Using Local Alignments for Relation Recognition

This paper discusses the problem of marrying structural similarity with semantic relatedness for Information Extraction from text. Aiming at accurate recognition of relations, we introduce local alignment kernels and explore various possibilities of using them for this task. We give a definition of a local alignment (LA) kernel based on the Smith-Waterman score as a sequence similarity measure and proceed with a range of possibilities for computing similarity between elements of sequences. We show how distributional similarity measures obtained from unlabeled data can be incorporated into the learning task as semantic knowledge. Our experiments suggest that the LA kernel yields promising results on various biomedical corpora outperforming two baselines by a large margin. Additional series of experiments have been conducted on the data sets of seven general relation types, where the performance of the LA kernel is comparable to the current state-of-the-art results.

💡 Research Summary

The paper tackles the long‑standing challenge of relation recognition in information extraction: how to jointly exploit structural similarity (the way entities and their contexts are organized) and semantic relatedness (the meaning conveyed by the words). Traditional approaches either focus on syntactic patterns—such as dependency‑tree kernels or handcrafted regular expressions—or on semantic similarity derived from lexical resources or distributional models, but they rarely integrate the two sources in a principled manner.

To bridge this gap, the authors introduce a Local Alignment (LA) kernel inspired by the Smith‑Waterman algorithm, a classic method for finding optimal local alignments between biological sequences. In the textual setting, each candidate relation is represented as a sequence of tokens (or token embeddings) extracted from the sentence that contains the two target entities. The LA kernel computes a similarity score by summing over all possible local alignments between two sequences, weighting each alignment with an exponential of the alignment score. Crucially, the alignment score (s(i,j)) is not a simple character match; it is defined as a semantic similarity between the i‑th token of the first sequence and the j‑th token of the second sequence. This similarity can be any distributional measure—pointwise mutual information (PPMI), latent semantic analysis (LSA), word2vec, GloVe, or more recent contextual embeddings—trained on large unlabeled corpora. By doing so, the kernel rewards alignments that pair semantically related words even when their surface forms differ, thereby capturing “soft” structural matches that traditional string‑based kernels miss.

Formally, the kernel is expressed as

\