Highlights and Paper Summaries from ACL 2018 Conference

The 2018 ACL conference was held in Melbourne, Australia from July 15‑20, attracting about 1,500 participants. A total of 1,544 submissions were received, with 258 long papers and 126 short papers accepted, yielding an overall acceptance rate of 24.9%.

Best Paper Awards : The award‑winning works focused on new problem formulation and dataset creation, including SQuAD 2.0 (unanswerable questions), ranking clarification questions from StackExchange, and detecting adverbial presupposition triggers.

Tutorial – Neural Semantic Parsing : The tutorial covered classic datasets (GeoQuery, ATIS, CoNLL‑A), decoding algorithms, training methods (maximum margin, structured learning, reinforcement learning), and practical tools such as AllenNLP for building semantic parsers.

Machine Translation : Highlighted papers combined RNN‑based NMT with self‑attention mechanisms, introduced dynamic sentence sampling for efficient training, and proposed coverage‑aware NMT models that incorporate attention‑based coverage features.

Semantic Parsing : Summarized approaches included coarse‑to‑fine encoder‑decoder decoding, structured VAE for latent semantic representations, and graph‑based generation of semantic parses using RNNs.

Question Answering : Discussed methods that integrate commonsense knowledge via external memory, multi‑paragraph reading comprehension with cross‑passage answer verification, and analyses of model understanding using attribution techniques.

Domain Adaptation : Presented a strong baseline using tri‑training for semi‑supervised learning under domain shift, showing competitive performance against adversarial methods.

Machine Learning : Described the SPIGOT algorithm for back‑propagating through structured argmax operations, enabling differentiable parsing and other structured prediction tasks.

Text Classification : Covered the integration of regular expressions with neural networks, joint word‑label embeddings, universal language model fine‑tuning (ULMFiT), and unsupervised random‑walk sentence embeddings.

Summarization : Reviewed soft‑template based neural summarization with retrieval, reranking, and rewriting, as well as reinforcement‑learning‑driven sentence rewriting for abstractive summarization of long documents.

Dialogue Systems : Presented exemplar encoder‑decoder models for conversation generation, memory‑augmented task‑oriented dialog (Mem2Seq), and planning‑integrated reinforcement learning for dialogue policy (Deep Dyna‑Q).

Generation : Introduced cooperative discriminators that evaluate generated text on repetition, entailment, relevance, and lexical style to improve language generation quality.

Data : Described MojiTalk, a large‑scale dataset of emoji‑annotated Twitter conversations for training emotional response generators.

Tools : Provided a link to the C++ open‑source Neural Machine Translation toolkit Marian.