A Discrete Hard EM Approach for Weakly Supervised Question Answering

Sewon Min, Danqi Chen, Hannaneh Hajishirzi, Luke Zettlemoyer

Abstract

Sewon Min, Danqi Chen, Hannaneh Hajishirzi, Luke Zettlemoyer. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP).

2019.

Introduction: The Challenge of Spurious Solutions

In many modern question answering (QA) tasks, models are trained under weak supervision. This means we know the final answer (e.g., "Robert Schumann" or "4"), but we do not have the gold-standard "derivation" or "solution path" that leads to that answer.

This lack of intermediate supervision leads to the "spurious solution" problem. For instance, in multi-mention reading comprehension, a correct answer might appear multiple times in a text, but only one specific mention is the logical response to the question. Similarly, in discrete reasoning tasks like DROP, a numeric answer like "4" could be reached through the correct logic (e.g., "40 - 36") or a coincidental, spurious calculation (e.g., "41 - 37").

Examples of spurious solutions in reading comprehension and discrete reasoning tasks.

Standard training methods often rely on Maximum Marginal Likelihood (MML), which sums the probabilities of all paths leading to the correct answer. However, MML can inadvertently encourage the model to distribute its probability mass across many paths, including these spurious ones, rather than identifying the single logically correct derivation.

Methodology: Discrete Hard EM

To address these issues, the authors propose a Hard EM (Expectation-Maximization) training scheme that treats the solution path as a discrete latent variable. The approach involves two primary stages:

  1. Precomputing Solution Sets ($Z$): For every training instance, the system precomputes a set $Z$ containing all possible solutions (such as text spans, arithmetic equations, or SQL queries) that yield the correct final answer.
  2. Hard EM Learning: Instead of averaging over all solutions in $Z$, the model identifies the most likely solution $\tilde{z}$ according to its current parameters. It then updates its parameters to maximize the likelihood of only that specific solution.

To prevent the model from prematurely committing to a sub-optimal spurious solution (the "cold start" problem), the authors employ an annealing schedule. Training begins with the standard MML objective and gradually transitions to the Hard EM objective.

Robustness to Solution Set Size

A key advantage of the Hard EM approach is its robustness as the complexity of the search space increases. In scenarios where the set of possible solutions $Z$ is large, MML's performance typically degrades because it attempts to account for an increasing amount of "noise" from spurious paths. In contrast, Hard EM remains stable by focusing on the single most probable path.

Accuracy comparison between Hard EM (Ours) and MML as the number of possible solutions (|Z|) increases on the DROP dataset.

Experimental Results

The researchers evaluated this method across six major datasets spanning reading comprehension (TriviaQA, NarrativeQA), discrete reasoning (DROP), and semantic parsing (WikiSQL).

  • State-of-the-Art Performance: The Hard EM approach achieved SOTA results on 5 out of 6 benchmarks at the time of publication, providing absolute gains of 2–10% over MML and other common heuristics.
  • WikiSQL Breakthrough: On the WikiSQL dataset, the method outperformed previous reinforcement learning (RL) approaches by a significant 13% absolute margin. This suggests that precomputing solution sets is often a more effective strategy than searching for them via RL.
  • Qualitative Accuracy: Manual error analysis revealed that when the model predicted the correct answer, it used the correct logical derivation in 88-98% of cases, effectively minimizing the impact of spurious solutions.

Conclusion

The Discrete Hard EM approach offers a simple yet powerful alternative to MML for weakly supervised QA. By enforcing a "single correct path" prior and aligning training more closely with inference, it achieves superior accuracy and robustness. Future work may involve extending this logic to continuous latent spaces or developing more sophisticated methods for enumerating solution sets in complex program synthesis tasks.

Summary updated

BibTeX
@inproceedings{min2019discrete,
  author = {Sewon Min and Danqi Chen and Hannaneh Hajishirzi and Luke Zettlemoyer},
  title = {{A Discrete Hard EM Approach for Weakly Supervised Question Answering}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2019},
  url = {https://openreview.net/forum?id=_yAB4rKzRY},
}
Figure