1. Introduction
Linguists have long been discussing the concept that intransitive verbs can be divided into at least two verb classes, according to the characteristics exhibited by verb meanings and associated grammatical structures (Perlmutter, Reference Perlmutter and Jaeger1978; Burzio, Reference Burzio1981, Reference Burzio1986). Unergatives are intransitive verbs whose subjects show grammatical and semantic similarities to the subjects of transitive verbs. On the other hand, unaccusatives are those whose subjects bear these similarities to the objects of transitive verbs. This division, known as unaccusativity or split intransitivity, forms a syntax-semantics nexus to bridge the argument structure and verbal semantics.
Several sentence typesFootnote 1 can be used to identify the verb class. Perhaps the most famous of these is auxiliary selection, appearing in many European languages (Burzio, Reference Burzio1986; Cennamo & Sorace, Reference Cennamo, Sorace and Aranovich2007; Hoekstra, Reference Hoekstra1984; Levin & Rappaport-Hovav, Reference Levin and Rappaport-Hovav1995; Sorace, Reference Sorace2000; Van Valin, Reference Van Valin1990). In example (1), the Italian verb telefonare “telephone” takes the auxiliary verb “have” to occur with the subject, which is characteristic of the subject of a transitive verb. Meanwhile, the verb arrivare “arrive” takes “be” to occur with the subject instead, and this co-occurrence is similar to the object of a transitive verb and the verb. The distinct auxiliary selection between “have” for unergatives and “be” for unaccusatives serves as a way to syntactically distinguish the two verb classes, and the sentence type is termed a “diagnostic.” When an intransitive verb occurs with “to have,” it is classified as an unergative, and when an intransitive verb occurs with “to be,” it is classified as an unaccusative. It is through these diagnostics that linguists determine the verb classes of intransitive verbs.
In English, the grammatical suffix “-er,” serves as such a diagnostic unergatives and transitive verbs are typically compatible with this suffix. Examples in (2) show that unergatives “walk,” “talk,” and “work” can form nouns with “-er,” resulting in “walker,” “talker,” and “worker” and the suffix is also possible for transitive verbs “kill,” “reform,” and “love” forming “killer,” “reformer,” and “lover.” On the other hand, the suffix “-er” cannot be grammatically attached to unaccusatives, as seen in the ungrammatical nouns “*faller” from “fall,” “*escaper” from “escape” and “*arriver” from “arrive.”
In Mandarin Chinese, a common diagnostic is to test if the subject of an intransitive verb can appear postverbally as in (3). If it can, it is an unaccusative; otherwise, it is an unergative (Huang, Reference Huang, Reuland and Ter Meulen1987; Paul et al., Reference Paul, Lu and Lee2020; Yang, Reference Yang1999; Yu, Reference Yu, Camacho and Choueiri1995).
The subject of an unaccusative is similar to the object of a transitive verb, which is base-generated initially in the internal-argument position, and later moves to the external-argument position. This allows the subjects of unaccusatives to stay postverbally. The subject of an unergative does not have this option and is base-generated in the external-argument position initially, similar to the subject of a transitive verb (Burzio, Reference Burzio1981, Reference Burzio1986). Perlmutter (Reference Perlmutter and Jaeger1978) explicates the unergative-unaccusative distinction in his Unaccusativity Hypothesis and discusses the semantics-syntax correlation between verbal semantics and their grammatical representations. Verbs such as “fall” and “escape” include an endpoint in the events and take themes as their subjects. Verbs with such telic semantics and patientivity are likely to be unaccusatives. In contrast, verbs such as “walk” and “talk” do not entail an endpoint in their events and require agentivity in their subjects. Such atelic semantics and agentivity are likely to form unergatives.
Subsequent research (e.g., Rappaport-Hovav & Levin, Reference Rappaport-Hovav, Levin, Butt and Geuder1998; Sorace, Reference Sorace2000, Reference Sorace, Alexiadou, Everaert and Anagnostopoulou2004) has found that more articulated semantic subclasses can explain the variations of unaccusativity. For instance, Sorace (Reference Sorace2000)’s Auxiliary Selection Hierarchy (ASH) is a scale that assigns different likelihoods of mapping between the semantics of verbs and their corresponding auxiliaries and is therefore interpreted to explain the relationship with corresponding verb classes (unergative or unaccusative, or both). It positions core unaccusatives on one end and core unergatives on the other. The verbs denoting change-of-location, with the most stable telicity, are considered the most likely to select “be” and classified certainly as unaccusatives, while controlled non-motional process verbs, with the most atelicity, are likely to select “have” and classified certainly as unergatives. The ASH also accommodates other semantic subclasses that fill in the space between these two main subclasses, which exhibit gradient levels of mapping between verb meanings and their possible verb classes.
This hierarchy accommodates the fluidity of verbs and captures their varying syntactic manifestation in different sentences. Verb semantics that can be mapped onto both unergative and unaccusative, or those that manifest different behaviours (unergative or unaccusative) in different sentences, are situated between the two extremes. For instance, “move” in (5) is a controlled motional process in the middle of ASH, and its verb classes can therefore have more variations in different sentences, as shown in (5a) and (5b).
Building on this understanding, the process by which children acquire the verb classes of unergative and unaccusative becomes an intriguing area of study. In the domain of language acquisition, previous studies have found that children acquire the distinction between unergative and unaccusative from a very young age (e.g., Costa & Friedmann, Reference Costa, Friedmann, Everaert, Matin and Siloni2012; Friedmann, Reference Friedmann2007; Lorusso et al., Reference Lorusso, Caprin, Guasti, Brugos, Clark-Cotton and Ha2005; Snyder et al., Reference Snyder, Hyams, Crisma and Clark1995). Some studies, however, have found that the acquisition of unaccusativity is delayed (Babyonyshev et al., Reference Babyonyshev, Ganger, Pesetsky and Wexler2001) or delayed for certain semantic classes (Lin & Deen, Reference Lin and Deen2021). Other research (Lu, Reference Lu2019; Randall et al., Reference Randall, van Hout, Weissenborn, Baayen, Alexiadou, Anagnostopoulou and Everaert2004) has focused on how children acquire verb classes, showing that children use verb meanings to facilitate a semantics-syntax link in grasping unaccusativity. However, an additional facet that can significantly contribute to this process is the distributional information present in language input. In this study, we investigate the role of distributional information in children’s acquisition of verb classes of unaccusative-unergative. Our analysis reveals that the distributional patterns of unergatives and unaccusatives across various syntactic structures exhibit significant distinctions, making these distributional cues potentially informative for the acquisition of verb classes. We further present empirical evidence demonstrating that young Mandarin-speaking children can use these distributional cues to classify verbs. Our findings suggest that such distributional information may serve as a critical factor in the categorization of novel verbs, complementing other cues such as semantics. Based on these findings, we posit that incorporating distributional information into the theoretical framework might elucidate the atypical behaviours observed with certain Mandarin verbs, and may yield a more comprehensive account of cross-linguistic variation in split intransitivity.
1.1. Distributional cues and unaccusativity in language acquisition
Previous studies on language acquisition have found that children are sensitive to various distributional cues, such as prosodic, transitional, grammatical, and semantic cues (e.g., Ambridge et al., Reference Ambridge, Pine, Rowland and Young2008; Brooks et al., Reference Brooks, Braine, Catalano, Brody and Sudhalter1993; Cartwright & Brent, Reference Cartwright and Brent1997; Gleitman, Reference Gleitman1990; Mintz et al., Reference Mintz, Newport and Bever2002; Pinker, Reference Pinker1984, a.o.). Children attend to the frequency, co-occurrence, and contextual patterns of words and phrases in language input, to uncover underlying grammatical rules and semantic relationships in a language. Some research has shown that argument positions of verbs in language input influence how children use verbs during production (Olguin & Tomasello, Reference Olguin and Tomasello1993); transitional probabilities can serve as an important cue for identifying word boundaries for learners (Saffran et al., Reference Saffran, Newport and Aslin1996), and distributional information plays an important role in acquiring various grammatical categories (Redington et al., Reference Redington, Chater and Finch1998). It is widely accepted that distributional cues in language input are useful for learning linguistic knowledge.
Regarding unaccusativity, previous studies focus on the age at which children have acquired the distinction between unergatives and unaccusatives (e.g., Costa & Friedmann, Reference Costa, Friedmann, Everaert, Matin and Siloni2012; Friedmann, Reference Friedmann2007; Lorusso et al., Reference Lorusso, Caprin, Guasti, Brugos, Clark-Cotton and Ha2005; Snyder et al., Reference Snyder, Hyams, Crisma and Clark1995, a.o.). Most of them found that children were able to distinguish between unergatives and unaccusatives at an early age. For instance, Snyder et al. (Reference Snyder, Hyams, Crisma and Clark1995) conducted a CHILDES corpus study and found that French- and Italian-speaking children had low error rates in using auxiliaries with reflexive and non-reflexive clitics, which are diagnostics for unaccusatives and unergatives. The findings suggested that children can distinguish between these two verb types as early as ages 2–3. However, a few studies have found that this ability is delayed. Babyonyshev et al. (Reference Babyonyshev, Ganger, Pesetsky and Wexler2001) showed that Russian-speaking children aged 3;0 to 6;6 had difficulties in producing the genitive-of-negation in which a genitive case is required for the underlying subject arguments with unaccusatives. They concluded that the A-movement, which involves moving the underlying subject of an unaccusative to the external-argument position, requires more time to mature. Children initially treat all intransitive verbs as unergatives and are unable to distinguish between unergatives and unaccusatives.
A few child-language acquisition studies have focused on the role of distributional cues in verb meanings and sentences during the acquisition of verb classes. Randall et al. (Reference Randall, van Hout, Weissenborn, Baayen, Alexiadou, Anagnostopoulou and Everaert2004) tested telicity and agentivity—two features argued to be crucial in distinguishing verb classes—in German and Dutch. In a cloze task involving Dutch and German-speaking adults, along with children aged 4–5 and 7–8, cues of telicity and/or agentivity were provided within verbal semantics or/and prepositional phrases. Participants were presented with scenes depicting novel actions involving telicity and agentivity and were introduced to these novel verbs within simple present-tense sentences. They were then asked to help a fictional monster produce the past tense of these novel verbs, which required selecting appropriate auxiliaries. Their choices of auxiliaries would show the verb class they assigned to novel verbs. The results showed that while children’s responses were not as clear as those of adults, the presence of telicity and/or agentivity did influence the choice of auxiliaries in the reformulated sentence. Notably, telicity had a stronger influence than agentivity in determining unaccusativity, because verbs occurring with an endpoint and an agent were often classified as unaccusatives (i.e., occurring with “be” auxiliary) rather than unergatives, suggesting that telicity took precedence over the influence of agentivity. Overall, this study demonstrated that telicity and agentivity (either within the verb meaning or in prepositional phrases) play a significant role in the selection of auxiliary verbs.
Randall et al.’s (Reference Randall, van Hout, Weissenborn, Baayen, Alexiadou, Anagnostopoulou and Everaert2004) research illustrates how children use various semantic and grammatical cues for verb categorization. However, this paper only examined novel verbs within simple sentences. Other types of sentences, with different distributional cues, may also influence the categorization of unaccusativity. These sentences, although not directly related to unaccusativity at first glance, could still have implicit effects on unaccusativity because various distributional information that interacts with unaccusativity, such as telicity and agentivity, is still present in these sentences. Therefore, there is a need to investigate various sentence types beyond simple transitive sentences to gain a comprehensive understanding of how different distributional cues in different sentences influence the categorization of unaccusativity. Furthermore, the study focused on European languages, and there is a necessity to explore languages from other linguistic families to broaden the scope of this type of research. Mandarin Chinese, for instance, is a language known to sometimes rely on sentential contexts to determine grammatical categories of nouns, verbs, and adjectives (cf. Zhu, Reference Zhu1982). It is possible the speakers might rely more on the various distributional cues to categorize intransitive verbs as well, and their influence might be easier to observe (e.g., Li et al., Reference Li, Jin and Tan2004). These considerations motivate our exploration of Mandarin in this paper, and the next step is to examine the distributional cues across various sentence types and whether children are sensitive to these distributional cues.
Regarding Mandarin unaccusativity, Wang et al. (Reference Wang, Yang and Shi2024) conducted two eye-fixation experiments to measure children’s looking time when hearing grammatical/ungrammatical sentences. In one experiment, children were exposed to one postverbal-subject sentence with either an unergative or an unaccusative in each trial. In another experiment, children were exposed to a preverbal-subject sentence with either verb class. Among the four patterns of sentences, only postverbal-subject sentences with unergatives were ungrammatical; the others were grammatical. The goal of this study was to measure children’s looking time after exposure to these sentences and examine their sensitivity to different verb classes within these sentences. The results showed that postverbal-subject sentences with unaccusatives were gazed at for a shorter time than the ungrammatical unergative counterparts. On the other hand, no such difference in looking time was found when the two verb classes occurred within grammatical preverbal-subject sentences. The authors claimed that children under the age of 2 were sensitive to the positions in which different verb classes can occur, suggesting that distributional cues in language input were likely to contribute to differentiating between unergatives and unaccusatives.
Wang et al. (Reference Wang, Yang and Shi2024)’s study provides a solid basis for understanding the acquisition of Mandarin unaccusativity and is informative of our current research. Specifically, we have preliminary evidence that Mandarin-speaking children can detect the distributional differences between unergatives and unaccusatives within postverbal and preverbal-subject sentences. However, the scope of this study is limited to sensitivity to only two types of sentences, while verbs can occur in many other sentence types within a language. Some of these sentences might not appear to be related to unaccusativity, but could still possess useful distributional cues that children can use for learning. Moreover, in the context of Taiwanese Mandarin, it was found that postverbal-subject sentences with unaccusatives are extremely rare and limited to a few verbs (Lin, Reference Lin2024) in children’s input. This poses a question regarding how this limited distributional difference between unergatives and unaccusatives in postverbal-subject sentences can contribute to the categorization in a real-world setting. Lastly, Wang et al. (Reference Wang, Yang and Shi2024)’s study only tested real verbs. Other confounding factors, such as the frequency of the verb, might serve as clues to distinguish between the two verb classes. Although they provided an account of frequency issue, it remains difficult to fully determine whether it was the distributional cues in these sentences or the frequency of the verbs themselves that triggered children’s successful distinctions. To address these concerns, our study employs novel verbs in both computational and behavioural experiments to control for these factors, which may provide further insights into the role of the distributional factor in learning Mandarin unaccusativity.
1.2 Word2vec and distributional environment
We leverage a machine learning model that utilizes distributional semantics, as these models can quantify the relationship between words and their surrounding words into computable representations. Word2vec model (Mikolov et al., Reference Mikolov, Chen, Corrado and Dean2013) is a shallow neural network that has shown effectiveness in a wide range of natural language processing tasks. The goal of the model is to learn the distribution of the input data and compute the vector of each word. During the training, the algorithm updates the weight to reflect the associations between a word and its surrounding words, thus learning the distributional pattern of a word and generating its word vector. After the model generates word vectors, we can compare the proximity of two word vectors to show their contextual similarity. This provides a way to quantify relationships among words, which shows the levels of similar contexts that words share. Consequently, Word2vec is a useful tool for exploring how a verb’s distributional meaning is influenced by its surrounding words in a language, which, eventually, impacts its clustering.
Using Word2vec models to investigate the argument structure is not unprecedented, and its efficacy has been demonstrated in prior studies (e.g., Lin & Washington, Reference Lin and Washington2023; You et al., Reference You, Bickel, Daum and Stoll2021). Research has shown that the performance of such models aligns well with human evaluations (e.g., Baroni et al., Reference Baroni, Dinu and Kruszewski2014; Mandera et al., Reference Mandera, Keuleers and Brysbaert2017; Mikolov et al., Reference Mikolov, Chen, Corrado and Dean2013), and Word2vec has been used as a computational model to investigate lexical semantics and contextual influences in language development (e.g., Alhama et al., Reference Alhama, Rowland and Kidd2023; Fourtassi, Reference Fourtassi2020; Huebner & Willits, Reference Huebner and Willits2018). This efficiency contrasts with behavioural experiments in the domain of language acquisition, in which only a few items can be tested, and the need for recruiting participants is often a cumbersome process in experimental studies. The modelling also yields preliminary results that can be validated and expanded upon through subsequent studies involving human subjects.
Next, we present a behavioural study with children to further demonstrate children’s sensitivity to distributional information, building upon previous Word2vec modelling. The entire experiment mimicked Randall et al. (Reference Randall, van Hout, Weissenborn, Baayen, Alexiadou, Anagnostopoulou and Everaert2004)’s design, where the position of the verb within a sentence type was filled with a novel verb. The verb class of novel verbs, which had no occurrence in children’s natural input, could only be decided by the verbal meaning (and other distributional information in the sentence type) we provided. Results indicated that children with additional access to distributional information in sentences provided ratings closer to expectations. Such findings from actual language learning support the notion that the categorization of verbs is not solely dependent on the meaning of a verb but can be reinforced by the distributional cues surrounding a verb.
In summary, this paper aims to explore the distributional information on learning verb classes of unergative and unaccusative, using a broader range of sentence types than previous studies. By examining and quantifying the effects of distributional cues across various sentences within the language input, including those seemingly unrelated to unaccusativity, we recommend that a theory of unaccusativity consider the role of distributional information in a language, given that the co-occurrences of different distributional cues can lead to distinct categorization results. The paper is structured as follows. Section 2 presents the Word2vec modelling experiment. Section 3 introduces a behavioural experiment on Mandarin children that adds more evidence to the results from Section 2. Section 4 concludes the paper and presents a theoretical account based on these conclusions.
2. Computational experiment (Word2vec modelling)
In this computational experiment, our goal is to train a Word2vec model using the sentences already present in the corpus, along with novel verbs occurring within various sentence types (refer to Materials and data for more explanations). Hence, the Word2vec model learns the distribution of words in language input, but the generated vectors of novel verbs only represent the surrounding cues (words) within the specific sentence types. Subsequently, we compare the vectors of real unergatives and unaccusatives with those of the novel verbs occurring within each sentence type, to quantify the distributional distance between the novel verbs and real verbs from two verb classes. This comparison allows us to assess how unergative (or unaccusative) a novel verb can be affected by the distributional cues within each sentence type, if learners are sensitive to such cues. Through this distributional semantics analysis, our aim was to evaluate the impact of different distributional information on distributional meanings that affect verb categorization.
2.1 Methods
Materials and data. Table 1 includes the 22 sentence types tested in this experiment, each designed to provide respective distributional information. We broadly classified these sentence types into three categories: unergative, unaccusative, or neutral. These categories were based on a body of prior research, which considered the positions of verbs, the agentivity of subjects, and the telicity and other factors in sentence types (e.g., Huang, Reference Huang2006; Laws and Yuan, Reference Laws and Yuan2010; Li & Thompson, Reference Li and Thompson1981; Liu, Reference Liu and Aranovich2007; Tsang, Reference Tsang1981; Wang, Reference Wang2010, a.o.).
Sentence types, expected categories, and translation with glosses. Subj (=subject), obj (=object), and loc (=locative) are bootstrapped sampled, and the positions of novel verbs are in bold. Dur = durative, imperf = imperfective, perf = perfective. For start+V, want+V, V + dur, imperf +V, will+V, going to+V, V refers to the novel verb that appears before/after the grammatical component (i.e., start, want, dur, imperf, will, going to). See Appendix 2 and the behavioural experiment section for more reasoning behind the expected categories.
To illustrate, consider sentences with a single argument and a verb marked with the imperfective aspect, referred to henceforth as the sentence type imperf+V in (6). Such sentences denote durative or ongoing events (Li & Thompson, Reference Li and Thompson1981). This duration of events is more incompatible with telic events (telic events being typically encoded by unaccusatives), making unaccusatives more incompatible with this sentence type.
We acknowledge that there exist some unaccusatives that can occur in this sentence type. We labelled it as “unergative” to represent that the imperf+V sentence type is more compatible with “unergatives” and the cues in it are more likely to exert unergative effects on the verb, but it by no means implies that this sentence type is exclusive to unergatives. This labelling was also a tentative prediction, and we were hoping to demonstrate the effect of these distributional cues as we proceeded through the studies. The distributional information in this sentence type, i.e., the co-occurrence with the imperfective marker zai4, can therefore be an important cue for recognizing a novel verb as more likely to be unergative during the categorization. Likewise, we identified 21 other sentence types (making a total of 22 sentence types) which we can assign to a preferred category in the same manner we did for the imperf+V sentence type. The full list of sentence types and their expected categories is listed in Table 1. Note that the element ba is a “preposition” that preposes the affected argument by the following compound verb (such as in 1st and 2nd resultatives as well) (cf. Huang et al., Reference Huang, Li and Li2009). The suffix -guo is an experiential marker indicating a person’s prior experience, and is followed by a perfective marker -le in our sentence type Perf(guo)-le. “1st” and “2nd” in resultatives indicate the position of the novel verb in the compound verb (V-V) construction, in which the second verb typically conveys a result of the first verb. LVS refers to a sentence type following a locative + verb + subject order; see Section 2.2 for further discussion of resultatives and LVS. We also included their reasoning for expected categories in the behavioural experiment and Supplementary Appendix 2.
Some sentence types, such as Resultative (2nd), Resultative (2nd) + object, and Resultative (1st), appeared multiple times in Table 1 (#2, 3; #4, 5; #11–16). These sentence types included a second verb, that is, the other verb (either transitive or intransitive) in the compound verb, which might cause different distributional results from the other similar sentences due to Word2vec algorithm’s sensitivity to surrounding words. To understand this potential influence, we included duplicates of these sentence types in our training data but replaced the second verb with an alternative verb to see if these differences affect the semantic vectors of novel verbs. However, as we will see later, our results indicated that the words in these sentence types showed minimal differences in their results.
We also examined whether the total occurrence number of a sentence type affected the categorization of the novel verb in it, so we trained another two models, which yielded a total of three models in this computational experiment (LOW, MIDDLE, HIGH). In the LOW model, we created four sentences per sentence type using the same novel verb, resulting in a total of 88 sentences of the 22 sentence types. We employed a bootstrapped sampling on the nouns in these 22 sentence types, and the nouns were randomly chosen from three curated lists containing 30 subjects, 10 objects, and 10 locations, as shown in Appendix 1. Refer to Table 1 for the expected categories of each sentence type, example sentences, and their translations. For the other two models, MIDDLE and HIGH, each sentence type was replicated to appear 8 and 32 times, respectively, with nouns chosen through the same bootstrapped samplingFootnote 2. This modelling method allowed us to evaluate the influence of occurrence frequency on the categorization of verbs.
2.2 Experiment setup
We employed the Continuous Bag of Words (CBOW) method from the Gensim package (Rehurek & Sojka Reference Rehurek and Sojka2011) in Python to train three Word2vec models. The CBOW approach is relevant to our objectives as it predicts target words based on their surrounding words rather than predicting surrounding words based on target words, as in the Skip-gram method. Using the CBOW algorithm, therefore aligned with our purpose to obtain vectors of target words based on surrounding words. We utilized a suite of Taiwanese Mandarin data from the CHILDES database (MacWhinney, Reference MacWhinney2000), namely the Chang1, Chang2, ChangPlay, ChangPN, TCCM, and TCCM-Readings corpora, so as to match the Taiwanese data from our behavioural experiment. These corpora contained the speech from both children and adults during their interactions. The datapoints in this training dataset consisted of 173,673 sentences, and additional sentences containing novel verbs were added to this number. Text segmentation into words and tokenization were conducted using the jieba package (Sun, Reference Sun2022) and a self-defined dictionary. To avoid initial randomization of weights, which can potentially affect distributional representations during each time of training, we set the Python hash environment variable and Word2vec hash function to zero and hardcoded the default parameters of Word2vec (window = 5, min_count = 3, workers = 1). Each model (LOW, MIDDLE, HIGH) was trained 20 times to ensure the reliability of the results.
After we obtained the word vectors of the 22 novel verbsFootnote 3 at each time of training, we compared their cosine distances (as employed in Mikolov et al., Reference Mikolov, Chen, Corrado and Dean2013) with real unaccusatives, specifically diao4 “to drop” and lai2 “to come.” These verbs were chosen because their subjects were found postverbally in the corpus. Similarly, we compared the vectors of 22 novel verbs with those of real unergatives from the corpus, specifically tiao4 “to jump,” wan2 “to play,” and ku1 “to cry,” because those verbs were found in the durative aspect sentence in the corpus.Footnote 4 These real verbs were also deemed as core unergatives and unaccusatives according to Sorace (Reference Sorace2000).
The rationale and the assumption of this experiment are as follows: we suppose real unergatives and unaccusatives occur systematically within these sentence types. The co-occurrence with specific words in a sentence type becomes an obvious distributional cue to indicate a verb being unergative or unaccusative. Learners of this language are supposed to be sensitive to these distributional cues and the co-occurring verb classes, so when a novel verb occurs in a sentence type with the unergative or unaccusative distributional cues, the novel verb is more possibly regarded as either verb class, and hence the distributional cues affect the categorization of novel verbs. Based on this, the cosine similarity measures between the novel verbs and real (unergative or unaccusative) verbs are predicted to show that novel verbs from certain sentence types are more similar to the reference class (unergative or unaccusative) than those from other constructions. In other words, the measured cosine similarities should reveal that novel verbs in certain constructions are distinct from novel verbs in other constructions—forming distributional separation—rather than forming a single undifferentiated group.
2.3 Results
The cosine distances between the 22 novel verbs and real unaccusatives and unergatives were calculated over 20 training sessions, and the distances for each novel verb were averaged across these sessions, resulting in 22 final averaged distances and standard deviations in each model condition. The results are illustrated in Figures 1 and 2, which depict the means and standard deviations of the cosine similarities for the three models. Figure 1 shows the comparison with real unaccusatives, while Figure 2 shows the comparison with real unergatives. In Figure 1, sentence types are arranged so that unaccusative categories appear on the left, unergative categories on the right, and neutral categories in the centre. We predicted a left-to-right downward trend in Figure 1 and a left-to-right upward trend in Figure 2. This prediction was borne out by the results: novel verbs in sentence types #1-#6 exhibited greater cosine similarities in Figure 1 and lower similarities in Figure 2. On the contrary, novel verbs in types #18-#22 displayed the opposite trend, while those in types #7-#17 showed intermediate values. Additionally, we only observed slight but not huge differences between certain sentence types (#2, 3; #4, 5; #11–16) that differed by only one word, which suggests that minor lexical variations did not largely affect the distributional results.
Average cosine similarities of novel verbs compared to real unaccusatives. The Figures are divided into three partitions,each representing an expected category (unergative/neutral/unaccusative) as shown in the box.
Average cosine similarities of novel verbs compared to real unergatives. The Figures are divided into three partitions, each representing an expected category (unergative/neutral/unaccusative) as shown in the box.
The differences in cosine similarities between verbs in types #1-#6 and those in types #18-#22 across three models were statistically significant when compared with both real unaccusatives and unergatives, which indicates a clear separation (Student t-test, t(31) = 4.85, p < .001 for unaccusative in Figure 1 and t(31) = −8.42, p < .001 for unergative comparisons in Figure 2). These results indicate that novel verbs from certain sentence types were more similar to the reference class than those from other sentence types. We interpret this as distributional cues in certain sentence types having more distributional effects in pulling the novel verbs into the unaccusative or unergative verb class than in the other sentence types.
Further examination revealed that the distributional effect within these sentence types not only aligned with our expectations but also exhibited variability. Even within the same expected categories, the effects were not uniform. For instance, novel verbs in some unaccusative sentences (e.g., #1 LVS) had more distributional similarities to real unaccusatives than others (e.g., #2, #3 resultatives (2nd)), as shown in Figure 1. A similar pattern can be observed among unergative sentences, with some (e.g., #22 want +V) having more similarities than others (e.g., #18 start +V), as shown in Figure 2. Their distributional effects formed gradients rather than fitting neatly into two or three categories as initially listed in Table 1. The co-occurrence with the distributional cues in some sentence types served as a stronger cue (e.g., LVS, resultative (2nd), want+V, imperf+V) to pull the novel verbs into unaccusative or unergative than in others, and exhibited gradients even within the same expected categories.
Additionally, the results indicated that the total occurrence number of a sentence type was associated with the stability of verb categorization. When the training data included fewer occurrences of each sentence type (as in the LOW model in Figures 1 and 2), the cosine similarity values showed more fluctuation. However, with an increased occurrence of each sentence type (as in the MIDDLE and HIGH models), the cosine similarity values tended to stabilize, resulting in smoother curves. The mean standard deviations of the three models decreased (for LOW, MIDDLE, and HIGH: std = 0.05, 0.04, 0.02), respectively, when compared with real unaccusatives, with their differences reaching statistical significance (LOW versus MIDDLE, t(42) = 2.14, p < .05; LOW versus HIGH, t(42) = 6.06, p < .001; and MIDDLE versus HIGH, t(42) = 4.72, p < .001). Similar means of standard deviations were observed and the statistical analyses on real unergatives were similar to those of real unaccusatives (LOW, MIDDLE, and HIGH: std = 0.06, 0.04, 0.02, LOW versus MIDDLE, t(42) = 2.91, p < .001; LOW versus HIGH, t(42) = 7.11, p < .001; MIDDLE versus HIGH, t(42) = 5.53, p < .001). These findings underscore that a higher frequency of sentence occurrences can be associated with a more stable categorization of novel verbs, and frequency is a significant factor in shaping the categorization of unaccusativity.
Different surrounding nouns can affect the vectors of novel verbs and might affect the distributional vectors as well, so we further present the results without employing bootstrapping sampling in Figures 3 and 4. Across all three models, the nouns used in each sentence type remained consistent. An increased fluctuation in the lines representing each group was observed in these figures, as compared to the bootstrapping sampled results. This fluctuation could be attributed to the fact that when novel verbs were consistently paired with specific nouns in sentence types, their distributional representations became closer to specific nouns. This contrasted with the real unergatives and unaccusatives, which occurred with a greater variety of nouns in the corpus, so the observed fluctuation in these Figures is expected. Notably, the condition with high frequency (HIGH model) showed the greatest stability in its results, and we observed more fluctuating lines in the low-frequency condition (LOW model). However, these differences did not affect the overall pattern, as we still observed the statistical difference in cosine similarity between unergative and unaccusative sentences. In Figure 3, this difference between unergative and unaccusative sentences across three models was significant (t(31) = 3.61, p ≤ .001), and a similar pattern was significant in Figure 4 (t(31) = −6.09, p ≤ .001). These findings indicate that, even after eliminating the confounding factor of bootstrapping sampling, novel verbs from certain sentence types were closer to the distribution of either real unergatives or unaccusatives. Such similarity could provide cues for the process of categorizing a novel verb into one of these two classes.
Average cosine similarities of novel verbs compared to real unaccusatives, without bootstrapping. The figures are divided into three partitions, each representing an expected category (unergative/neutral/unaccusative) as shown in the box.
Average cosine similarities of novel verbs compared to real unergatives, without bootstrapping. The figures are divided into three partitions, each representing an expected category (unergative/neutral/unaccusative) as shown in the box.
2.4 Discussion
In our computational modelling, we have demonstrated that the Word2vec model can represent the distributional information of the novel verbs in each sentence type, and the representations form similar distributions to real unergatives and unaccusatives. Specifically, novel verbs in certain sentence types tend to show the most distributional similarities to real unaccusatives, and the others show the most similarities to real unergatives. We suggest that if learners are sensitive to the distributional cues in these sentence types, these sentence types exhibit effects on verb categorization and can be important cues for the process of categorizing verb classes. Our findings also suggest that these effects on categorization may not be strictly categorical; rather, they exhibit gradients. This indicates that, even within the same category of unergative or unaccusative sentence types, when novel verbs occur in certain sentence types, their co-occurrence with surrounding words in these sentence types provides stronger clues of the verbs being unergative, while the other sentence types provide weaker clues of the verbs being unaccusative. Those sentence types that are sometimes viewed as unrelated to the unergative-unaccusative distinction (non-diagnostic sentences) can, in fact, play an important role in categorization and are beneficial to language learners. Importantly, our experiment reveals that the stability of categorization is correlated with the occurrence number of a sentence type. This correlation underscores frequency as an additional and significant factor for verb categorization. While our computational modelling has provided evidence that distributional cues can affect verb categorization, one needs to demonstrate that humans are sensitive to distributional information through real-world experiments. Therefore, we designed and conducted a child language acquisition experiment to further investigate the distributional information of sentence types on verb categorization and validate children’s sensitivity to the distributional information.
3. Behavioural experiment
To bolster the findings from the previous Word2vec modelling, we conducted a behavioural study following the approach of Randall et al. (Reference Randall, van Hout, Weissenborn, Baayen, Alexiadou, Anagnostopoulou and Everaert2004) but with modifications. In our experiment, we observed participants’ learning of novel verbs by exposing them to a wider array of sentences beyond simple transitive sentences. Participants, after a short exposure to the meanings (and sentential environments) of novel verbs, were asked to rate the acceptability of test sentences (two diagnostics), to determine the verb classes of novel verbs, rather than completing sentences as in Randall et al. (Reference Randall, van Hout, Weissenborn, Baayen, Alexiadou, Anagnostopoulou and Everaert2004). Since participants had no prior knowledge about novel verbs, their categorization could only be based on verb meanings (and distributional information) provided in the experiment. This experiment aimed to examine whether the distributional information in various sentences can reinforce the categorization in addition to verbal meanings. Including this behavioural experiment allowed us to bridge the gap between computational modelling and real-world categorization processes.
3.1 Methods and materials
The behavioural experiment had two groups: the non-environment group and the environment group, and was an acceptability judgment task modelled on Ambridge (Reference Ambridge and Hoff2011). By “environment,” we mean the distributional cues in sentences. The non-environment group received only the meanings of verbs, while the environment group received both meanings and sentential environments. This distinction served to examine the extent to which distributional cues in specific sentential environments can reinforce the categorization.
The experiment consisted of two phases for each critical item: the exposure phase and the test phase. Each critical item involved a single novel verb. During the exposure phase, participants in the non-environment group were presented with animations that approximately depicted the telicity of events, for example, an arrow that strikes a target depicts telicity, while a person running continuously depicts atelicity. This telic/atelic distinction is the most important verbal semantics associated with (Mandarin) unaccusativity, as suggested in Section 1 and previous literature (e.g., Borer, Reference Borer and Shimron2000; Dowty, Reference Dowty1991; Hoekstra, Reference Hoekstra1984; Liu, Reference Liu and Aranovich2007; Lu, Reference Lu2019; Randall et al., Reference Randall, van Hout, Weissenborn, Baayen, Alexiadou, Anagnostopoulou and Everaert2004; Tenny, Reference Tenny1994; van Hout, Reference van Hout, Alexiadou, Anagnostopoulou and Everaert2004; Van Valin, Reference Van Valin1990 a.o.)Footnote 5. We suspected that participants would not be able to grasp and leverage the distributional cues in sentences with totally meaningless verbs (as in our computational modelling), so the inclusion of verb meanings was to facilitate the task by simulating the natural setting of the language environment participants were exposed to. To avoid the potential association of the novel verbs with real verbs, each animation consisted of two actions while depicting either a telic or an atelic event overall. For instance, the animation for the novel unaccusative verb mi2 depicted a series of events like “hit the wall” and “stops on the wall,” illustrating a telic scenario where “an arrow hit the wall and stops on the wall.”
The environment group, on the other hand, received not only the animations depicting events (the intended meanings), but also one or two sentential environments containing the same novel verb during the exposure phase. These sentential environments were the single novel verb occurring within one or two of the sentence types used in the computational experiment above, that is, unergative or unaccusative sentence types, and were intended to provide plausible and natural sentence contexts in which the novel verb could occur. Thus, the environment group received information on the meanings of each verb in addition to the distributional cues in the sentential environment, while the non-environment group received information on only the meanings of the verb. This design allowed us to observe whether the additional sentential environments facilitated the categorization of the novel verbs above and beyond what the verb meanings alone permitted. In the test phase, which was identical in both groups, participants heard the same novel verb within a test sentence (either in a postverbal-subject diagnostic or a durative aspect sentence) and were asked to rate the acceptability of the test sentence. This test phase was to show participants’ classification of each novel verb after exposure to stimuli.
Six novel verbs were createdFootnote 6, including two unergative and four unaccusative-targeted verbs, as shown in (7). We created more unaccusative-targeted verbs because we would like to test two sentence types and their additive effects on categorization, as Mandarin unaccusatives seem to often show inconsistent behaviours in previous studies (e.g., Laws & Yuan, Reference Laws and Yuan2010). These verbs were included in both the non-environment and environment groups. Three sentence types were used in the environment group: two for the unaccusative (LVS and resultatives(2nd)) and one for the unergative (imperf+V), since these sentence types were the most effective in categorization in the computational experiment. Notably, two unergative-targeted verbs occurred in the sentence imperf+V, while two unaccusative-targeted verbs occurred in either of LVS or resultatives(2nd), and two occurred in both sentence types (LVS and resultatives(2nd)). Verbs appearing in both sentence types, LVS and resultatives(2nd) were used to explore the additive distributional effects of including two sentential environments.
The selection of these six novel verbs was based on a norming test involving five native Taiwanese Mandarin adults. The adults watched the animations and attempted to match them with three real verbs provided or any other real verbs they thought of. Although participants were given three real verbs as options for each animation, no real verb was chosen or raised more than three times with each animation (See Appendix 3 for the three verbs provided). This indicated that the adults found it challenging to directly link the animations to any real verbs, thereby affirming the distinctiveness of the novel verbs from real Mandarin verbs.
The distributional cues in imperf+V (imperfective) are noted in the previous section. As for potential cues in LVS and resultatives, LVS consists of a locative, a verb, and a subject. In Laws and Yuan (Reference Laws and Yuan2010), it is noted that this sentence type is most compatible with unaccusatives. Many cues can be identified in this sentence type, for instance: the non-canonical word order, which begins with a locative, and the subject is inverted to the postverbal position. This word order is similar to the postverbal-subject diagnostic in Mandarin and is characteristic of unaccusatives because the base-generated position for the subject of an unaccusative is argued to be postverbal (Burzio Reference Burzio1981, Reference Burzio1986). Also, this sentence type contains a perfective marker -le. Despite the results in our previous experiment that suggest the co-occurrence of this aspect marker is not exclusive to unaccusatives, verbs occurring with a perfective marker often suggest the event is read as telic (e.g., Liu, Reference Liu and Aranovich2007). Therefore, this perfective marker might be a distributional cue to the identification of unaccusatives. One important fact to note here is that even though LVS and the postverbal-subject diagnostic are similar in these aspects, their inner structure is argued to be different (cf. Paul et al., Reference Paul, Lu and Lee2020), so we included LVS as a sentential environment different from the postverbal-subject diagnostic.
As for Resultatives (2nd) + object, previous studies (e.g., Cheng & Huang, Reference Cheng, Huang, Chen and Tzeng1994; Liu, Reference Liu2021; Wang, Reference Wang2010) suggest that the second verb of a resultative requires a (change of) state interpretation and imposes the result interpretation on the verb. This sentence type also has an object that can often serve as the subject of the second verb in the resultative, and this subject follows the second verb linearly in Resultatives (2nd) + object. It also requires a perfective aspect that follows the second verb. The word order and aspect marker are similar to the postverbal-subject diagnostic, which allows for unaccusatives to appear, so all these can be supporting cues for classifying the verb at this position as an unaccusative.
Each participant was randomly assigned to one group, environment, or non-environment. The same six novel verbs occurred in both groups, and the categorization was determined by the two types of test sentences, postverbal-subject and durative test sentences. Each novel verb was crossed with the two types of test sentences, yielding a total of 12 critical items in one group (2 for each verb). In addition, five filler items were created using real verbs, yielding 17 items in each group. Each group, environment, or non-environment was given the same 17 stimuli, thus yielding 34 items in total across both groups. Some of these filler sentences were made ambiguously acceptable or unacceptable, in order to prevent participants from assuming that all real verb sentences were acceptable. We interspersed the critical items with five filler items after every two critical items, to keep participants engaged and reduce the predictability of our research direction. See the details of items in Appendix 4.
3.2 Procedure
The entire procedure was modelled on the acceptability judgment tasks in Ambridge (Reference Ambridge and Hoff2011) and Randall et al. (Reference Randall, van Hout, Weissenborn, Baayen, Alexiadou, Anagnostopoulou and Everaert2004), with modifications. Participants and the investigator were seated together at a table (all children) or in a Zoom meeting room (all adults), looking at PowerPoint slides. Participants were briefed that their task was to assist a cartoon dog in learning Mandarin by assessing the acceptability of various sentences. During the exposure phase, participants were introduced to the sounds of a novel verb they were going to learn and watched an animation illustrating this verb (and heard this verb in a sentential environment if in the environment group). After the exposure phase, we asked participants to repeat the sounds of this novel verb to ensure their attention to the stimuli. During the test phase, they were asked to rate the acceptability of this novel verb in a test sentence. Participants were shown a paper with a circle, denoting “acceptable,” and a cross, denoting “unacceptable,” and they were asked to record their answers by placing stickers on the corresponding symbol (all children) or noting down their answers (all adults). They were encouraged to use their intuition to rate the test sentence, even when encountering an unfamiliar verb.
To ensure participants were comfortable with the task, the session began with four additional practice trials, using items separate from the 17 critical items, following the procedure of Ambridge (Reference Ambridge and Hoff2011). The four practice trials consisted of two trials with real verbs and two trials with novel verbs, and were the same across the two groups. Participants were asked to follow the same procedures as they would in the main experiment, but they received feedback on their selection, regardless of whether their selections were correct. For example, in practice trial 1, participants were always told that the co-occurrence of diao4 “to drop” and shang4 “up” was not plausible, and this test sentence should be rejected. In practice trial 2, participants were told that the test sentence was plausible and that the phrase fei1-shang4-qu4 “fly upward” also sounded possible. Successful completion of all practice items was required to progress to the main experiment, which consisted of 17 items. If errors were made during the practice trials, the trials were repeated to ensure full comprehension. The entire experimental session lasted less than 20 minutes, designed to be engaging yet concise enough to maintain the attention of the child participants.
An example item in the environment group is provided in Figure 5. Each item contained two slides (a) and (b), each paired with corresponding audio and visual stimuli. Only the environment group had the shaded part shown in the figure. In the first slide (exposure phase, slide a), participants in the environment group watched a scenario that depicted the meaning of a novel verb mi2, and at the same time heard the verb in a resultative sentence (resultatives(2nd)). This implied that the verb mi2 can occur in the resultative sentence type. In the second slide (test phase, slide b), they were asked to rate the acceptability of a test sentence. This test sentence can show the verb class participants assigned to the novel verb.
The stimuli for the novel verb mi2 in the environment group.
a. (first slide, participants watched an animation depicting the verb mi2 and heard the verb occur in a resultative sentence). b. (second slide, a dog used the novel verb mi2 within a test sentence, and participants rated the acceptability of the postverbal-subject test sentence).
3.3 Participants
Sixty-six Mandarin-speaking children, aged between 5;3 to 6;5 (M = 5;9, Female = 32, Male = 34) were recruited from a kindergarten in Taipei, Taiwan. Forty-two native Taiwanese Mandarin adult speakers (aged 18–40, M = 28, Female = 21, Male = 21) were recruited online, as a control group to confirm the effectiveness of the design. The child and adult participants were equally distributed between the two experimental groups. Children’s age range was selected based on findings by Lin & Deen (Reference Lin and Deen2021), who found that (Taiwanese) Mandarin-speaking children did not exhibit adult-like differentiation abilities between unaccusatives or unergatives until around the age of 6 years.
3.4 Prediction
We expect the adult control group to perform more “accurately” (closer to the predictions in Table 2) than the children overall. Should the cues in sentence types effectively reinforce the categorization of novel verbs, we expect the environment group to provide ratings on test sentences that are significantly more “accurate” than the non-environment group and closer to the adult control results. More specifically, for the durative sentence, we expect the environment group to show higher rates of rejection for unaccusative-targeted novel verbs and greater acceptance for unergative-targeted novel verbs than the non-environment group, while for the postverbal-subject sentence, unaccusative-targeted novel verbs should be rated as more acceptable and unergative-targeted novel verbs should be rated as less acceptable (See Table 2).
The predicted acceptability of test sentences in the behavioural experiment
On the other hand, if the cues in sentence types do not assist in the categorization of verbs, we expect no significant difference between the environment and non-environment groups (See Appendix 4 for the prediction of each item).
3.5 Results
We first analyse the experimental data to see if children can distinguish between unaccusative and unergative-targeted novel verbs in both groups.
We conducted logistic regression analyses using the Generalized Linear Mixed Model (glmer) in R (Bates et al., Reference Bates, Mächler, Bolker and Walker2015). The independent variables were verb_type (unerg/unacc) and test sentence (postverbal-subject/durative diagnostic), with the children’s acceptability ratings as the dependent variableFootnote 7. The results showed statistically significant interaction effects between verb_type and test sentence in both the non-environment and environment groups (Table 3, β = −1.73, SE = 0.45, p < .001; Table 4, β = 1.58, SE = 0.46, p < .001). The significant interaction effects suggest that children were capable of distinguishing between unergatives and unaccusatives using the depicted telicity of the event. Such a finding lends support to the notion that verb meanings alone played a crucial role in categorization, and the approximate meanings were correctly designed to trigger the intended categorization.
Model statistics for the model answer ~ unerg_unacc*test_sentence (non-environment group, children)
Model statistics for the model answer ~ unerg_unacc* test_sentence (environment group, children)
To examine the categorization of two groups, ratings of test sentences were converted into binary labels of “accuracy.” Ratings that aligned with the expected categories were labelled as “accurate,” while those that did not were labelled as “inaccurate.” Figure 6 depicts the accuracy percentage for unergative-targeted verbs. We see that the environment group consistently exhibited higher accuracy rates, with both postverbal-subject and durative test sentences rated more accurately (durative: 47% vs 60.6% vs postverbal-subject: 34.8% vs 60.6%). Figure 7 depicts a similar trend with unaccusative-targeted verbs (durative: 45.5% vs 62.9% vs postverbal-subject: 31.8% vs 51.5%). The higher accuracy from durative test sentences than from postverbal-subject sentences aligns with previous studies using real verbs, that durative test sentences were rated correctly earlier (e.g., Lin & Deen, Reference Lin and Deen2021). Figure 8 further illustrates the mean accuracy across children and adults: in children, the environment group again showed higher accuracy than the non-environment group (58.3% vs 39.4%). The adult data showed a similar trend, though at a generally higher “accuracy” level (73.9% vs 63.6%).
Accuracy rates with the two test sentences with unergative-targeted verbs in two groups (children).
Accuracy rates with the two test sentences with unaccusative-targeted verbs in two groups (children).
Children and adults’ “accuracy” in the non-environment and environment groups.
We employed the same Generalized Linear Mixed Models (glmer) on children’s data, and “accuracy” was the dependent variable. The independent variables included “group.” The analysis revealed that the “group” variable was a statistically significant predictor of “accuracy” (Table 5, β = 0.77, SE = 0.15, p < .001), which indicates that sentence environments in which novel verbs occurred led to a significant improvement in children’s categorization.
Model statistics for the model accuracy ~ group (children)
In Figure 9, we further see that sentence-level mean “accuracy” across the two groups exhibited similar trends. The sentences in the non-environment group exhibited 30–50% mean accuracies, while sentences in the environment group exhibited 50–70% accuracies, showing a general increase in accuracy rate. Further logistic regression analyses were conducted with sentence_type and group as independent variables (and their interaction) and “accuracy” as the dependent variable. As shown in Figure 9, pairwise comparisons revealed that all sentence types, LVS, resultative, and imperfective, and the combination of LVS and resultatives were significant predictors of “accuracy” (Table 6, LVS: β = −0.87, SE = 0.04, p < .001; imperfective: β = −0.80, SE = 0.03, p < .001; LVS + resultative: β = −0.56, SE = 0.03, p < .001; resultative: β = −1.07, SE = 0.04, p < .001). However, we found that the exposure to two sentence types did not result in much higher accuracy. This is supported with our logistic regression model in Table 7 when we fit sentence_type and group as independent variables and accuracy as a dependent variable, we see that the improvement from the LVS + resultatives conditions was the worst (Table 7, LVS + resultative*group: β = −0.31, SE = 0.05, p < .001, LVS was the reference level). We suspect this may be due to the relatively short exposure to two sentential environments, where processing two sentence types created additional cognitive load for children during the short exposure phase. On the other hand, resultatives had the strongest effect, as also demonstrated in the model in Table 7 (Resultative*group: β = 0.20, SE = 0.06, p < .01). We think that this may be due to the greater frequency of resultatives in children’s language input (cf., Lin, Reference Lin2024), and the familiarity with this sentence type makes it easier for them to identify unaccusatives in this sentence type compared to the less frequent LVS.
Sentence-level mean “accuracy” across two groups (children).
Model statistics for the model accuracy ~ sentence_type*group (children)
Model statistics for the model accuracy ~ sentence_type*group (children)
The adult groups were originally included as a control group to identify any unexpected patterns under the experimental conditions. Overall, similar results were observed, except that the accuracy was generally higher and the improvement in the environment group was less obvious than that of the children. The results showed high accuracy with unergative-targeted verbs within both test sentences across both groups (Figure 10), while unaccusative-targeted verbs exhibited relatively lower accuracy within both test sentences across both groups (Figure 11). Statistical analyses using the same variables (answer was the dependent variable) and the same model as in children’s condition, showed that in both non-environment and environment groups, there were interaction effects from test sentence and verb_type (Table 8, β = 3.95, SE = 0.71, p < .001; Table 9, β = 5.52, SE = 0.76, p < .001), which suggest adults were able to distinguish between unergative and unaccusative-targeted verbs as well.
Accuracy rates with the two test sentences with unergative-targeted verbs in two groups (adults).
Accuracy rates with the two test sentences with unaccusative-targeted verbs in two groups (adults).
Model statistics for the model answer ~ unerg_unacc* test_sentences (non-environment group, adults)
Model statistics for the model answer ~ unerg_unacc* test_sentences (environment group, adults)
A similar trend was observed with adults’ sentence-level performances as well (Figure 12). With all the sentence types, there was improvement in the environment group, compared with the non-environment group, although the improvement seems less than in the children’s group. Our logistic regression models showed that the environment group was still a statistically significant predictor of “accuracy” (Table 10, β = 0.48, SE = 0.19, p < .01), and pairwise comparisons revealed that all the sentence types were also significant predictors in the sentence-level results (Table 11, LVS: β = −0.48, SE = 0.07, p < .001; imperf+V: β = −0.21, SE = 0.07, p < .01; LVS+ resultative: β = −0.79, SE = 0.05, p < .001; resultative: β = −0.31, SE = 0.07, p < .001). We attributed the less reinforcement effect in the environment group of adults to their already higher performance in the non-environment group, and the room for improvement in the environment group may have been limited.
Sentence-level mean “accuracy” across two groups (adults).
Model statistics of accuracy ~ group (adults)
Model statistics of accuracy ~ sentence_type*group (adults)
One interesting point to note is that unergative-targeted verbs received much higher accuracy – and less improvement – than unaccusative-targeted verbs overall. This was also borne out in a logistic regression model, where we fit sentence_type and group as independent variables and accuracy as the dependent variable: the improvement from the imperfective was significantly less than the other sentence types (Table 12, LVS: β = 0.27, SE = 0.10, p < .01; LVS + Resultative: β = 0.58, SE = 0.08, p < .001), except for the resultative (β = 0.10, SE = 0.10, p = 0.3). We attribute this trend to the fact that in schools in Taiwan, teachers sometimes introduce an oversimplified rule—namely, that “a verb is compatible with the durative aspect and the subject cannot occur postverbally.” As a result, adults in the non-environment group may have found it easier to recall and apply this “rule of thumb” to unergative-targeted verbs in our experimental setting. Moreover, the strongest reinforcement effect was observed with LVS + resultative (Table 12, LVS + resultative: β = 0.58, SE = 0.08, p < .001), likely due to the additive effect of two sentence types. The greater effect from LVS compared with the resultative (Table 12) may arise from its structural similarity to postverbal-subject test sentences. Hearing LVS may thus help adults recognize its compatibility with postverbal-subject sentences. These comparisons of results should be further examined in future research to confirm their robustness and to assess the extent to which they generalize across experimental settings.
Model statistics of accuracy ~ sentence_type*group (adults)
While a comparison among sentence types across two experiments is tempting, we did not observe a consistent trend. When we compared the results with those of computational modelling, we did see a discrepancy. In the children’s behavioural experiment, resultatives had a stronger reinforcement effect than LVS, whereas the computational results indicated that LVS was more effective in categorizing novel verbs as unaccusative. This discrepancy is likely due to the inclusion of verbal semantics and the modelling strategies. We also found participant-related differences. For instance, LVS + resultatives most significantly enhanced the recognition of unaccusatives in adults’ experiments, but this effect was the least obvious in the children’s group, perhaps because processing two sentence types simultaneously was too demanding for children. We also found that LVS reinforced categorization more strongly for adults than for children, likely because adults have had greater exposure to LVS. Since LVS is structurally similar to postverbal-subject sentences, this similarity and greater exposure may have contributed to more accurate categorization among adults. Future research should apply multiple quantification methods to confirm these effects across different experiments and participants. Nevertheless, the overall findings still suggest that all sentence types have effectively reinforced the verb classes that verbal semantics are linked to, and our research purpose was validated.
3.6 Interim discussion
In this behavioural experiment, we have shown that distributional cues in three sentential environments can reinforce verb categorization, compared to results based solely on verbal meanings. While verbal meaning alone provides a preliminary categorization of unergatives and unaccusatives, the “accuracy” is only around 39% in the children’s group and 64% in the adults’ group. It is only after the sentential environment with distributional information is provided that categorization becomes more successful. This finding might shed light on certain behaviours of unaccusatives in Mandarin, and we provide our discussion below.
4. Discussion and conclusion
In this paper, we have explored whether the distributional cues in various sentences affect the categorization of unergatives and unaccusatives. We conducted two experiments using a machine learning model and a child language acquisition method. Through Word2vec modelling, we have found that different sentence types in which novel verbs occur can impact their distributional representations, subsequently affecting verb categorization measured by the cosine distance with real unergatives and unaccusatives. These results reveal gradient effects from different sentence types, indicating that the effect from distributional cues, if used by language learners, is not strictly unergative or unaccusative but rather exhibits gradients that vary between unergative and unaccusative. Additionally, our findings underscore a significant association between the overall occurrences of sentence types and the stability of their categorization. This highlights the crucial role of repeated exposure of verbs within sentential environments in determining how verbs are classified. Following this, we selected three sentence types from the computational experiment and conducted a behavioural experiment with Mandarin-speaking children. This real-world experiment aimed to examine the effect of distributional cues on verb categorization in a natural language setting. We have demonstrated that children can categorize verbs into unaccusative or unergative using the observed meanings in animations. Moreover, the “accuracy” of categorization was significantly enhanced when participants were provided with sentence environments that were compatible with the intended verb classes that verbal semantics linked to. This substantiates the proposal that distributional cues in respective sentence types are critical in the categorization of unaccusativity, and children can make use of these distributional cues to categorize verbs into unaccusative or unergative, even though these sentence types might not seem related to unaccusativity at first sight.
A pertinent inquiry arises regarding the potential of any sentence type to reinforce the categorization linked to verb semantics. In our behavioural experiment, participants in the environment group were exposed to cues including both the semantic content and sentential contexts of novel verbs. This design raises the question of whether any sentential environment, irrespective of its specific properties, could reinforce categorization, and whether verb semantics alone determines categorization outcomes. However, our computational findings mitigate this concern. The diverse sentential environments employed in our computational experiment exhibited distinct distributional cues, which in turn yielded differentiated verb categorization results. This pattern suggests that the specific nature of distributional cues, rather than mere exposure to sentential environments, drives the clustering of novel verbs. Consequently, it is improbable that children would utilize different distributional cues to achieve the same categorization outcomes.
The findings of this study provide a theoretical perspective that may accommodate the behaviours of several unaccusatives in Mandarin. Previous theories have often linked the unaccusative–unergative distinction to lexical semantic properties (e.g., Levin & Rappaport-Hovav, Reference Levin and Rappaport-Hovav1995; Sorace, Reference Sorace2000). For example, core verbs involving change-of-location and change-of-state events (telic) are likely to be unaccusatives, while core verbs involving non-motional and motional processes (atelic) are likely to be unergatives (Sorace, Reference Sorace2000) within or across languages. Furthermore, intermediate verbs with a meaning of existence-of-state can shift classes in different constructions and exhibit more class variations within or across different languages.
However, we see below that this flexibility still cannot account for the following Mandarin examples, and their predictions of verb behaviours may overlook language-specific considerations. For instance, in Mandarin Chinese, the acceptability of some formal and double-syllable unaccusatives in the postverbal-subject diagnostic remains puzzling. Many double-syllable verbs, such as li2-kai1 (“leave-open”) “to leave” (Lu, Reference Lu2019:109), diao4-luo4 (“drop-drop”) “to drop,” and formal verbs like jia4-beng1 (“emperor-collapse”) “(emperor) to die” and jia4-dao4 (“emperor-arrive”) “(emperor) to come,” despite their telic meanings compatible with unaccusative characteristics, might sound odd in the postverbal-subject diagnostic. In contrast, their one-syllable and colloquial counterparts, such as zou3 “to leave,” diao4 “to drop,” si3 “to die,” and lai2 “to come,” are acceptable in the same diagnostic. The former types of verbs are often disregarded in many unaccusativity studies in Mandarin, but their lower acceptability in the diagnostic sentence may provide important insights into theories of unaccusativity.
In (8), we see that these verbs are not compatible with the durative test sentence, which suggests that these verbs are semantically telic, as the durative sentence can approximately test the telicity of such change-of-location and some change-of-state verbs (Liu, Reference Liu and Aranovich2007).
We also see that these verbs sound unnatural in the postverbal-subject diagnostic in (9).
In such cases, these verbs may be misclassified as unergative due to their lower acceptability within the postverbal-subject diagnostic. The flexibility in Sorace (Reference Sorace2000)’s framework does not straightforwardly account for this issue. These core verbs do not have the meaning of existence-of-state and should exhibit stable behaviours of unaccusatives. Also, according to Sorace (Reference Sorace2000), verbs that exhibit flexibility of being both unergative and unaccusative in different sentences should be acceptable in a diagnostic that singles out unaccusatives as well as in a diagnostic that singles out unergatives. However, the fact that these verbs are not accepted in the postverbal-subject diagnostic suggests that the unaccusativity is not consistently reflected.
Our findings may provide a preliminary explanation based on our behavioural experiment: when only atelic/telic semantics is provided, even though ratings can distinguish between unergatives and unaccusatives, the “accuracy” of unaccusative-targeted verbs is very low. A semantically telic verb requires occurrences in sentence types we labelled as “unaccusative,” such as LVS, resultative (2nd), and the postverbal-subject diagnostics themselves, to be recognized as an unaccusative syntactically in a diagnostic. Table 13 shows the occurrences of these verbs within sentence types in the largest Taiwanese Mandarin corpus, Sinica Corpus (n.d.). We found almost zero occurrences for the verbs occurring in LVS, and resultatives and postverbal-subject sentences, and most of these verbs occur in subject + verb constructions. These double-syllable and formal verbs rarely appear in sentences/phrases that reinforce the unaccusativity, so perhaps Mandarin speakers are not able to use the distributional information in these sentence types to categorize them as unaccusative.
Occurrences of examined verbs within various sentence types in the Sinica corpus
We propose that distributional information may be a crucial factor in how a verb is syntactically classified and categorized. Semantic clues from a verb alone, such as telicity and agentivity, may not guarantee the verb is successfully assigned to the verb class linked to its meaning. Instead, it is important to consider distributional factors, such as the frequency of verb occurrences and their presence in various sentence types in the language.
This insight invites a theoretical account cross-linguistically, particularly if accounting for most verb behaviours is the goal of the theories. When speakers of a language consistently place an intransitive verb with a change-of-location meaning (unaccusative) in non-unaccusative sentences but rarely in the constructions we labelled as unaccusative, this verb may not be classified by speakers as unaccusative in an unaccusative diagnostic, especially in languages like Mandarin. On the contrary, for a verb to be classified according to the verb class to which its semantics links, speakers need to experience substantial occurrences of the verb within various sentences of the same category; otherwise, the verb may be classified into the other verb class. Substantial occurrences in relevant constructions are required for a verb to be classified into the verb class its semantics links to in a diagnostic. Theoretical frameworks might benefit from considering the distributional factor, as this may help address certain verb behaviours in some languages. We acknowledge that this observation is preliminary and that enriching the content of theoretical frameworks requires more specifications, but we suggest that future studies consider including this distributional factor to enrich the framework on unaccusativity.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0305000926100531.
Data availability
The research materials and data for all experiments are available in the Open Science Framework repository: https://osf.io/e5smq/?view_only=7c66989920f74a9499b752eaa833f842
Acknowledgements
We express our gratitude to the Data Science Statistical Cooperation Center of Academia Sinica (AS-CFII-111-215) for statistical support. We also appreciate the Institute of Linguistics, Academia Sinica, for their support in publication. The author(s) gratefully acknowledge the grant support for CHILDES – NICHD HD082736.
Competing interests
The author(s) has/have no competing interests to declare.
Ethics and consent
This study was conducted in accordance with the ethical standards outlined in the Declaration of Helsinki and was approved by the Institutional Review Board (IRB) of the University of Hawaiʻi at Mānoa. Prior to participation, all subjects were informed about the purpose, procedures, potential risks, and benefits of the study. Written informed consent was obtained from all participants or, in the case of minors, from their legal guardians. Participants were assured of the confidentiality of their data, and their participation was voluntary, with the option to withdraw from the study at any time without any penalty.
Open access
