ESSMArT Way to Manage User Requests

Quality and market acceptance of software products is strongly influenced by responsiveness to user requests. Once a request is received from a customer, decisions need to be made if the request should be escalated to the development team. Once escalated, the ticket must be formulated as a development task and be assigned to a developer. To make the process more efficient and reduce the time between receiving and escalating the user request, we aim to automate of the complete user request management process. We propose a holistic method called ESSMArT. The methods performs text summarization, predicts ticket escalation, creates the title and content of the ticket used by developers, and assigns the ticket to an available developer. We internally evaluated the method by 4,114 user tickets from Brightsquid and their secure health care communication plat- form Secure-Mail. We also perform an external evaluation on the usefulness of the approach. We found that supervised learning based on context specific data performs best for extractive summarization. For predicting escalation of tickets, Random Forest trained on a combination of conversation and extractive summarization is best with highest precision (of 0.9) and recall (of 0.55). From external evaluation we found that ESSMArT provides suggestions that are 71% aligned with human ones. Applying the prototype implementation to 315 user requests resulted in an average time reduction of 9.2 minutes per request. ESSMArT helps to make ticket management faster and with reduced effort for human experts. ESSMArT can help Brightsquid to (i) minimize the impact of staff turnover and (ii) shorten the cycle from an issue being reported to an assignment to a developer to fix it.

💡 Research Summary

The paper presents ESSMArT (Escalation and Summarization Automation), a holistic framework that automates the entire lifecycle of a user request in a software development environment: from the moment a request arrives, through summarization, escalation decision, ticket creation (title and body), priority assignment, and finally developer allocation. The authors motivate the work with a real‑world case study at Brightsquid, a secure health‑care communication provider, where the current process involves a CRM staff member manually summarizing requests, a manager deciding whether to escalate, and a project manager generating and assigning Jira tickets. Analysis of 4,114 historic tickets revealed that the escalation decision alone consumes 26.6 % of the total lead time, making it a clear bottleneck.

ESSMArT is composed of five sequential steps. Step 1 performs extractive summarization of the raw user request. The authors evaluate several representation techniques (topic‑based, indicator‑based) and train supervised classifiers on domain‑specific data, achieving the highest F1 score among state‑of‑the‑art baselines. Step 2 predicts whether the request should be escalated. Three classifiers—Naïve Bayes, Support Vector Machines, and Random Forest—are compared; Random Forest, fed with both the original conversation and the extractive summary, yields the best performance (precision = 0.90, recall = 0.55). Step 3 generates a concise ticket title using an abstractive summarization approach that extracts key phrases and maps them onto a predefined template. Step 4 constructs the ticket body by expanding the summary with structured sections (problem description, reproduction steps, expected outcome) via rule‑based text generation and a domain thesaurus. Step 5 assigns the ticket to an available developer by matching the new ticket’s metadata against historic tickets using a nearest‑neighbor algorithm that considers module, priority, and past assignee.

The internal evaluation on the 4,114 tickets confirms that domain‑specific supervised models outperform generic methods for both summarization and escalation prediction. The external evaluation involves 315 live user requests processed by 21 CRM experts and 33 project managers using a prototype implementation of ESSMArT. The generated tickets align with human‑crafted tickets 71 % of the time, and the end‑to‑end automation reduces average handling time by 9.2 minutes per request (approximately a 15 % reduction).

The authors acknowledge several limitations. The extractive summarizer, while accurate, still requires human verification for nuanced business requirements. The escalation classifier’s recall of 0.55 indicates that a substantial fraction of escalations may be missed, suggesting a need for cost‑sensitive learning or additional features. The developer allocation step optimizes only similarity to past tickets and does not simultaneously consider workload balance, expertise depth, or sprint constraints. Future work is proposed in three directions: (1) employing deep neural abstractive summarization models to improve title generation; (2) integrating cost‑sensitive or ensemble methods to boost escalation recall without sacrificing precision; and (3) applying reinforcement learning or multi‑objective optimization for dynamic, load‑aware developer assignment.

In conclusion, ESSMArT demonstrates that a carefully engineered combination of NLP and machine‑learning techniques can substantially streamline user request management in a real‑world software organization. By automating repetitive, time‑consuming steps, the framework not only shortens the response cycle but also mitigates the impact of staff turnover, offering a scalable solution especially valuable for startups and mid‑size firms seeking to maintain high customer satisfaction while keeping development resources focused on core product work.