QP Fest 2016: The Thirteenth Annual QP Fest

1:00pm-1:30pm Ignacio Cases
"On the Effective Use of Pretraining for Natural Language Inference"

Abstract: Neural networks have excelled at many NLP tasks, but there remain open questions about the performance of pretrained distributed word representations and their interaction with weight initialization and other hyperparameters. We address these questions empirically using attention-based sequence-to-sequence models for natural language inference (NLI). Specifically, we compare three types of embeddings: random, pretrained (GloVe, word2vec), and retrofitted (pretrained plus WordNet information). We show that pretrained embeddings outperform both random and retrofitted ones in a large NLI corpus. Further experiments on simpler word-level entailment shed light on the contexts for which retrofitted embeddings can be useful. We also explore two principled approaches to initializing the rest of the model parameters, Gaussian and orthogonal, showing that the latter yields gains of up to 2.9% in the NLI task.

1:30pm-2:00pm Matt Lamm
"Detecting Syntactic Parallelism Using Sequence Alignment"

2:00pm-2:30pm Rob Voigt
"Prosody, Gesture, and the Joint Construction of Social Meaning"

Abstract: Theories of embodiment propose – and experimental results in linguistic research demonstrate – a powerful and necessary connection between acoustic prosody and bodily movements, particularly as relates to the communication of emotive and social meanings. In this qualifying paper I present a pipeline for fully automatic acoustic and visual prosodic analysis of spoken monologues and provide evidence in a large, naturalistic dataset for joint variation between several key body movements and acoustic prosodic variables. I examine individual cases of speakers using head tilt in particular to show the diverse range of social meanings in which these variables can participate, consider preliminary analyses of the role of gender and genre, and offer a theoretical proposal for the crucial role of joint, compositional variation in the production of socially meaningful expressive units in talk.

2:30pm-2:45pm Coffee break

2:45pm-3:15pm Lelia Glass
"Grounding the Social Meaning of a Strong Necessity Modal in its Semantics"

Abstract: Need tois less ambiguous than have toabout the source of the reported obligation; need to is said to require the obligation to stem from someone’s priorities or internal needs, whereas have tocan tie the obligation to any contextually plausible source (Rubinstein 2012). This paper investigates the social reasons that a speaker might choose or avoid the less ambiguous form. In view of the semantics of need to, the speaker who utters you need tounambiguously acts as if she is familiar with the hearer’s priorities and licensed to tell him what is good for him – a socially risky move. I therefore predict that you need towill be more appropriate and thus more common from people with knowledge about the relevant domain, people in authority over the hearer, and people who play a mentoring role in the hearer’s life because these people are more likely to be licensed to tell the hearer what is good for him in the context. I find evidence consistent with these predictions in corpora and experiments.

3:15pm-3:45pm Simon Todd
"Something from Nothing: Pragmatic Parsing of Partitive Possessives"

Abstract: Sentence structure is often ambiguous. Models of human parsing typically assume that such ambiguity is resolved through cues that are present in the signal. Here, I present experimental evidence that the absence of an overt cue can be meaningful for resolving parsing ambiguities, based on a freeform selection task with partitive possessives like "two of the boys' cars", in which the syntactic structure conditions a morphophonological alternation (Nevins, 2011). This result is parallel to the use of negative evidence evoked by comparison to available alternatives in tasks like reference (Frank and Goodman, 2012; Stiller et al., 2015), which is motivated by pragmatic principles (Grice, 1975; Levinson, 2000), and I argue that it cannot readily be explained by a parser which lacks the capacity to incorporate such principles.

3:45pm-4:00pm Coffee break

4:00pm-4:30pm Daniel Galbraith
"Dative Subjects and Case-Marking in Faroese: A Linking Theory Account"

Abstract: At least since the seminal paper of Zaenen, Maling and Þráinsson (1985) on Icelandic, non-nominative subjects have been a consistent focus of debate in the generative literature, and yet much remains to be understood. A closely related language, Faroese, has a lot to offer the case theorist. The puzzle presented by Faroese is that unlike Icelandic, in the presence of a non-nominative subject, the lower object argument will be marked with accusative case, not nominative. The Faroese data pose problems for all major theories of case as they stand, since the co-occurrence of dative-marked true subjects with structural accusative object case has not previously been predicted by theorists. In this paper, I will argue that the Faroese dative-accusative pattern can be accounted for by an agreement requirement that forces Icelandic to mark objects nominative in the relevant predicates, which in Faroese is trumped by a constraint enforcing canonical object case. I formulate my analysis within a linking theory of case (Kiparsky 1997 et seq.), which I implement in Optimality Theory (Prince and Smolensky 1993): this offers a ready-made apparatus for capturing constraint interaction, allowing the difference between Icelandic and Faroese to be captured by reranking a single pair of constraints. I argue that the diachrony of ergative case-marking in Indic languages (Deo and Sharma 2006) is somewhat analogous to the Faroese scenario, and that Kiparsky’s theory provides a basis for explaining both the synchronic facts and the diachronic trajectory observed. This demonstrates that far from being a bizarre outlier, the Faroese pattern is in fact the outcome of constraint interactions of a typical kind.

4:30pm-5:00pm Kate Lindsey
"A Syntactic Account for Bare Nominals in Turkish"

Abstract: The status of Turkish bare arguments is an interesting puzzle for the study of case and argument incorporation. In this talk, I will show that objects and even subjects can appear without morphological case immediately before the verb. I will argue that the empirical facts describe two distinct structures. The most productive construction allows nonspecific phrasal objects to remain in first merge position without raising to get case. The more restricted construction allows subjects or objects to remain in the verb phrase provided that they form a lexically recognized N+V combination. Both structures license their arguments via surface adjacency. Finally, I will suggest what my analysis would predict for the prosody and semantics of these constructions.