Information Retrieval and Web Search

1
Information Retrieval and Web Search

• Keyphrase Extraction
• Instructor Rada Mihalcea
• Invited Lecturer Andras Csomai
• Class web page http//lit.csci.unt.edu/classes/C
  SCE5200/

2
Keyphrase Extraction

• Set of keywords representing the topic of a
  document
• Dense summary
• Purpose
• Topic/content identification
• Indexing
• Browsing
• Back of book
• Classification
• Surfing
• . . .
• Stems from early roman times

3
(No Transcript)
4
(No Transcript)
5
Extraction vs. assignment

• Assignment the keyphrases are not necessarily
  present in the document
• controlled vocabulary
• uncontrolled vocablary
• Extraction - On average 75 of human expert
  assigned keywords present in the document
• Conclusion extraction is a feasible approach

6
General framework of supervised keyword extraction
7
Candidate extraction

• ngrams
• Sequence of n consecutive words from text
• 90 of the keywords present in the document
  contain up two 3 terms (3grams)
• Depends on domain, style, etc.
• Boundaries.?
• Np-chunks
• 50 extracted
• Helps increase precision
• Syntactic patterns
• Using linguistic knowledge
• Extract only patterns that are likely to be
  keyphrases (e.g. noun noun vs. verb adj is )

8
Machine Learning component

• The choice of a machine learning method is of
  secondary importance
• Decision trees
• Rule induction
• Naïve Bayes
• GenEx
• The most important part is the good feature set
• Training data

9
Features

• TFIDF
• The ubiquitous IR feature
• Distance
• The distance of the candidate from the beginning
  of the document
• Better performance in well structured text (e.g.
  scientific articles)
• POS pattern
• Linguistic information
• The pos pattern of the candidate phrase
• Length
• E.g. 13.7 unigram, 51.8 bigram, 25.4 trigram

10
Features contd

• Phraseness
• Tells you if a sequence of words constitutes a
  phrase. (e.g. information retrieval vs. wood
  desk)
• PMI (point-wise mutual information)
• Probabilities via MLE
• For ranking purposes ignore constants, use simple
  counts

11
Features contd.

• Syntactic elements
• E.g. adj. at the end of sequence, common verb
• Binary feature
• Special locations or emphasized style
• Binary feature showing if the candidate appears
  in a title, metadata, etc.
• If it is italic, bold, etc.
• Domain specific features
• Taxonomies, ontologies
• Use the relations of the hierarchical structure
  as binary features. E.g. the presence of
  neighboring items in text
• Or simple domain specific keyphrase collection
  (Turney)

12
More features

• Coherence
• the PMI of the top K phrases and the rest of the
  candidates (Turney)
• WordNet based features
• Are you familiar with WordNet?
• PMI point-wise mutual information
• Probabilities via MLE

13
Performance of the state of the art

• Precision of .23 (average) (Turney, Gutwin)
• Large collection, longer articles
• F-measure of .33 (best system) with precision of
  .25
• Only abstractseasier?
• Acceptability 80 (good no answer), 62 good
  (Turney)

14
Unsupervised

• When you have no training data
• Language and domain independent
• TextRank (Mihalcea)
• Unsupervised, simple, powerful.
• Efficient
• Uses a variation of PageRank
• Represent text as graph

15
Representing text as graph
16
Traditional Unsupervised Methods

• TFIDF plus some improvement
• Coherence based on WordNet
• PMI is language independent
• Heuristics
• Ranking based on the TFIDF or some composite
  score
• Select the top portion of the ranked list (score
  or keyphrase number cutoff)

17
Back of the Book indexing

• Information likely to be sought by a user
• Present at the end of almost every book.

18
Style

• Index terms and references (location, cross
  reference, etc )
• Alphabetical ordering (to facilitate the search)
• Phrase heads brought up front (inverted index)
• illustrations, indexing of
• Cascaded style (organized information shows
  topical relatedness or specific instances of a
  more general concept)
• illustrations, indexing of,
• in newspaper indexes,
• in periodical indexes,

19
Automated indexing

• Observation indexes tend to contain a large
  number of Named Entities and keyphrases
• Minimally supervised indexing system
• Candidate extraction
• Ranking
• Postprocessing

20
Candidate Extraction

• 1-4grams
• no span over sentence boundaries
• Named Entities
• Using Snow
• Heuristics
• If candidate starts or ends with a stop word,
  drop it
• If candidate starts or ends with a common word,
  drop
• Identify candidates that are paraphrases

21
Ranking

• TF from document and IDF from BNC
• Named Entities receive an IDF of 1
• Rank according to the TF/IDF scores
• Heuristics
• Index length (0.44 or 0.35 of text depending
  on target)
• Distribution of entries by length

22
Post processing the index

• Eliminate paraphrases
• Look for paraphrases in the candidate list, and
  keep only the instance with the highest TFIDF
  score
• Targeted paraphrase types
• Lexical synonymy trace the river / follow the
  river
• PP attachment a plant in Alabama / the
  Alabama plant
• Morpho-syntactic variants inductive phenomena
  / phenomena of induction, plural forms, etc

23
Detecting paraphrases

• Create an extended set for all non-common words
  of the entry including
• the word itself
• the stem (using Porters algorithm)
• The synonyms of the first sense of the word, or
  its variations in other parts of speech
• Two entries are paraphrases if
• number of non-common words is equal
• There is a one-to-one matching of the extended
  sets of the two entries

24
Postptrocessing contd

• Generate inversions following human style
• Find the word with the highest TFIDF component
• Split the entry in three entry(A,B,C), where B
  is the most important word, A or B may be empty
• Reorder into B,A,C
• E.g.
• centres of innervation / innervation, centres of
• thrust of the sting / sting, thrust of the
• concentrated nervous sytem /nervous,
  concentrated, system