A Light-Weight Automated Approach To Classify and Annotate English Sentences

Abstract: Natural languages like English, on the one hand, offer a lot of flexibility to express a piece of information in a variety of ways, but, on the other hand, pose challenges for automatic information extraction as the relevant information is embedded at different places in different types of sentences. This necessitate the NLP applications to first identify the sentence types so that the relevant information can be extracted from them accordingly. The training, testing and evaluation of such applications need large corpora annotated with the sentence types; however, the available open/free corpora are not annotated with sentence types. In this paper, we propose an automated approach along with a prototype tool support (AutoSA) to classify the English sentences into four types (Declarative, Imperative, Interrogative and Exclamatory), and to classify each type of the sentence further into the four subtypes (simple, compound, complex and compound-complex). First, the approach uses the Stanford Natural Language parser API to generate parse trees of the sentences. Then, it scans the parse tree of each sentence to search for the patterns that we have framed for the classification. Based on the pattern (s) found in the parse tree of the sentence, the approach annotates the sentence with the corresponding type and subtype. We present the two case studies that we have conducted for the validation of the proposed approach. The results of the case study conducted on 2182 sentences of Tatoeba Corpus showed 98.6% precision and 99.5% recall, and the results of the case study conducted on 1650 sentences of Penn Treebank Corpus showed 92.4% precision and 99.4% recall.