Event Schema Induction for Information Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Information retrieval systems face challenges in identifying relevant technical documents for user queries due to lengthy and verbose technical documents, semantic gaps between query and document terminology, and the difficulty in capturing relevant parts of larger documents using existing models like TF-IDF, LSTM, and Transformer models, especially when labeled data is scarce.
Innovation Solution
A system comprising a processor that executes components for deriving an event schema from unstructured or semi-structured data using graph neural networks and Abstract Meaning Representation (AMR) graphs, which facilitates event extraction and improves retrieval by clustering vector representations of events into clusters to identify event types and adjust the schema based on feedback data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If bag-of-word models like TF-IDF are used to identify relevant documents, then word frequency matching is simple and fast, but word order and semantic context are ignored leading to poor retrieval accuracy
Solution Approach 1:
The patent segments documents into discrete events with specific schemas (participants, triggers, objects) rather than treating them as continuous text. This segmentation allows the system to extract and match specific event components independently, improving retrieval accuracy while maintaining computational efficiency through structured data representation.
Solution Approach 2:
The patent transitions from one-dimensional bag-of-words representation to multi-dimensional event schema representation. By organizing information into structured dimensions (event type, participants, triggers, temporal relationships), the system captures semantic context and word order information that TF-IDF misses, significantly improving retrieval precision.
2Measurement precision
If LSTM models and Transformer models are used to capture relevant parts of documents, then semantic context is better understood, but large amounts of labeled data are required for fine-tuning which is generally unavailable for technical documents
Solution Approach 1:
The system performs self-service by automatically inducing event schemas from unlabeled technical documents without requiring manual annotation. The schema induction process extracts event patterns directly from the document corpus, enabling the retrieval system to learn semantic structures from raw text alone, eliminating the need for expensive labeled training data.
Solution Approach 2:
The patent performs preliminary schema induction on the document corpus before retrieval operations. By pre-processing documents to extract event schemas and organizing them into structured formats, the system prepares the data in advance, allowing efficient retrieval without requiring complex fine-tuning of deep learning models during deployment.
3Ease of operation
If entities are extracted from user queries and matched with entities from technical documents, then entity identification is straightforward, but actions involving those entities are missed
Solution Approach 1:
The patent merges entity extraction with action extraction by unified event schema induction. Instead of separately identifying entities and their actions, the system simultaneously extracts both components as integrated event structures, ensuring that actions involving entities are captured together in a cohesive representation that improves retrieval completeness.
Data Source
AI summary
Systems, devices, computer-implemented methods, and/or computer program products that facilitate event schema induction from unstructured or semi-structured data. In one example, a system can comprise a processor that executes computer executable components stored in memory. The computer executable components can comprise a schema component and a retrieval component. The schema component can derive an event schema for a document corpus using parsing results obtained from the document corpus. The retrieval component can populate a response to a query with a document of the document corpus using events extracted from the query and the document using the event schema.