REPOSITORY > RESULTS

Abstract

Text Understanding is a long term goal of Artificial Intelligence. Its mission is to create
algorithms that will fully automatically understand human-composed text. In this dissertation
we propose, implement and evaluate machine learning approaches and meaning
representations to make progress towards Text Understanding. Meaning representations
are used to encode the information contained in a sentence. Furthermore, they can be used
in applications that require detailed understanding of the sentence. The focus of our work
is on the intrinsic meaning (semantics) of sentences rather than their context or syntactic
characteristics.
An ontology is a conceptualization that can be used to express the meaning of text.
Predicate logic, which connects concepts from the ontology, is one of the meaning representations
that we use. We propose two rule-based approaches for mapping text into
predicate logic. One approach is based on textual patterns, while the other approach is
based on a syntax parser. Rules are manually constructed and the approaches work on the
assumption of a complete ontology. This led us to develop a grammar induction approach
for semantic parsing and ontology learning. The induced context-free grammar parses a
sentence of text into a semantic tree, which is a meaning representation, where each node
has its own semantic category, e.g. person, location, profession. Furthermore, some of the
nodes can be aligned with ontology concepts. The trees are used as a source for relation
extraction and instance discovery, while a taxonomy is extracted from the grammar. Thus,
both grammar and semantic trees are used to extend the ontology, so that it also represents
the meaning of the input sentences.
Text understanding can be both learned and evaluated through the task of Natural
Language Inference (NLI), where the goal is to determine whether one sentence (hypothesis)
entails, contradicts or is neutral in respect to the other sentence (premise). Instead of
focusing on this classification task, we propose several text generative neural networks for
generating hypotheses which allow the construction of new NLI datasets. To evaluate the
models, we propose a new metric for the accuracy of the classifier trained on the generated
dataset and tested on the original, manually constructed dataset. The accuracy obtained
by our best generative model is only 2.7% lower than the accuracy of the classifier trained
on the original dataset. The model learns a mapping embedding for each training example.
Furthermore, combining the best generated and original dataset results in the highest accuracy.
Our metric awards diverse, accurate, non-trivial and comprehensible examples. By
comparing various metrics we show that datasets that obtain higher ROUGE or METEOR
scores do not necessarily yield higher classification accuracies. We also provide an analysis
of what are the characteristics of a good dataset including the indistinguishability of the
generated dataset from the original one.