Mayfield et al, CoNLL 2003

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

[1]

Summary

In this paper the authors address the problem of Named Entity Recognition using 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 Named Entity Recognition system, considering hundreds of thousands of language-independent features, which SVM's decide if are relevant or not.

In most 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 Viterbi algorithm to find the best path and decide on the set of tags of the sentence.

This strategy outperforms the 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

Each sentence is processed individually. 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 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 Viterbi to to find the most likely path, which identifies the final tag for each word of the sentence.

However, standard SVM's do not provide such probabilities. Used a parametric model to fit the posterior $P(y=1|f)$ directly. Since the class-conditional densities between the margins of a SVM are exponential and the Baye's rule on two exponentials suggests using a parametric form of a sigmoid, comes that:

$P(y=1|f)={\frac {1}{1(1+\exp(Ax+b))}}$

The authors fix A=-2 and b=0.

A different SVM model is trained for each transition type.

To evaluate a test set, 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 patter, etc.). Each classifier 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 Viterbi algorithm computation takes place.

Experimental Results

The authors evaluated their approach in the CoNLL-2003 training and test sets. The authors compare 4 entity taggers: TnT tagger, 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 this tests are present in the next table:

The improvement from TnT to TnT + subcat is merely due to a reduction of the considered classes. SVM-Lattice achieves a slighlty better improvement over the TnT tagger, and the ultimate result results an improvement of about 3% over the TnT tagger. However 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.