[Paper Review] Knowledge Graph Anchored Information-Extraction for Domain-Specific Insights

Vivek Khetan, K. M. Annervaz+4 more|arXiv (Cornell University)|Apr 18, 2021

Topic Modeling6 references4 citations

TL;DR

This paper proposes a knowledge graph-anchored information extraction pipeline that uses state-of-the-art NLP models—bi-LSTM-CRF for entity extraction, attention-based SRL, and verb-based relationship extraction—to automatically extract task-specific regulatory changes from unstructured financial documents. The method enables real-time insights by integrating extracted data into a domain-specific knowledge graph, significantly reducing human effort in monitoring regulatory updates with promising summarization efficiency.

ABSTRACT

The growing quantity and complexity of data pose challenges for humans to consume information and respond in a timely manner. For businesses in domains with rapidly changing rules and regulations, failure to identify changes can be costly. In contrast to expert analysis or the development of domain-specific ontology and taxonomies, we use a task-based approach for fulfilling specific information needs within a new domain. Specifically, we propose to extract task-based information from incoming instance data. A pipeline constructed of state of the art NLP technologies, including a bi-LSTM-CRF model for entity extraction, attention-based deep Semantic Role Labeling, and an automated verb-based relationship extractor, is used to automatically extract an instance level semantic structure. Each instance is then combined with a larger, domain-specific knowledge graph to produce new and timely insights. Preliminary results, validated manually, show the methodology to be effective for extracting specific information to complete end use-cases.

Motivation & Objective

• To address the challenge of timely detection of regulatory changes in fast-evolving domains like finance, where manual monitoring is infeasible.
• To reduce human cognitive load by extracting only task-relevant information from large volumes of unstructured regulatory text.
• To build a scalable, semi-automated system that integrates extracted information into a domain-specific knowledge graph for continuous insight generation.
• To enable actionable notifications and dynamic updates based on user-defined criteria and semantic similarity.
• To develop a reusable framework applicable beyond finance, particularly in regulatory and compliance domains.

Proposed method

• A bi-LSTM-CRF model is trained on a combined dataset of custom-labeled regulatory articles and CoNLL 2017 data to improve entity extraction performance.
• An attention-based deep SRL model identifies predicates and assigns semantic roles (e.g., agent, theme) to tokens in sentences.
• A verb-based relationship extraction method identifies multiple relationships between entity pairs using syntactic and semantic cues from verbs.
• A filtering step ensures only entities detected by both the SRL and custom entity extractor are retained, improving precision.
• Extracted triples are enriched using heuristics from OpenIE and ClauseIE to expand the knowledge graph beyond the data model’s scope.
• The final knowledge graph integrates domain-specific data from the NIC database, including institutions, assets, and regulatory relationships, and is updated continuously with new article streams.

Experimental results

Research questions

• RQ1 Can a hybrid NLP pipeline effectively extract task-specific regulatory events—particularly threshold changes—from unstructured legal and regulatory texts?
• RQ2 To what extent can a knowledge graph anchored to extracted semantic triples improve the timeliness and relevance of domain-specific insights?
• RQ3 How effective is the integration of multiple NLP models (entity, SRL, relationship extraction) in reducing human effort for regulatory monitoring?
• RQ4 Can semantic similarity and rule-based systems generate accurate, personalized notifications without requiring full domain ontology engineering?
• RQ5 How does the summarization ratio of extracted information compare to original document size in real-world regulatory use cases?

Key findings

• The system successfully extracted 7 entity types from approximately 4,500 sentences across 131 regulatory articles, with instance counts ranging from 561 for 'regulated_activity_threshold' to 6,752 for 'regulatory_authority'.
• The pipeline achieved high recall with acceptable precision, enabling users to easily discard irrelevant results, which supports the goal of reducing human effort.
• The summarization ratio—defined as input document size divided by output extracted structure size—was found to be promising, though exact values are still under evaluation.
• The integration of heuristic-based OpenIE and ClauseIE triplets significantly expanded the knowledge graph beyond the initial data model, enriching node and relationship coverage.
• Notifications were successfully generated via user-defined subscription rules and WordNet-based semantic similarity, enabling role-based alerting.
• Over time, the evolving knowledge graph enabled the generation of domain-level insights through continuous data ingestion and querying.