Literature Extraction: Entities and Relations

1
Literature ExtractionEntities and Relations

• ChengXiang (Cheng) Zhai
• Department of Computer Science
• Institute for Genomic Biology
• Statistics
• Graduate School of Library Information Science
• University of Illinois at Urbana-Champaign

ABC Workshop, UIUC, Dec. 5-6, 2007
2
Outline

• General Background on Text Information Management
• Information Extraction Entities Relations
• Towards Automated Gene Annotation

3
Text Management Technologies
Mining
Access
Select information
Create Knowledge
Add Structure/Annotations
Organization
4
Text Information Access

• Search Works well if you know exactly what you
  want (e.g., PubMed)
• Navigation More useful for exploring information
  space or when you cant formulate a good query
• Recommendation Push information to a user

5
Text Information Organization

• Summarization
• Single document vs. multiple documents,
  unstructured vs. structured
• Helps digest information
• Categorization
• Classify text into predefined categories (e.g.,
  different GO terms)
• Adds structures to text helps browsing direct
  prediction
• Clustering
• Group similar text segments or articles into the
  same cluster
• Helps reveal underlying structures

6
Text Mining

• Information Extraction
• Pulling out entities (e.g., genes and proteins)
  and relations (e.g., protein interactions)
• Helps semantic analysis and inferences
• Topic Extraction
• Tease out subtopics/themes in text (e.g.,
  separate multiple subtopics in the same article)
• Helps summarization and browsing
• Inferences
• Attempts to create new knowledge (e.g., Gene A
  has function F Gene A and Gene B are similar ?
  Gene B has function F)
• Pattern Discovery
• Discover outliers, correlations/associations,
  clusters, trends, (e.g., gene A and gene B tend
  to occur in similar context)
• Helps create new knowledge

7
Key Technique Information Retrieval

• Text/topic representation Vector-space,
  probabilistic models
• Term weighting TF-IDF
• Text matching (similarity) Vector/distribution
  similarity functions
• User feedback (learning about what a user is
  interested in) machine learning

8
Key Technique Machine Learning

• Computation YF(X)
• Knowledge-based Specify a recipe (program)
• Data-driven Provide examples of (X,Y) pairs
• Basic setup
• Given training data (many (X,Y) pairs)
• Assume some kind of function relation between X
  and Y often with parameters
• Fit the function to the training data to set
  parameters
• Hope the learned function to be able to compute Y
  for new X
• Generally require many training examples
  (supervised learning)
• May also work without training examples
  (clustering)

9
Key Technique Natural Language Processing

• Basic Tasks
• Part-of-speech tagging (recognizing nouns, verbs,
  )
• Syntactic parsing (recognizing sentence
  structure)
• Semantic analysis (trying to get the meaning)
• State of the art methods rely on machine learning
  supplemented by some limited linguistic knowledge
  
• Generally a very difficult task, but easier for a
  specific domain such as biomedical literature

10
Massive Entity Recognition

• Class1 Small Variation (Dictionary/Ontology)
• Organism, Anatomy , Biological Process, Pathway,
  Protein Family
• Class2 Medium Variation
• Gene, cis Regulatory Element
• Class3 Large Variation
• Phenotype, Behavior

11
Massive Relation Extraction

• Expression Location
• the expression of a gene in some location
  (tissues, body parts)
• Homology/Orthology
• one gene is homologous to another gene
• Biological process
• one gene has some role in a biological process
• Genetic/Physical/Regulatory Interaction
• one gene interacts with another gene in a certain
  fashion (3 types of relations)
• a simple case Protein-Protein Interaction (PPI)

12
Generating New Knowledge

• Entity Relation Graph Mining
• Logic-based Inferences

13
Example of Interactive Graph Mining
Behavior B2
isa
isa
Co-occur-fly
Co-occur-bee
Behavior B1
Gene A1
Behavior B4
Behavior B3
Orth-mos
Co-occur-mos
Co-occur-fly
Gene A1
Gene A2
Gene A3
Reg
Reg
Reg
orth
Reg
Gene A4
Gene A4
Gene A5

• 1.XNeighborOf(B4, Behavior, co-occur,isa)
  B1,B2,B3
• 2. YNeighborOf(X, Gene, c-occur, orth
  A1,A1,A2,A3
• 3. YY A5, A6 A1,A1, A2, A3,A5,A6
• 4. ZNeighborOf(Y, Gene, reg) A4, A4

14
Logic-Based Discovery

• Encode all kinds of knowledge in the same
  knowledge representation language
• Perform logic inferences
• Example
• Regulate (GeneA, GeneB, ContextC). Literature
  mining
• SeqSimilar(GeneA,GeneA) Sequence mining
• Regulate(X,Y,C)? Regulate(Z,Y,C)
  SeqSimilar(X,Z) Human knowledge
• ? Regulate(GeneA,GeneB,ContextC)
• ADD InPathway(GeneB, P1)
• InPathway(X,P)? Regulate(X,Y,C) InPathway(Y,P)
  Human knowledge
• ? InvolvedInPathway(GeneA,P1)

15
Towards Automated Annotation

Genome
Name 1

Name 2
Name k
2. Gene Summarization
6. Relation Retrieval
1. Literature Search
4. Gene-Term Association
Literature
7. Inferences
Relevant Text
3. GOTerm Summarization
Entities Relations
Gene Ontology
5. New Term Suggestion
Term1 Term2 . Term n
16
Annotation Support Technologies

• Level 1 Literature Search
• document filtering (locate relevant articles)
• relevant passage retrieval (locate relevant
  passages)
• Level 2 Semi-automatic annotation
• Gene summarization
• GO term summarization (profiling)
• Gene-Term association analysis
• New GO term suggestion
• Level 3 Automatic annotation Gene?GO
• Relation retrieval (direct mentioning)
• Inferences/Relation mining (Inferred knowledge)

17
Thank You!