View source for Mayfield et al, CoNLL 2003

From Cohen Courses

← Mayfield et al, CoNLL 2003

You do not have permission to edit this page, for the following reason:

---

You can view and copy the source of this page.

Being edited by Rui Correia == Citation == James Mayfield, Paul McNamee, and Christine Piatko. 2003. Named Entity Recognition using Hundreds of Thousands of Features. In Proceedings of CoNLL-2003. == Online version == [http://www.cnts.ua.ac.be/conll2003/pdf/18487may.pdf] == Summary == In this [[Category::paper]] the authors address the problem of [[AddressesProblem::Named Entity Recognition]] using [[UsesMethod::Support Vector Machines]] to capture transition probabilities in a lattice: a method they called SVM-lattice. The main goal is to provide a language independent [[AddressesProblem::Named Entity Recognition]] system, considering hundreds of thousands of language-independent features, which [[UsesMethod::Support Vector Machines|SVM]]'s decide if are relevant or not. In most [[AddressesProblem::Named Entity Recognition|NER]] systems, handling large numbers of features is expensive and might result in overtraining, demanding for a wise and informed feature selection. The solution proposed by the authors is to built a lattice for each sentence (where the vertices are the name taggs and the edges are the possible transitions) and compute the edges' transitions probabilities. When these transitions are computed, the authors apply the [[UsesMethod::Viterbi]] algorithm to find the best path and decide on the set of tags of the sentence. This strategy outperforms the [[UsesDataset::CoNLL-X|CoNLL-2003 Shared Task]] baseline and a classic [[HMM]] approach as well. The fact that it can include many features with little concern for their relevance or independence makes this solution particularly well suited for language-neutral entity recognition. == Brief Description of the Method == A lattice is built for each sentence where each column contains one vertex for each possible tag and is connected by an edge to every vertex in the next column that represents a valid transition. To compute these transitions, the authors exploit some important properties of [[UsesMethod::Support Vector Machines|SVM]]'s: being able to handle very high dimensional spaces and being resistant to overfitting. When these transitions are finally estimated and applied to the lattice, the authors run [[UsesMethod::Viterbi]] to find the most likely path, which identifies the final tag for each word of the sentence. However, standard [[UsesMethod::Support Vector Machines|SVM]]'s do not provide such probabilities. To accomplish this, the authors used a parametric model to fit the posterior <math>P(y=1|f)</math> directly. Since the class-conditional densities between the margins of a [[UsesMethod::Support Vector Machines|SVM]] are exponential and the Baye's rule on two exponentials suggests using a parametric form of a sigmoid, the formulation of the posterior can be written as: <math> P(y=1|f) = \frac{1}{1(1+ \exp(Ax +b))} </math> where the authors fix <math>A=-2</math> and <math>b=0</math>, recognizing that a estimation of these parameters could give best results. In order to classify a given sentence, after training, each word of the input is represented by a vector of features (such as the word itself, character n-grams, word length and position in the sentence, capitalization pattern, etc.). Each classifier (one for each transition type) is then applied to this vector to produce a margin, that is then mapped to a probability estimate. When all the probabilities have been computed, the [[UsesMethod::Viterbi]] algorithm computation takes place. == Experimental Results == The authors evaluated their approach using the [[UsesDataset::CoNLL-X|CoNLL-2003]] training and test sets. In this paper 4 entity taggers are compared: TnT tagger [http://arxiv.org/pdf/cs/0003055] , TnT + subcat (reduction over the possible categories, and therefore transitions, removing the less common ones), SVM-Lattice (uses the same categories as TnT+subcat) and SVM-Lattice+ (that uses the result of the TnT+subcat version as a feature). The results of the F-measure for these tests are present in the next table: [[File:MayfieldResults.png]] The improvement from TnT to TnT + subcat is merely due to a reduction of the considered classes. However, SVM-Lattice achieves a slightly better improvement over the TnT tagger, and the ultimate result consists of an improvement of about 3% over the TnT tagger. The authors never comment on the significance of these results, and do not present results of the SVM-Lattice method with the complete set of standard taggs used in NER competitions. ==Related Papers == [[RelatedPaper::Thorsten Brants. 2000. TnT – a statistical part-of-speech tagger. In ANLP 6, pages 224–231.]][http://arxiv.org/pdf/cs/0003055]

Return to Mayfield et al, CoNLL 2003.