Sentence encoding for Dialogue Act classification

Abstract In this study, we investigate the process of generating single-sentence representations for purpose Dialogue Act (DA) classification, including several aspects text pre-processing and input representation which are often overlooked or underreported within literature, example, number words to keep in vocabulary sequences. We assess each these with respect two DA-labelled corpora, using a range supervised models, represent those most frequently applied task. Additionally, compare context-free word embedding models that transfer learning via pre-trained language based on transformer architecture, such as Bidirectional Encoder Representations from Transformers (BERT) XLNET, have thus far not been widely explored DA classification Our findings indicate considerations do statistically significant effect accuracy. Notably, found viable sequence lengths, sizes, can be much smaller than is typically used experiments, yielding no improvements beyond certain thresholds. also show some cases contextual sentence generated by reliably outperform methods. Though BERT, its derivative improvement over approaches, previous work classification.