Introduction
The 2009 Conference on Empirical Methods in Natural Language Processing (EMNLP), held in Singapore, marked a significant period of growth and transition for the NLP community. With a record-breaking 475 submissions and a selective 34% acceptance rate, the proceedings provide a comprehensive snapshot of a field moving toward web-scale statistics, sophisticated Bayesian modeling, and the early adoption of crowdsourcing for evaluation.
Addressing the Web 2.0 Challenge
In 2009, the primary challenge for researchers was scaling empirical methods to handle the explosion of data from the "Web 2.0" era. The community was grappling with how to move beyond small, hand-annotated corpora—such as the Penn Treebank—to utilize noisier, larger-scale resources like Wikipedia, online forums, and search engine logs.
Key research areas addressed across the 163 accepted papers included:
- Statistical Machine Translation (SMT): Improving fluency and structural accuracy.
- Semantic Parsing: Transitioning from syntax to unsupervised or semi-supervised meaning representation.
- Sentiment Analysis: Moving beyond binary classification to understand subjectivity and conditional sentiment.
- Resource Scarcity: Developing cross-lingual methods to leverage data from resource-rich languages for resource-poor ones.
Methodological Trends
The proceedings highlight a "state-of-the-art" toolbox dominated by classical statistical learning before the deep learning revolution. Key methods included:
- Discriminative and Generative Models: Extensive use of Conditional Random Fields (CRFs), Support Vector Machines (SVMs) with complex kernels, and Bayesian models for grammar induction.
- Machine Translation Architectures: A focus on synchronous grammars, lattice parsing, and phrase-based SMT.
- Crowdsourcing: A pivotal moment for the field was the validation of Amazon Mechanical Turk as a viable tool for evaluating translation quality and scaling annotation.
- Active Learning: Strategic selection of data points for human labeling to reduce annotation costs.
Data and Resources
Research in this volume relied on a mix of classic benchmarks and emerging web resources. While the Penn Treebank, PropBank, and Europarl remained foundational, there was a clear shift toward web-scale data. The Google Web 1T n-gram corpus and Wikipedia became essential repositories for frame information and linguistic statistics. Additionally, the community began exploring diverse datasets derived from email, blogs, and interactive media, reflecting the broadening scope of NLP applications.
Institutional and Conceptual Shifts
EMNLP 2009 solidified several practices and insights that have shaped the field:
- The Author Response Era: This conference established the use of an author response period in peer review, allowing researchers to correct factual errors before final decisions—a practice that remains standard.
- Legitimacy of Crowdsourcing: The acceptance of work using non-expert annotation signaled a shift in how the community viewed scaling NLP tasks.
- The Semantic Frontier: While earlier years focused heavily on Part-of-Speech tagging and basic parsing, 2009 showed a clear push toward unsupervised semantic parsing and coreference resolution.
Conclusion and Historical Context
The 163 papers in this volume represent the pinnacle of feature-engineered statistical NLP. While the field faced significant challenges in domain adaptation and computational constraints, EMNLP 2009 set the stage for the neural shift that would arrive a few years later. It remains a vital record of a community successfully scaling its methods to meet the demands of a rapidly expanding digital world.
Summary updated