Conversational Document Prediction to Assist Customer Care Agents

Abstract

A frequent pattern in customer care conversations is the agents responding with appropriate webpage URLs that address users' needs. We study the task of predicting the documents that customer care agents can use to facilitate users' needs. We also introduce a new public dataset 1 which supports the aforementioned problem. Using this dataset and two others, we investigate state-of-the-art deep learning (DL) and information retrieval (IR) models for the task. We also analyze the practicality of such systems in terms of inference time complexity. Our results show that an hybrid IR+DL approach provides the best of both worlds.

Conversational Search with Mixed-Initiative - Asking Good Clarification Questions backed-up by Passage Retrieval

“…using a disjunctive query over all words in the conversation C j . Following (Ganhotra et al, 2020), we treat the dialog query as a verbose query and apply the Fixed-Point (FP) method (Paik and Oard, 2014) for weighting its words. Yet, compared to "traditional" verbose queries, dialogs are further segmented into distinct utterances.…”

Q2R: A Query-to-Resolution System for Natural-Language Queries

“…In terms of ensembling multiple document retrieval approaches, the recent focus has been on the computational cost, where faster techniques are used to pre-filter the large document pool, to be re-ranked by computationally more expensive but more accurate techniques. A good example is the work by Ganhotra et al (2020), which combines a series of traditional IR techniques with neural approaches.…”

“…We evaluate the proposed method along with baseline methods on datasets in which labeled examples are available. 1 In particular, we use the Twitter and Telco datasets described in Ganhotra et al (2020) along with an IBM dataset.…”

Conversational Search with Mixed-Initiative -- Asking Good Clarification Questions backed-up by Passage Retrieval

“…Given a conversation context C j , Passage retrieval is performed in two steps. First, top-k documents are retrieved using the conversation context as a verbose query (Ganhotra et al, 2020). Next, candidate passages are extracted from those top-k documents using a sliding window of fixed size with some overlap.…”

“…For document retrieval, we create a disjunctive query from all words in the conversation C j . Following (Ganhotra et al, 2020), we treat the dialog query as a verbose query and apply the Fixed-Point (FP) method (Paik and Oard, 2014) for weighting its words. Yet, compared to "traditional" verbose queries, dialogs are further segmented into distinct utterances.…”