Abstract
In this paper, we describe a method for automatic creation of a knowledge source for text generation using information extraction over the Internet. We present a prototype system called PROFILE which uses a client-server architecture to extract noun-phrase descriptions of entities such as people, places, and organizations. The system serves two purposes: as an information extraction tool, it allows users to search for textual descriptions of entities; as a utility to generate functional descriptions (FD), it is used in a functional-unification based generation system. We present an evaluation of the approach and its applications to natural language generation and summarization.
Introduction
In the late 1990s, natural language processing (NLP) faced a significant hurdle in automated news summarization: the "context gap." When a news story breaks, it often involves entities—people, organizations, or locations—that may be unfamiliar to the reader or new to the public eye. Static knowledge bases, while reliable, are inherently limited by their inability to keep pace with the rapid cycle of global news. This paper introduces PROFILE, a system designed to bridge this gap by treating the Internet as a dynamic, real-time knowledge source.
The Challenge of Dynamic Context
Traditional summarization systems often lacked the background information necessary to provide context for new entities. For instance, a report on a downed pilot might require an explanation of their rank or affiliation—details that might not be present in the primary article. To address this, the authors argue for a "Generation Knowledge Source" that is both dynamic (updated in real-time) and syntactically aware. The latter is crucial; a system must do more than simply copy-paste text—it must understand the grammatical structure of descriptions to integrate them seamlessly into new, grammatically correct sentences.
The PROFILE System Architecture
PROFILE utilizes a client-server architecture to extract noun-phrase descriptions from Internet-accessible newswires (such as Reuters, Yahoo, and CNN). The pipeline consists of four primary stages:
- Entity Extraction: Using Part-of-Speech (POS) tagging and the
CREPregular grammar tool, the system identifies potential proper nouns. - Semantic Filtering with WordNet: To distinguish between unique entities and common nouns (e.g., "Prime Minister"), the system cross-references candidates against WordNet. If a term is listed as a common noun, it is filtered out, ensuring the system focuses on specific entities.
- Description Extraction: Once an entity is confirmed, PROFILE searches the corpus for appositions or pre-modifiers (e.g., identifying "Italy's former prime minister" as a description for "Silvio Berlusconi").
- Functional Description (FD) Conversion: Unlike systems that store "canned" text, PROFILE converts extracted snippets into Functional Descriptions. These are structured syntactic representations compatible with the
FUF/SURGEgeneration system, allowing for sophisticated grammatical transformations.
Key Innovations: Syntactic Flexibility and Semantic Triggering
A core contribution of this work is the move away from static strings. By storing descriptions as FDs, the system can perform complex operations, such as merging "President Yeltsin" and "President Clinton" into the pluralized "Presidents Yeltsin and Clinton" without losing grammatical integrity.
Furthermore, the system employs "semantic triggering." By tracing WordNet hypernyms, PROFILE can categorize entities. For example, if a description contains the word "minister," the system can trace this up the hierarchy to the concept of "leader." This allows a generation system to query the database for specific types of information, such as an entity's occupation or nationality, rather than just returning a random snippet.
Evaluation and Results
The system was evaluated using a seed corpus of Reuters articles and live data from sources like ClariNet and early web portals. From an initial pool of 2,150 candidates, the WordNet filtering process successfully isolated 421 high-confidence entity names.
In a manual evaluation of 611 randomly selected descriptions, PROFILE achieved a 90.2% precision rate. Errors typically stemmed from POS tagging inaccuracies or complex noun-phrase attachments that exceeded the grammar's capabilities. Despite these hurdles, the system demonstrated a robust ability to capture a wide variety of descriptions, ranging from simple titles to complex biographical appositions like "Maurizio Gucci, the former head of Italy’s Gucci fashion dynasty."
Limitations and Future Directions
While pioneering, the 1997 version of PROFILE faced several limitations that would define future research in the field:
- Entity Aliasing: The system could not yet resolve that "Bill Clinton" and "William Jefferson Clinton" were the same individual, maintaining separate profiles for each string.
- Polysemy: Ambiguous terms (e.g., "Rice" in "Rice University") occasionally led to misclassification because the system might categorize the entity based on the food definition found in WordNet.
- Integration: At the time of publication, the system had not been fully integrated into a closed-loop, end-to-end multi-document summarizer.
Conclusion
PROFILE represented an early and successful attempt to harness the "live" web as a structured knowledge base for NLP. By focusing on syntactic representations rather than simple text extraction, the authors laid the groundwork for modern systems that require both factual accuracy and grammatical fluidity in automated text generation.
Summary updated