Multi-dimensional analysis of Czech

Václav Cvrček

Zuzana Komrsková

David Lukeš

Petra Poukarová

Anna Řehořková

Adrian Jan Zasina

24^th September 2018

Multi-dimensional analysis

Principles of multi-dimensional analysis (MDA)

systemic & functional variability (× random, sociolinguistic…)
motivated by context & situation
registers × genres (~ intratextual × extratextual perspective)
text production process involves interrelated choices

→ multi-dimensional approach seems a good fit (Biber 1991; Biber & Conrad 2009)
expected challenges / highlights of MDA…
- … in Slavic languages – specific morphology, inflection, free word order
- … in Czech – situation bordering on diglossia (Bermel 2014): Literary × Common Czech

Methodology

corpus compilation
features: operationalization & extraction
statistical analysis (factor analysis, FA)
interpretation of results

Data

Koditex corpus

guiding principles: diverse, contemporary, text length control
- “diversified” stratified sampling
- after 1990, majority from 2007–2014
- text excerpts = chunks (not whole texts)
annotation: lemmas, tags, multi-word unit & named-entity recognition
tools: KonText, MorphoDiTa, NameTag
10 million words, 3,334 text chunks, 2000–5000 words each
3 modes – wri, spo, web
- 8 divisions, 45 classes, \(\approx\) 200,000 words per class
- letters, administrative texts: shorter chunks (1000 words)
- posts (web-mul): aggregated according to author and time
- spo: one speaker within conversation

Features

Features and their operationalization I

Originally: 140+ features, final list: 122, e.g.:

phonetics – narrowing é > í, diphthongization ý > ej, average word length…
morphology – freq. of cases, numbers, moods, tenses…
derivation – adjectives denoting similarity, verbal nouns, diminutives…
lexicon – indefinite pronouns, reporting verbs, verbs of thinking, semantically bleached nouns…

Features and their operationalization II

pragmatics – contact expressions, fillers, intensifiers, downtoners…
syntax – types of attributes, clusters of POS, types of dependent clauses…
text/discourse – questions, phraseology, word repetition…

Type-based features – inventories of pronouns, prepositions, conjunctions (relativized using zTTR, Cvrček & Chlumská 2015)

Lexical richness – Yule’s K, thematic concetration (Popescu et al. 2007), unigrams & bigrams (zTTR)

Statistical evaluation

Factor analysis

input: 122 features × 3292 text chunks
factor analysis:
- R environment, using fa function from psych package
- parameters:
  - rotation: promax (oblique)
  - factoring method: generalized weighted least squares (GLS)
  - number of factors/dimensions (!): 8
output:
- loadings – “correlations” of features and dimensions
- factor scores – positions of texts (chunks) within dimensions
variance explained: 56%

Interpretation

Dimensions of variability

dynamic (+) vs. static (-)
spontaneous (+) vs. prepared (-)
higher (+) vs. lower (-) level of cohesion
polythematic (+) vs. monothematic (-)
higher (+) vs. lower (-) amount of addressee coding
general/intension (+) vs. particular/extension (-)
prospective (+) vs. retrospective (-)
attitudinal (+) vs. factual (-)

Dim 1: dynamic (+) × static (-)

Explain plot. Tries to aid interpretation by showing how different externally specified text categories are positioned within a dimension. I.e., we know part of our corpus contains administrative texts (it’s one of the 45 classes), and this plot shows it’s the most negatively scoring class in our data in terms of the first dimension. That tells us something about the first dimension, and it’s part of how we interpreted it and came up with the label dynamic × static.

Y-axis shows factor scores, i.e. position within dimensions. Plus is up.

first dimension, most variance explained
verbal vs. nominal constructions
+ load feats: verbs as a part of speech and several verbal subcategories; function words (pronouns and conjunctions); adverbs
- load feats: nouns and adjectives in various functions; secondary prepositions; word length; passives; thematic concentration
opposing strategies: elaboration of clause members (-) or adding new clauses → clausal–phrasal (Biber)
0 load feats: dim indifferent to preparedness of speakers/writers: associated with spontaneous spoken comm. or lexical richness
+ fact score texts: two shades of “verbality”, narrative (e.g. various kinds of novels) + reflective (verbs of thinking in pri cor or web forums)
- fact score texts: information-dense chunks – official documents, hard science papers, encyclopedias

Dim 2: spontaneous (+) × prepared (-)

reflects differences in conditions of production: wri vs spo
wri: endless editing and refining possible
spo: online production and perception under heavy time constraints → speakers juggle tradeoffs between eloquence, intelligibility, timeliness, propriety, etc.
+ feats: contact expressions, fillers, pronouns
- Czech – pro-drop → explicit use of pronouns ~ redundancy of information
- confirmed by additional feats: word reduplication, Yule’s K, formulaic language
- redundancy: flaw of online production…
- … but also feature for successful communication over noisy channel
- specifically Czech: non-standard forms of the Common Czech variety
  - long-established spellings
  - but not standard
  - still stigmatized in formal settings → diglossia?
+ texts: spo int inf (stands out), pri cor, web fcb / for
- texts: prepared monologues (!), administrative texts, economic news, Wikipedia
- feats: both freq and type inventory of prepositions, clauses with interrogative or relative adverbs, lexical richness (zTTR), nouns, longer words

2D-plot: dim 1 and dim 2

Dim 5: higher (+) × lower (-) amount of addressee coding

+ feats: questions, verbs 2nd pers. sg. and pl., future tense, vocative case, imperative mood
- typical of dialogue × monologue?
- feats: only avg sent length in tokens (makes sense in theory but unreliable) → absence of + feats
presence of and interaction with communication partner
cf. distribution of spo classes: elicited speech (- → interview answers) × informal private dialogues (+ → turn-taking mechanisms)
BUT + texts: especially fiction! (screenplays, poetry, sci-fi, crime, romance, fantasy and short stories)
WHY?
- exaggeration of typical features ~ artistic representation: natural characteristics exhibited by real-life dialogues are over-emphasized in fictional ones
- making dialogues accessible to outsiders (viewers, readers)
- e.g. vocative (not needed all that often in real embodied multi-party interaction)

Dim 8: attitudinal (+) × factual (-)

+ feats:
- particle-related: downtoners, restrictors, intensifiers, text organizing particles
- adverbials, conjunctions (including concessive and adversative connectives)
- modality across POS categories
→ careful evaluation of quality / relations + attempt to present speaker’s point of view
- feats: only coordination (clausal or phrasal) → no straightforward interpretation due to versatility:
- substitute for complex syntactic relations (esp. in online production)
- enumeration of items in list
+ texts: web comments sections + forums (→ presenting viewpoint), pri cor, eli (monologues giving opinions), blogs
- most “positive” written class: popular non-fiction from humanities
- tendency to overtly mark message as representing speaker’s opinion / stance
- texts: administrative, Wikipedia, screenplays, poetry/lyrics, tabloids
- common traits: enumeration (lists of names or things), lack of hedges / explicit stance markers

Final remarks

Conclusion

Key features of Czech MD model:
- “universal” dimensions (Biber 2014):
  - clausal/oral–phrasal/literate \(\approx\) dynamic × static + spontaneous × prepared (← diglossia? spo data?)
  - narrative–non-narrative \(\approx\) prospective × retrospective
- “specific” dimensions:
  - polythematic × monothematic (← type-based features)
Current work:
- studying elicited texts from psychology through MD model
- comparison of corpora (traditional vs. web-crawled)
Future plans:
- how robust is the MD model?
- intratextual classification and annotation in CNC

Comparison of corpora from the perspective of how much linguistic variability they cover (= what types of research they can be used for).

supports Biber’s 2014 proposal of universal dimensions:
- clausal vs. phrasal
- narrative vs. non-narrative
first dimension usually characterized by opposition between verbal and nominal
- most prominent text category on “verbal” extreme: almost always private conversation
- → features functionally associated with personal involvement, interactivity and/or online production - ours is a bit different: clear verbal vs. nominal distinction associated predominantly with written genres (especially romance novels and letters on the positive side)
- past tense verbs are most salient verbal feature → our dim 1 marks both personal involvement and narration in contrast to description (therefore, more general label “dynamic”)
narr. vs non-narr. is similar
- contrast between past and present tense
- conspicuously separates fiction from rest
every MDA uncovers idiosyncrasies of target language → specific sociolinguistic situation of Czech highlighted by dim 2 (spontaneous vs. prepared), which isolates spoken, interactive and private genres
3 groups of features:
1. interactivity and online production (contact expressions, fillers, demonstratives, word repetition)
2. informality (expressive particles, interjections)
3. conventionalized non-standard Common Czech morphonological variants (~ diglossia)
QUESTION: influence of interactivity + online production usually visible right in dim 1 of MD models
- is the presence of Common Czech features responsible for this re-arrangement?
- or caused by nature of our spo data? (scope – broad, includes real informal spoken language, transcription decisions)
polythematic vs. monothematic
- separates newspaper articles from rest
- undoubtedly present in other languages as well, but only manifests itself upon the inclusion of various language-agnostic measures of lexical richness (unigrams and bigrams on the positive pole, Yule’s coefficient and thematic concentration on the negative) and type-based features
robustness: if we leave out Common Czech features, will dims 1 and 2 join?
different names:
1. try to avoid bias from previous research, jump to conclusions
2. came up with names that worked well in Czech first

References

Bermel, N. (2014): Czech Diglossia: Dismantling or Dissolution? In J. Árokay et al. (eds), Divided Languages? Springer.
Biber, D. (1991): Variation across speech and writing. Cambridge: Cambridge University Press
Biber, D. & Conrad, S. (2009): Register, Genre, and Style. New York, NY: Cambridge University Press.
Biber, D. (2014): Using multi-dimensional analysis to explore cross-linguistic universals of register variation. Languages in Contrast 14(1), 7–34.
Cvrček, V. & Chlumská, L. (2015): Simplification in translated Czech: a new approach to type-token ratio. Russian linguistics 39(3), 309–325.
Popescu, I., Best, K. & Altmann, G. (2007): On the dynamics of word classes in texts. Glottometrics 14, (p. 58–71).

Acknowledgments

This research was supported by the ERDF project Language Variation in the CNC no. CZ.02.1.01/0.0/0.0/16_013/0001758.

It builds upon work made possible by the Czech National Corpus project (LM2015044) funded by the Ministry of Education, Youth and Sports of the Czech Republic within the framework of Large Research, Development and Innovation Infrastructures.

logolink