Whamit!

Title: Discussion of Romero & Han 2004: On Negative Yes/No Questions
Date and time: Monday November 13, 1-2pm
Location: 7th Floor Seminar room
Abstract:

Preposed negation yes/no (yn)-questions like Doesn’t John drink? necessarily carry the implicature that the speaker thinks John drinks, whereas non-preposed negation yn-questions like Does John not drink? do not necessarily trigger this implicature. Furthermore, preposed negation yn-questions have a reading “double-checking” p and a reading “double-checking” ≠ p, as in Isn’t Jane coming too? and in Isn’t Jane coming either? respectively. We present other yn-questions that raise parallel implicatures and argue that, in all the cases, the presence of an epistemic conversational operator VERUM derives the existence and content of the implicature as well as the p/≠ p-ambiguity.

The discussion will be led by Milica.

Speaker: Erin Olson (MIT)
Title: Influence of schwa duration on stress/pitch patterns in Passamaquoddy
Date/Time: Monday, 13 November, 5:00-6:30pm
Location: 32-D831
Abstract:

Passamaquoddy-Maliseet (Eastern Algonquian; Maine and New Brunswick) is well-known for the fact that reduced vowels are in many cases invisible to the stress/pitch accent system (LeSourd 1988, 1993; Hagstrom 1995). While previous analyses have assumed that this invisibility is due to some sort of structural deficiency of these vowels, I assume that it is at least in part due to the phonetics of stress/pitch accent. I will present a preliminary phonetic study of pitch and vowel duration in the language, based on data from the Passamaquoddy-Maliseet Dictionary Project (Language Keepers & Passamaquoddy-Maliseet Dictionary Project 2016). I will show that the syllables marked as stressed in the literature crucially involve a rise in pitch, and I hypothesize that the odd behaviour of schwa can be derived from the fact that it is too short to host either this rise in pitch or a fall from one pitch to the next. A brief sketch of this hypothesis in OT is provided.

Speaker: Michelle Yuan (MIT)
Title: Last Resort licensing in Inuktitut
Date and time: Tuesday November 14, 1-2pm
Location: 32-D461
Abstract: 

The traditional Case Filter (Chomsky 1981, et seq.) regulates the syntactic distribution of nominal arguments; nominals cannot appear in positions where they are inaccessible for licensing (e.g. Case assignment). It has also been argued that nominals in such positions may nonetheless be later licensed by a Last Resort mechanism that rescues the derivation. In much of this literature, Last Resort Case assignment is modeled as the countercyclic insertion of a Case-bearing head directly onto the nominal in question (e.g. Harley & Noyer 1998, Rezac 2011, Levin 2015, van Urk 2015).

I argue for the existence of this kind of Last Resort Case assignment in Inuktitut (Eskimo-Aleut), thus providing evidence in favour of this idea more generally. Along the way, I also show how this analysis sheds light on various other aspects of Inuktitut morphosyntax. In Inuktitut, regular nominal licensing strategies may be blocked or disrupted in certain environments. When this happens, these nominals uniformly receive a certain kind of oblique case. Crucially, this oblique case morphology must be countercyclically inserted once the clause structure has been fully built. In Inuktitut, this derivational logic is particularly transparent: phi-agreement, the language’s regular nominal licensing strategy, is in C0, meaning that failure to license a lower nominal via phi-agreement can only be seen once the CP layer is Merged. Finally, I argue that, despite surface similarities, this Last Resort Case-insertion process is distinct from default case (e.g. Schutze 2001).

Michelle Sheehan (Anglia Ruskin University) will be visiting the department this week. In addition to her Colloquium talk on Friday, she will be offering a mini-course on the evidence for and against an ‘escape hatch’ for successive-cyclic A-movement. Details below:

Speaker: Michelle Sheehan (Anglia Ruskin University)
Title: No escape hatch for A-movement: evidence and issues
Time: Wednesday, November 15th, 1:00pm-2:30pm and Thursday, November 16th, 4:00-5:30 pm
Place: 32-D461 (Wed), TBD (Thurs)
Abstract:

This mini-course explores the proposal that A-movement does not proceed through the phase edge. We begin by looking at simple clause-internal A-movement and note that, even if we accept Legate’s (2003) evidence that passive vP is a phase, if we also adopt Chomksy’s (2001) PIC2, there is no need for A-movement to proceed through the phase edge in such contexts. Turning to (more insightful) interclausal contexts, we examine places where A-movement is possible vs. contexts where it is blocked (notably with causative/perception ECM verbs) and note that many of these contrasts can be attributed to the lack of an escape hatch in the v- or C-related phase. For example: 

1. 1. *Kim_i was made/had/let seen/heard/witnessed/listened to [t_i sing]
   2. Kim_i was made/seen/heard [ t_i to sing].
   3. Kim was seen/heard/witnessed/listened to [ t_i singing].
   4. Sam_i was made [ t_i angry] by the news.

 Using independent diagnostics for the size of these ECM complements, we propose that A-movement is blocked in (1a) because it must cross two phase heads. A-movement is permitted in (1b-c) because the complement is larger and so a T-related EPP feeds successive cyclic A-movement. Finally, in (1d) the small clause complement is too small to be phasal. The same effect is observed in instances of raising, and possibly (some instances of) control. Where a phasal complement is embedded and there is no T-projection to feed A-movement before the next phase head is merged, the result is ungrammaticality. Once we have shown that this approach works pretty well for English-type languages, we will address a number of elephants in the room: (i) passivisation of causatives/perception verbs in other languages (the variation problem); (ii) languages which appear to lack an English-style EPP to feed successive cyclic A-movement; (iii) hyperraising and (iv) apparent interactions between A- and A-bar movement that appear problematic for our central claim (including Holmberg, Sheehan and van der Wal 2016, oh dear). We’ll see that many languages have restrictions on the passivisation of causatives and/or perception verbs (Italian, French, Spanish, European and Brazilian Portuguese, Danish, Swedish, Dutch, German, Icelandic, Hungarian), but the patterns are complex and there is a lot of variation both within and across languages. Some languages are also notable in permitting passives of causatives, e.g., Japanese (with complications), Zulu and Sotho. We’ll see how the approach can handle this variation, making testable predictions about the size of complementation in each language. There will be many puzzles left open for participants to solve and they will be encouraged to think about these issues in other languages that they know or work on. 

Speaker: Tanya Bondarenko (MIT)
Title: Results, Repetitives and Datives: towards an account of the crosslinguistic variation
Date and time: Wednesday November 15, 4-5pm
Location: 32-D461
Abstract:

Repetitive adverbs like ‘again’ are ambiguous in sentences with dative (= indirect) arguments in some languages (ex., in English, (1)), but not in others (ex., in Russian, (2)):

(1) Thilo gave Satoshi the map again.
REP(etitive): ‘Thilo gave Satoshi the map again, and that had happened before.’
RES(titutive): ‘Thilo gave Satoshi the map, and Satoshi had had the map before.’
(Beck & Johnson 2004: 113)

(2) Maša opjat’ otdala Vase kartu.
Masha.NOM again gave Vasja.DAT map.ACC
REP(etitive): ‘Masha gave Vasja the map, and that had happened before.’
RES(titutive): *’Masha gave Vasja the map, and Vasja had had the map before.’

Following the structural ambiguity approach to repetitives ((von Stechow 1996), (Beck 2005), among others), I will argue that the availability of the restitutive reading in constructions with dative arguments depends on the syntactic configuration. The restitutive reading in ditransitives has been previously attributed either to presence of a small clause in the structure (Beck & Johnson 2004), or on the presence of an applicative phrase (Bruening 2010). I will explore another hypothesis: that the restitutive reading arises due to existence of a prepositional phrase in the structure.

Speaker: Tom Roeper (UMass-Amherst)
Title: Strict Syntax/Semantic Interfaces and Ellipsis: an explanatory role for acquisition theory
Date and time: Thursday November 16, 12:30-1:50pm
Location: 32-D461
Abstract: 

It has often been said that linguistic theory should account for acquisition. Modern work extends this claim to the more important goal: account for steps on the acquisition path itself, where micro-operations are often visible. The abstract claim has particular relevance when it can resolve long-standing questions in theory. We show that the acquisition path for ellipsis must entail a shift from an Open Interface, dependent upon inference, to a Strict Syntax/semantic Interface which involves a perfect correspondence of syntax and semantics and minimizes pragmatics:

Strict Interface Hypothesis: The acquisition device assumes that there should be an isomorphic connection between LF-semantics, syntax, and phonology.

This is based upon an acquisition principle:

Minimization Goal: Minimize pragmatic inference and maximize the information determined by explicit grammar.

One tradition, epitomized in the work of Hardt (1999), claims that a broad empty category like pro and general cognitive biases toward parallelism (Jackendoff and Culicover (2005)) are needed to account for ellipsis. Another tradition is captured in Kennedy’s (2008) Syntax-Semantics Identity Hypothesis:

A recovered syntactic representation generates an LF form with variables that allow sloppy readings.

Merchant (2015, to appear) and Craenenburg (2014) argue for a “hybrid” account which is theoretically unsatisfying. They claim the system is strict, but then the system assumes pronominal forms (or just do that) when necessary. The literature has also shown an ongoing debate about less than full reconstruction in syntactic representations, given various deviations known as “mismatches”. The acquisition path perspective makes this outcome natural.

The empirical basis of this argument works with evidence about VP ellipsis, NP-ellipsis, and Partitive interpretation. Children overgenerate possible reconstructions beyond sloppy identity and then must somehow retreat. eg. They allow John patted his dog and so did his grandfather to mean grandfather pats an uncle’s dog. This creates a learnability problem since interpretations both include and exceed the adult grammar. Evidence on early VP-ellipsis from Santos, and Wijnen and Roeper (2003) on NP-ellipsis and Hulk and Schleeman (2014) on partitives, and Perez et al (to appear) on empty pro objects support this account.

The acquisition path claim is: Children will step by step, following Minimize Pragmatics, replace a pronominal representation with an explicit syntactic representation and a fixed link to semantic representations.

This Strict Interface perspective is deliberately much more stringent than a claim that interfaces involve “Third Factor” effects (Chomsky (2005) which weakens learnability. A big challenge is to develop a notation that makes the connection transparent and simple for the child and natural within UG.

Speaker: David Alvarez Melis (MIT CSAIL)
Title: Interpretability for black-box sequence-to-sequence models
Date/Time: Thursday, November 16th, 5:00-6:30pm
Location: 46-5165
Abstract:

Most current state-of-the-art models for sequence-to-sequence NLP tasks have complex architectures and millions —if not billions—of parameters, making them practically black-box systems. Such lack of transparency can limit their applicability to certain domains and can hamper our ability to diagnose and correct their flaws. Popular black-box interpretability approaches are inapplicable to this context since they assume scalar (or categorial) outputs. In this work, we propose a model to interpret the predictions of any black-box structured input-structured output model around a specific input-output pair. Our method returns an “explanation” consisting of groups of input-output tokens that are causally related. These dependencies are inferred by querying the black-box model with perturbed inputs, generating a graph over tokens from the responses, and solving a partitioning problem to select the most relevant components. We focus the general approach on sequence-to-sequence problems, adopting a variational autoencoder to yield meaningful input perturbations. We test our method across several NLP sequence generation tasks.

On Friday, Donca Steriade was honored as a Professor Honoris Causa by the Faculty of Letters at the University of Bucharest. Congratulations Donca!