Enhancing In-vehicle Multi-turn Dialogue Text Rewriting with Lightweight Text Filtering

Deep learning has advanced intelligent vehicle development, improving communication and language comprehension. The Transformer model is currently the mainstream approach for text rewriting(coreference resolution) in vehicle dialogue systems, aiming to detect and rephrase sentences that require modification. Nonetheless, the Transformer model, consisting of a stack of Transformer blocks, is complex and extensive, making it inefficient for processing tasks. Accordingly, this paper proposes a proficient approach, denoted as FM-Transformer, for text rewriting of multi-round in-vehicle dialogues. The improvement idea is to add a lightweight text filtering module before the Transformer model, which can quickly classify whether the text of vehicle dialogues needs to be written or not, and only send the sentences that need to be modified into the Transformer model. After analysis, we adopt Fasttext as the lightweight text filtering module. Before and after the improvement of the model, we carried out a control experiment to verify the effectiveness of the FM converter by using a multi-turn vehicle voice conversation data set. The results showed positive results in terms of accuracy and time efficiency, highlighting the potential of FM converters to generate higher quality and more efficient response text.