PhD Proposal: Identifying Asymmetric Semantic Divergences for Improved Multilingual Natural Language Understanding

A fundamental goal of natural language processing is to build computational models that can capture the meaning of text generated by humans. Such models are essential to many NLP tasks such as question answering and summarization. A key component of these models is the ability to detect semantic divergences, i.e. mismatches in meaning of texts. Such divergences can also be asymmetric in nature -- for instance, when a piece of text is more specific than another, we know that the latter implies the former, but not vice versa. Detecting such asymmetric semantic divergences can improve natural language understanding even when exact equivalence does not exists.In this proposal, we focus on detecting asymmetric semantic divergences to improve multilingual natural language understanding. We start by providing empirical evidence that semantic divergences exist in the wild and we can detect them using a model of symmetric semantic similarity to help downstream applications such as machine translation. We then turn to the question of detecting fine-grained asymmetric lexical divergences by identifying cross-lingual hypernymy, and show that word embedding based models tailored to this task perform better than generic word embeddings. In our final piece of completed work, we move from the type word type level to the token level and identify hypernymy between words used in context, where the context consists of a full sentence.Having observed the impact of semantic divergences on downstream applications, in our first proposed work, we ask how we can exploit the presence of such divergences to improve performance on a downstream task, viz. natural language inference. We propose two extensions of a neural model for this task that explicitly take into account knowledge of lexical relations. In our second proposed work, we aim to investigate how we can further improve detection of such asymmetric lexical divergences using word embeddings. Generic word embeddings can identify general-purpose semantic similarity between words but they are unsuitable for detecting more specific relations which are usually asymmetric in nature. We propose to construct a semi-supervised, multilingual model of word embeddings using a combination of limited labeled data and large amounts of raw, unlabeled corpora that improves identification of asymmetric lexico-semantic relations within and across languages.

Examining Committee:

Chair: Dr. Marine Carpuat Dept. rep: Dr. David Jacobs Members: Dr. Jordan Boyd-Graber Dr. Philip Resnik Dr. Ido Dagan