Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018

EMNLP 2018 · DBLP

Abstract

The 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP) represented a landmark event in the NLP community, witnessing a nearly 50% increase in submissions and a definitive consolidation of neural methodologies. This article synthesizes the key research directions, dataset innovations, and systemic challenges presented in the conference proceedings, highlighting the field's shift toward complex reasoning, cross-lingual transfer, and social accountability.

Introduction

The EMNLP 2018 proceedings document a field in the midst of an "explosion." With 2,231 valid submissions, the volume reflects the massive influx of interest in deep learning for language. Beyond the logistical challenge of scaling peer review, the research within these pages addresses core bottlenecks of the era: the data-intensive nature of neural networks, the "black box" problem of model interpretability, and the integration of text with other modalities like vision and video.

Methodological Innovations

The 2018 proceedings were dominated by several "meta-methods" that refined the neural paradigm:

  • Sequence Modeling: A heavy focus on improving the Transformer architecture and exploring non-autoregressive decoding to accelerate translation.
  • Adversarial Training: The use of adversarial techniques to remove demographic bias from datasets and to harden models against input perturbations.
  • Cross-Lingual Mapping: Innovations such as Gromov-Wasserstein Alignment allowed for mapping word embeddings between languages without bilingual dictionaries, a boon for low-resource language processing.

Benchmark Datasets and Grounded Inference

Several datasets introduced at EMNLP 2018 have become foundational to the field. These benchmarks shifted focus from simple pattern matching to more complex linguistic tasks:

  • SWAG: A large-scale adversarial dataset for grounded commonsense inference, designed to challenge the models of the time.
  • MultiWOZ: A multi-domain Wizard-of-Oz dataset that became a standard for task-oriented dialogue modeling.
  • Spider: A complex, cross-domain semantic parsing dataset for text-to-SQL tasks.
The Allen Institute for AI (AI2), creators of the SWAG dataset, representing the institutional drive for new benchmarks.

Key Insights: Semantics and Real-World Application

Invited speakers and researchers alike emphasized a return to semantics. Johan Bos argued for "language-neutral meaning representations" to achieve transparent logic in neural systems. Simultaneously, NLP began to move into high-stakes social applications, including finance, deception detection in speech, and monitoring gang-involved escalation on social media.

A critical result from the conference was the questioning of "human parity" in machine translation. Research by Samuel Läubli and colleagues provided quantitative evidence that parity claims were premature when evaluated at the document level rather than the sentence level.

Systemic Challenges and the "Compute" Barrier

The organizers identified several looming challenges for the community. The "Compute" Barrier emerged as a significant concern, as the increasing size of models created a divide between well-funded industry laboratories and academic researchers. Furthermore, the surge in submissions led to "reviewer fatigue," prompting calls for more automated or hierarchical review structures.

Salesforce Research, an example of the industry labs contributing to the rapid scaling of NLP models.

Conclusion

EMNLP 2018 captured a field at a crossroads, balancing rapid performance gains with the need for deeper understanding and broader inclusivity. The proceedings remain a vital record of the transition toward the large-scale, neural-dominated landscape of modern NLP.

Summary updated

BibTeX
@inproceedings{2018proceedings,
  title = {{Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2018},
}
Figure