Daniele Pighin

A comparison of unsupervised methods to associate colors with words

Colors have a very important role on our perception of the world. We often associate colors with various concepts at different levels of consciousnes and these associations can be relevant to many fields such as education and advertisement. However, to the best of our knowledge, there are no systematic approaches to aid the automatic development of resources encoding this kind of knowledge. In this paper, we propose three computational methods based on image analysis, language models, and latent semantic analysis to automatically associate colors to words. We compare these methods against a gold standard obtained via crowdsourcing. The results show that each method is effective in capturing different aspects of word-color associations.

Chunk-lattices for verb reordering in Arabic---English statistical machine translation

Syntactic disfluencies in Arabic-to-English phrase-based SMT output are often due to incorrect verb reordering in Verb–Subject–Object sentences. As a solution, we propose a chunk-based reordering technique to automatically displace clause-initial verbs in the Arabic side of a word-aligned parallel corpus. This method is used to preprocess the training data, and to collect statistics about verb movements. From this analysis we build specific verb reordering lattices on the test sentences before decoding, and test different lattice-weighting schemes. Finally, we train a feature-rich discriminative model to predict likely verb reorderings for a given Arabic sentence. The model scores are used to prune the reordering lattice, leading to better word reordering at decoding time. The application of our reordering methods to the training and test data results in consistent improvements on the NIST-MT 2009 Arabic–English benchmark, both in terms of BLEU (+1.06%) and of reordering quality (+0.85%) measured with the Kendall Reordering Score.

Evaluating the impact of syntax and semantics on emotion recognition from text

In this paper, we systematically analyze the effect of incorporating different levels of syntactic and semantic information on the accuracy of emotion recognition from text. We carry out the evaluation in a supervised learning framework, and employ tree kernel functions as an intuitive and effective way to generate different feature spaces based on structured representations of the input data. We compare three different formalisms to encode syntactic information enriched with semantic features. These features are obtained from hand-annotated resources as well as distributional models. For the experiments, we use three datasets annotated according to the same set of emotions. Our analysis indicates that shallow syntactic information can positively interact with semantic features. In addition, we show how the three datasets can hardly be combined to learn more robust models, due to inherent differences in the linguistic properties of the texts or in the annotation.