We propose an approach to adjective-noun composition (AN) for corpus-based distributional semantics that, building on insights from theoretical linguistics, represents nouns as vectors and adjectives as data-induced (linear) functions (encoded as matrices) over nominal vectors. Our model significantly outperforms the rivals on the task of reconstructing AN vectors not seen in training. A small post-hoc analysis further suggests that, when the model-generated AN vector is not similar to the corpus-observed AN vector, this is due to anomalies in the latter. We show moreover that our approach provides two novel ways to represent adjective meanings, alternative to its representation via corpus-based co-occurrence vectors, both outperforming the latter in an adjective clustering task. 1

Models of lexical semantics are core paradigms in most NLP applications, such as dialogue, information extraction and document understanding. Unfortunately, the coverage of currently available resources (e.g. FrameNet) is still unsatisfactory. This paper presents a largely applicable approach for extending frame semantic resources, combining word sense information derived from WordNet and corpus-based distributional information. We report a large scale evaluation over the English FrameNet, and results on extending FrameNet to the Italian language, as the basis of the development of

Most attempts to integrate FrameNet in NLP systems have so far failed because of its limited coverage. In this paper, we investigate the applicability of distributional and WordNetbased models on the task of lexical unit induction, i.e. the expansion of FrameNet with new lexical units. Experimental results show that our distributional and WordNet-based models achieve good level of accuracy and coverage, especially when combined.

Abstract. A Polish WordNet has been under construction for two years. We discuss the organisation of the project, the fundamental assumptions, the tools and the resources. We show how our work di ers from that done on EuroWordNet and BalkaNet. In a year we expect the network to reach 20000 lexical units. Some 12000 entries will have been completed by hand. Work on others will be automated as far as possible; to that end, we have developed statistics-based semantic similarity functions and methods based on a form of chunking. The preliminary results show that at least semi-automated acquisition of relations is feasible, so that the lexicographers ' work may be reduced to revision and approval. 1 Organisation of the project Ever since the initial burst of popularity of the original WordNet [1, 2], there has been little doubt how useful wordnets are in Natural Language Processing. For those who work with a language that lacks a wordnet, the question is not whether, but how and how fast to construct such a lexical resource. The construction is costly, with the bulk of the cost due to the high linguistic workload. This appears to have been the case, in particular, in two multinational wordnetbuilding projects, EuroWordNet [3] and BalkaNet [4]. The recent developments in automatic acquisition of lexical-semantic relations suggest that the cost might be reduced. Our project to construct a Polish WordNet (plWordNet) explores this path as a supplement to a well-organized and well-supported e ort of a team of linguists/lexicographers.

We propose an approach to corpus-based semantics, inspired by cognitive science, in which different semantic tasks are tackled using the same underlying repository of distributional information, collected once and for all from the source corpus. Task-specific semantic spaces are then built on demand from the repository. A straightforward implementation of our proposal achieves state-of-the-art performance on a number of unrelated tasks. 1

We introduce a way to represent word pairs instantiating arbitrary semantic relations that keeps track of the contexts in which the words in the pair occur both together and independently. The resulting features are of sufficient generality to allow us, with the help of a standard supervised machine learning algorithm, to tackle a variety of unrelated semantic tasks with good results and almost no task-specific tailoring. 1

semantic relatedness with vector space models and random walks

Thesauri, that list the most salient semantic relations between words have mostly been compiled manually. Therefore, the inclusion of an entry depends on the subjective decision of the lexicographer. As a consequence, those resources are usually incomplete. In this paper, we propose an unsupervised methodology to automatically discover pairs of semantically related words by highlighting their local environment and evaluating their semantic similarity in local and global semantic spaces. This proposal differs from all other research presented so far as it tries to take the best of two different methodologies i.e. semantic space models and information extraction models. In particular, it can be applied to extract close semantic relations, it limits the search space to few, highly probable options and it is unsupervised. 1

Abstract. We describe a system to tackle the Lexical Substitution task that exploits, as its only resource, co-occurrence statistics from a large PoS-tagged corpus. The system exploits the word space model formalism, and represents the word to be substituted by a composite vector that takes into account both the overall distribution of the word in the input corpus and its local context. As far as the precision and recall are concerned, the system is ranked among the highest positions in the Evalita competition, while it results winner in the mode p and mode r ranking. Key words: word space models, composition in word space models, corpus-based semantics 1

The aim of this paper is to present a computational model of the dynamic composition and update of verb argument expectations using Distributional Memory, a state-of-the-art framework for distributional semantics. The experimental results conducted on psycholinguistic data sets show that the model is able to successfully predict the changes on the patient argument thematic fit produced by different types of verb agents. 1