Authorship verification is the task of determining if two distinct writing samples share the same author and is typically concerned with the attribution of written text. In this paper, we explore the attribution of transcribed speech, which poses novel challenges. The main challenge is that many stylistic features, such as punctuation and capitalization, are not informative in this setting. On the other hand, transcribed speech exhibits other patterns, such as filler words and backchannels (e.g., 'um', 'uh-huh'), which may be characteristic of different speakers. We propose a new benchmark for speaker attribution focused on human-transcribed conversational speech transcripts. To limit spurious associations of speakers with topic, we employ both conversation prompts and speakers participating in the same conversation to construct verification trials of varying difficulties. We establish the state of the art on this new benchmark by comparing a suite of neural and non-neural baselines, finding that although written text attribution models achieve surprisingly good performance in certain settings, they perform markedly worse as conversational topic is increasingly controlled. We present analyses of the impact of transcription style on performance as well as the ability of fine-tuning on speech transcripts to improve performance.

Using novel approaches to dataset development, the Biasly dataset captures the nuance and subtlety of misogyny in ways that are unique within the literature. Built in collaboration with multi-disciplinary experts and annotators themselves, the dataset contains annotations of movie subtitles, capturing colloquial expressions of misogyny in North American film. The dataset can be used for a range of NLP tasks, including classification, severity score regression, and text generation for rewrites. In this paper, we discuss the methodology used, analyze the annotations obtained, and provide baselines using common NLP algorithms in the context of misogyny detection and mitigation. We hope this work will promote AI for social good in NLP for bias detection, explanation, and removal.

Lexical Speci cation and Meaning in Use: Traditional Views on the LexiconTraditionally, the lexicon is the list of words of a language or communication system.From a biological perspective, the lexica of human languages are special with respect to animal communication systems because they can be freely extended, both by the creation of new (lexical) forms and by the extension of the meaning of these words.Notwithstanding this extensibility, the construction of dictionaries as xed lists of words with their meanings pays o for the purposes of language education, reliable communication and translation.Computational models of language largely share this view, as it seems initially reasonable to think that such a lexicon could be part of what characterizes the human ability to code thoughts into linguistic expressions and to recover thoughts from such expressions.What is needed for conventional dictionaries-in the tradition started by such eminent scholars as Samuel Johnson and Jacob Grimm-is precision with respect to the characterization of the word senses described.Such a task would be * We wish to express our thanks to audiences at Szklarska Poreba, the Workshop Bridging Formal and Conceptual Semantics, and Formal Semantics Meets Cognitive Semantics as well as to Katrin Erk, to Louise McNally and, especially, to an anonymous reviewer of these proceedings for many useful comments, more, in fact, than we could integrate into this preliminary report.In addition, Lotte Hogeweg and Sander Lestrade would like to thank the the Netherlands Organization for Scienti c Research (NWO) for their nancial support.All errors are our own.Authors after the rst author are listed in alphabetical order.

Loftus (1975) and subsequent established that presupposing new information during witness questioning can influence subsequent eyewitness reports. We report on a study that attempts to replicate these results for another language (French) in order to better understand some of the factors that facilitate (or not) disinformation. Our sixty participants watched a video of an attempted robbery and answered questions in which the existence of an element was (a) a true presupposition, (b) a false presupposition, or (c) not presupposed. A week later, all participants received a second questionnaire where the critical information was questioned, followed by a working memory task and a demographic survey. Our analysis (by binomial logistic regression) tests the proportion of false responses to the second survey by condition. Our results show that presuppositions of true information significantly decrease the rate of false responses (p < .01) and that false presuppositions increase the rate of false responses, though not significantly (p = .09), as compared to the control condition. We conclude that while questions after an event will have an effect on witness reports, it is not valid to assume that the specific effects will be universal across languages and cultures. Finally, we discuss the implications of these results for various interview techniques used across Francophone nations that could diminish the integrity ofjudicial testimony.

This paper reports on the first detailed experimental investigation of the perception of the felicity of evidential markers in context. We investigated the Japanese evidentials -rashii, -sooda, and -yooda in various discourse environments by manipulating two key variables: (a) whether there was any conjecture required, and (b) whether the information source was accessible firsthand to the speaker. This work provides a baseline against which future studies of other discourse variables can be measured, and our results present some challenges to established conceptions. For example, -rashii was found to be compatible with reportative utterances, building on its traditional categorization as a conjectural evidential. We situate our findings with respect to the typological literature and contemplate how the results may inform semantico-pragmatic theories of evidentiality. We further propose a slight modification to McCready and Ogata (2007) to account for the felicity of bare propositions with indirect information sources.

Loftus (1975) and subsequent established that presupposing new information during witness questioning can influence subsequent eyewitness reports. We report on a study that attempts to replicate these results for another language (French) in order to better understand some of the factors that facilitate (or not) disinformation. Our sixty participants watched a video of an attempted robbery and answered questions in which the existence of an element was (a) a true presupposition, (b) a false presupposition, or (c) not presupposed. A week later, all participants received a second questionnaire where the critical information was questioned, followed by a working memory task and a demographic survey. Our analysis (by binomial logistic regression) tests the proportion of false responses to the second survey by condition. Our results show that presuppositions of true information significantly decrease the rate of false responses (p

One might think that Sam's utterance in (1) is a subjective one, essentially expressing that he personally finds the cake tasty, in which case one would not expect significant meaning differences between (1) and a similar utterance where the subjectivity is made explicit, such as (3).

We present a unified categorial analysis of several types of English comparative, superlative, and THE SAME/DIFFERENT (S/D) sentences, thereby accounting for parallels among these constructions first noted in Heim ms. Our analysis, couched in a linear-logic-based from of categorial grammar along the lines of Oehrle 1994, builds on the basic insights underlying Barker's (2007) `parasitic scope' analysis of internal readings of THE SAME, but is simpler and more general than Barker's. Ours is also the first unified analysis of all three kinds of phenomena. Our analysis of phrasal comparatives captures their essential similarity to associate-remnant S/D constructions such as ANNA READ THE SAME BOOK AS BILL.

Projective contents, which include presuppositional infe rences and Potts’ (2005) conventional implicatures, are contents which may project when a construct ion is embedded, as standardly identified by the Family of Sentences diagnostic (e.g. Chierchia an d McConnell-Ginet 1990). This paper establishes distinctions among projective contents on the basis of a series of diagnostics, including a variant of the Family of Sentences diagnostic, that can be a pplied with linguistically untrained consultants in the field and the laboratory. These diagnosti cs are intended to serve as part of a toolkit for exploring projective contents across language s, thus allowing generalizations to be examined and validated cross-linguistically. We apply the di agnostics in two languages, focussing on Paraguayan Guaranı́ (Tupı́-Guaranı́), and comparing th e results to those for English. Our study of Paraguayan Guaranı́ is the first systematic exploration o f projective content in a language other than English. Based on the application of our diagnostics to a wide range of constructions, four subclasses of projective contents emerge. The resulting ta xonomy of projective content has strong implications for contemporary theories of projection (e.g . Karttunen 1974; Heim 1983; van der Sandt 1992; Potts 2005; Schlenker 2009), which were develop ed for the projective properties of particular subclasses and fail to generalize to the full set of projective contents. ∗

AbstractThis paper introduces a novel approach to the question of how sociophonetic information might be stored cognitively that makes use of the tools of formal semantics. This approach involves applying conventional semantic tests for types of lexicalized meanings (e.g. presuppositions, conventional implicatures) to sociophonetic variables, with the hypothesis that, insofar as sociophonetic meaning patterns like lexical meaning, it should be stored in the same way. Two examples of how this approach can be implemented experimentally are given, applying projections tests (Frege, Letter to Peano, University of Chicago, 1896) (i.e., the ‘Family of Sentences’ tests, Chierchia & McConnell-Ginet 1990) and the ‘Hey, Wait a Minute!’ test (Shannon, Foundations of language 14: 247–249, 1976, von Fintel, Would you believe it? The king of France is back! Presuppositions and truth-value intuitions, Oxford University Press, 2004) to patterns of /æ / tensing and retraction in Minnesotan English and /aI / monophthongization in Southern English, respectively. Preliminary results of these experiments indicate that sociophonetic meaning patterns like secondary entailments (such as presuppositions and conventional implicatures) from the semantics literature, being both conventional and subsidiary to the proffered content of an utterance, and thus should be considered lexical. The primary goal of the paper, however, is to present a new way of thinking rather than to provide conclusive laboratory evidence for a specific position.

Munson et al. [J. Phonetics (2006)] found that 11 self-identified gay and 11 heterosexual men produced different variants of the vowel /æ/, with gay men producing lower, more retracted variants and heterosexual men producing higher, more tense variants. Listeners’ performance in a perception task in which they rated these talkers’ sexual orientation was correlated with these measures: talkers with higher, more-tense /æ/ were rated as sounding more heterosexual than talkers with lower, more retracted /æ/. However, Smith et al. [New Ways of Analyzing Variation (2008)] found the opposite pattern in an experiment in which listeners rated the sexual orientation of productions by 10 trained talkers of sentences containing either tense or retracted /æ/ variants. A different group of listeners showed the same pattern when rating the /æ/ words excised from these sentences, though these listeners replicated Munson et al.’s original finding when presented with the /æ/ words from the original 22 talkers. An attempt to reconcile these findings is made through detailed acoustic analysis of the /æ/ productions in the two experiments. Results underscore the importance of doing careful acoustic analyses in concert with careful phonetic transcription when conducting experiments on perception of social variables and speaker attributes.

We review Potts’ influential book on the semantics of conventional implicature (CI), offering an explication of his technical apparatus and drawing out the proposal’s implications, focusing on the class of CIs he calls supplements. While we applaud many facets of this work, we argue that careful considerations of the pragmatics of CIs will be required in order to yield an empirically and explanatorily adequate account.

Despite the notion that clefting is a cross-linguistic constituency test, Japanese allows some nonconstituent exceptions. There is, however, a certain restriction on the degree of flexibility; some constituents are more tightly connected (and thus less likely to be separated by clefting) than others. We refine Kubota and Smith’s (2006) CCG account in terms of Multi-Modal CCG (Baldridge, 2002): finer-grained modal control provides a means for capturing different degrees of connectedness between an argument and its functor. We then demonstrate how a MMCCG system that finds independent motivation from syntactic complex predicate data interacts with a simple analysis of clefting to account for the full range of clefting patterns. This in turn suggests that what seems to pose problems for a simple analysis of a given phenomenon (clefting) can be overcome once interactions with other phenomena are taken into account.

In the Japanese cleft construction, strings composed of multiple phrases that apparently do not form constituents can occupy the focus position. This poses a significant challenge to mainstream syntactic frameworks that have the notion of phrase structure as their back- bone. In fact, in the Minimalist tradition, there have been three proposals in the recent literature (Koizumi 1999, 2000; Takano 2002; Fukui and Sakai 2003) that introduce differ- ent operations to treat this phenomenon. While each of these works is suggestive in several ways, capturing some aspect of this construction within the assumed framework, none of them stand up to the full range of data without running into inconsistent consequences, as we will see below. Categorial grammar provides a particularly attractive framework for this problem because its independently motivated theoretical assumptions lead to a grammar that nat- urally licenses the kind of unusual constituents we find in this construction. The goal of the present paper is to develop an analysis of Japanese nonconstituent clefting in Combina- tory Categorial Grammar (CCG) (Steedman 1996, 2000b), in which a straightforward and precise syntax-semantics-information structure account of this construction is given. In constructing our analysis, we avail ourselves of only those assumptions and mechanisms that have been proposed in the literature of CCG and that find empirical motivation else- where in the grammar of Japanese. We argue that the resultant analysis is simpler and more explicit than any of the previous analyses and that it accounts for all the relevant data while

We critically examine a widely-entertained assumption that the semantics of Japanese Internally Headed Relative Clauses (IHRCs) involves a special kind of anaphora called E-type anaphora. We first summarize motivations for such an approach and a specific formulation by Shimoyama (1999). We then present novel data that pose problems for E-type approaches, followed by further data suggesting a previously-unnoticed parallel between IHRCs and null pronouns in Japanese. Based on these observations, we conclude that the anaphoric relations found in IHRCs and null pronouns should be given a unified treatment, a goal not attainable in any E-type approach to IHRCs.