Content uploaded by Katherin Perez
Author content
All content in this area was uploaded by Katherin Perez on Mar 24, 2015
Content may be subject to copyright.
From Translation Research Projects 5, eds. Esther Torres-Simon and David Orrego-Carmona,
Tarragona: Intercultural Studies Group, 2014. pp. 51-62.
http://isg.urv.es/publicity/isg/publications/trp_5_2014/index.htm
Automatically building translation memories
for subtitling
KATHERIN PÉREZ ROJAS
Universitat Autònoma de Barcelona, Spain
University of Wolverhampton, United Kingdom
This article describes a methodology to automatically create translation
memories for subtitling, using translated books adapted into films and
recognising extra-linguistic markers to differentiate character
interventions from narration. This methodology includes the automatic
identification, extraction, and alignment of the dialogues. The aligned
bi-texts served as translation memories in the subtitling of the adapted
films. Results show an overall 95% extraction rate for English dialogues
and 85% for Spanish dialogues. Alignment showed an accuracy of 90%.
Results for the translation memory performance showed that hits between
70% and 100% matches accounted for 15% of the corpus. The results
reinforce the claim that dialogues in books can be used as reference
material for the translation of subtitles.
Keywords: audiovisual translation; translation memories; subtitling;
natural language processing
Introduction
The use of automatic tools to help in the translation process is a field with a
great amount of academic research and well proven usability in the market.
However, in audiovisual translation, and more specifically in subtitling, there
are very few tools to support this process by granting access to the creation of
subtitles (Mejías 2010). The lack of resources for automatic subtitle creation
may result from the difficulty of processing an ever changing language.
Unlike scientific, legal, or even literary texts, audiovisual material does not
follow guidelines and cannot be framed in a single type of language.
This research focuses on developing a methodology to build,
automatically, translation memories to assist in the subtitling process.
52 Katherin Pérez Rojas
Literature review
The Code of Good Subtitling Practice (Carroll and Ivarsson 1998) establishes
guidelines for the subtitling task. The authors propose a standardisation of
technical features including length of subtitles, number of characters per line,
duration on screen, font and size of the letters, and of language features like
coherence, cohesion, and treatment of interjections. However, with the
increasing demand for subtitling, short deadlines and tight budgets, translators
often need the help of more automated or semi-automated methodologies to
help them follow good practices and meet deadlines (SUMAT 2011).
In recent decades, researchers in audiovisual translation and machine
translation have begun to combine the two fields in order to provide
audiovisual translation techniques with tools that focus not just on the
technical aspects but also on the translation. The support computer-based
systems offer to the subtitling tasks mainly focuses on mechanical aspects
such as time coding and word processing, while the possibility to reuse
previous translations in new translation assignments remains unaided.
Previous studies have approached the reusability of translations in
audiovisual material and automated subtitling. The STAR project developed a
rule-based machine translation system to produce Japanese subtitles for
English news programmes and Japanese subtitles for newswire translation
services (Sumiyoshi et al. 1995: 4). The project Global Translation Systems
(GTI), Inc. (Díaz Cintas and Remael 2007: 20-21) uses SYSTRAN to
translate, in real time, English subtitles into Spanish subtitles on selected
television programmes. SUMAT (2011) is an on-going project to develop an
online service for subtitling through machine translation in nine different
European languages, with the aim being to semi-automatize the subtitle
translation processes on a large scale.
However, when it comes to films, specifically those that are adapted
from novels, the translator could rely on the same source the scriptwriters
used: the written novel. Harrington (1977) estimates that a third of all films
ever made have been adapted from novels, without including other literary
forms (such as plays or short stories). The Writing Studio (2001-2004)
confirms that over fifty percent of feature-length films for both cinema and
television are adaptations of novels, short stories, plays or nonfiction
journalism, which together account for 25 percent of all adapted feature films.
It is safe to say that almost all great literary works have been adapted at
least once in cinema history. When adapting a novel into a script for a film,
and especially when writing a screenplay, “[t]he essence of dialogue and
subtext should stay the same […] despite several tricks to “cull and shape the
cinematic elements” (Online Film School). When adapting, the writer should
compress all dialogues “so that it has the economy and directness of screen
dialogue” (ibid). Dialogue in literature is mimetic (as opposed to diegetic), the
writer tries to “create the illusion that it is not he who speaks”, therefore
Automatically building translation memories for subtitling 53
dialogue in novels is generally direct discourse; a quotation of a character’s
words (Rimmon-Kenan, 1983). Also, Lotman (1989) (as quoted by Rauma
2004) argues that cinematic dialogue is equivalent to dialogue in novels or
plays and is thus an indistinctive property of the film medium.
Methodology
It may be possible to use the data contained in novels that have been adapted
into films as reference material for the translation of the subtitles for the films.
The present research thus seeks to ascertain whether that material can create
translation memories suitable for the production of translated subtitles.
To work on this hypothesis, a series of experiments were designed and
conducted with the objective of defining a method to search for relevant data
in the novels, extract it and create translation memories. We assessed the
efficacy of the search and extraction of the information, the alignment, and the
matching percentages in the translation memory.
The automatic creation of the translation memories consisted of four
stages. The first stage analyzed five film-adapted novels in English and their
Spanish translations and extracted the common dialogue features in both
languages. The second stage focused on the creation of rules to edit and
standardize the texts for processing, the creation of the dialogue extraction
scripts and the testing of eight further novels in English and their translations
in Spanish. The third stage focused on the alignment of the dialogues
extracted and the creation of a TMX file, fine-tuning the pre-processing and
extraction scripts as well as the revision of the extraction using seven
additional novels in English and their translations into Spanish. The final
stage focused on collecting and analyzing the data, and a first evaluation of
the translation memory. In all, 20 novels in English and their corresponding
translations were analyzed and tested for dialogue extraction.
Definition and characterization of dialogues in English and Spanish novels
Literary theorists (Bakhtin 1986; Maranhão 1990) define dialogue as the
conversation, or the literary work in the form of a conversation, between
characters, often used as a mechanism to reveal characters and to develop and
make the plot advance. Dialogues are the lines that a character speaks in any
literary work.
Dialogues can be diferentiated from the surrounding prose by using
punctuation marks. Sherlock (2011) states that “in standard American usage,
opened and closed double quotation marks [“ ”] indicate when narration has
stopped and a character’s dialogue begins.” This is supported by Elson and
Mckeown (2010), who state that, in English literary novels, quoted speech or
54 Katherin Pérez Rojas
dialogue is “considered to be a block of text within a paragraph, falling
between quotation marks”.
According to Portolés et al. (2009), Spanish novel dialogues tend to be
written directly, without any clarifying introduction, and with very few
comments or detailed explanations of the mental state or features of the
speaker. In Spanish novels, dialogues open with an em-dash, raya, [—] at the
beginning of the sentence (but it is not repeated at the sentence closure), and
when indicating the speaker; closing only when there is a clarification in the
middle.
Tables 1 and 2 present the most common standard combinations of
punctuation, narration and character intervention (CI) or dialogue sequences
found in English and Spanish novels.
Table 1: Types of narration and character intervention (CI) in English novels
English
Example
“CI”
“Oh, yes! I know that! I know that, but do
you know what day it is?”
“CI” narration.
“She’s packing them,” explained Mother.
narration "CI".
He said to the driver, “You are early tonight, my
friend.”
“CI” narration “CI”.
“But where?” he asked. “Where are we going
exactly? Why can't we stay here?”
narration “CI”
narration.
She saw, I suppose, the doubt in my face, for she
put the rosary round my neck and said, “For your
mother's sake,” and went out of the room.
Table 2: Types of narration and character intervention (CI) in Spanish novels
Spanish
Example
―CI
―Los Potter, eso es, eso es lo que he oído…
―CI ―narration.
―Tendremos a papá y a mamá y a nosotras mismas
―dijo Beth alegremente desde su rincón.
narration ―CI
Hubo un momento de silencio, hasta que Padre
dijo:
—¿Y bien? ¿Qué opinas?
―CI ―narration―. CI
―No ―respondió en tono cortante―. ¿Por qué?
―CI ―narration―. CI.
―narration― CI.
—Hemos recibido denuncias sobre hombres y
mujeres vagabundos que desaparecieron el mes
pasado —intervino Banks—. Al principio pensamos
que podría ser uno de ellos, pero no es así. —añadió
en tono dramático—. La víctima fue una de esas
personas de anoche.
―CI ―narration―. CI.
― CI2.
—Pero ¿adónde? —preguntó—. ¿Adónde nos
vamos? ¿Por qué no podemos quedarnos aquí? —Es
por el trabajo de tu padre. Ya sabes lo importante
que es, ¿verdad?
Automatically building translation memories for subtitling 55
English and Spanish pre-processing of the novels
English dialogues are represented by single (‘ ’) or double (“ ”) quotation
marks. There are two ways of expressing them: 1) The ones with identical
form (neutral, vertical, straight, typewriter, or “dumb” quotation marks),
typewriter double quotes (" ") and typewriter single quotes (' '). 2) The ones
with left and right hand distinction (typographic, curly) typewriter double
quotes curly (“ ”) and typewriter single quotes curly (‘ ’). For the sake of
standardization, and to enhance recognition of the dialogues, it was decided to
convert all quotes to the typewriter double quotes curly (“ ”) format.
The main feature of the Spanish dialogues is the em-dash (―), but it
needs to be in accordance with another feature (punctuation mark, capital
letter) to distinguish the caracter intervention from the narrative part. For
technical reasons we changed the em-dashes into en-dashes (–).
English and Spanish dialogue extraction scripts
The scripts to extract the dialogues from the English and Spanish texts were
designed to read each line of the novel and recognize the dialogues based on
the previously mentioned marks.
Figure 1 is a flow chart graphically showing the process and the
description of the modules included in the methodology.
Figure 1: Pipeline of the methodology for creating TMX files from novels
The most common error in the extraction phase of the English texts was
due to problems in recognizing the marks when the extraction involved very
long segments. These segments contained several character interventions as
56 Katherin Pérez Rojas
well as narration that required editing in the alignment phase. The most
common error in the Spanish extraction was when there was a character
intervention that did not follow the specified mark rules; therefore it was not
recognized and extracted. However, since most of these errors usually took
place in the second part of the character intervention, the first part of it was
always extracted, providing the first sentence of the dialogue.
Evaluation results and discussion
For the dialogue extraction, the notion of accuracy is understood as the
amount of the original dialogue that the script is able to recognize and extract.
For the TMX file results, accuracy is in terms of the fuzzy matches and the
match percentage.
The results of the method will be presented in three different categories:
(i) Results of the dialogue extraction,
(ii) results of the alignment, and
(iii) first results of the TM performance when translating the scripts.
Results of the dialogue extraction
The total amount of dialogues was retrieved using ReGex to recognise the
proposed patterns and then they were manually checked to compare the
correct recognition of all of them, especially for the Spanish novels were the
dialogues needed to be checked to determine the end of a dialogue and the
beginning of narrative. Table 3 shows the difference rates between English
orignal and extracted amount of CI, Spanish original and extracted amount of
CI and the difference rate between both original sets.
On average, the dialogue sequences differ by 14%, which is not
surprising because in English there are more dialogue markers (quotation
marks) for the same dialogue in Spanish. For English, the difference between
the dialogue in the original novel and the dialogues recognised and extracted
by the script is 5%. For Spanish, the average extraction difference is 18%.
Again, this was expected because the Spanish script recognizes all the
character interventions and merges them into one, outputting just one chain of
dialogue.
Automatically building translation memories for subtitling 57
Table 3: Percentage difference between the total numbers of English and Spanish dialogues
Title
Original
CI Eng
Extracted
CI Eng
Org/Ext
CI Eng
%
Original
CI Spa
Extracted
CI Spa
Ori/Ext
CI Spa
%
% Differ
Org/CI
Eng/Spa
Atonement
1010
998
1,2
932
836
10
8
The bone
collector
4372
4230
3,2
4050
3489
14
7
The boy in the
striped
pyjamas
1451
1394
3,9
1306
935
28
10
One flew over
the cuckoo’s
nest
1713
1679
2,0
1554
1150
26
9
The devil
wears Prada
2312
2254
2,5
2139
1654
23
7
Dracula
1150
1098
4,5
1048
927
12
9
East of Eden
7623
7543
1,0
5898
5361
9
23
The great
Gatsby
1438
1348
6,3
1153
922
20
20
Memoirs of a
geisha
2875
2740
4,7
2415
1896
21
16
The
hitchhiker’s
guide to
galaxy
1862
1828
1,8
1453
1272
12
22
Harry Potter 1
2396
2230
6,9
2205
1700
23
8
Harry Potter 2
2919
2748
5,9
2431
1846
24
17
Harry Potter 3
3826
3708
3,1
3510
2481
29
8
Harry Potter 4
6004
5365
10,6
4940
3773
24
18
Harry Potter 6
6124
5491
10,3
5523
3977
28
10
Little women
1723
1614
6,3
1611
1608
0,1
7
Murder on the
Orient Express
2572
2346
8,8
2377
2075
13
8
Pride and
prejudice
1783
1597
10,4
1298
1190
8
27
Sense and
sensibility
1580
1481
6,3
1089
993
9
31
The hunger
games
1537
1536
0,1
1398
1094
22
9
Average
%
5
18
14
Results of the alignment
The results of the alignment of the extracted dialogues showed that most of
the books followed a pattern according to which the number of sentences, the
disposition of lines, and the continuity of the speech are similar in both
languages. Since there is no gold standard for the evaluation of the alignment
of these dialogues, the alignment was measured taking into account a shuffled
sample corresponding to 10% of each aligned bitext. On average, the
58 Katherin Pérez Rojas
alignment was precise for about 90% of the cases, with an erroneous
alignment rate of 10%.
First results of the TM performance when translating the scripts
This step aims at providing preliminary findings on how many complete
matches or partial matches were retrieved from the TM. These results were
obtained using OmegaT 2.6.3 and the entire corpus of TMs against the subtitle
corpus of all the adaptated movies.
Table 4 shows the total result of repetitions, exact matches, fuzzy
matches and no matches found when using the translation memories as
reference material for the translation of the subtitles.
Table 4: General results and percentages of Match Statistics with OmegaT
Type of segment
Total
Occurrences
%
Repetition
2105
5.9%
Exact Matches
1129
3.2%
Segm 75-99%
1960
5.8%
Segm 50-74%
23553
69.8%
No Match
5334
15.2%
Total Segments
34081
The general results show that of 34,081 analyzed segments, repetitions
account for 5.9% (2,015 segments were repeated). The repetition feature of
the translation memory means that there is one translation per repeated
segment, if the translation memory system finds the exact segment in the
translation memory file; this 5.9% can be automatically translated. This also
accounts for the Exact Matches, which were 3.2% of the corpus (1,129
segments).
In literary narrative, dialogues are one of the main guides for plot
development, but the neighboring description is in charge of directing the
references thereby limiting the dialogues to their function of continuity. This
is reflected in the matching of subtitles with their extracted dialogues,
especially when the character interventions in the subtitles and in the novel
are the same. In these cases, there may be a 100% match, however, that match
will not occur frequently along the text: we see that 9.1% was found to be
repeated, rendering it difficult to rely on automatic translation as a resource to
speed up the translation process.
Note that 5.8% of the corpus (1,960 segments) obtained a match between
75% and 99%. This range is considered highly relevant for translation because
these segments need minor edition and, as a whole, save the translator a great
amount of time (Somers 2003, Bowker 2005).
In total, 69.8% (23,552 segments) of the corpus found a match in the
50%-74% group. Although it can be assumed that matches below 70% are not
Automatically building translation memories for subtitling 59
very useful for translation memory systems, many TM software products state
the minimum match value between 60% and 75%, and recommend starting
with a low percentage such as 50%. With these values in mind, the hits found
in the fifth group could be considered partial matches (O’Brien 1998). As
shown by the alignment and preliminary fuzzy match analysis, the tendency is
that most of the dialogues stored in the TM will serve as contextual
information for the subtitles.
The remaining 15.2% of the corpus (5334 segments) obtained no match
with the subtitle TM file. It can be assumed that these dialogue segments
either belong to films with a high degree of free adaptation (change of plot,
change of characters, etc.) or to novels whose dialogues tend to be especially
longer in the original and have been significantly cut.
Conclusions and future work
The aim of this paper was to design a method to automatically create
translation memory files for subtitling. The method searches for dialogue
marks in the novels and sets the boundaries to extract each character
intervention. These character interventions are then aligned to create a TMX
file that is used as translation memory when translating the subtitles.
It is possible to automatically recognize and extract dialogues from the
English and Spanish novels. This process is language-dependent, requiring a
defined set of pre-processing steps and a specific set of scripts per language.
The alignment results show that the books and their translations followed
similar organization patterns and speech continuity.
The preliminary translation memory test showed that a high percentage
of partial matches were in the 50% to 74% group. Probably these low
percentages were due to the transformations from written novel to
cinematographic script, but an initial analysis showed that they contain
enough contextual information to help in the translation process.
In regard to future work, it is necessary to improve the features of the
dialogue extractor, taking into account the special needs of Spanish narrative,
by marking the end of dialogues to allow full automatic recognition. It is also
necessary to test the resulting TM files using different translation memory
systems. Further ways of automatically shortening the longer dialogues might
be useful to obtain higher matching scores in the TM.
References
Bakhtin, Mikhail. 1986. Speech Genres and Other Late Essays. Translated by
Vern W. McGee. Austin: University of Texas Press.
60 Katherin Pérez Rojas
Bowker, Lynne. 2005. “Productivity vs Quality: A pilot study on the impact
of translation memory systems”. Localisation Focus 4(1): 13-20.
Carroll, Mary, and Jan Ivarsson. 1998. Code of Good Subtitling Practice.
Approved at the meeting of the European Association for Studies in
Screen Translation in Berlin.
Díaz Cintas, Jorge, and Aline Remael. 2007. Audiovisual translation:
subtitling. Manchester: St. Jerome.
Elson, David, and Kathleen McKeown. 2010 “Automatic Attribution of
Quoted Speech in Literary Narrative”. Proceedings of the Twenty-Fourth
AAAI Conference on Artificial Intelligence (AAAI 2010), Atlanta.1013-
1019.
Harrington, John. 1977. Film and/as literature. Englewood Cliffs NJ:
Prentice-Hall
Maranhão, Tullio. 1990. The Interpretation of Dialogue. Chicago: University
of Chicago Press.
Mejías Moreno, Raquel. 2010. Herramienta para la traducción asistida en la
industria audiovisual. Proyecto de Fin de Carrera. Escuela Politécnica
Superior. Universidad Carlos III de Madrid.
O’Brien, Sharon. 1998. Practical experience of computer-aided translation
tools in the localization industry. Lynne Bowker, Michael Cronin,
Dorothy Kenny and Jennifer Pearson (eds) Unity in Diversity: Current
Trends in Translation Studies. Manchester: St. Jerome. 115-122.
Online Film School. Script adaptation.
http://filmschoolonline.com/sample_lessons/sample_lesson_adaptation.ht
m Visited September 2014.
Portolés, Elisa. 2009. “Cómo escribir diálogos”. Online didactic resource
http://www.alquimistasdelapalabra.com/dialogos/teoria_dialogo/index.ht
ml. Visited September 2014.
Rauma, Sara I. 2004. Cinematic dialogue, literary dialogue and the art of
adaptation. Pro Gradu Thesis. University of Jyväskyla.
https://jyx.jyu.fi/dspace/bitstream/handle/123456789/7349/G0000703.pd
f?sequence=1Visited. September 2014.
Rimmon-Kenan, Shlomith. 1989. Narrative fiction: Contemporary poetics.
London: Methuen.
Sherlock, Karl. 2011. “Some practical advice about using punctuation and
markers in creative writing”. English 126 Creative Writing, Punctuation
and Prose
http://www.grossmont.edu/karl.sherlock/English126/punctuation.html.
Visited September 2014.
Somers, Harold. 2003. “Translation memory systems”. In Harold Somers (ed.)
Computers and Translation: A Translator’s Guide. Amsterdam and
Philadelphia: Benjamins. 31-47
SUMAT An Online Service for Subtitling by Machine Translation. 2011
Annual Public Report 2011.
Automatically building translation memories for subtitling 61
http://ec.europa.eu/information_society/apps/projects/factsheet/index.cfm
?project_ref=270919. Visited September 2014.
Sumiyoshi, Hideki, Hideki Tanaka, Nobuko Hatada and Terumasa Ehara.
1995 “Translation workbench for generating subtitles for English TV
news”. Tokyo: NHK Science and Technical Research Laboratories.
Writing Studio. 2001. The art of adaptation.
http://www.writingstudio.co.za/page62.html.Visited September 2014.
