The diachronic change of English relativizers: a case study in the State of the Union addresses across two centuries

2.1 Data

This study utilized the corpus of the State of the Union (SOTU) addresses delivered by American presidents. The dataset has been widely used in previous research concerning the change of linguistic features (e.g., Lei and Wen [2020] and Burgers and Ahrens [2020]). The corpus is composed of 232 SOTU addresses delivered by 44 presidents,^[1] from President George Washington’s first address in 1790 to President Joseph Biden’s address in 2023.^[2] It consists of a total of 2,079,563 tokens and 74,174 sentences, with an average of approximately 28 tokens per sentence.

The corpus was chosen for our study for the following reasons. First, the timespan of the SOTU addresses covers 234 years, which provides sufficient temporal coverage to observe the possible evolution of linguistic features. In addition, the one-year intervals enable a fine-grained analysis of potential change trajectories. Second, all texts are freely available, and have already undergone proofreading prior to release (Lei and Wen 2020; Savoy 2015). Third, it should be noted that despite the comparable registers, i.e., political speeches (Burgers and Ahrens 2020), the delivery method of the SOTU addresses have varied. As shown in Figure 1, before 1933, the SOTU messages were primarily written reports. After President Franklin D. Roosevelt’s first speech in 1934, the modern practice of delivering spoken addresses began. That is, from 1934 onwards, the SOTU addresses can be considered primarily as oral speeches rather than written reports.^[3] Given that the method of delivery (spoken vs. written) may affect the use of relativizers (Biber 1986; Romaine 1982), the factor of the delivery method was considered in our study, which may provide us an opportunity to examine the effect of genres (i.e., written reports vs. oral speeches) on the changes of relativizers.

Figure 1:

Delivery methods of SOTU (1790–2023).

2.2 Extraction of relative clauses: a new method^[4]

The corpus texts underwent initial tokenization using Natural Language Toolkit (NLTK) and then parsing with SuPar. SuPar is a highly efficient dependency parser, which can help parse a text with an accuracy of 93.37 % for the unlabeled attachment score (UAS) and that of 91.27 % for the labeled attachment score (LAS) (Zhang et al. 2020). It parses the texts and identifies 37 universal syntactic relations between words in a sentence based on the Universal Dependency (UD) Scheme v2 (De Marneffe et al. 2021). To extract relative clauses from our data, we identified sentences containing the relation acl:relcl, which refers to the finite clause modifying a noun phrase (Chen et al. 2021). For example, in sentence (1), the relative clause that I proposed was successfully detected by the parser, represented by the subtree of the label acl:relcl (see Figure 2 for the dependency relations of word pairs in the sentence).

Figure 2:

Dependency relations of sentence (1).

Our proposed method also effectively detected zero relative clauses. For example, in sentence (2) where the relativizer is omitted, the acl:relcl relation between saw and man was accurately identified (see Figure 3). Therefore, regardless of the presence of relativizers, all possible relative clause candidates can be extracted from the texts as long as sentences with the acl:relcl relations were identified.

Figure 3:

Dependency relations of sentence (2).

After the extraction of the relative clauses, we also extracted other linguistic features, i.e., the head nouns, the clause verbs, the relativizers, the restrictiveness of relative clauses, the role of head nouns, and the animacy of head nouns. The flowchart for the extraction of candidate relative clauses and linguistic features is depicted in Figure 4.

1. Head noun and clause verb: the dependency relation acl:relcl connected the head noun and the clause verb. The identification of the dependency relation acl:relcl allowed for the extraction of both head nouns and clause verbs.

2. Relativizer: we extracted the relativizers by matching them with a predefined list of known relativizers. The list includes which, whichever, that, who, whom, whose, whoever, whomever, however, when, whence, whenever, where, wherein, whereon, whereof, whereby, wherever, why, what, whatever, whatsoever, than, as, and in case. To ensure accuracy in our analysis, we required that the identified relativizers directly connect with the clause verbs. For example, in “…programs which that president proposed”, only which was extracted, because that functioned as a pronoun referring to president. In instances where no relativizers were identified, the absence was considered a relativizer omission. Consequently, the relativizers were designated as zero.

3. Restrictiveness of relative clauses: the restrictiveness of a relative clause was determined based on the presence or absence of a comma separating the relativizer from the head noun. A comma indicated a non-restrictive relative clause, while its absence suggested a restrictive clause.

4. Role of head nouns: the roles of head noun were extracted by means of the dependency relation of relativizers (e.g., nsubj, nsubjpass, obj), because the relativizer assumed the role of the head noun in the relative clause.

5. Animacy of head nouns: The head noun was assigned one of the binary animacy values (animate or inanimate) by referring to an animacy dictionary (Ji and Lin 2009). Words lacking entries in the dictionary, such as pronouns (e.g., those, all, and some) and infrequent words (e.g., Qaida), were manually checked and assigned animacy values.

Figure 4:

Flowchart for the extraction of candidate relative clauses.

After extracting all candidate relative clauses, we identified prototype relative clauses, defined as definite and restrictive relative clauses (Comrie 1989 [1981]; Givón 2001 [1984]: 24). Non-restrictive relative clauses were excluded due to their exclusive use of wh-relativizers. Among prototype relative clauses, we also excluded adverbial relative clauses and nominal relative clauses because this study did not consider adverbial relativizers^[5] (e.g., when, where, why) or nominal relativizers (e.g., what). Cleft sentences were also excluded because they only employed that. After the exclusion of relative clauses of the aforementioned types of clauses, our study narrowed its focus to subject and object relative clauses.

We restricted the analysis to inanimate head nouns, since animate head nouns employ zero and that but not which. For subject relative clauses, we focused on the use of which and that as in the sentence (3), since zero relativizers are not allowed to introduce subject relative clauses. For object relative clauses, we focused on the use of which, that and zero as in the sentence (4). The selection of subject and object relative clauses was driven by two considerations. First, they have been extensively examined and constitute the largest proportion of all relative clauses according to noun phrase accessibility hypothesis (Comrie 1989 [1981]). Our study was also not an exception with subject relative clauses accounting for 56.34 % and object relative clauses for 21.93 % of all relative clauses. Second, only under such circumstances could these relativizers be used interchangeably without grammar restriction (Sauerland 2003). The interchangeability allows for the observation of the natural changes in relativizers, facilitating the detection of other diachronic effects.

The detailed procedures for exclusion are listed in Table 1.

To validate our proposed method, we manually checked the recall rate and the precision value of the extracted relative clauses. First, we selected six sample texts from the SOTU corpus (1,237 sentences, 27,805 tokens). The samples were randomly chosen from various periods of time, with one in every half a century, namely 1792, 1809, 1871, 1927, 1965, and 2002. Next, two of our authors independently read the texts and manually identified the relative clauses. Any discrepancies in manual identification were discussed and resolved to reach an agreement. Our manual identification yielded 140 eligible relative clauses, while our proposed method automatically extracted 143 relative clauses. The proposed method missed one relative clause and mistakenly extracted four relative clauses, with a recall rate of 99.29 % and a precision value of 97.20 %.

Meanwhile, we manually verified the accuracy of the extracted features in sample texts. The extraction of head nouns, relativizers, roles of head nouns in relative clauses (e.g., subject and object), restrictiveness (restrictive vs. nonrestrictive relative clauses), and clause verbs all achieved a precision value of 100 %. The detection of animacy achieved a precision rate of 97.84 % when relying solely on the animacy dictionary (Ji and Lin 2009). The precision increased to 100 % when incorporating both dictionary-based assignment and manual annotation of words not found in the dictionary. In conclusion, the manual verification of the automatic extraction confirmed the reliability and validity of our proposed method.

2.3 Data processing and analysis^[6]

We processed and analyzed the data in the following steps. The data processing and analyses of the data were conducted with the programming language R. First, we calculated the proportions of which and that in subject relative clauses per year and the proportions of which, that, and zero in object relative clauses per year. The proportions from 1790 to 2023 were then visualized to illustrate the trends of relativizers within subject and object relative clauses. Then, simple linear regression (Wen and Lei 2022) was used to assess whether the proportions of relativizers were subject to significant changes over the two hundred years.^[7]

Second, to investigate the factors that may affect the use of relativizers, we identified 11 factors based on prior research, including studies by Hinrichs et al. (2015) and Grafmiller et al. (2018). The factors can be categorized into three types, i.e., those concerning linguistic complexity, those pertaining to text styles, and other external factors (see Table 2). To be specific, factors such as relative clause length, filler-gap dependency distance, and head noun length were the complexity indicators of relative clauses (Fox and Thompson 2007; Tagliamonte et al. 2005). Preceding relativizers are related to structural persistence or priming, which affects the ease of production (Szmrecsanyi 2006). Mean word length and mean sentence length serve as context indicators, with the former reflecting lexical complexity and the latter measuring syntactic complexity (Flesch 1948; Lei and Shi 2023). Text style factors included type-token ratio reflecting lexical richness (Biber 1986) and personal pronouns indicating involved or less formal styles (Biber 1986). It is worth noting that mean word length also pertains to text styles, as longer words often convey more specific, specialized meanings than shorter words (Zipf 1949). Precise lexical choice is not easy and rarely fully accomplished in speech (Chafe and Danielewicz 1987), leading to stylistic variation in mean word length (Biber 1986).

To investigate the factors that may affect the use of relativizers, we conducted a series of simple logistic regressions, examining the significance and effect size of each independent factor following the steps of minimal adequate model.^[8] Factors demonstrating significance value below 0.05 (p < 0.05) were considered as relevant factors and included in the subsequent multivariate modelling.

We employed a mixed-effects multivariate logistic regression analysis to predict the use of which and that in subject relative clauses and the use of which/that and zero in object relative clauses.^[9] It should be noted that although object relative clauses exhibited a ternary alteration (which, that, and zero), we aggregated which and that to examine the alteration between which/that and zero. The method of the aggregation of which and that employed in this study aligns with established methodologies outlined in previous research by Hinrichs et al. (2015). It also facilitated the distinction between explicit and zero relativizers. To ensure comparability across factors, the factors relative clause length, mean sentence length and personal pronouns were logarithm-converted. The transformation can improve the interpretability of the models and prevent variables from dominating others due to larger eigenvalue ratios (Tomaschek et al. 2018). During the modelling process, the external factor year was excluded due to its strong correlation with other factors (e.g., relative clause length, filler-gap dependency distance, and mean sentence length), which introduced multicollinearity issues. After the multivariate modelling, we analyzed the evolution of predicative factors to determine the effect of time on these factors and their subsequent effects on relativizers.

Third, we examined whether the rate of the change is consistent throughout the time spans or varies in different periods of time. To identify possible abrupt changes, we employed the method of change point detection, i.e., the Pettitt test for single change point detection (Lanzante 1996; Shi and Lei 2022). The Pettitt test was a widely used statistical test applicable across various disciplines for the detection of significant change points or abrupt changes in time-series data (Verstraeten et al. 2006).

3.1 Diachronic change of relativizers

To investigate the diachronic change of relativizers, we first analyzed the use of which and that in subject relative clauses, and then analyzed that of which, that and zero in object relative clauses.

Table 3 presents the proportions of relativizers in subject and object relative clauses. The results showed that approximately two thirds of the subject relative clauses were introduced by which, while approximately a half of the object relative clauses were zero relative clauses.

The proportions of which and that in subject relative clauses from 1790 to 2023 are shown in Figure 5. The linear regression analysis revealed a significant decrease in the use of which (t = −24.18, F(1,230) = 584.8, p < 2.2e⁻¹⁶) and a significant increase in the use of that (t = 24.18, F(1,230) = 584.8, p < 2.2e⁻¹⁶) in subject relative clauses. The results suggested that that gradually replaced which in subject relative clauses in the State of the Union (SOTU) addresses.

Figure 5:

Diachronic change of relativizers in subject relative clauses. The dots represent the proportion of which (blue) and that (red) used in the subject relative clauses. The solid line represents smoothed trends of the dots using the locally weighted scatterplot smoothing, with the shadow indicating the 95 % confidence interval.

The proportions of which, that and zero in object relative clauses from 1790 to 2023 are reported in Figure 6. The results show a significant downtrend of which (t = −16.59, F[1,229] = 275.4, p < 2.2e⁻¹⁶). In contrast, upward trends were observed for that (t = 8.84, F[1,229] = 78.09, p < 2.2e⁻¹⁶) and zero (t = 10.58, F[1,229] = 111.9, p < 2.2e⁻¹⁶) in object relative clauses. By the end of the observed period, zero emerged as the dominant relativizer in place of which and that in object relative clauses in the SOTU addresses.

Figure 6:

Diachronic change of relativizers in object relative clauses. The dots represent the proportion of which (blue), that (red) and zero (green) used in the object relative clauses. The solid line represents smoothed trends of dots using the locally weighted scatterplot smoothing, with the shadow indicating the 95 % confidence interval.

A comparison of the proportions of zero relative clauses at two time points, 1960 and 1990 revealed inconsistencies with Hinrichs et al.’s (2015) findings. While their dataset demonstrated stability in the proportions of zero (69 % in 1960 vs. 70 % in 1990), our dataset showed an increase (57 % in 1960 vs. 88 % in 1990). Notably, our data revealed fluctuations in the use of zero relative clauses from 1960s to 1990s, characterized by a plunge from 1961 to 1969 followed by a rebound (see Figure 7). The discrepancy of Hinrichs et al.’s (2015) and our findings underscored the importance of a finer-grained investigation. Without it, the complex changes in the use of relativizers would have remained unnoticed (Lee 2020; Leech et al. 2009).

Figure 7:

The fluctuation in the use of zero relative clauses from 1960s to 1990s. The dots represent the proportion of zero used in the object relative clauses.

3.2 Factors affecting the diachronic change of relativizers

The results of the mixed-effects multivariate logistic regression model concerning subject relative clauses are summarized in Table 4. The model achieved an 80.6 % prediction accuracy of relativizers with a concordance value (C) of 0.81. The result indicated the model’s ability to discriminate which and that in subject relative clauses (Hinrichs et al. 2015). The inclusion of mean word length in the multivariate model was favored due to a strong negative correlation (R = −0.82) between mean word length and personal pronouns.^[10] Personal pronouns were subsequently excluded from the model, resulting in improved predictive performance. In addition, the variance inflation factors^[11] of all factors were lower than 1.01, well under the threshold 5 for detecting the collinearity issues (Tomaschek et al. 2018). Factors without a significant value were excluded in the model. It should be noted that although the delivery method of the addresses could predict relativizers in the simple logistic regression (z = 37.28, p < 2e⁻¹⁶), it failed to reach statistical significance (p = 0.99) in the multivariate model. A possible explanation is that the effect of delivery methods was diminished by indirect factors such as mean word length. In other words, its effect may be mediated or affected indirectly by its relationship with mean word length (MacKinnon et al. 2007).

To summarize, relativizer alternations (which vs. that) in subject relative clauses were affected by factors related to linguistic complexity and text styles. First, factors such as relative clause length, filler-gap dependency distance, and head noun length are indicators of linguistic complexity of relative clauses. Longer length or greater distance indicated higher complexity of relative clauses, leading to a greater probability of using the explicit relativizer which. Second, preceding relativizers served as an indicator of structural priming and persistence (Szmrecsanyi 2006). It was of interest to find that the use of that predicted subsequent use of that, but which failed to predict itself. The finding suggested that the effect of structural priming remained for that, but was mitigated by other complexity factors for which. For example, in cases where the relative clause length was short enough to use that, which may not be used irrespective of the preceding relativizer. Zero did not exhibit a priming effect, consistent with findings in syntactic priming literature (Gries 2005; Hinrichs et al. 2015). Third, mean word length served as an indicator of lexical complexity within contexts, affecting the use of which and that. The reason may be that mean word length not only reflected lexical complexity (Flesch 1948) but also demonstrated stylistic variation given its strong correlation with personal pronouns.

The findings of the mixed-effects multivariate logistic regression model in object relative clauses were presented in the Table 5. The model achieved an 82.3 % correct prediction rate and a concordance value of 0.82 in the use of relativizers in object relative clauses. The factors affecting relativizer alternations (which/that vs. zero) were related to linguistic complexity. For example, relative clause length and the filler-gap dependency distance were the measures of the complexity of relative clauses. Longer lengths and distances were associated with the use of explicit relativizer (which/that) and discouraged the use of zero. In addition, the significance of mean sentence length, an indicator of the overall complexity of context, suggested that contextual complexity also predicted the use of relativizers. Contexts characterized by higher linguistic complexity were likely to use explicit relativizers and prevent the use of zero.

Diachronic changes of the predicative factors are listed in Table 6. Factors concerning linguistic complexity, i.e., relative clause length, filler-gap dependency distance, head noun length, mean word length, and mean sentence length had undergone a decreasing trend. The trend suggested that language simplification occurred in terms of various perspectives of speeches in the SOTU addresses (Lei and Wen 2020). Besides, factors associated with language styles such as mean word length showed a decreasing trend, particularly after the 1950s. The trend aligned with the general trend in English towards colloquialization (Leech et al. 2009). To be specific, the trend toward a more colloquial style affected factors such as mean word length, and potentially increased the use of zero relativizers. Also, the use of personal pronouns, which exhibited a “V” curve growth, largely mirrored the change in delivery methods of the addresses. During the first decade (1790–1800), the addresses were delivered as spoken speeches, accompanied by a large number of personal pronouns. From 1801 to 1912, a span featuring a shift to written reports, the use of personal pronouns steadily declined. After 1933, with the return to spoken speeches, the number of personal pronouns surged again. It may be President Franklin Roosevelt (presidency: 1933–1945) who ushered the modern tradition of the frequent use of we and our (Shogan 2016). The increased use of personal pronouns can also be attributed to the effect of colloquialization (Hundt and Mair 1999; Leech et al. 2009) and the need to establish a closer relation with the audience.

Table 6:

The diachronic changes of predicative factors of relativizer alteration.

Factors	Diachronic changes
Relative clause length
Filler-gap dependency distance
Head noun length
Mean word length
Mean sentence length
Personal pronouns

3.3 The abrupt change in the use of relativizers

Although a general upward trend of zero and a downward trend of which and that has been observed in the examined period, it remains unclear whether the rate of the change is consistent throughout the time series, or any sudden change occurred at specific phases. The Pettitt test for single change point detection (Verstraeten et al. 2006) revealed significant shifts of which, that, and zero relative clauses (p < 2.2e⁻¹⁶) in the examined timespan. The change points for the three types of relative clauses are listed in Table 7. It is of interest to find that the turning points for all the three types of relative clauses clustered around the year 1940, which revealed a faster descend of which relative clauses in 1939, an accelerated growth of that relative clauses in 1939 and zero relative clauses in 1942.

Table 7:

The change points in different types of relative clauses.

Type of relative clauses	Change points	Figures
Which relative clauses	1939
That relative clauses	1939
Zero relative clauses	1942

4.1 Diachronic changes of relativizers

The first finding of our study pertains to the diachronic change of relativizers in the SOTU addresses across the past 234 years. The study revealed an increase of that over which and an increase of zero over which and that. The results of the multivariate model suggested that the observed changes may not solely reflect grammatical alternations in relativizer forms. Instead, these shifts could be indicative of broader external factors influencing the textual habitat (Szmrecsanyi 2016). To be specific, the textual habitat has undergone a decrease in linguistic complexity across various aspects, indicating a trend towards simplification in the SOTU addresses. The complexity principle posited a preference for explicit relativizers in complex linguistic contexts (Rohdenburg 1996), because explicit relativizers could help facilitate clarifying the structure and meaning of complex relative clauses (Tagliamonte et al. 2005). In other words, in simpler relative clauses, explicit relativizers may not be necessary, which results in their reduced frequency. Hence, due to the trend of linguistic simplification in the textual habitat, the occurrence of explicit relativizers has become less frequent, resulting in less which and more that, and more zero than which/that. In addition, style factors such as personal pronouns also exhibited a trend towards a more colloquial style in the SOTU addresses. Considering the stylistic difference of the relativizers (Biber 1986, 1995), the colloquialization in texts led to the increase of that and the decrease of which. In summary, relativizers evolved with an increase of that over which and zero over which/that, which appeared to be influenced by factors such as simplification and colloquialization in the textual habitat.

In addition, our findings provided valuable insights concerning the change of various linguistic features from the perspective of the American presidents’ addresses. First, the sentence length has undergone a decreasing trend, which corroborated the previous findings in public speaking (Tsizhmovska and Martyushev 2021) and scientific discourses (Hundt et al. 2012). Second, the decrease in filler-gap dependency distance, which is related to the integration and retrieval costs in processing relative clauses (Futrell et al. 2015; Gibson 2000), reflected the trend of cost saving in human communication. The finding provided further evidence for dependency distance minimization (Futrell et al. 2020; Lei and Wen 2020) from the perspective of non-adjacent dependency. Last, with more zero relativizers, noun phrases with modifier construction (Nouns + relative clauses) (Hoffmann and Trousdale 2013) tended to be compressed, allowing speakers to convey the same message with the minimum number of words (Baker 2011; Zhu and Lei 2018). The results support the claim that sentence structure is optimized for efficiency, similar to observed patterns at both lexical (e.g., omission of infinitive marker) (Berdicevskis et al. 2024) and phonetic levels (e.g., omission of explosive sounds) (Mahowald et al. 2018; Manin 2008).

The findings of diachronic change of relativizers also reflected a trend towards explicitness reduction (Biber and Gray 2016). The trend manifested in the increased frequency of less explicit relativizers. For example, the less explicit relativizer that was increasingly favored over which in subject relative clauses. Similarly, within object relative clauses, the least explicit option zero exhibited a growing prevalence, surpassing both which and that. In 1790, both zero and that only held a small proportion, but they increased gradually and overtook which around 1960. The post-1960 distribution (zero > that > which) inverted the degree of their explicitness in the referents (which > that > zero). The observed trend of explicitness reduction aligned with the findings of Biber and Gray (2016), who reported the decrease of explicitness across various linguistic elements, such as connectors (e.g., from linking adverbials to colons) and modifiers (e.g., from clausal modifiers to phrasal modifiers). Our finding, showing a decrease in relativizer explicitness, corroborated Biber and Gray’s (2016) observations regarding linguistic change.

4.2 What may accelerate language change?

The growth pattern of language change is also of interest. According to the Piotrowski Law (Piotrowskaja and Piotrowski 1974), the development of linguistic units (e.g., vocabulary growth) usually follows a path of initial slow growth, subsequent acceleration, and eventual deceleration (Sun and Baayen 2021; Wu et al. 2016). The path shape is similar to an “S”, also known as an S-curve model (Sun and Baayen 2021). The S-curve model describes changes in frequency at morphological, semantic, and syntactic levels (Sun and Baayen 2021). Our study indicated that the development of relativizers also adhered to the S-curve model, exhibiting a gradual start and subsequent acceleration. However, the change of relativizers appeared to have not yet reached the final stable phase. The change points of the relativizers which, that, and zero have revealed an acceleration in their diachronic change around 1940. If the principle of the least effort as a universal constraint on language development (Millward and Hayes 2012 [1988]; Zipf 1949) was considered, the abrupt change in the rate or the acceleration of the relativizers around 1940 can be explained as follows.

The first possible explanation pertains to the issue of prescriptivism. In 1926, Henry Fowler, renowned for A Dictionary of Modern English Usage, advocated the use of that in restrictive relative clauses to balance which, the sole permissible relativizer in non-restrictive relative clauses (Fowler 1926). The prescriptive rule, disseminated through education, dictionaries, and proofreading tools, gradually permeated everyday language use (Hinrichs et al. 2015). However, the widespread adoption of such a prescriptive rule can be a protracted process (Schmidt and McCreary 1977), which may explain the time lag between the proposal of prescriptive rules in 1926 and the accelerated shift in the relativizer change in the 1940s.

Technological innovations in media around the 1940s may also have sped up the change of relativizers. One pivotal development was the invention of color television in 1940, which profoundly impacted the life of the American public (Murray 2018). This groundbreaking technology broke down geographical barriers, facilitated language communication, and fostered extensive language contact. The surge in language contact, in turn, could provide opportunities for audiences to encounter diverse linguistic structures and usage patterns (Blount and Sanches 1977), thereby expediting the formation of new relativizer distributions.

Another possible catalyst that could accelerate the change of the relativizers is the transition in delivery styles of the SOTU addresses. The year 1940 was in the middle of the administration of President Franklin D. Roosevelt, who has significantly shaped the modern oral practice of the Annual Message (Shogan 2016). The stylistic analyses by Savoy (2015) revealed that since Roosevelt, the presidents tended to adopt a style distinct from their predecessors, such as the use of more adjectives (e.g., Eisenhower), pronouns (e.g., Clinton), or verbs (e.g., Obama). Consequently, the shift towards a modern oral speech could constitute another possible factor that accelerated the change of relativizers.

While we have discussed the aforementioned factors, it is important to acknowledge that other factors may also contribute to its acceleration, such as the popularization of education after the second world war. However, the factors which can affect language change are so complex and intricate that we cannot exhaust all of them in this study.

4.3 Extracting zero relative clauses based on dependency grammar

Methodologically, our analysis highlighted the benefits of using a long-term diachronic corpus and an improved method of automatic extraction of relative clauses.

First, the long-term diachronic corpus is useful for revealing both the general trend and the nuanced fluctuations. An important finding is that the increase of zero in our datasets appeared to contradict the stable state of zero reported in previous studies such as Hinrichs et al. (2015). The observed difference may arise from disparities in time span and sampling intervals. Hinrichs et al. (2015) focused on comparison of two periods (1960 and 1990), a timeframe of 30 years that may not be sufficient enough for a comprehensive observation of relativizer change. While the use of zero may appear consistent at both ends of the 30-year period, fluctuations in its occurrence could have transpired that were not captured by the two designated time points. Therefore, for one thing, the long-term corpus in our study, encompassing successive periods over more than 200 years, enriches our understanding of the long-term development of American presidents’ speeches (Hilpert and Cuyckens 2016; Hilpert and Gries 2016). For another, the intervals are also of importance because the dynamics of diachronic change can range from relative stability to substantial reorganization (Buchstaller 2009), necessitating a fine level of observational granularity (Hilpert and Gries 2016: 44). These even and fine-grained intervals enable us to scrutinize ups and downs, even when the proportions at either end appear similar (Ball 1996).

Second, the automatic retrieval and annotation of relativizers and relative clauses, particularly zero ones, have proven invaluable for large-scale research on relative clauses. Researchers previously manually classified and extracted relative clauses from a limited number of texts (Olofsson 1981). With the advent of tree banks, an exhaustive compilation of regular expressions was used to search for specific forms of relative clauses (Roland et al. 2007). Subsequently, supervised machine learning was introduced to extract relative clauses utilizing pretrained data. The method, nevertheless, has yielded barely satisfactory recall and precision values of 0.85 and 0.60 respectively (Frazee et al. 2015). This issue has been addressed in our study with the use of a dependency parser to identify the relation labeled as acl:relcl (clausal modifiers of nouns). The approach facilitates the extraction of all relative clauses with high accuracy and efficiency (with recall and precision rates both above 0.9). Hopefully, this efficient retrieval approach can be adopted to explore large corpora and to explore the use and change of different sophisticated syntactic structures.