Filtering-Ranking Perceptron Learning for Partial Parsing

Abstract

Abstract. This work introduces a general phrase recognition system based on perceptrons, and a global online learning algorithm to train them together. The method applies to complex domains in which some structure has to be recognized. This global problem is broken down into two layers of local subproblems: a filtering layer, which reduces the search space by identifying plausible phrase candidates; and a ranking layer, which builds the optimal phrase structure by discriminating among competing phrases. A recognition-based feedback rule is presented which reflects to each local function its committed errors from a global point of view, and allows to train them together online as perceptrons. As a result, the learned functions automatically behave as filters and rankers, rather than binary classifiers, which we argue to be better for this type of problems. Extensive experimentation on partial parsing tasks gives state-of-the-art results and evinces the advantages of the global training method over optimizing each function locally and independently.

A Two-Phase Framework for Learning Logical Structures of Paragraphs in Legal Articles

“…Baseline: Filter-Ranking Perceptron Algorithm. We chose the Filter-Ranking (FR) Perceptron algorithm proposed by Carreras et al [2002] and Carreras and Marquez [2005] as our baseline model because of its effectiveness on phrase recognition problems, especially on problems that accept the embedded relationship 18 . We use the FR-perceptron algorithm to recognize logical parts in law sentences one by one in an input paragraph.…”

“…To learn parameter vectors for these functions, a perceptron-like algorithm was introduced [Carreras and Marquez 2005;Carreras et al 2002]. The main advantage of the FR-perceptron algorithm is that it can learn parameters for both classification functions and score function simultaneously.…”

RelHunter: a machine learning method for relation extraction from text

“…Carreras et al [4] propose the Filtering-Ranking Perceptron (FRP) system for two phrase recognition problems: phrase chunking and clause identification. The FRP modeling is strongly related to the two previously cited systems.…”

“…Carreras et al [4] propose the Filtering-Ranking Perceptron (FRP) system for this general task. The FRP task modeling is strongly related to previous proposals [3,21]; however, FRP simultaneously learns to solve the three subtasks.…”

“…To learn how to extract the entities and the relation, the method uses two additional schemes: task decomposition and interdependent classification. The RelHunter's task decomposition strategy is inspired by the modeling in [21] for Phrase Chunking, in [3] for Clause Chunking, and in [4] for Phrase Recognition. RelHunter uses Entropy Guided Transformation Learning (ETL) [6,16] as its basic learning engine.…”