Parallel Clustering of English Verbs into Levin Classes

1
Parallel Clustering of English Verbs into Levin
Classes

• 6.338/18.337 Final Project
• Melanie Goetz Andrew Hogue
• May 13, 2004

2
Background

• Levin 1993 hand-classified verbs
• 3086 verbs into 264 classes (with overlaps)
• Utilized verb arguments and alternations
• E.g. the glass broke or broke the glass
• Classes correlated with semantic meaning of verbs

3
Our Approach

• Automatically classify verbs
• Build graph G with node for each word, edges if
  words appear in same sentence
• First, build bipartite graph with verbs and
  prepositions
• Extend with subject nouns, object nouns
• Use spectral partitioning to divide verbs into
  classes

4
Our Approach
5
Our Approach
6
Parallel Implementation

• Three components
• Extract meaningful words from parsed corpus
• Merge per-processor sparse matrices without
  bringing data to front end
• Run parallel spectral partitioning on full graph

7
Parsing

• Embarrassingly parallel
• Wall Street Journal corpus of 99 documents
• Each processor separately extracts tree from
  corpus and relevant words from tree

8
Indexing

• Need to combine matrices from separate processors
  into one indexing scheme
• Bringing to front end is inefficient
• Solution share vocabulary lists between
  processes
• Allows each process to use the same index for
  each word

9
Indexing
10
Indexing
11
Partitioning

• Based on specpart.m from Meshpart toolkit
• Serial version uses Cholesky decomposition
• Our parallel version uses eigs() function as we
  only need a few eigenvalues

12
(No Transcript)
13
Results

• Clustered 3317 sentences from Wall Street Journal
  corpus
• 2827 unique words
• Included subjects, verbs, objects, prepositions

14
Results - Parsing
15
Results - Indexing
May 13, 2004
6.338/18.337 Final Project
15
16
Results - Partitioning
May 13, 2004
6.338/18.337 Final Project
16
17
Results - Clustering
May 13, 2004
6.338/18.337 Final Project
17
18
Future Work

• Parse other corpora (Project Gutenberg)
• Restrict word types to verb/preposition or
  subject/verb/object
• Other ways to use eigenvectors for partitioning
  into more than 2 parts