We describe a new probabilistic Sentence Tree Language Modeling approach that captures term dependency patterns in Topic Detection and Tracking's (TDT) Story Link Detection task. New features of the approach include modeling the syntactic structure of sentences in documents by a sentence-bin approach and a computationally efficient algorithm for capturing the most significant sentence level term dependencies using a Maximum Spanning Tree approach, similar to Van Rijsbergen's modeling of document-level term dependencies.

We describe a new probabilistic approach in the language modeling framework that captures adaptively the local term dependencies in documents. The new model works by boosting scores of documents that contain topic-specific local dependencies and exhibits the behavior of the unigram model in the absence of such dependencies. Contributions of the current work include adapting van Rijsbergen 's [14] work in the classical probabilistic framework to the language modeling framework and adaptive modeling of withinsentence dependencies.

We describe a new probabilistic graph-based language Model that captures adaptively the local term dependencies in documents. The new model works by boosting scores of documents that contain topic-specific local dependencies and exhibits the behavior of the unigram model in the absence of such dependencies. New features of the approach include modeling the syntactic structure of sentences in documents and a computationally efficient algorithm for capturing the most significant within-sentence term dependencies using a Maximum Spanning Tree approach, similar to Keith van Rijsbergen's [13] modeling of document-level term dependencies.