Research Proposal

1
Research Proposal

• by
• Navendu Garg

2
Research Question

• Can Named Entities improve Query Precision in
  the Biomedicine domain?

3
Observations

• Phrasing techniques have not been fine tuned to
  the requirements of the biomedicine domain.
• Most of the phrasing techniques have been based
  on n-gram models.
• Research in the area of phrasing and biomedicine
  specific IR systems has been done in isolation.
• Due to the diversity of the vocabulary of the
  biomedicine domain, existing phrasing techniques
  may not improve query precision.
• Biomedicine is a specialized domain with
  vocabulary rich in named entities.
• Named Entity Recognition systems have shown
  remarkable precision in the news article domain.

4
Hypotheses

• Adding named entities as phrases to the index of
  an IR system can increase the accuracy because
• The biomedicine domain has a rich and diverse
  collection of named entities.
• Users of such systems invariably use named
  entities for their search. e.g.
  N-acetyl-cysteine, 84 kDa proteins

5
Prior Work

• Named Entity Extraction
• Information Extraction task (MUC 1995)
• Methods
• Support Vector machines
• Hidden Markov Models
• Maximum Entropy Models
• Good Performance
• Primarily news articles

6
Prior Work

• Same Methods used for Biomedicine NE
• Not comparable results due to lack of training
  data.
• Evidence of improvement in results over a period
  of time.
• Research in Phrasing
• Done in isolation with IR Systems for biomedicine
• Researches using collocation has been done for
  NE.
• No known research about using NE for improving
  Query Precision

7
Methods

• Named Entity Tagging system
• Identify features to extract named entities.
• Train the supervised learning algorithm with the
  training corpus.
• Extract named entities using classification
  algorithm by tagging corpus with the named
  entities.
• Post process the named entities extracted.
• Information Retrieval system
• Extract tokens including named entities and build
  the inverted index.
• Preprocess the query.
• Run the processed query on the IR engine.
• Get the ranked documents.

8
Methods

• Support Vector Machines
• By far best results in biomedicine domain.
• Maximum Entropy Tagger with
• Boosting algorithm.
• Voted perceptron.
• Select the best result from N possible Named
  Entities.
• Flexibility in incorporating features
• Scores are discriminating for correct hypotheses.

9
Experiment Setup

• An existing IR System specific to biomedicine
  domain.
• GENIA corpus will be used to train the algorithm.
• 10,000 commonly biomedicine terms will be used
  for named entity extraction.
• 10,000 commonly used prefixes and suffixes are
  used for named entity extraction.
• MEDLINE abstracts will be used to extract Named
  Entities and build the index and queries will be
  formed out these abstract to test the system.
• To extract Named Entities certain features will
  be used such as
• Orthographic features
• Parts of Speech
• Prefixes and suffixes
• Contextual words on the left and right side of
  the possible entity.

10
Performance Estimation

• Named Entity Extraction/ Query
• Precision/Recall
• F-Measure

11
Experimentation

• Build the index
• Without Named Entities
• To get base perfomance
• With Named Entities
• To get experiment results
• Use the 3 proposed methods to arrive at a result.

12
References

• NLP in Biomedicine ACL 2003 Workshop Program
• Lee, K.J., Hwang, Y.S. Rim, H.C. "Two-Phase
  Biomedical NE Recognition based on SVMs"
• Koichi, T. Collier, N. "Bio-medical Entity
  Extraction Using Support Vector Machines"
• Shen, D., Zhang, J., Zhou, G., Su, J. Tan, C.L.
  "Effective Adaptation of Hidden Markov
  Model-based Named Entity Recognizer for
  biomedical Domain"
• Hou, W. J. Chen, H.H. " Enhancing Performance
  of Protein Name Recognizers Using Collocation"
• Computational Linguistics (ACL), Philadelphia,
  July 2002
• Michael Collins, ATT Labs-Research, Florham
  Park, New Jersey, mcollins_at_research.att.com,
  Ranking Algorithms for NamedEntity Extraction
  Boosting and the Voted Perceptron