**Event Date**: Thursday, 26 April, 2018, 11 a.m.

**Location**: Via Santa Maria, 36, Pisa, PI, Italia [2nd floor seminar room]

**Speaker**: Prof. Denis Paperno (Lorraine Lab of Computer Science and its Applications, CNRS)

**Title**: Predictability in the semantic space: towards a distributional characterization of the inflection-derivation distinction

**Abstract**: The literature is divided on the relationship between inflection and derivation. Many morphologists, of various theoretical inclinations, hold that inflection and derivation are essentially the same thing (Robins, 1959; Di Sciullo & Williams, 1987; Bochner, 1993; Booij, 1996; KÅ“nig, 1999), or that there is at most a gradient distinction between canonical inflection and canonical derivation (Dressler, 1989; Corbett, 2010; Spencer, 2013). Many others hold on the contrary hold that there is an irreducible difference between the two (Matthews 1965, 1974; Anderson 1982; Perlmutter 1988; Aronoff 1994; Stump 2001). We focus on one of the criteria that are discussed in the literature, which we term stability of contrast. In a nutshell, inflection is supposed to be stable in its syntactic and semantic effects across lexemes (books is to book as cats is to cat), while derivation is expected to be less so (delegation is not to delegate as election is to elect). The idea that inflection and derivation differ in this way is intuitively compelling, and has been stated repeatedly (Robins 1959: 125-126; Matthews 1974, 49-52; Wurzel 1989, 36; Stump 1998). However, to our knowledge, no previous study has attempted to define stability of contrast in an operational fashion, and to test on a large scale the validity of a difference between inflection and derivation: rather, all studies discuss intuitive semantic contrasts between hand-picked series of pairs of words.To test the contrast stability of different morphological relations, we construct word form triples where R1 is an inflectional relation and R2 is derivational. By hypothesis, we expect the vector offsets for derivationally-related pairs to be more diverse than those for inflectionally-related pairs. We employ the Euclidian distance between the vector offsets and the mean vector for the same relation as our main measure of diversity. This is illustrated schematically in Figure 2. The null hypothesis in t-test analysis is that the means of the two paired samples are identical. The contrast stability hypothesis will be confirmed if the deviation from average is greater for derivational relations than for inflectional ones with a significant t-test value.Our experiment compares different kinds of relations between word forms in a controlled, paired-sample setting. The results support that there is a systematic contrast between derivational and inflectional relations. This finding fully agrees with existing theoretical literature and is meant to inform the ongoing debate on the status of derivation and inflection in language and cognition.

**Denis Paperno** is researcher an the National Center of Scientific Research (CNRS), France. He received a PhD in Linguistics from UCLA (2012) and did postdoctoral work at the University of Trento. Having worked on language documentation, formal semantics, and semantic typology, he currently focuses on distributional semantic models and their applications to linguistically motivated tasks