The goal of entity set expansion is deceptively simple: given a few "seed" examples of a category—such as Texas and Oregon—an algorithm should be able to scan a massive corpus of text and find all other members of the "U.S. States" category.
In theory, this works through a self-reinforcing loop called bootstrapping. You find patterns that the seeds share (e.g., "the governor of [X]"), use those patterns to find new entities (California, Florida), and then use those new entities to find even more patterns. However, in practice, this loop is fragile. A single mistake—like identifying Chicago as a state because it also appears in the pattern "the governor of [X]"—can cause the system to pivot. In the next iteration, it might start extracting cities, then perhaps countries, eventually losing the original category entirely.
Researchers call this phenomenon semantic drift. Across over a hundred papers, the central challenge is how to keep this iterative engine on the tracks.
The Geometry of Drift
To understand why drift happens, we can look at it through the lens of graph theory. Mamoru Komachi and Masashi Shimbo demonstrated in Graph-based Analysis of Semantic Drift that semantic drift is mathematically equivalent to "topic drift" in web search algorithms like HITS.
When an algorithm treats entities and patterns as nodes in a bipartite graph, drift occurs when the system wanders into "hubs"—patterns that are too generic. For example, the pattern "[X] is a place" is a hub that connects to almost everything. Once the bootstrapping process touches a hub, it loses its specificity. To counter this, researchers like Zeqiu Wu and Jiawei Han in SetExpan propose rigorous context-feature selection to penalize these overly generic patterns, ensuring that only the most "exclusive" features are used to pull in new candidates.
Establishing Semantic Boundaries
If drift is caused by a lack of focus, one solution is to define not just what a category is, but what it is not.
Tara McIntosh introduced the concept of negative categories in Unsupervised Discovery of Negative Categories. The intuition is that drift isn't random; it usually moves toward a coherent but incorrect category (e.g., from "States" to "Cities"). By automatically identifying these "drift-prone" clusters and treating them as negative evidence, the system creates a semantic boundary.
This idea was modernized by Lingyong Yan and Xianpei Han in BootstrapGAN. They framed bootstrapping as a game: a generator proposes new entities, while a sequence of discriminators learns to distinguish the true seeds from the increasingly difficult "fake" or drifted examples. This adversarial pressure forces the model to refine its category boundaries in every iteration.
Lookahead and Global Reasoning
Traditional bootstrapping is "short-sighted"—it picks the best-looking entity in the immediate next step without considering if that choice will lead to a dead end or a noisy cluster ten steps later.
To solve this, Lingyong Yan and colleagues applied Monte Carlo Tree Search (MCTS) to the expansion process in Learning to Bootstrap for Entity Set Expansion. By simulating multiple steps into the future, the system can estimate the "delayed feedback" of adding a specific entity. If adding Chicago eventually leads to a flood of non-state entities in future simulations, the system can reject it now, even if it looks like a strong candidate in the short term.
Similarly, the Global Bootstrapping Neural Network (GBN) moves away from local pattern matching toward global corpus semantics. By using a Graph Attention Network (GAT), the model considers the entire structure of how entities and patterns relate across the whole dataset, making it much harder for a single noisy instance to corrupt the process.
Guidance from Pre-trained Knowledge
Recent work has shifted toward using the latent knowledge already stored in Large Language Models (LLMs) to anchor the bootstrapping process.
In CGExpan, Yunyi Zhang and Jiaming Shen use a language model to "probe" for a class name. If the seeds are Texas and Oregon, the system asks the LLM to name the category (e.g., "U.S. States"). Once the category has a name, it acts as a permanent semantic anchor. Every new candidate must not only match the patterns in the text but also be a plausible member of the named category "U.S. States" according to the LLM.
Robustness through Forgetting
Finally, some researchers argue that since some drift is inevitable, we need mechanisms to "wash away" the noise. Payam Karisani introduced Robust Self-Training (RST) in Neural Networks Against (and For) Self-Training, which leverages "catastrophic forgetting." By periodically resetting parts of the model or using specific uncertainty metrics, the system can effectively forget the noisy, low-confidence labels it learned during a drifted phase, allowing the core semantic signal from the original seeds to prevail.