SAT-Level Sentence-Completion Questions

Sentence Completion using Natural Language Processing Techniques

The primary goal of this project is to automatically answer SAT/GRE style sentence completion questions. For this we have to build a model that will deal with semantic coherence on a sentence level. Initially, we start out by applying Natural Language Processing Techniques on simple sentences. Later, we will see how these Natural Language Processing Techniques scale to complex analogy-based sentences. [Sample sentences are listed in dataset section]

This project will be done as a two-member team -

1) Sai Chaitanya Mallampati [109597269]

2) Paavan Kumar Sirigiri [109596437]

In the research paper {f ``Computational Approches to Sentence Completion"}, the authors Geoffrey Zweig, John C. Platt, Christopher Meek, Christopher J.C. Burges, Ainur Yessenalina, Qiang Lu study the problem of sentence level semantic coherence by answering SAT level sentence completion questions. They build various models using backoff n-gram language model, maximum entropy class-based n-gram language model, recurrent neural net language model, latent semantic analysis and a combination of n-gram and latent semantic analysis. Some of the key observations are -

1) Latent semantic analysis has better accuracy than recurrent neural network [which is better than the n-gram language model]. Since latent semantic analysis takes into picture the global coherence, it was concluded that global coherence is an important parameter to weigh in.

2) A combination of n-gram language model and latent semantic analysis performs better than the individual models. This hybrid model performs significantly better than random selection but is no where close to the levels of human accura...

... middle of paper ...

...of above models

5) We will also explore the feasibility of integrating LSA with labelled dependency language model in order to attain better performance.

We intend to use the n-gram language model results as our baseline for other language models being developed.

For synonym and antonym detection, we will use the WordNet tool. [url {http://wordnet.princeton.edu/}]

1) Computational Approches to Sentence Completion by Geoffrey Zweig, John C. Platt, Christopher Meek, Christopher J.C. Burges [Microsoft Research], Ainur Yessenalina [Cornell University] , Qiang Lu [University of California, Irvine]

2) Dependency language models for sentence completion by Joseph Gubbins and Andreas Vlachos [University of Cambridge]

3) Sentence Completion Task using Web-scale Data by Kyusong Lee and Gary Geunbae Lee [Pohang University of Science and Technology]