Register evaluation in lexicon?

Language corpora as a source of information for

• meaning
• typical context / collocations
• frequency
• …
• why not style/usage markers?

Style/usage markers/labels

• variety – standard vs. non-standard (× Homoláč & Mrázková 2014)
• communicative situations – corpus metadata (text types, genres)
  • text types contain mixture of styles (cf. newspaper)
  • what about corpora without metadata? (web-crawled, Baroni et al. 2009; Benko 2014; Jakubíček et al. 2013; Davies 2018)
• register affiliation

Outline of the talk

1. multi-dimensional analysis and establishing registers
2. how to measure affinity of a lexeme and register
3. empirical verification
   • extraction of register-typical words
   • verification of established labels (case of ASSČ)

Principles of multi-dimensional analysis (MDA)

Biber 1995; Biber & Conrad 2009

• model of systemic & functional variability (× random, sociolinguistic…)
  • motivated by context & situation
• text production process involves interrelated choices
• dimensions of variation (“intratextual” perspective)
• establishing registers (clusters of text with similar linguistic features)
• model of register variation can be used for evaluation of texts (“additive” MDA)

Methodology of MDA

Establishing MD model:

1. corpus compilation
2. features: operationalization & extraction
3. statistical evalaution: factor analysis \(\rightarrow\) dimensions
4. interpretation of results (labelling dimensions)
5. clusters of texts \(\rightarrow\) registers

CNC: MDA team

MDA of Czech

Mini-portal https://www.korpus.cz/mda

Cvrček, V. – Komrsková, Z. – Lukeš, D. – Poukarová, P. – Řehořková, A. – Zasina, A.J. (2021): From extra- to intratextual characteristics: Charting the space of variation in Czech through MDA. Corpus Linguistics and Linguistic Theory 17(2), p. 351-382.

Cvrček, V. – Laubeová, Z. – Lukeš, D. – Poukarová, P. – Řehořková, A. – Zasina, A. J. (2020): Registry v češtině. Praha: Nakladatelství Lidové noviny, (233 p.).

MDA of Czech

• inspiration from English and other languages
• 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

Data: Koditex corpus

• “traditional” carefully designed corpus covering all available text types
• guiding principles: diverse, contemporary, text length control
  • text excerpts = chunks (not whole texts)
  • 3 modes – wri, spo, web
    • 8 divisions, 45 classes, \(\approx\) 200,000 words per class

Koditex: composition

Features and their operationalization

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

• phonetics – narrowing é > í, vowel breaking 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…
• pragmatics – contact expressions, fillers, intensifiers, downtoners…
• syntax – types of attributes, clusters of POS, types of dependent clauses…
• text/discourse – questions, phraseology, word repetition…

Statistical evaluation: Factor analysis

• 122 features × 3292 text chunks
• factor analysis:
  • identifying latent factors
  • R environment, using fa function from psych package
  • parameters:
    • rotation: promax (oblique)
    • factoring method: generalized weighted least squares (GLS)
    • number of factors/dimensions: 8
    • variance explained: 56 %

Factor analysis – output

• loadings – “correlations” of features and dimensions
  • participation of a feature on a dimension
• factor scores – positions of texts (chunks) within dimensions
  • linguistic characteristics of a text
• variance explained: 56 %

Interpretation: Dimensions of variability

Dim 1: dynamic (+) × static (-)

Features dim 1: dynamic (+) × static (-)

Positive loading features:

• past tense verbs, finite verbs, verbs: indicative forms, perfective verbs, 3rd person pronouns (personal + possessive), semantically bleached verbs, function words (pron., num., prep., conj. & part.), adverbs of time, pronouns, verbs: 1st person, verba dicendi, sentence negation…

Negative loading features:

• nominal post-modifiers without agreement, adjectives, abstract nouns, noun pre-modifiers with agreement, nouns: genitive, adjective clusters, noun clusters, clusters of same-case adjectives, nouns, average word length (number of syllables), verbal nouns complex prepositions…

Dim 2: spontaneous (+) × prepared (-)

Features dim 2: spontaneous (+) × prepared (-)

Positive loading features:

• features marking a) interactivity and online production (contact expressions, fillers, demonstratives, pronouns, word repetition) and b) informality (expressive particles, interjections) attract c) conventionalized non-standard Common Czech morphonological variants, symptomatic of diglossia

Negative loading features:

• both frequency 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

Dim 8: attitudinal (+) × factual (-)

1. Projection on MD model

Additive MDA (cf. Berber Sardinha et al. 2019, 165):

• using previously established general model for new data
• texts from other corpora (e.g. web-crawled) can be analysed with the same set of features that were used for the original MDA
• relative frequencies of features can be converted to positions on dimensions

\(\Rightarrow\) register information for any text of a language

2. Establishing registers

Register

• a variety defined by (a) set of distinct pervasive and functional linguistic features and (b) communicative situation
• operationalization: set (cluster) of texts sharing the same position on dimensions

Register classification

Static registers

• analysis: static monothematic
• popularization: static polythematic general
• journalism: static mixed/indeterminate
• facts: static polythematic particular
• argumentation: static cohesive

Dynamic registers

• question answering: dynamic without addressee coding
• conversation: dynamic spontaneous
• commentary: dynamic attitudinal
• screenplay: dynamic with addressee coding
• narration: dynamic retrospective

Narration: dynamic retrospective register

Average position of texts within the narrative cluster (9)

Words in registers

Usage based approach to style labels:

• preference/affiliation/affinity of words to a register \(\sim\) association

\(\Rightarrow\) using frequency and association measures

Association with registers

Instead of co-occurrence of words x and y we work with word x and its presence in texts of register R:

\[\text{MI-score} = \log_2 \frac{f(xy) \times N}{f(x)\times f(y)} \Rightarrow \log_2 \frac{f(xR) \times N}{f(x)\times f(R)}\]

\[\text{logDice} = 14 + \frac{2f(xy)}{f(x) + f(y)} \Rightarrow 14 + \frac{2f(xR)}{f(x) + f(R)}\]

where:

• \(f(xR)\) …occurrences of x in texts of register R
• \(f(R)\) …number of tokens in texts of register R

Top 3 associations

source: Koditex

Simple modification

Number of texts (chunks) instead of occurrences/frequencies

\[\text{MI-score} = \log_2 \frac{texts(xR) \times texts(corpus)}{texts(x)\times texts(R)}\]

\[\text{logDice} = 14 + \frac{2\times texts(xR)}{texts(x) + texts(R)}\]

where:

• \(texts(xR)\) …number of texts of register R containing x
• \(texts(R)\) …number of texts of register R

Modified results

source: Koditex

Examples from a larger corpus

Applied to SYN2015 (100m corpus of written Czech)

Verification of established labels

Labels in Akademický slovník současné češtiny (Academic dictionary of contemporary Czech) – currently compiled at the CLI (words: A–G)

Usage labels in ASSC

Style characteristics:

• colloquial
• colloquial with a tendency to become neutral
• slang and professional terms
• vulgar expressions (too few occ. in corpus)
• pejorative

RQ: Where can we typically find them?

Data: Koditex (SYN2015)

Colloquialisms

Which register has the strongest association with words labeled as colloquial (source: Koditex)?

n = 129

Colloquialisms with a tendency to become neutral

Which register has the strongest association with words labeled as colloquial with a tendency to become neutral (source: Koditex)?

Slang and professional terms

(source: Koditex)

Pejorative expressions

Where are these expressions used in SYN2015?

Conclusions I

• What are usage labels?
  • warning signs or
  • descriptions of words’ typical habitats
• Parallels with: prescriptive/proscriptive vs. descriptive approach
  • the former is demanded by the users
  • the later is preferred by (empirical) linguists

Conclusions II

Data-driven (descriptive) approach:

• text-linguistic approach (MDA) to variation \(\rightarrow\) registers based on linguistic features \(\rightarrow\) more objective labels
  • applicable to corpora without metadata (web-crawled, cf. Sharoff 2018)
• post-hoc check of the labels – where the words with label X live
• the concept of association can be extended from the lexical level (collocations) to other domains (although modifications are necessary)

References