Christopher Potts

A large annotated corpus for learning natural language inference

Understanding entailment and contradiction is fundamental to understanding natural language, and inference about entailment and contradiction is a valuable testing ground for the development of semantic representations. However, machine learning research in this area has been dramatically limited by the lack of large-scale resources. To address this, we introduce the Stanford Natural Language Inference corpus, a new, freely available collection of labeled sentence pairs, written by humans doing a novel grounded task based on image captioning. At 570K pairs, it is two orders of magnitude larger than all other resources of its type. This increase in scale allows lexicalized classifiers to outperform some sophisticated existing entailment models, and it allows a neural network-based model to perform competitively on natural language inference benchmarks for the first time.

The expressive dimension

Abstract Expressives like damn and bastard have, when uttered, an immediate and powerful impact on the context. They are performative, often destructively so. They are revealing of the perspective from which the utterance is made, and they can have a dramatic impact on how current and future utterances are perceived. This, despite the fact that speakers are invariably hard-pressed to articulate what they mean. I develop a general theory of these volatile, indispensable meanings. The theory is built around a class of expressive indices. These determine the expressive setting of the context of interpretation. Expressive morphemes act on that context, actively changing its expressive setting. The theory is multidimensional in the sense that descriptives and expressives are fundamentally different but receive a unified logical treatment.

No country for old members: user lifecycle and linguistic change in online communities

Vibrant online communities are in constant flux. As members join and depart, the interactional norms evolve, stimulating further changes to the membership and its social dynamics. Linguistic change --- in the sense of innovation that becomes accepted as the norm --- is essential to this dynamic process: it both facilitates individual expression and fosters the emergence of a collective identity. We propose a framework for tracking linguistic change as it happens and for understanding how specific users react to these evolving norms. By applying this framework to two large online communities we show that users follow a determined two-stage lifecycle with respect to their susceptibility to linguistic change: a linguistically innovative learning phase in which users adopt the language of the community followed by a conservative phase in which users stop changing and the evolving community norms pass them by. Building on this observation, we show how this framework can be used to detect, early in a users career, how long she will stay active in the community. Thus, this work has practical significance for those who design and maintain online communities. It also yields new theoretical insights into the evolution of linguistic norms and the complex interplay between community-level and individual-level linguistic change.

The syntax and semantics of As-parentheticals

This paper is a detailed investigation of the syntax and semantics of a single type of cross-linguistically commonparenthetical expression, here dubbed As-parentheticals (e.g., Ames, as you know, was aspy). I show that a treatment of such clauses as adverbial modifiers combines with a motivated semantic analysisto account for a wide range of ambiguities concerning negation in particular, but also tense, modal, and adverbialoperators. I provide a principled explanation for the impossibility of variable binding into, and extractionfrom, As-parentheticals, and argue that this construction yields novel support forthe view (of Ladusaw 1992, and others) that negative DPs like no one are actually non-negated indefinites licensed by anabstract, clause-level negation. Overall, the analysis shows that parentheticals, in addition to being a richsource of puzzles in their own right, provide a useful probe into clause structure in general.

Harmonic grammar with linear programming: From linear systems to linguistic typology

Harmonic Grammar (HG) is a model of linguistic constraint interac- tion in which well-formedness is calculated in terms of the sum of weighted constraint violations. We show how linear programming algorithms can be used to determine whether there is a weighting for a set of constraints that fits a set of linguistic data. The associated software package OT-Help provides a practical tool for studying large and complex linguistic systems in the HG framework and comparing the results with those of OT. We first describe the translation from harmonic grammars to systems solvable by linear programming algorithms. We then develop an HG analysis of ATR harmony in Lango that is, we argue, superior to the existing OT and rule-based treatments. We further highlight the usefulness of OT-Help, and the analytic power of HG, with a set of studies of the predictions HG makes for phonological typology.

Did it happen? the pragmatic complexity of veridicality assessment

Natural language understanding depends heavily on assessing veridicality—whether events mentioned in a text are viewed as happening or not—but little consideration is given to this property in current relation and event extraction systems. Furthermore, the work that has been done has generally assumed that veridicality can be captured by lexical semantic properties whereas we show that context and world knowledge play a significant role in shaping veridicality. We extend the FactBank corpus, which contains semantically driven veridicality annotations, with pragmatically informed ones. Our annotations are more complex than the lexical assumption predicts but systematic enough to be included in computational work on textual understanding. They also indicate that veridicality judgments are not always categorical, and should therefore be modeled as distributions. We build a classifier to automatically assign event veridicality distributions based on our new annotations. The classifier relies not only on lexical features like hedges or negations, but also on structural features and approximations of world knowledge, thereby providing a nuanced picture of the diverse factors that shape veridicality.“All I know is what I read in the papers”—Will Rogers

The Lexical Semantics of Parenthical‐as and Appositive‐which

Despite their superficial similarities, nonrestrictive relatives and as-parentheticals show contrasting behavior in a range of apparently disparate domains, including (i) equative constructions (section 4); (ii) selective island contexts (section 5); and (iii) clause-internal “niching” (section 7). Additionally, as-parentheticals allow a wider range of interpretations relative to their antecedents than do appositive relatives (section 6). This paper offers a unified account of these differences and others based largely on the respective semantic types of the gaps these clauses define: as-clause traces are propositional; nonrestrictive relative traces are individual denoting (i.e., nominalized propositions). The type distinction follows from the lexical denotations of these morphemes (section 2) and combines with independently motivated principles to predict the clauses’ divergent behavior. The analysis also unifies the various kinds of appositive-relatives and similarly informs our understanding of predicate-type as-clauses (e.g., Sue hates parties, as does Ali); see section 8.

Expressives and Identity Conditions

EXPRESSIVES AND IDENTITY CONDITIONS Christopher Potts Ash Asudeh Seth Cable Yurie Hara Eric McCready Luis Alonso-Ovalle Rajesh Bhatt Christopher Davis Angelika Kratzer Tom Roeper Martin Walkow Müller, Gereon. 2004. Verb-second as vP-first. Journal of Comparative Germanic Linguistics 7:179–234. Nilsen, Øystein. 2003. Eliminating positions. Doctoral dissertation, OTS, Utrecht. Pafel, Jürgen. 1998. Skopus und logische Struktur. Arbeitspapiere des Sonderforschungsbereichs 340, Bericht 129. Tübingen/Stuttgart: University of Tübingen/University of Stuttgart. Reinhart, Tanya. 1983. Anaphora and semantic interpretation. Chicago: University of Chicago Press. Sauerland, Uli, and Paul Elbourne. 2002. Total reconstruction, PF movement, and derivational order. Linguistic Inquiry 33: 283–319. Thiersch, Craig. 1985. Some notes on scrambling in the German Mittelfeld, VP and X-bar theory. Ms., University of Connecticut, Storrs, and University of Cologne.

Model theory and the content of OT constraints

We develop an extensible description logic for stating the content of optimalitytheoretic constraints in phonology, and specify a class of structures for interpreting it. The aim is a transparent formalisation of OT. We show how to state a wide range of constraints, including markedness, input–output faithfulness and base–reduplicant faithfulness. However, output–output correspondence and ‘ intercandidate ’ sympathy are revealed to be problematic : it is unclear that any reasonable class of structures can reconstruct their proponents’ intentions. But our contribution is positive. Proponents of both output–output correspondence and sympathy have offered alternatives that fit into the general OT picture. We show how to state these in a reasonable extension of our formalism. The problematic constraint types were developed to deal with opaque phenomena. We hope to shed new light on the debate about how to handle opacity, by subjecting some common responses to it within OT to critical investigation.

A Fast Unified Model for Parsing and Sentence Understanding

Tree-structured neural networks exploit valuable syntactic parse information as they interpret the meanings of sentences. However, they suer from two key technical problems that make them slow and unwieldyforlarge-scaleNLPtasks: theyusually operate on parsed sentences and they do not directly support batched computation. We address these issues by introducingtheStack-augmentedParser-Interpreter NeuralNetwork(SPINN),whichcombines parsing and interpretation within a single tree-sequence hybrid model by integrating tree-structured sentence interpretation into the linear sequential structure of a shiftreduceparser. Ourmodelsupportsbatched computation for a speedup of up to 25◊ over other tree-structured models, and its integrated parser can operate on unparsed data with little loss in accuracy. We evaluate it on the Stanford NLI entailment task and show that it significantly outperforms other sentence-encoding models.