Selection of Passages for Information Reduction

AAAI 1996

TL;DR

This paper introduces SPIRE, a system that tackles the bottleneck of extracting structured information from large text documents for case-based reasoning (CBR) systems. SPIRE leverages existing case-base annotations to generate targeted queries, efficiently retrieving only the most relevant passages from new texts, thereby significantly reducing manual effort and improving information extraction.

What Problem Does This Paper Address?

Extracting structured information from unstructured text is a significant challenge, particularly for domains like law and medicine where case-based reasoning (CBR) is crucial. CBR systems rely on symbolic representations of cases, often built from textual data. Manually converting large volumes of text into these structured formats is time-consuming and error-prone. Existing automated information extraction techniques often require extensive training data with manually encoded answers, which is also a bottleneck. This paper addresses the need for a more efficient way to identify and extract key pieces of information (slots and fillers) from new texts to populate or augment CBR systems, without requiring full text comprehension or massive manual annotation efforts.

Methods

The core of SPIRE is the integration of a case-based reasoner with an Information Retrieval (IR) engine. When a CBR system encounters a new problem, it first identifies similar cases in its existing knowledge base. SPIRE then uses "annotations" – descriptive words and phrases associated with slot fillers in the existing cases – to construct targeted IR queries. These queries are designed to retrieve small, relevant passages from new documents that are likely to contain information pertinent to the slots in the current case. This approach effectively "reduces" the amount of text a user or an automated system needs to process by focusing on the most important segments.

Datasets

Not discussed in the paper. The paper refers to "pre-existing texts" and "novel texts" but does not specify any particular datasets used for training or evaluation.

Key Insights and Ideas

  • Leveraging existing annotations is key: Instead of starting from scratch, SPIRE cleverly reuses the descriptive annotations already present in a CBR system's case base to guide information retrieval.
  • Targeted passage retrieval saves time: By generating queries that aim to find specific slot fillers, SPIRE avoids the need to process entire documents, leading to substantial time savings.
  • Bridging CBR and IR: The paper demonstrates a practical synergy between CBR systems (which need structured data) and IR systems (which excel at finding relevant text segments).
  • Annotations as query seeds: The concept of using slot-filler descriptions as seeds for IR queries is a novel way to make IR more context-aware and domain-specific.

Key Results

The paper highlights that SPIRE significantly reduces the time and effort required for information extraction. By presenting users with only relevant passages, it transforms the task of reading an entire document into examining several sets of sentences. This leads to a "tremendous savings in time and effort." While specific quantitative results are not provided, the qualitative benefit of focused information retrieval is emphasized.

Limitations and Future Work

The paper acknowledges that full-text understanding is not yet feasible and that current information extraction techniques rely on large amounts of manually encoded training data. SPIRE's approach aims to mitigate these issues by reducing the reliance on full-text processing and manual annotation for new texts. Future work could involve exploring more sophisticated methods for generating annotations and queries, as well as evaluating SPIRE on a wider range of domains and text types. The paper implicitly suggests that further automation of the annotation creation process itself would be beneficial.

Summary updated

BibTeX
@inproceedings{1996selection,
  title = {{Selection of Passages for Information Reduction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1996},
}
Figure