Hierarchy Builder: Organizing Textual Spans into a Hierarchy to Facilitate Navigation

Itay Yair, Hillel Taub-Tabib, Yoav Goldberg

Introduction

Information extraction (IE) systems have become remarkably proficient at scanning vast corpora to identify relevant snippets. However, this success often results in a new bottleneck: the "flat list" problem. When a query for medical causes or symptoms returns thousands of individual strings, researchers face an overwhelming task of manual categorization. Hierarchy Builder addresses this by transforming flat, redundant lists into navigable Directed Acyclic Graphs (DAGs), allowing for more efficient exploratory search.

Traditional IE systems typically rank results by frequency. While useful for identifying common occurrences, this approach buries rare but significant details and fails to group semantically related concepts. For example, a search for "sciatica" causes might return "herniated disc," "disc herniation," and "lumbar disc protrusion" as separate, disconnected entries. This redundancy increases the cognitive load on users attempting exploratory search—a scenario where the user needs a high-level overview of the data's facets rather than a single specific answer. Hierarchy Builder aims to mirror how humans naturally categorize information by inducing a logical structure over these textual spans.

Methodology: The Hierarchy Construction Pipeline

The system employs a six-stage pipeline to construct a DAG. Unlike a strict tree structure, a DAG allows a single concept to have multiple parent categories (e.g., "brain tumor" can be categorized under both "brain" and "tumor"), which more accurately reflects complex medical relationships.

  1. Expansion & Grouping: The system decomposes long phrases into smaller sub-strings and groups them into "equivalence sets" using lemmas and medical thesauruses (UMLS/WordNet). This ensures that variations like "herniated disc" and "disc herniation" are treated as a single node.
  2. Initial DAG Construction: Edges are created based on lexical specificity. If string B contains all the words of string A (plus additional modifiers), B is positioned as a more specific child of A.
  3. Semantic Merging: Using SAP-BERT, a specialized language model for medicine, the system identifies and merges nodes that are semantically identical but linguistically distinct, such as "administration of insulin" and "insulin injection."
  4. Taxonomic Enrichment: The system integrates high-level abstract categories from the Unified Medical Language System (UMLS). This provides the "glue" to group disparate strings; for instance, it might add "cardiovascular disease" as a parent node to group "ischemia" and "hypotension" even if that specific phrase never appeared in the source text.
  5. Pruning: To prevent the graph from becoming an unreadable "hairball," a Greedy Set Cover algorithm removes redundant edges and nodes. This ensures every original input string remains reachable through the most compact and informative path possible.
  6. Entry-Point Selection: Finally, the system selects the top-k most representative "root" nodes to serve as a clean "table of contents" for the user.

Evaluation and Results

The researchers evaluated the system in the biomedical domain using over 3,000 strings related to the causes of jaundice and chest pain extracted from PubMed abstracts. Medical experts were tasked with finding known etiologies within the generated hierarchy versus a standard frequency-ranked flat list, using gold standards from the NCBI and WebMD as benchmarks.

  • Efficiency: Experts located information significantly faster using the DAG. In flat lists, many important causes were buried hundreds of items deep; in the hierarchy, they were often accessible within a few clicks.
  • Coverage: For the "jaundice" dataset, 28 out of 38 known causes were present in the extracted data, and 25 of those were easily reachable through the top 50 entry points of the hierarchy.
  • Quality: Human evaluators rated the logical paths in the hierarchy very highly, with most paths receiving a 5/5 score for making sense to a medical professional.

Discussion and Future Work

A key insight from this research is that hierarchy is contextual; the system's pruning stage ensures that categories are only introduced if they actually help organize the specific data at hand. Furthermore, the use of a DAG over a tree structure proved essential for intuitive navigation.

However, limitations remain. The current implementation is heavily reliant on the UMLS, making it domain-specific to medicine. Future work is required to adapt the system to other fields like law or finance by integrating different taxonomic backbones. Additionally, the authors aim to improve the system's ability to generate complex abstractions (e.g., grouping various antibiotics under a single heading) and move beyond exact lexical matching when linking to external hierarchies.

Summary updated

BibTeX
@inproceedings{yair2023hierarchy,
  author = {Itay Yair and Hillel Taub-Tabib and Yoav Goldberg},
  title = {{Hierarchy Builder: Organizing Textual Spans into a Hierarchy to Facilitate Navigation}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
  doi = {10.18653/v1/2023.acl-demo.27},
}
Figure