Marc Kupietz, Beata Trawiński, Nils Diewald

DeutUng in the Context of the European Reference Corpus EuReCo

DeutUng final workshop, 16.12.2021, Mannheim / online

Background

• importance of corpora for linguistic research, both in single language as well as in cross-linguistic contexts

• the number of linguistic studies based on corpus data is increasing

• the number (and size) of corpora is growing

• linguists are often faced with a choice between several different types of corpora

• options available for cross-linguistic research →

Corpora for language comparison

Kupietz et al. (2020), Trawiński and Kupietz (2021)

Corpora for language comparison

Kupietz et al. (2020), Trawiński and Kupietz (2021)

Monolingual corpora

• texts in only one language

• usually in the original language, therefore of high quality

• usually linguistically annotated (language-specific)

• examples of large national reference corpora:

  • DeReKo , ANC, BNC, CNC, NKJP, RNC, CoRoLa, HNC

• frequently used in cross-linguistic research

  • Augustin (2017), Taborek (2018, 2020), GDE etc.

Methodological questions

• To what extent are the results of studies based on monolingual corpora comparable?

  • The results are comparable at a meta-level (theoretical level / level of generalizations).

  • At the empirical level (data level) they are less comparable.

  • Reason: Diversity of monolingual corpora (Trawiński and Kupietz 2021)

Conclusion: Monolingual corpora

• low matching with regard to size, text types, topics, etc. (also morphosyntactic annotation)

• 

• but high linguistic quality

• 

Parallel corpora

• Parallel corpora consist of original texts in one language (source language) and their translations in other languages (target languages).

• Texts in all languages aligned at sentence level

• partially linguistically annotated (mainly language-specific annotation)

• large multilingual parallel corpora:

  • OPUS, Europarl, InterCorp

Advantages of parallel corpora

• Parallel data: sequences of linguistic units (words, sentences) in two or more languages,

  • which are translation equivalents of each other and as such convey the same meaning

  • are used in the same contexts

  • occur in the same text types from the same time periods etc.

• perfect basis for identifying functional equivalence between linguistic structures (James 1980, Chesterman 1998) → tertium comparationis

• allow to gain insights into cross-linguistic similarities and differences that might be overlooked when working with monolingual corpora

Linguistic work

contrastive, typological, translational

• Johansson (2007), Altenberg and Granger (2002), Granger (2010), Languages in Contrast (International Journal for Contrastive Linguistics) etc.

• Cysouw and Wälchli (2007) etc.

• Granger et al (2003) etc.

Problems with parallel corpora

• relatively small size

  • the more languages, the smaller and less differentiated is the corpus

• unbalanced in terms of original texts and translations

• specific properties of translations (a third code)

  • Laviosa (1998), Baker (1995), Teich (2003): shining through, normalization, simplification etc.

Conclusion: Parallel corpora

• high comparability in terms of size and content (but not in terms of morphosyntactic annotation)

  • 

• lower quality of the linguistic material

  • 

Need for comparable corpora

• monolingual and parallel corpora alone are not suitable for finer grained linguistic research

  • because they lack either comparability or linguistic quality

• possible remedy:

  • use a combination of parallel and monolingual corpora

  • disadvantages:

    • (quantitative) findings not directly assessable

• desideratum: comparable corpora of high quality

• 

Comparable corpus

Definition (McEnery & Xiao 2007)

• "a comparable corpus can be defined as a corpus containing components that are collected using the same sampling frame and similar balance and representativeness [...], e.g. the same proportions of the texts of the same genres in the same domains in a range of different languages in the same sampling period"

Actual or Practical comparability

is also relevant, we think

• 

• through rich metadata

  • which should be mappable among the corpora

• through linguistic annotations

  • ideally also mappable e.g. by using Universal Dependencies annotations

• through appropriate (and common) analysis tools

  • which make all these possibilities usable in the first place

Available general comparable corpora

... that include German

• currently only Aranea - family of comparable gigaword web corpora (Benko 2014)

  • great and readily usable via NSkE

  • but composition not controlled

  • 

• in future: International Comparable Corpus (ICC)
  (Kirk et al. 2017, Čermáková et al. 2021).

  • approach complementary to EuReCo : using small corpora with controlled composition along the lines of the ICE (Greenbaum 1991)

  • ideally joined with EuReCo in the future

EuReCo - European Reference Corpus

• open initiative founded in 2013 by the academies in Poland, Romania and Hungary and the IDS

• pilot projects (Humboldt Research Group Linkage Programmes):

  • DRuKoLA: Romanian-German (2016-2018): CoRoLa (Tufiş et al. 2019).

  • DeutUng: Hungarian-German (2017-2021): HNC (Váradi 2002)

• EuReCo is based on two core assumptions …

1. dedicated comparable corpora econom. infeasible

• even monolingual universal corpora often cannot be permanently maintained / extended

• dedicated multilingual comparable corpora would multiply the already unrealistic costs

• dedicated comparable corpora cannot be built from scratch and maintained sustainably

EuReCo's approach

• use the existing national and reference corpora

• which are maintained and sometimes extended by sustainable institutions

• instead of trying to build new ones

Expected benefits of the EuReCo approach

• more economical, scalable and sustainable

  • especially since one can also benefit from ongoing and future extensions and improvements of these corpora

• high linguistic quality and sufficient size to be expected in national corpora

• 

2. general comparability is not achievable

• corpora with reasonable size and diversity cannot be perfectly comparable in general

  • there will always be a criterion against which the corpora are not comparable

  • whether an unequal distribution with regard to such a variable is relevant depends on the specific research question

• generally comparable corpora are not a reasonable goal

2b. General representativeness is not possible

• more important: single language corpora cannot be generally representative

  • since population=language is not generally definable

  • whether a corpus is sufficiently representative depends on the research question and the language domain

• generally comparable and representative corpora are not meaningful objectives

EuReCo: dynamically definable, virt. comparable corpora!

In analogy to DeReKo's ›primordial sample‹ approach (Kupietz 2016)

• EuReCo users are invited to ...

• use predefined (comparable) corpora or

• define themselves corpora that are
  suitably representative and comparable
  with regard to their respective research question

Basic approach

(Cosma et al. 2016) cf. McEnery & Xiao (2007).

• draw sub-corpora from the monolingual corpora

• so that they have similar token distributions with respect to metadata variables like:

  • topic area

  • text type

  • publication date

  • …

Refinement: Iterative (and question specific)

for gradual approximation to sufficient comparability

1. begin as described above

2. carry out comparative case studies

3. if the findings appear to be artefacts of comparability criteria, refine the mapping and start again with 2

4. 

In principle possible with KorAP-VC-Builder

but not yet practicable due to missing downsampling function

Comparable CoRoLa/DeReKo corpus

just based on token distribution wrt topic domain (according to DeReKo's top level taxonomy)

Composition by year of publication

was not controlled, but is also quite similar

Comparable corpus usable with KorAP

➞ KorAP on DRuKoLa-VC - ➞ KorAP on CoRoLa: https://korap.racai.ro/

Also HNC fully queryable via KorAP

https://korap.nlp.nytud.hu/ – but so far without controlled comparability

Comparison of syntagmatic patterns

and their contexts of use in German, Romanian & Hungarian

• using collocation analysis

• to investigate e.g. light verb constructions (cf. Taborek 2018)

  • depending on text-external variables

• with manifold 'secondary' objectives:

  • investigate comparability criteria, consequences …

  • identify helpful extensions for KorAP

  • further develop the methodology to identify interesting syntagm. patterns (Belica & Perkuhn 2015)

Collocation analysis with KorAP's R library

not yet supported in the UI, but very flexible with R library (Kupietz et al. 2020b)

library(RKorAPClient)
corola <- new("KorAPConnection", KorAPUrl = "https://korap.racai.ro/")
dereko <- new("KorAPConnection", verbose = T)
vc_drukola <- "referTo drukola.20180909.1b_words"
 
in_NN_setzen <- collocationAnalysis(
  dereko,
  node = "focus(in [tt/p=NN] {[tt/l=setzen]})",
  vc = vc_drukola,
  leftContextSize = 1, # refers to {} in focus()
  rightContextSize = 0
)
 
pune_in_NN <- collocationAnalysis(
  corola,
  node = "focus({[drukola/l=pune] în} [drukola/p=noun])",
  leftContextSize = 0,
  rightContextSize = 1
)
 
 

Example LVC comparison Romanian-German

Kupietz & Trawiński (forthcoming)

➞ Cohesion strengths strongly dependent on domain

Collocate rankings of "pune în ..." in domain = / ≠ law: ϱ(N=39) << 0.58

Collocation analysis with HNC

to identify LVC with hoz (=bring) and nouns in sublative or illative

hulvc <- function(q, leftContextSize, rightContextSize) {
  hnc %>%
    collocationAnalysis(
      q,
      leftContextSize = leftContextSize,
      rightContextSize = rightContextSize,
      exactFrequencies = TRUE,
      withinSpan = "",
      maxRecurse = 0,
      searchHitsSampleLimit = 1000,
      topCollocatesLimit = 30
    )
}
 
hoz_left <- hulvc(
  'focus([hnc/p="FN.(SUB|ILL)"] {[hnc/l=hoz]})',
  leftContextSize = 1,
  rightContextSize = 0
)
hoz_right <- hulvc(
  'focus({[hnc/l=hoz]} [hnc/p="FN.(SUB|ILL)"])',
  leftContextSize = 0,
  rightContextSize = 1
)
 
hoz <- hoz_left %>% bind_rows(hoz_right) %>%
  arrange(desc(logDice))
 
 
 

Light Verb Constructions in Hungarian

hoz (=bring) with noun in sublative or illative – focus([hnc/p="FN.(SUB|ILL)"] {[hnc/l=hoz]})

Larger synt. patterns by »recursive collocation analysis«

hoz (=bring) with recursion depth=2 (sorted by logDice)

Preliminary meta findings

• when comparing syntagmatic patterns, the corpus composition itself can play a greater role than comparability

• corpus access via scripts is essential

  • for reproducible results which are essential not only in the iterative corpus refinements

  • but for all analyses consisting of multiple steps

  • for custom-tailored visualizations

• practical comparability is currently more essential than ›theoretical‹ comparability

Desirable Improvements

• HNC-KorAP-integration: we need more metadata

  • to define comparable corpora

  • to detect dependencies on text external variables

  • but: some great practical comparability is already achieved
    just by the HNC/KorAP integration

• KorAP

  • access to token annotations should be supported by client libraries

    • lemma access essential for agglutinative languages

KorAP Support for EuReCo

• VC for virtual comparable corpora already mentioned

• support for

  • arbitrary metadata

  • arbitrary annotations

    • token-bound, hierarchical, relational

  • different query languages (Bingel & Diewald 2015)

  • arbitrary licences (Bański et al 2014)

• But: Focus on DeReKo and limited resources

Levels of Access

Kupietz et al. (forthcoming)

0

User Interface

• the web user interface (Diewald et al. 2019)

1

Web Service API

• accessible directly or via client libraries (Kupietz et al 2020b)

2

Plugin

• user interface plugins

3

Instance

• independent access by fully customized installations

4

Open Source

• new features by source code contributions

5

Corpus

• direct access to corpus data (without KorAP system)

Plugin Level

Plugin Level

Functional areas of the UI

• search input

  • e.g. lemma expansion

• definition of virtual corpora

  • e.g. corpus visualization

• search results

  • e.g. export

• individual matches

  • e.g. annotation visualizations

• …

Next steps in KorAP development

• extension and publication of core functionalities
  (currently only available via API client libraries)

  • e.g. sorting, grouping, distribution

• establishing plugin interfaces

• publication of plugins

Conclusion

• we would like to have comparable corpora:

  • 

• EuReCo offers a realistic approach to achieve this

  • by virtually joining existing large corpora

  • by means of user-defined, dynamically constructable comparable corpora

  • by providing the required sustainable research platform

Outlook

Next EuReCo steps

• integrate further large corpora

  • we are currently working on the NKJP!

• continue the KorAP development

  • especially also the client libraries

• continue working on ICC and its Integration

References

Monolingual corpora

in linguistic research

• source of evidence, also for quantitative studies, e. g.

  • Augustin (2017): contraction of prepositions and articles in German ( DeReKo ) and Italian (CORIS: Corpus di italiano scritto, PAISÀ corpus)

  • Taborek (2018, 2020): light verb constructions in German ( DeReKo ) and Polish (NKJP)

• GDE-N: DeReKo , BNC, COCA, GloWBe, FRANTEXT, ABU, NKJP, HNC (Gunkel et al. 2017)

GDE-V: monolingual corpora for single case studies

Multilingual corpora

It has often been said that through corpora we can observe patterns in language that we did not know before (...) My contention is that this is especially true for multilingual corpora. We can see how languages differ, what they have in common, and - perhaps eventually - what characterizes language in general.

Johansson (2007)

Case Study:

(Non-)finite complementation and control

• comparative study: German-Swedish-Dutch

  • Hartmann/Mucha/Trawiński/Wöllstein (in preparation)

  • Baseline: Wöllstein (2015), Brandt/Trawiński/Wöllstein (2016), Brand (2019).

• Subject:

  • Structures with verbs segregating propositional / verbal finite and non-finite complements

• Questioning:

  • Is there a correlation between preferences for (non)finite complementation and control ratios?

Case study: Background assumptions

• Referential cohesion and event integration (Givón 1990:527).

  • The more the two events coded in the main and complement clauses share their referents, the more likely they are to be semantically integrated as a single event; and the less likely is the complement clause to be coded as an independent finite clause.

• In relation to German (Rapp et al. 2017:197):

  • The more the embedding verb lexically tends to semantic control, the more often to-infinitives occur.

  • The less the embedding verb lexically tends to semantic control, the more frequently that-clauses occur.

Case study: The corpora

• Corpora:

  • DeReKo / KoGra-DB (IDS Mannheim, Kupietz et al. 2010)

    • → German

  • Språkbanken / Moderna (Gothenburg, Borin et al. 2012)

    • → Swedish

  • LASSY Large (Groningen / Lueven, van Noord et al. 2013)

    • → Dutch

Example: The monolingual corpora in the study

... on DE, SW, NL - Hartmann/Mucha/Trawiński/Wöllstein (forthcoming)

• Object of study: distribution of different verb classes with different complement types:

  • with finite complements

  • with infinite complements with a complementizer

  • with infinite complements without a complementizer

Results for the German

Verb categories 1: Swedish

Verb categories 2: Swedish

Verb categories 3: Swedish

Results for Swedish

Results for Dutch

Summary and conclusion

• The corpus studies show that there is a correlation between selection preferences and control ratios

  • the results confirm the hypothesis of referential cohesion and event integration

  • and show that it has cross-linguistic validity

• Methodological question: Are the results for German, Swedish and Dutch comparable?

• The answer: yes and no

  • the results are comparable on a meta-level (level of generalizations)

  • at the empirical level (data level) they are less comparable

Example: The monolingual corpora in the study

... to the DE, SW, NL - Hartmann/Mucha/Trawiński/Wöllstein (forthcoming)

Small parallel resources (selection)

Contrastive and typological studies

• Bengt Altenberg (1999). Adverbial connectors in English and Swedish: Semantic and lexical correspondences. In Hasselgård & Oksefjell (eds.) Out of Corpora. Amsterdam: Rodopi, 249-268.

• Hilde Hasselgård (2007). Using the ENPC and the ESPC as a parallel translation corpus: adverbs of frequency and usuality. Nordic Journal of English Studies 6:1.

• Sandrine Zufferey & Bruno Cartoni (2012). English and French causal connectives in contrast. Languages in Contrast, Volume 12, Issue 2, 2012, pages 232 -250.

Typological studies

Discussions

Case study: Imperative

• The underlying assumption: imperatives are used canonically in utterances expressing direct commands, requests, instructions, advice, etc.

• require an agent: an entity capable of conscious and deliberate action

  • → we can only ask someone to do something who has direct control over the action (Potsdam 1996, Jensen 2003).

• Expected corpus distribution: imperative markers occur significantly more often with agentive verbs than with non-agentive verbs.

Case study: Imperative

• Two goals:

  • test the agentivity hypothesis on the basis of corpus data

  • to test the cross-linguistic validity of the hypothesis

• 4 languages: English, German, Polish and Czech

  • Does the degree of relationship matter?

Data Source: InterCorp 6. Via KonText

Procedure

• Extraction of imperative word forms with the help of CQP queries

• Identification and selection of 50 most frequent lemmas (minimum frequency 10) from all lemmas underlying the imperative forms.

• Mapping of all selected lemmas (language specific) to abstract event concepts based on FrameNet frame index (Baker et al., 1998).

The relevant concepts

Distribution via concepts

The overall distribution

• The results confirm the hypothesis of agentivity:

  • Imperative markers are used significantly more often with agentive than with non-agentive verbs.

Language comparison

Example: The parallel data in the study

... to the imperative in DE, EN, PL, CZ - Trawiński (2016a, b)

Properties of translations: Laviosa (1998)

• a relatively lower proportion of lexical units over functional units

• a relatively higher proportion of high-frequency words over low-frequency words

• a relatively greater repetition of the most frequent words

• less variety in the words that are most frequently used

Properties of translations: Baker (1995)

• simplification

  • translations tend to use simpler language

• explicitation

  • translations show a tendency to spell things out

• normalization

  • translations tend to conform to the typical patterns of the target language and to overuse its features

Properties of translations: Teich (2003)

• shining-through (translations let the source language shine through)

• normalisation (translations obey the target language norm more than comparable target language original texts)

• Case studies for the German-English language pair: passive voice, transitivity, etc.

Case study: Passive (Teich 2003)

• Passive is more typical in English than in German

• Passive alternatives (man, sich lassen, -bar sein) are more typical in German than in English.

Case study: Passive (Teich 2003)

• Hypotheses for E-ORI-G-TL

• In translations from English into German, shining-through in the passive range is present if significantly more passives occur in G-TL than in comparable German texts (G-ORI).

• In translations from English into German, normalization in the passive domain is present if significantly more passive alternatives occur in G-TL than in comparable German texts (G-ORI).

Case study: Passive (Teich 2003)

• Hypotheses for G-ORI-E-TL

• In translations from German into English, shining-through in the passive domain is present if significantly more passive alternatives occur in E-TL than in comparable English texts (E-ORI).

• In translations from German into English, normalization in the passive range is present if significantly more passives occur in E-TL than in comparable English texts (E-ORI).

Case study: Passive (Teich 2003)

• Weak shining-through in G-TL

• No normalization effects in G-TL

Case study: Passive alternatives (Teich 2003)

• no shining-through in E-TL

• Normalization in E-TL

Possible requests

[drukola/m=pos:particle & drukola/m=type:negative] {,5} decît

Thematic composition of the corpus

according to CoRoLa taxonomy

Flat quantitative characteristics

Flat quantitative features: Word Types

Graphics all generated with KorAP-R client package

library(RKorAPClient)
library(tidyverse)
library(highcharter)
library(kableExtra)
 
D <- new("KorAPConnection", verbose=T)
B <- D
DRuKoLAVC <- "referTo+drukola.20180909.1b_words"
baseVC = "corpusSigle=/U[0-9][0-9]/ | corpusSigle=/W.D17/"
R <- new("KorAPConnection", KorAPUrl = "http://89.38.230.10:5555/", verbose=T)
Dsize <- corpusStats(D, vc=DRuKoLAVC)@tokens
Rsize <- corpusStats(R)@tokens
Bsize <- corpusStats(D, vc=baseVC)@tokens
queryResultToHtml <- function(r) {
  link <- slice(r, which.min(f))$webUIRequestUrl # use the query with less results
  text_spec(round(r[1,]$f,2), color="blue", link=link, tooltip = r[1,]$query %>% str_replace_all('"', '&#34;')) %>% str_replace(">", 'target="korap">')
}
 
add_comp <- function(.data, cat, n1, n2, b1) {
 
 
 
 

Design challenges

by "drukola-1b"

• automatic and dynamic construction of the comparison corpus not yet possible with KorAP

  • Downsampling function is still missing

• Text types classified differently and distributed very differently

• Topic taxonomies each 2 levels but defined differently

  • DeReKo based on Open Directory Project (dmoz) (Weiß 2005, Klosa et al. 2012).

  • CoRoLa (Tufiş et al 2016) based on Universal Decimal Classification (UDC) and Wikipedia top-level domains (Gîfu et al 2019).

Construction of "DRuKoLA-1b"

• Translation of the topic taxonomy of CoRoLA to that of DeReKo

  • e.g. Religion ➞ State/Society:Church, Art and Culture ➞ Culture, Medicine ➞ Health-Nutrition:Health

• illustrations could be found for 89% of the CoRoLa texts

  • some categories incomplete, others inaccurately mapped

• DeReKo large enough to fully mimic CoRoLa's theme distribution.

• instead of sampling CoRoLa and DeReKo , only DeReKo was sampled.

Comparability?

• whether corpora are sufficiently comparable cannot be decided in general, but depends, among other things, on the question to be answered

• dynamically definable virtual comparable corpora, are a good approach to solve this problem.

  • especially since this can also adjust the composition as a whole

• most important for now seems the practical comparability with the help of a uniform tool, mappable metadata and annotations

Levels of Access

Kupietz et al. (forthcoming)