Loss-sensitive Training of Probabilistic Conditional Random Fields

We consider the problem of training probabilistic conditional random fields (CRFs) in the context of a task where performance is measured using a specific loss function. While maximum likelihood is the most common approach to training CRFs, it ignores the inherent structure of the task’s loss function. We describe alternatives to maximum likelihood which take that loss into account. These include a novel adaptation of a loss upper bound from the structured SVMs literature to the CRF context, as well as a new loss-inspired KL divergence objective which relies on the probabilistic nature of CRFs. These loss-sensitive objectives are compared to maximum likelihood using ranking as a benchmark task. This comparison confirms the importance of incorporating loss information in the probabilistic training of CRFs, with the loss-inspired KL outperforming all other objectives.

💡 Research Summary

This paper addresses a fundamental mismatch in Conditional Random Field (CRF) training: the most common objective, conditional maximum likelihood (ML), optimizes the model’s probability distribution without regard to the structured loss function that will ultimately be used to evaluate performance. To bridge this gap, the authors propose four loss‑sensitive training objectives that explicitly incorporate the loss into the probabilistic learning process.

1. Loss‑Augmented Maximum Likelihood (LA) – The energy function is modified by subtracting the loss, E_LA(y)=E(y)−ℓ_t(y). Plugging this into the usual log‑likelihood yields an objective that upper‑bounds the average loss. This mirrors the margin‑re‑scaling technique of structured SVMs but retains a probabilistic interpretation.

2. Loss‑Scaled Maximum Likelihood (LS) – Here the loss acts as a weight on the energy difference between a candidate y and the ground truth y_t: E_LS(y)=ℓ_t(y)