A News Analysis System

Robert J. KuhnsCOLING 1988

TL;DR

This paper introduces a prototype News Analysis System (NAS) that classifies and indexes news stories in real-time. By parsing sentences and mapping syntactic structures to a concept base, NAS creates a detailed index, enabling efficient routing of relevant information to users and significantly reducing information overload.

What Problem Does This Paper Address?

In today's information-saturated world, professionals in fields like finance and news reporting face a critical challenge: processing vast amounts of incoming data rapidly and accurately. Traditional text processing methods, while useful for general information retrieval, often fall short when immediate, nuanced understanding and categorization are required. The sheer volume of news, coupled with the need to identify specific, pertinent details within seconds, creates a bottleneck. This paper addresses the need for a system that can automatically classify, index, and route news stories in real-time, thereby alleviating the burden of sifting through irrelevant information and ensuring users receive only what is most relevant to them.

Methods

The core of the NAS is a modular architecture designed for real-time news analysis. It comprises four sequential subsystems:

  • Stream Filter: This component initially scans incoming news feeds (e.g., from satellite dishes) to isolate textual reports from non-textual elements like financial quotations.
  • Lexical Scanner: It then processes the textual reports, breaking them down into words, numbers, punctuation, and identifying sentence boundaries. Crucially, it associates each recognized word with its lexical information.
  • Parser: This is the heart of the system, employing a deterministic, principle-based Government-Binding (GB) parser. This parser constructs syntactic structures for each sentence by applying linguistic principles and constraints (like Case, bounding, X-bar, thematic, trace, and binding theories). Its deterministic nature, while sacrificing some ambiguity resolution capabilities, allows for a high processing speed of approximately 100 words per second. The parser outputs a set of "licensing relations" that describe the syntactic relationships within a sentence.
  • Semantic Processor: This subsystem takes the syntactic structures from the parser and maps them onto a concept base. It uses a f-relation interpreter to match predicate-argument structures with concepts and designators (subtopics) within the concept base. This allows for the identification of general topics (e.g., "merger/acquisition") and specific details (e.g., "company names," "industries").

The system is designed to be highly modular, allowing for easy integration of different news feeds or concept bases without affecting other components.

Datasets

The paper mentions that the system was initially developed and tested using stories from Reuters. The system's lexicon was built by collecting distinct "words" from stories received from a satellite feed. The paper also notes that a group of professional indexers provided feedback on the system's output, contributing over 400 topics for the concept base.

Key Insights and Ideas

  • Real-time, Deep Analysis: The system moves beyond simple keyword matching to perform a deeper syntactic and semantic analysis, enabling more accurate and nuanced classification.
  • Government-Binding Theory for Practical NLP: The successful application of a principle-based GB parser in a real-time, practical NLP task demonstrates the viability of complex linguistic theories for applied problems.
  • Modularity for Adaptability: The system's modular design is a key strength, allowing for easy adaptation to new data sources or evolving domain knowledge without requiring a complete system overhaul.
  • Concept Base Abstraction: The multi-level concept base (names, entity types, general concepts) allows for both specific detail extraction and broader categorization, providing a rich indexing capability.
  • Implicit Concept Identification: NAS can identify concepts even when they are not explicitly stated in the text, relying on the semantic processor's ability to infer meaning from syntactic structures and world knowledge.

Key Results

The paper reports on the capabilities of the system rather than providing extensive quantitative results from a large-scale evaluation. However, it highlights:

  • The ability to classify stories into over 200 concepts and designators.
  • A pre-prototype version achieved over 70% accuracy on a small sample of stories.
  • The current version is implemented in Symbolics Common Lisp and achieves a parsing speed of approximately 100 words/second.
  • Qualitative examples demonstrate NAS's ability to generate relevant indexes, correctly identifying concepts like "Strikes and Lockouts," "Airlines Industry," "Divestment," "Industry Giftware," "Trade Glass," "Terrorism," and "Location Kuwait," even when specific keywords are absent from the text.

Limitations and Future Work

  • Error Handling: While NAS has error handling routines for unknown words and incomplete parses, a general categorization might be missed if syntactic relationships cannot be fully established.
  • Lexicon Size: The lexicon currently contains less than 15,000 members, which, while extensive for news reports, might still miss some specialized terms.
  • Future Work: The authors plan to focus on establishing quantitative benchmarks for speed and accuracy, enhancing the semantic processor and parser, and exploring inferencing across classified stories. They also envision enabling users to query processed stories and developing automatic summarization capabilities.

Summary updated

BibTeX
@inproceedings{kuhns1988news,
  author = {Robert J. Kuhns},
  title = {{A News Analysis System}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1988},
}
Figure