An open discussion and exchange of ideas

1
Natural Language Processing (NLP) Visualization
and Virtual Reality (VVR)
School of Computing FACULTY OF ENGINEERING

• An open discussion and exchange of ideas
• Introduced by Eric Atwell, Language Research
  Group

2
Saman Hina (NLP seminar coordinator)

• Eric will present aspects of NLP research
  projects which involve "visualisation" of text,
  to seek advice on further visualisation
  techniques NLP researchers should consider and
  other NLPers can ask about visualisation
  techniques they could use. The VVR "angle" may be
  that current visualisation methods work mainly
  for numerical datasets, so the VVR people might
  benefit from ideas on text analytics techniques
  which might "turn text into numbers what sorts
  of number-vectors can represent meanings of
  texts, and how to extract them.

3
Typical NLP research

• NLP research often involves developing an
  algorithm to automatically process some text and
  output analysis, eg
• For each word, its Part of Speech (or semantic
  class, or)
• For each sentence, its grammatical structure
  (parse-tree)
• For each text, its classification Genre,
  sentiment, CoD, interesting wrt specific
  task/users
• Often this is done by Machine Learning given a
  training dataset of example words/sentences/texts,
  each marked (beforehand) with its Class learn
  a Classifier which can predict the Class of any
  new, unseen word/sentence/text.
• The algorithm is automatic, so where does
  Visualisation fit?

4
Visualisation of feature space?

• Machine Learning is automatic (eg using WEKA
  toolkit), the classification is not done by
  humans
• BUT ML relies on mapping each word/sentence/text
  into a set of FEATURES which characterise the
  data
• Visualisation may guide the researcher in
  exploring the dataset, to choose useful
  features?
• OR ML with different parameter-settings can
  produce different classification models
  Visualisation may help the researcher to compare
  the models?

5
Lexical semantic space
6
by Justin Washtell
7
Typical NLP dataset a CORPUS (plural Corpora or
Corpuses)

• Quran English translation interesting subset
  of verses
• Leeds Arabic NLP http//www.comp.leeds.ac.uk/arabi
  c/
• Arabic morphological analysis tools
• Quranic Arabic Corpus http//corpus.quran.com/
• Verbal Autopsy interviews narrative text
  yes/no, numbers
• SNOMED-CT Systematized Nomenclature of Medicine
  Clinical Terms adopted by UK NHS and US health
  authorities

8
Verbal Autopsy Dataset

• Verbal Autopsy interview of mother after death
  of her baby.
• Data collected as part of a main trial over 7
  year period
• 10,000 interview reports Data collected
  includes
• Signs and symptoms that led to the death
• History of any ailments
• Socio economic characteristic
• Care seeking and treatment
• Fertility and obstetric history
• Classification of Cause of Death by doctors at
  LSHTM - London School of Hygene and Tropical
  Medicine, Uni London
• Based on signs, symptoms and expert knowledge

9
Problems with VA data

• Both quantitative and qualitative
• Missing values (-)
• 215 variables (plus narrative text)
• Entries can have opaque codes
• sex 1, 2, 8 or 9
• Weight 1.45, 9.99 or 8.88
• Continuous revision of questionnaire created
  blank values for some variables
• Visualization of decision tree is problematic
  (size 1043, leaves601) also other classifier
  outputs, eg Naïve Bayes

10
Visualising Corpus Linguistics

• Paul Rayson presented overview of techniques at
  CL2009 International Conference on Corpus
  Linguistics
• Paul Rayson and John Mariani, 2009.
• Visualising Corpus Linguistics
• I like the Key Word Clouds from CL2001 CL2009
  !!!
• Wordle etc make pretty pictures, for PR etc
• BUT do word clouds actually help guide NLP
  research???

11
Open to discussion

• Over to you
• NLPers can ask about visualisation techniques
  they could use
• VVRers can ask about ideas on text analytics
  techniques which might turn text into numbers
• And/or any other ideas?
• THANK YOU for your participation