Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, EMNLP 2014, October 25-29, 2014, Doha, Qatar, A meeting of SIGDAT, a Special Interest Group of the ACL

EMNLP 2014 · DBLP

Abstract

The 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) marked a pivotal moment in the history of the field. Held in Doha, Qatar, the proceedings document a discipline in the midst of a massive growth spurt and a fundamental shift toward neural architectures. With a record-breaking 790 submissions and the introduction of dedicated "Neural Net Mixer" sessions, EMNLP 2014 solidified the transition from traditional statistical methods to the deep learning techniques that define modern artificial intelligence.

The Scaling Challenge

As a collection of proceedings, this document addresses the "scaling problem" of the NLP community circa 2014. General Chair Alessandro Moschitti noted that attendance had surged to over 500 paying attendees by the previous year, necessitating a move to a more professionalized, large-scale organizational structure. To manage this volume, the conference introduced nine parallel poster sessions and a five-day format including tutorials and workshops—a structure that persists in major AI conferences today.

Scientifically, the proceedings tackle the challenge of extracting meaning from increasingly diverse data sources. Researchers moved beyond curated news corpora to the "wild west" of social media (Twitter, Weibo), microblogs, and multi-modal data combining text with images or brain activity.

The "Neural Turn" and Methodology

While the volume contains hundreds of specific methods, the overarching "meta-method" described in the prefaces is a rigorous peer-review process where 790 submissions were filtered down to 155 long and 70 short papers (an acceptance rate below 30%).

The most significant trend was the emergence of the "Neural Turn." The program schedule highlighted Neural Net Mixers, featuring then-novel approaches such as:

  • Recurrent Neural Networks (RNNs) for machine translation.
  • Convolutional Neural Networks (CNNs) for image embeddings.
  • Recursive Neural Networks for discourse parsing.

Furthermore, many papers focused on "joint modeling," such as combining word segmentation with POS tagging or punctuation prediction with disfluency detection, attempting to integrate deep learning into established linguistic tasks.

Key Insights: From Discrete to Continuous

One of the most influential contributions in these proceedings was the introduction of GloVe (Global Vectors for Word Representation). This era saw the definitive transition from discrete "one-hot" representations to continuous vector spaces. The "Neural Net Mixer" sessions signaled a shift away from complex, hand-engineered feature pipelines toward models that could learn representations directly from data.

There was also a notable increase in "NLP for Social Good." The prefaces highlight an increasing interest in real-world applications, moving NLP out of the lab and into tools for analyzing political debates, stock movements, and detecting deceptive reviews.

Institutional Support and Results

The 2014 conference achieved an unprecedented scale, supported for the first time by a government research foundation, the Qatar National Research Foundation (QNRF). This professionalization was mirrored by significant industry involvement from global tech leaders, indicating the growing commercial importance of empirical NLP methods.

Sponsors and supporting institutions for EMNLP 2014, reflecting the conference's growth and industry integration.

Quantitative gains were reported across the board in translation quality (BLEU scores), parsing accuracy, and sentiment classification. While Machine Translation, Semantics, and Syntax remained the most popular research areas, "NLP for the Web and Social Media" emerged as a major new pillar of the conference.

Limitations and Future Directions

Program Committee Co-Chairs Bo Pang and Walter Daelemans acknowledged the "inevitable time and space limitations" that forced the rejection of many worthwhile papers, highlighting the field's struggle to keep pace with its own rate of discovery.

In 2014, the identified frontiers for future work included:

  • Multi-modal Integration: Better combining text with vision and sensor data.
  • Low-Resource Languages: Improving processing for languages where large training sets are unavailable.
  • Discourse-Level Understanding: Moving from sentence-level analysis to document-level context.
  • Architecture Refinement: The continued evolution of neural models to fully replace traditional statistical frameworks.

Summary updated

BibTeX
@inproceedings{2014proceedings,
  title = {{Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, EMNLP 2014, October 25-29, 2014, Doha, Qatar, A meeting of SIGDAT, a Special Interest Group of the ACL}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2014},
}

Loading related content...

Figure