Automatic Summarization Based on Sentence
Morpho-Syntactic Structure: Narrative Sentences
Mehdi Yousfi-Monod and Violaine Prince
LIRMM, UMR 5506, 161 rue Ada, 34392 Montpellier Cedex 5, France
Abstract. We propose an automated text summarization through sentence com-
pression. Our approach uses constituent syntactic function and position in the
sentence syntactic tree. We first define the idea of a constituent as well as its role
as an information provider, before analyzing contents and discourse consistency
losses caused by deleting such a constituent. We explain why our method works
best with narrative texts. With a rule-based system using SYGFRAN’s morpho-
syntactic analysis for French [1], we select removable constituents. Our results
are satisfactory at the sentence level but less effective at the whole text level, a
situation we explain by describing the difference of impact between constituents
and relations.
1 Introduction
The amount of information available on the Web or in some compagnies, administra-
tions and laboratories is always increasing, thus hardening information retrieval on such
resources. Automatic summarization, aiming at considerably reducing the size of such
data, appears to be a good solution to ease this search. It does so by introducing a
smaller but relevant text, and thus shortens the choice duration of a request, concerning
text relevance acceptance.
The main idea of our research is to find text contraction bounds by sentence com-
pression without major content loss. Its originality relies on constituents syntactic func-
tion and position in the syntactic tree, to select deletable constituents.
In next section, we enumerate the main automatic summarization approaches, then
we compare those working at a finer granularity level (section 2); we then outline our
sentence compression method (section 3); we illustrate its effectiveness with experi-
menting a prototype application-applied to story/short novel texts (section 4); and fi-
nally we discuss about the results of this experiment and draw some perspectives (sec-
tion 5).
2 Summarization by sentence compression
In this article, we only focus on sentence compression.
[2] tackles the sentence compression problem by using a noisy-channel model con-
sisting in making the following assumption: “We look at a long string and imagine
Yousfi-Monod M. and Prince V. (2005).
Automatic Summarization Based on Sentence Morpho-Syntactic Structure: Narrative Sentences Compression.
In Proceedings of the 2nd International Workshop on Natural Language Understanding and Cognitive Science, pages 161-167
DOI: 10.5220/0002570201610167
that it was originally a short string, and then someone added some additional, optional
text to it. Compression is a matter of identifying the original short string.”. The aim is
then to locate this optional text and to remove it. To do so, the authors use a Bayesian
probabilistic model trained on a corpus composed by documents with their summary.
[3] focuses on detecting and removing relative clauses which are preceeded by
clauses like NP
P rep NP
, where NP
and N P
are noun phrases and P rep is
a preposition. The purpose is to correctly attach the relative referent by choosing a wide
or local attachment.
These two approaches based on textual units shorter than sentences do not take into
account the sentences constituents syntactic function and position in the syntactic tree.
In fact, function and position are naturally useful to help choosing the constituents to
be removed. Moreover, such a technique is easily checked by human examination.
3 Compression by pruning the syntactic tree
The starting point of our approach was the insight that the sentence constituents syn-
tactic function and position in the syntactic tree, plays a weighty role in the con-
stituents importance for text understanding. This insight comes from logical gram-
matical analysis. Indeed, some adjective phrases, adverbials, etc, are not systematically
needed to understand the main sentence meaning,
This approach needs a sentence morpho-syntactic analysis tool (section 3.1) and a
survey on constituents importance relative to their syntactic function and position in the
syntactic tree (section 3.2). We present our system architecture in the section 3.3.
3.1 The morpho-syntactic analyser
Since our working language is French, our experiments have been run on this language.
However, the same methods can be easily transposable to English or other languages
for which syntactic parsers have been developed.
We use the French morpho-syntactic parser called SYGFRAN, based on the op-
erational system SYGMART, both defined in [1]. SYGFRAN uses a transformation
rules set of structured elements, based on French grammar rules. It transforms a sen-
tence (raw text) in a syntactic tree (structured element) enriched with information about
constituents. This parser has the following advantages: the fastness: the analysis com-
plexity is O(k n log
(n)) where k is the rules number and n the text length. the
robustness: SYGFRAN manages to produce a correct structure for at least 30% of the
different cases of French sentences syntaxes, for other cases, SYGFRAN provides a
partial but workable analysis. the production of a syntactic tree: much of the existing
syntactic analysis systems only achieve a basic linear tagging and those providing a tree
are not robust enough relatively to the body of existing syntactic constructions.
SYGFRAN takes a raw text input and produces a bracketed structure, correspond-
ing to the morpho-syntactic tree of each text sentence, in which many variables are
acquainted on the different constituents natures, syntactic functions, canonical forms,
grammatical categories, tense, gender, number, etc.
3.2 Function and Position
The constituents deletion test is addressed by many French grammar works to help
in attaching a syntactic function to a constituent. The test is validated if the resulting
sentence remains grammatically consistent. However, linguistic texts dealing with the
constituent importance in the sentence according to their syntactic function are rather
uncommon. Some recommendations are provided by linguists, but there is no funda-
mental rule.
So we have proceeded in the following way. We have considered these recom-
mendations as working assumptions and we have tried to support them empirically.
cuk, in his contemporary French analysis, speaks about syntactic functions known
as governement (in the aftermath of Chomsky’s works). Constituents are said to be
governors, if they are mandatory to grammatical coherence and to sentence semantics.
The sentence subject and its verbal group are governors in a grammatical coherence
We have noted three non-governor constituent categories likely to be deleted, ac-
cording to their syntactic function and their position: adverbials, epithets and apposi-
tions. As we can see, they have a medium granularity level. Appositions, when trans-
formed in relative clauses (noun complement) get a wider granularity level, thus in-
creasing the final compression ratio.
Adverbials. We have noticed that the most important adverbials where temporal and
purpose ones. They do answer the questions we deem the most important namely
“When ?” and “In which purpose ?” In the case where a location adverbial is present
after the verb “to be”, deletion cannot be done. ”to be” is a particular verb, and must be
cautiously dealt with.
However, if several location adverbials are consecutive, all but one can be deleted
without major content loss : John is
in the car, in the car park, near to the sweet shop.
At last, adverbials located in interrogative sentences appear to be extremely important
since they do issue the question.
Epithets. Adjectives, adjective phrases and some relative clauses (noun complement)
have an epithet function. In a way similar to adverbials, when an epithet si located after
the verb “to be”, and more generally after a stative verb, its importance considerably
increases, making deletion impossible.
Also, we have noticed that when the epithet is located in a noun phrase in which the
determiner is a definite article, then its deletion is difficult. The reason is the definite
article is used to speak about a specific entity and , thus the noun epithet allows to dif-
ferenciate this entity from others.
Appositions. Apposition may be of different types and might appear asa noun phrase, a
pronoun, a relative clause, a present participle clause, a past participle clause, an infini-
tive clause.In the first three cases, constituents can be easily deleted. Participle clauses
can be deleted too, but with a more important content loss. In the latter case, deleting the
clause appears to be more difficult, because the infinitive clause systematically provides
an important information completing the subject.
3.3 Architecture
Our system architecture is outlined in figure 1. It relies on all considerations provided
in the preceding section about the importance of constituents in a sentence. It is based
on a parser output in the form of syntactic trees, and produces as an output, a text
coloration of the deletable segments according to constituents hierarchy. The way the
system works is: source text is fed to SYGFRAN, which in turn produces syntactic trees.
Then, the textual segment selection/coloration module uses the following information to
accomplish the selection: the source text, syntactic trees and variables/values provided
by SYGFRAN, the size/loss ratio threshold not to exceed, being provided by the user or
defined by the application type and last, the constituents selection rules set, to achieve
the different constituents selection iterations until the size/loss ratio is satisfied. The
selected constituents are then deleted.
syntactic analysis
Source text
Syntactic trees and
associated variables
Contracted text
Colored text
If CC and ...
If epithet and ...
If apposition and ...
selection rules
User or
Constituents selection
iterations until
size/ loss
ratio satisfaction
Selection module
trees and
Fig.1. From the source text to the compressed text: our sentence compression system
4 Experiment
We have have implemented a part of our theory in a computer program to assess the
effectiveness of such an approach. We have defined a system using basic rules, based
on our experimental survey’s results (section 3.2).
Our current prototype only performs one iteration. The first step consists in coloring
deletable constituents. A color is assigned to each constituent type. So it is easy to assess
rules quality on the processed text before actually deleting these constituents.
In the second step, colored segments are deleted to produce the summary. The cho-
sen textis a French Haitian story. We have chosen a French text because the current rules
set of SYGFRAN allows it to analyze only French sentences. The reason of choosing
this story is that SYGFRAN produces a correct syntax for all the sentences of this text
and because it is a well-sized, good representative of what is a narrative text. The col-
oration result of a story part is presented in the figures 2 (the orignial French version)
and 3 (the English translated one).
5 Discussion
With our current rule set, our method has allown us to delete approximatively 34% of the
full text. We can note a light discursive content and coherence loss, which is more than
satisfactory relatively to current automatic summarizers. Moreover, the grammatical
consistency is preserved. We think our rules can be more refined, but there is a lack
of linguistic information in this domain. For this text, SYGFRAN provides us correct
syntactic trees, but variable values are not systematically true and full. For adverbials,
SYGFRAN only specifies the object semantics for the temporal and locative ones. The
other somehow lack semantic information.
Selecting rules of deletable constituents can be more refined according to con-
stituent function and especially to text types. Concerning this subject, we project to
carry out experiments on more texts dealing with more different types. However, sen-
tence compression is not sufficient to produce a summary of a satisfying size in most
application cases. As we have already seen, compression greatly depends on the text
type. So we consider our intra-sentential approach as one of the tasks to perfom in
the automatic summary production, in complement with other approaches working at a
granularity level at least as big as sentences.
6 Conclusion
Current automatic summarization approaches use information such as term frequency,
lexical relations, POS tags, probabilistical learning engines, texts rhetorical structure,
however, none of them use both constituents syntactic function and position in the
syntactic tree as our is able. Our approach has started by a survey on the sentence
constituents importance. The deletion criterion evaluates the contents and coherence
loss generated by constituents deletion. The selection criterion is based on constituents
syntactic function and position in the syntactic tree. Narrative texts (novels, stories, ...)
appeared to be the most suitable for such an approach. We have modeled a sentence
Au bout d’un moment elle bougea et marmonna: “Quelle sorte de nuit est-ce donc pour
durer si longtemps ?” Mais elle se rendormit
parce qu’il faisait aussi noir qu’au cœur
de la nuit dans la maison. Finalement elle se r
eveilla en sursaut et se mit
a chercher ses
Courant de tous c
es, elle arracha ce que Maui avait fourr
e dans les fentes.
Mais c’
etait le jour! Le grand jour! Le soleil
etait d
a haut
dans le ciel ! Elle s’empara
d’un morceau de tapa
pour se couvrir et se sauva de la maison, en pleurant
la pens
ainsi tromp
ee par ses propres enfants . Sa m
ere partie, Maui bondit pr
es du
store qui se balanc¸ait encore de son passage et regarda par l’ouverture. Il vit qu’elle
a loin
, sur la premi
ere pente de la montagne. Puis elle s’arr
eta, saisit
pleines mains
un arbuste de tiare Tahiti, le souleva d’un coup: un trou apparut, elle s’y engouffra et
remit le buisson en place
comme avant.
Maui jaillit
de la maison aussi vite qu’il put, escalada la pente abrupte, tr
ebuchant et
tombant sur les mains car il gardait les yeux fix
es sur l’arbuste de tiare. Il l’atteignit
finalement, le souleva et d
ecouvrit une
egende :
ements circonstanciels, proposition au g
Fig.2. Our text coloration/compression, original French version
After a moment, she stirred and muttered; ”what type of a night it is to be so long” ?
But she went back to sleep
because it was as dark in the house as in the core of the night
. Finally she woke up with a start and began to look for her clothes. Running everywhere
she tore up what Maui had slipped into the holes. It was day ! The full bright day ! The
sun was already high
up in the sky! She took a piece of tapa to cover herself and fled
from home,weeping at the thought that she had been so deceived by her own children.
His mother gone Maui jumped close to the window shade that was still moving after
her and looked through the openning. He saw that she was already far away ,on the first
slope of the mountain. Then she stopped, grabbed a Tahiti tiara bushtree with her whole
arms and lifted it up completely : a hole appeared, and she rushed in and then put the
bushtree back
like before.
Maui sprang up
from the house as quickly as possible, climbed up the abrupt slope ,
stumbling and falling on his hand, because his eyes were kept on the tiara bushtree.
He finally reached it, lifted it up and found a
Legend :
adverbials, gerund clause,
Fig.3. Our text coloration/compression, English translated version
compression system based on constituents deletion. The creation of a rule system based
on our model has allowed us to assess the feasibility of such an approach. We first
have colored the constituents according to selection rules, in order to judge the rele-
vance of each rule. Our method managed to delete approximatively 34% of the test text,
while preserving a good grammatical coherence. We thus conclude that our compres-
sion could be useful when used as one of the tasks of a wider automatic summarization
process, either as a first-phase running summarization, or as a post-phase, after having
removed larger chunks of text. We plan to augment accuracy of text sentences prun-
ing by running our system on important narrative text corpora, find heuristics for wider
portions of text deletion based on rethorical information use text types or domains to
introduce specific summary rules (scientific articles in which titles might help to delete
wide portions of text). All this, naturally, will be sorted out and put into a more sophis-
ticated system to provide a better set-up for summarization by compression.
1. Chauch
e, J.: Un outil multidimensionnel de l’analyse du discours. In: Coling’84, Standford
University, California (1984) 11–15
2. Knight, K., Marcu, D.: Summarization beyond sentence extraction: a probabilistic approach
to sentence compression. Artificial Intelligence archive 139(1) (2002) 91–107
3. Siddharthan, A.: Resolving relative clause attachment ambiguities using machine learning
techniques and wordnet hierarchies. In: 5th National Colloquium for Computational Linguis-
tics in the UK (CLUK 2002). (2002) 45–49