Comparison of Error Correction and Extraction Approaches

We compare different approaches for error correction detection and error correction. For the error correction detection task, the inputs are the last two utterances of a user and the output is whether there is an error correction in the last utterance. The error correction task gets the same inputs, but the output is the correction of the second last utterance according to the error correction in the last utterance and the extracted pairs of reparandum and repair entity. There are two advantages when using the compared approaches as utility component for a dialog system. It can be avoided to collect corrections for every new domain, and the extraction of the reparandum and repair pairs offers the possibility to learn from them. As benchmark for our comparison, we use an adapted version of the EPIC-KITCHENS-100 dataset. The best approach, a pipeline approach with a fine-tuned sequence labeling BERT model for error correction detection and a fine-tuned sequence-to-sequence T5 model for error correction, has an accuracy of 96.40% on synthetic validation data and an accuracy of 77.81% on human-created real-world test data.