JU CSE NLP @ SemEval 2017 Task 7: Employing Rules to Detect and Interpret English Puns

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 432–435, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics JU_CSE_NLP at SemEval 2017 Task 7: Employing Rules to Detect and Interpret English Puns Dipankar Das and Aniket Pramanick Department of Computer Science and Engineering Jadavpur University, Kolkata, India dipankar.dipnil2005@gmail.com abcdefgh.ABCD65@gmail.com Abstract The problem of detection and interpreta- tion of English puns falls under the area of word sense disambiguation in natural language processing, which deals with the sense of a word used in a sentence from the readers’ perspective. We have tried to design a system to identify puns from a sen- tence by developing a cyclic dependency– based system which is implemented based on some rules which are actually statistical inferences taken from a set of random data collected from the Web. 1 Introduction Extensive research has been done in the field of modelling and detecting puns (Hempelmann, 2003; Miller and Gurevych, 2015). The context or the sense depends largely on the perspective and knowl- edge of the reader about a particular language. For example, in the sentence, ‘I was a banker but I lost interest,’ the word in italics conveys two dif- ferent meanings or ‘senses’ in the sentence. So, the word ‘interest’ could be called a pun. A pun is the exploitation of the various meanings of a word or words with phonetic similarity but different meanings. Our system is a rule-based implementation of a dependency network and a hidden Markov model. 2 Dataset and Preprocessing In the present shared task (Miller et al., 2017), participants are provided with a trial dataset and a test dataset. No training data was supplied due to the large cardinality of such words or contexts in general. In case of the subtask on pun detection (Subtask 1), the test data was subdivided into two sets: a homographic set containing 2250 contexts, and a heterographic set containing 1780 contexts. On the other hand, in case of Subtask 2, another set of data was provided and that set was also subdivided into homographic and heterographic sets. For training purposes, or rather to analyze the contexts statistically, a dataset was collected from random sources, mostly form Project Gutenberg and used in our present experiments. This dataset contains 413 sentences and is not subdivided into homographic and heterographic subsets. The given data containing English contexts is preprocessed and each word of a sentence is tagged with its part of speech using NLTK, an open-source package for NLP written in Python. For example, the sentence, ‘I was a banker but I lost interest.’ is tagged using the Stanford NLP parser as follows: [(‘I’, ‘PRP’), (‘was’, ‘VBD’), (‘a’, ‘DT’), (‘banker’, ‘NN’), (‘but’, ‘CC’), (‘I’, ‘PRP’), (‘lost’, ‘VBD’), (‘interest’, ‘NN’), (‘.’, ‘.’)] We also generate the parse tree for the sentence, which looks as in Figure 1. Using such parse trees, the clauses are identified and used for our further tasks. 3 System Framework We have used a hidden Markov model (Ghahra- mani, …

Content truncated to 500 words