Slot Recognition and Correction in Voice Robots: Methods, Models, and Experimental Results

Background In voice robots, slot (entity) extraction is essential for dialogue control, but ASR errors cause misrecognition of domain‑specific terms.

Slot Recognition Process The task is treated as a sequence labeling problem using IOB2/IOBES tags. Data annotation involves defining an ontology (e.g., car_brand), crawling candidate texts, initializing entities with a trie, and fine‑tuning using rasa‑nlu.

Model Training Two architectures were evaluated: IDCNN+CRF and BiLSTM+CRF. Experiments on 12,000 annotated utterances showed BiLSTM+CRF achieving higher accuracy and recall for both car brand and series slots.

Slot Correction Process Errors caused by homophones or similar pinyin are corrected by mapping recognized slots to a curated knowledge base of car brands and series. A pinyin‑based edit‑distance similarity algorithm, threshold filtering, and regex rules are applied to obtain the correct slot values.

Results After correction, the online system’s brand and series slot accuracy improved by 14.43 % and recall by 8.29 % in a used‑car follow‑up scenario.

Conclusion The paper demonstrates practical implementations of slot recognition and correction in voice bots and suggests future joint training of the two modules to further boost overall performance.