’s concurrence with quotatives does suggest, however, that it has undergone some degree of chunking (Bybee 2010; Tantucci 2017a), phonetic reduction and entrenchment (Tantucci and Di Cristofaro 2019), as in the English formula and-I-said-look, which could be arguably considered the internal constituent of a more schematic quotation-related construction [and Subj said look, P]. What may be interesting in this respect is that, in the BNC files for which we could access the audio, said and look often form an intonation unit that is followed by a break and subsequently the quotation.

A final observation based on Table 2 is the positive correlation of look in quotations to concur with vocatives (β = 1.81, Z = 4, p = 6.21e−05). In other words, vocatives may not be predictors of any differences in the parenthetical between languages, but they do appear more frequently with look in than outside quotations. This finding should not come as a surprise. A quotative move often hinges on the attempted reproduction of a past interaction with interlocutors who are usually physically absent from the here-and-now of the conversation. What a vocative does in such a context is increase the deictic accessibility, for the current interlocutors, to the participants of the original exchange. Even in interviewee B’s imagined past interaction in (33), for example, lads can be said to make the hypothetical situation more tangible for interviewer A, particularly through its informal character.

3.5 Clause and turn position and visual perception

In this section, we explore the cross-linguistic variation of look appearing in different positions across clauses and whole turns-at-talk. We start by fitting a conditional inference tree (see Section 3.2) with the aim of disentangling the concurring relationship between visual perception and whether look occurs in clause-initial, -medial or -final position. Conditional inference trees result from binomial “decisions” based on statistical significance that inform the outcome of the response variable, which in our case was language diversity. Conditional inference trees are visually more accessible than random forests and complement the latter in providing a good illustration of the conditionals that lead to the outcome variable (see also Gries 2020; Levshina 2021). The present variables’ selection was centered on the hypothesis that the flexibility of look’s positioning within the clause may vary cross-linguistically depending on the degree of reanalysis of the construction, hence the degree to which look would encode non-perceptual meaning. Consider Figure 4.

Figure 4:

Conditional inference tree for clause position, visual perception and language.

This tree is made of five nodes and two splits. The higher the split in the figure, the stronger the statistical significance of that conditional decision, hence the smaller the p-value. A useful way to approach conditional inference trees is to think of them as a series of paths, each leading to a final distribution of outcomes. Each path is made up of a number of stages (nodes) that correspond to statistical decisions that are necessary for the significant distribution of the bars at the bottom of the plot. As we can observe, the most significant partition here is in node 1 (p < 0.001), i.e. between look occurring in clause-final versus clause-initial/medial position. In fact, when look does not occur clause-finally, it shows a very even distribution across languages, with almost absent cross-linguistic variation, as shown by the bars in node 2, at the bottom left corner of the plot, each ranging from 20 to 30% of the distribution under those conditions. Conversely, when clause-final position is taken into account, three important observations are in order. First, Chinese and Dutch simply display no uses of look displaced at the right hand-side of the clause. Second, Italian is the only language that allows look to occur in clause-final position with a bleached meaning not expressing visual perception (see node 4 at the center of the plot). Third, look occurring in clause final position expressing a visual perceptual meaning appears to be a distinctive feature of English, and only marginally of Italian (see node 5 at the bottom right corner of the plot). Together, these last two observations mean that, while English and Italian have the possibility of a clause-final look in common, look only occurs in that position with its original meaning – like in (27), repeated here as (33) – but guarda typically fulfils a less compositional and more pragmatic function there – like in (11), repeated as (34).

With reference to the plot above, clause-final literal look in (33) illustrates the significant transition from node 1 to node 3 to, finally, node 5 (1 > 3 > 5). On the other hand, clause-final pragmatic guarda in (34) underpins the transition from node 1 to node 3 to, finally, node 4 (1 > 3 > 4).

The conditional inference tree for turn position in Figure 5 shows that a very similar pattern is at play in the cross-linguistic variation of look across turns.

Figure 5:

Conditional inference tree for turn position, visual perception and language.

There is a higher number of nodes in Figure 5 than in Figure 4, which entails higher illocutional complexity (Tantucci and Wang 2018) underpinning the positioning of look throughout turns-at-talk clauses. What is still remarkable is the absence of usages also in turn-final position in Chinese and Dutch, with Italian again distinctively favoring a less compositional, non-perceptual meaning and English exclusively expressing visual perception in that position. This is shown in nodes 5 and 6. Perception is indeed an important predictor of cross-linguistic diversity in turn-final (node 4) and turn-medial position (node 7) but not at the beginning of turns (node 2). This, in turn, crucially indicates that the employment of look in turn-initial position is more schematic, as it is no more subject to semantic differentiation (cf. De Smet et al. 2018) determined by that specific positioning within utterances. This fact is understandable, as the general function of look across languages appears to be the one of attention-getting, which can indeed reasonably be expected to occur at the beginning of a turn. Confirmation comes from the observation that the concurrence of look with turn-initial position has the highest frequency of all (node 3, n = 359).