Abstract: This study tackles literal to metaphorical sentence generation, presenting a framework that can potentially lead to the production of an infinite number of new metaphors. To achieve this goal, we propose a complete workflow that tackles metaphorical sentence classification and metaphor reconstruction. Unlike similar research works regarding metaphor generation, our approach does not require any customor closed-source model, hence with this work we introduce a complete literal to metaphorical open-source model. The obtained results show that a good ratio of originally literal sentences, coming from different data sources and topics, are turned to metaphorical. Human evaluation shows that our constructed metaphors are considered more fluent, creative and metaphorical than figurative statements created by a real person. Furthermore, by using our artificial data to increase the training size of a metaphorical sentence classification dataset, we register an improvement of 3% over the baseline.
0 Replies
Loading