CLIP Colloquium: Marianna Apidianaki (LIMSI-CNRS)
Title: Senses, paraphrases and substitutes: exploring boundaries in translation induced meaning representations
Abstract: Translations of words in parallel corpora are a valuable source of indirect supervision for lexical semantic analysis. They can be used to detect word senses on the premise that different meanings are translated differently, and to identify synonyms and paraphrases sharing translations in other languages. In this talk, I will show that complementing translation information with distributional evidence can lead to higher quality sense representations and paraphrases. Distributional models might even suffice for some semantic processing tasks without explicitly accounting for sense. Then, I will present results of a recent study showing that different sense representations might be appropriate for different lemmas depending on the partitionability of their senses. I will show how translation and paraphrase data can be used with clusterability measures to estimate how easily a word’s usages can be partitioned into discrete senses. This type of analysis can help to determine the optimal representation for the automatic processing of different words and to plan annotation tasks.
Bio: Marianna Apidianaki is a CNRS Researcher at the LIMSI lab working on multilingual semantics and translation. Her research addresses both theoretical and application-oriented aspects of semantics, ranging from explorations on the nature of automatically acquired sense representations and their adequacy for automatic processing, to the integration of semantic models in machine translation systems and evaluation metrics. Since October 2016, Marianna is a Visiting Scholar in the Computer and Information Science Department at the University of Pennsylvania in the team of Chris Callison-Burch.
