Yves Peirsman

Deterministic coreference resolution based on entity-centric, precision-ranked rules

We propose a new deterministic approach to coreference resolution that combines the global information and precise features of modern machine-learning models with the transparency and modularity of deterministic, rule-based systems. Our sieve architecture applies a battery of deterministic coreference models one at a time from highest to lowest precision, where each model builds on the previous models cluster output. The two stages of our sieve-based architecture, a mention detection stage that heavily favors recall, followed by coreference sieves that are precision-oriented, offer a powerful way to achieve both high precision and high recall. Further, our approach makes use of global information through an entity-centric model that encourages the sharing of features across all mentions that point to the same real-world entity. Despite its simplicity, our approach gives state-of-the-art performance on several corpora and genres, and has also been incorporated into hybrid state-of-the-art coreference systems for Chinese and Arabic. Our system thus offers a new paradigm for combining knowledge in rule-based systems that has implications throughout computational linguistics.

Metonymy as a prototypical category

Abstract A definition of metonymy that has gained some popularity in Cognitive Linguistics contrasts metonymical semantic shifts within a domain or domain matrix with metaphorical shifts that cross domain boundaries. In the past few years, however, this definition of metonymy has become subject to more and more criticism, in the sense that it relies too much on the vague notions of domains or domain matrices to be fully reliable. In this article, we address this problem by focusing on a nonunitary, prototypical definition of contiguity (the concept that used to be seen as the defining feature of metonymy before Cognitive Linguistics introduced domains and domain matrices). On the basis of the traditional pre-structuralist literature on metonymy, we identify a large number of typical metonymical patterns, and show that they can be classified in terms of the type of contiguity they are motivated by. We argue that metonymies, starting from spatial part-whole contiguity as the core of the category, can be plotted against three dimensions: strength of contact (going from part-whole containment over physical contact to adjacency without contact), boundedness (involving an extension of the part-whole relationship towards unbounded wholes and parts), and domain (with shifts from the spatial to the temporal, the spatio-temporal and the categorial domain).

Advances in Cognitive Sociolinguistics

Cognitive Sociolinguistics is a novel and burgeoning field of research which seeks to foster investigation into the socio-cognitive dimensions of language at a usage-based level. Advances in Cognitive Sociolinguistics brings together ten studies into the social and conceptual aspects of language-internal variation. All ten contributions rely on a firm empirical basis in the form of advanced corpus-based techniques, experimental methods and survey-based research, or a combination of these. The search for methods that may adequately unravel the complex and multivariate dimensions intervening in the interplay between conceptual meaning and variationist factors is thus another characteristic of the volume. In terms of its descriptive scope, the volume covers three main areas: lexical and lexical-semantic variation, constructional variation, and research on lectal attitudes and acquisition. It thus illustrates how Cognitive Sociolinguistics studies both the variation of meaning, and the meaning of variation.

The English genitive alternation in a cognitive sociolinguistics perspective

As a corpus-based inquiry into the probabilistic nature of lectal variation, the present study seeks to explore how language-external determinants of linguistic variation – real time, geography, text type – interact with language-internal determinants of linguistic variation, and in so doing shape cognitive and probabilistic grammars. The concrete empirical attention of this study will be directed toward the English genitive alternation as an instructive case study. The evidence suggests that the probabilistic grammar underlying the system of genitive choice is fundamentally the same across sampling times, geographic varieties of English, and text types. This overall qualitative stability notwithstanding, the importance of individual conditioning factors varies across different data sources, and this variability is shown to be mediated by language-external factors.

Predicting Strong Associations on the Basis of Corpus Data

Current approaches to the prediction of associations rely on just one type of information, generally taking the form of either word space models or collocation measures. At the moment, it is an open question how these approaches compare to one another. In this paper, we will investigate the performance of these two types of models and that of a new approach based on compounding. The best single predictor is the log-likelihood ratio, followed closely by the document-based word space model. We will show, however, that an ensemble method that combines these two best approaches with the compounding algorithm achieves an increase in performance of almost 30% over the current state of the art.

Semantic relations in bilingual lexicons

Bilingual lexicons, essential to many NLP applications, can be constructed automatically on the basis of parallel or comparable corpora. In this article, we make two contributions to their induction from comparable corpora. The first one concerns the creation of these lexicons. We show that seed lexicons can be improved by adding a bootstrapping procedure that uses cross-lingual distributional similarity. The second contribution concerns the evaluation of bilingual lexicons. It is generally based on translation lexicons, which corresponds to the implicit assumption that (cross-lingual) synonymy is the semantic relation of primary interest, even though other semantic relations like (cross-lingual) hyponymy or cohyponymy make up a considerable portion of translation pair candidates proposed by distributional methods. We argue that the focus on synonymy is an oversimplification and that many applications can profit from the inclusion of other semantic relations. We study what effect these semantic relations have on two cross-lingual tasks: the cross-lingual projection of polarity scores and the cross-lingual modeling of selectional preferences. We find that the presence of non-synonymous semantic relations may negatively affect the former of these tasks, but benefit the latter.

Don't let metonymy be misunderstood: An answer to Croft

Abstract Misunderstandings are common, even among semanticists. Indeed, after having read William Crofts answer to our article, we believe that most of his criticisms arise from a misunderstanding of our intentions, and more broadly, of the type of model that we have developed. In this answer to Croft, we will first argue that our prototype model, as any prototype model in general, does not claim to put forward any necessary or sufficient attributes for the concept that it describes. Second, we will show that our use of contiguity is less controversial and less problematic than Croft claims. Third, we will counter Crofts criticism that our model is undermined by its own use of domains. Fourth, we will wrap up with some final remarks on Crofts defense of domain highlighting.

Word Space Models of Lexical Variation

In the recognition of words that are typical of a specific language variety, the classic keyword approach performs rather poorly. We show how this keyword analysis can be complemented with a word space model constructed on the basis of two corpora: one representative of the language variety under investigation, and a reference corpus. This combined approach is able to recognize the markers of a language variety as words that not only have a significantly higher frequency as compared to the reference corpus, but also a different distribution. The application of word space models moreover makes it possible to automatically discover the lexical alternative to a specific marker in the reference corpus.