Συντάχθηκε 11-10-2013 15:30
από Balasi Panagiota
Email συντάκτη:
Ενημερώθηκε:
-
Ιδιότητα: -.
ΠΟΛΥΤΕΧΝΕΙΟ ΚΡΗΤΗΣ
Σχολή Ηλεκτρονικών Μηχανικών και Μηχανικών Υπολογιστών
Πρόγραμμα Προπτυχιακών Σπουδών
ΠΑΡΟΥΣΙΑΣΗ ΔΙΠΛΩΜΑΤΙΚΗΣ ΕΡΓΑΣΙΑΣ
Αροδάμη Χωριανοπούλου
με θέμα
Investigating Paraphrasing Algorithms with Application to Spoken Dialogue Systems
Τετάρτη 16 Οκτωβρίου 2013, 17:00
Αίθουσα Συνεδριάσεων, Κτίριο Επιστημών, Πολυτεχνειούπολη
Εξεταστική Επιτροπή
Επίκουρος Καθηγητής Πολυχρόνης Κουτσάκης (επιβλέπων)
Αναπληρωτής Καθηγητής Αλέξανδρος Ποταμιάνος
Καθηγητής Ευρυπίδης Πετράκης
Abstract
The task of paraphrasing is an important part of natural language processing (NLP) and is being employed in several NLP applications. A paraphrase is an alternative surface form in the same language expressing the same semantic content as the original form. In this thesis, we investigate paraphrase recognition and generation via a composition of Finite State Machines (FSMs) combined with english resources. Additionally, we propose several string similarity measures for paraphrase recognition of sentences.
An FSM defines a language by accepting a string of input tokens in the language and rejecting those that are not included in the language. In our case, the tokens are the words of the language. Our motivation was to create an alignment between sentences by using FSMs. This linking between two sentences allows as to identify whether or not they are paraphrases. Then, by adding a language model we manage to generate paraphrases.
To test our idea, we experimented on domain-specific paraphrasing, and more specifically, in the travel domain field, by employing paraphrase recognition and
generation on prompts. For this purpose, we have used two different types of corpora. Our first corpus contains web documents harvested from the web while the second was created from the CrowdFlower crowdsourcing service.
To evaluate the performance of the method we used the datasets from the SemEval workshop of 2012, in order to estimate the semantic equivalence between
sentences. The comparison of the results at the evaluation stage with scores from string-based similarity metrics, shows that our implementation achieves encouraging
performance.
