We present a novel multi-modal unspoken punctuation prediction system for the English language which combines acoustic and text features. We demonstrate for the first time, that by relying exclusively on synthetic data generated using a prosody-aware text-to-speech system, we can outperform a model trained with expensive human audio recordings on the unspoken punctuation prediction problem. Our model architecture is well suited for on-device use. This is achieved by leveraging hash-based embeddings of automatic speech recognition text output in conjunction with acoustic features as input to a quasi-recurrent neural network, keeping the model size small and latency low.

Automatically extracting information from social media is challenging given that social content is often noisy, ambiguous, and inconsistent. However, as many stories break on social channels first before being picked up by mainstream media, developing methods to better handle social content is of utmost importance. In this paper, we propose a robust and effective approach to automatically identify microposts related to a specific topic defined by a small sample of reference documents. Our framework extracts clusters of semantically similar microposts that overlap with the reference documents, by extracting combinations of key features that define those clusters through frequent pattern mining. This allows us to construct compact and interpretable representations of the topic, dramatically decreasing the computational burden compared to classical clustering and k-NN-based machine learning techniques and producing highly-competitive results even with small training sets (less than 1'000 training objects). Our method is efficient and scales gracefully with large sets of incoming microposts. We experimentally validate our approach on a large corpus of over 60M microposts, showing that it significantly outperforms state-of-the-art techniques.

Unsupervised template induction over email data is a central component in applications such as information extraction, document classification, and auto-reply. The benefits of automatically generating such templates are known for structured data, e.g. machine generated HTML emails. However much less work has been done in performing the same task over unstructured email data.

Applying classical time-series analysis techniques to online content is challenging, as web data tends to have data quality issues and is often incomplete, noisy, or poorly aligned. In this paper, we tackle the problem of predicting the evolution of a time series of user activity on the web in a manner that is both accurate and interpretable, using related time series to produce a more accurate prediction. We test our methods in the context of predicting signatures for online petitions using data from thousands of petitions posted on The Petition Site - one of the largest platforms of its kind. We observe that the success of these petitions is driven by a number of factors, including promotion through social media channels and on the front page of the petitions platform. We propose an interpretable model that incorporates seasonality, aging effects, self-excitation, and external effects. The interpretability of the model is important for understanding the elements that drives the activity of an online content. We show through an extensive empirical evaluation that our model is significantly better at predicting the outcome of a petition than state-of-the-art techniques.