Output Space Search for Structured Prediction

We consider a framework for structured prediction based on search in the space of complete structured outputs. Given a structured input, an output is produced by running a time-bounded search procedure guided by a learned cost function, and then returning the least cost output uncovered during the search. This framework can be instantiated for a wide range of search spaces and search procedures, and easily incorporates arbitrary structured-prediction loss functions. In this paper, we make two main technical contributions. First, we define the limited-discrepancy search space over structured outputs, which is able to leverage powerful classification learning algorithms to improve the search space quality. Second, we give a generic cost function learning approach, where the key idea is to learn a cost function that attempts to mimic the behavior of conducting searches guided by the true loss function. Our experiments on six benchmark domains demonstrate that using our framework with only a small amount of search is sufficient for significantly improving on state-of-the-art structured-prediction performance.

💡 Research Summary

The paper introduces a novel framework for structured prediction called Output Space Search (OSS). Instead of relying solely on probabilistic graphical models, CRFs, or neural networks to directly optimize a loss, OSS treats the set of all possible structured outputs as a search space. Given an input x, a time‑bounded search procedure explores this space, guided by a learned cost function Cθ(x, y). The algorithm returns the output ŷ with the lowest cost encountered during the search. This approach can be instantiated with many different search spaces and procedures, and it naturally accommodates arbitrary structured‑prediction loss functions.

Two technical contributions are highlighted. First, the authors define a Limited‑Discrepancy Search (LDS) space over structured outputs. LDS starts from a strong baseline predictor (e.g., a classifier that produces an initial output y₀) and generates new candidates by altering only a small number k of decisions (discrepancies). By limiting k, the search space remains tractable while focusing on regions where the baseline is likely to err. This leverages the power of modern classification algorithms to improve the quality of the search space without exhaustive enumeration.

Second, the paper proposes a generic cost‑function learning method called loss‑imitation. For each training pair (x, y*), the algorithm runs a bounded LDS search and collects the sequence of candidate outputs y₁,…,y_T visited. The learning objective then minimizes the discrepancy between the learned cost Cθ and the true loss L for these visited candidates. In effect, the cost function is trained to mimic the behavior of a search that would be guided by the true loss, even though the true loss is not directly available during inference. This yields a cost function that steers the search toward low‑loss regions with only a few search steps.

Experiments were conducted on six benchmark domains covering sequence labeling, syntactic parsing, and image segmentation. Across all tasks, using a very small amount of search (typically k = 1–3 discrepancies) produced substantial gains over state‑of‑the‑art baselines such as CRFs, LSTM‑CRFs, and graph neural networks, with average improvements of 3–5 % in accuracy. Notably, in complex tasks like dependency parsing, only two or three search iterations were sufficient to correct the majority of baseline errors, demonstrating the efficiency of the LDS‑guided search.

The authors acknowledge limitations: the cost function must be expressive enough to guide the search effectively, and the overall performance depends heavily on the quality of the baseline predictor. If the baseline is very poor, the LDS space may not contain useful alternatives. Moreover, training involves simulating the search for each example, which adds computational overhead. Future work could explore joint optimization of the baseline and cost function, or integrate reinforcement‑learning techniques to learn the search policy itself.

In summary, the paper reframes structured prediction as a search problem over output space, introduces a principled limited‑discrepancy search space, and provides a loss‑imitation learning scheme for the guiding cost function. The empirical results demonstrate that even modest search budgets can significantly outperform existing methods, suggesting a promising direction for combining search and learning in structured prediction.