NLP and Sentiment Analysis with Watson

Thomas Allen, Qian Wang, Jonathan Tizon, Yan Ren

History

• Definition: Interaction between human (natural) language and computer (artificial) language
• Components
• Semantic
• Syntax
• Context
• Timeline
• Machine Translation (1950’s)
• Artificial Intelligence (ELIZA)
• http://www.manifestation.com/neurotoys/eliza.php3
• Statistical NLP

Some Uses of NLP

• Sentiment analysis
• Machine translation
• Text classification
• Text summarization
• Artificial intelligence

Semantics

Lexical Analysis or Tokenization

A person or natural language speaker

• Naturally separate a words by spaces or punctuation.
• Can easily identify linguistic units

A computer would simply see a string object.

• Will need to have linguistic units defined such as words, punctuation, alpa-numerals etc.
• Most have delimiters defined for segmentation.

White Space Tokenization

The words in this string can easily be tokenized by splitting it by whitespaces

• “The quick brown fox jumps over the lazy dog.”
• [“The”, “quick”, “brown”, “fox”, “jumps”, “over”, “the”, “lazy”, “dog.”]

​

Difficulties in Tokenization

• Dirty Text
• Grammatical/Syntactic errors in original document
• Special characters, Abbreviations, Contractions
• “He’s” could mean “he is” or “he has”
• CAT could mean “cat” or “ct scan”
• “60 second long videos” or “60-second long videos”

Syntactic Analysis

Categorize segments through context free grammars that define Natural Language

Basic grammar defining english

• [Sentence] -> [Noun Phrase] [Verb Phrase]
• [Noun phrase] -> [Article]|[Adjective] [Noun]
• [Verb phrase] -> [Verb][Noun Phrase]|[Prepositional Phrases]
• [Verb phrsae] -> [Verb][Noun Phrase][Prepositional Phrases]
• [Verb Phase] -> [Verb][Noun Phrase][Adverb]
• [Prepositional Phrase] -> [Preposition][Noun Phrase]

Parse Tree

• A dictionary of definitions isn’t enough to determine the definition of a word.
• Location in tree will help determine its part of speech
• Part of speech helps determine word’s meaning
• Buffalo from Buffalo that other buffalo from Buffalo bully [themselves] bully buffalo from Buffalo.

​

Pragmatics

Pragmatic Analysis

We always pragmatically say more than we semantically say

Eliminates ambiguity by explaining meaning not found in plain text/semantics

• Done through contextual information
• Applies real world knowledge

Sentiment Analysis

We want smarter algorithms!

• Sentiment Analysis accuracy is broken down into two categories
• Polarity: Was the text positive or negative (Bag of Words)
• Degree: How positive or negative the text is (Naive Bayes Classification)
• An intelligent approach to analyze human emotion
• Allows the extraction of key entities and semantics

The simplest example

“It’s rather like a lifetime special -- pleasant, sweet, and forgettable.”

​

• Sentiment analysis is not about understanding the sentence in any meaningful way
• Think of pattern analysis of characters without any external influences or intuition
• How should we leverage data to classify this sentence?
• Most obvious, find similar sentences in data set
• Simply the presence of good words, such as sweet and pleasant, can be a good indicator

​

Now let’s assume...

“It’s rather like a lifetime special -- pleasant, sweet, and forgettable.”

​

​

Good: 46

-Goodness = 46/(46+22) = 0.68

Bad: 22

-Badness = 22/(46+22) = 0.32

​

Data from Sentdex

Other polarized words ex:

“It’s rather like a lifetime special -- pleasant, sweet, and forgettable.”

​

​

Good: 46

-Goodness = 46/(46+22) = 0.68

Bad: 22

-Badness = 22/(46+22) = 0.32

​

Good: 15

-Goodness = 15/(6+15) = 0.71

Bad: 6

-Badness = 6/(6+15) = 0.29

​

Good: 10

-Goodness = 10/(14+10) = 0.41

Bad: 14

-Badness = 14/(14+10) = 0.59

​

Data from Sentdex

What about neutral words?

“It’s rather like a lifetime special -- pleasant, sweet, and forgettable.”

​

​

Good: 46

-Goodness = 46/(46+22) = 0.68

Bad: 22

-Badness = 22/(46+22) = 0.32

​

Good: 15

-Goodness = 15/(6+15) = 0.71

Bad: 6

-Badness = 6/(6+15) = 0.29

​

Good: 10

-Goodness = 10/(14+10) = 0.41

Bad: 14

-Badness = 14/(14+10) = 0.59

​

Good: 506

-Goodness = 506/(506+507) = 0.49

Bad: 507

-Badness = 507/(506+507) = 0.51

​

Data from Sentdex

Goodness:

• 5.22

Badness:

• 4.8

Bag of Words model

• Sentences are classified based on the number of positive and negative words contained.
• The test sentence would be classified as positive
• Note that this approach does not make use of the sentence’s order
• Polarity: Was the text positive or negative
• Thus coined the name “Bag of Words”
• Example:
• “It’s rather like a lifetime special -- pleasant, not sweet, and forgettable.”
• Sweet would still contribute .68 positivity
• The algorithm would still operate around 80% accuracy without actually understanding the sentence

Introduction to Bayes’ Algorithm

• An extension of the Bayes algorithm
• Assumes each data point is independent from others
• This allows for infinite properties to be taken into consideration
• Theorem -
• States that the Probability of event B given A is equal to the probability of the event A given B multiplied by the probability of B.
• Equation:

​

​

P(A|B) - Conditional probability of event A occurring given event B is true.

P(B|A) - Conditional probability of event B occurring given event A is true.

P(A) and P(B) - Probabilities of event A and event B occurring respectively

Naive Bayes Algorithm

• The algorithm calculates the the conditional probability of an object with i feature vectors
• Machine learning classification problems:
• Classes (C_I) - Occurrences that are attempting to be classified through machine learning
• Feature Vectors (x_i) - Base features, of syntax, that are independent from each other’s influence.
• Often paired with n-grams to increase accuracy

N-grams

• A contiguous sequence of n items from a given data sample.
• Independent assumptions are made on each word based on the last n-1 words in the series.
• Pro: Groups words that are unknown to the language interpreter to better understand the open nature of languages.
• Cons: Without smoothing imbalances can occur between infrequent grams (proper names), frequent grams (common words), and non-present grams (not seen in the data set)
• Top five 3-grams, based on the largest publicly-available, genre-balanced corpus of English.
• Much the same
• Much more likely
• Much better than
• Much more difficult
• Much of the

Data from Sentdex

NLP Library

Node.js

• Twitter-text
• Knwl.js
• Retext
• Natural

Python

• TextBlob
• SpaCy
• Gensim
• ScatterText
• AllenNLP
• PyNLPl

Java

• Stanford NLP
• OpenNLP
• ClearNLP

NLP Services

Watson NLP APPs

Concept

​

Category

• /News
• /international news
• /art and entertainment

Emotion

• Anger
• Disgust
• Sadness
• Fear
• Joy

Entity

• People
• Places
• Events

Keywords

Metadata

• Author
• Page title
• Publication date

Relations

Semantic Roles

• Subject
• Action
• Object

Sentiment

• Positive
• Negative

• Identify tones at the sentence and document level
• This insight can be used to refine and improve communications
• Emotion
• Sad, Disgust, Fear, Joy
• Social Propensities
• Openness, Conscientiousness, Agreeableness
• Language Styles
• Analytical, Confident, tentative

• Classification of text input
• Train your own classifier with specific training data

Set Up IBM Cloud

1. Sign in or Sign Up for https://www.ibm.com/watson/.
2. Menu → Developer → Cloud → My Cloud Console
3. Create Service
4. Add Credentials

pip install --upgrade watson-developer-cloud

Thank you

Questions?