training common sense - day 3

adjectives

../training-common-sense/sources/pictures/good-bad.png

../training-common-sense/sources/pictures/sentiWordNet.png

adjectives in WordNet

• 
  
• * AI
  • - George A. Miller + Philip N. Johnson-Laird
  • - componential lexicon
    • --> looking for significant atoms
  • - ±1976 - 1985
• 
  
• * Cognitive Studies, Psychology
  • - relational lexicon
    • --> IS-A-KIND-OF relations
  • - ±1985 and onwards
  • WordNet developed at Princeton University, at Cognitive Studies

• - Philip N. Johnson-Laird
  • 266 pairs of antonymous adjectives ------------> 25 classes of nouns
• - Christiane Fellbaum
  • 266 pairs of antonymous adjectives ------------> 25 classes of verbs
• - Kitty Miller
  • descriptive adjectives 

Philip N. Johnson-Laird

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3250179/bin/pnas.1117174108unfig01.gif

• --> searched for it, but cannot find it, also no references to it in articles.

adjectives in Pattern

• 1. Introduction
  • - "Textual information can be broadly categorized into two types: objective facts and subjective opinions (Liu, 2010)." --> first sentence of the article
  • - two approaches for Sentiment Analysis: 
    • - using subjectivity lexicons (Taboada et al., 2011)
    • - using machine learning text classification (Pang et al., 2002)

commodification of difference

to compute the semantic similarity/relateness/mathematical-relation between adjectives

• - 1.100 adjectives manually annotated on polarity and subjectivity
• - then, they want to extend that set to 5.000 adjectives

• 3.1. Using Distributional Extraction
• There   is   a   wellknown   approach   in   computational linguistics  in  which  semantic  relatedness  between  words is extracted from distributional information (see e.g. Van de  Cruys  2010  for  an  example  for  Dutch).  The  method uses a vector space model with adjectives as vectors (i.e., matrix  rows)  and  nouns  as  vector  features  (i.e.,  matrix columns).  The  value  for  each  vector  feature  represents the  frequency  an  adjective  precedes  a  noun.  It  is  then possible  to  take  the  cosine  of  the  angle  between  two vectors  as  a  measure  of  similarity  (cosine  similarity).  In other  words,  adjectives  followed  by  the  same  nouns  are more  semantically  related  than  adjectives  followed  by different adjectives nouns. The method then uses dimensionality reduction  and  clustering  by  cosine  distance  to  create groups of semantically related words.

• - Looking at the relation between adjectives and nouns, is a way to have a sense of the context that a word appears in.
• - What nouns could adjectives modify?
• - you can calculate the relation between a certain adjective to a certain noun (a1-n1)

• --> these nouns fall under the same class of nouns

../training-common-sense/sources/pictures/similarity-of-adjectives-related-to-nouns_from_mining-for-meaning-2.png

• a1 [n1]
• a1 [n2]
• a1 [n3]

• For each gold1000 --> fantastic
• appears in the gold1000 with 10 different nouns --> a1[n1, n2, n3, n4, n5, n6, n7, n8, n9, n10] --> is the adjective vector
• 
  
• great, beautiful, excellent, fine, nice, good
• --> are already hand-annotated / or not
• --> have adjective vectors as well
• 
  
• --> are these two adjective vectors then compared?
• 
  

en-sentiment.xml

fr-sentiment.xml

how about going forwards?

• - what is better?
  • - less suggestive?
  • - more efficient?
  • - ... 
• - does it become better by 
  • - giving critique on it?
  • - play with it? being goofy?

• - as a form of annotation of Pattern
  • - there is already annotation present from the Pattern team, but this is focused on an urge to show how the data speaks
• - additional files that go into the package, reacting on the files of Pattern
• - which might include a check-me.txt file 
• - in some packages there is an ADD-TO-DO or ADD-FIX-ME file

Google Group question: sentiment_score()

../training-common-sense/sources/pictures/if-happy-is-0.0-something-is-wrong.png

Pattern on the web

backing-up: Cqrrelations