A bi-threshold model for PP-attachment disambiguation through backing off to 2-tuples directly

We present a bi-threshold approach to prepositional phrase attachment disambiguation which is based on a pseudo-backed-off model through backing off to 2-tuples directly. We use 2-tuples counts only to make attachment decision relied on two threshold comparisons. The model alleviates the sparse data problem because 4-tuples and 3-tuples counts are not used. It also achieves good performance with 85.02% accuracy on the test data of the IBM data set. Our experiments show that 4-tuples and 3-tuples counts are unnecessary for resolving prepositional phrase attachment.