At the nexus between computation and theory:

In his dissertation research, Yohei Oseki investigated how to model the recognition of visually presented morphologically complex words.  In the most promising model that he confronted with behavioral and MEG data, words were parsed by a probabilistic context free grammar of morphological structure, from left to right in the word.  Cumulative surprisal over the word, given the correct parse, correlated with reaction time in lexical decision and with brain responses around 170ms post-stimulus onset around the Visual Word Form Area of the left hemisphere.

In his PCFG, category nodes dominating stems and derivational affixes were treated as non-terminals, with the orthographic form of morphemes identified as the terminals.  So in parsing, say, “parseability,” at the second morphological form in the parse, -abil, the relevant rules would be one that expands an adjective into a verb and an adjective head ([_Adj[_v[parse]]Adj]) and an “emission” rule that expands the Adj node into -abil. At -ity, the emission rule that takes the non-terminal N node to -ity would not be sensitive to the presence of “-abil” under the Adj node.  From the perspective of a speaker’s knowledge of English, this is wrong – speakers know that “able/abil” potentiates nominalizing -ity.

If we were to improve the model, we probably would want to stick closely to the actual structure of Distributed Morphology, since the development of DM involves maximum consideration of all potentially relevant types of data – formal choices within the theory are motivated by empirical considerations.  In DM, the phonological/orthographic form of a morpheme is a vocabulary item. Vocabulary insertion is part of the Phonology, separate from the Merger operation that builds structures from morphemes.  So, we would like to model “emission” not as a context free operation, as might be appropriate for structure building, but as a context sensitive insertion rule (so, in early versions of transformational syntax, lexical insertion was a transformation, rather than a PS rule).  On this approach, the category nodes of the morphological tree are potentially terminal nodes – vocabulary insertion doesn’t create a (non-branching) tree structure but elaborates the terminal node into which it is inserted.  This matters for locality:  we want -ity to be inserted as a sister to -able, so -able must reside at the A node, not below it.

A context dependent vocabulary insertion rule seems formally equivalent to a treelet – N <-> -ity in the context of [_Adj able ] = [_N[_Aable] ity]  Or, rather, the rule looks like a fragment tree in a Fragment Grammar, since the -able can be the head of the adjective, sister to a verbal head – the relevant subtree is not a proper tree structure.  Which raises the question of the formal connection between PCFGs with contextual rules of vocabulary insertion and TAG and Fragment Grammar formalisms.

For the moment, I’m thinking about a couple quasi empirical questions.  First, if we’re considering continuations of, say, a verb “parse,” does the frequency in which adjectives are made from verbs overall in English really contribute to the processing of “parseable” over and above the transition probability from “parse” to “able”?  One could ask similar questions about sentential parsing, and perhaps people have – is the parsing of a continuation from a transitive verb to its direct object modulated by the probability of a transitive verb phrase in general in English, independent of the identity of the head verb?

Second is a theoretical question about the representation of derivational morphemes, with possible computational consequences.  It’s clear that many derivational affixes contribute meanings beyond those associated with bare category nodes.  So “-able” has a meaning beyond that of a simple adjective head.  A possibility being explored in the morpho-phonological literature is that (many) derivational heads include roots.  However, this possibility comes with the additional possibility that there’s a split between root-full derivational heads and derivational morphemes that are the pure spell out of category heads.  So, for example, maybe the little v that verbalizes many Latinate stems in English and is spelled out as -ate is a pure category head, without a root.  A further speculation is that such bare category heads might be classed with inflectional heads, and exhibit contextual allosemy – meaning that they could have null interpretations.  Jim Wood’s recent work on Icelandic (and English) complex event nominalizations suggests that nominalizers of verbs may have null interpretation – these nominalizers, then, would be candidates for the bare category head type of derivational suffix, while contentful nominalizers such as -er would always involve roots and would not show the null semantic contextual allosemy of the bare nominalizers.

Over the next several months, I’ll be posting thoughts as I prepare for, and teach, an advanced Intro to Morphology.  I’ve been teaching versions of this class for at least 25 years, and I believe that several of the readings I assign for the class these days were written to address questions raised in earlier incarnations of the class (former students:  you can confirm or deny this claim in the “comments”).  From my point of view, the field has made great progress in morphology since the early 1990’s.  In particular, what were relatively hazy theoretical questions 25 years ago are now sharply realized research programs.  See, for example, the progress on “root suppletion” since “”Cat” as a phrasal idiom.”  This theoretical progress has enabled a research program in behavioral and brain experiments that can (at long last, perhaps) feed back into theoretical thinking. However, there is sufficient empirical and theoretical work, both within and outside of Distributed Morphology, that is still insufficiently incorporated into general morphological theorizing to warrant a sustained group educational endeavor, like that of the Morphology classes at MIT in the early 1990’s.  To the great benefit of the field, clued-in morphologists have Distributed and spawned, so I’d like to expand the discussion digitally beyond a particular room in a particular place on a particular day of the week (just fyi, the “Introduction to Morphology…” class will meet on Mondays and Wednesdays at NYU this Fall).

I’ve been beginning my Morphology courses with a discussion of what a “morpheme” is and what “Morphology” is about.  Once one has a working definition of the “morpheme,” it’s straightforward to present Morphology as treating “allomorphy” (the phonological and/or orthographic realization of morphemes, and, these days, the semantic interpretation of morphemes as in contextual allosemy) and “morphotactics” (the positioning and ordering of morphemes in words and sentences). I won’t bother to maintain the fiction that I can/should present a “pretheoretical” definition of a “morpheme,” or that the job of a linguist is to discover what morphemes are.  Rather, I’ll state the definition of a “morpheme” within Distributed Morphology (and other realizational versions of a general Minimalist Program):  the minimal atom of Merge in the Syntax.  Within standard DM, morphemes are feature “bundles,” whose internal structure is NOT formed by Merge, and therefore, not explained directly by Syntax.  Within Nanosyntax, on the other hand, morphemes are privative features – any “bundling” or combination of features must be handled by the (Nano)syntax. The general theoretical consequences of asyntactic bundling vs. morphemes as features are of course a central topic for Morphology in the 21^stcentury.

DM does claim that the “syntactic atom” definition of “morpheme” allows a theory that covers the sort of data and generalizations that have traditionally been the bread and butter of Morphology.  That is, DM WANTS to draw connections with the structuralist tradition in linguistics, rather than declaring a clean break, e.g., via an announcement that “there are no morphemes” or “morphemes don’t exist.” I get the sense that some linguists, Jim Blevins in particular, think that DM is cheating here.  For these linguists, morphemes simply ARE the minimal (in a phonological sense) units of meaning in a language, and contemporary thinking (particular in the Word and Paradigm tradition) have shown that morphemes don’t exist (within an explanatory linguistics).  That is, in the classic trope, DM says morphemes are broccoli, and Blevins replies, “I say it’s spinach and I say the hell with it” https://en.wikipedia.org/wiki/I_say_it%27s_spinach  Needless to say, I don’t think fighting over whether DM’s morphemes are actually morphemes is going to particularly rewarding.  A question to the reader:  do you feel that linguists are confused by DM’s use of the term, “morpheme”? Just asking.

Theoretical progress depends on comparing alternative theoretical options.  What general alternatives to DM will figure in my Fall course?  First, there are theories that claim that syntax plays no role in word formation.  These range from “lexicalist” theories in which the syntax deals with words that are formed by processes independent of general syntactic operations to theories like Koopman’s in which the syntax only provides an ordering and hierarchy of morphemes and all phonological word formation is post and a-syntactic.  “Lexicalist” theories range from Kiparsky’s recent work in Stratal OT to Bruening’s anti-DM mss.  Second, there are neo-traditional theories like Kayne’s that deny “separation” – the morphemes in the syntax combine syntactic features, semantic import, and phonology. Third, there is Nanosyntax and related frameworks in which “vocabulary items” (units of phonological realization of morphemes) correspond to “spans” of morphemes that do not necessarily correspond to nodes in the syntactic tree.  Interestingly, the Kayne view and the Nanosyntax view both lead to syntactic structures with more morpheme positions than within standard DM, e.g., multiple nodes involved in English past tense inflection and English plural, for example.  While both theories make similar claims about the structures behind stem suppletion, they are diametrically opposed on the nature of the phonological forms that realize structure.  For Nanosyntax, allomorphy generally involves vocabulary items of different sizes – spelling out different amounts of hierarchical structure – while for Kayne similar patterns of allomorphy involve the distribution of zero morphemes, of which there are many. Although DM will claim that both these approaches are wrong, the arguments for more morpheme positions in the explanation of stem allomorphy may hold within DM as well, although with more Kayne’ian phonological zeros.

Welcome to the MorphLab blog. Check back soon for blog content written by Dr. Alec Marantz.