The amount of readily available on-line text has reached hundreds of billions of words and continues to grow. Yet for most core natural language tasks, algorithms continue to be optimized, tested and compared after training on corpora consisting of only one million words or less. In this paper, we evaluate the performance of different learning methods on a prototypical natural language disambiguation task, confusion set disambiguation, when trained on orders of magnitude more labeled data than has previously been used. We are fortunate that for this particular application, correctly labeled training data is free. Since this will often not be the case, we examine methods for effectively exploiting very large corpora when labeled data comes at a cost.

Traditionally, Information Extraction (IE) has focused on satisfying precise, narrow, pre-specified requests from small homogeneous corpora (e.g., extract the location and time of seminars from a set of announcements). Shifting to a new domain requires the user to name the target relations and to manually create new extraction rules or hand-tag new training examples. This manual labor scales linearly with the number of target relations. This paper introduces Open IE (OIE), a new extraction paradigm where the system makes a single data-driven pass over its corpus and extracts a large set of relational tuples without requiring any human input. The paper also introduces TEXTRUNNER, a fully implemented, highly scalable OIE system where the tuples are assigned a probability and indexed to support efficient extraction and exploration via user queries. We report on experiments over a 9,000,000 Web page corpus that compare TEXTRUNNER with KNOWITALL, a state-of-the-art Web IE system. TEXTRUNNER achieves an error reduction of 33% on a comparable set of extractions. Furthermore, in the amount of time it takes KNOWITALL to perform extraction for a handful of pre-specified relations, TEXTRUNNER extracts a far broader set of facts reflecting orders of magnitude more relations, discovered on the fly. We report statistics on TEXTRUNNER's 11,000,000 highest probability tuples, and show that they contain over 1,000,000 concrete facts and over 6,500,000 more abstract assertions.

We describe the architecture of the AskMSR question answering system and systematically evaluate contributions of different system components to accuracy. The system differs from most question answering systems in its dependency on data redundancy rather than sophisticated linguistic analyses of either questions or candidate answers. Because a wrong answer is often worse than no answer, we also explore strategies for predicting when the question answering system is likely to give an incorrect answer.

Extractive summarization techniques cannot generate document summaries shorter than a single sentence, something that is often required. An ideal summarization system would understand each document and generate an appropriate summary directly from the results of that understanding. A more practical approach to this problem results in the use of an approximation: viewing summarization as a problem analogous to statistical machine translation. The issue then becomes one of generating a target document in a more concise language from a source document in a more verbose language. This paper presents results on experiments using this approach, in which statistical models of the term selection and term ordering are jointly applied to produce summaries in a style learned from a training corpus.

We present a new HMM tagger that exploits context on both sides of a word to be tagged, and evaluate it in both the unsupervised and supervised case. Along the way, we present the first comprehensive comparison of unsupervised methods for part-of-speech tagging, noting that published results to date have not been comparable across corpora or lexicons. Observing that the quality of the lexicon greatly impacts the accuracy that can be achieved by the algorithms, we present a method of HMM training that improves accuracy when training of lexical probabilities is unstable. Finally, we show how this new tagger achieves state-of-the-art results in a supervised, non-training intensive framework.