The many dimensions of Information Mining

1
The many dimensions of Information Mining

• Searching textual information sources
• Thimal Jayasooriya
• thimal_at_cs.york.ac.uk

2
Introduction

• Exponentially increasing amounts of material are
  available
• Finding and making sense of this material is
  potentially useful, but difficult with present
  search technology.
• Information mining includes information
  extraction, information retrieval, natural
  language processing and document summarization.
  Dixon1997
• Many diverse research areas and conferences cover
  the topic.
• TREC (Text Retrieval Conference)
• MUC (Message Understanding Conference)
• SIGIR-ACM (ACM Special Interest Group on
  Information Retrieval)

3
Introduction continued

• Some problems which search can currently solve
• Retrieving a document with known attributes. A
  conference paper can be found by using the author
  name and the conference title.
• Retrieving documents which have been manually
  annotated or reviewed as relevant.
• Retrieving documents which contain exact matches
  for specific words or phrases.

4
Introduction - continued

• And, some problems that arent so easy
• Finding the most relevant information for a
  particular topic
• Finding out opinions (positive, negative or
  neutral)
• Inferring the bias of a writer based on his
  articles
• Cross referencing domain knowledge with another
  field
• Finding other relevant resources or similar
  documents
• Yang2002, vanRijke2003

5
Why isnt search easier ?

• Unstructured and freeform nature of text
• Not always possible to distinguish content from
  fluff.
• The content refreshing problem for web search
• Content continually changes and leaves no time
  to discover new material.
• Ambiguity
• Lack of semantic knowledge
• Difficult to automatically discern the sense
  of a sentence.
• Lack of domain knowledge

6
Why isnt search easier - continued

• Manning2000 defines several steps that are
  required to disambiguate a corpus of text
• Distinguish word sense
• Define the word category
• Decide the syntactic structure
• Resolve semantic ambiguity and scope

7
Ambiguity in text
Time flies like the wind, fruit flies like banana
8
The search procedure
Gather data source (corpora of text)
Filter and index content
End user queries indices
Retrieves documents matching query
9
Using metadata effectively

• Most search techniques rely on metadata to
  increase relevance.
• For instance, frequently occurring words are more
  likely to be the topic of a particular paper.
  vanRijsbergen1979, Manning2000
• Other forms of metadata can include formatting
  and markup. Header tags and bold text, for
  example, may be more relevant.
• HTML and some other document types can link to
  other pages. This was used by Brin1998 to
  calculate pagerank. The higher the pagerank,
  the higher the relevance.

10
Using metadata effectively - continued

• Hawking Hawking2000 extends this concept to
  searching email archives.
• Users could constrain searches by metadata, as
  well as by content.
• Search for all emails with SIGIR facm.org
  (with acm.org in the from field)
• A precursor to these papers is seen in
  InfoHarness Shklar1995 which proposes a generic
  framework for organizing any sort of metadata.

11
Data mining

• Fayyad et al. notes that data mining techniques
  can be adopted for finding useful patterns in
  information. Fayyad1996
• The use of KDD (Knowledge Discovery in Databases)
  and techniques to analyse interesting patterns
  can also be adopted for text search.
• A related concept is data organization in
  warehouses and decision support systems.
• Data is organized into dimensions, and each
  dimension is analysed using a special algorithm.
• Some analysis techniques include LSI (Latent
  Semantic Indexing), SVD (Singular Value
  Decomposition) and CFA (Correspondence Factorial
  Analysis) Mothe2000

12
Document dimensions

• In datawarehousing terminology, a dimension is "a
  structure which categorizes data in order to
  allow users to answer questions".
• Mothe employs dimensions to cluster and find
  similar documents in astronomical data.
  Mothe2000
• Mothe uses mainly five elements of a document for
  similarity detection. They are
• The title of the document
• The author name
• The journal name
• The date published
• The text of the document
• Each dimension is then analysed using Singular
  Value Decomposition or its derivatives (such as
  LSI or CFA).

13
Document dimensions - continued

• Limitations of Mothes dimensioning for plain
  text search
• Treats the entire textual content of a document
  as a single dimension.
• Has a single set of dimensions for all documents,
  regardless of type. Only five dimensions were
  discovered and used.
• Merely focuses on plotting of dimensions over
  time.

14
Enhancements to dimensions

• The accessibility dimension Roellke2001
  addresses some of the limitations in Mothes
  model.
• The content of a document is mapped to a tree
  structure.
• Each segment of text is analysed for its
  relationship to the given search topic.
• A segment of text is broken up into sub contexts
  and super contexts.
• Each search is modelled according to its term
  frequency (tf) to the inverse document frequency
  (idf) to its accessibility dimension (ac).

15
Proposed enhancements to dimensions
Entities (names, places and objects)
Dates (signifying timelines)
Opinions (positive, negative)
16
Potential uses of document dimensions

• Document dimensions can be applied to a number of
  natural language processing problems.
• Opinion classification
• Opinion classification is a mining technique used
  to discover branding and product information in
  public media.
• Traditionally requires human analysis to
  determine subject area and bias for a particular
  area.
• Document clustering
• Document clustering is used to group similar
  documents.
• It can be employed before search to reduce the
  searchable document set vanRijsbergen1979, and
  also as a means of organizing result documents.

17
Opinion classification

• Majority of attempts employ machine learning
  techniques.
• A definitive work is Das and Chen Das2001,
  which classifies investor sentiment for certain
  stocks.
• Finn et al. Finn2002 followed up with product
  classification for entertainment (movies, theatre
  and football).
• Multiple techniques (Part of Speech tagging, Bag
  of Words and Text statistics) were evaluated
• He observed that a machine learning approach had
  reduced domain transferability compared to
  linguistic techniques.
• Dini Dini2002 attempted a hybrid technique,
  combining tokenization of words with machine
  learning.

18
Opinion classification - continued

• A conclusion of both Dini and Dave et. al
  Dave2003 is that statistical methods are
  unlikely to give larger than 90 accuracy at any
  time.
• They conclude that combining non statistical
  natural language techniques and semantic
  understanding might provide better results.

19
Conclusion and implementation

• Retrieving relevant, precise information from a
  corpora of text is a hard AI problem.
• Increasing the amount of metadata stored about a
  corpus is another means of increasing search
  relevance.
• Document dimensions aim to integrate existing NLP
  techniques into a higher level metadata
  framework.
• An existing search product (Jakarta Lucene 1.3)
  will be extended and subclassed to provide
  dimension based functionality.