Alberto Massidda

How Meta manages to take offline millions of pictures, videos and text that violate its community standards, all of them adversarially engineered, in a catalog that counts in the trillions. We'll talk about open source technologies that embrace vector search, state of the art in neural and non-neural embeddings, as well as turnkey solutions.

Content moderation is a problem that affects every service that hosts user uploaded media. From the avatars to a personal collection of pictures, the platform holds the responsibility of removing the violating content. The problem can be tackled with clssifiers, human moderators and by comparing media signatures; this presentation will be about the latter. Similarity Detection is an approach that tries to detect media based on an archive of "definitions" (yes, like the antiviruses) of things that have already been classified as violating. But how do we measure similarity between images from the perspective of a machine (not to mention video/audio clips of different lenghts)? The answer is not MD5... We'll talk how we do it, what technologies you can use too and how we can leverage a public, crowdsourced archive of signatures to defeat various threats, from terrorism to misinformation to Child Exploitation.

We present: 1) a full pipeline for unsupervised machine translation training (making use of monolingual corpora) for languages with low available resources; 2) a translation server making use of that unsupervised MT with an HTTP API compatible with Moses toolkit, a once prominent MT system; 3) a Docker packaged version of the EU funded free Computer Aided Translation (CAT) tool MateCAT for ease of deployment. This full translation pipeline enables a non technical user, speaking a non-FIGS language for which there is scarcity of parallel corpora, to start translating documents and software following translation industry standards.

Localization within community suffers from the fragmentation of technologies (too wide wedge between commercial Computer Aided Translation tools and free ones), available language resources (making difficult to train a Machine Translation) and lack of clear and robust pipelines to get started. Low resource language communities suffer the most, since MT systems require training corpora of millions of words and industry has settled to expecting the massive corpora available to FIGS (French, Italian, German, Spanish) languages. Moreover, the community suffers from a lack of adoption of established technologies and workflows, leading to reinventing the wheel and suboptimal efforts’ outcomes. Today we would like to present a connector for the implementation of an unsupervised MT (made by Artetxe et al.), that claims a BLEU of 26 on limited language resources (which is enough as a support system) integrated with MateCAT, an industry level, free, web based tool funded by EU, in order to provide a more viable alternative to resorting to Google Translate and commercial LSPs.

We present a suggester of git commit messages based on the files diff: by reading the commit patch, the system outputs a message in natural language describing the subject of the commit. While high level intent guessing is out of the scope of this project, this may provide further insights to a CI system in refusing pull requests with poor commit messages or not explaining the subject matter.

When it comes to commit messages, we all have witnessed the worst of our kind: while code is (hopefully) carefully crafted, the urge to submit our work for peer review can play badly with the helpfulness of the commit message attached (leaving a helpless reviewer in the dark and calling for the developer presence to explain the meaning of the coded feature). Machine translation techniques come to the rescue, by providing a viable aid to suggesting better commit messages from a golden standard of various situations: by analyzing thousands of commit messages along with their patches, we show that a machine can understand the semantics of a patch thus providing a way to automatically tell good messages from bad messages. We can employ this in CI environments to rule out insufficiently documented patches or to propose better alternatives to the developers.