Machine Assisted Study of Writers’ Rewriting Processes
Julien Bourdaillet
, Jean-Gabriel Ganascia
and Ir
ene Fenoglio
University Pierre and Marie Curie - LIP6
104 Quai Kennedy, 75016 Paris, France
Ecole Normale Sup
erieure - ITEM
45 rue d’Ulm, 75005 Paris, France
Abstract. This paper presents a joint work between artificial intelligence and lit-
erary studies. As part of the humanities, textual genetic criticism deals with writ-
ers’ rewriting processes. By studying drafts and manuscripts issued from these
processes, the genesis of the text is discovered. When draft comparison is done
manually, it requires a huge amount of work. The introduction of the machine
provides a high gain on efficiency and enables to focus on the interpretative work.
The application we developed relies on a sequence alignment algorithm close to
the ones used in molecular biology. This paper describes the textual alignment al-
gorithm, presents an experimental validation, and illustrates the textual analysis
with two genetic studies.
1 Introduction
Textual genetic criticism is a school of literary studies born during the seventies in
France [1]. Its main contribution is to study all the versions of a text, the final edition
becoming one version among others. This broadening of the scope of the study intro-
duces a temporal dimension by dealing with a sequence of versions. Finally, the goal is
to make the genesis of the text emerge from the drafts in order to understand the writers’
rewriting processes.
Two draft examples are presented in Figure 1. From these manuscripts, linearized
transcriptions are extracted, resulting in a sequence of versions. The comparison of
these versions enables the identification of invariants and differences between them.
When the author inserts or deletes a sentence or a paragraph, the comparison is not
so hard; but when spelling faults are corrected or micro-modifications are done, this
becomes tedious. Further, authors work on their style, they sometimes want to “free” a
word or a phrase in order to use it at another position in the text, that is to move a word
or a phrase. This move detection is very hard for humans, but it is a requirement for
genetic interpretation because moves are one of the four operations defined by genetic
criticism, along with insertions, deletions and replacements.
In fact, this corresponds to the notion of edit distance with moves which has been
studied in theoretical computer science [2]. This model is an extension of the usual
edit distance which can be computed in a quadratic time; but the introduction of the
Bourdaillet J., Ganascia J. and Fenoglio I. (2007).
Machine Assisted Study of Writers’ Rewriting Processes.
In Proceedings of the 4th International Workshop on Natural Language Processing and Cognitive Science, pages 222-227
DOI: 10.5220/0002429002220227
move operation makes the problem NP-complete and there exists no polynomial time
algorithm [3].
It means that the automation of this textual comparison work is a difficult prob-
lem for computer scientists. In this paper we present our application named MEDITE
dedicated to this work [4], an experimental validation, and two genetic studies.
2 Algorithm
In this section we present our algorithm named MEDITE; it is closely related to frag-
ment alignment commonly used in bioinformatics [5]. The two texts are considered
as two character sequences A and B which are processed in five steps. The first step
is a pre-processing step where character equivalence classes are set between identical
upper- and lower-case characters, accentuated or not characters and separators. Geneti-
cists choose which equivalence classes to use depending of the analysis type.
The second step identifies repeated character blocks between A and B by building
a generalized suffix tree between A and B [6]. This data structure enables to find all
repeated character blocks between A and B. The size of this set of blocks is exponential
and only a subset is interesting: the subset of super maximal exact matches (SMEMs),
that is matches of maximal size which are not included in other SMEMs.
The third step aligns these SMEMs in order to determine which are invariant and
which are moved. The pairwise alignment of SMEMs enables to make the decision:
aligned SMEMs are considered as invariant and unaligned as moved. Because the space
of possible alignments is combinatorial, we use an A
heuristic algorithm based on the
computation of the symmetric difference between remaining block to align.
The fourth step consists in looping over the pairings resulting from step 3 and in
considering the subsequences in A and B between each pair of aligned blocks. These
subsequences are processed again with steps 2 and 3. It allows the pairing of new blocks
which are then included in the main alignment. This recursive step enables the pairing
of blocks which would otherwise have been left unaligned.
The last step is the deduction of insertions, deletions and replacements. Insertions,
deletions and replacements can then be deduced. Deletions are non-repeated blocks in
A and insertions are non-repeated blocks in B. Further, when there is a deleted block
d in A and an inserted block i in B between two pairs of aligned blocks, and the ratio
|d|/|i| reaches a threshold t, then d and i are transformed in replacements r
and r
meaning that r
has been replaced by r
. t is arbitrarily set to 0.5.
3 Experimental Validation with Synthetic Data
In this section we present an experimental validation of the algorithm. The goal is to
evaluate the quality of MEDITE on synthetic data when a reference alignment exists.
Given a text and a noise model, a second text is generated by altering the first one;
the alignment between the texts is recorded during the alteration process. Then, it is
possible to evaluate the quality of the algorithm by comparing its results with the refer-
ence alignment. This kind of experiment is not possible with texts issued from genetic
criticism because there exists no reference alignment.
This experiment compares MEDITE with GREEDY which is a greedy algorithm
for computing the edit distance with moves [3]. GREEDY pairs the longest blocks first
and considers them as moves, finally it computes the classical edit distance by dynamic
programming. Five modified documents were generated with the noise model from a
20 KB text. Then the documents were aligned with MEDITE and accuracy rates eval-
uated. Two series of tests with different modification ratios were conducted: in the first
one there are 5 % of insertions, 5 % of deletions, 5 % of replacements, and 5 % of
moves, which means that there are 20 % differences between original and modified
texts; in the second series, the ratio is set to 10 %, meaning that there are 40 % dif-
ferences. For each of the five kinds of characters (insertions, deletions, replacements,
moves and invariants) the accuracy rate is defined as the number of correctly aligned
characters / the total number of characters. Then the average accuracy rate is calculated
for these five rates. For the average weighted accuracy rate, the four accuracy rates are
weighted with their ratio of the texts’ sizes. For example, for the first test series, we
have weighted accuracy = 0.80 Invariant acc. + 0.05 Insertion acc. + 0.05
Deletion acc. + 0.05 Replacement acc. + 0.05 Move acc.. The average runtime
of alignments is also calculated. Figure 1 presents the results of this experiment.
Table 1. Synthetic data alignment with GREEDY and MEDITE.
Modification ratio
5 % 10 % 5 % 10 %
Average accuracy 39.20 % 37.72 % 84.87 % 80.19 %
Av. weighted accuracy
75.16 % 59.75 % 93.45 % 86.26 %
Average runtime
19mn 25s 48mn 59s 0mn 9s 0mn 14s
For the average precision, GREEDY’s results are around 40 points inferior to ME-
DITE’s results, and for the average weighted precision around 25 points inferior. Fur-
ther, the runtimes are an order of magnitude different.
4 Machine Assisted Genetic Analysis
4.1 Short Text Analysis
We present a genetic analysis of a short passage of a novel named “La Robe Noire” (The
Black Dress), by Andr
ee Chedid a Franco-Egyptian contemporary writer, where she
describes her relations with her mother when she was a teenager. Figure 1(a) presents
the manuscript of which several linearized transcriptions have been extracted: they are
the versions of the text. These versions are compared automatically using MEDITE,
then the following manual interpretations are done.
In this text, most of the modifications are related to the mother character “Elle”
(She). The main fact discovered during this study is the complete transformation of the
writer’s viewpoint to her mother, between the first and the last version.
Fig.1. Manuscript exemples: (a) “La Robe Noire” by Andr
ee Chedid, and (b) “F
ete des Chants
du Marais” by Pascal Quignard.
The first version describes the mother as a loving and maternal mother. It starts by
the sentence “Elle nous aimait.” (She loved us.). But the set of modifications transforms
her in an imposing and brilliant character. The first sentence strength is reduced be-
coming “Nous nous sentions aim
es par elle” (We felt loved by her). In the first version,
the phrase “l’avant sc
ene qu’elle occupait avec
eclat et assurance” (the first place she
was occupying with brightness and assurance) was clearly minimized by the preced-
ing passage “gommant sa flamboyance pour nous offrir l’avant-sc
ene.” (minimizing her
flamboyant style to offer us the first place). In the last version, the minimization is re-
stricted: “ Il lui arrivait pourtant de s’effacer pour un temps, de reculer, d’abandonner”
(Sometimes, she kept in the background for a while, leaving). The phrase “parfois
avec abn
egation” (sometimes with self-abnegation) becomes “avec intelligence et sans
egation” (with intelligence and without self-abnegation). The following sentence
is inserted: “Elle jouait alors pour un temps, les seconds r
oles, exaltait - exag
parfois les qualit
es de celle ou de celui qu’elle d
ecidait de placer sous les feux de la
rampe.” (For a while she was playing the second roles, exalting, sometimes excessively,
the qualities of the one she decided to give the first place.); it clearly shows that she
controls and manipulates the social positions inside the family.
This short text shows the interest of genetic analysis. The meaning of the first and
the last versions is completely different and can not be guessed by studying only the last
edited version. MEDITE enables to make easier the comparison work and to discover
this kind of genetic evolution.
4.2 Genetic Folder Analysis - Global Viewpoint of a Text Genesis
We present a genetic analysis with the help of MEDITE of a tale named “F
ete des
Chants du Marais” (Swamp Song Feast) by Pascal Quignard, a French contemporary
writer. The tale relates a song contest for children in the Marais (a Paris district) during
the sixteenth century. The author has produced five manuscripts (or versions) of the text
and each version has been annotated giving different sub-versions, called states; this
gives 15 different states of the text. The set of these states is called the genetic folder.
The early analysis consists in comparing the first and the last state of the text, the
following facts can be observed:
The rewriting process is expansive; there are more insertions than deletions.
By examining replacements, we remark that the words “Palaiseau” (main charac-
ter’s first name) and “gouverneur” (governor character) were added several times.
These additions reinforce the weight of those characters in the text and focus the
attention on them.
Moves involve only single words, there are no phrase moves.
This first analysis enables to browse modifications and to discover interesting ones.
It is then possible to look for the genetic moment of a modification. For example, if
we wonder when the following sentence appears: “On monte le cr
ane au Grenier des
es” (The skull is carried up to the Reformeds’ Attic), by comparing the versions,
we observe that between the second and the third version the sentence “On monte le
ane au Grenier” is added, and between the third and the fourth version “des R
is added. This enables to observe the genesis of specific text portions and to understand
that the final version results from multiple rewritings.
More generally, we analysed the 15 states with MEDITE by comparing the states
pairwise and chronologically; three rewriting campaigns have been discovered:
1. Between the first and the second version: The main characteristic of this campaign
is to focus on the style; the author improves his style by bringing the text closer to
the linguistic norm.
For example, the following modifications have been done: “du visage” (of the face)
becomes “de son visage” (of his face); “soit” becomes “f
ut”, present subjunctive to
imperfect subjunctive of to be; “- Pourquoi ? Tu doutes” (- Why ? You doubt) be-
comes “Le gouverneur: - Douterais-tu” (The governor: - Would you doubt), moving
from oral style to written style.
2. Between the second and the third version: This campaign focuses on the beauty
theme. A lot of terms related to beauty are inserted.
For example, the following expressions are added: “qui
etait tr
es beau” (who was
very beautiful); “le Beau Palaiseau” (the Beautiful Palaiseau); “Son bel air s’ajoute
a sa beaut
e” (His beautiful attitude improves his beauty”); “radieux, le bel adoles-
cent” (shining, the beautiful teenager).
Further, some expressions are magnified:
a l’
eglise” (to the church) becomes
la chantrerie de la basilique” (to the basilica’s chant-school); “le cr
ane” (the skull)
becomes “la t
ete de mort” (the death’s head).
3. Between the third and the fifth version, the campaign focuses on passion. The fol-
lowing expressions are added: “passion
ement catholique” (passionately catholic);
“Le Palaiseau souffre et pleure.” (The Palaiseau suffers and cries.); “Je les d
(I hate them); “Nous ne l’aimons gu
ere” (We dislike it).
Further, by considering the positions of these expressions, we can remark that they
enable to strengthen the distinction and the opposition between Catholics and Re-
formeds which is latent in the text.
By comparing the last state of the first version and the first state of the second ver-
sion, we discovered that there are a lot of modifications. They were typed by the writer:
MEDITE makes appear these modifications that are not visible on the manuscript (these
ones where there are no pen corrections). They correspond to the first stylistic rewrit-
ing campaign. On the other hand, for the third campaign on passion, corrections are
not typed: the author corrects directly on the manuscript, with his pen, because more
considerations are needed to emphasize the focus on passion.
5 Conclusion
In this paper, MEDITE has been presented; it addresses the problematic of textual align-
ment issued from genetic criticism. The algorithm is based on character sequence align-
ment using a suffix tree and an A* heuristic alignment algorithm. It handles fine-grained
modification detection and move identification. The experimental validation on syn-
thetic data shows good results. Two genetic analyses were presented; they showed how
the machine helps the geneticist in the study of rewriting processes. MEDITE is now
used by text geneticists for literary studies.
1. Deppman, J., Ferrer, D., Groden, M., eds.: Genetic Criticism - Texts and Avant-textes. Uni-
versity of Pennsylvania Press (2004)
2. Lopresti, D.P., Tomkins, A.: Block Edit Models for Approximate String Matching. Theoretical
Computer Science 181 (1997) 159–179
3. Shapira, D., Storer, J.A.: Edit Distance with Move Operations. In: CPM. Volume 2373 of
Lecture Notes in Computer Science., Springer (2002) 85–98
4. Bourdaillet, J., Ganascia, J.G.: Alignment of Noisy Unstructured Text Data. In: Proc. of the
IJCAI Workshop on Analytics for Noisy Unstructured Text Data (AND 2007) of the 20th
International Joint Conference on Artificial Intelligence (IJCAI 2007). (2007) pp. 139–146
5. Bray, N., Dubchak, I., Pachter, L.: AVID: A Global Alignment Program. Genome Res. 13
(2003) 97–102
6. Gusfield, D.: Algorithms on Strings, Trees and Sequences: Computer Science and Computer
Biology. Cambridge University Press (1997)