Many applications that process social data, such as tweets, must extract entities from tweets (e.g., "Obama" and "Hawaii" in "Obama went to Hawaii"), link them to entities in a knowledge base (e.g., Wikipedia), classify tweets into a set of predefined topics, and assign descriptive tags to tweets. Few solutions exist today to solve these problems for social data, and they are limited in important ways. Further, even though several industrial systems such as OpenCalais have been deployed to solve these problems for text data, little if any has been published about them, and it is unclear if any of the systems has been tailored for social media. In this paper we describe in depth an end-to-end industrial system that solves these problems for social data. The system has been developed and used heavily in the past three years, first at Kosmix, a startup, and later at Wal-martLabs. We show how our system uses a Wikipedia-based global "real-time" knowledge base that is well suited for social data, how we interleave the tasks in a synergistic fashion, how we generate and use contexts and social signals to improve task accuracy, and how we scale the system to the entire Twitter firehose. We describe experiments that show that our system outperforms current approaches. Finally we describe applications of the system at Kosmix and WalmartLabs, and lessons learned.

This paper presents and evaluates several original techniques for the latent classification of biographic attributes such as gender, age and native language, in diverse genres (conversation transcripts, email) and languages (Arabic, English). First, we present a novel partner-sensitive model for extracting biographic attributes in conversations, given the differences in lexical usage and discourse style such as observed between same-gender and mixedgender conversations. Then, we explore a rich variety of novel sociolinguistic and discourse-based features, including mean utterance length, passive/active usage, percentage domination of the conversation, speaking rate and filler word usage. Cumulatively up to 20% error reduction is achieved relative to the standard Boulis and Ostendorf (2005) algorithm for classifying individual conversations on Switchboard, and accuracy for gender detection on the Switchboard corpus (aggregate) and Gulf Arabic corpus exceeds 95%.

This paper presents novel improvements to the induction of translation lexicons from monolingual corpora using multilingual dependency parses. We introduce a dependency-based context model that incorporates long-range dependencies, variable context sizes, and reordering. It provides a 16% relative improvement over the baseline approach that uses a fixed context window of adjacent words. Its Top 10 accuracy for noun translation is higher than that of a statistical translation model trained on a Spanish-English parallel corpus containing 100,000 sentence pairs. We generalize the evaluation to other word-types, and show that the performance can be increased to 18% relative by preserving part-of-speech equivalencies during translation.

No abstract

Challenging the implicit reliance on document collections, this paper discusses the pros and cons of using query logs rather than document collections, as self-contained sources of data in textual information extraction. The differences are quantified as part of a large-scale study on extracting prominent attributes or quantifiable properties of classes (e.g., top speed, price and fuel consumption for CarModel) from unstructured text. In a head-to-head qualitative comparison, a lightweight extraction method produces class attributes that are 45% more accurate on average, when acquired from query logs rather than Web documents.