Difference between revisions of "Lin and Wu. 2009. Phrase Clustering for Discriminative Learning."

From Cohen Courses
Jump to navigationJump to search
Line 20: Line 20:
  
 
The phrase vectors are then clustered using [[UsesMethod::K-means]] clustering algorithm, which can be easily parallelized. The "similarity" between two phrases is simply the Euclidean norm of the phrase vectors in the feature space. What this means is that phrases are considered to be more similar if they share more context words.
 
The phrase vectors are then clustered using [[UsesMethod::K-means]] clustering algorithm, which can be easily parallelized. The "similarity" between two phrases is simply the Euclidean norm of the phrase vectors in the feature space. What this means is that phrases are considered to be more similar if they share more context words.
 +
 +
=== Soft clustering ===
  
 
In addition to just performing hard clustering (where each phrase can only belong to one cluster), they perform soft clustering as well by assigning phrases to cluster centroids that are within a threshold distance. The motivation behind doing soft clustering is that it may be better able to model the fact that phrases may contain several "senses".
 
In addition to just performing hard clustering (where each phrase can only belong to one cluster), they perform soft clustering as well by assigning phrases to cluster centroids that are within a threshold distance. The motivation behind doing soft clustering is that it may be better able to model the fact that phrases may contain several "senses".

Revision as of 03:34, 27 September 2011

Phrase clustering for discriminative learning, by D. Lin, X. Wu. In Proceedings of the Annual Meeting of the Association for Computational Linguistics and the International Joint Conference on Natural Language Processing of the Asian Federation of Natural Language Processing, 2009.

This Paper is available online [1].

Summary

This paper makes use of phrase clustering to improve on the state of the art for the Named Entity Recognition problem. They obtained 1 F-score improvement over NER systems on the CoNLL benchmark (in 2009). In their paper, they used phrases that occur more than 100 times in a 700 billion token web corpus (Lin et al., 2008).

This paper leveraged on large amount of unlabeled data to induce phrase clustering, which provided an advantage over word clustering features used in current NER systems.

Brief description of the method

Due to the large number of possible phrases, the authors used Bloom filters to decide whether a sequence of tokens is considered a phrase.

Phrases as feature vectors

Each phrase is represented as a vector of its context words. The frequency count of words appearing within a fixed-sized window is aggregated and converted into pointwise mutual information(PMI) values.

Parallel K-Means using MapReduce

The phrase vectors are then clustered using K-means clustering algorithm, which can be easily parallelized. The "similarity" between two phrases is simply the Euclidean norm of the phrase vectors in the feature space. What this means is that phrases are considered to be more similar if they share more context words.

Soft clustering

In addition to just performing hard clustering (where each phrase can only belong to one cluster), they perform soft clustering as well by assigning phrases to cluster centroids that are within a threshold distance. The motivation behind doing soft clustering is that it may be better able to model the fact that phrases may contain several "senses".

Experimental Result

The effectiveness of phrase clustering is evaluated on NER problem. For NER, they used 1-word context window and hard clustering, and a linear chain CRF with standard NER features. The baseline features contains a total of 48 feature templates.

The results on CoNLL'03 test set are as follows:

Conll results.png

The authors evaluated their phrase clustering system on the KDDCup 2005 competition. The task is to categorize 800k internet user search queries into 67 topical categories. They treated the problem as 67 separate binary classification task and trained logistic regression classifiers with regularization. Their system were on par with the winning KDDCup 2005 system.

Kddcup results.png