Aspect construal in Mandarin: a usage-based constructionist perspective on LE

1.1 The problem of le

What makes le such a big problem? First, there seem to be two homophonous and homographic morphemes associated with le. One is typically described as a verbal suffix marking perfective aspect. The other is considered a final particle conveying state change.^[3] These are respectively illustrated in (1a) and (1b), retrieved from the BCC corpus.

Second, the temporal reference of le depends on contextual information (Ren 2008; Smith and Erbaugh 2005). While the default time reference inferred from (1a) is past time, other time references are possible, as in (2) where the event is located in a future time.

In (3), the situations described in both clauses ending in the final le are located in the past.

Third, the particle le can be used with adjectives serving as predicates, as in (4).^[4]

The adjective hăo in (4a) describes a new physical state and in (4b) it describes the speaker’s subjective state in response to the new physical state. Similarly, the verbal suffix le is not always a verbal suffix. It can also attach to an adjective, as in (5b) where it follows the adjective dà ‘large’ and precedes a complement of quantity.

Finally, le can be used in both positions in the same sentence, as in the second clause of (6) where the perfective le precedes a noun of temporal duration and the utterance ends in the particle le. The entire sentence is interpreted as having a perfect continuous reading with an inference that the event is perceived as excessive (Chappell 1986).

Scholarly treatments of le fall into two general models, a two-morpheme model and a single-morpheme model, which we now discuss in turn.

1.2 Two-morpheme treatments

Most accounts of le in Chinese linguistics espouse a two-morpheme view of le (e.g., Chao 1968; Li and Thompson 1981; Lü 1980; Sun 2006; Xiao and McEnery 2004; Zhu 1982; Zhu 2019; Zhu and Gao 2013). At first glance, this analysis distinguishes the functions of le based on its sentential position, namely clause-internal versus clause-final. This general distinction, however, is not based on a theoretical commitment to form-function pairing that consistently associates the constructional properties of le with its functions. Despite what appears to be a constructionist approach, the two-morpheme model is more concerned with the lexical aspects of the predicates than with the larger constructional context of which le is a part.

As examples (1)–(6) show, there is no one single lexical class that is uniquely associated with either the clause-internal or the clause-final le. A brief summary of how the major treatments of le handle the variability of the collocating lexical categories will highlight the inconsistencies and ambiguities that are inevitable on a morpheme-based approach. Wang (1965), Chao (1968), Lü (1980), and Zhu (1982) all treated the verbal suffix le as perfective aspect in the sense of “completed action”, which does not account for the adjectival uses such as (5b). Lü (1980: 355) recognized this problem and acknowledged that the adjectival uses denote “deviation of a property from a standard”. However, it is unclear how this meaning can be reconciled with completion.

Li and Thompson (1981), Chappell (1986, 1988, and Sun (2006) emphasized that the perfective aspect should not be understood in terms of completion of an action, but in the sense that a situation is viewed in its entirety, or as a bounded event. The viewpoint-oriented perspective made it possible to accommodate adjectival uses. That is, the property described by dà ‘large’ in (5b) is viewed as bounded by virtue of being quantified. As for the final le, Li and Thompson (1981: 244) coined the term “Currently Relevant State” (CRS) as the general semantics with various contextual entailments. These include both propositional meanings, e.g., “changed state” and “what will happen next”, and interactional functions such as correcting a wrong assumption and signaling speaker contribution. However, while CRS can explain many uses of the clause-final le, there are exceptions that are analyzed as signaling boundedness, an inconsistency to which we shall return shortly.

Zhu and Gao (2013) and Zhu (2019) treated the clause-internal le as a perfective aspect marker and the clause-final le as a transformative aspect marker, stating that the former signifies the occurrence or completion of an action and the latter indicates the end of a state. Faced with the same problem of explaining the adjectival uses in terms of completion of an action, the authors treated these uses instead as part of the final le category. Their rationale is that the latter normally admits stative lexical input.

Liu (1988) challenged the notion that the perfective le denotes completion. He proposed instead actualization of an event to accommodate stative predicates with the perfective le, which describe situations that in real-world scenarios do not end after their actualization. Prioritizing a truth-conditional view of verb lexical semantics over construal at the constructional level, Liu insisted that such situations be considered imperfective. The actualization analysis found support among truth-conditionally minded grammarians, leading to the proposal that the atelic uses with the internal le denote imperfective aspect (e.g., Jin 1998, 2002; Lin 2003). As we shall argue in this article, the actualization analysis ignores the viewpoint from which different situation types described by different verbs are perceived through the shared lens of a single constructional pattern paired with the construal of that viewpoint. In doing so, it fundamentally misunderstands linguistic representation.

Xiao and McEnery (2004) drew on Smith’s (1983, 1991) distinction between situation and viewpoint aspects and provided a two-component approach in an attempt to take care of both situation or lexical aspect and viewpoint or grammatical aspect. They proposed that the perfective le signals the “actual” viewpoint in terms of an actualized single whole, “irrespective of the final endpoint of the situation” (p. 17). Their treatment allows the actual viewpoint to be applied to atelic situations, unlike the actualization model that treated these as imperfective. This is a significant improvement. However, this upgrade comes with a tradeoff. Xiao and McEnery rejected the notion of bounding as part of the actual viewpoint, apparently worried that it would undermine the ability of their proposal to accommodate the atelic uses that they consider “inherently unbounded” (p. 17). Critically, this dilemma arose because boundedness was treated as related solely to situation aspect but not to viewpoint aspect. Thus, although their data showed that le is “sensitive to the endpoint of a situation” (p. 19), the authors resisted interpretating that data as evidence of boundedness in the actual viewpoint. Apart from the empirical challenge of compromised data interpretation, it is conceptually problematic to disregard event boundaries from the perspective of cognitive psychology. There is converging evidence that event perception is analogous to object perception whereby segmentation at the location of event boundaries is “a core, domain-general mechanism of cognitive control” with consequences for long-term memory and learning (Zacks et al. 2001a, 2001b; Zacks et al. 2007; Zacks and Tversky 2001: 276). Thus, it is implausible to perceive events without paying attention to the location of event boundaries.

When it comes to the final particle le, single verb predicates pose a huge analytical problem. When le follows a single verb and occurs in the clause-final position, it apparently has the dual role of a verbal suffix and a final particle. So, what exactly is it? Chao (1968) suggested that this le is a fusion of both the perfective le and the particle le. Lü (1980) and Zhu (1982) contended that it can be a verbal suffix, a final particle, or the fusion of both. Similarly, Zhu and Gao (2013: 425) treated it as either perfective (completion of action) or transformative (end of state), though they added that it “expresses the meanings of both aspects”. Li and Thompson (1981: 195) proposed that this kind of le can have three different readings depending on the meaning of the preceding verb: a perfective marker, a CRS marker, or both. They maintained that “some verbs represent specific, bounded events by virtue of their meaning”, such as sĭ ‘die/dead’, and that the clause-final le denotes perfective aspect rather than CRS when preceded by such verbs. That is, it should be treated in the same way as the clause-internal le describing bounded events. Xiao and McEnery (2004) shared this view and glossed this kind of le as DBL “double”, allowing for both interpretations. Interestingly, Li and Thompson chose the example in (7) but not the one in (8) to illustrate their argument.

Note that qùnián ‘last year’ is added in (7) to explicitly locate the situation described by sĭ ‘die/dead’ in the past to support the inference of a bounded reading. In a non-past context, e.g., (8), a stative reading, namely “S/he is dead (not alive anymore)” would be natural. This example suggests that sĭ ‘die/dead’ possesses what Dahl (1985: 26) calls “aspectual potential” in the sense of being capable of alternative construals. It can be construed as a permanent state, it can also be construed as a bounded event with the extra help of a past adverbial, in the same way the English verb know has the aspectual potential to be construed as either state or achievement, as in (9)–(10), respectively, discussed in Croft (2012: 38, ex. 12, 15):

Croft (2012: 38) pointed out the way know is interpreted in a sentence depends on the “tense-aspect constructions” in which it occurs. In the Simple Present tense-aspect construction in (9) know denotes state. In the Past tense-aspect construction in (10), it can obtain an achievement interpretation but needs the support of the punctual adverbial suddenly. Critically, grammatical aspect at the construction level trumps lexical aspect at the verb level, not the other way round. Confusion about multiple readings only arises because these two levels are treated as equal, as we see in the literature on le.

1.3 Single-morpheme treatments

Chappell (1988) argued against the two-morpheme approach. She maintained that the verbal enclitic le and the final particle le have a synchronically identical meaning in the sense that both signal boundedness but have different discourse functions. Specifically, the verb-final le bounds the main event as a perfective marker and is structurally associated with [V-le-NP]. On the other hand, the sentence-final le, structurally most clearly identifiable in [V-NP-le], signals inception of a state and is used to mark episode boundaries at the discourse level. However, the majority of Chappell’s data on the final le has the [V-le] pattern without an intervening NP between the verb and le. Chappell considered this structure indistinguishable from the perfective le following a single verb, thus facing the same problem of ambiguity that troubles the two-morpheme model discussed in the foregoing section. Chappell (1988: 119) recognized that the inceptive final le has “the whole clause in its scope of modification” in contrast to the perfective le being “constrained by the inherent aspectual meaning of the verb it modified”. Unfortunately, this insight did not lead to what would be a logical conclusion, namely that the inceptive le prevails over the perfective le in [V-le]. Instead, Chappell (1988: 120) maintained that the semantic distinction is neutralized in the sense that the perfective le and the inceptive le are “one and the same morpheme, identical in semantic content”, and that its function is to be contextually but not structurally distinguished. Because the neutralization proposal does not consistently associate function with form, we believe it suffers the same weaknesses as the two-morpheme studies previously reviewed.

Shi (1990) also criticized the two-morpheme model, arguing that there is just one le and that, whether verbal or sentential, it signals relative anteriority rather than perfectivity and inchoativity. When le goes with a bounded event, it gives rise to the inference that the event is completed by virtue of being relatively anterior; when it goes with an unbounded event, the resulting inference is inchoativity. However, Shi did not provide independent criteria for boundedness. He took the bounding features of the complement NP, e.g., quantified object, as the criterium in some cases. Whenever this criterium contradicted the assumed verb lexical aspect, he relied on the latter. For example, (11) and (12) both have a quantified object, but he treated (11) as bounded and (12) as unbounded (Shi 1990: 107, Ex. (23a) and (24a).

Shi said that (12) is unbounded because the main verb yŏu ‘have’ is a state verb. Essentially, this verb-centered treatment of aspect suffers the same inconsistencies as the other morpheme-based approaches described previously.

The above overview shows that the existing accounts of le, whether two-morpheme or single-morpheme oriented, are rife with arbitrariness and ambiguities. Despite the constructionist intuition, especially on the two-morpheme model, that in general the two different positions of le are conventionally associated with two different ways of event representation, the analyses are morpheme-based and verb-centered. As such, they get bogged down in the local lexical aspects of the predicates the variability of which gets in the way of a unitary theoretical account that explains the data in its totality. Specifically, when the presumed aspectual type of a predicate is incoherent with the presumed role of le, the constructionist intuition is invariably abandoned. Croft (2012: 31) pointed out, “Aspect is manifested both grammatically and lexically”. However, the manifestation of grammatical aspect cannot be represented without some theory that recognizes that grammatical constructions have meanings independent of lexical input. In the absence of such a theory, a unitary account of le remains out of reach.

Finally, the confusion surrounding le is as much methodological as it is theoretical. To our knowledge, few of the claims reviewed above have been empirically tested. With the exceptions of Chappell (1988) using a small sample of narrative data and Xiao and McEnery (2004) employing a Chinese newspaper corpus of limited size, previous studies relied on introspective data and an intuition-based analysis. In our opinion, the methodological limitations have impeded theoretical advancements in this area of research. We believe a unitary account of le requires a theoretical leap from a morpheme-based and verb-centered analysis to a constructionist framework as well as a methodological leap toward a greater empirical commitment. It is the goal of this study to take these leaps.

3.1 Data and methods

The corpus data for this study was retrieved from BCC Corpus developed by the Beijing Language and Culture University Corpus Center (Xun et al. 2016). BCC is a balanced POS tagged corpus with a size of ten billion words collected from various genres including fictions, newspaper articles, popular science texts, microblogs etc. The search platform supports both keyword- and syntax-based retrievals. For example a search for ‘v 了 w’ (v=verb, w=punctuations) retrieves all tokens of the utterance final le following a verb. The system features a “statistics” function that automatically generates a frequency rank list of all the types of collocates that occupies an open slot of interest in the search syntax. For example, the statistics function ranks all the different verbs returned from the search for ‘v 了 w’ by their token frequency. For each search, BCC allowed users to download the first 1,000 results of types ranked by token frequency. All of our corpus data was retrieved from the multi-genre domain (多领域) of BCC.

For the Internal Le Construction, we first ran two exploratory searches for [v 了∼] and [a 了∼] where ‘∼’ stands for any word to identity possible subtypes of the NP following le. This process identified six subtypes of NP for the verb type, e.g., noun, classifier_noun, numeral_noun, numeral_classifier_noun, adjective_noun, and adjective_de_noun. The first five of these NP subtypes were also identified for the adjective type, but not the last one. We ran 11 syntax-based searches, as shown in the “search syntax” column of Table 2. A frequency rank list of the automatically identified types in the returned concordances was generated in BCC and downloaded into excel for annotation. Table 2 shows the raw frequencies of the retrieved concordances from the 11 searches.

Manual annotations were performed to identify false positives. For example, chū le gòngxiàn ‘out le contribution’ was retrieved from searching [v le n], though it is an incomplete phrase without a verb because BCC mistakes the resultative chū ‘out/exit’ for the main verb, which it is not in this context. So, we went back to BCC to conduct a follow up search for [v chū le gòngxiàn] to retrieve the missing verb. We then downloaded the new rank list to replace the false positive type in the master list of concordances of [v le n]. This process was performed on all identified false positive items such as this for all eleven searches. A combined rank list of the top 100 most frequent types of [V/A-le-NP] was generated.

For the Final le Construction, we first ran three exploratory syntax-based searches, shown in the top three rows of the “search syntax” column in Table 3. We then examined the three initial rank lists for noise including identical types with different punctuation marks and incomplete strings due to features of POS tagging in BCC that (1) obscure the distinction between a verb/adjective as predicate and one serving as a directional/resultative suffix and that (2) do not distinguish idiomatic combinations from regular compositions. For example, the [verb_le] search returned zŏu le ‘go away_le’ many tokens of which were a part of ná-zŏu le ‘take away_le’ or some other verbal strings where zŏu is a directional suffix meaning ‘away’. Similarly, the [adjective_le] search returned hăo le ‘good_le’ with many tokens being a part of tài hăo le ‘too good_le’. To sort out ambiguous cases like these, subsequent searches were conducted to identify subtypes, e.g., the bottom five items in the search syntax column of Table 3, to distinguish types that are full utterances from those that are part of larger strings. The results were further manually processed in excel to (1) combine identical types with different punctuations and to (2) retrieve any missing heads, as described in the previous paragraph. A combined rank list of the top 100 most frequent types of [V/A-le-NP], totaling 188,041 tokens, was generated in R Studio (Version 1.1.463) with the package tidyverse.

3.2 Results

Figure 1 visualizes the top 100 types of the Internal le Construction [V/A_le_NP].

Figure 1:

Top 100 types of [V/A_le_NP].

The data in Figure 1 shows a strong representation (83%) of accomplishments, including physical accomplishments, e.g., tàn le kŏu qì (sigh le CLF air) ‘heave a sigh’ and abstract accomplishments, e.g., zuò-chū le gòngxiàn (make le contribution) ‘make a contribution’ and chōngmăn le xìnxīn (infuse-full le confidence) ‘filled with confidence’. The rest describe achievements, e.g., chū le mén (exit le door) ‘go out the door’ and mí le lù (confuse le way) ‘lose orientation’. A third of the top 100 types have a numeral (yī ‘one’) and/or a classifier (CLF) in the complement NP whereby 75% of the classifiers are verbal classifiers that quantify actions and activities.

Figure 2 shows the top 100 types of the Final le Construction, which fall into three subtypes. The first (42%) conveys the inception of speaker intersubjectivity toward what is being evaluated. Of these 88% feature [tài A le] with the intensifier tài ‘too, very’ that expresses a heightened excitement about a situation that compels the listener’s attention. Through frequent uses, [tài A le] has clearly become a highly productive formula, as can be seen in its dominance among the high-frequency types. A second subtype (44%) describes the inception of a new eventuality, e.g., lèi le (tired le) ‘tired now’, zhăo-dào le (find le) ‘found now’. The use of ‘now’ in the glossing makes explicit the initial boundary or inception of the eventualities being described. Depending on the context, hăo le (good-le) can describe the inception of a physical state, e.g., ‘well now, recovered, back to normal now’ or convey the inception of an intersubjective state, e.g., ‘it’s alright now’. A third type (8%) features [yào V le] where the epistemic modal yào ‘going to’ announces the high probability of an event to inchoate in the imminent future, as in yào sĭ le ‘is going to die, is dying’, yào lái le ‘is going to come, is coming’, yào kū le ‘is about to cry, on the verge of tears’. The formulas wăn-ān le (night peace le) ‘good night’ and zàijiàn le (again see le) ‘goodbye’ initiate social rituals.

Figure 2:

Top 100 types of [XP_le].

Lexical aspectual collocates have been shown to influence the interpretation of event types (Croft 2012; Vendler 1967). A follow-up analysis was conducted to determine whether the two aspect constructions differ in their preference of such collocates. Specifically, we examined zhōngyú ‘finally’, yào ‘going to’, yĭjīng ‘already’, kāishĭ ‘begin’, and zăojiù ‘already’ that occur immediately before the verb in [V le N] and [V N le] in BCC. A Pearson’s chi-square test of independence on the association between construction and lexical collocates was significant (X² = 10948.156, df = 4, p < 0.0001). As Table 4 shows, [V N le] strongly prefers yào ‘going to’ and kāishĭ ‘begin’ and disprefers yĭjīng ‘already’ whereas [V le N] strongly prefers yĭjīng ‘already’ and disprefers yào ‘going to’ and kāishĭ ‘begin’. The results for zhōngyú ‘finally’ and zăojiù ‘already’ were not as strong.

Given that yào ‘going to’ denotes high probability or imminence of an event and kāishĭ ‘begin’ denotes the inchoation of an event, it is unsurprising that they are preferred by [V N le] which construes the initial boundary of event. The adverbs zhōngyú ‘finally’, yĭjīng ‘already’, and zăojiù ‘already’ convey viewpoints relative to the time of speech and speaker expectation. Zhōngyú ‘finally’ conveys a viewpoint from which the event is viewed as occurring later than expected. Yĭjīng ‘already’ conveys that the event in question has occurred by the moment of speech and is perceived as earlier than expected. The glossing of zăojiù as ‘already’ may suggest it is synonymous to yĭjīng ‘already’. However, zăojiù explicitly and more strongly conveys a viewpoint from which the event in question is seen as earlier than expected and is indifferent to the actually temporal boundary of the event with which yĭjīng ‘already’ is concerned. This explains why yĭjīng ‘already’ is strongly preferred by [V le N] which construes final boundary of event whereas zăojiù ‘already’ is compatible with both construals. Similarly, because zhōngyú ‘finally’ is more about the perception of deviation from an expectation than the actual temporal boundary of an event, it has the potential to go with either the onset or closing of a long awaited event. Hence its marginal effect, as indicated by the smaller adjusted residual, makes sense.

3.3 Event sequencing

Li and Thompson (1981) and Sun (2006) observed that an important indication of boundedness is event sequencing in discourse whereby the event marked by the perfective le precedes another event the occurrence of which depends on the perceived closure of the earlier event. Our hypothesis of the two aspect constructions representing two aspectual construals takes this observation a step further and predicts that the Internal le Construction that construes closing of event is more likely than the Final le Construction to participate in sequencing as the first event regardless of its actual closure in reality. Consider (13) from BCC as an example.

Here the event ‘have a child’ takes the internal le in (13a). As such it is given a final boundary after which another situation begins, namely temperament change, described with the final le in (13b). Note that it is equally felicitous to describe ‘have a child’ with the final le to mark the inception of parenthood. In fact, there are 355 uses of yŏu háizi le (have child le) in BCC. However, none collocates with zhīhòu ‘thereafter’ or yĭhŏu ‘after’, which, as (13a) shows, explicitly separates the first episode from the subsequent situation. By contrast, of 1,208 tokens of yŏu le háizi (have le child), 29 collocate with zhīhòu ‘thereafter’ and 62 with yĭhŏu ‘after’. The preference of the internal le clause is due to the construal of final boundary whether or not such a boundary makes any sense in reality.

To further test our hypothesis, we examined a sample of 100 concordances each of [v_le_n] and [v_n_le] from BCC within a search window of 30 characters on each side of the target to include usage context. The samples were independently coded for sequencing (YES, NO) by two researchers based on the criteria: if the -le clause precedes and stands in any of three relationships (temporal, conditional, causal) to another clause such that the occurrence of the event in this second clause presupposes the event in the le clause, a YES coding was obtained, otherwise NO. Cohen’s Kappa was calculated for intercoder reliability (k = 0.898 for [v_le_n] and k = 0.786 for [v_n_le]). Intercoder discrepancies occurred with the coding of 11 items (5 in [v_le_n] and 6 in [v_n_le]). These items were coded by a third coder not aware of the other two’s results. The discrepancies were resolved in favor of the majority coding decisions.

A chi-square of independence on the association between construction and sequencing was statistically significant (X² = 27.00, df = 1, p < 0.0001). As shown in Figure 3, [v_le_n] is significantly more frequently involved in sequencing as an antecedent than [v_n_le], confirming our prediction based on the functions of the two aspect constructions.

Figure 3:

Frequencies of sequencing by construction.

To summarize the corpus analysis, the Internal le Construction [V/A_le_NP] is overwhelmingly used to describe accomplishments with a propensity for event quantification, bounding adverbials, and event sequencing. The Final le Construction [XP_le] is used to convey the inception of an event, be it at the intersubjective, the perceptual, or the social level, prefers inchoative adverbials, but is incompatible with event sequencing. These distributions are consistent with their respective functions of aspectual construal: namely the construal of final boundary and the construal of initial boundary.

4.1 Sentence sorting tasks

4.2 Similarity judgment task I

Li and Thompson (1981: 195) argued that when the final le collocates with a verb with an “inherent bounded meaning”, it does not denote CRS, but boundedness. This hypothesis would predict that native speakers, when given two choices to match this kind of use in a similarity judgment – one with the internal le that describes a bounded event and the other with the final le that describes CRS, would choose the former and reject the latter. Other scholars (e.g., Chao 1968; Lü 1980; Sun 2006; Xiao and McEnery 2004; Zhu and Gao 2013) claimed that le used in this way is a fusion of both the perfective le and the particle le. This claim would predict that participants are equally likely to choose one or the other sentence as similar to the target sentence. Our constructionist hypothesis predicts that participants will make a family resemblance judgment by picking the final le choice as similar to the final le target.

4.3 Similarity judgment task II

Recall the disagreement in the literature on the status of the internal le used with adjectives that precede an NP of quantity. Some grouped it under the perfective aspect (e.g., Chao 1968; Lü 1980; Xiao and McEnery 2004) and others treated it as equivalent to the final le describing state change (e.g., Zhu and Gao 2013). A third view held that the internal le can denote both perfective and imperfective aspects depending on the type of verbs (Jin 1998; Lin 2003, 2017). For this last school, the verbs that trigger an imperfective reading of the internal le include also activity verbs in addition to adjectives. The sorting task in 4.1 has confirmed that participants overwhelmingly grouped sentences by construction by resisting verb meaning. The first similarity judgment task targeted verbs with a final le. It remains unclear from a statistical perspective if native speakers perceive sentences with an internal le suffixed to adjectives and activity verbs as similar to (1) internal le sentences that denote boundedness, or to (2) final le sentences that signal a new state. Similarity judgment task II was designed to test these competing hypotheses.