Michael Levison
Computing and Information Service
Queen's University at Kingston
Kingston, Onterio
K7L 3N6 Canada
levison@cs.queensu.ca
The measures of productivity of word-formation devices which have been used to date, such as lexicographical data (Dubois 1962) and corpus data (Baayen and Renouf 1996, Baayen and Lieber 1991) do not lend themselves well to the study of L2 productivity. Among other things, L2 corpora show relatively small amounts of productivity (Lessard, Levison, Maher and Tomek 1994, Boeder et alii 1993). Despite issues of reliability and interpretation (see Birdsong 1989, Gass 1994, Coppetiers 1987), elicitation appears to offer a potentially useful measure.
However, to our knowledge, little research has been done on elicitation to test lexical productivity in French, particularly among L2 learners. Hawkins (1985) forms one exception, but he was concerned primarily with the class of past participle markers, which are at the borderline of affixation and derivation.
The research described here draws upon and extends previous work on native speaker judgements of lexical productivity, including Aronoff and Schvaneveldt 1978, Gorska 1982, Anshen and Aronoff 1991, Levison and Lessard, 1995a, Fowler and Liberman 1995 and Keane and Costello 1997. It should be noted however that very little previous work used a computational environment, whereas this is central to the work presented here, and is based on the VINCI natural language generation environment.
In the first, judgements of acceptability were elicited from native speakers of French for derived forms in -able, -age, -ment, -tion and -ure. In a nutshell, a French lexicon sorted by frequency was used to provide verbs of the first conjugation. Suffixes were added automatically, and at random. A randomized subset of the resulting derived forms was presented to each subject, along with the base form of each verb. Subjects were required to identify which verbs were known to them (almost all, in the case of the native speakers) and which derived forms they found acceptable. Results defined a continuum of relative acceptability with -able at the top of the list, followed by -age, -ment, -tion and -ure in that order. It is important to note that the variable being measured was neither the correctness of the judgements (whether an existing derived form corresponded to those seen as acceptable) nor the individual lexical items, but rather the ranking of suffixes in terms of the number of derived forms they were seen to be capable of producing.
In the second stage, the same protocol was applied to non-native speakers ranging from some with high school training in French to some with significant university level studies in French. Results of these tests showed that as knowledge of verb bases decreased, non-native speakers showed increasing discrepancy in their judgements with respect to native speaker rankings.
The third stage addressed problems found in the second: the absence of an external measure on which to rank the linguistic skills of the subjects tested, and the relatively high level of linguistic competence of almost all the subjects tested. In response to these problems, the experiment was repeated using the same protocols. Subjects tested were students in oral French classes at Queen's University. Five levels were represented, ranging from 016 to 320, where 016 represented those with essentially no previous knowledge of French, while 320 represented those with near-native proficiency. Placement in these classes had been done on the basis of an oral interview.
As an illustration, the results for three of the five suffixes are reproduced in the following table. In the table, the columns suffix ok and suffix not ok represent the average of all responses for each class on the basis of 20 questions. In principle, each row should sum to 20, however because some students responded to less than 20 questions, some small variations are found. Verb known and verb not known represent subjects' claimed knowledge of the base verb.
| Suffix | Course | Verb known | Verb not known | ||
| Suffix ok | Suffix not ok | Suffix ok | Suffix not ok | ||
| 016 | 2 | 2.25 | 2.75 | 12.75 | |
| 017 | 3.75 | 1.5 | 8.75 | 6 | |
| -able | 118 | 6.5 | 2.5 | 7 | 4 |
| 219 | 9.4 | 3.7 | 2.4 | 3.8 | |
| 320 | 9.1 | 4.1 | 2.1 | 3.7 | |
| 016 | 2 | 3 | 1.5 | 13 | |
| 017 | 4.75 | 3.25 | 8.5 | 3.5 | |
| -ment | 118 | 5 | 4.5 | 9.5 | 1 |
| 219 | 8.7 | 4.3 | 2.7 | 3.6 | |
| 320 | 4.6 | 8.1 | 2.1 | 4.6 | |
| 016 | 1.5 | 4.25 | 2.5 | 11.75 | |
| 017 | 1.75 | 4.75 | 3.5 | 9.75 | |
| -ure | 118 | 2.5 | 9 | 2 | 6 |
| 219 | 1.3 | 10.9 | 1.3 | 5.7 | |
| 320 | 2.9 | 10 | 1 | 5.4 | |
The table shows that in all cases, the number of base verbs known rises with level, suggesting that knowledge of verbal base and placement interview results are measuring comparable things. As well, bearing in mind that the acceptance rates by native speakers for derived forms based on -able, -ment and -ure were 77%, 39% and 10% respectively, we see in the non-native speaker data a gradual convergence on native speaker-like judgements. Thus, in the case of -able, while 016 students find little to choose between accepting of rejecting derived forms for verbs they know (2 acceptances and 2.25 rejections) and reject strongly derived forms for verbs they don't know (2.75 acceptances versus 12.75 rejections), more advanced students tend to accept derived forms for verbs they know (in the case of 219, 9.4 acceptances versus 4 rejections) while rejecting somewhat derived forms in -able for verbs they don't know.