Fast Phrase Querying with Combined Indexes

1
Fast Phrase Querying with Combined Indexes

• Hugh E. Williams, Justin Zobel, and Dirk Bahle
• hugh, jz, dbahle_at_cs.rmit.edu.au
• School of Computer Science and
• Information Technology, RMIT University

2
Overview

• An overview of phrase querying
• Word-position inverted indexes
• Use in evaluating phrase queries
• Phrase indexes for phrase querying
• Partial phrase indexes
• Nextword indexes
• Combined indexes for fast, space-efficient phrase
  querying
  
• Results
• Conclusions

3
Phrase Querying

• Searching in text databases is enhanced by
  providing a rich set of alternative methods for
  expressing information needs
  
• A phrase query matches documents or phrases that
  have the same word ordering. Unambiguous.
  
• For example, a query may be to find all documents
  containing the exact phrase All the President's
  men
  
• Collection documents are usually indexed so that
  punctuation defines a phrase
  
• For example, the query will not match the
  document fragment all the presidents. Men,
  
  
• Phrase queries might also be partially specified
  All the Pres' or All the men

4
Phrase Querying

• Around 8 of queries from the Excite query logs
  we used in our experiments contain phrases
  enclosed in quotation marks
  
• Median phrase length is 2 words and mean is 2.5
  around 34 contain three or more words
  
• Around 41 of the remaining non-phrase queries
  actually match a phrase in a 21Gb Web collection
  we used (derived from the 1997 TREC VLC2)
  
• Observation users often pose phrase queries
  without explicitly using the query operator
  (usually quotation marks)
  
• Investigated in our recent TREC terabyte track
  experiments (the talks next week!)

5
More Observations...

• Around 11 of the explicit phrase queries contain
  one of the three commonest words the, to, and
  of
  
• But only 0.4 of these are terminated by the
  common word
  
• Around 14.5 include one of the twenty commonest
  terms
  
• Common words play different roles
• Structural only tower of london
• Important flights to london, flights from
  london
  
• Essential the who, the the, this is the
  end
  
• Ignoring common words leads to very fast query
  evaluation (details later)

6
What if we ignore common words?

• Acceptable the query tower -- london may
  match
  
• tower of london ...
• tower in london ...
• tower, while london
• Unacceptable examples
• The query -- -- -- end matches any document
  containing three stopwords and then end
  
• The query the who may be completely stopped
• Conclusion stopping of common words may be
  important for efficiency, but can have an
  unpredictable effect on user searching

7
A Word-Position Inverted Index

• Inverted indexes are the standard method of
  supporting querying of large text databases
  
• Example hampshire occurs in 53 documents, the
  first of which is ordinal document 9, where it
  occurs 3 times at word positions 3, 8, and 90

8
Evaluating Phrase Queries

• Resolving phrase queries is straightforward, but
  that does not mean it is fast
  
• The phrase new hampshire can be identified by
  retrieving and processing the postings lists for
  new and hampshire
  
• A logical AND of the lists would identify the
  phrase beginning at position 7 in document 9 (and
  perhaps in other documents) but not in document
  5
• The evaluation process
• 1. Essential sort terms from rarest to most
  common
  
• 2. Retrieve the shortest postings list
  (hampshire)
  
• 3. Decode and create an in-memory list of
  candidate matches
  
• 4. Using the in-memory list
• 4a. Retrieve the next shortest postings list
  (new)
  
• 4b. Merge the new postings list with the
  in-memory list to identify (partial) phrases
  
• 4c. Repeat from 4a until there are no more lists

9
Evaluation Costs

• The following graph shows the results of running
  132,276 phrase queries on a 21.9Gb collection of
  Web data (from the TREC Web track data)
  
• Result Stopping only three words reduces query
  time by 60!

10
Evaluation Costs

• The following graph shows how stopping with a
  hand-crafted list of 490 stopwords affects the
  speed of queries of different lengths (same
  queries as previously, but the 1997 TREC WT10g
  web collection)
• Result Stopping improves speed five-fold, and is
  more effective for longer queries

11
Phrase Indexes

• Phrase indexes are designed to improve the speed
  of phrase querying, ideally with low main-memory
  and disk space overheads
  
• A phrase index is an inverted index, where the
  vocabulary entries are phrases
  
• Cannot be used for other purposes word-level
  inverted index still needed
  
• It stores entries for selected phrases, and is
  therefore useful for only those phrases complete
  phrase indexes are unlikely to be practical
  
• Two types
• Complete index of phrases of a fixed length l
  (the nextword index discussed later is of this
  type, for l2)
  
• Incomplete phrase index for phrases of arbitrary
  length
  
• We experiment with both, and discuss the latter
  first

12
Partial Phrase Indexes

• A partial phrase index stores phrases that are
  exact matches to queries (Gutwin et al., 1998)
  
• The following is an example for five common
  phrases
  
• We do not store word positions, as we only
  experiment with exact matching
  

13
Using partial phrase indexes

• Query evaluation is straightforward
• For each phrase query, check the partial phrase
  vocabulary
  
• If the phrase is found (for example, Apple
  Macintosh), then the query is resolved by
  retrieving its list
  
• If the phrase is not found, follow the
  conventional evaluation strategy using the
  word-based inverted index
  
• This structure affords the greatest savings for
  phrases containing common words
  
• For example, the list for new york is at most
  as long as the list for york and probably much
  shorter. Its certainly much shorter than the
  list for new

14
A Useful Index

• A useful partial phrase index is dependent on
  being able to successfully predict future
  queries
  
• A simple approach is to store only frequent past
  queries, and this what we did in our
  experiments
  
• For example, in our logs, thumbnail post was
  the most common (198). Other common queries were
  jennifer lopez (126), the cranberries (79),
  and santa claus (76)
• In practice, an LRU or LFU strategy might be
  better, but this requires experimental validation
  with much larger query sets (which we now have)

15
Experiments

• We experimented with a partial phrase index,
  combined with a complete inverted index
  
• The partial phrase index is trained using the
  first 66,000 queries from our Excite log, and
  tested using the remaining 66,000
  
• We tested keeping the 100, 1000, and 10000 most
  common queries in the partial phrase index
  
• For 10000, there are many singularly-occuring
  phrases (choice is arbitrary)
  
• Experiments were on the WT10g web collection

16
Results

• Storing 10,000 queries in the partial phrase
  index reduces query evaluation costs by around 15

17
Results

• Around 70 of the queries in the 10,000 word
  index are two words in length, 21 are three
  words, and 6 are four words
  
• The scheme works best for short two word queries,
  which improve by around 29
  
• The largest index is only 12.8 Mb (or 0.1 of the
  collection size)

18
Nextword Indexes

• Williams et al. (1999) proposed a nextword index
  as a structure to support fast phrase querying
  
• Advantages no additional main-memory costs
  works well for phrases of length two

19
Nextword Indexes...

• Nextword indexes can be used to evaluate two word
  phrase queries as follows
  
• 1. Find the word new in an in-memory vocabulary
  and retrieve the pointer to a nextword list on
  disk
  
• 2. Process the nextwords of new to find
  hampshire and retrieve the pointer to the
  inverted list on disk
  
• 3. Retrieve and process the inverted list (and
  then either display the document identifiers and
  offsets, or retrieve and display the documents)
  
• For longer queries, each word need only be
  evaluated as a word or as a nextword the
  evaluation process is the same as for a
  conventional inverted index
• Example the query the cat sat on the mat can
  be evaluated by retrieving and merging the lists
  for the cat, sat on, and the mat
  
• For queries with an odd number of words, there is
  a choice as to which of the non-end pairs to
  evaluate

20
A Subtle Change

• Since it was originally proposed in 1999, we have
  experimented with many optimisations to the
  nextword structure
  
• In the experiments we describe here, we store
  every nth nextword in memory on disk, the
  nextwords are front-coded for compact storage
  
• This allows the short, in-memory list to be
  scanned for a near match, and then the on-disk
  list to be directly accessed rather than
  completely decoded from its beginning

21
Evaluation Costs

• The nextword index structure is large
• For our collection of 10Gb, the compressed
  nextword index is 2.75Gb in size, almost exactly
  twice the size of the inverted index alone
  
• But the saving in query evaluation time is
  dramatic
  
• Average evaluation time is 0.02 seconds, or
  around 50 times faster than with the inverted
  index
  
• Two word queries are 35 times faster
• Five word queries are 15 times faster
• Observation not surprisingly, most of the speed
  gain is for non-rare words there is little
  improvement for rare words
  

22
Combined Nextword Inverted Indexes

• Proposal Combine a partial nextword index with a
  complete conventional index to speed phrase
  queries

23
Combined Nextword Inverted...

• In a combined inverted nextword index, only
  common firstwords have a nextword list
  
• We call this a top frequency based scheme
• This leads to the following strategy
• 1. Find all words with nextword lists
  (firstwords). Record inverse document frequency
  (IDF)
  
• 2. Find all remaining words not in a
  firstword-nextword pair. Record IDF.
  
• 3. Process unique words by increasing IDF (using
  nextword lists where possible)
  
• 4. Display results
• Query evaluation stops when
• There are no lists left to process
• There are no candidates left in the in-memory list

24
Evaluation Costs

• The combined indexing strategy works well
• An additional around 192Mb (or 13 of the
  inverted index size) for the three most-common
  words
  
• Query time is halved, and is only 0.3 seconds
  slower on average per query than with 490-word
  stopping
  

25
Combining all three

• We examined combining a partial nextword, partial
  phrase, and complete inverted index
  
• Advantages phrase containing common words can
  use the nextword index, while common phrases
  (which rarely contain common words) can be
  evaluated using the partial phrase index.
  Difficult cases use the inverted index only
• This works as follows
• Look for the phrase in the partial phrase index
  vocabulary. If found, return answer if not,
  continue
  
• Evaluate the phrase query using the combined
  nextword inverted strategy

26
Three-way results

• Results for a 10,000 phrase partial phrase index
  on WT10g, with varying numbers of terms in the
  nextword index (0, 3, 6, and 192)

27
Three-way Results

• The schemes are complementary
• The partial phrase and partial nextword schemes
  improve different sets of queries
  
• Queries of length two improve by 60 compared to
  phrase inverted, and by 15 compared to
  nextword inverted
  
• Depending on the choice of the number of
  nextwords, the three-way combination is 60 to
  80 faster than an unstopped inverted index alone

28
Conclusions

• Phrase queries are an important class of
  queries
  
• Around 8 of queries explicitly involve a phrase
• Many more can be evaluated as phrase queries
• However, phrase queries are costly to evaluate
• Phrase indexes offer an efficient solution to
  phrase querying
  
• With a complete nextword index, querying is 50
  times faster, but the index is twice the size as
  an inverted index
  
• With a partial phrase index of 10,000 words,
  querying is 15 faster and index is 1 of the
  size of an inverted index

29
Conclusions

• Solution we propose a combined strategy use
  partial phrase, then partial nextword inverted
  indexes
  
• With only three common words, this adds 13 to
  the index size and reduces phrase query
  evaluation times by 60
  
• Conclusion Using a combined index, queries with
  stopwords can be efficiently evaluated with only
  a small increase in index size. Stopping is
  unnecessary
• Questions?

30
Pointers ( advertising!)

• The Search Engine Group, http//www.seg.rmit.edu.a
  u/
  
• The zettair search engine, http//www.seg.rmit.edu
  .au/zettair/
  
• My home page, http//www.cs.rmit.edu.au/hugh
• H.E. Williams, J. Zobel, and D. Bahle, Fast
  Phrase Querying with Combined Indexes'', ACM
  Transactions on Information Systems, 22(4),
  573-594, 2004.
• D. Bahle, H.E. Williams, and J. Zobel,
  Efficient Phrase Querying with an Auxiliary
  Index'', In Proc. of the ACM-SIGIR International
  Conference on Research and Development in
  Information Retrieval (Tampere, Finland), K.
  Jarvelin, M. Beaulieu, R. Baeza- Yates, and S. H.
  Myaeng, Eds., 215-221, 2002.
• D. Bahle, H.E. Williams, and J. Zobel,
  Compaction techniques for nextword indexes.
  In Proc. of the String Processing and Information
  Retrieval Symposium (San Rafael, Chile). IEEE
  Computer Society Press, 33--45, 2001.  
• D. Bahle, H.E. Williams, and J. Zobel,
  Optimised phrase querying and browsing in text
  databases. In Proc. of the Australasian
  Computer Science Conference (Gold Coast,
  Australia), M. Oudshoorn, Ed. Conferences in
  Research and Practice in Information Technology,
  11--19, 2001.
• H.E. Williams, J. Zobel, and P. Anderson,
  What's next? Index structures for efficient
  phrase querying. In Proc. of the Australasian
  Database Conference (Auckland, New Zealand), M.
  Orlowska, Ed. Springer-Verlag, 141--152, 1999.