Committee: Alec Marantz (chair), Ailís Cournane, Richard Kayne, Liina Pylkkänen, Charles Yang (University of Pennsylvania)
Sentences are uncontroversially represented as hierarchical syntactic structures, not linear strings (Chomsky, 1957). The hierarchical nature of sentences has also been demonstrated in sentence processing at both behavioral (Roark et al., 2009; Fossum and Levy, 2012) and neural (Brennan et al., 2016;Nelson et al., 2017) levels. In contrast, despite the theoretical agreement on hierarchical syntactic structures in the domain of derivational morphology among “word-based” (Anderson, 1992; Aronoff, 1994) and “morpheme-based” (Lieber, 1992; Kayne, 1994) lexicalist as well as anti-lexicalist (Halle and Marantz, 1993) theories, words have been formally proposed to require less generative capacity than sentences (Heinz and Idsardi, 2011) and have been computationally implemented as linear strings of morphemes (Beesley and Karttunen, 2003). Consequently, the hierarchical nature of words has not been sufficiently addressed in morphological processing (with a few exception; e.g. Libben 2003, 2006) and even the psychological reality of morphemes has been denied (Baayen et al., 2011).
In this dissertation, extending the computational psycholinguistic approach developed in sentence processing to morphological processing (Marantz, 2013), we investigate whether morphological processing tracks hierarchical syntactic structures of words. Methodologically, computational models implemented in natural language processing are employed as models of human grammars to define probabilities over syntactic structures (Yang, 2004) and quantify processing costs via the informational-theoretic complexity metric called surprisal (Hale, 2001; Levy, 2008). The inventory of computational models investigated in this dissertation includes Markov Models with no syntactic structures, Hidden Markov Models with linear syntactic structures, Probabilistic Context-Free Grammars with hierarchical syntactic structures, and recent models of productivity such as Fragment Grammar (O’Donnell, 2015) and Tolerance Grammar (Yang, 2016). Then, the model predictions are statistically compared with the experimental observations obtained through five experiments in order to reverse-engineer the most “human-like” computational model.
Chapter 1 starts the dissertation by explaining the theoretical background and spells out the central hypothesis called the Hierarchical Morphological Processing Hypothesis that morphological processing tracks hierarchical syntactic structures of words. The overview of the dissertation is also provided.
Chapter 2 reviews theories of morphological structure and processing and linking hypotheses to connect the two. Specifically, we propose the grammatical architecture developed in the Minimalist Program (Chomsky, 1995) as the “single-route” model of morphological processing where syntactic structures are inevitable to connect forms and meanings (Marantz, 2005). Then, we exhibit formal grammars with different generative capacities situated in the Chomsky hierarchy and introduce computational models as probabilistic generalizations of formal grammars.
Chapter 3 presents two computational simulation experiments (Experiments 1 and 2) with bimorphemic derived words available from the English Lexicon Project (Balota et al., 2007), a “shared task” in morphological processing. The results of the experiments demonstrate that hierarchical models accurately explain reaction times in visual lexical decision experiments on top of various control predictors including surface frequency. Additionally, the residual errors of the hierarchical and sequential models will also be explored with respect to the distinction between Class I vs. II derivational affixes.
Chapter 4 conducts two crowdsourced behavioral experiments (Experiments 3 and 4), acceptability judgment on Amazon Mechanical Turk and visual lexical decision on IbexFarm, with novel trimorphemic derived words (Halle, 1973) developed to solve methodological limitations of the previous experiments. The results of the experiments indicate that hierarchical models outperform sequential models, replicating the computational simulation experiments, and predict morphologically complex words with center-embedding structures better than those with left-branching structures.
Chapter 5 performs an magnetoencephalography experiment (Experiment 5) to investigate what stage of morphological processing is sensitive to hierarchical syntactic structures of words, which cannot be addressed with time-insensitive behavioral measures. The experiment shows that the first stage of morphological processing indexed by the M170 (Pylkkänen and Marantz, 2003) and proposed to reflect morphological decomposition (Solomyak and Marantz, 2010) may parse morphologically complex words into hierarchical syntactic structures of words.
Chapter 6 concludes the dissertation by summarizing the results and discussing theoretical implications. Overall, the five experiments presented in this dissertation converge on the conclusion that morphological processing tracks hierarchical syntactic structures of words, strongly suggesting that there is no escape from syntax in morphological processing.