Research Scientist Talk:
Islam Beltagy
Title: Natural Language Semantics using Probabilistic Logic
Abstract:
Being able to automatically read and understand natural language is useful because the web has a huge amount of textual knowledge that we need to acquire automatically, and because it enables seamless interaction with human users. Current language understanding systems rely mostly on shallow represents for text, like bag of words or syntactic structure which do not work adequately for tasks that require deeper semantics understanding like question answering. A better semantic representation would be able to represent entities, events, the relations between them, logical operators and quantifiers, and would support a form of automated reasoning that can automatically draw conclusions from text to perform a task of interest.
In this talk, I will present a framework for deep semantic understanding. It integrates logical representation of sentences with uncertain, corpus-based information in order to overcome the brittleness of logical approaches. It represents tasks as probabilistic logic inference problems then solve them using tools like Markov Logic Networks (MLNs) and probabilistic Soft Logic (PSL). The knowledge base is collected from various sources like WordNet, PPDB and distributional information, in addition to learning dataset-specific rules from the training set. I applied this framework to textual entailment achieving a state of the art result on one dataset, applied it to textual similarity and currently working on applying it to open-domain question answering. Different inference algorithms for MLNs and PSL were developed to scale up inference and solve the challenges that each task introduces.
I will also talk about my work on semantic parsing, the task of translating natural sentences to an executable representation. The task in hand is translating a short sentence to an IF-Then statement. We explored various neural network architectures to solve the problem and achieved a new state of the art result.
