🧑‍🔬 Iñigo Lopez-Gazpio

Iñigo Lopez-Gazpio works on semantic textual similarity, the problem of scoring how close in meaning two short pieces of text are. Their papers here deal with designing and running shared evaluation tasks. That means assembling sentence pairs, gathering graded judgments from crowd workers, and checking how closely automatic systems track human ratings. The work carries this kind of evaluation beyond English and touches on Multilingual & cross-lingual NLP and Interpretability.

The most cited of these papers is SemEval-2017 Task 1. It moved the similarity task's emphasis away from English toward multilingual and cross-lingual comparisons, with Arabic, Spanish and Turkish as the added languages. The bulk of its sentence pairs were taken from the SNLI corpus and picked using GloVe vector similarity. They were rated through Amazon Mechanical Turk and then translated. The earlier SemEval-2015 Task 2 prepared fresh English and Spanish test material, drawing on headlines, captions of images, answers from a tutoring system and forum posts, and it added a pilot on interpretability. Candidate pairs there were sampled evenly across ranges of string similarity. After crowd labelling, an automatic step kept the pairs that best balanced score spread, annotator agreement and difficulty for simple string matching.

Between 2015 and 2017 the work moves from English and Spanish data toward wider cross-lingual coverage, with the newer task built around translated pairs across several languages. Lopez-Gazpio's co-authors on these task papers include Eneko Agirre, Daniel Cer and Mona Diab, along with Rada Mihalcea, Lucia Specia, Carmen Banea and German Rigau.

Iñigo Lopez-Gazpio has 2 papers in our library, in 2015–2017, most of them on Multilingual & cross-lingual NLP, Interpretability; often with Daniel Cer, Eneko Agirre, Mona Diab.