Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Tutorial Abstracts

André F. T. Martins, Tsvetomila Mihaylova, Nikita Nangia, Vlad Niculae, Alexander Koller, Stephan Oepen, Weiwei Sun, Shafiq Joty, Giuseppe Carenini, Raymond T. Ng, Gabriel Murray, Goran Glavašš, Federico Nanni, Simone Paolo Ponzetto, Jen‐Tzung Chien, Joty Shafiq, Mihaylova Tsvetomila, Nanni Federico, Marius Paşca, Sebastian Ruder, Alexander, Krishnan, Laha, Anirban, Mishra, Gabriel, Vlad, Stephan, Paolo, Sankaranarayanan, Karthik, Sgaard, Anders, Vuli, Ivan, Wilker Aziz, Philip Schulz, Jihun Choi, Min Kang, Sang-Goo Yoo, Lee, Caio Corro, Ivan Titov, Jacob Devlin, Ming-Wei Chang, Kristina Toutanova, Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, Noah Smith, Serhii Havrylov, Germn Kruszewski, Armand Joulin, Eric Jang, Shixiang Gu, Ben Poole, Yoon Kim, Carl Denton, Loung Hoang, Alexander Rush, Sam Wiseman, Yang Liu, Mirella Lapata, Andriy Chris J Maddison, Yee Mnih, Teh, Jean Maillard, Stephen Clark, Chaitanya Malaviya, Pedro Ferreira, F Andr, Ramn Martins, Astudillo Fernandez, Arthur Mensch, Mathieu Blondel, Samuel Bowman, Mathieu Martins, Claire Blondel, Cardie, Goran GlavašOpenalex Unknown 2019 · DOI · OpenAlex

Introduction

The 57th Annual Meeting of the Association for Computational Linguistics (ACL 2019) represented a pivotal moment in the evolution of Natural Language Processing. The collected tutorial abstracts from this conference serve as a high-level roadmap, documenting a significant shift in the field: the move away from purely "black-box" deep learning toward models that incorporate latent linguistic structure, graph-based meaning, and unsupervised cross-lingual transfer. These tutorials provide the theoretical and practical foundations for making neural networks more interpretable, structured, and globally applicable.

The Quest for Latent Structure and Interpretability

As deep learning became the standard in NLP, a gap emerged between high-performing "end-to-end" neural models and the human need for structural correctness. A primary focus of the ACL 2019 tutorials was the discovery of hidden linguistic structures—such as parse trees—within neural networks without requiring explicit supervision.

Researchers highlighted several methodologies to bridge this gap:

  • Latent Structure Modeling: Utilizing Reinforcement Learning (REINFORCE), surrogate gradients, and end-to-end differentiable methods to allow models to "discover" their own internal hierarchies.
  • Deep Bayesian NLP: Combining the flexibility of neural networks with the formal uncertainty and prior knowledge of Bayesian inference, specifically through Variational Auto-Encoders (VAEs) and Dirichlet processes.

By treating structure as an inductive bias, researchers can lead models toward better generalization, making them more interpretable for humans while maintaining the performance benefits of deep learning.

Cross-Lingual Transfer and Resource Scarcity

A major challenge in modern NLP is the disparity between high-resource languages (like English) and the thousands of languages that lack massive labeled datasets. The tutorials explored Unsupervised Alignment techniques, such as the MUSE architecture, which uses adversarial training to map word embeddings from different languages into a shared space without the need for a bilingual dictionary.

This approach relies on the Isomorphism Hypothesis: the observation that word vector spaces in different languages often share similar geometric shapes. This allows for mathematical alignment without direct translation pairs, enabling unsupervised machine translation that, in some cases, rivals earlier supervised methods.

Beyond the Sentence: Discourse, Graphs, and Argumentation

The tutorials also signaled a move beyond sentence-level analysis toward document-level coherence and complex semantic representations.

Graph-Based Parsing

Traditional tree structures are increasingly being replaced by "unanchored" graph representations, such as Abstract Meaning Representation (AMR). These graphs capture deeper semantic nuances that simple hierarchies miss, moving semantic parsing from a niche research area to a staple of the field.

Argument Mining and Political Analysis

While sentiment analysis focuses on what a user feels, Argument Mining seeks to extract the underlying logic and premises of a claim. This is vital for identifying misinformation and analyzing political discourse. The field has seen explosive growth, expanding from a few papers in 2010 to thousands by 2019. This is complemented by specialized work in political science, using the Manifesto Corpus and the Comparative Agendas Project to analyze policy positions and electoral programs.

Data-to-Text Generation

The transition from structured data to natural language (NLG) was another key theme. Using datasets like WIKIBIO and WEBNLG, researchers are refining how models generate human-like narratives from raw data. However, this field faces significant evaluation challenges; traditional metrics like BLEU or ROUGE often fail to capture the narrative flow or factual accuracy required for high-quality storytelling.

Limitations and Future Directions

Despite the progress documented at ACL 2019, several hurdles remain:

  • The "Distant Language" Gap: Unsupervised cross-lingual methods struggle when language pairs are not isomorphic (e.g., English and Chinese).
  • Supervision Tension: There is an ongoing debate between using "off-the-shelf" linguistic parsers, which may carry human bias, and letting models learn latent structures, which may result in representations that are difficult for humans to interpret.
  • Evaluation Metrics: As NLP tasks become more subjective (like storytelling or argumentation), the need for more robust, human-centric evaluation metrics becomes critical.

In summary, the ACL 2019 tutorials reflect a field maturing beyond simple pattern recognition toward a more nuanced, structured, and linguistically informed approach to artificial intelligence.

Summary updated

Abstract

is a PhD student at New York University, advised by Samuel Bowman. She is working on building neural network systems in NLP that simultaneously do structured prediction and representation learning. This work focuses on finding structure in language without direct supervision and using it for semantic tasks like natural language inference and summarization.
BibTeX
@inproceedings{martins2019proceedings,
  author = {André F. T. Martins and Tsvetomila Mihaylova and Nikita Nangia and Vlad Niculae and Alexander Koller and Stephan Oepen and Weiwei Sun and Shafiq Joty and Giuseppe Carenini and Raymond T. Ng and Gabriel Murray and Goran Glavašš and Federico Nanni and Simone Paolo Ponzetto and Jen‐Tzung Chien and Joty Shafiq and Mihaylova Tsvetomila and Nanni Federico and Marius Paşca and Sebastian Ruder and Joty Shafiq and Koller and Alexander and Krishnan and Laha and Anirban and Mihaylova Tsvetomila and Mishra and Murray and Gabriel and Nanni Federico and Niculae and Vlad and Oepen and Stephan and Marius Pasca and Paolo and Sebastian Ruder and Sankaranarayanan and Karthik and Sgaard and Anders and Vuli and Ivan and Wilker Aziz and Philip Schulz and Jihun Choi and Min Kang and Sang-Goo Yoo and Lee and Caio Corro and Ivan Titov and Caio Corro and Ivan Titov and Jacob Devlin and Ming-Wei Chang and Kenton Lee and Kristina Toutanova and Chris Dyer and Adhiguna Kuncoro and Miguel Ballesteros and Noah Smith and Serhii Havrylov and Germn Kruszewski and Armand Joulin and Eric Jang and Shixiang Gu and Ben Poole and Yoon Kim and Carl Denton and Loung Hoang and Alexander Rush and Yoon Kim and Sam Wiseman and Alexander Rush and Yang Liu and Mirella Lapata and Andriy Chris J Maddison and Yee Mnih and Teh and Jean Maillard and Stephen Clark and Chaitanya Malaviya and Pedro Ferreira and Andr Ft Martins and F Andr and Ramn Martins and Astudillo Fernandez and Arthur Mensch and Mathieu Blondel and Nikita Nangia and Samuel Bowman and Vlad Niculae and Mathieu Blondel and Vlad Niculae and F Andr and Mathieu Martins and Claire Blondel and Cardie and Andr Ft Vlad Niculae and Andre Martins and Goran Glavaš},
  title = {{Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Tutorial Abstracts}},
  booktitle = {Openalex Unknown},
  year = {2019},
  doi = {10.18653/v1/p19-4},
}
Figure