Syntactic Simplification for Machine Translation

Jos CAMACHO COLLADOS

Centre Tesnire, Universit de Franche-Comt, France
Research Group in Computational Linguistics, University of
Wolverhampton, United Kingdom
Laboratori fLexSem, Universitat Autnoma de Barcelona, Espaa
Abstract
This paper presents a new rule-based Syntactic Simplification system for the
Spanish language. The system basically consists of splitting complex
sentences into several simple ones preserving its original meaning. It is
envisioned to be a preprocessing tool for other Natural Language Processing
applications such as Text Summarization, Information Extraction, Parsing
and Machine Translation. According to the evaluation, the application of the
syntactic rules proposed clearly improves Machine Translation from
Spanish to far target languages such as Korean and Chinese. A parallel
corpus of Spanish original and simplified texts has been manually built and
validated by native Spanish speakers for this research. In addition, an
experiment was carried out among two groups of Spanish learners
comparing their reading comprehension of original and simplified texts.
Their answers show a null correlation between simplification and text
comprehension. The automatic Syntactic Simplification system has also
been intrinsically evaluated obtaining promising results.
Key-words
Text Simplification; Syntactic Simplification; Natural Language Processing;
Spanish; Machine Translation.
1. Introduction
Text Simplification (TS) is a Natural Language Processing (NLP) task
consisting of reducing the complexity of a sentence in order to make it
simpler, but without changing the general meaning of it (Siddharthan, 2002:
1). This research focuses on the syntactic level. The syntactic simplifications

Jos CAMACHO COLLADOS

Syntactic Simplification for Machine Translation

considered are basically based on splitting a complex sentence into several

simpler ones without damaging its original meaning. The motivations for
Text Simplification are basically divided in two main groups. The first one
is the one which uses TS in order to make a text accessible for certain
people with problems in reading complex structures. This research includes
an evaluation of the system regarding this application but, more specifically,
for foreigners learning Spanish. The second main group includes the systems
aiming to improve other NLP applications, which is the main purpose of this
research. Some of the applications which may be improved from this
approach are explained below:
Text Summarization: The idea for the TS in Text Summarization is to
reduce the information extracted by each sentence, keeping just the relevant
one. Since Text Summarization is sometimes based on the information
extracted from different sentences, having the information better divided in
sentences might be extremely useful for the task (Chandrasekar et al., 1996;
Siddharthan, 2002).
Information Extraction: As far as Information Extraction is concerned,
automatic systems may work better when the complexity of a sentence is
low (Evans 2011, Klebanov et al., 2004). TS may help split different types
of information inside a complex sentence into several simple sentences.
Parsing: The goal of TS in parsing is due to the improvement of
performance if given a short sentence as input (Siddharthan, 2002). Longer
sentences give more cases of ambiguity, so that TS aims to reduce the length
of the sentences while preserving the meaning as much as possible.
Machine Translation: One of the evaluations of the Syntactic
Simplification (SS) system carried out on this research is extrinsic and it is
done regarding the improvement of some MT systems taking Spanish as
source language. The target languages taken into account in this research are
Chinese and Korean. The far languages (such as Spanish and either Chinese
or Korean) and long sentences are nowadays two of the main drawbacks of
MT, so it may be quite useful to have a monolingual tool simplifying
complex sentences regardless of the target language.

BULAG N XX, 201X

Natural Language Processing and Human Language Technology 201X

This paper is organized as follows: Section 2 shows previous researches

related to Text Simplification and Machine Translation; Section 3 explains
SS splitting rules considered along this research; Section 4 introduces the
Spanish parallel corpus of original and syntactically simplified sentences
and its validation; Section 5 shows the development and architecture of the
automatic SS system by applying the rules presented in Section 3; Section
includes the both extrinsic and intrinsic evaluations of the system; and
Section 7 gives the conclusions of the work and address some possible
directions for further research on the subject.
2.Related work
2.1 First steps in Text Simplification
Chandrasekar et al. (1996) was the first serious attempt in order to create a
general architecture for a TS system. In the research, they comment the
usefulness of the TS for a wide variety of applications. The main
applications have been explained in Subchapter 1.1. Apart from helping
NLP applications such as MT, parsing and Text Summarization,
Chandrasekar et al. (1996) introduce the possibility that TS could help
Information Retrieval. Not only are NLP applications addressed as possible
beneficiaries of the TS, but also TS could make the text clear and easier to
understand. TS may be used as a preliminary step in order to create this
artificial constrained language like controlled languages, which will be
handled in Sub-section 2.4. Nevertheless, the research was focused on
improving the performance of a parser, which consequently would be
beneficial for the other applications. Regarding this task of improving a
parser, the main problem was to reduce the complex syntactic structures that
we can find in long sentences. This was achieved by opportunistically
splitting the complex sentences into two or more simple ones.
Siddharthan (2002) proposes a new architecture for a TS system. It follows
the work of Chandrasekar et al. (1996) but introducing quite remarkable
improvements. The first remarkable difference from the previous work is the
introduction of a new stage on the TS process: regeneration. This is the main
improvement of this research and where Siddharthan tries to focus on. The
discourse level problems raised by Chandrasekar et al. (1996) are partially
solved by this new stage in the system. The new architecture is then based in
three stages: analysis, transformation and regeneration. As far as SS is

Jos CAMACHO COLLADOS

Syntactic Simplification for Machine Translation

concerned, Siddharthan (2002) considered several syntactic phenomena

such us relative clauses, adverbial clauses, coordinated clauses,
subordinated clauses, correlated clauses, participial phrases, appositive
phrases and passive voice. Afterwards, he analyzes and evaluates each stage
separately, giving more emphasis to the new regeneration stage.
2.2 Corpus-based Text Simplification
Several rule-based systems were created after Siddharthans work. Some of
them already gave some external evaluation depending on the purpose of the
system (Petersen and Ostendorf, 2007; Evans, 2011). There are also works
addressing the need of a parallel corpus of original and simplified sentences
(Petersen and Ostendorf, 2007; Alusio et al., 2008; Specia et al., 2009;
Specia, 2010). The corpus, apart from potentially being used to develop a
Machine Learning algorithm, would be useful to carry out a deeper analysis
of the task leading to some new ideas or improvements of rule-based TS
systems.
Specia et al. (2009) followed this direction and worked on building a
Brazilian Portuguese parallel corpus of original and simplified sentences
with both lexical and syntactic simplifications. The main goal of the corpus
is to help people with low level of literacy or some other cognitive
disabilities. Specia (2010) experimented afterwards with a quite simple
corpus-based TS approach for the Brazilian Portuguese. The goal of this
system is the same as in Specia et al. (2009), for people with some problems
at reading. She basically used a Statistical Machine Translation (SMT)
method to deal with the TS problem and check the results. The SMT was
carried out without making many changes, so that the approach could be
easily improved by adapting the framework to the particular TS problem.
2.3 Spanish Text Simplification
Bott and Saggion (2011a) studies the problem of TS for the Spanish
language. The goal of Bott and Saggion (2011a) is to create a text easy to
read for people with learning disabilities, so it differs in this point from the
main goal of this research. It is a preliminary study where they analyze a
corpus of news and the respective simplified one. They address the need of
getting a parallel corpus in order to be able to use it for the creation of a
reliable TS system.

BULAG N XX, 201X

Natural Language Processing and Human Language Technology 201X

Bott and Saggion (2011b) follows their last paper and explain in detail an
algorithm to align a parallel corpus of news and their simplified ones at a
sentence level. As there is not training data for task (there is no manually
aligned parallel corpus), they rely on unsupervised learning.
Drndarevi et al. (2013) presents a two-component (syntactic and lexical)
automatic text simplification system for the Spanish language in order to
make the text easier to read for people with cognitive disabilities. The
system managed to get simpler sentences without seriously damaging their
grammaticality and preservation of meaning with the original sentences.
However, they do not propose any solution to the parsers errors, which
caused most of the errors found in the system.
2.4 Controlled Languages and Machine Translation
Cardey et al. (2004) addressed the problem of MT from French to two far
languages such as Chinese or Arabic. One of the main conclusions is that
many issues concerning a pair of languages should be studied separately for
each different language pair, which makes these MT systems really
language-independent. This is a problem as many pairs of minority
languages will not receive the needed attentions to solve all the specific
problems, especially on long sentences. This may be also the case of SMT,
since SMT usually backfires when translating long and complex sentences
(Koehn, 2010). SMT tends to backfire when the source sentence is long and
have several verb phrases, no matter the target language. The problems of
MT often come, apart from the length of the sentence, from the ambiguity
that such sentence could have. MT systems should deal with this problem,
which is one of the main ones for every system. Even deeply treated,
ambiguity is really complicate to be totally solved in order to create a high
quality MT system. That is one of the main reasons why controlled
languages were developed (Kaji, 1999; Mitamura, 1999; Cardey, 2011).
Controlled languages are artificially created subset of the natural language
where the ambiguity is eliminated and complexity is reduced. This clearly
improves MT and it is especially useful where used for different target
languages.
Although controlled languages belong to natural language, they are
artificially created. They need to be created and people using it should know

Jos CAMACHO COLLADOS

Syntactic Simplification for Machine Translation

all the restrictions in order to be able to use it. As explained in Mitamura

(1999), this is not an easy task because usually there is a big restriction in
both grammar and vocabulary, so it should be deeply studied before using it.
As controlled languages have been proved to be a successful approach for
MT on many cases, TS may be used as a tool to reach the desirable
controlled language. Temnikova (2012) shows how the post edition by a
human translator becomes an easier task by applying TS prior to the MT.
3. Syntactic Simplification splitting rules
The simplification considered in the corpus is basically reduced in splitting
long and complex sentences in simple ones. Since this research is aimed to
be used to build an automatic simplification system, all the rules have been
carefully selected in order to achieve this goal in the future (Camacho
Collados, 2013). To begin with, the first condition to split a sentence is the
number of conjugated verbs. The sentences which are to be simplified must
have at least two conjugated verbs. It is important to notice that in most
sentences there are different types of structures to be simplified
simultaneously, which makes the task harder. Sub-section 3.1 handles the
coordinate sentences and Sub-section 3.2 summarizes some simple
subordinate cases.
3.1 Coordination
The original sentence is split on the position of the coordination nexus or
articulation point as called by Chandrasekar et al. (1996). It is usually
suggested repeating some noun phrases in order to improve the
understandability and a best processing by other NLP applications. Example
(3) is an illustrative example of how the simplification is handled on
coordinate sentences. Several examples on this section are given in English
for a better understanding of the reader.
(3) a.
b.

Fishes swim in the sea and butterflies fly in the sky.

Fishes swim in the sea.
Butterflies fly in the sky.

3.2 Subordination

BULAG N XX, 201X

Natural Language Processing and Human Language Technology 201X

Non-restrictive relative clauses: This kind of structure is split without

adding any other element which might cause a mistreatment by other NLP
applications. The relative clause begins with a comma followed by a
connector such as que [which] or quien [who]. Example (4) is a simple and
representative example for the Spanish language of this type of
simplification.
(4) a.
Juan, que es an muy joven, consigui el premio. [Juan, who is
still very young, got the prize.]
b.

Juan es an muy joven.

Juan consigui el premio.

[Juan is still very young.]

[Juan got the prize.]

Effect: This type of structure contains the cause-effect relation. They are
connected by a conjunction which indicates the end of the cause and the
beginning of the effect. Therefore, the splitting of the sentence will be done
at the conjunctions position. Another effect connector such as therefore in
English is introduced at the beginning of the second sentence, as we can
appreciate in Example (5). In Spanish the connector introduced will be Por
lo tanto,.
(5) a.
b.

The cat ate poisoned food, so it died.

The cat ate poisoned food.
Therefore, the cat died.

Causal: The same kind of relation cause-effect appears in this structure.

However, the cause is placed after the effect in this case. They are connected
by a different causal connector such as because on the English language
(porque on the Spanish language) - Example (6). The output sentences
include the effect in the first position and the cause or reason in the second
sentence.
(6) a.
b.

Dogs cant fly because they dont have wings.

Dogs cant fly.
Reason: They dont have wings.

Indirect speech to direct speech: This structure is currently simplified

just on very specific cases. More precisely, in the cases where a

Jos CAMACHO COLLADOS

Syntactic Simplification for Machine Translation

communication verb such as decir, comunicar or explicar are followed by

the relative que and a sentence concerning the communication of the
speaker. Most of the communication verbs included for this task were
extracted from Diccionario combinatorio del espaol contemporneo
(Bosque, 2004). The original sentence is split into two sentences, the second
one introduced by :, as we can observe in Example (7):
(7) a.
El jugador dijo que el presidente estuvo con el equipo antes del
partido. [The player said that the president was with the team before the
match.]
b.

El jugador dijo:
[The player said:]
El presidente estuvo con el equipo antes del partido. [The
president was with the team before the match.]

4.Corpus
4.1 Creation of the parallel corpus
The corpus chosen as a reference was the AnCora Corpus (Taul et al.,
2008), which consists of Spanish newspaper texts annotated at syntactic and
morphological level. Newspaper texts are quite representative of the natural
language and complex enough for the simplification task. Once the corpus
was transferred from the original XML format to text format, the sentences
from the AnCora Corpus were manually simplified as explained in Section 3
in separate text files. On this way a parallel corpus of original and simplified
texts was built. The original sentences and the simplified ones are easily
aligned, as each original sentence is separated from each other on a new
line, which is respected on the simplified part no matter how many new
sentences have been created. The corpus currently counts with 3000 original
sentences and their respective simplified ones (Camacho Collados, 2013). It
is already available for research purposes if required.
4.2 Corpus validation
The validation of the corpus has been done regarding two issues
(grammaticality and preservation of meaning), similar to the one used in
Drndarevi et al. (2013) for the evaluation of their SS system. To carry out
this, some preliminary results were obtained from six native Spanish
speakers (three of them linguists and three of them holding a non-related

BULAG N XX, 201X

Natural Language Processing and Human Language Technology 201X

university degree). They required to fill three excel sheets. The first one was
to evaluate the grammaticality of the original sentences from AnCora
corpus; the second one about the grammaticality of the simplified sentences
from AnCora Corpus; and the third one concerning the preservation of
meaning of original and simplified sentences. To do so, thirty complex
sentences were randomly selected from thirty different texts of the corpus.
All the evaluations were done on a 1-5 scale. For the grammaticality
measure, 1 means that sentence is completely a grammatical and the 5 that
the sentence is completely grammatical. For the preservation of meaning, 1
means that there is no preservation of meaning at all and 5 that the meaning
of the simplified sentence and the original is identical. Table 1 shows the
total average for each evaluation and from the results we can appreciate how
grammaticality is not damaged on the simplified sentences (4.74 without
simplification 4.66 with simplification). A paired two-sample t-test at the
0.05 level suggests that the difference is not statistically significant
(t(179)=1.513; p-value=0.132). There are even a few cases where the
simplification actually improves the grammaticality of the original
sentences.
Table 1: Corpus validation (grammaticality and preservation of meaning)
Grammaticality
Original
Sentences

Grammaticality
Simplified
Sentences

Preservation
of meaning

Average

4.74

4.66

4.8

Positive (4-5)

95.6%

97.2%

98.9%

Neutral (3)

3.3%

2.8%

1.1%

Negative (1-2)

1.1%

0%

0%

Regarding the inter-rater reliability for the preservation of meaning task, a

statistical test at a 0.05 level was carried out among the six annotators. The
results (F(5,174)=1.577, p-value= 0.169) show an inter-rater agreement
really high.

Jos CAMACHO COLLADOS

Syntactic Simplification for Machine Translation

5.Automatic Syntactic Simplification system

5.1 Analysis stage
This stage takes the original sentence as input and decides whether
simplifying it or not. It mainly relies on Part of Speech (PoS) tagging, word
searches and, in a specific case, lemmatization. PoS taggers are only
necessary in order to find the number of verbs in the sentence. As explained
before, this research will include the simplification of sentences containing
two conjugated verbs, since sentences with only one conjugated verb will be
considered as simples already. This stage was based on a decision-tree
model taking into consideration indicators (Medero and Ostendorf, 2011) of
possible simplification but also indicators suggesting not to simplify
(cuando, donde).
5.2 Transformation stage
Once the analysis stage has decided to simplify the input sentence, the
transformation stage comes into place. At this stage, the system basically
chooses a position to split the sentence and introduces new connectors if
necessary. This differs a bit from the transformation stage of Siddharthans
text simplification architecture (Siddharthan, 2002), where the analysis stage
were the one selecting the boundaries between clauses and the main
sentence. This stage does not take into account the relation within the
sentences nor the anaphoric references to be considered in the output
sentences. This will be considered later on the regeneration stage.
5.3 Regeneration stage
This is the last and most challenging task of the SS process. After the
splitting of the sentence on the transformation stage, many assumptions need
to be taken in order to make these new simplified sentences readable and
understandable. For instance, there are some anaphoric references which
may be lost during the transformation process and they need to be fixed on
this stage. For the anaphoric references and noun phrase repetition within
the sentences, this stage relies mainly in noun chunking rather than
dependency structures (Siddharthan, 2011). The sentence reordering is
based on a simple algorithm regarding the original position of the sentences
to be split. This algorithm is improved for the complicate case of nonrestrictive relative clause containing embedded structures.

10

BULAG N XX, 201X

Natural Language Processing and Human Language Technology 201X

6. Evaluation
6.1 Syntactic Simplification for Machine Translation
In this Sub-section, the effect of SS in MT will be evaluated by a
comparison between direct MT systems and MT systems which previously
have used SS on the source languages. There will be two different
evaluations concerning different target languages, the first one taking
Korean as the target language and the second one taking Chinese as the
target language. Spanish will be in both cases the source languages. The
experiments settings are similar for both experiments. Thirty complex
sentences have been randomly selected from thirty different texts of the
parallel corpus created along this research (see Section 4), including their
respective simplified ones. The thirty original sentences were translated
respectively by a Spanish-Korean and a Spanish-Chinese translator, which
was considered as Gold Standard. The Spanish-Korean and Spanish-Chinese
Google Translate was the MT system for the evaluation. First, the original
sentences were introduced directly on the automatic translator and the
output sentences were stored, what we will call simply MT. Second, SS was
used prior to the translation. This second output will be called SS+MT. The
simplified part of the thirty original sentences (taken from the parallel
corpus) was introduced on Google Translate to produce three new outputs.
An evaluation of both ways was then carried out among three native Korean
and three Chinese. The native speakers were asked to evaluate the
grammaticality of the output sentence and compare each output with the
Gold Standard regarding its preservation of meaning. The grammaticality
was evaluated by following a 1-5 scale where 1 indicates a totally non
grammatical sentence and 5 indicates a fully grammatical sentence. The
preservation of meaning was also done by following the 1-5 scale in a
similar way. 1 means that the output sentence does not preserve at all the
meaning of the Gold Standard sentence and 5 indicates a total preservation
of meaning.
There are some preliminary results obtained from three native Korean
evaluators holding at least a university postgraduate degree. Spanish-Korean
MT systems have not been as developed as other pair of languages.
According to the native Korean evaluators of this research, the output

11

Jos CAMACHO COLLADOS

Syntactic Simplification for Machine Translation

sentences given by the Spanish-Korean Google Translate MT system are

poorly constructed and in many cases not understandable. This is reflected
on the results which are summarized on Table 2.
Table 2: Spanish-Korean MT results
SpanishKorean

Grammaticality

Meaning
preservation

MT

SS+MT

MT

SS+MT

Average

2.6

3.07**

2.44

3.02**

Positive (4-5)

15.6%

31.1%

18.9%

28.9%

Neutral (3)

40%

44.4%

24.4%

35.6%

Negative (1-2)

44.4%

24.4%

56.7%

35.6%

**: difference statistically significant at the level 0,001

According to a paired t-test, the difference between the grammaticality score
averages of MT and SS+MT output sentences is statistically significant
(t(89)=-5.205, p-value<0.001). This is a considerable difference taking into
account that the Gold Standard sentences have been translated directly from
the original sentences and not the simplified ones. The results from the
preservation of meaning task are also summarized on Table 3. In this case,
the meaning is better preserved by SS+MT. The difference is also
statistically significant according to a t-test (t(89)=-6.183, p-value<0.001).
As far as the Spanish-Chinese MT experiment is concerned, some
preliminary results were obtained from three graduated native Chinese
speakers. The results obtained from the evaluation about the grammaticality
task are summarized on Table 3. According to the results, the
grammaticality is improved when applying SS prior to MT. The difference
is statistically significant at the 0.05 level according to a paired t-test
(t(89)=-2.074 , p-value=0,041). The meaning preservation results were even
more promising than the one regarding the grammaticality task. The
meaning preservation average obtained by applying direct MT got the

12

BULAG N XX, 201X

Natural Language Processing and Human Language Technology 201X

extremely low score 1.82, with 81.1% of the sentences obtaining a negative
attitude from the evaluator. These results, even being still low for a MT
system, show an important improvement by applying SS prior to the
translation. The average reaches a 2.18 score and the percentage of
sentences with a neutral or positive attitude by the evaluators is considerably
raised, whereas the percentage of sentences with a negative attitude from the
evaluator decreases until 65.6%. The average difference between MT and
SS+MT for the meaning preservation task is even statistically more
significant that in the grammaticality task according to an paired t-test
(t(89)=-3,734, p-value<0.001).
Table 3: Spanish-Chinese MT results
SpanishChinese

Grammaticality

Meaning
preservation

MT

SS+MT

MT

SS+MT

Average

2.41

2.62*

1.82

2.18**

Positive (4-5)

12.2%

12.2%

4.4%

8.9%

Neutral (3)

31.1%

44.4%

14.4%

25.6%

Negative (1-2)

56.7%

43.3%

81.1%

65.6%

*: difference statistically significant at the level 0,05

**: difference statistically significant at the level 0,001
6.2 Foreigners learning Spanish
Twenty four Spanish learners at the Faculty of Traduccin e Interpretacin
in Granada University agreed to collaborate in this project. The participants
had a native language other than Spanish. They were divided in two groups
of twelve participants each: A and B. Group A participants were given two
original texts from AnCora Corpus and were asked to answer eight multiplechoice questions about them. It was a reading-comprehension test. Group B

13

Jos CAMACHO COLLADOS

Syntactic Simplification for Machine Translation

students were asked to do the same but from the group As simplified texts.
These simplified texts belong to the parallel corpus created along this
research. The participants of both groups had a similar level of English: five
participants with an advanced level (C1-C2 according to the Common
European Framework of Reference for Languages) of Spanish and seven
with an Intermediate level (European Framework of Reference for
Languages).
The results obtained from the test were processed and summarized on Table
4. Surprisingly, the participants from group A (original texts) got a higher
amount of right answers than the participants from group B (simplified
texts). Group A obtained 82.3% of correct answers in contrast to the 79.2%
obtained by group B. However, after applying an un-paired t-test at 0.05
level (t(22)=-0.457, p-value=0.652), we concluded that this difference is not
statistically significant. Therefore these results suggest that the text
comprehension has not been affected by SS.
Table 4: Results from the reading-comprehension test
Correct answers

A: Original text

B: Simplified text

Total

82,3%

79,2%

Intermediate

82,1%

74,5%

Advanced

82,5%

87,5%

6.3 Automatic Syntactic Simplification system

The test set selection (40 sentences containing two conjugated verbs) was
done by following a random process. Every sentence was randomly selected
from an AnCora Corpus text. There is no more than one sentence selected
from a specific text. Therefore 40 different texts were necessary for the
extraction of all the sentences to be used as a test set. Since these 40 texts
belong to the original part of the parallel corpus, the simplified part of the
corpus was taken as a reference (Gold Standard) in order to evaluate the

14

BULAG N XX, 201X

Natural Language Processing and Human Language Technology 201X

system. The evaluation was then divided in three parts, one for each
different stage.
The analysis stage takes an individual sentence from the test set as input and
decides whether simplifying the sentences or not. The output of a single
sentence takes only two values: True for a sentence to be simplified and
False for a sentence which does not need simplification. As we can
appreciate from Table 5, there were 12 sentences which were not simplified
by the system and 28 correctly identified as simplifiable. In terms of
precision, the analysis stage obtains a remarkable 100%. This means that
there is no a single sentence simplified when it should not be simplified.
Table 5: Analysis stage results
Simplification

No Simplification

Gold Standard (n. of sentences)

30

10

Correct

28

10

Wrong

Precision:

100%

Recall:

93.3%

The splitting of the original sentence is carried out on the transformation

stage. There are 28 sentences from the test set to be split at this stage. The
system correctly selects the positions of the clauses and the simplification
cases in 100% of the sentences from and makes just a single mistake in the
transformation. Therefore, the accuracy of the transformation stage
reaches 96.43%, which is quite promising at this point.
For the regeneration stage, the same test set is used. 27 sentences are already
correctly split on the transformation phase and need to be fixed at this stage.
A general evaluation was done for the regeneration stage. Each output from
the 27 already split sentences was given a 0-2 score. 2 means that the output
is exactly same as the Gold Standard. 1 means that the output has the same

15

Jos CAMACHO COLLADOS

Syntactic Simplification for Machine Translation

meaning as the Gold Standard but something has been modified on a

different way, just affecting the readability. 0 means that the meaning of the
original sentence is on some way changed by the regeneration. The results
obtained by the regeneration module are shown on Table 6.
Table 6: Regeneration stage results
Score

Number of sentences

18

Percentage

3.7%

29.6%

66.7%

Average:

1.63

As we can appreciate on Table 6, 18 out of the 27 sentences (66.7%) were

correctly regenerated according to the Gold Standard, which is a remarkably
high percentage taking into account the complexity of this module. The
score average of the system is 1.63. This result is quite promising, as a few
improvements may lead to an almost perfect score (2) by the system. To
sum up, the general results obtained by the system are shown on Table 7.
The general results obtained are promising, as 92.9% of the simplified
sentences were positively handled by the system.
Table 7: Simplification system general results
Total

Percentage

Input sentences

40

100%

Sentences simplified

28

70%

Positively simplified

26

92.9%

Incorrectly simplified

7.1%

Perfectly simplified

18

64.3%

16

BULAG N XX, 201X

Natural Language Processing and Human Language Technology 201X

12

Sentences not simplified

30%

7. Conclusions and future work

The results from the evaluation state that a reliable syntactic simplification
system is feasible by considering splitting rules proposed on this paper. A
few adjustments need to be made especially on the regeneration stage,
focusing on particular features from the Spanish language. Further research
should focused on the development of the Syntactic Simplification system
and extend it to handle every kind of sentence from the natural language.
Different rules could be also added in the system and evaluate its impact on
extrinsic applications other than Machine Translation (Text Summarization,
Information Extraction and parsing, for instance). Modules at the lexical,
discourse and semantic level could be implemented in the system in order to
solve some ambiguity problems on these extrinsic applications. A possible
way to improve the system will be to extend the parallel corpus in order to
apply Machine Learning techniques. The future system could be hybrid by
using the syntactic rules and Machine Learning in some complicated cases.
As far as Machine Translation is concerned, Syntactic Simplification has
been proved to be a quite reliable and easy to implement monolingual
resource for underdeveloped systems. Spanish-Chinese and Spanish-Korean
Statistical Machine Translation systems have been evaluated on this
research. The preliminary results state that the application of Syntactic
Simplification prior to the translation has been proved to be really beneficial
on both cases. However, other languages should be taken into consideration
as target languages in the future. Further research should also focus on
finding accurate evaluation metrics, either automatic or human, for the task.
Acknowledgements
This research was supported by the European Commission, Education &
Training, Erasmus Mundus: EMMC 2008-0083, Erasmus Mundus Masters
in NLP & HLT programme.
I would like to acknowledge my supervisors: Prof. Sandrine Fuentes from
Universitat Autnoma de Barcelona; Prof. Richard Evans from the
University of Wolverhampton; and Prof. Sylviane Cardey-Greenfield from

17

Jos CAMACHO COLLADOS

Syntactic Simplification for Machine Translation

Universit Franche-Comt. I would also like to thank to all the people who
helped me in some way during my stay in the three universities.
Thank you to Obra Social La Caixa, which partially supported this
masters by a La Caixa grant for master studies in Spain.
References
Alusio, S., Specia, L., Pardo, T., Maziero, E., and Fortes, R. (2008)
Towards Brazilian Portuguese Automatic Text Simplification Systems.
ACM Symposium on Document Engineering, pp. 240248.
Bosque, I. (dir.) (2004) Redes. Diccionario combinatorio del espaol
contemporneo, Ediciones SM, Madrid.
Bott, S. and Saggion, H. (2011) Spanish Text Simplification: An
Exploratory Study. Revista de la Sociedad Espaola para el Procesamiento
del Lenguaje Natural.
Bott, S. and Saggion, H. (2011) An Unsupervised Alignment Algorithm for
Text Simplification Corpus Construction. Proceedings of the ACL
Workshop on Monolingual Text-to-Text Generation, pp. 20-26.
Camacho Collados, J. (2013) Splitting complex sentences for Natural
Language Processing applications: Building a Simplified Spanish, V
International Conference on Corpus Linguistics, Alicante, Spain, in press.
Cardey, S., Greenfield, P., Alsharaf, H., and Shen, Y. (2004) Problems and
Solutions in Machine Translation Involving Arabic, Chinese and French.
Proceedings of the International Conference on Information Technology:
Coding and Computing (ITCC04).
Cardey, S. (2011) Machine Translation of Controlled Languages for More
Reliable Human Communication in Safety Critical Applications.
Proceedings of the 12th International Symposium on Social Communication
- Comunicacin Social en el Siglo XXI, Santiago de Cuba, Cuba, Vol. II, pp.
953-958.

18

BULAG N XX, 201X

Natural Language Processing and Human Language Technology 201X

Chandrasekar, R., Doran, C., and Srinivas, B. (1996) Motivations and

methods for text simplification. Proceedings of the 16th International
Conference on Computational Linguistics (COLING 96), pp. 10411044.
Evans, R. (2011). Comparing methods for the syntactic simplification of
sentences in information extraction. Literary and Linguistic Computing,
26(4), pp.371-388.
Drndarevi, B., tajner, S., Bott, S., Bautista, S., and Saggion, H. (2013),
Automatic Text Simplification in Spanish: A Comparative Evaluation of
Complementing Modules. Computational Linguistics and Intelligent Text
Processing, 7817, pp. 488-500.
Goldberg, M. (1999). An Unsupervised Model for Statistically Determining
Coordinate Phrase Attachment. Proceedings of the 37th annual meeting of
the Association for Computational Linguistics. College Park, Maryland.
Kaji, H. (1999) Controlled languages for machine translation: state of the
art. Proceedings of MT Summit VII MT in the Great Translation Era.
Kent Ridge Digital Labs, Singapore, pp. 37-39.
Klebanov, B. B., Knight, K., and Marcu, D. (2004). Text Simplification for
Information-Seeking Applications. On the Move to Meaningful Internet
Systems. Berlin: Springer-Verlag, pp. 73547.
Koehn, P. (2010) Statistical Machine Translation. Cambridge University
Press. Textbook.
Medero, J. and Ostendorf, M. (2011). Identifying Targets for Syntactic
Simplification. ISCA International Workshop on Speech and Language
Technology in Education (SLaTE 2011).
Mitamura, T. (1999) Controlled Language for Multilingual Machine
Translation. Proceedings of Machine Translation Summit VII.
Petersen, S., and Ostendorf, M. (2007) Text simplification for language
learners: a corpus analysis. Proceedings of Workshop on Speech and
Language Technology for Education.

19

Jos CAMACHO COLLADOS

Syntactic Simplification for Machine Translation

Siddharthan, A. (2002) An Architecture for a Text Simplification System.

Proceedings of the Language Engineering Conference (LEC'02):
Hyderabad, India, December 13-December 15, IEEE Computer Society,
London, United Kingdom, pp. 64.
Siddharthan, A. (2011). Text Simplification using Typed Dependencies: A
Comparision of the Robustness of Different Generation Strategies.
Proceedings of the 13th European Workshop on Natural Language
Generation. Nancy, France.
Specia, L., Caseli, H.M., Pereira, T.F., Pardo, T.A.S., Gasperin, C., and
Alusio, S.M. (2009). Building a Brazilian Portuguese parallel corpus of
original and simplified texts. In Alexander Gelbukh (ed), Advances in
Computational Linguistics, Research in Computer Science, 41, pp. 59-70.
10th Conference on Intelligent Text Processing and Computational
Linguistics (CICLing-2009). Mexico City.
Specia, L. (2010). Translating from Complex to Simplified Sentences. 9th
International Conference on Computational Processing of the Portuguese
Language (Propor-2010). Lecture Notes in Artificial Intelligence, 6001,
Springer, pp. 30-39. Porto Alegre, Brazil.
Taul, M., Mart, M., and Recasens, M. (2008) AnCora: Multilevel
Annotated Corpora for Catalan and Spanish. Language Resources and
Evaluation-LREC 2008.
Temnikova, I. (2012) Text Complexity and Text Simplification in the Crisis
Management domain. Ph. D. Thesis. University of Wolverhampton, United
Kingdom.

20