Two papers by the LSV group were accepted to this year’s “ACL 2018 Workshop on Deep Learning Approaches for Low Resource Natural Language Processing (DeepLo)”. Michael Hedderich and Dietrich Klakow present a method to train a neural network in a low-resource setting by using automatically annotated data and handling the noise in the labels. Clayton Greenberg, Mittul Singh and Dietrich Klakow explore the quality of embeddings of rare words, finding that below a certain number of occurrences, SWORDSS (developed at LSV in 2016) can generate better embeddings than standard word2vec.
A paper by Marc Schulder and Michael Wiegand of the LSV group and their colleague Josef Ruppenhofer from the Institute for German Language, Mannheim, was accepted at 27th International Conference on Computational Linguistics (COLING 2018). In the paper, the authors present methods for automatically identifying verbs that, like negation, can change the sentiment polarity of a phrase (e.g. “abandon hope” or “alleviate pain”). They expand the previously established mono-lingual methods by introducing cross-lingual methods that use bilingual dictionaries and cross-lingual word embeddings.