Accurate and Scalable Matching of Translators to Displaced Persons for Overcoming Language Barriers

Divyansh Agarwal, Yuta Baba, Pratik Sachdeva, Tanya Tandon, Thomas Vetterli, Aziz AlghunaimarXiv 2020 · arXiv

Introduction

In humanitarian crises, language barriers often prevent aid workers from providing essential services to displaced persons. While bilingual volunteers are numerous, professional interpreters are scarce and expensive. Tarjimly addresses this gap by connecting approximately 20,000 volunteers with 30,000 users via a mobile app. The core technical challenge is scalability: how to notify a small, relevant subset of volunteers (typically 30-60) from a massive pool to ensure a rapid response without causing notification fatigue.

The Matching Pipeline

To solve this "needle in a haystack" problem, the researchers developed a lightweight matching pipeline centered on a logistic regression classifier.

The Tarjimly translator matching pipeline illustrates the flow from request to ranking and feedback.

The process begins with filtering the volunteer pool based on hard constraints such as language pairs, time zones, and specific requester preferences (e.g., gender). The remaining candidates are then ranked by the probability of responding "yes" to a notification. The top n translators are pinged, and the first to accept is matched. Logistic regression was chosen for its low latency in real-time production environments and its high interpretability, allowing the team to understand exactly which factors drive volunteer engagement.

Data and Feature Importance

The model was trained on an internal dataset of approximately 400,000 "pings." Despite a significant class imbalance—where only 6% of pings result in a "yes"—the model achieved high predictive power with an AUC of 0.91 on the test set. Analysis of the model reveals that historical behavior is the strongest predictor of future responsiveness.

Odds ratios for various translator features, highlighting the dominance of historical response rates.

As shown in the odds ratio analysis, the "overall response rate" is the most significant feature by several orders of magnitude, followed by "periodic response rate" (responsiveness at specific times of day). Profile features like formal experience or specific skills (medical, legal) also contribute but are less influential than past activity. To handle the "cold start" problem for new users, the system employs Bayesian-style smoothing, starting new users at a default response rate that adjusts as they interact with the platform.

Key Results and Impact

The implementation of this system has had a measurable impact on humanitarian aid delivery:

  • High Match Rate: 82% of all requests are successfully matched with a volunteer.
  • Rapid Response: The median response time is just 59 seconds, which is critical for aid workers in high-pressure environments.
  • Efficiency: The system maintains high accuracy while minimizing the number of pings sent, preserving the volunteer pool's long-term engagement.

Limitations and Future Directions

While effective, the current model risks "over-utilizing" a small group of highly responsive "super-volunteers," potentially leading to burnout. To address this, future iterations may incorporate an epsilon-greedy algorithm to "explore" the potential of newer or less active volunteers. Additionally, while the current system optimizes for responsiveness, future work will integrate user ratings to also account for translation quality. Finally, the authors emphasize the need for ongoing audits to ensure algorithmic fairness across different volunteer demographics.

Summary updated

BibTeX
@misc{agarwal2020accurate,
  author = {Divyansh Agarwal and Yuta Baba and Pratik Sachdeva and Tanya Tandon and Thomas Vetterli and Aziz Alghunaim},
  title = {{Accurate and Scalable Matching of Translators to Displaced Persons for Overcoming Language Barriers}},
  year = {2020},
  eprint = {2012.02595},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2012.02595},
}
Figure