Translation-driven mapping of semantic fields of inceptiveness

Translation-driven mapping of semantic fields of inceptiveness Using bidirectional parallel corpus data for measuring and visualizing distances between lexemes in the semantic field of inceptiveness Lore Vandevoorde, Gert De Sutter & Koen Plevoets

Research goal To operationalize and to measure semantic differences in translated vs non-translated language

• Does the semasiological structure of polysemous and near-synonymous lexemes in translated Belgian Dutch differ from their structure in original Belgian Dutch? • What is the role of source language and register? pag. 2

Today’s research question Can we find a technique that enables us to retrieve and visualize semantic fields AND in addition allows for:   

register-sensitive comparison; cross-linguistic comparison; intra-linguistic comparison (translated vs. original language)?

pag. 3

CORPUS: Dutch Parallel Corpus -

Ten-million-word Sentence aligned Parallel and comparable Balanced for 5 text types (= registers / genres) -

-

External communication, journalistic texts, instructive texts, administrative texts, literature

Balanced for 4 translation directions -

Dutch -> French, French -> Dutch Dutch -> English, English -> Dutch pag. 4

TECHNIQUE: Semantic Mirroring Semantic Mirroring A technique for meaning differentiation that uses translational data from parallel corpora.

Dyvik, H. (1998). A translational basis for semantics. In S. Johansson & S. Oksefjell (Eds.), Corpora and cross-linguistic research: theory, method, and case studies (pp. 51-86). Amsterdam: Rodopi. Dyvik, H. (2004). Translations as semantic mirrors from parallel corpus to Wordnet. In K. Aijmer & B. Altenberg (Eds.), Advances in Corpus Linguistics (pp. 311-326). Amsterdam & New York: Rodopi.

pag. 5

Semantic Mirroring Advantages  Not predetermined selection of lexical items  Expansive: create semantic field

Dyvik (2004) “semantically closely related words ought to have strongly overlapping sets of translations” Overlapping sets of translations should reveal semantic relations

pag. 6

Semantic Mirroring

pag. 7

Semantic Mirroring: example Initial lexeme: bank

First T-image: English translations for bank bench, bank, desk

pag. 8

Semantic Mirroring: example Inverse T-image Dutch translations for bench zitbank, bank

Dutch translations for bank oever, geldbedrijf, bank

Dutch translations for desk werktafel, schoolbank pag. 9

Case study semantic field of inceptiveness in Dutch (BEGINNEN) and French (COMMENCER)

pag. 10

Initial lexeme

Select a concise set of nearsynonyms using lexicographic data and substitution testing

pag. 11

Initial lexeme

Look up synonyms of beginnen in 8 dictionaries (no meta-information)

19 / 104 synonyms in at least 3 dictionaries

Substitution testing for 19 synonyms

Set of 5 synonyms

pag. 12

Initial lexeme Beginnen Aanvangen Starten Een aanvang nemen Van start gaan Aanvatten

pag. 13

Apply semantic mirroring

pag. 14

Create First T-image

Check all French translations of the SET of variants of BEGINNEN

74 different translations

Minimal overlap criterion + frequency threshold 5

T-image of 12 French lexemes

pag. 15

Apply semantic mirroring

Query T-image lexemes from the corpus

Check translations back into Dutch

Apply minimal overlap criterion + frequency threshold 5

Inverse T-image of 22 Dutch lexemes

pag. 16

SELECTION

22 Dutch lexemes of inceptiveness aanvang-

invoeren

aanvangen

lanceren

aanvankelijk

openen

aanvatten

oprichten

begin

opstarten

begin-

start

beginnen

start-

eerst (adj)

starten

gaan

van start gaan

ingaan

vanaf

instellen

vertrekken

pag. 17

Insert frequencies

VISUALIZATION

BUT

Inverse T-image = TL frequencies

Additional step

pag. 18

Visualize Data are summarized in frequency tables commencer début débuter démarrer départ aanvang0 4 0 0 aanvangen 0 1 3 1 aanvankelijk 1 9 0 0 aanvatten 1 0 0 2 begin 2 264 2 0

…

entamer instaurer lancer 0 0 0 0 1 0 3 0 0 0 8 1 9 0 0

lancer, se ouvrir 0 0 0 0 0 0 0 0 1 0

prendre cours

partir 0 0 0 0 0

0 1 0 0 0

0 3 0 0 0

Correspondence Analysis a statistical technique designed for the visualization of frequency tables (Greenacre, 2007) pag. 19

2 1

aanvangbegin beginaanvankelijk

openen

0

2

startstart eerst (adj)

-1

aanvangen vertrekken ingaanbeginnen gaan starten vanaf van start gaan aanvatten opstarten lanceren oprichten

-2

invoeren instellen

n = 5323

-3

-2

-1 1

0

1

pag. 20

First-order Correspondence Analysis Same word class lexemes seem to cluster together This would imply that: - aanvankelijk [firstly] and begin [beginning] are semantically closer to each other than begin [beginning] to beginnen [to begin]

pag. 21

Solution?  

Insert earlier stages of the Semantic Mirroring into the analysis Exclude possible interaction with other semantic fields

Base the semantic space on the six lexemes and their first-order translations

pag. 22

Second-order Correspondence Analysis Anchor 12 points (French lexemes) with respect to the 6 Dutch lexemes (Initial lexeme)

pag. 23

1 2

-1

0

lancer, se commencer départ beginnen début ouvrir starten partir lancer entamer démarrer débuter van start gaan instaurer aanvatten

-2

aanvangen prendre cours

-4

-3

een aanvang nemen

-3

-2

-1

0 1

1

2

pag. 24

Second-order Correspondence Analysis Visualize the 22 Dutch lexemes based on the 12 anchored points (“supplementary” points)

pag. 25

1 0

eerst (adj) beginnen lancer, se commencer aanvankelijk départ startaanvangstart openen begin début ouvrir begin- gaan starten partir vertrekken lancer entamer démarrerlanceren débuter instaurer van start gaan opstarten oprichten

2

-1

aanvatten invoeren instellen

-2

prendre cours

-3

aanvangen

-4

vanaf ingaan

-3

-2

-1

0 1

1

2

pag. 26

1 0

eerst (adj) beginnen aanvankelijk startaanvangstart openen begin begin- gaan starten vertrekken lanceren van start gaan opstarten oprichten

-2

2

-1

aanvatten invoeren instellen

-3

aanvangen

-4

vanaf ingaan

n = 5323

-3

-2

-1

0 1

1

2

pag. 27

- Separate clusters are now meaningbased and formed independently of the word class of each lexeme - Lexemes follow a descending line towards more outlying, more formal lexemes

pag. 28

Semantic field of BEGINNEN: journalistic texts

pag. 29

1

Journalistic texts

eerst (adj) beginnen

0

start aanvankelijk beginopenen begin startgaan vertrekken starten oprichten lanceren invoeren van start gaan

-3

-2

2

-1

opstarten

-4

ingaan

n = 1468

-3

-2

-1

0 1

1

2

pag. 30

Semantic field of BEGINNEN: external communication

pag. 31

1

External communication

beginnen

0

startbegin start eerst begin(adj) openen vertrekken starten

aanvangen lanceren

-1

gaan van start gaan

2

aanvatten invoeren

-4

-3

-2

opstarten

n = 928

-3

-2

-1

0 1

1

2

pag. 32

Register-specific plots provide us with interesting insights into the influence of register on the structure of semantic fields

pag. 33

Cross-lingual comparison BEGINNEN and COMMENCER

pag. 34

0.5

1.0

partir

débuter

0.0

lancer démarrer mettre sur pied

lancement

départ créer lancer, se

commencer ouvrir début

-0.5

entamer

-1.0

2

entreprendre

-1.0

-0.5

0.0 1

0.5

1.0

pag. 35

Intra-lingual comparison translated and original BEGINNEN

pag. 36

original 1

1

translated beginnen startaanvankelijk aanvangeerst (adj) start gaan begin openen begin-

0

0

eerst (adj) beginnen aanvankelijk startaanvangstart openen begin begin- gaan starten vertrekken

vanaf vertrekken oprichten lanceren starten opstarten van start gaan aanvatten

lanceren van start gaan opstarten oprichten

-1

invoeren instellen aanvangen

-2

-2

2

-1

aanvatten invoeren instellen

-3

-3

aanvangen

ingaan

-4

-4

vanaf ingaan

-3

-2

-1

0 1

1

2

-3

-2

-1

0 1

1

2

pag. 37

- Many similarities between the translated and original semantic field - Noteworthy differences: To what extent can translated language represent original language? pag. 38

Conclusion The developed translational method is bottom-up, quantitative, corpus based and enables us to measure and visualize translated and original semantic fields, even register-specifically, which was our main methodological concern. BUT: - Extra testing with a different pivot language - Carry out a Semantic Mirroring with a single lexeme

pag. 39

Questions for the future If we are able to visualize semantic structures (including register-specific ones) in translated and non-translated texts, then we can aim at providing evidence for the Gravitational Pull Hypothesis (Halverson 2003, 2007).

Apply semantic mirroring to a field with more salient semantic differences pag. 40

Possibilities for the future - Automatization of the retrieval process (using word-aligned corpora) - Add morpho-syntactic ID-tags to the translation annotated sentences

pag. 41

Thank you! [email protected] [email protected] [email protected]

pag. 42