Department of Computer Science, Princeton University, Princeton, USA.

---

Abstract—SAT sentence completion questions are designed to assess knowledge of the English language. The ability to answer such types of questions has wide implications in optical character recognition, speech recognition, and word-suggestion programs. In our study, we analyze several statistical corpus-based methods through which to answer such questions, including normalized pointwise mutual information, co-occurrence frequencies, latent semantic analysis, and the word2vec neural net implementations of continuous bag of words (CBOW) and continuous skip-gram (CSKIP) models. We find that the co-occurrence frequency method has a strong performance with 52% correctness and that combining the co-occurrence frequency method with CBOW and CSKIP results in a state-of-the-art performance of 59%. The results of this study demonstrate that local context is a fairly strong measure in determining how well a word fits in a sentence and that exploration of non-similarity based methods may be required to further enhance the ability of computers to answer such questions.

Index Terms—MSR sentence completion challenge, question answering, SAT, sentence completion.

[PDF]

Cite: Eugene Tang, "Assessing the Effectiveness of Corpus-Based Methods in Solving SAT Sentence Completion Questions," Journal of Computers vol. 11, no. 4, pp. 266-279, 2016.