Language modeling is the art of determining the probability of a sequence of words. This is useful in a large variety of areas including speech recognition, optical character recognition, handwriting recognition, machine translation, and spelling correction (Church, 1988; Brown et al., 1990; Hull, 1992; Kernighan et al., 1990; Srihari and Baltus, 1992). The most commonly used language models are very simple (e.g. a Katz-smoothed trigram model). There are many improvements over this simple model however, including caching, clustering, higherorder n-grams, skipping models, and sentence-mixture models, all of which we will describe below. Unfortunately, these more complicated techniques have rarely been examined in combination. It is entirely possible that two techniques that work well separately will not work well together, and, as we will show, even possible that some techniques will work better together than either one does by itself. In this...

Maximum entropy models are considered by many to be one of the most promising avenues of language modeling research. Unfortunately, long training times make maximum entropy research difficult. We present a novel speedup technique: we change the form of the model to use classes. Our speedup works by creating two maximum entropy models, the first of which predicts the class of each word, and the second of which predicts the word itself. This factoring of the model leads to fewer nonzero indicator functions, and faster normalization, achieving speedups of up to a factor of 35 over one of the best previous techniques. It also results in typically slightly lower perplexities. The same trick can be used to speed training of other machine learning techniques, e.g. neural networks, applied to any problem with a large number of outputs, such as language modeling.

A Bloom filter (BF) is a randomised data structure for set membership queries. Its space requirements are significantly below lossless information-theoretic lower bounds but it produces false positives with some quantifiable probability. Here we explore the use of BFs for language modelling in statistical machine translation. We show how a BF containing n-grams can enable us to use much larger corpora and higher-order models complementing a conventional n-gram LM within an SMT system. We also consider (i) how to include approximate frequency information efficiently within a BF and (ii) how to reduce the error rate of these models by first checking for lower-order sub-sequences in candidate ngrams. Our solutions in both cases retain the one-sided error guarantees of the BF while taking advantage of the Zipf-like distribution of word frequencies to reduce the space requirements. 1

We propose a succinct randomized language model which employs a perfect hash function to encode fingerprints of n-grams and their associated probabilities, backoff weights, or other parameters. The scheme can represent any standard n-gram model and is easily combined with existing model reduction techniques such as entropy-pruning. We demonstrate the space-savings of the scheme via machine translation experiments within a distributed language modeling framework. 1

Incorporating features extracted from clickthrough data (called clickthrough features) has been demonstrated to significantly improve the performance of ranking models for Web search applications. Such benefits, however, are severely limited by the data sparseness problem, i.e., many queries and documents have no or very few clicks. The ranker thus cannot rely strongly on clickthrough features for document ranking. This paper presents two smoothing methods to expand clickthrough data: query clustering via Random Walk on click graphs and a discounting method inspired by the Good-Turing estimator. Both methods are evaluated on real-world data in three Web search domains. Experimental results show that the ranking models trained on smoothed clickthrough features consistently outperform those trained on unsmoothed features. This study demonstrates both the importance and the benefits of dealing with the sparseness problem in clickthrough data.

Trigram language models are compressed using a Golomb coding method inspired by the original Unix spell program. Compression methods trade off space, time and accuracy (loss). The proposed HashTBO method optimizes space at the expense of time and accuracy. Trigram language models are normally considered memory hogs, but with HashTBO, it is possible to squeeze a trigram language model into a few megabytes or less. HashTBO made it possible to ship a trigram contextual speller in Microsoft Office 2007.

A Bloom filter (BF) is a randomised data structure for set membership queries. Its space requirements fall significantly below lossless information-theoretic lower bounds but it produces false positives with some quantifiable probability. Here we present a general framework for deriving smoothed language model probabilities from BFs. We investigate how a BF containing n-gram statistics can be used as a direct replacement for a conventional n-gram model. Recent work has demonstrated that corpus statistics can be stored efficiently within a BF, here we consider how smoothed language model probabilities can be derived efficiently from this randomised representation. Our proposal takes advantage of the one-sided error guarantees of the BF and simple inequalities that hold between related n-gram statistics in order to further reduce the BF storage requirements and the error rate of the derived probabilities. We use these models as replacements for a conventional language model in machine translation experiments. 1

We present a system for on-line, cumulative learning of hierarchical collections of frequent patterns from unsegmented data streams. Such learning is critical for long-lived intelligent agents in complex worlds. Learned patterns enable prediction of unseen data and serve as building blocks for higher-level knowledge representation. We introduce a novel sparse n-gram model that, unlike pruned n-grams, learns on-line by stochastic search for frequent n-tuple patterns. Adding patterns as data arrives complicates probability calculations. We discuss an EM approach to this problem and introduce hierarchical sparse n-grams, a model that uses a better solution based on a new method for combining information across levels. A second new method for combining information from multiple granularities (n-gram widths) enables these models to more effectively search for frequent patterns (an on-line, stochastic analog of pruning in association rule mining). The result is an example of a rare combination---unsupervised, on-line, cumulative, structure learning. Unlike prediction suffix tree (PST) mixtures, the model learns with no size bound but using less space than the data. It does not repeatedly iterate over data (unlike MaxEnt feature construction). It discovers repeated structure on-line and (unlike PSTs) uses this to learn larger patterns. The type of repeated structure is limited (e.g., compared to hierarchical HMMs) but still useful, and these are important first steps towards learning repeated structure in more expressive representations, which has seen little progress especially in unsupervised, on-line contexts.

This paper presents several practical ways of incorporating linguistic structure into language models. A headword detector is first applied to detect the headword of each phrase in a sentence. A permuted headword trigram model (PHTM) is then generated from the annotated corpus. Finally, PHTM is extended to a cluster PHTM (C-PHTM) by defining clusters for similar words in the corpus. We evaluated the proposed models on the realistic application of Japanese Kana-Kanji conversion. Experiments show that C-PHTM achieves 15 % error rate reduction over the word trigram model. This

In statistical language modeling, one technique to reduce the problematic effects of data sparsity is to partition the vocabulary into equivalence classes. In this paper we investigate the effects of applying such a technique to higherorder n-gram models trained on large corpora. We introduce a modification of the exchange clustering algorithm with improved efficiency for certain partially class-based models and a distributed version of this algorithm to efficiently obtain automatic word classifications for large vocabularies (>1 million words) using such large training corpora (>30 billion tokens). The resulting clusterings are then used in training partially class-based language models. We show that combining them with wordbased n-gram models in the log-linear model of a state-of-the-art statistical machine translation system leads to improvements in translation quality as indicated by the BLEU score. 1