Abstract
The NAMIC project introduces a system for "Multilingual Authoring"—the automated creation of hypertext links between news articles across different languages. By leveraging Information Extraction (IE) and a shared multilingual ontology (EuroWordNet), the system identifies core events and entities to connect related stories in English, Spanish, and Italian, moving beyond simple keyword matching to true semantic linking.
Introduction
As the volume of electronic information grows, news agencies face a "needle in a haystack" problem: how to help users navigate similar stories published in different languages. Traditional Information Retrieval (IR) is often too broad, returning a flood of documents that happen to share a keyword. Conversely, traditional Information Extraction (IE) is usually too narrow, designed for very specific domains.
The NAMIC (News Agencies Multilingual Information Categorisation) project identifies a gap: the need to automatically structure news streams so that a reader interested in a specific event, such as a corporate acquisition in an English paper, can be automatically directed to related reports in Spanish or Italian. The core challenge lies in "event matching"—recognizing that "Intel bought NKT" and "Intel ha acquistato Xircom" are the same type of event, despite linguistic differences.
System Architecture and Methodology
The NAMIC system uses a distributed, modular architecture to process text through several stages, as illustrated in Figure 1.
Text Processing and Information Extraction
The system employs language-specific processors that perform morphosyntactic analysis to identify parts of speech and syntactic relations. It utilizes components of the LaSIE (Large-scale Information Extraction) system, including:
- Named Entity Matcher: Identifies people, organizations, locations, and dates.
- Discourse Processor: Translates semantic representations into instances within a "World Model" and handles coreference resolution (e.g., realizing "the company" refers back to "Intel").
Ontological Modeling
The heart of the multilingual bridge is EuroWordNet. By using an Inter-Lingual Index (ILI), the system maps language-specific terms to a common set of "Base Concepts." This allows the system to recognize that different verbs in different languages refer to the same underlying concept, such as a "financial acquisition."
Distributed Implementation
Built using CORBA and Java, the system employs language-specific processors that feed into a central Multilingual Authoring Engine (M-AE). This architecture allows for scalability by adding more language-specific servers as news volume increases.
Key Insights
- Events over Keywords: Linking based only on proper nouns creates too much noise. Effective authoring requires matching the roles within an event (Agent, Action, Object).
- The "World Model" as a Bridge: By populating a language-independent ontology with extracted facts, the system creates an "objective representation" of the news, making the linking process language-agnostic.
- Coreference is Critical: To understand an event, the system must resolve anaphora and synonyms. Knowing "Intel" and "the chipmaker" are the same entity is vital for accurate event mapping.
- Semi-Automating Knowledge Engineering: NAMIC proposes using statistical techniques and "Galois lattices" to automatically learn subcategorization patterns for verbs, easing the porting of the system to new domains.
Results and Evaluation
The system was developed and tested using real-world news streams from major European providers: EFE (Spain), ANSA (Italy), and The Financial Times (UK), focusing on the corporate and industrial news domain.
Key results include:
- Cross-Lingual Connectivity: Successful generation of "Multilingual Links" for navigation between English, Spanish, and Italian news streams based on shared semantic content.
- Robustness: Shallow parsing and modular components allow the system to handle the "noise" of real-world news text better than rigid, deep-parsing alternatives.
Limitations and Future Work
While the architecture is general, the current implementation is tuned to corporate news. Expanding to other areas requires defining new event classes. Furthermore, the system is designed as an assistive tool; the authors note that users should retain the ability to refuse or ignore suggested links. Future work aims to make the system more "adaptive" by allowing users to specify new types of events for the IE system to track.
Summary updated