[Paper Review] Fine-Grained Entity Typing with High-Multiplicity Assignments

Maxim Rabinovich, Dan Klein|arXiv (Cornell University)|Apr 25, 2017

Natural Language Processing Techniques8 references3 citations

TL;DR

This paper proposes a structured prediction approach for fine-grained entity typing in high-multiplicity settings, using set-valued type predictions with features capturing type correlations. Evaluated on a new Wikipedia-based corpus with rich, denoised type assignments, the model outperforms unstructured baselines by explicitly modeling type sets, achieving a macro F1 of 54.5 with full feature integration.

ABSTRACT

As entity type systems become richer and more fine-grained, we expect the number of types assigned to a given entity to increase. However, most fine-grained typing work has focused on datasets that exhibit a low degree of type multiplicity. In this paper, we consider the high-multiplicity regime inherent in data sources such as Wikipedia that have semi-open type systems. We introduce a set-prediction approach to this problem and show that our model outperforms unstructured baselines on a new Wikipedia-based fine-grained typing corpus.

Motivation & Objective

• To address the lack of realistic, high-multiplicity entity typing datasets in existing research, particularly those reflecting real-world knowledge resources like Wikipedia.
• To model fine-grained entity typing as a structured set-prediction problem rather than a multilabel classification task.
• To improve performance on high-multiplicity entity typing by incorporating type correlations and set-level features.
• To construct and release a new, denoised Wikipedia-based corpus with high-multiplicity type assignments for benchmarking.

Proposed method

• The model frames entity typing as structured prediction over sets of types, using a linear model with a weight vector w to score candidate type sets T.
• It employs a feature function φ(x, e, T) that combines entity-specific features, pairwise type features, graph-based features (e.g., co-occurrence or hierarchy), and set size.
• The model uses greedy decoding to search over type sets, iteratively adding types that improve the score, with optional connectivity constraints to reduce search space.
• A key innovation is the use of projected types—syntactic heads of Wikipedia categories—to reduce sparsity and improve generalization.
• The approach allows for efficient inference by restricting candidate sets to connected components in a type constraint graph.
• The model is trained end-to-end using stochastic gradient descent on the structured prediction objective, with a margin-based loss function.

Experimental results

Research questions

• RQ1 Can structured prediction over sets of types improve performance on fine-grained entity typing in high-multiplicity settings compared to unstructured baselines?
• RQ2 How effective are set-level features such as pairwise type interactions and graph structure in modeling type correlations?
• RQ3 Does using a semi-open, Wikipedia-derived type system with high multiplicity yield better generalization than coarsened Freebase-type systems?
• RQ4 To what extent do performance gains from structured modeling persist across rare and frequent types in high-multiplicity settings?
• RQ5 Can type projection via WordNet heads improve model robustness and reduce sparsity in real-world entity typing?

Key findings

• The structured prediction model with full feature integration (pairwise, graph, and set size) achieved a macro F1 of 54.5 on the entity-level task, outperforming the unstructured baseline (52.9).
• Adding pairwise type features led to a significant performance gain, while graph-based features provided minimal improvement, suggesting pairwise correlations may suffice for moderate type set sizes.
• The model maintained strong performance on rare types, as shown by micro-averaged F1 scores that remained high even for low-frequency types.
• The use of projected types—syntactic heads of Wikipedia categories—enabled better generalization and reduced sparsity in the type space.
• The corpus construction process successfully created a realistic, high-multiplicity dataset by denoising Wikipedia categories using WordNet and hierarchical constraints.
• Greedy decoding with connectivity constraints reduced search space without sacrificing performance, enabling scalable inference on large type sets.