We provide a comparative study between neural word representations and traditional vector spaces based on cooccurrence counts, in a number of compositional tasks. We use three different semantic spaces and implement seven tensor-based compositional models, which we then test (together with simpler additive and multiplicative approaches) in tasks involving verb disambiguation and sentence similarity. To check their scalability, we additionally evaluate the spaces using simple compositional methods on larger-scale tasks with less constrained language: paraphrase detection and dialogue act tagging. In the more constrained tasks, co-occurrence vectors are competitive, although choice of compositional method is important; on the largerscale tasks, they are outperformed by neural word embeddings, which show robust, stable performance across the tasks.

This paper presents a series of experiments in applying compositional distributional semantic models to dialogue act classification. In contrast to the widely used bag-ofwords approach, we build the meaning of an utterance from its parts by composing the distributional word vectors using vector addition and multiplication. We investigate the contribution of word sequence, dialogue act sequence, and distributional information to the performance, and compare with the current state of the art approaches. Our experiment suggests that that distributional information is useful for dialogue act tagging but that simple models of compositionality fail to capture crucial information from word and utterance sequence; more advanced approaches (e.g. sequence-or grammar-driven, such as categorical, word vector composition) are required.

No abstract

No abstract

Twitter has become a rich source for linguistic data. Here, a possibility of building a trilingual Latvian-Russian-English corpus of tweets from Riga, Latvia is investigated. Such a corpus, once constructed, might be of great use for multiple purposes including training machine translation models, examining cross-lingual phenomena and studying the population of Riga. This pilot study shows that it is feasible to build such a resource by collecting and analysing a pilot corpus, which is made publicly available and can be used to construct a large comparable corpus.