Tal Linzen, Johns Hopkins University, will present “What inductive biases enable human-like syntactic generalization?” in the Linguistics zolloquium series at 3:30 Friday April 17. An abstract follows. All are welcome! The Zoom link has already beensent out on department mailing lists. If you did not receive it and would like attend, please email Brian Dillon for the link.
Abstract
Humans apply their knowledge of syntax in a systematic way to constructions that are rare or absent in their linguistic input. This observation, traditionally discussed under the banner of the poverty of the stimulus, has motivated the assumption that humans are innately endowed with inductive biases that make crucial reference to syntactic structure. The recent applied success of deep learning systems that are not designed on the basis of such biases may appear to call this assumption into question; in practice, however, such engineering success speaks to this question in an indirect way at best, as engineering benchmarks do not test whether the system in fact generalizes as humans do. In this talk, I will use established psycholinguistic paradigms to examine the syntactic generalization capabilities of contemporary neural network architectures. Focusing on the classic cases of English subject-verb agreement and auxiliary fronting in English question formation, I will demonstrate how neural networks with and without explicit syntactic structure can be used to test for the necessity and sufficiency of structural inductive biases, and will present experiments indicating that human-like generalization requires stronger inductive biases than those expressed in standard neural network architectures.
From Sarah Gibbons, of the College of Humanities and Fine Arts:
Set to take place on the University of Massachusetts Amherst campus from March 19 through 21, the 33rd Annual CUNY Human Sentence Processing Conference seemed doomed in the face of mounting concerns and decisions to cancel events and transition to remote working and learning in response to the spread of COVID-19. Determined to carry forward with decades of interdisciplinary scholarly tradition and connection, UMass Amherst community members worked fast beginning on March 8 to transition the conference, also known as CUNY 2020, completely online.
Associate Professor Brian Dillon (top left) and other participants of CUNY 2020.
Led by Brian Dillon, associate professor in the UMass Amherst Department of Linguistics, a 24-person team of UMass and Five College faculty members, graduate students, and CUNY community members from other institutions worked quickly to cancel on-site arrangements, communicate with attendees and speakers, organize video and webinar technology, and move the conference online in ten days. No small feat, as the team had been planning the physical conference since the summer of 2018. “Our decision to cancel had the full support of the Provost and event services, and this support was critical in allowing to cancel as early as we did,” said Dillon.
Dillon, who hasn’t missed a CUNY conference since 2006, emphasized that this feat was “very definitely” a team effort. “One person whose work should be especially recognized is Jon Burnsky. Jon is a PhD student in Psychological and Brain Sciences here, and he is our official CUNY Graduate RA. He did a lot of work to set up live-streaming for the physical conference; his work laid the groundwork for our virtual conference. We were able to virtualize so quickly because of the work he had put in ahead of time,” said Dillon.
The online format added an unexpected tone of lightheartedness and relief to many experiencing serious disruption in their lives. “I was surprised at how connected this allowed us to feel,” recalls Dillon. “This allowed us to feel like we were making connections and giving a sense of community and normalcy in difficult times.” “It was also just plain fun in ways I hadn’t expected,” he said, feeling at times like he and the co-organizers were “thrown into the role of radio show host, rather than organizers of an academic conference. We would banter, crack jokes, take questions from the audience during downtime. It allowed for a sense of levity that I think some participants appreciated at this moment.”
People from 44 countries participated in CUNY 2020. “The top four were the USA, Germany, the UK, and China,” said Dillon. “The large Chinese participation is notable because of the 12-hour time difference. Chinese colleagues reported staying up late into the night to catch talks.”
One conference attendee and session chair, Colin Phillips, Professor of Linguistics at the University of Maryland and Director of the Maryland Language Science Center, called Dillon’s leadership of CUNY 2020, “an incredible success.”
“The conference normally draws 200-300 in-person attendees,” said Phillips. “There was a broad concern that few people would want to participate in the remote format. In contrast, over 1,000 people took part over the course of the 3 days, with peak participation of 300-350 simultaneous participants in some sessions.” According to Phillips, about 20 people at the University of Maryland participated in a virtual watch party and communicated through the remote workplace communication tool, Slack. Because of this, Phillips explained, “the conference even helped us feel more connected locally. We heard of various other similar activities at other institutions.”
Reflecting on the unique and challenging circumstances facing universities during the COVID-19 crisis, Phillips went on to say, “The past couple of weeks have seen unprecedented changes in how universities operate. It seems like everything is changing from day to day. Brian and the UMass team made people feel that they were connected, despite the upheaval. But they also made people feel like they were witnessing something special that they hadn’t seen before. This conference has built up a strong following in its 33 years. Many people, myself included, have attended almost all of them. At this point I think there’s little doubt that this was the most memorable of them all.”
The CUNY conference, which was originally founded at the City University of New York Graduate Center by Janet Dean Fodor, is a major event in the psycholinguistics subfield of linguistics (it is now primarily hosted by other institutions, but retains the name). It is highly interdisciplinary, with strong contributions from researchers in Linguistics, Psychology, Neuroscience, Cognitive Science, Computer Science, Education, and Philosophy. “It is one of the central North American venues for researchers interested in psycholinguistics,” explains Dillon, “it is well attended by both linguists and by psychologists, and creates lots of productive interdisciplinary cross-talk.”
“UMass Amherst is a special place to host CUNY for many reasons,” said Dillon. “First, we have a very strong tradition in psycholinguistics. Our community goes back to the late 70’s and early 80’s, lead by researchers in psychology (Chuck Clifton Jr., Keith Rayner, Alexander Pollatsek, and Jerry Myers), and in linguistics (Lyn Frazier, Tom Roeper).” The UMass Amherst linguistics department, ranked No. 2 worldwide, is known for having a strong psycholinguistics program, and, “hosting CUNY further increases our community’s visibility,” said Dillon.
The UMass Amherst community has been involved with the CUNY conference since the beginning, but this is the first year that the campus has hosted since 1993. “Back then,” remarks Dillon, “incidentally, our conference coincided with a famous blizzard that stranded many people and prevented them from coming to the conference.”
Huge congratulations are due to both Sakshi and Ivy, who arrived to UMass in the same PhD cohort. So it’s a special joy to get to make this double announcement. We’re proud of you both: Best of luck in the next phase of your careers!
Annika Andersson, Associate Professor at Linnaeus University, will present “Cross-linguistic influence on processing of fine-grained placement verb semantics as recorded by ERPs and appropriateness ratings” and talk about some of her research with second language learners on Wednesday March 18th 12:00 in ILC N400. An Abstract follows.
Abstract
Second language (L2) learners typically experience challenges when semantics differ across source and target languages, and often display CLI in speech production and behavioral comprehension studies (e.g., Jarvis & Pavlenko, 2008). However, in studies using ERPs, CLI has rarely been reported, probably because these studies typically examine the processing of gross semantic violations (e.g., Kutas & Hillyard, 1980). We explored how English and German learners of Swedish process fine-grained L2 verb semantics that are either shared or not shared with their first language. Three Swedish placement verbs (sätta ‘set’, ställa ‘stand’, lägga ‘lay’), obligatory for describing placement on a surface with support from below (Viberg 1998) were examined. In difference to Swedish, English has one general placement verb (put), whereas German has specific verbs similar to Swedish (Narasimhan et al., 2012). In contrast to previous ERP studies of semantic processing qualitative differences in semantic processing were related to non-native processing. However, more interestingly neurophysiological processing of fine-grained semantics was strongly related to CLI both offline and online.
Matthew Tyler, Yale University, will present “Internal arguments disguised as external arguments: Lessons from an active alignment system” in the Linguistics colloquium series at 3:30 Friday March 6. An abstract follows. All are welcome!
Abstract
Active alignment describes a morphological alignment pattern where the lone argument of an intransitive verb is marked sometimes like the subject of a transitive verb, and sometimes like the object. Many generative accounts of active alignment hold that this morphological distinction is rooted in the syntactic distinction between external arguments, merged as the specifier of a functional head Voice (or v), and internal arguments, merged as an argument of the lexical verb. However, on the basis of novel fieldwork with Choctaw, a language with an active agreement system, I show that an argument’s morphological marking must be dissociated from its syntactic position: the marking that is characteristic of canonical external arguments is, exceptionally, found with certain internal arguments too. Nevertheless, I show that these internal arguments receive their exceptional marking only if they can form an uninterrupted syntactic (i.e. Agree/case-assignment) relation with the Voice head.
The implications of these findings are twofold. Firstly, active alignment is argued to be a consequence of Voice forming a syntactic relation with some arguments and not with others, rather than a direct consequence of the differing syntactic positions of internal vs. external arguments. This provides a new way of understanding lexical and configurational exceptions to the dominant alignment pattern of a language. Secondly, by studying the particular circumstances under which internal arguments receive exceptional marking, I argue that the agreement/case-assignment properties of a single Voice head can vary contextually according to the syntactic material in its immediate neighborhood, including the lexical root and other functional heads. This brings the agreement/case-assignment properties of functional heads in line with how we often think about their morphological properties: that is, they can have default and contextually-conditioned variants.
Michael Yoshitaka Erlewine (mitcho), National University of Singapore, will present “Bikol clefts and topics and the Austronesian extraction restriction” (joint work with Cheryl Lim) in the Linguistics colloquium series at 3:30 Friday February 28. An abstract follows. All are welcome!
Abstract
Many Austronesian languages exhibit an extraction restriction whereby only one particular DP — the “pivot” argument, the choice of which is reflected by morphology on the verb — can be A’-extracted. We show that such extraction restrictions can vary between different A’-constructions in Bikol: local clefting is limited to the pivot, whereas topicalization can target pivots and non-pivot agents, but not other non-pivot DPs. Following the phase-theoretic, locality-based approach to such extraction asymmetries in related Austronesian languages, we propose that clefting and topicalization differ in the featural specifications of their probes, but must always attract their closest matching goal. Evidence for this approach comes from interactions between clefting, topicalization, and hanging topic left dislocation in long-distance configurations. Such data motivates the view that the classic Austronesian pivot-only extraction restriction is best characterized in terms of syntactic locality, rather than as a restriction on the grammatical function or morphological case of movement targets.
Faruk Akkuş, University of Pennsylvania, will present Lessons from “make” causatives in Sason Arabic in the Linguistics colloquium series at 3:30 Friday February 21. An abstract follows. All are welcome!
Abstract This talk investigates the syntax of an indirect causative construction embedded under the verb “make” in Sason Arabic, with a focus on the syntax of the embedded structure, and the syntactic and semantic status of the implicit embedded agent. I demonstrate that this construction embeds both an active and passive VoiceP despite the absence of any morphological reflex. I also contend that the implicit agent in the active complement of “make” may be introduced (i) as a full DP in Spec,VoiceP, being subject to Romance ECM-type restrictions, and thus providing striking evidence of A’-movement feeding licensing relationships, or (ii) as a free variable à la Heim (1982) generated on the Voice head itself. The latter possibility raises implications regarding licensing, suggesting that licensing of a thematic object is dissociated from the projection of a specifier
Sandhya Sundaresan, Universität Leipzig, will present “Indexical Shift and the grammar of intensionality” in the Linguistics colloquium series at 3:30 Friday February 14. An abstract follows. All are welcome!
Abstract
In classic cases of indexical shift, attested in languages like Amharic, Zazaki, Nez Perce, Turkish, and many others, a sentence like “Jill said that I am sick”, uttered by Marie, can actually be a statement about Jill’s sickness rather than Marie’s. I.e. the reference of the indexical pronoun ‘I’ is context-shifted, such that it doesn’t refer to Marie (speaker of the utterancecontext) but to Jill (speaker of the context associated with the matrix speech event).
In this talk, I will present two types of evidence that show that the landscape of indexical shift is far more nuanced than is typically assumed:
(i) exceptions to Shift Together (the restriction that all shiftable indexicals in a local domain must shift together) in Tamil, Korean and potentially other languages including Late Egyptian, which crucially co-exist with Shift Together holding as a robust restriction in many languages; and
(ii) evidence, from dialectal microvariation in Tamil (based on personal fieldwork) and crosslinguistic variation from 28 languages, which shows not only that indexical shift is subject to considerable selectional variation, but also that such variation is implicationally structured, privileging speech predicates over all others (making indexical shift an embedded root phenomenon).
I show that current theories of indexical shift cannot handle these challenges and develop a new syntax and semantics of this phenomenon, which does. This new theory derives indexical shift without overwriting the utterance-context but also makes such shift sensitive to Relativized Minimality; it also redefines the contextual operator or “monster” that effects shift as a special type of intensional complementizer, merged at different heights along the clausal spine, reflecting differences in the nature of the embedded attitude. The new model also makes several testable empirical predictions which are fulfilled:
a. that indexical shift cannot obtain in structures that lack a complementizer;
b. that it can obtain in the absence of attitude verbs; and
c. that it interacts with other embedded root phenomena (like allocutivity).
Through it all, we will see that the new theory also has the welcome consequence that it demystifies indexical shift: this is no longer an esoteric phenomenon that applies to “exotic” languages. Rather, all languages are indexically shifting in some way, with variation simply being relegated to which contextual coordinates are shifted, and which indexicals are shiftable, in a given structure or language. Indexical shift can thus be fruitfully deployed as an empirical lens to diagnose the structures involved in intensionality and finiteness across dialects and languages.
Bronwyn Bjorkman, Queen’s University, will present “Realizing Syntax” in the Linguistics colloquium series at 4:00pm Tuesday Feb. 11. An abstract follows. All are welcome!
Abstract
This talk looks at interactions among linearization, prosody, and vocabulary insertion, focusing on cases of verb doubling that appear to be motivated not by syntactic movement, but by the need for an otherwise-unsupported clitic to have a host.
Drawing on examples of verb doubling in Ingush (Nakh-Dagestanian) and Breton (Celtic), I argue first that the linearization of syntactic structures is accomplished via the interaction of ranked and violable constraints, as in OT, rather than via a deterministic linearization algorithm of the type often assumed in syntax. Second, I argue that linearization and prosodification proceed in parallel, allowing verb doubling as a trade-off between prosodic well-formedness (the need of a clitic for a host) and optimal linearization—but that this evaluation occurs prior to both Vocabulary Insertion and the subsequent competition of segmental phonology.
The final sections of the talk discuss the implications of this model for doubling more generally, and more particularly for our ability to explain the fact that certain movement configurations appear to lead to doubling in some languages but not in others. I discuss verb doubling in predicate focus, clitic doubling, and several other instances of apparent multiple realization.
Simon Charlow, Rutgers University, will present “Givenness and local contexts” in the Linguistics colloquium series at 3:30 Friday December 6. An abstract follows. All are welcome!
Abstract
Givenness plays a central role in the licensing of anaphoric reduction phenomena like deaccenting and ellipsis. I argue that Givenness is assessed compositionally, in a local context. In other words, I give evidence for anaphoric Givenness operators in syntax, and I argue that such operators are sensitive to local manipulations of the context (specifically, the assignment). This gives a theory of the syntax-semantics of Givenness intermediate in a sense between the systems of Schwarzschild and Rooth, and converges with conclusions reached in recent work by Kratzer & Selkirk.
I show how this approach leads to significant simplifications in the theory of reduction licensing, allowing us to dispense with otherwise necessary stipulations (e.g., Heim’s prohibition of “Meaningless Coindexing” or Sag’s non-standard definition of alphabetic variance), and making it feasible at last to treat the relationship between an elided phrase and its antecedent as one of strong semantic identity.
I consider several consequences of my proposal for the formulation of Givenness operators, arguing that it compels us to take their dynamic, anaphoric character seriously. I argue that Givenness domains are maximized, and against the notion that any non-F-marked node must be Given. And, time permitting, I explore some consequences of these moves for restrictions on antecedent-contained deletion and the puzzling phenomenon of focused bound pronouns.
