Lithuanian researchers are working on enriching the existing corpora; they are also looking for ways to make the corpora inter-operable and co-searchable through the annotation of discourse relations. One of the goals of the present research is working on the annotation of discourse relations in TED talks transcripts translated into Lithuanian and expanding the set of available resources in the Lithuanian language. A second goal is to compare cross-linguistically the annotated texts with the view of looking for translation tendencies in rendering discourse relations in the Lithuanian language. This, we believe, will open up a new research path in digital humanities leading to an understanding of translation tendencies in TED talks transcripts across languages. According to our research results, noteworthy translation tendencies embrace explicitation - a tendency to use more explicitly marked discourse relations in Lithuanian than the original transcripts, verbatim translations of discourse connectives, and also a tendency to use fewer alternative lexicalizations (a type of discourse-relational devices).
Lithuanian researchers are working on enriching the
existing corpora and are also looking for ways to make
the corpora inter-operable and co-searchable through the
annotation of discourse relations. One of the aims of
the current research is extending the available resources
and lexicons of discourse-relational devices in
Lithuanian cooperating with the international team of researchers
brought together by the European COST Project TextLink
(http://www.textlink.ii.metu.edu.tr/). The aim is partially
achieved by adding Lithuanian annotated texts to the
existing TED Multilingual Discourse Bank, or TED-MDB,
a parallel corpus annotated at the discourse level
following the goals and principles of Penn Discourse Treebank
The section provides some general insights on Lithuanian, describes discourse connectives (DCs), and briefly outlines the PDTB annotation scheme. It also describes the data and presents some observations about the data. 2.1.
Lithuanian is a very old Indo-European language. It is a Baltic language which has conservative morphology, e.g. it has preserved morphological aspects of the protolanguage, such as the word declensions. It is spoken by about 2,900,000 native Lithuanian speakers in Lithuania and about 200,000 abroad.
There are two main resources for modern Lithuanian: (a)
The 9-million-word Corpus Academicum Lithuanicum –
CorALit (http://coralit.lt) compiled by Vilnius University.
It contains academic texts from the fields of biomedical
sciences, humanities, physical sciences, social sciences,
and technological sciences. (b) The 102-million-word
online corpus of the Contemporary Lithuanian Language
(http://tekstynas.vdu.lt), which is of general character and
includes publicist texts, fiction, non-fiction, administrative
literature and spoken language. However, parallel
corpora involving Lithuanian are still insufficient; currently
only one parallel two-directional (English - Lithuanian and
Lithuanian - English) corpus exits comprising English
Lithuanian (70,813 parallel sentences) and Lithuanian
English (1,614 parallel sentences) (http://tekstynas.vdu.lt).
Furthermore, the corpus is not discourse-annotated. Such
scarcity of corpora resources is an obvious barrier for
machine translation
Discourse connectives signal the way the writer or speaker
would like the reader or listener relate the ideas that are
about to be said to the ideas that have been said before.
According to
The literature suggests that some languages tend to express
discourse relations (DRs) through complex structures while
others prefer to use simpler structures and mark discourse
relations explicitly, as for example, the difference between
English and Arabic illustrates
Our data comprise Lithuanian TED talks transcripts of the
original English texts included in TED-MDB (Table 1).
TED-MDB is created on the basis of PDTB 3.0 relation
hierarchy
Both the TED website and the WIT3 website are open
resources, which is attractive to research as they present
numerous advantages, e.g. subtitles are available in a
substantial number of languages, and the topics cover a wide span
of knowledge fields, making the data applicable in
multiple domains
In Lithuanian, explicit DCs include expressions from four
grammatical classes: subordinating conjunctions – e.g. kai,
kol, nes, kadangi (when, while, because, since),
coordinating conjunctions – ir, bei, o, tacˇiau (and, but, or, however),
sentential relatives – tam kad, tuo metu kai (so that, at the
time when), and discourse adverbials – faktiškai,
galiausiai (actually, eventually). The main task is to identify if
the words and phrases function as explicit DCs as they can
have other functions. As in the PDTB, five types of
relations are identified and annotated: Explicit relations,
implicit relations, alternative lexicalizations, entity relations,
and no relations. The argument annotation of explicit DCs
and alternative lexicalizations follows the rule that the
argument which appears as syntactically bound to the DC is
marked as Arg2; the other argument is annotated as Arg1.
As in TED-MDB, adverbials called “discourse markers”
1. Dabar kaip matote ˛itampa apie kuria˛ girde˙jome San Fransiske apie susiru¯pinima˛ de˙l bu¯sto kainu˛ ir gyventoju˛ išstu¯mimo ir technologiju˛ kompaniju˛, kurios atneša daug turto ir ˛isikuria, yra tikra. 2. Now you can see, though, that the tensions that we’ve heard about in San Francisco in terms of people being concerned about gentrification and all the new tech
Title/Speaker The investment of logic for sustainability (Chris McKnett)
Embrace the near win (Sarah Lewis)
A glimpse of life on the road (Kitra Cahana) Social maps that reveal a city’s intersections and separations (Dave Troy) Word count Eng./Lith.
1,614 (1,345) 1,772 (1,362) 694 (512) 1,053 (678) 5,133 (3,897) companies that are bringing new wealth and settlement into the city are real.
No relation (NoRel) is annotated when there is no DR inferred by the reader between the adjacent sentences: According to PDTB annotation guidelines, in annotating implicit DRs, the annotator has to insert a DC that best expresses the inferred relation between two adjacent sentences. This procedure is adopted, as in Lithuanian example 3 and its English equivalent in 4. In all the examples, Arg1 is shown in italics, Arg2 is shown in boldface.
3. Ji tokie sude˙tingi ir gali atrodyti mums tolimi, kad galime bu¯ti linke˛ daryti štai ka˛: sle˙pti galva˛ sme˙lyje ir negalvoti apie tai. [Implicit=Bet] Jei tik galite, priešinkite˙s tam. (Implicit) (Comparison: Contrast) 4. ...bury our heads in the sand and not think about it. [Implicit=But] Resist this, if you can. (Implicit) (Comparison: Contrast) Alternative lexalization (AltLex) includes cases of inferred DRs between adjacent clauses, where redundancy appears if an explicit DC is inserted. The reason for this is that the relation is already expressed by some alternatively lexicalized non-connective expression, e.g.
5. Se˙kme˙ mus motyvuoja, bet beveik pasiekta pergale˙ skatina mus leistis ˛i nuolatinius ieškojimus. [Viena˛ iš ryškiausiu˛ to pavyzdžiu˛ pastebime], kai žvelgiame ˛i skirtuma˛ tarp olimpinio sidabro laime˙toju˛ ir bronzos laime˙toju˛ rungtyne˙ms pasibaigus. (AltLex) (Expansion: Instatiation) 6. Success motivates us, but a near win can propel us in an ongoing quest. [One of the most vivid examples of this comes] when we look at the difference between
ter a competition. (AltLex) (Expansion: Instantiation) Entity relations (EntRel) are annotated between adjacent sentences when an entity in one argument is described further in the other argument, as in 7 and its English version in 8.
7. Jie ture˙tu˛ ˛ivertinti ir tuos efektyvumo rodiklius, kuriuos vadiname ASV: aplinkosauga, socialiniai klausimai ir valdymas. Aplikosauga apima energijos vartojima˛, prieiga˛ prie vandens, atlieku˛ tvarkyma˛ ir tarša˛ ir ekonomiška˛ ištekliu˛ naudojima˛. (EntRel) 8. Investors should also look at performance metrics in what we call ESG: environment, social and governance. Environment includes energy consumption, water availability, waste and pollution, just making efficient uses of resource. (EntRel) 9. Tai 4 milijardai viduriniosios klase˙s žmoniu˛, kuriems reikia maisto, energijosir vandens. Dabar ju¯ s tub u¯t klausiate save˛s: gal tai tik pavieniai atvejai. (NoRel) 10. That’s four billion middle class people demanding food, energy and water. Now, you may be asking yourself, are these just isolated cases. (NoRel) TED-MDB adds a new top-level category to the PDTB 3.0 relation hierarchy, called hypophora. This category aims to capture rhetorical question-response pairs, where the question is asked and answered by the speaker. TED-MDB annotates hypophora as a case of AltLex anchored by the question word. Where possible, the additional sense of the Q/R pair may be added.
As in TED-MDB, in Lithuanian, we annotate the question as Arg2, the answer as Arg1. We consider the question as Arg2 because the AltLex is part of the question. The question word (either the wh-word or ar, a specific question particle used in Yes/No questions, which can also serve as an explicit DC in Lithuanian) is selected as AltLex since it marks the DR holding between the question and the answer, as in example 11 and its equivalent in 12: 11. Niekas nepasikeis, [ar] mes bandysime pakeisti, [ar] tu nieko nebandysi (Explicit) (Expansion: Disjunction) 12. Nothing is going to change [either] we try to change something [or] you don’t try anything. (Explicit) (Expansion: Disjunction) In the following pairs of examples, we provide more cases of how hypophora is annotated in Lithuanian and English. Lithuanian Q/R pairs are annotated for a primary sense, and tagged as hypophora as the secondary sense. 13. [Ar] ˛imone˙s, atsižvelgiancˇios ˛i tvaruma˛, išties finansiškai se˙kmingos? galintis nustebinti atsakymas yra “taip" (Explicit) (Altlex: Ar; Expansion: Level-ofdetail:Arg1-as-detail; Hypophora). 14. [Do] companies that take sustainability into account really do well financially? The answer that may surprise you is yes. (AltLex: Do) (Hypophora) 15. [Kode˙l] kas nors apskritai rinktu˛si tok˛i gyvenima˛ - Atsakymas ˛i š˛i klausima˛ gali skirtis, kaip skiriasi ir žmone˙s sutinkami kelyje, bet keliautojai dažnai atsako vienu žodžiu: laisve˙. (Explicit) (Altlex: Kode˙l; Contingency: cause: Reason; Hypophora). 16. [Why] anyone would choose a life like this, under the thumb of discriminatory laws, eating out of trash cans, sleeping under bridges, picking up seasonal jobs here and there. The answer to such a question is as varied as the people that take to the road, but travelers often respond with a single word: freedom. (AltLex: Why)(Hypophora)
The stability of the annotation scheme is evaluated both
by intra- and inter-annotator agreement. One transcript
(Text ID 1978), which comprises approximately 25% of
the Lithuanian section of the data is reannoted by the
primary annotator after about 2 months of the first annotation,
and it is annotated independently by the secondary
annotator (cf. Table 2 for the distribution of the annotated,
reannoated and independently annotated DR types).1 We
measured F1 score, which evaluates agreement between the
annotators regarding the existence of a DR between the same
discourse units. To measure agreement on the types and
senses of these DRs, we calculated Cohen’s Kappa
Relation Type
AltLex NoRel Explicit Implicit EntRel
Primary annotator 1st annot 2nd annot - 2 15 15 105 107 48 53 28 30
Secondary annotator
In this section, we focus on the whole unit of the annotated texts in English and Lithuanian and present the frequencies of annotated DR types (Table 3) as well as the frequencies of the annotated top-level senses (Table 4). We then discuss the results.
In Table 3, the low frequency of AltLex annotations in
Lithuanian could reveal a certain tendency characteristic
reflecting the translators’ choices while translating the DCs
it appears that the translators tended to render DCs by the
variants provided by dictionaries rather than using AltLexs,
e.g. kai (when), kol (while), nes (because), nes (since), etc.
This resonates with Baker´s
1The primary and the secondary annotators are the first and the
third authors of the study.
Another interesting feature observed is that there are more
explicit DRs in the Lithuanian transcripts than in the
English versions. This might be explained by the translators’
effort to render the implicit DRs in English explicitly. There
are also cases where implicit DRs in English texts are
translated explicitly to Lithuanian, which goes in tune with
explicitation, as observed by
Finally, the annotation of EntRels also revealed some interesting cases. We observed that in some Lithuanian translations, the EntRel is present in two loosely related sentences as in 23, while in the source English text there is just one sentence lacking two separate arguments (see 24): 23. Tad pasakysiu kai ka˛, kas gali jus nustebinti: galios balansas, galintis išties paveikti tvaruma˛, yra instituciniu˛ investuotoju˛ rankose. Tai tokie didieji investuotojai kaip pensiju˛ fondai, kiti fondai ir labdaros fondai. (Entrel) 24. And here’s something that may surprise you: the balance of power to really influence sustainability rests with institutional investors, the large investors like pension funds, foundations and endowment.
Concerning the frequencies of the top-level senses of DRs, the distribution seems to be approximately equal for both languages as indicated in Table 4.
The research findings presented here represent our initial
observations and reveal certain tendencies in rendering the
discourse of English TED talks in Lithuanian. Our focus
has been on how DRs are expressed in Lithuanian
transcripts. We observed that there are more explicit DRs in
the Lithuanian transcripts, which might be explained by the
translators’ efforts to render the implicit DRs explicitly
this goes in tune with the observations of
In the future, by annotating more of the Lithuanian transcripts of the English texts in TED-MDB, we hope to reveal and specify more translation tendencies. Also, by exploring the transcripts further, we expect to find out what translation strategies (direct translation, transposition, etc.) are preferably employed by the translators and what this may add to the research field of digital humanities.
This research is funded by the European Social Fund under the No 09.3.3-LMT-K-712 “Development of Competences of Scientists, other Researchers and Students through Practical Research Activities” measure. For training in annotation and generating ideas for research, we acknowledge the support of the STSM grants by TextLink COST action IS1312.