As a step in a project whose final goal is to propose a Controlled Natural Language for requirements writing at CNES (Centre National d'Études Spatiales), we intend to build the grammar of the textual genre of the requirements. One of the main issues faced when analyzing our corpus is the (sometimes subtle) difference between the terms and syntactic structures pertaining to the genre and those linked to the domain (in our case, the development of space systems) - a difference that is generally not taken into account by automated tools. In this paper, we present a methodology aimed at detecting candidate terms and textual patterns specific to the genre by combining results obtained from a terminology extractor and a data mining tool with a validated resource in use for indexing documents at CNES. The results are then illustrated by a selection of examples from our corpus.
This study is part of a wider project aiming at improving the writing of requirements1 at CNES (Centre National d’Études Spatiales), the French Space Agency.
Indeed, the requirements (as well as the
specifications, that is, the documents in which they
are included) are mostly written in a natural
language – in this case, in French –, and as a
consequence they may sometimes contain well-known
related problems, such as ambiguity and
vagueness
However, in order for this CNL to be actually
applied, we believe that it should not be
unnecessarily restrictive and, in particular, not too far
removed from the way engineers are already
used to write the documents – otherwise, they
will probably merely ignore it. In other words,
we wish to propose a CNL inspired by already
existing data, following a corpus-driven and
corpus-based methodology that we describe
more in details in
This methodology relies on the existence of a
textual genre, which
In the present study, we will focus on the
results obtained by a terminological extraction.
More specifically, we will propose a method to
sort them (as we are interested only in the terms
pertaining to the genre, not in those pertaining to
the domain) and subsequently to use them as a
filter to retrieve textual patterns belonging to the
grammar of the genre. An example of similar
work, based on collocations and n-grams, is
given by the transdisciplinary scientific lexicon
Although this grammar should ideally be independent of the field (aerospace industry, aeronautics, software engineering, etc.), in practice, the distinction is not so simple as regards specifications2. While some features are indeed inherent in the nature of the documents (because they describe something that does not exist yet, but will have to exist and to conform with the requirements, the use of the future tense and injunctions, for instance, are common), others, however, are closely related to the field to which belongs the future “object” being described. It may reasonably be assumed that the lexical level – since it directly refers to the object in question – is most significantly affected by the domain, but we cannot reject the hypothesis that syntactic structures too may differ from one field to another.
For that reason, if we want to define a
terminology of requirements, we must keep in mind
that the candidate terms proposed by the
terminology extractors may actually belong either to
the genre or to the domain. Unfortunately,
although the possibility to filter terms by domain
has already been highlighted as a user need
Specifications are thus unusual, specialized corpora and they bring new challenges to terminology extraction in general. In particular, considering the fact that the candidate terms linked to the domain are probably more numerous than those linked to the genre, we want to find a way to exploit the results without a need for manually revising all of them. In the next section, we present the small experiment we conducted on our corpus of specifications as a possible way to reach this goal, but also to reuse these results to filter textual patterns identified by a text mining tool. 3 3.1
All the operations described hereafter were performed on two corpora of requirements in French extracted from several specifications provided by the CNES. (All tables and figures were removed from the requirements, because their automatic analysis would have been more difficult.) The first corpus concerns the project called “Pleiades” 3 (two very-high-resolution satellites for Earth observation) and is composed of nearly 120,000 words; the second corpus, related to the smaller project “Microscope”4 (a microsatellite, whose main objective is to verify a physical principle), contains nearly 44,000 words. Although the requirements were written under similar circumstances and represent the same levels of specifications for the two projects, it is worth noting that Pleiades and Microscope have totally different scales and purposes. Consequently, the fields to which they relate are at least partially distinct. 3.2
First of all, candidate terms for both corpora
were extracted using the terminology extractor
developed for the Talismane toolkit
Since they included candidate terms for the
genre and for the domain (see section 2), and
since we are interested only in the former, all the
entries present in a list of terms used at CNES
for indexing documents in their knowledge base
were removed. This list of domain terms (used
here as a “stop list”) has been augmented for
many years thanks to internal documents of
various types and carefully validated by domain
2 The distinction between genre and domain itself is
actually far from trivial
In order to remove even more candidate terms supposedly linked to the field, we decided to keep only entries present in both lists (Pleiades and Microscope). This resulted in a much shorter list of just 300 candidate terms (meaning 1,055 were exclusive to Pleiades and 298 to Microscope). This step makes sense because the specifications of Pleiades and Microscope are comparable at many levels, but also because, as already mentioned, the two projects are sufficiently distinct. Hence, whereas the first selection was useful to eliminate candidates related to the field at a more general level (e.g. “satellite” or “simulation”), here some of the candidates were not kept because they are more dependent to one of the two projects, and thus more specialized (e.g. “magnétomètre” ‘magnetometer’ or “masse interne” ‘internal mass’). (However, because the corpus of specifications from Pleiades is almost three times larger than the other corpus, it is also probable that some terms, such as “priorité” ‘priority’, could have appeared in the Microscope corpus as well.)
Lastly, we proceeded to a manual revision of the remaining candidate terms to eliminate some entries that were obviously noise. The final list contains 267 candidate terms (to be compared with the original list, which would have contained over 1,850 different candidates, or almost 2,000 if the extraction had been performed on the two corpora as a whole). Interestingly, the terms seem to concern both functional requirements (e.g. “fonctionnalité” ‘functionality’) and nonfunctional requirements (e.g. “disponibilité” ‘availability’). 3.3
Of course, a grammar of genre should not be limited to the lexicon, as it would be the case with the results of the terminological extraction. We would like to identify recurring syntactic structures or, at least, frequent textual patterns5 with the help of text mining tools.
For this purpose, we used SDMC
In order to reduce this number to a more reasonable proportion, we have decided to keep only patterns containing at least one of the remaining candidate terms (for the sake of simplicity, the noun phrases were reduced to their heads); indeed, we assume that the structures based on terms belonging to the genre are themselves more likely to be typical of this same genre. This restriction limited the number of patterns to approximately 6,000, among which “être connaître avec un [précision] 6 meilleur que (number)” ‘be know with a [precision] better than (number)’, “être conforme au [format]” ‘be consistent with the [format]” and “devoir respecter le [contrainte]” ‘must respect the [constraint]’.
The list can be further reduced by focusing on patterns containing a verb. In this way, we consider an intermediary level between the lexicon and the discourse.
To conclude this section, the main steps of the process we described are represented by Figure 1.
corpus 2
text mining
patterns
filtered
patterns
5 Patterns of this kind are the basis of the so-called
“boilerplates”
terms candidate terms terminological extraction
stop list common entries candidate terms candidate
terms candidate terms
In this section, we briefly discuss some of the results we obtained after applying the process described previously. 4.1
Some terms belonging to the space domain remain: initialisms (“ASH”, “DGAPC”), terms too general to be useful for indexation (“mission”, “centre de contrôle” ‘control center’), terms of the field (“tuyère” ‘nozzle’, “calibration”).
Others, by contrast, belong more to the genre. They may describe a need (“besoin de test+programmation+restitution” ‘need for a test+programmation+restitution’) or the characteristics of the objet that is described (“taille du buffer temporaire+du paquet TM” ‘size of the temporary buffer+TM packet”, “durée de désaturation+la manoeuvre” ‘duration of desaturation+the manoeuvre’); they can specify expected functions (“fonction de gestion+filtrage” ‘function of management+filtering’); or they can be related to the management of the project: possible problems (“défaillance” ‘failure’, “défaut” ‘defect’), necessary documentation (“rapport d’avancement+d’expertise” ‘progress+expertise report’), validation (“acceptation” ‘acceptance’, “confirmation”, “autorisation” ‘authorization’).
Some terms can belong either to the field or to the genre, depending on their modifier: “date de début du produit” ‘starting date of the product’ (genre) vs. “dates de début et de fin de vidage TM” ‘starting and ending dates of the emptying of the TM’ (field, because of the domain terms “vidage TM”). 4.2
The most frequent verbs in the patterns are: “être” ‘to be’, “devoir” ‘must’, “permettre” ‘to allow’, “mettre” ‘to put’, “prendre (en compte)” ‘to take (into account)’, “fournir” ‘to provide’, “pouvoir” ‘to be able’, “définir” ‘to define’, “passer (en mode+dans l’état)” ‘to enter (a mode+a state)’, “contenir” ‘to contain’, “donner” ‘to give’, “utiliser” ‘to use’, “gérer” ‘to manage’, “sélectionner” ‘to select’, “rejeter” ‘to reject’, “traiter” ‘to process’, “correspondre” ‘to correspond’, “générer” ‘to generate’, “décrire” ‘to describe’, “tenir” ‘to hold’, “exécuter” ‘to execute’, “vérifier” ‘to verify’, “calculer” ‘to calculate’.
Some structures based on these verbs are typical of the corpus:
[Det N permettre de (V+deverbal noun)]: “le DUPC permettra de modifier localement les paramètres du calcul”.
[Det N fournir Det N1 (à Det N2)]: “cette interface fournit les positions navigateur de l’instrument”.
[Det N utiliser Det N2 (pour V)]: “le système GIDE utilisera le protocole FTP pour effectuer les transferts”.
[Det N fournir (à Det N2) Det N3]: “le système de navigation fournira au système informatique central une réference de temps”.
[Sur réception de cette TC, le LVC exécute la procédure de mise ON+OFF de Det N (, par l’envoi de commandes (sur+vers+à Det N3))]: “sur réception de cette TC, le LVC exécute la procédure de mise ON de la carte IOT sélectionnée, par l’envoi de commandes discrètes sur l’OBMU” (only in Pleiades).
[Det deverbal noun doit s’exécuter (conditions)]: “la consolidation du scenario de travail au CECT doit s’exécuter en moins de 15 secondes” (only in Microscope).
[Det N (avoir la capacité de+être (capable de+autorisé à)) traiter Det N2]: “le CCC doit avoir la capacité de récupérer et traiter 291 Mo de TM par jour”.
These regular structures are therefore part of the grammar of the genre of requirements (at CNES). 5
As emphasized in section 2, specifications of space systems represent a particular type of corpus, because the terms of the domain and the terms of the genre are closely linked – making it difficult to automatically distinguish them. In section 3, we described the methodology we applied to keep only the terms belonging to the textual genre, using an existing resource (built for other needs) and a comparison between two corpora. This also allowed us to identify some structures (textual patterns) belonging to the grammar of the genre, which are used for writing functional requirements (describing expected functions) as well as for non-functional requirements (describing qualities or constraints applied to the system). The grammar could be refined thanks to existing guides to writing specifications that specify the various sections of the documents and the different types of requirements, which are likely to be expressed in different ways.
Nevertheless, it also appears that it is not always possible to draw a line clearly separating terms of the field and terms of the genre, since some terms may belong to both categories. In any case, the interpretation of the results remains dependent on the objective(s) being pursued.
Finally, we used this experiment as a proofof-concept; before we can generalize it, we would have to ask for validation by experts (experienced writers). It would also be very interesting to compare our corpus to specifications written in another domain.