Last week I attended the 11th International Conference on Computational Semantics (IWCS 2015). This conference is the bi-yearly meeting of the ACL‘s Special Interest Group on Semantics and this edition was hosted by Queen Mary University’s Computational Linguistics Lab. The topics discussed at the conference revolve around computing, annotating, extracting and representing meaning in natural language. The format of the conference consisted in a first day of workshops (I attended Advances in Distributional Semantics, ADS) followed by three days of main event.
It was nice to be back at Queen Mary, where I studied for my MSc and PhD in Information Retrieval, and it was nice to have the opportunity to mix with a different academic crowd. In fact, a part from the local organisers, I hadn’t met any of the attendees before, and I only knew a couple of famous names. In particular, Hinrich Schuetze (probably best known for co-authoring a book on Natural Language Processing and one on Information Retrieval) gave a talk at the ADS workshop about The case against compositionality, and Yoshua Bengio (one of the most influencial figures in Deep Learning) gave one of the keynote speeches, about Deep Learning of Semantic Representations.
To confirm a feeling that I already had, I have to say that small, single-track conferences are in general more enjoyable than huge ones. You might not have an open bar reception in a 5+ stars fancy hotel, but the networking is much more relaxed, people are in general more approachable, and QA sessions are usually spot on. Of course it really depends on the venues, but my non-statistically significant experience tells me so. Moreover, in bigger venues a lot of attention goes to improving some baseline of some 0.1% accuracy (or whatever metric) without many details on the theoretical foundations (of course with exceptions). Smaller venues usually have the chance to dig deeper into what it is that really makes a model interesting, even when the results are less solid or the evaluation is on a small-ish scale.
Talking about evaluation, this was in my eyes the biggest difference with Information Retrieval conferences: scalability and large-scale evaluation have been rarely, if ever, mentioned. I understand that other venues like EMNLP are probably more suitable for these topics, but it was something that I noticed.
In general, it’s difficult to mention one particular talk, as they were all more or less interesting in my eyes, but one quote that stood out for me was an answer given by Prof. Bengio at the end of his keynote, regarding negation and quantification, and how a Neural Network model deals with them: “I don’t know. But it learns to do what it needs to do.”
As a final side-note, the social event/dinner was a boat trip on the Thames: looking at some well-known London landmarks from a different point of view was absolutely amazing. Well done to the organisers!