Crosslinguistic Corpus Studies in Linguistic Typology

Crosslinguistic Corpus Studies in Linguistic Typology Stefan Schnell Nils Norman Schiborr

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

Crosslinguistic Corpus Studies in Linguistic Typology 2022 Annual Review of Linguistics 10.1146/annurev-linguistics-031120-104629 schnell2022-crosslinguistic-corpus-studies-in schnell2022-crosslinguistic-corpus-studies-in.txt, the project's staged copy 5d187c92b160e4b5bfd80aa99bdd48ce6e3d8f35 open-access linguistic-typology

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them existence Quotation withheld because this record is not cleared for redistribution. corpus-based typology Quotation withheld because this record is not cleared for redistribution. review review Quotation withheld because this record is not cleared for redistribution. mechanisms Conventionalised orders minimise total dependency length, and because a dependent tree then sits closer to the higher head it hangs from, harmonic (cross-domain aligned) word order falls out as a by-product. Quotation withheld because this record is not cleared for redistribution. mechanisms Keeping a head adjacent to the head of its dependent lets the hearer recognise higher-order constituency sooner, which is why languages converge on harmonic orders. Quotation withheld because this record is not cleared for redistribution. mechanisms Nouns anchor referring expressions and so open a wider set of lexical and referential choices, making them harder to plan, which shows up as slower speech and pausing just before them. Quotation withheld because this record is not cleared for redistribution. The authors' own exploratory comparison of noun-phrase lengths in Universal Dependencies versus Multi-CAST is presented as preliminary. what this limits: the DLM comparison specifically: MORE DETAILED INVESTIGATIONS ARE REQUIRED, these being INITIAL OBSERVATIONS Quotation withheld because this record is not cleared for redistribution. Corpus representativeness is identified as the field's most pressing unsolved problem. what this limits: every crosslinguistic corpus result in the field: THE MOST PRESSING ISSUE IN CBT IS THE REPRESENTATIVENESS OF CORPORA USABLE FOR CROSSLINGUISTIC RESEARCH Quotation withheld because this record is not cleared for redistribution. Spoken and especially interactional language use is badly under-represented in available corpora. what this limits: any claim about spoken or conversational language: THERE IS A BLATANT SHORTAGE OF SPOKEN-LANGUAGE DATA FROM A LARGER RANGE OF LANGUAGES, [and] EVEN LESS INTERACTIONAL LANGUAGE USE Quotation withheld because this record is not cleared for redistribution. The review names its field's outstanding problem as the corpora themselves: they are not yet crosslinguistically representative, and spoken or signed texts are what is missing. what this limits: any crosslinguistic generalisation drawn from these corpora: the review names more representative corpora including spoken or signed texts as a remaining challenge for the field Quotation withheld because this record is not cleared for redistribution. The texts that would be comparable are frequently not obtainable, and copyright is named as one of the reasons, so what can be compared is set by access rather than by the research question. what this limits: the coverage of stimuli-based elicited texts in this field: they are crucial for projects needing comparability, and the review states they are rarely published as organized web-accessible corpora and often not openly accessible Quotation withheld because this record is not cleared for redistribution. corpus-based typology corpus-based typology Quotation withheld because this record is not cleared for redistribution. utterance populations utterance populations Quotation withheld because this record is not cleared for redistribution. grammar-based typology grammar-based typology Quotation withheld because this record is not cleared for redistribution. language - AS A POPULATION OF UTTERANCE TOKENS: the corpus-based conception, in which an individual corpus is a SAMPLE of that population rather than the language itself Quotation withheld because this record is not cleared for redistribution. language - AS A MONOLITHIC ENTITY shared by all members of a speech community: the conception the corpus-based view is explicitly defined against Quotation withheld because this record is not cleared for redistribution. word order - a CATEGORICAL property in some languages and a probabilistic one in most: Tamil at 0% and Arabic at 100% VO, while most languages show variability with a preference Quotation withheld because this record is not cleared for redistribution. builds-on → Haig et al. (2011) Quotation withheld because this record is not cleared for redistribution. contradicts → Du Bois (1987) discourse ergativity Quotation withheld because this record is not cleared for redistribution. corpus-based typology —contrasted-with→ grammar-based typology corpus-based typology grammar-based typology strength of the relation: 1 Quotation withheld because this record is not cleared for redistribution. regression regression Quotation withheld because this record is not cleared for redistribution. mean mean Quotation withheld because this record is not cleared for redistribution. standard deviation standard deviation Quotation withheld because this record is not cleared for redistribution. Corpus-based typology largely reproduces the classifications of grammar-based typology, but adds finer-grained, quantitatively explicit detail and robustness. Quotation withheld because this record is not cleared for redistribution. In the authors' own comparison, long noun phrases are far commoner in the written Universal Dependencies corpora (about 36% of NPs at least four dependencies long) than in the spoken Multi-CAST corpora (about 12%). Quotation withheld because this record is not cleared for redistribution. Spoken texts seem to lack the prerequisite for DLM effects: phrases or local dependency trees long enough for dependency length to matter. Quotation withheld because this record is not cleared for redistribution. Across 12 spoken corpora, subjects inside a same-subject reference chain are about 15-20% likelier to be realised as zero than subjects elsewhere, regardless of the language's overall zero rate. Quotation withheld because this record is not cleared for redistribution. Nearly every language with a zero-subject rate above 70% has fully grammaticalised, transparent subject agreement, the exceptions being Mandarin and Jinghpaw. Quotation withheld because this record is not cleared for redistribution. Haig & Schnell (2016) and Schnell et al. (2021) find no support for discourse ergativity: S shows a lower rate of newness and full-NP expression than P. Quotation withheld because this record is not cleared for redistribution. Slower articulation and pausing are significantly more likely just before nouns than before verbs. Quotation withheld because this record is not cleared for redistribution. Word-order entropy trades off against case marking: the less confusable subject and object are by case, the freer their relative order. Quotation withheld because this record is not cleared for redistribution. The largest annotated multilingual resource is typologically skewed - 24 million tokens across 104 languages, more than half of them Indo-European. Quotation withheld because this record is not cleared for redistribution. Spoken-language data are scarce in corpus-based typology, and conversational data scarcer still. Quotation withheld because this record is not cleared for redistribution. corpus-based typology (CBT) corpus-based typology (CBT) Quotation withheld because this record is not cleared for redistribution. grammar-based typology (GBT) grammar-based typology (GBT) Quotation withheld because this record is not cleared for redistribution. parallel text corpora parallel text corpora Quotation withheld because this record is not cleared for redistribution. parallax text corpora parallax text corpora Quotation withheld because this record is not cleared for redistribution. original text corpora original text corpora Quotation withheld because this record is not cleared for redistribution. dependency length minimization (DLM) dependency length minimization (DLM) Quotation withheld because this record is not cleared for redistribution. discourse ergativity discourse ergativity Quotation withheld because this record is not cleared for redistribution. NP length NP length Quotation withheld because this record is not cleared for redistribution. proportions of zeros proportions of zeros Quotation withheld because this record is not cleared for redistribution. 36% ... 12% — share of noun phrases at least four dependencies long in the written UD corpora versus the spoken Multi-CAST corpora Quotation withheld because this record is not cleared for redistribution. 15–20% — extra likelihood that a chain-medial subject is zero rather than a pronoun, compared with subjects outside a same-subject chain Quotation withheld because this record is not cleared for redistribution. β1 = 0.84 — regression coefficient of the fit relating zero rates inside versus outside same-subject chains, with intercept 24.77 Quotation withheld because this record is not cleared for redistribution. 1,257 — languages covered by the Parallel Bible Corpus, averaging 10,707 verses per language Quotation withheld because this record is not cleared for redistribution. 464 — languages for which a Universal Declaration of Human Rights translation exists, averaging 1,694 words each Quotation withheld because this record is not cleared for redistribution. 24 million — word tokens in Universal Dependencies v2.7, spread over 104 languages of which over half are Indo-European Quotation withheld because this record is not cleared for redistribution. 131,574 — mean word-token size of an individual language corpus in Universal Dependencies Quotation withheld because this record is not cleared for redistribution. 143 ... 74 — UD corpora, and languages they come from, used in the authors' own NP-length analysis after excluding 37 with under 1,000 data points Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.1146/annurev-linguistics-031120-104629.