M. Kupietz, R. Cosma, D. Cristea, N. Diewald, B. Trawiński, D. Tufiş, T. Váradi, A. Wöllstein

Recent Developments in the European Reference Corpus EuReCo

UCCTS 2018, Louvain-la-Neuve, 2018-09-12

Overview

1. Introduction

2. European Reference Corpus EuReCo

3. KorAP Demo

4. Summary and Future Plans

Need for comparable corpora

• parallel corpora alone are not suitable for finer-grained contrastive research

  • because of several translation effects:

    • simplification, explication, normalization (Backer 1995)

    • distributional biases (Laviosa 1998)

    • shining through effect (Teich 2003)

• possible workaround: use a combination of parallel and monolingual corpora

  • however: complicated to handle for typical use cases

• high-quality comparable corpora would be nice to have
  

No large, truly comparable corpora available

• one exception:

  • Aranea – Family of Comparable Gigaword Web Corpora (Benko 2014)

  • however with uncontrolled comparability

• in future – hopefully:

  • International Comparable Corpus (ICC) (Kirk & Čermáková 2017)

• OTOH: there are many large national and reference corpora available:

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

Idea of the EuReCo-initiative

• use these existing national and reference corpora

• and some software / infrastructure

• to join sub-samples of them virtually and dynamically to (pairs of) comparable corpora

• to ideally enable the user to, e. g.:

  • build the largest possible German-Hungarian comparable corpus with equal distribution of text types, topics and publication years starting from year 2000

  • to perform contrastive analysis on these virtual comparable corpora

Virtual joining is crucial

… because corpora are typically tied to their hosting institutions

Approach to achieve comparability

• draw sub-samples of the existing corpora

  • based on metadata properties of individual texts:

    • topic domain

    • text type

    • publication date

    • …

  • so that the token distribution wrt. these metadata properties is similar in both sub-samples

• challenge: in existing corpora these metadata properties are unlikely to be 1:1-mappable

Current state of EuReCo

Two (pilot) projects running

• German-Romanian: DRuKoLA (2016-2018)

  • IDS, Bucharest University, Romanian Academy in Bucharest and Iaşi

  • state: first comparable corpora ready

• German-Hungarian: DeutUng (2017-2019)

  • IDS, Szeged University, Hungarian Academy

  • state: first HNC sample available via KorAP

• both projects funded by the Humboldt-Foundation

EuReCo’s software base: KorAP

• corpus analysis platform developed mainly at the IDS as COSMAS-successor

• beta running since 5/2017

• still some important missing features:

  • most quantitative functions

  • sampling function for the creation of virtual corpora
    (currently performed externally)

KorAP key features for EuReCo

• KorAP can work with physically distributed data

  • all corpus data can stay where it is

• allows dynamic definition of (comparable) sub-corpora

• no relevant restrictions on data size and annotations

• supports multiple query languages

• has sustainable funding

• is open source (BSD-licensed)

KorAP - Features for Comparable Corpora

DeReKo/CoRoLa

DeReKo/HNC - Contrastive Example

Molnár 2015

DeReKo/HNC - Demo

Summary

• we need corpora for language comparison

• big, overall goal:

  • 

• not met by existing corpora

• EuReCo & KorAP might provide a solution

  • by virtually joining existing national corpora
    to comparable corpora

Next EuReCo steps

• refine the current mappings by adding time and text type as additional comparability factors

• integrate the HNC

• explore differently defined comparable corpora

  • by their effect on quantitative distributions wrt case studies

• implement missing KorAP features

• draft an integrated UI for working with comparable corpora

• ask more corpus providers to join

  • don't hesitate to contact us if you are interested

References

Multilingual corpora

It has often been said that, through corpora, we can observe patterns in language which we were unaware of before (…) My claim is that this applies particularly to multilingual corpora. We can see how languages differ, what they share and – perhaps eventually – what characterises language in general.

Johansson (2007)

Comparable corpus

Definition

• a comparable corpus consists of two or more monolingual corpora that have similar compositions wrt relevant properties (time, genre, topic domain, …)

  • (ideally) contains original texts only

• early prominent example:

  • International Corpus of English (ICE) (Greenbaum 1991)

  • contains twelve corpora of different national or regional varieties of English with a controlled, similar composition

Problem: IPR and license contracts

… typically tie existing corpora to their hosting institutions

If the data cannot move …

…find ways to put the computation near the data

Jim Gray (2003)

Translated to EuReCo

… build some infrastructure to use the data from where it is.

How to achieve “comparability”?

Basic approach: (Cosma et al. 2016)

• draw comparable sub-samples of the existing corpora

  • using KorAP's virtual collection feature

• based on metadata properties of individual texts:

  • topic domain

  • text type

  • publication date

  • …

• comparable ≙ equal token distribution wrt. these metadata properties

Apply this approach iteratively

to gradually approximate good/sufficient comparability

1. start with a good mapping of metadata properties

2. define a preliminary comparable corpus pair

3. perform comparative case studies

4. refine mapping, if findings seem to be artifacts of comparability criteria and start over with 2

   1. side-question: how stable are the quantitative effects over different comparable corpora

5. 

Mapping of CoRoLa topic domains to DeReKo

In order to construct a first comparable corpus

Risks & Challenges 1: Comparability

• text property taxonomies are not easily 1:1-mappable between different corpora

• text type / genre: pretty hopeless

• topic domain

  • several internationally established taxonomies: e.g. UDC, Open Directory, Wikipedia Categories

  • but none of them alone seems ideal for corpora

Risks & Challenges 2:

Composition & Coverage

• comparable corpus composition and coverage depend on the composition of the respective underlying monolingual corpora:

  • each comparable stratum is limited to the size of the lesser populated one in the monolingual sources

• because of the decentral structure of EuReCo missing/underpopulated strata cannot be added easily

  • e.g. Wikipedia Talk Pages in CoRoLa

Risks & Challenges 3:

Reliance on a single software

• KorAP is open source and so are the APIs

• it's sometimes easier to deploy an additional corpus query tool in many places than to develop and maintain an equivalent API for many corpus tools (my lessons learned from CLARIN)

• solution: don't replace your existing query tool with KorAP, but run KorAP in addition

  • also at the IDS, KorAP is currently only a secondary query tool

Arbitrary number of annotation layers

Romanian: Adjective serialization

Assertive verbs with or without correlate

CoRoLa live

http://corola.racai.ro/

DRuKoLa/CoRoLa Comparation