1 INTRODUCTION
Morphology, like other fields of linguistics, benefits greatly from using corpora. By nature, it is perhaps the most able to deal with large amounts of electronic data, due to the ease with which word forms are gathered and processed by simple computer programs. This is why morphology is a linguistic domain whose evolution is directly connected to the amount of available data, and it is understandable that its evolution has partially paralleled the technological evolution in data retrieval, from paper word lists and dictionaries to text corpora and, most recently, to the World Wide Web.
A question which is often raised in corpus linguistics concerns the quantity vs. quality issue: more data may correspond to a less controlled corpus, and an increase in the amount of available data leads to increased noise, but also to more relevant data. However, this question can be considered in a different way from a morphologist's point of view. Since it focuses on the low-level mechanics of word coining (in our case, mostly suffixation), whose interactions with other linguistic phenomena are somewhat reduced, the study of (derivational) morphology can exploit raw material that could be considered unsuitable for other studies. On the contrary, more data is indeed better data. In fact, only very large corpora can allow the observation and the study of rare phenomena, which may be too little represented, if at all, in a traditional corpus (such as a list of dictionary entries). Data-intensive morphology can also lead to new perspectives on previously studied mechanisms, and easily confirms or invalidates intuition and conjectures.
This paper will first describe the techniques used for gathering word forms from the Web, providing researchers with quantities of new data, sometimes up to ten times more than with previous methods. We will then present a number of recent studies that have used such extensive data focusing on specific French suffixes (-esque, -este, -able, -ment). Each of these case studies will show how the use of greater quantities of data led to new insights on a given morphological phenomenon.
2 COLLECTING MORPHOLOGICAL DATA ON THE WEB
In the past, the data used for most morphological studies consisted of word-lists manually extracted from dictionaries or corpora. Increased access to digital resources has, however, made it possible to easily automate the gathering of such data: electronic dictionaries such as the TLFi (Trésor de la langue française informatisé) or text databases such as Frantext can provide large lists of words containing a given string of characters (corresponding e.g. to a suffix, such as -able) in a matter of seconds. Such a simple operation would have taken several months some fifteen years ago.
In addition to traditional digital corpora (mostly consisting, for French, of literary texts and newspaper archives), there is an increasing use of the World Wide Web as a resource for linguistics studies. Access to this resource can easily be obtained through generic all-purpose search engines such as Google or Yahoo, which allow researchers to instantly get contexts and frequency of use for a given word form.
Given the sheer size of the data searched, the Web can be used to extend the study of a particular pattern, giving access to many more occurrences of a target form, and providing a more intricate insight into its meaning(s) and use. Even more interestingly, the Web can also be used to discover word forms which have not been recorded in lexicographic works (due to their rareness and/or novelty), and whose very existence would have remained a simple conjecture. For the latter type of research, the Web is superior to all existing traditional corpora because of its size, variety and constant evolution. The Web has thus been successfully used to discover words previously considered to be unlikely or even theoretically impossible to coin. For example, one can find the prefix anti- attached to simple (non constructed) adjectives, such as in anti-triste (‘anti-sad’) or anti-obèse (‘anti-obese’), and even to adjectives following the V-able scheme, such as anti-inflammable (‘non-flammable’). This latter case was previously declared as theoretically impossible for instance in Fradin (Reference Fradin1997:100–101).
Using the Web for linguistic purposes, however, is subject to both methodological and technical difficulties.
2.1 Using the Web as a corpus for linguistics
Of course, the Web cannot be considered as just a huge data store which linguists can use without taking any precautions.
The advocates of the Web as a corpus emphasise its many advantages: its size, the presence of many different text types (including some that cannot be found anywhere else), the variety of languages used, its constant evolution, etc. According to them, every corpus study should use the Web as soon and as widely as possible. In fact, many researchers have been doing so for many different tasks, in domains ranging from translation to collocations. However, most of the studies of this kind use the Web with little questioning of its content and are mainly related to computational linguistics.
The opposite point of view is to consider that the Web is not a corpus at all. The following arguments have been used in this direction (Lüdeling, Evert and Baroni Reference Lüdeling, Evert, Baroni, Hundt, Nesselhauf and Biewer2007):
• it has not been constructed or balanced in any way (as compared to generic corpora such as the British National Corpus) and cannot be viewed as representative of anything except itself;
• one cannot find even the simplest information about it as a whole (the size of the Web cannot be measured, and only a part of it is accessible through search engines);
• most Web pages provide no information regarding their author (age, sex, nationality, date of writing, or even proficiency in the language used);
• the Web is constantly evolving (pages appear and disappear), making most experiments impossible to reproduce;
• only the crudest methods of access are available (restricted to simple keyword searches).
It is then understandable that some linguists simply refuse to consider data from the Web as linguistic material useful for research. In a less drastic way, many researchers are simply using the Web as a source for building traditional corpora, after a careful selection of documents. A great amount of work is also in progress for dealing with some of the above questions, for example efforts are made for archiving Web pages,Footnote 1 identifying Web text types (Santini Reference Santini, Mertens, Fairon, Dister and Watrin2006), and even building a search engine able to provide sophisticated modes of access to Web pages (Resnik and Elkiss Reference Resnik, Elkiss, Knight, Ng and Oflazer2005). However, none of these improvements can compete today with the generic search engines. Their speed, ease of use and amount of data available make them the major access point to Web data, despite their many limitations.
Not all of these problems can be solved, and as a result each occurrence has to be manually checked.
2.2 Using Web search engines
Any linguist can easily use a search engine to instantly verify her/his own intuitions, or check some uses of a word, without needing any specific software or computing skill: the only thing to do is type a word form on the home page of the search engine. The given results are the approximate number of Web pages known by the engine to contain the word, and the list of Web addresses of these documents illustrated with short text extracts.
Doing this on a larger scale, however, calls for specific programs that can automate the querying of search engines. This can be done on several levels. Tools such as Webcorp (www.webcorp.org.uk) are built on top of search engines for linguistic purposes. Webcorp provides concordances and collocates, which are very useful for syntax, semantics or discourse analysis.
But, morphological studies need to have access to subparts of words: the search for forms in this case cannot be done without technological means of handling substrings of graphic words. The basic tool needed should be able to retrieve word patterns, such as those commonly expressed by wildcards, e.g. “*able” which stands for “any word ending with able”. Search engines do not provide this, and so morphologists can greatly benefit from dedicated software such as Walim (Namer Reference Namer, Fradin, Dal, Kerleroux, Hathout, Plénat and Roché2003). This tool uses computational morphology techniques to automatically generate inflected word forms, given a list of bases and a derivation scheme (i.e. a suffix). It then uses the word forms as queries to a search engine, and only retains those for which at least one occurrence can be found.
Using the same principle, and adding some features for the discovery of new word forms, Webaffix (Hathout and Tanguy Reference Hathout, Tanguy, Rodríguez and Araujo2002) can be used in two different ways to gather derived words from the Web.
The first method, based on hypothesis testing, follows the same principle as Walim. The main difference between these tools is that Webaffix places emphasis on the filtering of the raw results from the search engine: it checks for the correct target language, typos, bad word divisions, etc. It can also perform a more restrictive selection, retaining only derived words that co-occur with their base forms in a Web page, an efficient criterion for the morphological link between the two. For instance, copolymérisable (‘copolymerizable’) will only be retained as a legitimate -able adjective form if the deduced base verb copolymériser (in any of its inflected forms) appears in its vicinity.
The second method takes advantage of some Web engines that allow limited use of wild cards in the queries (see above),Footnote 2 and is thus able to gather new word forms in an inductive way, without any assumption or knowledge of the base form. An automatic analysis is performed on each new word in order to calculate its base form, and the co-occurrence test is performed to ensure a correct analysis.
This second method, which makes no assumptions regarding the base form, has been extensively used for instance in Hathout, Plénat and Tanguy (Reference Hathout, Plénat and Tanguy2003) to gather many new -able words whose base cannot be found in any major dictionary (e.g. acronyms or proper names).
2.3 Caveats
One very common use of Web engines by linguists is searching for frequencies. In morphology, frequencies are useful for measuring the productivity of a given process, and for comparing two competing word forms (see Baayen Reference Baayen, Booij and van Marle1991; Fradin et al. to appear). Such experiments, if performed on the Web, mostly rely on the comparative numbers of occurrences of different word forms, as indicated by any search engine for a given query. The use of these raw numbers must be subject to a lot of caution for several reasons. First, it indicates a number of documents, not words, and several occurrences of the same word form in the same Web page only count as one. The other obstacle, more related to the nature of the Web itself, is the duplication of occurrences, which covers different phenomena such as quotes, plagiarism, or simply the verbosity of a given Web page author which can lead him to use a newly coined words dozens of times. The last problem that these numbers raise comes from the search engines themselves. These systems are conceived to deal with huge quantities of data and to give an instant answer to the user. This often leads them to produce approximate and unstable numbers (the same query can give different results even within a few hours). In addition, the commercial competition between the major search engines often makes them boast more pages for a given query than they can really display. All these facts should lead to increased carefulness when using these numbers in a quantitative analysis, such as the comparison of two word forms’ frequencies. If the difference between presence and absence can be taken as significant, as well as very large differences between two frequencies (such as 2 vs 100,000 documents), smaller quantitative differences (such as 10 vs 15 documents) should not.
Other limitations are due to the technology used by search engines for automatic language identification: many documents returned for a given query are not written in the selected language, and should not be considered as valid occurrences. The impact of this problem heavily depends on the studied phenomenon: -able tends to give English words, while -este and -esque are commonly used in other Romance languages (including Latin, whose presence on the Web has to be taken into account). However, documents written in these languages can be easily detected and discarded.
But the most annoying problem when dealing with Web data comes from the complete lack of information regarding the author and the context (as seen in 2.1.). This means that some detected forms are wrong for a number of reasons, and cannot be taken into account. Examples of encountered error types are:
• direct transfers from another language (most of the time the author's mother tongue). These include the results of machine translation;
• words coined for stylistic reasons (rhyme, pun, etc.);
• regionalisms or archaic words;
• plainly incomprehensible contexts, either from low-quality writing or technical jargon.
2.4 Methodological issues
Most of the problems mentioned above cannot be solved by automated means, even if Webaffix implements some heuristics to detect non-French Web pages and a few kinds of suspicious contexts. As a result, every new word form harvested by such a tool has to be manually examined. This examination makes full use of the context: the Web page itself, but sometimes other pages on the same site. As mentioned above, the main point is to sort out legitimate constructions from errors, to get their meaning and to identify their base forms.
The number of occurrences of a given word has not been taken into account: one legitimate and one interpretable context per target word is considered sufficient. Thus, frequency was a minor issue in the studies presented here. However, the search for “good” contexts often leads the linguist to examine several occurrences in order to get a clear view.
The variety of the Web confronts us with many different types of discourses, some of them making understanding difficult. In a few cases, the authors (if identifiable) have been contacted in order to get confirmation of their intent to coin a new word. In other cases, overly obscure contexts, and thus words, have been discarded. As can be seen in the examples described hereafter, the main advantage of Web data is the spontaneity found in the productions: many interesting findings occur in forums, blogs, and other types of pages where a community of speakers freely express themselves in informal contexts. This kind of data, as will be discussed below, is one of the most interesting aspects of using the Web as a corpus for morphology.
Perusing data automatically extracted from the Web remains a time-consuming task, but the experiments described in the following sections will demonstrate the utility of such work.
Another important point to be mentioned is the fact that the great majority of the examples retrieved and used in the works described below were consistent with native speakers’ intuitions and sounded ‘natural’. The advantage of the Web, in this case, is that it gives access to all (or at least the majority of) the cases we can find in relationship to a particular construction.
3 SAMPLE STUDIES USING EXTENSIVE MORPHOLOGY
There is no doubt that the efforts for automatic retrieval of word forms in the huge mass of electronic texts have brought about significant progress in our knowledge of morphology. Recent studies show that this ‘extensive’ approach permits new and finer generalisations, as the number of collected forms grows. In particular, extensive morphology allows us to record rare facts whose existence was uncertain due to the weakness of our intuitions. Finally, this approach makes it possible to design experiments capable of renewing, at least partially, the empirical bases of our discipline.
This section presents some of the results, in the fields of morphophonology and morphosemantics, obtained through extensive morphological studies.
3.1 New generalisations
The progress brought about by the extensive morphology approach has been rapid, in particular in the domain of morphophonology. Concerning French, the most spectacular results are certainly those regarding the phenomena of dissimilation. As an example, we describe below how Marc Plénat and his collaborators progressively uncovered the conditions under which certain rhymesFootnote 3 are deleted before the suffix -esque. As far as semantics is concerned, progress is slower, but some examples, such as derivation with -able, show that we can expect some questions to be completely reconsidered. The choice of these suffixes was not completely arbitrary: -esque is a suffix which can be very productively attached to proper nouns. Thus, through Web searches, we expected to find a large amount of forms not recorded in traditional lexicographic sources. -able was chosen because it has been often considered as a ‘simple’ element to describe, with relatively straightforward combinatorial properties and meaning, and all (apparently) deviant cases had been treated as ‘marked’ or ‘peripheral’.
3.1.1 Mid vowels before -esque
The data which provided the largest amount of new observations are certainly those collected in the database of derived words in -esque developed by the ERSS research group. This base currently includes some 3,000 different forms, each accompanied by one or more referenced examples. By comparison, the Robert électronique and the TLFi each contain less than 100 words derived with this suffix. This database has constantly grown through different sources during 15 years of work: starting with a simple selective reading of books and newspapers, it then took advantage of the availability of digital corpora, and more recently of the Web search techniques previously mentioned.
This database sheds new light on the morphophonology of French in many ways. To illustrate our point, we will only relate here the appearance and progressive clarification of a problem so far ignored: the behaviour of words ending in a mid front vowel (/e, ε, ø, æ/) followed by a fixed (i.e. stable, non latent) consonant before -esque.
If we trust standard lexicographic sources, bases ending with such a sequence do not pose any special problem, since the few words recorded in the above mentioned dictionaries are formed by simply concatenating the suffix to the base lexeme as it appears in its free form (cf. 1):
(1)
The idea that identity (or similarity) between the vowel in the suffix and the final vowel of the base lexeme could cause the deletion of the rhyme first appeared in the mid-90s, when the base contained some 800 items, and was put forth in Plénat (Reference Plénat1997: 168). The new list showed in fact that a rhyme in /ε/ followed by a fixed consonant may sometimes disappear when the base is at least four syllables long, be it a simple (2a) or a complex (e.g. a derivative in –eur, 2b) word:
(2)
The data also included one case in which the ending -eur disappeared in a base of only three syllables (tirailleur ‘sharpshooter’ → tiraillesque). The trisyllabic Cervantes was also shortened in cervantesque, but this was not different from the typical behaviour of trisyllabic words ending in /s/ (cf. 3):
(3)
Some years later, the retrieval of some 400 new forms did not lead to a significant advance: Plénat (Reference Plénat2000: 32) points out the relative weakness of endings in a mid front vowel + a fixed consonant before -esque for long bases, but is incapable of precisely determining the factors that cause the shortening.
The current database allows a much more precise description of data:
(i) It is confirmed that a rhyme consisting of a mid front vowel + a fixed consonant may be deleted when the base lexeme is at least four syllables long. It is also confirmed that with shorter bases such a rhyme is maintained. (4) shows some pairs we find in the database:
(4)
(ii) However, rhymes may be deleted not only when the last consonant is identical to one of the consonants in the suffix (as in cervantesque, see above, or in BTesque formed on BTS /beteεs/ ‘technical-scientific high-school diploma’), but also if the last consonant is already represented at least once in the base (cf. 5). In other words, the tiraillesque case is now explained: it is the presence of two /r/ in the base which causes -eur to disappear:
(5) Footnote 4
(iii) Finally, they are also deleted even if the base only has two syllables when its last fixed consonant is identical (or almost identical) to one of the consonants of the suffix. (6) displays examples containing a sibilant or a /k/:
(6)
Other rhymes in a sibilant may disappear at the end of a disyllabic base (cf. 7), but less systematically than when the last vowel of a base is /ã/. Moreover, rhymes in /k/ only fall in words with three or more syllables (cf. 8).
(7)
(8)
As we can see, the increase in the amount of available data is comparable to the introduction of the microscope for natural sciences. In those areas of morphology where the observation of the few recorded forms did not show anything interesting, an enlargement of 30× revealed an appreciable number of new facts, leading to new conclusions.
The observations above fit into a broader picture where conservative forces tend to preserve the integrity of the base lexeme and of the affix, and in the most basic cases impose the simple concatenation of the two. However, in French, two sorts of constraints oppose these forces: length constraints, which penalise forms of more than three syllables (Plénat to appear), and dissimilation constraints, such as the well-known Obligatory Contour Principle (OCP), put forth, among others, by McCarthy (1986). To resume, constraints of the second type penalise the repetition in the same form of two identical or similar segments (for French cf. Plénat Reference Plénat, Durand and Laks1996, Reference Plénat2000, Lignon and Plénat to appear). None of these constraints can stand alone against the conservative forces, but when they operate jointly, they can trigger the truncation of the base (cf. Burzio Reference Burzio and Boucher2002 for a similar view of ‘alliances’ between lower ranked constraints).
3.1.2 Semantic plasticity of derived adjectives in -able
Increasing the amount of available data is also a decisive factor for the semantic and categorial dimensions of morphological descriptions. The study of the -able derivation in French by Hathout, Plénat and Tanguy (Reference Hathout, Plénat and Tanguy2003) gives a good indication of the kind of advances made possible by the accumulation of new data.
Adjectives derived with -able have often been regarded as de-verbal adjectives with a passive meaning. In other words, the noun they modify is analysed as corresponding to the direct object or the patient of the base verb, depending on whether the relation is viewed as syntactic or thematic. This analysis has been questioned in previous studies on -able derivation by Leeman and Meleuc (Reference Leeman and Meleuc1990), Leeman (Reference Leeman1992) and Anscombre and Leeman (Reference Anscombre and Leeman1994), among others. These authors have put forward semantic arguments (the analysis does not capture the categorisation dimension of the derived adjectives in -able) as well as distributional ones (not all verbs which can be passivised have a derivative in -able). The three studies mentioned were mainly founded on lexicographical lists and recourse to grammaticality judgements on large scale. The size of the corpus was approximately 1,400 words, which roughly corresponds to the number of adjectives ending in -able that are listed in large dictionaries like the Grand Larousse de la langue française, the Grand Robert de la langue française or the Trésor de la langue française (T.L.F.).
For their study, Hathout, Plénat and Tanguy (Reference Hathout, Plénat and Tanguy2003) gathered a much larger list of about 5,000 adjectives. They made use of the two collecting methods provided by Webaffix and systematically analysed the usage of some of these derived words on the Web. The study proposes a new analysis of the -able derivation which generalises the traditional one while taking into account the counter-arguments mentioned above.
The collected data show that most of the -able derivatives indeed have a passive meaning. However, the noun they modify can also represent a variety of other participants in the process. This plasticity can be easily illustrated by looking at the possible nouns that a derived adjective such as pêchable ‘fishable’ (which, by the way, does not appear in the T.L.F.) can modify. Obviously, first among the things that can be said to be pêchable are fish and other kinds of seafood. However, places can be qualified as pêchable as well: (i) bodies of water (rivers, ponds, streams, etc.) and (ii) fishing spots like riverbanks, bridges, dams, etc. Depending on whether the fishing season is open or not, whether the weather is nice or bad, seasons, days and atmospheric conditions can also be said to be pêchable or impêchable (‘unfishable’). (9) presents some examples obtained from Google.Footnote 5
(9) 31 Aout Eau très haute (9,7 m3/s) et froide (9°C), premier jour pêchable depuis le 15 Aout. Quelques gobages, surtout des petits poissons, . . . sosdessoubre.free.fr/saison%202004.html
C'est vrai, la carte de pêche complète à 75€, rapportée aux nombres de jours pêchables, et même si ça augmente chaque année, ce n'est pas hors de prix. . . .www.achigan.net/msgforum.php?id_sujet=936&page=3
Jusqu’ à 14 ça va, au delà je sors pas car le vent devient trop gênant voir impêchable. maintenant peut-être que tu as des bras trop frêles pour résister à . . .
peche-en-mer-aquitai.1fr1.net/DIVERS-f5/LIENS-f18/METEO-COEF-f21/previsions-meteo-etat-de-l-ocean-t23.htm
Si le vent monte trop et que les conditions ne deviennent plus pêchables, plusieurs solutions s'offrent à vous:. – tout plier et attendre une accalmie . . . pechemed.free.fr/loupsurf1.htm
The authors have also found contexts where impêchable modifies fishing tackle (flies or nylon fishing lines, for instance, see (10)). Finally, not only the participants in the process, but their properties too can be characterised as pêchable or not: they came across examples where fish size is said to be pêchable (11):
(10) je remarque aprés quelques lancers (je peche generalement a 40 metres en etang) que mon nylon se met a vriller et devient impechable. . . .www.pechemaniac.com/forums/viewtopic_208.htm
(11) pêchable, l'ouverture du gisement à la pêche semble incompatible avec sa gestion durable. Compte tenu de la raréfaction des coques de taille pêchable la . . .www.reservebaiedesaintbrieuc.com/DOC/coques2002.pdf
Actually, the fishermen seem to be the only participants that cannot be said to be pêchable!
Discovering the existence and even the proliferation of these uses which had not been identified before helps us understand some better known adjunct uses:
• verbs denoting ‘building’ (12) or ‘movement’ (13) yield adjectives which modify nouns that designate locations:
(12)
(13)
• verbs denoting ‘work’ yield derivatives which modify nouns that designate periods of time:
(14)
• verbs denoting a punishment yield adjectives which modify nouns that denote charges:
(15)
We also better understand that a property like price can be said to be abordable ‘affordable’:
(16)
3.1.3 Denominal adjectives in -able
Classical descriptions of the -able suffixation report a number of derivatives coined from nominal bases. Gawelko (Reference Gawelko1977) had identified three small series of such adjectives, derived from names of taxes (17), vehicles (18) and titles (19):
(17)
(18)
(19)
This inquiry confirms the existence of these series. Modern taxes (20), recent vehicles (21) and titles of all sorts (22) yield numerous -able derivatives.
(20)
(21)
(22)
But there is more: the study brought to light other denominal adjectives which match some of the remarkable deverbal series. For instance, the adjectives derived from nouns of vehicles as in (18) and (21) can be connected to the ones derived from verbs denoting movement like in (11). In the same way, the denominal examples in (23) parallel the adjectives derived from verbs denoting building like the ones in (12).
(23)
We also can draw a parallel between adjectives derived from verbs denoting condemnation as in (15) and denominals like (24). The examples in (23) and (24) illustrate the fact that frequency does not play an important role in the decision to accept an example or to reject it: boxable occurs thousands of times and peinable de mort only once. Nevertheless, both are perfectly acceptable.
(24)
Other denominals are remarkable not because of the nouns they modify but because of the thematic role of their bases which can be a location as in (25) or a final state as in (26).
(25)
(26)
In summary, the inquiry clearly confirms that categorial constraints on the -able derivation have a semantic origin: -able derivatives usually select verbs as bases because they denote processes, but when a process does not have a specific corresponding verb, a nominal base will do quite nicely.
Building up a database like the one we just mentioned is time-consuming. Even if the harvesting process has become very fast, the validation of the collected data involves lengthy philological work. But we have demonstrated that the gamble is paying off: new generalisations have been revealed in phonology, categorial constraints and semantic interpretation.
3.2 Rare facts
Looking systematically for occurrences of new words does not conflict with intuition. On the contrary, our experience suggests that speakers can have very sharp and interesting intuitions, even for extremely rare configurations. For this kind of configuration, a Web search can confer the status of verifiable facts to intuitions that would otherwise have remained mere conjectures. A spectacular example of this type concerns the substitution of the suffix -esque with the ending -este.
Pichon (Reference Pichon1940) had proposed an analysis of an isolated example by Verlaine (27) where he suggested that the substitution resulted from a dissimilation phenomenon that takes place after velar consonants.
(27)
In the next half century, this conjecture has been cited many times but no other example in which -este was substituted to -esque after /k/ or /g/ had been found. Recently, Plénat et al. (Reference Plénat, Fradin, Kerleroux and Plénat2002) discovered half a dozen new examples, through Web searches carried out with Webaffix. For now, their database includes about thirty new words ending in -gueste and -queste, some of them being (very) well-attested, such as those in (28):
(28)
A point of detail in the morphology of adverbs ending with -ment provides us with another illustration of the Web's capacity to confirm uncertain intuitions (Dal Reference Dal, Hathout and Montermini2007; Plénat and Boyé to appear). It is well-known that these adverbs are based on the stem which is also used to build the feminine of their corresponding adjective (29).
(29)
It is also well known that, exceptionally, adjectives ending in -ant and -ent yield adverbs in -amment and -emment respectively (30). This characteristic can be explained historically: in Old French, these adjectives were epicene words (i.e. they had only one form for both genders).
(30)
However, not all adjectives in -ant and -ent yield adverbs in -amment and -emment. Yvon (Reference Yvon1996:164) notes the dubious acceptability of adverbs such as (31) (his suspicion is based on a remark by Molinier Reference Molinier1992).
(31)
Actually, we have the feeling that the derivatives in (31) are quite ungrammatical and that the adverbs derived from these adjectives must instead be coined from the feminine stems, as in (32). This unacceptability could be explained by the occurrence of two /m/ sounds and two similar vowels in two consecutive syllables (/mamã/): the choice of the feminine stem follows then from a dissimilation constraint.
(32)
For some time, this intuition was just a hypothesis due to the scarcity of arguments which supported it. The only one available was the presence of véhémentement among the four known exceptions to the general rule (33).
(33)
A recent Web search provided extra arguments in support of this view. One can find about fifty good examples of charmantement, even if charmamment appears five times too, mainly in blogs. Besides, there is one additional attestation of charmantement in Frantext by Albert Cohen, in his novel Mangeclous. As for the adverb derived from clément, the Web search provided two good examples of clémentement but only one of clémemment in a French-English dictionary which blindly applies the rule of grammar handbooks. Good attestations of two other exceptions (34) were discovered. We did also find two occurrences of (35), but both appear in bad (or machine) translations.
(34)
(35) alarmamment
‘alarmingly’
Of course, these few remarks do not exhaust the subject, but these examples lead us to believe that the hypothesis was correct and that French tends to resort to adverbs in -mantement and -mentement instead of -mamment and -memment.
The previous discussion also shows that native speakers’ intuition remains essential, and that the heterogeneity of the corpora we can extract from the Web prevents us from accepting all the material without closer examination.
4 CONCLUSION
The different experiments described in this article all used an unprecedented amount of data, which was gathered through automated means, mostly from the Web. The advances achieved by these data-intensive studies on several levels have been shown. The availability of these data opens the way towards further studies: finer descriptions using more homogeneous corpora; quantitative studies; large scale comparative studies (Web data are available for a large number of languages). So far, only a few suffixes have been investigated in this new way and almost all French affixes remain to be studied.
From a methodological point of view, Web data are of course different from what can be extracted from traditional corpora (text databases or newspaper archives), as the new word forms found on the Web can frequently be characterised as being of a lower register or from specialised terminology. The Web actually covers a variety of language usages, many of which have not been previously taken into account in wide spectrum linguistics studies (except for some popular magazines or paperback novels). The informal context of the Web, especially the lack of editorial filtering, indeed gives access to more spontaneous word coinage.
In no case should this ‘new’ kind of data be taken as a weakness for the theories deduced from it. Most spontaneous words only add to more acceptable ones in the same paradigm, and they can easily be used as a source for new hypotheses.
Using Web material leads to new skills in detecting the nature and status of some Web pages, which is crucial in manual filtering. Generic linguistic competence is also useful in judging the level of proficiency of a given Web page's author, not to mention hunting down and excluding automated translations.
Extensive morphology certainly has good prospects. However, the rapidity in the development of this field of research will depend on the emergence of a community of extensive morphologists able and willing to share their databases. New standards have to be proposed for these databases in order to ease their reuse and their merging.
ACKNOWLEDGMENTS
We would like to thank Marc Plénat for opening the way to extensive morphology, and for enthusiastically involving us in his research. We are also grateful to Jacques Durand and Jesse Tseng for helping us improve the quality of this paper. All remaining errors are of course ours.
