农业病虫害知识问答意图识别与槽位填充联合模型研究

Joint Intent Detection and Slot Filling of Knowledge Question Answering for Agricultural Diseases and Pests

农业病虫害领域的意图识别和槽位填充研究仍处于起步阶段,除语料严重匮乏外,还面临任务相互独立、忽略彼此相关性和未充分利用意图嵌入信息等问题。为此,提出了一种基于意图嵌入信息和槽位门控机制的意图识别与槽-位填充联合模型(AgIG-IDSF)。首先,该模型在共享编码模块引入了注意力机制用于丰富上下文语义特征;其次,提出了一种融合意图嵌入表示和槽位门控机制的意图-槽位交互方法用以增强意图信息指导槽位填充任务的能力,进而提高模型的整体识别性能。在包含22个意图类别、10个槽位类别和11 976条标注样本的自构建语料上进行了实验。结果表明,在该语料上AgIG-IDSF模型的意图识别准确率为94.41%,槽位填充F1值为94.01%,整体识别准确率高达88.07%,显著优于包含双向关联模型在内的多种基准模型,表明了该模型在识别农业病虫害意图与槽位方面的有效性。此外,在公共数据集上的实验结果还表明了该模型具有一定的泛化能力。

Joint intent detection and slot filling plays an important part in natural language understanding for knowledge answering, but it is still in its infancy in the field of agricultural diseases and pests. In addition to lack of corpus, it also faces several challenges such as task independence, mutual correlations ignoring, and intent embedded information neglection. To address the above questions, a novel joint intent detection and slot filling model based on intent embedding and slot-gated mechanism, named AgIG-IDSF was proposed. Firstly, the attention mechanism, which could extract the context-aware features, was introduced into shared encoder to further enrich the contextual semantic features. Secondly, an intent-slot interaction method with intent embedding and the slot-gated mechanism was designed to enhance the ability of intent detection to guide the slot filling task. Finally, the comparative experiments from various aspects were conducted on a self-constructed corpus named AGIS, which mainly contained 22 intent categories, 10 slot categories, and 11 976 annotated samples. The experimental results showed that AgIG-IDSF achieved the intent detection accuracy of 94.41%, slot filling F1-score of 94.01%, and overall semantic accuracy of 88.07% on the self-constructed corpus, which were significantly better than a variety of benchmark models, including bidirectional mutual models. It demonstrated the effectiveness of AgIG-IDSF in jointly identifying the intent and slots in the field of agricultural diseases and pests. In addition, the experimental results on public datasets, i.e., ATIS and SNIPS also showed that the model had a certain generalization ability.