generalizing pa erns in instrumented item and pa ern
play

Generalizing paerns in Instrumented Item-and-Paern Morphology Sarah - PowerPoint PPT Presentation

Generalizing paerns in Instrumented Item-and-Paern Morphology Sarah Beniamine and Olivier Bonami Universit Paris Diderot Laboratoire de linguistique formelle Labex EFL, opration Morph1 SNCL Workshop, May 30, 2016 0. Bonami & S.


  1. Generalizing paerns in Instrumented Item-and-Paern Morphology Sarah Beniamine and Olivier Bonami Université Paris Diderot Laboratoire de linguistique formelle Labex EFL, opération Morph1 SNCL Workshop, May 30, 2016 0. Bonami & S. Beniamine Generalizing paerns in Intrumented IPa May 2016 1 / 29

  2. Introduction Introduction: Item and Paern Morphology • Morphology is modeled directly in terms of surface alternations • Term due to Blevins (forthcoming); preferable to the ambiguous ‘Word and Paradigm’ • Consider French adjective paradigms: • Surface alternations between Lexeme . . . . forms lead to opacities that are  lokal loko lokal lokal problematic for speakers.  banal banal banal banal • Classical phonological and  ɡɛ ɡɛ ɡɛ ɡɛ morphological analyses do not  lɛ lɛ lɛd lɛd model these opacities, but try  ʁɛd ʁɛd ʁɛd ʁɛd to reduce them.  pʁɛ pʁɛ pʁɛt pʁɛ  nɛt nɛt nɛt nɛt  njɛ njɛ njɛz njɛz  obɛz obɛz obɛz obɛz  epɛ epɛ epɛs epɛs  ɛkspʁɛs ɛkspʁɛs ɛkspʁɛs ɛkspʁɛs 0. Bonami & S. Beniamine Generalizing paerns in Intrumented IPa May 2016 2 / 29

  3. Term due to Blevins (forthcoming); preferable to the ambiguous ‘Word and Paradigm’ Consider French adjective paradigms: Introduction Introduction: Item and Paern Morphology • Morphology is modeled directly in terms of surface alternations • . ∼ .: two paerns Lexeme . . . . 1 Xal ∼ Xo  lokal loko lokal lokal 2 X ∼ X  banal banal banal banal • This leads to uncertainty, as  ɡɛ ɡɛ ɡɛ ɡɛ some . in -al do not  lɛ lɛ lɛd lɛd alternate.  ʁɛd ʁɛd ʁɛd ʁɛd  pʁɛ pʁɛ pʁɛt pʁɛ • Thinking about morphemes (or  nɛt nɛt nɛt nɛt processes) does not help  njɛ njɛ njɛz njɛz address that uncertainty.  obɛz obɛz obɛz obɛz  epɛ epɛ epɛs epɛs  ɛkspʁɛs ɛkspʁɛs ɛkspʁɛs ɛkspʁɛs 0. Bonami & S. Beniamine Generalizing paerns in Intrumented IPa May 2016 2 / 29

  4. Term due to Blevins (forthcoming); preferable to the ambiguous ‘Word and Paradigm’ Consider French adjective paradigms: Introduction Introduction: Item and Paern Morphology • Morphology is modeled directly in terms of surface alternations • . ∼ .: numerous Lexeme . . . . paerns  lokal loko lokal lokal 1 X ∼ X  banal banal banal banal 2 X ∼ Xd  ɡɛ ɡɛ ɡɛ ɡɛ 3 X ∼ Xt  lɛ lɛ lɛd lɛd 4 X ∼ Xz  ʁɛd ʁɛd ʁɛd ʁɛd 5 X ∼ Xs  pʁɛ pʁɛ pʁɛt pʁɛ • This leads to more uncertainty.  nɛt nɛt nɛt nɛt  to  unpredictable C drop  njɛ njɛ njɛz njɛz  to  unpredictable epenthesis  obɛz obɛz obɛz obɛz • Thinking about underlying  epɛ epɛ epɛs epɛs representations does not help  ɛkspʁɛs ɛkspʁɛs ɛkspʁɛs ɛkspʁɛs address that uncertainty. 0. Bonami & S. Beniamine Generalizing paerns in Intrumented IPa May 2016 2 / 29

  5. Term due to Blevins (forthcoming); preferable to the ambiguous ‘Word and Paradigm’ Consider French adjective paradigms: Introduction Introduction: Item and Paern Morphology • Morphology is modeled directly in terms of surface alternations • Item and Paern Morphology Lexeme . . . . focuses on modeling  lokal loko lokal lokal alternations themselves.  banal banal banal banal • We can then quantify how  ɡɛ ɡɛ ɡɛ ɡɛ harmful opacity is.  lɛ lɛ lɛd lɛd  ʁɛd ʁɛd ʁɛd ʁɛd • We do not try to infer abstract  pʁɛ pʁɛ pʁɛt pʁɛ representations from which to  nɛt nɛt nɛt nɛt reconstruct the surface forms.  njɛ njɛ njɛz njɛz ☞ Not unfeasible or  obɛz obɛz obɛz obɛz uninteresting, but a different  epɛ epɛ epɛs epɛs enterprise.  ɛkspʁɛs ɛkspʁɛs ɛkspʁɛs ɛkspʁɛs 0. Bonami & S. Beniamine Generalizing paerns in Intrumented IPa May 2016 2 / 29

  6. Introduction Introduction: Intrumented IPa • Instrumented Item and Paern Morphology (IIPa) 1 Based on large, machine-readable datasets (corpora or lexica) (e.g. Albright, 2002) ⋆ Evaluating the prevalence of morphological phenomena is crucial ⋆ Enough data to see correct generalizations despite Zipfian distributions 2 Fully implemented analytic strategies (e.g. Albright, 2002; Stump and Finkel, 2013) ⋆ Systematization of descriptive practice ⋆ Cross-linguistic applicability 3 Focus on quantitative methods (Ackerman, Blevins, and Malouf, 2009; Ackerman and Malouf, 2013) ⋆ Gradience of morphological complexity • See among others Bonami and Boyé (2014), Bonami and Luı́s (2014), Sims (2015), and Malouf (2016) 0. Bonami & S. Beniamine Generalizing paerns in Intrumented IPa May 2016 3 / 29

  7. Introduction Introduction, 3 • This talk focuses on the notion of a paern of alternation that is at the heart of current work in Instrumented IPa. • The plan: 1 Present key analytic techniques in Instrumented IPa 2 Evaluate the importance of the choice of a particular classification of alternations 3 Outline a new algorithm 4 Present preliminary results on Zenzonpetec Chatino (Oto-Manguean) 0. Bonami & S. Beniamine Generalizing paerns in Intrumented IPa May 2016 4 / 29

  8. Results in Instrumented Item and Paern Morphology

  9. Results in Instrumented Item and Paern Morphology Results in Intrumented IPa 1 Evaluating the predictibility of inflectional paradigms 1 Implicative entropy Principal part systems 2 2 Inflectional classification 1 Inference of macro-classes Inflection systems as semi-laices of classes 2 0. Bonami & S. Beniamine Generalizing paerns in Intrumented IPa May 2016 6 / 29

  10. Results in Instrumented Item and Paern Morphology Predictivity in inflectional paradigms When a speaker knows only one form of a lexeme, how hard is it to predict the others? (Ackerman, Blevins, and Malouf (2009)’s Paradigm Cell Filling Problem) Consider French adjectives: . . • . ⇒ . is trivial • . ⇒ . is easy but not trivial, see /lokal/ ∼ /loko/ vs. /banal/ ∼ /banal/ • . ⇒ . is harder, see /lɛd/ ∼ /lɛ/ vs. /ʁɛd/ ∼ /ʁɛd/ • . ⇒ . is hardest, see /ɡɛ/ ∼ /ɡɛ/ vs. /lɛ/ ∼ /lɛd/ vs. /njɛ/ ∼ /njɛz/ vs. … . . 0. Bonami & S. Beniamine Generalizing paerns in Intrumented IPa May 2016 7 / 29

  11. Results in Instrumented Item and Paern Morphology Implicative entropy Lexeme . . alternation . class  lwajal lwajo X al ∼ X o C1  banal banal X ∼ X C1  kalm kalm X ∼ X C2  poli poli X ∼ X C2 Data sample: French masculine adjectives • For each pair of cells ( A , B ) , over a set lexicon: • Group lexemes by type of alternation: random variable A ∼ B • Group forms in A by shape, on the basis of which alternations these shapes are compatible with: random variable A A ∼ B • The implicative entropy from A to B is the conditional entropy of paerns of alternation given input cell. H ( A ⇒ B ) = H ( A ∼ B | A A ∼ B ) • In our example: H ( . ⇒ . ) = 0 . 5 0. Bonami & S. Beniamine Generalizing paerns in Intrumented IPa May 2016 8 / 29

  12. Results in Instrumented Item and Paern Morphology Using implicative entropy • Creole complexity Language Mauritian French 0.017 Average 0.744 0.446 . . 0.039 Minimum 0.563 0 6 0 Maximum 0.925 0.916 2 . 0 0.190 (Bonami, Boyé, and Henri, 2011) 0.190 0.528 0.206 0.561 0 . 5 • Prediction from multiple cells 0 2 . 5 8 6 1 French E. Portuguese 0 . . 1 predictor 0.174 0.205 0 Average: 0.252 2 predictors 0.054 0.106 3 predictors 0.021 0.076 (Bonami and Beniamine, 2015) 0. Bonami & S. Beniamine Generalizing paerns in Intrumented IPa May 2016 9 / 29

  13. Results in Instrumented Item and Paern Morphology Systems of principal parts • Principal part system: a set of perfect predictor cells • A traditional pedagogical tool • Stump and Finkel (2013): Cardinality of the smallest such set is an indicator of the complexity of an inflection system. • Can be deduced from implicative entropy ▶ Set of cells from which implicative entropy to all other cells is 0 Language 1 cell 2 cells 3 cells French conjugation 0 0 0 E. Portuguese conjugation 0 184 7884 Number of distinct categorical systems of principal parts 0. Bonami & S. Beniamine Generalizing paerns in Intrumented IPa May 2016 10 / 29

Recommend


More recommend