[Paper Review] Practical and Ethical Considerations in the Effective use of Emotion and Sentiment Lexicons

Saif M. Mohammad|arXiv (Cornell University)|Nov 6, 2020

Sentiment Analysis and Opinion MiningComputer Science30 references37 citations

TL;DR

The paper outlines practical and ethical considerations for using word–emotion association lexicons, drawing on the NRC Emotion Lexicon and related resources. It discusses coverage, sense priors, biases, aggregation, and usage guidelines.

ABSTRACT

Lexicons of word-emotion associations are widely used in research and real-world applications. As part of my research, I have created several such lexicons (e.g., the NRC Emotion Lexicon). This paper outlines some practical and ethical considerations involved in the effective use of these lexical resources.

Motivation & Objective

• Explain what emotion lexicons are and how they are used across disciplines and applications.
• Identify practical limitations and ethical concerns in constructing and applying emotion lexicons.
• Provide guidance to mitigate bias and misinterpretation when analyzing text with lexicons.
• Promote responsible use by discussing provenance, crowd-sourcing, and transparency in annotations.

Proposed method

• Review existing emotion lexicons and their creation methods, including crowdsourced and automatically generated entries.
• Enumerate coverage, sense priors, connotations versus denotations, and temporal stability of associations.
• Highlight socio-cultural biases and issues in aggregation and translation of lexicons.
• Offer practical recommendations and “pro-tips” for applying lexicons in text analysis and research.

Experimental results

Research questions

• RQ1 What are the key practical limitations of current emotion and sentiment lexicons in terms of coverage and sense representation?
• RQ2 What ethical considerations arise from crowd-sourced annotation, socio-cultural biases, and translation of lexicons?
• RQ3 How should researchers interpret and apply lexicon scores to avoid misattributing emotions to speakers?
• RQ4 What best practices can mitigate bias and improve reliability when using these lexicons in research and applications.

Key findings

• Coverage varies across lexicons; high-coverage sets exist but no lexicon covers all language terms.
• Words may carry dominant senses that differ across domains, affecting associations.
• Associations reflect connotations, not denotations, and are not immutable over time.
• Socio-cultural biases arise from annotator pools and text sources, impacting lexicon perceptions.
• Majority-vote aggregation can obscure minority or context-specific associations; disaggregated data is available.
• Translation and automatic generation introduce errors and cultural differences in emotion mappings.