I recently attended the Empirical Methods in Natural Language Processing (EMNLP) conference. In this post I write about the most remarkable stuff presented there and in the co-located events, from the point of view of a neural machine translation researcher. These are just my opinions, feel free to disagree.
Feedback is very welcome. Please leave your comments as replies to this tweet.
If, after reading this post, you want to know more about what happened at EMNLP 2018, I recommend searching for hashtag #emnlp2018 on twitter as there was plenty of live tweeting.
Before the main conference, there were two days of co-located events. The three most interesting ones were the Conference on Machine Translation (WMT), the SIGNLL Conference on Computational Natural Language Learning (CoNLL) and the Blackbox NLP Workshop.
WMT started as a workshop on statistical machine translation, with several shared task competitions were you could take place it. We participated in the news translation shared task, specifically in Estonian-English and Finnish-English (our system).
English to German translation, and especially the WMT14 dataset, has lately become the standard benchmark for machine translation quality (followed by English to French). This year’s news translation task included English to German and Facebook AI Research won in that translation direction with their massive use of backtranslation. In general, the Transformer model was dominant architecture in the news translation task, with a special mention to Marian implementation, which seems to be gaining momentum due to its very high performance.
You can check the full WMT18 news translation task findings report for all the details of the competition.
Apart from the news task, this year I found interesting the noisy corpus filtering task. Partitipants were provided a huge corpus, for which they have to score each sentence pair. Based on the scores, the organization sampled one smaller and one larger subset and trained statistical and neural MT system and evaluated the translation quality. I think it’s worth mentioning Microsoft’s Dual Conditional Cross-Entropy Filtering, which seemed to be simple yet very effective. The full report of the corpus filtering task is here
I could not attend CoNLL, but I would like to mention the article that won the best paper award: Uncovering Divergent Linguistic Information in Word Embeddings with Lessons for Intrinsic and Extrinsic Evaluation, which proposes a method to adapt pretrained embedding vectors to work better for semantics/syntax tasks or for similarity/relatedness tasks.
EMNLP stands for “Empirical Methods…”. The incorporation of the deep learning black box into NLP has lead to an emphasis in the empirical part, ruling out most of the interpretability derived from symbolic approaches and from the modular structure of statistical methods. Many pieces of research just try to characterize the behaviour of models and the effects observed when they are subjected to explorative experiments, but the analyses undergone to explain them are sometimes superficial or are merely justified by intuition, and there are few conclusions that can be applied to contexts other than those very specific experiments.
Blackbox NLP Workshop tried to shed some light on the inner working of deep NLP models. The workshop gained a lot of attention, with both the oral sessions and the poster sessions being packed.
Despite the attempts to make NLP neural networks more interpretable, I think that the black box nature of the currently dominant models imposes a hard non-interpretability wall that prevents us from actually understanding their behaviour completely, and hence we just can resort to characterizing them under different conditions. I hope that at some point we devise models with built-in interpretability where we no longer need to trade interpretability for effectiveness.
EMNLP: Main Conference
The trend from last conferences to try to leverage linguistic knowledge was not very strong at EMNLP. The most remarkable articles were Linguistically-Informed Self-Attention for Semantic Role Labeling and On Tree-Based Neural Sentence Modeling.
Cross-lingual learning was present in a lot of different articles presented at the conference, both regarding word embeddings, machine translation. You can take a look at the accepted papers and search for “cross-lingual” to get an idea.
The recent enthusiams about transfer learning with pretrained models like AllenNLP’s ELMo, Google’s BERT, OpenAI’s GPT and FastAI’s ULMFiT was not reflected at EMNLP, but will probably do at next conferences, for which there’s still time to build new systems making use of them.
- Transformer and backtranslation are the standard machine translation toolbox.
- Cross-lingual and low resource scenarios are gaining momentum.
- Currently, unsupervised SMT works better than unsupervised NMT.