CARE: A Clue-guided Assistant for CSRs to Read User Manuals

TL;DR

CARE introduces a clue-guided reading assistant that enables CSRs to read and respond from information-rich user manuals with explicit explanation paths. By representing manuals as heterogeneous graphs and aligning questions to a question clue node, CARE performs joint procedural and factual inference via beam-search-driven clue chains and adaptive convolution scoring. Self-supervised data construction addresses limited labeled data, enabling robust reasoning without extensive annotation. In offline and online evaluations, CARE outperforms baselines on standard metrics and significantly reduces CSR reading time while preserving high service quality, indicating strong practical value for online customer service. The work emphasizes explainability and safety by making inference chains explicit, which helps CSRs verify and act on model predictions.

Abstract

It is time-saving to build a reading assistant for customer service representations (CSRs) when reading user manuals, especially information-rich ones. Current solutions don't fit the online custom service scenarios well due to the lack of attention to user questions and possible responses. Hence, we propose to develop a time-saving and careful reading assistant for CSRs, named CARE. It can help the CSRs quickly find proper responses from the user manuals via explicit clue chains. Specifically, each of the clue chains is formed by inferring over the user manuals, starting from the question clue aligned with the user question and ending at a possible response. To overcome the shortage of supervised data, we adopt the self-supervised strategy for model learning. The offline experiment shows that CARE is efficient in automatically inferring accurate responses from the user manual. The online experiment further demonstrates the superiority of CARE to reduce CSRs' reading burden and keep high service quality, in particular with >35% decrease in time spent and keeping a >0.75 ICC score.

Figures (11)

• Figure 1: Illustrations of (a) a log of online custom service, where the user prefers to chat with CSR; and (b) the proposed clue-guided reading assistant (CARE), which provides clues explaining how to arrive at proper responses.
• Figure 2: Part of the heterogeneous graph derived from the user manual in Figure \ref{['fig:assistance']}. Each action on the graph is decorated with corresponding arguments, and the arguments possessed by each entity and the sub-entity relations between different entities are clearly represented by corresponding links. Best viewed in color.
• Figure 3: The backbone model of CARE, which first aligns the user question to the graph of the user manual and then infers procedural and factoid clues over the graph from the question clue node.
• Figure 4: The construction of a sample for self-supervised training.
• Figure 5: Results of the ablation study