Proceedings of the Third Conference on Empirical Methods for Natural Language Processing, Palacio de Exposiciones y Congresos, Granada, Spain, June 2, 1998

EMNLP 1998 · DBLP

Introduction

The Proceedings of the Third Conference on Empirical Methods in Natural Language Processing (EMNLP 1998) represent a foundational moment in the history of computational linguistics. Held in Granada, Spain, this conference marked the first time EMNLP was hosted outside the United States, signaling a deliberate shift toward international collaboration and the application of data-driven methods to a global array of languages.

Addressing the Rule-Based Bottleneck

In the late 1990s, the field of Natural Language Processing (NLP) faced a significant transition. Traditional systems, largely reliant on hand-crafted linguistic rules, were increasingly viewed as "brittle"—unable to scale effectively to the noise and variety of real-world data. Furthermore, research was heavily skewed toward the English language.

Organizers Nancy Ide and Atro Voutilainen identified a critical research gap: while statistical methods were showing promise, their scalability across different languages remained unproven. EMNLP 1998 sought to bridge this gap by focusing on how empirical techniques—such as lexical tagging, parsing, and anaphora resolution—could be adapted for a multilingual landscape.

Methodological Innovations

The 1998 proceedings showcased several emerging empirical techniques that have since become standard in the field. These methods moved away from rigid syntactic rules toward models that accounted for specific word behaviors and statistical probabilities:

  • Head-Lexicalized Probabilistic Context-Free Grammars (PCFGs): These were utilized for valence induction, helping models understand how verbs interact with their arguments based on corpus data.
  • Maximum Entropy and Minimum Divergence: These mathematical frameworks provided a rigorous basis for feature selection, allowing researchers to identify the most informative data points within a corpus.
  • Lexicalized Statistics: Applied specifically to Japanese dependency structure analysis, these models allowed for more nuanced parsing by focusing on the statistical relationships between specific words rather than abstract categories.
  • Parallel Text Alignment: Researchers presented techniques for aligning clauses across different languages, a vital precursor to modern machine translation.

Multilingual Applications and Results

The conference highlighted successful applications of these empirical models across diverse linguistic tasks. Notable achievements included:

  • Japanese Dependency Analysis: The use of lexicalized statistics enabled more accurate structural analysis of Japanese sentences.
  • Automatic Accent Insertion: Empirical models were successfully applied to restore missing accents in French text, demonstrating the practical utility of statistical approaches in orthographic correction.
  • Robust Anaphora Resolution: Research by Mitkov et al. emphasized the importance of "robustness"—developing methods that, while perhaps not 100% accurate, remained resilient when processing the varied data found in multilingual corpora.
  • Coreference-Based Summarization: New approaches used the relationships between words (coreference) to generate dynamic summaries of longer documents.

The Rise of Statistical Translation

A defining moment of the conference was the invited talk by Kevin Knight, "Statistical Translation: Where It Went." This presentation signaled the growing dominance of the statistical approach to translation, a paradigm that would eventually underpin major modern technologies like Google Translate. The talk underscored the conference's broader theme: the future of NLP lay in its ability to process multilingual information through empirical data rather than manual rule-writing.

Conclusion and Legacy

EMNLP 1998 established empirical evaluation as the gold standard for the field. By prioritizing the evaluation of strengths and weaknesses through data, the conference moved NLP toward a more scientific and reproducible discipline. While the organizers noted a persistent "resource gap"—the need for more high-quality, annotated corpora in non-English languages—the proceedings laid the groundwork for the multilingual, data-rich era of modern artificial intelligence.

Summary updated

BibTeX
@inproceedings{1998proceedings,
  title = {{Proceedings of the Third Conference on Empirical Methods for Natural Language Processing, Palacio de Exposiciones y Congresos, Granada, Spain, June 2, 1998}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {1998},
}
Figure