Wooclap time

• Sentiment =

  • Feelings, Attitudes, Emotions, Opinions

  • A thought, view, or attitude, especially one based mainly on emotion instead of reason

• Subjective impressions, not facts

• Use of natural language processing (NLP) and computational techniques to automate the extraction or classification of sentiment from unstructured text

• Other terms

  • Opinion mining
  • Sentiment mining
  • Sentiment classification

• Subjectivity (neutral vs sentimental text)
• Opinion retrieval (opinion for a given query)
• Comparative opinion analysis
• Emotion detection (e.g., happiness, anger, sadness)
• Stance detection (in favor or against)
• Reputation analysis
• Sarcasm/Irony detection
• Hate-speech

• Can be applied in every topic & domain (non exhaustive list):

  • Book: is this review positive or negative?
  • Humanities: sentiment analysis for German historic plays.
  • Products: what do people think about the new iPhone?
  • Blog: how are people thinking about immigrants?
  • Politics: who is going to win the election?
  • Social Media: what is the trend today?
  • Movie: is this review positive or negative (IMDB, Netflix)?
  • Marketing: how is consumer confidence? Consumer attitudes?
  • Healthcare: are patients happy with the hospital environment?

• Regular opinions: Sentiment/opinion expressions on some target entities

  • Direct opinions:

    • “The touch screen is really cool.”

  • Indirect opinions:

    • “After taking the drug, my pain has gone.”

• Comparative opinions: Comparison of more than one entity.

  • E.g., “iPhone is better than Blackberry.”

• An opinion is a quintuple
  
  ( entity, aspect, sentiment, holder, time)
  
  where

  • entity: target entity (or object).

  • aspect: aspect (or feature) of the entity.

  • sentiment: +, -, or neu, a rating, or an emotion.

  • holder: opinion holder.

  • time: time when the opinion was expressed.

• Simplest task:

  • Is the attitude of this text positive or negative?

• More complex:

  • Is the attitude of this text positive, negative or neutral?
  • Label the attitude of this text from 1 to 5

• Advanced:

  • Detect the target, source, or complex opinion types
  • Implicit opinions or aspects

• Classify a document (e.g., a review) based on the overall sentiment of the opinion holder
  • Classes: Positive, negative (possibly neutral)
    • Neutral means no sentiment expressed
    • “I believe he went home yesterday.”
    • “I bought a iPhone yesterday”
• An example review:
  • “I bought an iPhone a few days ago. It is such a nice phone, although a little large. The touch screen is cool. The voice quality is great too. I simply love it!”
  • Classification: positive or negative?
• It is basically a text classification problem

• Classify the sentiment expressed in a sentence
  • Classes: positive, negative (possibly neutral)
• But bear in mind
  • Explicit opinion: “I like this car.”
  • Fact-implied opinion: “I bought this car yesterday and it broke today.”
  • Mixed opinion: “Apple is doing well in this poor economy”

• Harder than topical classification, with which bag of words features perform well

• Must consider other features due to…

  • Subtlety of sentiment expression
    • irony ( What a great car, it stopped working in the second day.)
    • expression of sentiment using neutral words ( The concert didn’t meet my expectations.)
  • Domain/context dependence
    • words/phrases can mean different things in different contexts and domains ( long queue vs long battery life)
  • Effect of syntax on semantics

• Explicit aspects: Aspects explicitly mentioned as nouns or noun phrases in a sentence

  • “The picture quality is of this phone is great.”

• Implicit aspects: Aspects not explicitly mentioned in a sentence but are implied

  • “This car is so expensive.”
  • “This phone will not easily fit in a pocket.”
  • “Included 16MB is stingy.”

An implicit aspect should satisfy the following conditions:

• The related aspect word does not occur in the review sentence explicitly.

• The aspect can be discovered by its surrounding words (e.g. opinion words) in the review sentence.

• Lexicon-based methods
  • Dictionary based: Using sentiment words and phrases (e.g., good, wonderful, awesome, troublesome, cost an arm and leg)
  • Corpus-based: Using co-occurrence statistics or syntactic patterns embedded in text corpora
• Supervised learning methods: to classify reviews into positive and negative.
  • Naïve Bayes
  • Support Vector Machine
  • Deep learning
  • …
• Large Language Models
  • BERT
  • …

• Lists of words that are associated with sentiment scores
• Can have binary scores (1, -1) or intensity scores (from 0 to 1)
• Positive/negative polarity, emotions, affective states, negation lists
• Manually annotated or created from our corpus

• Detect sentiment in two independent dimensions:

• Positive: {1, 2,… 5}

• Negative: {-5, -4,… -1}

• Example: “He is brilliant but boring”

  • Sentiment(‘brilliant’) = +4
  • Sentiment(‘boring’) = -2
  • Overall sentiment = +2

• 2,300 words, >70 classes

• Affective Processes

  • negative emotion (bad, weird, hate, problem, tough)
  • positive emotion (love, nice, sweet)

• Cognitive Processes

  • Tentative (maybe, perhaps, guess), Inhibition (block, constraint)

• Pronouns, Negation (no, never), Quantifiers (few, many)

• VADER (Valence Aware Dictionary and sEntiment Reasoner) is a lexicon and rule-based sentiment analysis tool designed specifically for social media text. Contains a pre-built lexicon of words that are associated with sentiment scores ranging from -4 to +4

• Five generalizable heuristics based on grammatical and syntactical cues:

  • Punctuation: “The food here is good !!!” vs “The food here is good.”
  • Capitalization: “The food here is GREAT!” vs “The food here is great!”
  • Degree modifiers: “The service here is extremely good” vs “The service here is good”
  • The conjunction “but”: “The food here is great, but the service is horrible” has mixed sentiment
  • For negation examine the tri-gram preceding a sentiment lexical feature: “ The food here isn’t really all that great”

• WordNet: A lexical database of semantic relations between words including synonyms, antonyms, hyponyms

• The synonyms are grouped into synsets with short definitions and usage examples

• Create positive (“good”) and negative seed-words (“terrible”)

• Find synonyms and antonyms

  • Positive set: Add synonyms of positive words (“well”) and antonyms of negative words
  • Negative set: Add synonyms of negative words (“awful”) and antonyms of positive words (”evil”)

• Bing Liu’s Page on Opinion Mining

• 6,786 words

  • 2,006 positive
  • 4,783 negative

• Start with 30 adjectives that you know the semantic orientation (positive adjectives: great, fantastic, nice, cool and negative adjectives: bad, dull)

• Use WordNet to predict the orientations of all the adjectives in the opinion word list

• https://github.com/aesuli/SentiWordNet

• All WordNet synsets automatically annotated for degrees of positivity, negativity, and neutrality/objectiveness

• [estimable(J,3)] “may be computed or estimated”
  \[\operatorname{Pos\ \ 0\ \ \ Neg\ \ 0\ \ \ Obj\ \ 1} \]

• [estimable(J,1)] “deserving of respect or high regard” \[\operatorname{Pos\ \ .75\ \ \ Neg\ \ 0\ \ \ Obj\ \ .25} \]

1. Extract a phrasal lexicon from reviews
2. Learn polarity of each phrase
3. Rate a review by the average polarity of its phrases

The JJ tags indicate adjectives, the NN tags are nouns, the RB tags are adverbs, and the VB tags are verbs

• Positive phrases co-occur more with “excellent”

• Negative phrases co-occur more with “poor”

• But how to measure co-occurrence?

• PMI between two words:
  • How much more do two words co-occur than if they were independent?

• P(word) estimated by      hits(word)/N
• P(word1,word2) by     hits(word1 NEAR word2)/N^2

• Intuition
  • Start with a seed set of words (“good”, “poor”)
  • Find other words that have similar polarity:
    • Using “and” and “but”
    • Using words that occur nearby in the same document
    • Using WordNet synonyms and antonyms
  • Using rules based on punctuation, emoticons

• Advantages:
  • Can be domain-independent with general purpose lexicons
  • Can become domain-dependent
  • Can be easy to rationalise prediction output
  • Can be applied when no training data is available
• Disadvantages:
  • Compared to a well-trained, in-domain ML model they typically underperform
  • Sensitive to affective dictionary coverage

• Pre-processing and tokenization
• Feature representation
• Feature selection
• Classification
• Evaluation

• The Porter stemmer identifies word suffixes and strips them off.

• But:

  • objective (pos) and objection (neg) -> object

  • competence (pos) and compete (neg) -> compet

• The problem has been studied by numerous researchers.

• Key: feature engineering. A large set of features have been tried by researchers. E.g.,

  • Terms frequency and different IR weighting schemes
  • Part of speech (POS) tags
  • Opinion words and phrases
  • Negations
  • Stylistic
  • Syntactic dependency

Add NOT_ to every word between negation and following punctuation:

• “terrible” vs “wasn’t terrible”
  • The movie was terrible
  • The movie was bad but wasn’t that terrible as they said
• The degree of the intensity shift varies from term to term for both positive and negative terms

• A supervised statistical text classification approach based on surface, semantic, and sentiment features.
• For negation: estimate sentiment scores of individual terms in the presence of negation
• One lexicon for words in negated contexts and one for words in affirmative

• Features:
  • ngrams
  • character ngrams
  • all-caps: the number of tokens with all characters in upper case
  • POS
  • the number of negated contexts
  • sentiment lexicons
  • the number of hashtags, punctuation, emoticons, elongated words
• Classifier: linear-kernel SVM

• Advantages
  • Lead to better performance compared to lexicon based approaches
  • The output can be explained (most of the times)
• Disadvantages
  • They need training data (distant supervision comes with limitations)
  • They can’t capture the context
  • Based on feature engineering that is a tedious task
  • Not good performance in multiclass classification

• Continuous word representations model the syntactic context of words but ignore the sentiment of text

• Good vs bad: They will be represented as neighboring word vectors

• Solution: Learn sentiment specific word embedding, which encodes sentiment information in the continuous representation of words

• Three neural networks to effectively incorporate the supervision from sentiment polarity of text in their loss functions

• Distant-supervised tweets

• Start with a set of pre-trained word vectors and a sentiment lexicon
• Calculate the semantic similarity between each sentiment word and the other words in the lexicon based on the cosine distance of their pre-trained vectors
• Select top-k most similar words as the nearest neighbors and re-rank according to sentiment scores

• Refine the pre-trained vector of the target word to be:
  • closer to its sentimentally similar neighbors,
  • further away from its dissimilar neighbors, and
  • not too far away from the original vector.

• Sentiment analysis was one of the tasks in the BERT paper

https://huggingface.co/blog/sentiment-analysis-python

• Twitter-roberta-base-sentiment is a roBERTa model trained on ~58M tweets and fine-tuned for sentiment analysis (https://huggingface.co/cardiffnlp/twitter-roberta-base-sentiment)

• SST-2 BERT: Fine-tuned on the Stanford Sentiment Treebank (SST-2) which consists of sentences from movie reviews. The model is well-suited for general sentiment analysis tasks. (https://huggingface.co/distilbert-base-uncased-finetuned-sst-2-english)

• Bert-base-multilingual-uncased-sentiment is a model fine-tuned for sentiment analysis on product reviews in six languages: English, Dutch, German, French, Spanish and Italian (https://huggingface.co/nlptown/bert-base-multilingual-uncased-sentiment)

• Distilbert-base-uncased-emotion is a model fine-tuned for detecting emotions in texts, including sadness, joy, love, anger, fear and surprise (https://huggingface.co/bhadresh-savani/distilbert-base-uncased-emotion)

• Analyzed around 126,000 tweets
• Annotated with NRC lexicon

• Three different approaches for calculating the emotional signals of the claims:
  • emoLexi
  • emoInt
  • emoReact

• Are systems that detect sentiment biased?

• Hypothesis: a system should equally rate the intensity of the emotion expressed by two sentences that differ in the gender/race

What about biases in LLMs?

• DistilBERT base uncased finetuned SST-2:
  • “This movie was filmed in France” -> 0.89
  • “This movie was filmed in Afghanistan” -> 0.08

• “This movie was filmed in {country_name}”

From Aurélien Géron colab

• Sentiment analysis
• Lexicon-based methods
• Learning-based methods
• Sentiment aware word embeddings
• Other aspects regarding sentiment analysis

• Crawl your own data from Twitter:
  • https://developer.twitter.com/en/docs/twitter-api
• SemEval Datasets: 2012-now
  • https://semeval.github.io/
• Stanford Twitter Sentiment (STS):
  • http://help.sentiment140.com/ (Go et al. 2009)
• Sanders Corpus:
  • https://github.com/zfz/twitter_corpus
• IMDB movie reviews (50K)
  • https://ai.stanford.edu/~amaas/data/sentiment/
• Datasets from Bing Liu’s group:
  • https://www.cs.uic.edu/~liub/FBS/sentiment-analysis.html
• Amazon review data
  • https://nijianmo.github.io/amazon/index.html
• iSarcasm
  • https://github.com/dmbavkar/iSarcasm

• VADER (Hutto and Gilbert, 2014)
  • https://github.com/cjhutto/vaderSentiment
• LIWC
  • https://www.liwc.app/
• Bing Liu
  • https://www.cs.uic.edu/~liub/FBS/sentiment-analysis.html
• Multi-Perspective Question Answering - MPQA (Wiebe et al., 2005)
  • https://mpqa.cs.pitt.edu/lexicons/subj_lexicon/
• SentiWordNet (Esuli and Sebastiani, 2006)
  • https://github.com/aesuli/SentiWordNet
• NRC Lexicons
  • http://saifmohammad.com/WebPages/lexicons.html
• AFFINN (Nielsen, 2011)
  • https://github.com/fnielsen/afinn

• Sentiment analysis in huggingface
• Sentiment analysis with BERT

• Bagheri, A., Saraee, M, and De Jong, F. 2013. Care more about customers: Unsupervised domain-independent aspect detection for sentiment analysis of customer reviews. Knowledge-Based Systems. 52, 201–213
• Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805
• Esuli, A. and Sebastiani, F. 2006. SENTIWORDNET: A Publicly Available Lexical Resource for Opinion Mining. In Proceedings of the Fifth International Conference on Language Resources and Evaluation (LREC’06)
• Ghanem, B., Rosso, P., & Rangel, F. 2020. An emotional analysis of false information in social media and news articles. ACM Transactions on Internet Technology (TOIT), 20(2), 1-18
• Giachanou, A., Rosso, P., & Crestani, F. 2019. Leveraging emotional signals for credibility detection. In Proceedings of the 42nd international ACM SIGIR conference on research and development in information retrieval. pp. 877-880
• Hu, M., & Liu, B. 2004. Mining and summarizing customer reviews. In Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining. pp. 168-177

• Hutto, C. and Gilbert, E. 2014. Vader: A parsimonious rule-based model for sentiment analysis of social media text. In Proceedings of the international AAAI conference on web and social media, vol. 8, no. 1, pp. 216-225
• Kiritchenko, S., and Saif M. 2018. Examining Gender and Race Bias in Two Hundred Sentiment Analysis Systems. In Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics, pp. 43-53. 2018
• Kiritchenko, S., Zhu, X., & Mohammad, S. M. 2014. Sentiment analysis of short informal texts. Journal of Artificial Intelligence Research, 50, 723-762
• Narayanan, R., Liu, B., & Choudhary, A. 2009. Sentiment analysis of conditional sentences. In Proceedings of the 2009 conference on empirical methods in natural language processing. pp. 180-189
• Nielsen, F. A. 2011. A new ANEW: Evaluation of a word list for sentiment analysis in microblogs. Proceedings of the ESWC2011 Workshop on ‘Making Sense of Microposts’: Big things come in small packages 718 in CEUR Workshop Proceedings 93-98
• Tang, D., Wei, F., Yang, N., Zhou, M., Liu, T., & Qin, B. 2014. Learning sentiment-specific word embedding for twitter sentiment classification. In ACL. pp. 1555-1565

• Tausczik, Y. R., & Pennebaker, J. W. 2010. The psychological meaning of words: LIWC and computerized text analysis methods. Journal of language and social psychology, 29(1), 24-54
• Turney, P. 2002. Thumbs Up or Thumbs Down? Semantic Orientation Applied to Unsupervised Classification of Reviews. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics. pp. 417-424
• Vosoughi, S., Roy, D., & Aral, S. 2018. The spread of true and false news online. science, 359(6380), 1146-1151
• Wiebe, J., Wilson, T., & Cardie, C. 2005. Annotating expressions of opinions and emotions in language. Language resources and evaluation, 39(2), 165-210
• Yu, L. C., Wang, J., Lai, K. R., & Zhang, X. 2017. Refining word embeddings for sentiment analysis. In Proceedings of the 2017 conference on empirical methods in natural language processing. pp. 534-539