Introduction
Relation Extraction (RE) is a fundamental task in natural language processing, aimed at identifying semantic relationships between entities in text. For instance, recognizing that "Steve Jobs co-founded Apple" implies a founder relationship. While effective, supervised RE requires vast amounts of manually labeled data. To address this, Distant Supervision (DS) automatically generates labels by aligning text with knowledge graphs.
However, DS introduces significant noise. A sentence like "I spent my summer vacation in Paris, France" might be incorrectly labeled with a capital_of relation simply because a knowledge graph contains that fact, even if the sentence doesn't express it. Current Multi-Instance Learning (MIL) strategies, such as averaging (AVG), attention (ATT), or selecting the single best sentence (ONE), often fail because they treat sentences as independent entities, ignoring the intrinsic associations within a "bag" of sentences.
Methods: The HFMRE Framework
The proposed HFMRE (Huffman Tree for Relation Extraction) framework shifts from simple aggregation to an iterative denoising process that leverages sentence consensus.
Sentence Encoding and Circular Cosine Similarity (CCS)
Sentences are first transformed into high-dimensional vectors using a Pre-trained Language Model (PLM) like BERT. The core innovation is Circular Cosine Similarity (CCS). Unlike standard attention mechanisms that compare sentences to an external query, CCS calculates the similarity of each sentence to every other sentence in the bag. Sentences that are similar to many others are considered more likely to express the true relation, while outliers are treated as noise.
Huffman Tree Construction and Denoising
HFMRE treats sentences as leaf nodes in a Huffman tree, with their CCS scores serving as weights. The algorithm iteratively merges the two nodes with the lowest weights—the most probable noise. During merging, a reduction factor ($\alpha$) is applied to the lower-weight node to dampen its influence. This process continues until a single root node, the "excellent instance," is formed, representing the distilled features of the entire bag.
Experimental Results
The model was evaluated on three major benchmarks: NYT10d, NYT10m, and Wiki20m.
Performance Metrics
HFMRE achieved state-of-the-art results on the NYT10d dataset, reaching an Area Under the Curve (AUC) of 54.0, a 3.2-point improvement over the previous best model (CIL). It also demonstrated high reliability with a Precision at M (P@M) score of 86.4.
Ablation Studies
Ablation tests confirmed the importance of the Huffman tree structure. Removing this component led to a significant performance drop (e.g., 12.2 points in AUC on NYT10m), indicating that the tree-based denoising is the primary driver of the model's success, rather than the choice of encoder alone.
Conclusion and Future Work
HFMRE provides a robust method for denoising distantly supervised data by identifying "excellent instances" through consensus. While the current implementation faces some computational overhead for very large bags, it offers a general-purpose improvement for MIL tasks. Future research may explore multilingual applications and the simultaneous construction of positive and negative trees to further enhance discriminative power.
Summary updated