1 of 86

1

Translation Mining and semantic maps: a new approach to cross-linguistic research?

Bert Le Bruyn Prague, 2025-4-2

2 of 86

Setting the stage

3 of 86

Setting the stage

Corpus research is about recognizing and interpreting patterns in data.

But what do relevant patterns in parallel corpora look like?

Rather than showing you, I invite you to try and come up with it by yourself!

Parallel corpus research is no exception.

4 of 86

Setting the stage

Here’s a fun fact about plural nouns in English and Romance:

In English, bare plurals can be used as indefinites (1) and in generic statements (2).

In Romance, bare plurals can be used as indefinites (3) but not in generic statements (4) (examples from Spanish). In generic statements, Romance languages rely on definite plurals.

  1. I bought cookies.
  2. Cookies are yummy!
  3. Compré bizcochos.
  4. *(Los) bizcochos son deliciosos.

What kind of pattern do you think these facts give rise to in a parallel corpus with English and Spanish in it?

5 of 86

Setting the stage

Tip:

Think of translation corpora as collections of contexts and of contexts as coming in one of four possible flavors:

<English Bare Plural; Spanish Bare Plural>

<English Definite Plural; Spanish Definite Plural>

<English Bare Plural; Spanish Definite Plural>

<English Definite Plural; Spanish Bare Plural>

6 of 86

Setting the stage

Spanish Definite Plurals

Spanish Bare Plurals

English Bare Plurals

English Definite Plurals

  1. I bought cookies.
  2. Cookies are yummy!
  3. Compré bizcochos.
  4. *(Los) bizcochos son deliciosos.

7 of 86

Setting the stage

Spanish Definite Plurals

Spanish Bare Plurals

English Bare Plurals

English Definite Plurals

  1. I bought cookies.
  2. Cookies are yummy!
  3. Compré bizcochos.
  4. *(Los) bizcochos son deliciosos.

8 of 86

Setting the stage

Is this the pattern we find?

Corpus: all contexts that have a bare or a definite plural in English and Spanish in subject or object position in chapters 1, 16 and 17 of the first Harry Potter volume (N=65).

23

12

29

1

Isn’t this pattern accidental?

9 of 86

The chessmen seemed to have been listening , because at these words a knight , a bishop and a castle turned their backs on the white pieces and walked off the board leaving three empty squares which Harry , Ron and Hermione took .

Las piezas parecieron haber escuchado porque , ante esas palabras , un caballo , un alfil y una torre dieron la espalda a las piezas blancas y salieron del tablero , dejando libres tres cuadrados que Harry , Ron y Hermione ocuparon .

(English, definite)

(Spanish, definite)

Setting the stage

10 of 86

Voldemort had powers I will never have.

Voldemort tenía poderes que yo nunca tuve.

(Spanish, indefinite)

(English, indefinite)

Setting the stage

11 of 86

Scars can come in useful.

Las cicatrices pueden ser útiles.

(Spanish, definite)

(English, indefinite)

Setting the stage

12 of 86

(Spanish, definite)

(English, indefinite)

But on the edge of town, drills were driven out of his mind by something else.

Pero en las afueras ocurrió algo que apartó los taladros de su mente.

Setting the stage

13 of 86

Training

14 of 86

15 of 86

16 of 86

17 of 86

Construction 1

Construction 2

Construction 3

Construction 4

1, 2, 4, 6, 7, 10, 11, 15, 16, 17, 20

3, 5, 8, 9, 12, 13, 14, 18, 19

1, 3, 4, 6, 8, 11, 12, 13, 15, 19, 20

2, 5, 7, 9, 10, 14, 16, 17, 18

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

1

Construction 1

Construction 2

Construction 3

Construction 4

18 of 86

Construction 1

Construction 2

Construction 3

Construction 4

1, 2, 3, 6, 7, 14, 17, 19

8, 13, 15, 16, 20

4, 5, 9, 10, 11, 12, 18

2, 6, 8, 13, 14, 15, 16, 19, 20

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

1

Construction 1

Construction 2

Construction 3

Construction 4

Construction 5

1, 3, 4, 7, 9, 17, 18

Construction 6

5, 10, 11, 12

Construction 5

Construction 6

19 of 86

Data analysis

20 of 86

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

1

Construction 1

Construction 2

Construction 3

Construction 4

Construction 5

Construction 6

Translation Mining consists in breaking up this complex relation into sets in which all contexts select the same constructions.

21 of 86

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

1

Construction 1

Construction 2

Construction 3

Construction 4

Construction 5

Construction 6

Translation Mining consists in breaking up this complex relation into sets in which all contexts select the same constructions.

22 of 86

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

1

Construction 1

Construction 2

Construction 3

Construction 4

Construction 5

Construction 6

Translation Mining consists in breaking up this complex relation into sets in which all contexts select the same constructions.

The researcher then proceeds to analyzing what it is that makes each of these sets of contexts different from the other sets.

23 of 86

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

1

Construction 1

Construction 2

Construction 3

Construction 4

Construction 5

Construction 6

Translation Mining consists in breaking up this complex relation into sets in which all contexts select the same constructions.

The researcher then proceeds to analyzing what it is that makes each of these sets of contexts different from the other sets.

The insights that follow from this analysis are subsequently used to propose analyses of the different constructions that account for why each of them appears in a number of sets but not in others.

24 of 86

Preparing the MDS move

25 of 86

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

1

Construction 1

Construction 2

Construction 3

Construction 4

Construction 5

Construction 6

Figuring out how to even draw the relations between the different constructions probably took you a little while longer than for the other datasets.

And yet, it’s just two languages with three constructions each and a mere 20 datapoints.

What would happen if we were to have – say – 5 languages, up to 8 constructions per language and about 400 datapoints?

26 of 86

Even though multiple representations are possible, we know, among other things, that:

>

should be positioned further away from

than from

>

the contexts belonging to the same colored set should be positioned closer to one another than to contexts belonging to different colored sets;

;

>

should be positioned further away from

than from

.

All of this amounts to positioning contexts on the basis of the constructions they appear in: contexts that appear in construction 1 in language A and in construction 5 in language B should be positioned closer to each other than contexts appearing in construction 5 in language B but in construction 3 in language A.

At the same time, contexts that appear in construction 3 in language A and in construction 6 in language B should be positioned even further away.

27 of 86

In order to be able to position contexts automatically on the basis of the constructions they appear in, it suffices to record for each context which constructions it appears in and let the computer handle the positioning.

The computer – and more specifically a Multi-Dimensional Scaling algorithm – uses similarities and differences between the above tuples to position the different contexts with respect to each other and the result will be exactly the way we want it to be.

28 of 86

For a more complex dataset than ours (including 5 languages, up to 8 constructions per language and 433 contexts), this looks as follows:

Each dot in this representation corresponds to a context in the dataset and sets of contexts appearing in the exact same constructions appear as bigger dots with the number. This number indicates the number of contexts in the set.

Drawing the different Venn diagrams representing the constructions of the different languages would make the figure uninterpretable but we can add color markup for one language at a time and compare the different constructions in the different languages.

29 of 86

No matter how complex the dataset, the Translation Mining strategy will always be the same: we analyze the contexts in the different sets to understand how the sets differ from one another and we take these insights as the basis for our analyses of the different constructions across the different languages.

30 of 86

Some technicalities

31 of 86

Context 1

Context 4

Context 2

Context 3

Context 65

32 of 86

https://sites.google.com/site/blebruyn/writings-and-presentations

> data

> dissimilarity_matrix

> cmdscale_input

> ggplot_input

33 of 86

R commands

> install.packages(“ggplot2”)

> library(ggplot2)

Definite and bare plural data

import file ‘definite_bare_plural_data’ > tab ‘cmdscale_input’

cmdscale(as.dist(definite_bare_plural_data),k=2)

import file ‘definite_bare_plural_data’ > tab ‘ggplot_input’

> m<-ggplot(definite_bare_plural_data,aes(Dimension_1,Dimension_2))

> m+geom_jitter(aes(colour=Language1))

> m+geom_jitter(aes(colour=Language2))

34 of 86

The Perfect

35 of 86

36 of 86

Passé Composé

37 of 86

Passé Composé

Perfekt

38 of 86

Passé Composé

Perfekt

VTT

39 of 86

Passé Composé

Perfekt

VTT

Pretérito

Perfecto

Compuesto

40 of 86

Passé Composé

Perfekt

VTT

Pretérito

Perfecto

Compuesto

Present Perfect

41 of 86

42 of 86

43 of 86

44 of 86

‘Classical’ Perfect contexts

45 of 86

46 of 86

‘Classical’ Perfect contexts

Narrative contexts

47 of 86

Nilsson (2016)

48 of 86

‘Classical’ Perfect contexts

Narrative contexts

Narrative contexts with stative verbs

49 of 86

50 of 86

‘Classical’ Perfect contexts

Narrative contexts

Narrative contexts with stative verbs

Non-narrative

past referring

contexts

51 of 86

52 of 86

Michaelis (1994)

53 of 86

‘Classical’ Perfect contexts

Narrative contexts

Narrative contexts with stative verbs

Non-narrative

past referring

contexts

Presuppositionality/

Interaction with

adverbs

54 of 86

Methodological reflections

55 of 86

Basic assumptions of parallel corpus research

  • The distributional hypothesis: distribution <> meaning.
  • The constancy of meaning hypothesis: translation is (by and large) meaning preserving.
  • The target language representativeness hypothesis: translations are representative of their target languages.

56 of 86

Caveats

The constancy of meaning hypothesis

  • Translation is a creative process (Rojo 2017) that cannot but lead to meaning alterations.
  • Whether or not this is a problem depends on the phenomena one is interested in.
  • Our work on the perfect shows that there are minor alterations in how translators compose the same meaning but that variation in tense use is typically dictated by the grammars of the different languages.
  • In Le Bruyn et al. (2024), we argue that – overall – the same holds for other tense-aspect categories.

57 of 86

58 of 86

59 of 86

Caveats

The constancy of meaning hypothesis

  • Translation is a creative process (Rojo 2017) that cannot but lead to meaning alterations.
  • Whether or not this is a problem depends on the phenomena one is interested in.
  • Our work on the perfect shows that there are minor alterations in how translators compose the same meaning but that variation in tense use is typically dictated by the grammars of the different languages.
  • In Le Bruyn et al. (2024), we argue that – overall – the same holds for other tense-aspect categories.
  • Role for ‘control’ items.
  • Potential opposition between more grammatical and more lexical items.

60 of 86

Caveats

The target language representativeness hypothesis

  • ‘translated language is at best an unrepresentative special variant of the target language’ (McEnery et al. 2006).
  • This claim builds on McEnery & Xiao (1999)’ s work on the CEPC-health corpus (35k words) and the C-health corpus (35k words). They found that the joint frequency of aspect markers le and guo is massively different between the translated and the untranslated corpus (98 vs. 213, LL value = 49.11, p < 0.001).

61 of 86

Caveats

The target language representativeness hypothesis

  • Similar claims by convinced Parallel Corpus researchers like Johansson for – among others – hate and love verbs in translated and untranslated Norwegian.

62 of 86

Caveats

The target language representativeness hypothesis

  • Later work makes similar claims for a range of phenomena (selection):

Olohan & Baker (2000) ENGLISH: that realization with say and tell

Teich (2003) ENGLISH | GERMAN: e.g., passive vs. active, full relative clauses vs. dense modification

Delaere et al. (2012) DUTCH: e.g., akkoord gaan met vs. akkoord zijn met, te veel vs. teveel

Evert & Neumann (2017) ENGLISH | GERMAN: e.g., frequency of imperatives, modal adverbs, subordination

Kruger & De Sutter (2018) and Kruger (2019) ENGLISH |SOUTH AFRICAN: that omission

De Baets et al. (2020) DUTCH: inchoative verbs

63 of 86

Caveats

The target language representativeness hypothesis

  • We take the qualms about target language representativeness seriously.
  • They are the reason the Translation Mining research cycle aims for replication with different source languages, allowing us to have – eventually –untranslated texts for every language.
  • At the same time, we’re convinced that the assumption of target language representativeness is far less problematic for cross-linguistic research than is typically assumed.

64 of 86

Caveats

The target language representativeness hypothesis

  • Claims about the lack of target language representativeness of translations build on the fact that are differences between translated texts and different (!) untranslated texts.

65 of 86

66 of 86

Caveats

The target language representativeness hypothesis

  • Claims about the lack of target language representativeness of translations build on the fact that are differences between translated texts and different (!) untranslated texts.
  • Claims about the lack of target language representativeness build on comparisons between small samples of translated and untranslated texts.

67 of 86

Le

guo

LCMC (1m words)

untranslated

9054

1168

ZCTC (1m words)

translated

8748

1175

Xiao & Hu (2015)

68 of 86

69 of 86

Caveats

The target language representativeness hypothesis

  • Claims about the lack of target language representativeness of translations build on the fact that are differences between translated texts and different (!) untranslated texts.
  • Claims about the lack of target language representativeness build on comparisons between small samples of translated and untranslated texts.
  • Claims about the lack of target language representativeness are often about phenomena that display optionality.

70 of 86

71 of 86

Kruger (2019)

72 of 86

Related paradigms

  • Corpus-based contrastive linguistics

‘Balanced corpus’: multiple authors, multiple translators, multiple genres.

The design is also meant to allow for an easy evaluation of target language representativeness through the inclusion of translated and untranslated texts for both languages.

Typical corpus: the ENPC

The design is conceived of in such a way that replication ought to be unnecessary.

73 of 86

Related paradigms

  • Primary Data Typology

Exemplary study: Wälchli & Cysouw (2012)

  • Lexical typology study on motion verbs.
  • Selection of 360 motion verb contexts from the gospel by Mark in a typologically diverse sample of 100 languages.
  • Goal: determine the main dimensions of variation across languages.
  • Output: MDS map with 30 dimensions + interpretation of the first 12 dimensions.

74 of 86

75 of 86

Related paradigms

  • Primary Data Typology

Exemplary study: Wälchli & Cysouw (2012)

  • ‘Convenience corpus’: the bible is one of the few texts that is really translated to a huge sample of languages, including underrepresented ones.
  • The fact that a (or the only available) convenience corpus is used might also explain why replication is typically not undertaken.
  • Analysis of higher order dimensions rather than of individual markers in individual languages.
  • Potential (smaller) influences of translation in individual languages are unlikely to have an impact on the analysis.
  • No other design feature that allows one to check target language representativeness.

76 of 86

Corpora and experiments

Two roles for experiments:

  • Triangulation.
  • Probe distinctions that are difficult to address on the basis of corpora alone.

77 of 86

78 of 86

Participants���Materials

  • N = 10
  • Students age 19-24
  • Native speakers of Dutch

  • Picture-based storytelling task
  • Cartoons of 6 frames
  • Brief instructions

PILOT EXPERIMENT

79 of 86

OVT/VTT Alternations in Dutch Sample Stimulus

PILOT EXPERIMENT

80 of 86

Procedure

  • Participants got randomly placed in one of two conditions (narrative vs. non-narrative).
  • Brief instructions based on condition:

  • Online, through voice-recording

NARRATIVE: Vertel het verhaal van Maria’s trip van gisteren naar het winkelcentrum. Begin met: “Er was eens…”

NON-NARRATIVE: Som de activiteiten van Maria’s trip van gisteren naar het winkelcentrum op.

PILOT EXPERIMENT

81 of 86

PILOT EXPERIMENT

82 of 86

Peter and Theresa are planning to go to a concert next weekend. Peter offers to go get the tickets later today, but Theresa tells him:

+T, +D: “I purchased / have purchased mine this morning

+T, -D: “I purchased / have purchased mine at midnight

-T, + D: “I purchased / have purchased mine last month

-T, -D: “I purchased / have purchased mine in November

It was cheaper that way”.

83 of 86

84 of 86

Replicating TM studies

85 of 86

https://sites.google.com/site/blebruyn/writings-and-presentations

86 of 86

Tense data

import file ‘cmdscale_input’

cmdscale(as.dist(cmdscale_input),k=2)

import file ‘ggplot_input’

> library(ggplot2)

> m<-ggplot(ggplot_input,aes(Dimension_1,Dimension_2))

> m+geom_jitter(aes(colour=English))+

+ scale_colour_manual(values=c("forestgreen","chartreuse","orange2","yellow","dodgerblue","deepskyblue"))

> m+geom_jitter(aes(colour=Dutch))+

+ scale_colour_manual(values=c("orange2","forestgreen","dodgerblue"))

> m+geom_jitter(aes(colour=French))+

+ scale_colour_manual(values=c("darkorchid","chartreuse","dodgerblue","cyan","firebrick","orange2"))

> m+geom_jitter(aes(colour=German))+

+ scale_colour_manual(values=c("dodgerblue","orange2","forestgreen"))

> m+geom_jitter(aes(colour=Italian))+

+ scale_colour_manual(values=c("chartreuse","dodgerblue","orange2","firebrick"))

> m+geom_jitter(aes(colour=Spanish))+

+ scale_colour_manual(values=c("chartreuse","darkorchid","deeppink","orange2","forestgreen","dodgerblue","firebrick"))