Dies ist das Newsblog des Sprachwissenschaftlichen Instituts an der Ruhr-Universität Bochum.





Ruhr-Universität Bochum
Sprachwissenschaftliches Institut



Powered by PivotX - 2.3.11 
XML-Feed (RSS 1.0) 
XML: Atom Feed 

« Vortrag von Tristan M… | home | Vortrag von Jana Häus… »

Vortrag von James Henderson am Dienstag, 24.10.2017 (16:00 Uhr - GB 3/159)

Dienstag, 17. Oktober 2017. Aus der Kategorie 'Vortragsreihe'.

Das Sprachwissenschaftliche Institut lädt herzlich ein zum Vortrag von

James Henderson (Idiap Research Institute & Université de Genève)

Inducing representations for lexical entailment
with entailment-based distributional semantics

Distributional semantics uses the distribution of words in their contexts in text to induce vector-space representations of words (word embeddings). It has been very successful at capturing semantic similarity between words. But previous work has had difficulty using these representations to capture lexical entailment, such as hyponymy, which is an asymmetric relation measuring information inclusion.

This talk presents distributional semantic models where the induced vector representations are designed to capture entailment. These models use the recently proposed entailment-vectors framework to model the semantic relationship between a word and its context. Training these models on a large corpus of text induces entailment-vector representations of words, which indicate when one word’s vector entails another word’s vector. We evaluate these entailment predictions on the task of hyponymy detection, achieving a substantial improvement over previous results.