Search-based Structured Prediction

We present Searn, an algorithm for integrating search and learning to solve complex structured prediction problems such as those that occur in natural language, speech, computational biology, and vision. Searn is a meta-algorithm that transforms these complex problems into simple classification problems to which any binary classifier may be applied. Unlike current algorithms for structured learning that require decomposition of both the loss function and the feature functions over the predicted structure, Searn is able to learn prediction functions for any loss function and any class of features. Moreover, Searn comes with a strong, natural theoretical guarantee: good performance on the derived classification problems implies good performance on the structured prediction problem.

💡 Research Summary

The paper introduces Searn (Search‑based Structured Prediction), a meta‑algorithm that unifies search and learning to tackle complex structured prediction tasks across natural language processing, speech, computational biology, and computer vision. Traditional structured learning methods such as Conditional Random Fields (CRFs), structured SVMs, or perceptron‑based approaches require the loss function and feature functions to be decomposed over the output structure. This decomposition is often impossible or highly inconvenient when the loss is non‑decomposable (e.g., BLEU score) or when features capture global properties of the output. Searn eliminates this restriction by reducing any structured prediction problem to a series of simple binary classification problems, allowing any off‑the‑shelf binary classifier to be plugged in.

Core Idea

Searn treats the construction of a structured output as a sequential decision‑making process. At each step the algorithm is in a state (the partially built structure) and must choose an action (the next token, label, or sub‑structure). The choice of action is cast as a binary classification problem: given a feature representation of the current state, predict whether a particular action is optimal. The optimality of an action is measured by the cost obtained from a roll‑out: after taking the action, the current policy is used to complete the rest of the structure, and the total loss of the completed output is recorded. By enumerating all possible actions from a state, Searn obtains a cost for each and treats the minimum‑cost action as the correct label for training.

Learning Procedure

1. Initialize a policy – often a random or heuristic policy.
2. Generate training examples – For each training instance, simulate the current policy to produce a trajectory of states. At each state, perform roll‑outs for all admissible actions, compute their costs, and create a binary example (state features, “action is optimal” vs. “action is not optimal”).
3. Train a binary classifier on the accumulated examples.
4. Update the policy by mixing the newly trained classifier with the old policy (e.g., ε‑greedy mixing).
5. Iterate the above steps for several rounds. As the policy improves, roll‑out costs become more accurate, and the classifier receives higher‑quality training data.

Theoretical Guarantees

The authors prove a regret bound: let T be the number of decision steps, ε the classification error of the learned binary classifier, and β the mixing probability with the old policy. Then the expected structured loss L of the final policy satisfies

\