What ACL 2016 Tutorials Reveal About the Future of NLP and Deep Learning

Deep learning has driven major advances in AI, bringing speech recognition close to a game‑changing threshold and enabling computer vision in security, robotics, and autonomous driving. Natural language processing (NLP) is the next frontier, as highlighted at ACL 2016.

Conference Overview

ACL 2016 was held from August 7‑12, 2016 at Humboldt University in Berlin, attracting about 1,200 participants. The event included a tutorial day, three days of the main conference, and several workshops. The Lifetime Achievement Award went to Prof. Joan Bresnan of Stanford.

Tutorial Summaries

1. Computer Aided Translation

Presented by Prof. Philipp Koehn (Johns Hopkins), this tutorial reviewed the CASMACAT project, which integrates statistical machine translation advances such as confidence scores, paraphrasing, visual word alignment, and translation‑option arrays into computer‑aided translation tools, improving translator efficiency.

2. Neural Machine Translation

Speakers Christopher Manning (Stanford), Thang Luong, and Kyunghyun Cho described the evolution of NMT, from early attention models to recent LSTM and sub‑word techniques, and presented a detailed encoder‑decoder framework with maximum‑likelihood training and beam‑search decoding.

3. Semantic Representations of Word Senses and Concepts

Researchers from Roma University and Cambridge discussed knowledge‑based and unsupervised methods for sense representation, highlighting applications such as semantic similarity, disambiguation, information mining, and clustering.

4. Understanding Short Texts

Presented by Wang Zhongyuan (formerly MSRA, now Facebook) and Wang Hai‑xun, the tutorial covered challenges of short‑text sparsity, noise, and ambiguity, and introduced explicit (ERM) and implicit (IRM) representation models, including embedding‑based deep‑learning approaches.

Selected Papers Discussed

Machine Translation

Multimodal Pivots for Image Caption Translation – proposes retrieving similar target‑language images and re‑ranking translations using image captions, improving caption translation quality.

Modeling Coverage for Neural Machine Translation – introduces a coverage vector to track translated words, addressing over‑ and under‑translation in attention‑based NMT.

Phrase‑Level Combination of SMT and TM Using Constrained Word Lattice – combines statistical MT and translation memory via a constrained lattice, preserving TM usefulness in modern systems.

Machine Reading Comprehension

ACL featured papers such as “A Thorough Examination of the CNN/Daily Mail Reading Comprehension Task” and “Text Understanding with the Attention Sum Reader Network”, which evaluate neural versus traditional classifiers on reading‑comprehension benchmarks.

Language and Vision

Two notable works explored grounding language in images: “Learning Prototypical Event Structure from Photo Albums” (automatic album segmentation and labeling) and “Generating Natural Questions about an Image” (producing human‑like questions for images).

Further tutorials on dialogue systems, information extraction, sentiment analysis, syntactic parsing, and word‑vector representations were announced for the next installment.