Article Preview
TopIntroduction
With the change of people’s consumption concept and the release of consumption demand, the loan scale of Chinese financial institutions continues to expand. According to the data released by The People's Bank of China, the loan amount in China increased by 19.95 trillion yuan in 2021, with a year-on-year increase of 315 billion yuan. The huge loan scale has led to the explosive growth of business consulting volume, and manual customer service is facing great business pressure(Jindi Ai & Huiyu Liu, 2021). In addition, manual customer service consumes a lot of resources and has limited efficiency, while QA platforms obtain higher efficiency at a lower cost. Therefore, it makes sense to build a domain knowledge-based QA platform to reduce the pressure on manual customer service(Xin Peng, 2021).
In the field of loans, the rapid adjustment and update of products make loan data have complex relationships. Therefore, it is a great choice to store these data in KG(Yongzhi Zheng et al., 2022). Google put forward the concept of KG in November 2012, and then published a large-scale knowledge graph based on Freebase(Bollacker et al., 2008) and Wikipedia, which provides a reference for the storage of complex data(Singhal & others, 2012). KGs focus more on the construction and visualization of relations. Therefore, they can be used for knowledge reasoning and quickly mine new knowledge and relationships from entities and concepts(Yun et al., 2021). In the era of big data, data intensive industries are booming(Jaswal & Malhotra, 2022; Pawar et al., 2022). Using KGs to store and represent a large number of complex data provides strong support for management and utilization of data(Kar, 2022).
The QA technology based on KG has been studied for a long time abroad. QA platforms based on the semantic web, such as Aqualog(Lopez et al., 2005), Orakel(Cimiano et al., 2007) and Pythia, provided reference methods for constructing QA platforms based on KG. TBSL(Unger et al., 2012) proposed a template-based QA method. Although the template-based QA method consumes resources, it can well improve the accuracy of QA platforms. Although QA technology based on KG developed relatively late in China, there have been some achievements in recent years. For example, Ma et al. used MADDPG algorithm and TICC algorithm to construct a opponents’s behavior QA platform based on KG to infer the intention of opponent(Ma et al., 2021). Wu et al. used BiLSTM(Zhou et al., 2016) to construct a QA platform based on KG for suicide tendency detection, which can effectively answer questions related to psychological counseling(Shuzhao Wu et al., 2021). Zhang et al. used CRF and recommendation algorithm to construct an Android QA platform based on agricultural technology KG to facilitate farmers' work(Bokai Zhang & Xiang Li, 2021). In recent years, with the development of artificial machine learning and deep learning(Almomani et al., 2022; Sharma et al., 2022), many pre-training language models suitable for natural language processing have been born, such as ELMO(Peters et al., 2018), GPT(Radford et al., 2018), BERT(Devlin et al., 2018) and ERNIE(Sun et al., 2019). Domestic scholars also began to combine these pre-training language models with QA platforms(Fei Yuan, 2021). For example, Xu(Jinjin Xu, 2021) used BERT to construct a medical QA platform.