Multi-Perspective Dialogue Non-Quota Selection with loss monitoring for dialogue state tracking

The task-oriented dialogue system is dedicated to helping users achieve specific goals. Within this system, dialogue state tracking (DST) is the core module to obtain the key compressed information of the dialogue. In the field of DST, effective management of dialogue history assumes paramount importance. However, existing models consistently employ a uniform dialogue history throughout the process of tracking states, irrespective of the slot that is being updated. This may lead to insufficiency or redundancy of information for different slots. To address this issue, we introduce DNQS-DST, a novel approach intended for dynamically choosing the pertinent dialogue contents matching each slot to update the state. Our method operates by initially retrieving turn-level utterances from the dialogue history. Subsequently, it evaluates the relevance of these utterances to the target slot through a three-perspective evaluation process and then yields a non-quota dialogue selection. To fully exploit the effectiveness of the dialogue selection module, we also propose a Loss Monitoring-based Supervision Reinforcement module to achieve directly supervised training of the dialogue selection module without given labels. After that, only the chosen dialogue content is fed into the State Generator. Our proposed DNQS-DST, for the first time, breaks the limitation of using fixed dialogues when updating slots, which effectively filters out irrelevant information while preserving the integrity of the input data. In addition, the proposed supervision-strengthened module also offers a general solution for optimization schemes in all similar unlabeled stages. The experimental results demonstrate that this approach outperforms baseline models across mainstream datasets.¹