Πλοήγηση ανά Συγγραφέα "Tassias, Panagiotis"

A B C D E F G H I J K L M N O P Q R S T U V W X Y Z

Α Β Γ Δ Ε Ζ Η Θ Ι Κ Λ Μ Ν Ξ Ο Π Ρ Σ Τ Υ Φ Χ Ψ Ω

Τώρα δείχνει 1 - 1 από 1

A prompting-based encoder-decoder approach to intent recognition and slot filling
(2021-12-31) Tassias, Panagiotis; Τασσιάς, Παναγιώτης; Athens University of Economics and Business, Department of Informatics; Vassalos, Vasilios; Malakasiotis, Prodromos; Androutsopoulos, Ion
In recent years, there is an increasing interest in developing advanced conversationalagents that facilitate users to accomplish specific goals. Natural LanguageUnderstanding (NLU), a subfield of Natural Language Processing, is at the core ofthese task-oriented dialogue systems. In this thesis, we experimented with differentways of tackling NLU problems, focusing on the sub-tasks of Intent Recognitionand Slot Filling. By conducting various experiments on the publicly available ATISand SNIPS datasets, we confirm that in cases where there is explicit slot labelalignment, fine-tuning large Language Models like BERT, seems to be the goldstandard approach. However, regarding the Slot Filling problem, in most real-worldcases this method is not feasible due to the absence of human-annotated B-I-O tagsand it additionally performs poorly in few-shot settings, where there is a limitedset of labeled data. In order to overcome these limitations, we propose using anencoder-decoder approach that incorporates the concept of prompting. Specifically,we utilize the T5 Language Model along with natural language templateswhich the model is prompted to fill in with the relevant information. This methodachieves 98% intent accuracy and 95.9% slot micro-F1-score on the SNIPS dataset.More importantly, it provides substantial performance improvements in few-shotsettings and displays great adaptability to different intents and domains, whencompared to its counterpart that does not embody prompts.