Home - MJB90/ConversationLabelling GitHub Wiki

Customer engagement analysis through social media has become an important component in customer care analysis for tracking the engagement and measure the effectiveness of agents in addressing the issues raised by customer on online social media such as Twitter. Customers interact with agents to express their grievances, opinions, and inquiries about new products on social media. Agents then engage the customers through online conversations which may continue for a long period of time where the agent may get back to customers after collecting some information. An agent simultaneously engages multiple customers and should come back only to those customers whose conversations are still open. Issues which have been successfully addressed, an agent need not spend time on checking those and should directly take up by the issues which are still pending. An automated algorithm for categorizing conversations can save a significant amount of time and effort on the agent’s side and thus enhancing the productivity and customer satisfaction. Conversation labeling is a generic problem and spans across different vertical of customer care where the content and formats of conversation may vary such as retail, finance, telecom domains. To categorize the conversations, classifiers are built in a supervised manner using training data from the domain of interest. These algorithms have an underlying assumption that the conversations they will encounter for categorization follow a similar/same distribution as that of the conversations on which these models were trained. However, this is a very restricted assumption and does not generally hold true in real applications across different domains. The basic style of conversations may remain similar, however, the content, topic and use of vocabulary may be specific to different domains due to which the model built for a specific domain does not yield desired results across different domains for the same task. For example, a conversation analyzer trained for labeling conversations in a telecom domain may not yield acceptable results for labelling conversations in a financial domain.