A Webbased Kernel Function for Measuring the Similarity of Short Text Snippets

1
A Web-based Kernel Function for Measuring the
Similarity of Short Text Snippets

• WWW 2006
• M.Sahami, T.Heilman
• Google Inc.

2
Determining the similarity of short text

• Traditional document similarity measures often
  produces inadequate results
• Need a method that captures more of the semantic
  context of the snippets rather than simply
  measuring their term-wise similarity.

3
Other means of determining query similarity

• Raghavan Sever, 1995
• Differences in the ordering of documents
  retrieved in response
• Fitzpatrick Dent, 1997
• Normalized set overlap (intersection) of the top
  200 documents retrieved for each query.
• In the context of Machine Learning, a great deal
  of work has extended to measuring semantic
  similarity between documents

4
The approachs advantage

• Based on query expansion techniques
• A lazy approach that we need not compute an
  expansion for a given text snippet until we want
  to evaluate the kernel function

5
Leverage the large volume of documents

• Treat each snippet as a query
• By examining documents that contain the text
  snippet terms we can discover other contextual
  terms
• Create a context vector that contains many words
  that tend to occur in context with the original
  snippet terms

6
Compute the query expansion
7
Some issues

• Create vectors using just the contextually
  descriptive text snippet for each document
• 1000 characters is sufficient
• Given two short text snippets x and y, the
  semantic similarity kernel between them is

8
Evaluation
9
Related query suggestion

• Initial repository Q of previously issued queries
• sampling search logs from the Google search
  engine
• For any newly issued query u, compute our kernel
  function K(u, qi)
• Do some post-filtering to eliminate queries that
  are too linguistically similar to each other
• Suggest related queries qi which have the highest
  kernel score with u

10
Query suggestion system
11
Conclusion and Comments

• The approach is relatively simple, but
  surprisingly quite powerful
• Works well even when the short texts being
  considered have no common terms