Title: How are you two related? Corpus-based Learning of Lexical Semantic Relations

Author: Vered Shwartz

Abstract: 

Recognizing lexical semantic relations between words is an essential
component in semantic applications such as question answering and
recognizing textual entailment. In order to overcome lexical
variability, such systems traditionally relied heavily on lexical
resources such as WordNet.

In the main part of the talk I will discuss our work on automatic
detection of lexical semantic relations from free text. This task
stems from the limited coverage of lexical resources, both in terms of
missing lexical items (proper names, new words) and missing relations
between existing items. Typical approaches to address this task are
either distributional, i.e. based on the word embeddings of the two
target words, or path-based (pattern-based) approach, based on the
words co-occurrences in the corpus. I will present our integrated
path-based and distributional method for recognizing lexical semantic
relations, which is currently the state-of-the-art in this task.

In the second part, I will raise some questions about the interplay of
WordNet and word embeddings: is external lexical knowledge obsolete in
the deep learning era? And if it isn't, then how can lexical knowledge
from WordNet and other resources be incorporated into neural models
for semantic applications?


Bio: 

Vered is a Computer Science PhD student in Natural Language Processing
lab at Bar-Ilan University, under the supervision of Prof. Ido
Dagan. Her research focuses on lexical semantics. Her recent work
involved an integrated distributional and path-based method for
recognizing lexical semantic relations, and an
automatically-constructed resource of predicate paraphrases. She
completed her B.Sc. (2013) and M.Sc. (2015) in Computer Science in
Bar-Ilan University.