Word order flexibility affects complementizer omission: a cross-linguistic investigation

1 Introduction

Linguistic alternations have drawn the attention of researchers from various subfields in linguistics and have emerged as a topic that has been extensively studied over the past few decades (Gries 2017). Informally, a linguistic alternation can be defined as the phenomenon of language users choosing between “pairs of semantically more-or-less equivalent expressions” (Gries and Stefanowitsch 2004: 97). The variety of alternations in languages reflects a remarkable degree of flexibility in how we can encode information with linguistic forms or signals.

One pervasive type of alternation is termed reduction, which refers to the choice of language users between the (a) full and (b) more or less reduced forms or signals (Jaeger and Buz 2018). At the phonetic level, for example, syllabic duration in English has been reported to be inversely related to language redundancy (Aylett and Turk 2004). Morphologically, optional case-marking in Japanese has been argued to be affected by communicative pressures (Kurumada and Jaeger 2015). In the present study, our focus is on a cross-linguistic syntactic phenomenon: complementizer omission. It has been observed in many languages that the complementizer at the beginning of a complement clause (CC) can be optionally omitted (English: Roland et al. 2006; Jaeger 2010; Spanish: Yoon 2015; Russian: Zou and Lin 2024). An example for English is provided in (1):

A large number of studies have been conducted to investigate complementizer omission, and several accounts of language processing have been proposed to explain the phenomenon. Among the most influential accounts are the availability-based production account, the uniform information density account, and the ambiguity avoidance account. Apart from language processing mechanisms, complementizer omission has also been argued to be under the influence of grammaticalization.

The idea of availability-based production is based on the Principle of Immediate Mention, which states that “production proceeds more efficiently if syntactic structures are used that permit quickly selected lemmas to be mentioned as soon as possible” (Ferreira and Dell 2000: 299). Several variables have been proposed to be related to availability-based production, including (a) the coreferentiality between subjects in matrix and complement clauses and (b) the number of disfluencies at the CC onset. Specifically, a complementizer is more likely to be omitted when the matrix subject and the CC subject refer to the same entity, and when there is no sign of disfluency at the CC onset. In addition, an effect of the frequency of matrix verbs is also compatible with the availability account (Jaeger 2010): less frequent matrix verbs tend to have higher rates of using an overt complementizer. According to Jaeger, the high processing load associated with the production of the less frequent verbs may “spill over” to the CC onset, thereby leading to the production of an overt complementizer.

The uniform information density (UID) account has received wide support since its first proposal in Jaeger (2010: 25), where it is formulated as follows: “within the bounds defined by grammar, speakers prefer utterances that distribute information uniformly across the signal (information density).” Regarding the complementizer omission phenomenon, the UID account predicts that the higher the information density at the CC onset, the more likely language users are to produce an overt complementizer. In Jaeger (2010), the information density at the CC onset is estimated with the subcategorization preference of matrix verbs. In Wulff et al. (2018), the same hypothesis was tested with a different variable (i.e., surprisal) that measures information density at a finer granularity. In either case, the UID account was strongly supported: speakers tend to lower the information density by producing an overt complementizer if the CC onset is highly surprising.

Unlike the above-mentioned two accounts, mixed results have been reported in previous research on the ambiguity avoidance mechanism. Based on previous observations on linguistic reductions, Frazier (1985) proposed the Impermissible Ambiguity Constraint: constructions that would lead to ambiguities or misanalyses on every occurrence tend to be prohibited. Specifically for complementizer omission, it has also been argued that the use of that is tailored to avoid temporary ambiguities (Temperley 2003). One example showing how an overt complementizer can avoid temporary ambiguities is shown in (2), where the matrix verb know allows both CCs and direct objects. The omission of the complementizer can thus lead to a garden path, in which the CC subject the story is incorrectly understood as a direct object of the verb.

A strong pattern of complementizer omission in English that supports the ambiguity avoidance account was found in Elsness (1984): an overt complementizer is less likely to be used when the CC subject is a pronoun. It was pointed out that in English, unlike nouns, some pronouns have distinct nominative and accusative forms. Therefore, a pronominal CC subject is less likely to be mistaken for a direct object of the matrix verb. At the same time, the ambiguity avoidance account also predicts a difference in omission rates between cases where the embedded subject is a case-distinguishing pronoun and cases where it is not. However, the predicted difference was observed neither in Elsness (1984) nor in Ferreira and Dell (2000). In other words, the evidence for and against the account is still inconclusive and further exploration is needed. In Temperley (2003), it is argued that the general tendency to omit complementizers in cases with pronominal CC subjects is itself an ambiguity avoidance mechanism at the syntactic level.

Broadly, the three above-mentioned accounts all align with the principle of communicative efficiency. A large amount of evidence has been provided cross-linguistically for the argument that language users tend to act efficiently, saving effort for processing and articulation, and that language structure and use reflect this tendency (for review, see Gibson et al. 2019; Jaeger and Buz 2018; Levshina and Moran 2021). In Kachakeche et al. (2021), for instance, it was shown that communicative pressures influence the use of adjectives, such that prenominal languages use adjectives at a higher rate compared to postnominal languages. In previous research on complementizer omission, however, only one specific language (primarily English) was studied at a time, and only language-internal factors have been explored. Here, “language-internal” factors contrast with “language-level” factors. Regarding complementizer omission, it has been shown in many studies that within a given language, various “language-internal” factors play a role in the omission of the complementizer (e.g., verb frequency, length of the CC). Cross-linguistically, although languages differ greatly in their overall omission rates, the question of which “language-level” properties may have contributed to the difference has rarely been explored. To the best of our knowledge, no cross-linguistic investigation on complementizer omission has been conducted. In the present study, we revisit the ambiguity avoidance account by analyzing complementizer omission cross-linguistically. Our study attempts to investigate the cross-linguistic difference in omission rates and to check whether it can be explained by the ambiguity avoidance account at a general strategic level.

Our study is motivated by the following hypothesis: word order flexibility of languages can influence the tendencies of their users to avoid ambiguities. Since human languages differ in the amount of word order flexibility they permit, we expect different languages to show different degrees of tendencies to avoid ambiguities. For instance, compared with English, German has a more flexible word order. In cases where an overt complementizer is omitted in German, the listener may have more difficulty in determining the grammatical role of the noun following the matrix verb. In German, the noun that follows the matrix verb can not only be its direct object and a CC subject but also a CC object, as shown in (3):

Cacoullos and Walker (2009) offered a review of the treatment of the complementizer that in English in the prescriptive grammatical tradition and concluded that prescriptive grammarians’ main argument for retaining that is to ensure clarity. More generally, it can be argued that the use of overt complementizers (in any case) is itself an ambiguity avoidance mechanism. We propose that when the complementizer is omitted, instances of complex clauses in languages with a more flexible word order should be potentially more ambiguous than those in languages with a more fixed word order. Therefore, users of languages with more flexible word order should be more reluctant to omit complementizers. Our prediction can be formulated as follows:

Word order flexibility affects complementizer omission, such that languages with a less flexible word order have a higher tendency to omit complementizers.

In order to reliably test our prediction, we conducted two corpus-based analyses: an in-depth analysis of three languages, and a broad analysis of 23 languages from 13 genera (language branches). The first (in-depth) analysis includes three languages that clearly differ in their word order flexibility: English, German, and Russian. An example is provided for each of the latter two languages in (4)–(5). The in-depth analysis enables us to accurately extract relevant cases and to compare the omission rates of languages while controlling for other potential influencing factors.

The broad analysis is conducted with 23 languages from the Universal Dependencies (UD) Treebank (version 2.14; Nivre et al. 2020). In the broad analysis, we quantitatively estimated the overall omission rates of languages and their word order flexibility. Results from the two parts of analysis complement each other and provide strong support for our hypothesis.

The remainder of this paper is structured as follows. Section 2 focuses on the dataset, method, and results of the in-depth analysis and ends with a few of its limitations. Section 3 reports the broad analysis that attempts to address the limitations. In Section 4, we summarize our findings and discuss a few general issues related to the results and the choices made in our study. Section 5 concludes the study.

2 In-depth analysis of three languages

2.1 Dataset

We downloaded corpora of the three languages (one for each language) from the Leipzig Corpora Collection (Goldhahn et al. 2012; https://wortschatz.uni-leipzig.de/en). The collection contains corpora that were constructed from different sources, with different methods, and at different time points. For the sake of consistency, for each language, we chose a news corpus constructed through web-crawling in 2019 that contains one million sentences.

It has long been suggested that different lexical items often have different preferences for certain constructions (because of their semantic or information-structural characteristics) (Gries and Stefanowitsch 2004; Stefanowitsch and Gries 2003). The lexical effect on complementizer omission has been reported to be significant in many studies (Cacoullos and Walker 2009; Jaeger 2010). For instance, in Dor (2005), it was mentioned that the omission of the complementizer is unnatural under manner-of-speaking verbs in English. In the present study, we assume that lexical semantics have the same effect direction in different languages. More specifically, if manner-of-speaking verbs disprefer complementizer omission in English, then so do their translational equivalents in other languages. In order to minimize the effect of verb semantics, we first chose two English verbs for analysis, and then got their translational equivalents from the other two languages. For English, we chose know and believe, which are the third and sixth most frequent verbs in the sample used in Jaeger (2010). Compared with the top two most frequent verbs (i.e., think and guess), know and believe much more actively participate in the alternation. Their equivalents are wissen and glauben in German, and znat’ and verit’ in Russian. Importantly, all six verbs allow both CCs and direct objects (the Russian verb verit’ takes nouns in the dative case, while the other five verbs take nouns in the accusative case). In other words, for all the chosen verbs, temporary ambiguities will arise when the complementizer is omitted.

For each language, to extract the target complex sentences with or without a complementizer, we first picked out sentences containing an inflected form of the target verbs. For instance, we considered 11 forms of the Russian verb znat’ (10 inflected forms and one infinite form; a list of the considered forms is provided in the Appendix). Participle forms in Russian (e.g., znajuščij) are not considered due to their extremely low frequency.

We then conducted part-of-speech (POS) tagging and dependency parsing on the picked-out sentences with the Stanza package (Qi et al. 2020) in Python. For simplicity, we only included declarative sentences (ending with a full stop) where the optional complementizer is used directly after the target verbs. As reported in Jaeger (2010), the matrix verb directly precedes the CC in 93.5 % of cases in his sample. Therefore, it can be estimated that only a small proportion of cases would be excluded from our analysis. An example of the English annotated target sentences is provided in Figure 1:

Figure 1:

An English complex sentence with an overt complementizer, annotated with POS and dependency information.

In our analysis, a complex sentence is tagged as having a complementizer if it satisfies the following conditions: (a) one of the target verbs is the root of the sentence; (b) the target verb is the head of the dependency relation ccomp; (c) the sentence has at least two subjects (one occurring before the matrix verb and in the matrix clause (MC), and the other occurring after the matrix verb and in the CC; the dependency relation with the main verbs in the respective clauses being nsubj); (d) the word following the target verb is a subordinating conjunction. On the other hand, a complex sentence is tagged as having no overt complementizer if it satisfies conditions (a), (b), and (c), but does not contain a subordinating conjunction. Items from the three languages were extracted with the same rules. For condition (d), to reduce the number of false hits, we specified the exact complementizer (subordinating conjunction) in the three languages: that in English, dass in German, and čto in Russian. Finally, we excluded sentences with interrogative pronouns as CC subjects. In total, our dataset contains 5,951 items. For each language, the number of items with and without a complementizer is presented in Table 1. The number of cases with each verb is shown in Table 2. It can be seen that the subcategorization bias of the verbs, which is calculated by dividing the number of target cases by their raw lemma frequency, is consistently low.

Table 1:

The number of extracted items for each language. The first and second columns indicate the number of items with and without an overt complementizer, respectively. The last column provides the omission rate.

Language	Present	Absent	Sum	Omission rate
English	1,290	2,835	4,125	0.687
German	569	411	980	0.416
Russian	609	237	846	0.280

Table 2:

Raw frequency, number of usages as matrix verbs, subcategorization preference, and estimated tendency for complementizer omission of the target verbs.

Verb	Raw frequency	# of target cases	Subcategorization
know	17,065	2,269	0.133
believe	7,248	1,856	0.256
wissen	5,028	521	0.104
glauben	2,819	459	0.163
znat’	6,610	679	0.103
verit’	1,264	167	0.132

2.3 Results

We used a mixed-effects logistic regression model to test the effects of our predictors with the lme4 package (Bates et al. 2015) in R (R Core Team 2023). The response is a binary variable with two levels: present and absent. The fixed-effects structure (FES) of the model contains our main predictor (i.e., language) and five controls, as introduced above. Length of CC and frequency are logged, and all the continuous variables are centered and scaled before being included in the model. Additionally, we included verb, which has 6 levels, into the random-effects structure (RES) of our model by adding by-verb random intercepts. Issues of multicollinearity and overdispersion are checked with helper functions provided in Gries (2013).

Results of our regression analysis indicate that all the predictors have a significant effect on complementizer omission. No issue of multicollinearity or overdispersion was found. Coefficients and results of significance tests for predictors are presented in Table 3. Marginal R² and Conditional R² of the model are 0.236 and 0.251, respectively.

Table 3:

Summary of results. Coefficient estimates, standard errors, z scores, and p values for all the predictors in the analysis.

	Estimate	SE	z score	p value
(Intercept)	−1.162	0.202	−5.751	<0.001
length	0.309	0.032	9.783	<0.001
surprisal	0.283	0.030	9.490	<0.001
frequency	−0.477	0.147	−3.234	0.001
language (en → de)	0.770	0.364	2.115	0.034
language (en → ru)	1.520	0.375	4.051	<0.001
CC subject (pronouns → nouns)	0.346	0.064	5.423	<0.001
matrix subject (first singular → other pronouns)	0.527	0.075	7.010	<0.001
matrix subject (first singular → nouns)	0.216	0.075	2.893	0.004

As predicted by the ambiguity avoidance account, there was a significant effect of language on complementizer omission. Specifically, the predicted probability of complementizer omission is the lowest in English and is significantly lower than that in German ( β ˆ  = 0.770, z = 2.115, p = 0.034). The probability of omission in German is significantly lower than that in Russian ( β ˆ  = 0.751, z = 2.536, p = 0.011). The effect holds even while other variables are controlled for. Figure 2 illustrates the effect of language. In other words, the language with the least flexible word order omits the complementizer most frequently, and the language with the most flexible word order omits the complementizer least frequently.

Figure 2:

Effect of the predictor language on the predicted probability of using an overt complementizer. Error bars denote 95 % CIs.

We then used likelihood ratio tests (LRTs) to check the contributions of predictors in our model. Since the variable verb in the RES contains language-related information, in order to compare the relative contribution of language to that of the other controls, we needed to refit a model with no RES. Results of LRTs indicate that language is the strongest predictor in terms of its contribution to the likelihood of the model (χ²(2) = 275.600, p < 0.001). Moreover, the improvement in model quality by including language is much higher than the improvement by including any of the controls. In other words, there exist large cross-linguistic differences in terms of complementizer omission, and their effects are much larger than those of language-internal factors. Word order flexibility at the cross-linguistic level emerges as the most important predictor of complementizer omission.

Next, we briefly discuss effects of the control variables in the model. As shown in Table 2, all the control variables have a significant effect on complementizer omission in the expected direction. First, the predicted probability of complementizer omission is significantly higher when the matrix subject is a first-person singular pronoun (I in English, ich in German, and ja in Russian) than when it is some other pronoun ( β ˆ  = 0.527, z = 7.010, p < 0.001) or noun ( β ˆ  = 0.216, z = 2.893, p = 0.004). The observed effect aligns with the grammaticalization account, according to which the frequent collocation of first person singular pronominal subject and complement-taking verb has been reanalyzed as an epistemic phrase that does not take subordinate CCs (Thompson and Mulac 1991). Second, the length of the CCs is strongly correlated with lower omission rates ( β ˆ  = 0.309, z = 9.783, p < 0.001). Longer CCs prefer the use of an overt complementizer. Similar results have also been reported in Jaeger (2010) and Gries (2021). As mentioned in Jaeger (2010), the significant effect of length of CC is surprising because it is not directly predicted by accounts of language processing. Moreover, it seems to contradict with the experimental evidence that language production is radically incremental (Brown-Schmidt and Konopka 2008). The observed effect of length of CC in our study provides further support for the explanation in Jaeger (2010: 42) that language users seem to have “at least heuristic weight or complexity estimates of material that is not yet phonologically encoded.” Third, the effect of surprisal ( β ˆ  = 0.283, z = 9.490, p < 0.001) indicates that the predicted probability of using an overt complementizer increases as the degree of surprisal at the CC onset increases, thereby providing support for the UID account. The effect of frequency ( β ˆ  = −0.477, z = −3.234, p < 0.001) shows that the higher the lemma frequency of the matrix verb, the more likely the complementizer is to be omitted, which is also in line with the grammaticalization account. Finally, the effect of the CC subject ( β ˆ  = 0.346, z = 5.423, p < 0.001) is compatible with both the ambiguity avoidance account and the availability account.

In sum, the comparison of omission rates across three languages provides initial support for our hypothesis. However, it has a few obvious limitations. First, although the choice of semantically equivalent verbs across languages avoids the potential influence of lexical semantics, it runs the risk of misrepresenting the language-level overall omission rates. For instance, it is potentially possible that the two Russian verbs chosen in the analysis happen to have the lowest omission rates in the language. In other words, the observed difference in omission rates could be an artefact of our choice of verbs. Second, the number of languages included in the analysis is small. It could be the case that the three languages just happen to differ in their omission rates in the expected direction. Third, there is a difference between potential and effective flexibility. Although Russian is theoretically more flexible in word order than German, their actual flexibility still needs to be determined empirically. In any event, in order to gather more robust evidence for our hypothesis, we think it is important to include more languages, to compare their overall omission rates, and to quantitatively correlate omission rates with word order flexibility. In the following section, we attempt to address these concerns through our analysis of 23 languages from the UD Treebank.

3 Broad analysis of the UD Treebank

3.1 Dataset

We used the UD Treebank, which contains 283 treebanks from 161 languages to test our hypothesis at a broad level. If a certain language has several treebanks, we group them into a larger one. The treebank is developed with cross-linguistically consistent POS and dependency annotations, so the target structures can be easily extracted with a Python script.

A few steps need to be taken to decide on the exact structures to analyze and the languages to include. First, we found that we cannot rely on dependency structures and POS tags to exclusively extract all the sentences containing a CC without an overt complementizer when the CC object is a pronoun. It can be seen in Figure 3 that, when the CC starts with an interrogative pronoun who, the entire complex sentence has the exact same annotations as the one in Figure 4. Therefore, to circumvent this issue completely, we decided to only include in our broad analysis sentences where the CC subject is a noun. Second, we wanted to exclude the languages that do not allow complementizer omission. Ideally, languages with a 0 or 100 percent omission rate should be excluded. However, our extraction method, which is described below, still returns some false hits. For example, our analysis shows that Cantonese, which does not have a complementizer similar to that in English, has an omission rate of 0.96. French, which does not allow complementizer omission (in its standard variety), has an omission rate of 0.06. Therefore, we only included in our analysis languages with an omission rate between 0.1 and 0.9. In addition, Arabic, which also does not allow omission, is excluded. Finally, we excluded languages containing too few target structures (less than 20) and ancient languages (Berdicevskis et al. 2018). After taking the above-mentioned steps, our sample contains 23 languages from 13 genera.

Figure 3:

An English complex sentence with an interrogative pronoun as the CC subject, annotated with POS and dependency information.

Figure 4:

An English complex sentence without an overt complementizer, annotated with POS and dependency information.

3.3 Results

To test our prediction across 23 languages, we fitted a linear mixed-effects regression model predicting the omission rate by word order entropy, with random intercepts by language family and genus. Information about language families and genera was taken from WALS Online (https://wals.info). As predicted and shown in Figure 5, we find a significant effect of word order entropy ( β ˆ  = −0.457, p = 0.010). Cross-linguistically, the predicted probability of the omission rate of a language tends to be lower if its word order entropy is higher. Importantly, the observed effect remains robust when we controlled for genealogical relations across languages and, therefore, provides strong support for our hypothesis.

Figure 5:

Relationship between word order entropy and omission rate across 23 languages in the UD Treebank. Each point represents a language, with point size denoting the number of target cases and point color denoting language genus. The correlation line and the confidence interval take into consideration genealogical relations between languages.

4 General discussion

The primary purpose of this study is to investigate cross-linguistic differences in complementizer omission and to test if the differences across languages can be explained in terms of their word order flexibility. Our hypothesis of ambiguity avoidance predicts that language users should show a higher preference for omission if the word order in their languages is more fixed and is thus less likely to cause ambiguities. As shown in our two analyses, the effect is indeed observed.

In previous studies on complementizer omission, only the effects of language-internal factors have been investigated. The present study goes beyond them by expanding the empirical base and exploring cross-linguistic differences. Each of the two analyses conducted in our study has its advantages and provides complementary evidence for our hypothesis. The in-depth analysis is based on large corpora of three languages and thus offers accurate estimates of the omission rates of the target verbs from the three languages. In extracting target cases, we relied on our knowledge of these languages, in addition to POS and dependency annotations, to improve our accuracy. For instance, we stipulated that the complementizer in Russian has to be čto (rather than other complementizers that cannot be omitted, like čtoby). Moreover, the analysis controlled for factors that have been reported to have an effect on complementizer omission, so that we can ensure that the language-level difference we observed was not due to their effects. In the in-depth analysis, we used language itself as a proxy for word order flexibility in the clause and found that the language with the least flexible order (English) has the highest omission rate, and that the language with the most flexible order (Russian) has the lowest omission rate. The results, therefore, constitute initial evidence that complementizer omission is affected by language users’ syntactic strategies to avoid ambiguities.

In our broad analysis, we addressed a few limitations of the first analysis by quantitatively estimating the word order flexibility and the complementizer omission rate of 23 languages. As mentioned above, the POS and dependency annotations do not allow us to filter out all the unwanted cases, so the accuracy of estimated omission rates would be lower. However, it has the advantages of including languages from various families and quantitatively estimating their word order flexibility with an entropy measure. As predicted by the ambiguity avoidance hypothesis, we found a significant negative correlation between word order flexibility and complementizer omission rate even after we controlled for genealogical relations between languages. Combining results from the two analyses, we, therefore, find strong support for the hypothesis.

According to Temperley (2003), ambiguity avoidance can be observed from three aspects. First, ambiguity avoidance could play a role in the formation of general syntactic principles. For instance, it was suggested in Bever (1970) that the (mandatory) requirement of a relative pronoun in subject relative clauses is due to the fact that such clauses would be highly ambiguous without an overt relative pronoun. See the contrast in (7) (Temperley 2003):

Second, ambiguity avoidance could influence syntactic strategies. One example of such strategies regarding complementizer omission is that language users show a preference for an overt complementizer when the CC subject is a pronoun (Elsness 1984). Third, ambiguity avoidance could come into play in a highly situation-specific, ad-hoc fashion. At the third level, whether or not a complementizer is needed to avoid ambiguity should be discussed case by case. One makes a decision by taking into account various (syntactic, semantic, and pragmatic) factors in a specific context.

In our study, the observed difference in omission rates across languages is attributed to the effect of word order flexibility and is in line with the ambiguity avoidance account at the second level (strategic ambiguity avoidance). Although some previous studies seemed to have failed in identifying the expected effects of ambiguity avoidance, our finding that word order flexibility can affect complementizer omission does not contradict with their results. In Elsness (1984), a prediction was made on the basis of the ambiguity avoidance account that the English complementizer that would be used more frequently with you as the subject of the CC than with other case-distinguishing pronouns. Yet no significant difference was detected. Similarly in Ferreira and Dell (2000), no specific preference for an overt complementizer was found in potentially ambiguous sentences, again casting doubt on the ambiguity avoidance account. However, as also pointed out in Temperley (2003), their studies mainly focused on specific lexical syntactic strategies (you vs. I and she), and the general tendency to omit complementizer in cases of a pronominal subject of the CC reported in Elsness (1984) can already be seen as an ambiguity avoidance mechanism. Further supporting evidence for the ambiguity avoidance account at the general syntactic level was provided in the analysis of the use of relative pronouns in Temperley (2003). In comparison, the effect of word order flexibility detected in our study is at an even higher and cross-linguistic level, which is in line with the claim that it is more likely to observe ambiguity avoidance mechanisms at general strategic levels than at specific and situational levels. The use of ambiguity avoidance strategies is much more feasible at the general syntactic level, especially if “their conditions for their application are defined in a fairly simple way (Temperley 2003: 482).”

Broadly, our findings are compatible with the communicative efficiency hypothesis, according to which languages are efficient in both production and perception to meet the communicative needs of their users (Jaeger and Buz 2018; Levshina and Moran 2021). The availability account was proposed as an account for efficient production. Previous research has argued that the effects of several language-internal factors, including the frequency of matrix verbs (Jaeger 2010), the type of subjects in complement clauses (Cacoullos and Walker 2009), and the coreferentiality between subjects in matrix and complement clauses (Elsness 1984), can be explained with the availability account. In our study, we attributed the effect of word order flexibility to the ambiguity avoidance account, which is related to efficient communication on the perception side. Ambiguity avoidance is regarded as a type of audience design, which refers to “when speakers fashion their utterances so as to cater to the needs of their addressees” (Ferreira 2019: 29). It should be noted that the observed effect in our study does not indicate that users of a certain language will choose to use an overt complementizer in a specific ambiguous situation. Rather, our results indicate that in order to avoid ambiguities, a general tendency (or strategy) is developed at the language level. Different languages thus show different tendencies. In brief, our results align well with the hypothesis that complementizer omission is not only conditioned by various factors to facilitate the production of language users but also to ensure the understanding of their interlocutors (or readers in the context of writing).

The idea that language-level properties can influence language use has been well documented in previous research (Jaeger and Buz 2018; Levshina and Moran 2021). For example, in Rubio-Fernandez et al. (2021), it is reported that speakers of English, which is a prenominal language (adjectives come before nouns), use more redundant adjectives than Spanish, which is a postnominal language. Kachakeche et al. (2021) quantified the propensity of languages to use adjectives prenominally and found that, across 74 languages, the ones that favor prenominal adjectives indeed exhibit higher rates of adjectival modification. Through miniature artificial language learning experiments, Fedzechkina et al. (2017) and Fedzechkina and Jaeger (2020) showed that learners tend to drop optional case markings in the fixed order language but retain them in the flexible order language. However, in previous studies on linguistic alternations, the direction of alternations has rarely been associated with language-level properties. In future research, it would be interesting to investigate other alternations cross-linguistically and to see if there is any systematic difference that can be explained with language-level characteristics.

5 Conclusions

In the present study, we proposed an account of ambiguity avoidance and made a prediction about complementizer omission at the cross-linguistic level: in order to avoid ambiguities, users of languages with more flexible word order should be more reluctant to omit the complementizer. The prediction was confirmed with an in-depth analysis of three languages (English, German, and Russian) and a broad analysis of 23 languages from 13 genera. To the best of our knowledge, our study is the first to explore cross-linguistic differences in complementizer omission. Our findings support the claim that language users have a strategic (cross-linguistic) tendency to produce optional linguistic forms when it helps to avoid ambiguities.

In conclusion, our study provides replicating and novel supporting evidence that language-processing mechanisms have a major influence on syntactic reduction. The observed effect of word order flexibility adds to the literature on ambiguity avoidance and is compatible with the hypothesis that language users consider the knowledge and processor state of their interlocutors (or readers in the context of writing) and adjust their language production accordingly in an attempt to successfully transfer the intended information. Ultimately, our findings on complementizer omission align well with the general hypothesis that language production is organized to transfer information efficiently.