Publications

Abstract: We present a novel text-to-text generative approach for complex word identification that leverages the power of transformer-based language models. Our method reformulates the task as a sequence-to-sequence problem, enabling the model to identify and suggest simpler alternatives for complex words in context. This approach significantly outperforms traditional classification methods on standard benchmarks...

Abstract: We introduce FABSA (Feedback ABSA), a large-scale aspect-based sentiment analysis dataset comprising over 10,500 user reviews across 10 different domains. Unlike existing datasets, FABSA provides fine-grained aspect annotations with hierarchical category structures, enabling more sophisticated analysis of customer feedback. Our dataset addresses the critical need for cross-domain evaluation in ABSA systems and includes comprehensive annotations for aspect terms, aspect categories, and their associated sentiments...

Abstract: We propose Control Prefixes, a novel parameter-efficient method for controllable text generation that enables fine-grained control over multiple attributes simultaneously. Our approach extends prefix tuning by introducing attribute-specific control codes that can be composed dynamically at inference time. This allows for flexible control without requiring separate models for each attribute combination...