Sentiment analysis on Twitter: an Indian election case study
India’s elections generate an enormous stream of public commentary across X, formerly known as Twitter. During the 2024 Lok Sabha election, users discussed parties, candidates, welfare schemes, unemployment, inflation, regional identity, religion and campaign speeches in several languages. Analysing this conversation can reveal how political narratives develop, although social media opinion should never be treated as a direct substitute for voting intention.
For analytics professionals, the value lies in combining natural language processing with election knowledge, careful sampling and transparent interpretation. An Indian election provides a useful case study because its electorate is linguistically diverse, its political debate is highly regional and its online population is unevenly distributed. These conditions produce lessons for analysts working with public opinion data in Australia, where political conversations also vary between cities, communities and platforms.
Defining the research question
A useful study begins with a narrow question. Instead of asking whether people “liked” a party, researchers might examine how sentiment towards major national parties changed during the campaign, which issues produced the strongest negative reactions, or whether online discussion differed between national and state-level audiences. The research question determines the keywords, time period, geographic filters and modelling approach.
For a case study of the 2024 Indian general election, the observation window could run from the announcement of the election schedule through the results. The period should be divided into meaningful phases: campaign launch, manifesto releases, polling rounds, major speeches, exit polls and result day. Comparing these periods helps distinguish persistent attitudes from short-lived reactions to breaking news.
Sentiment can be measured at several levels. A post may express positive emotion about a leader while criticising a policy, or praise a welfare programme while attacking the party that introduced it. A robust project should therefore track overall polarity, specific emotions such as anger or hope, and issue-level sentiment. Party mentions, hashtags and candidate references should be analysed separately so that a general mood is not incorrectly assigned to one political group.
Building a reliable X dataset
Data collection is one of the most difficult parts of social media research. X content is shaped by platform rules, API access, deleted posts, private accounts, rate limits and recommendation algorithms. A keyword search may capture highly active users rather than a representative sample of voters. Researchers should record the query design, collection dates, language filters, duplicate-handling rules and any changes in access conditions.
Search terms need to reflect India’s linguistic and political diversity. Queries can include party names, alliance names, candidate names, election slogans, constituency references and common hashtags in English, Hindi, Bengali, Tamil, Telugu, Marathi and other languages. Transliteration creates another complication: a Hindi phrase may appear in Devanagari, Roman script or a mixture of both. The same political figure may also be described using nicknames, initials or abbreviations.
A basic cleaning pipeline removes duplicate posts, obvious advertisements, repeated campaign material and automated accounts where possible. Bot detection should combine signals such as posting frequency, account age, repetition, coordination and unusual timing. High-volume accounts should not automatically be labelled bots; journalists, campaign workers and political commentators can post frequently. The study should report whether retweets and quote posts were included because they can heavily influence sentiment counts.
Understanding language and political context
A simple English sentiment dictionary performs poorly on multilingual Indian political content. Negation, sarcasm, code-switching and local idioms can reverse the apparent meaning of a sentence. A post containing a laughing emoji or an informal phrase may be supportive, mocking or hostile depending on context. Models trained on product reviews often misread political language because words such as “disaster”, “fight” and “corruption” carry a different meaning in campaign debate.
A stronger workflow combines multilingual transformer models with human annotation. A sample of posts should be labelled by trained reviewers for positive, negative, neutral, mixed or unrelated sentiment. Reviewers should also identify the target of the sentiment and, where possible, the relevant issue. Inter-annotator agreement provides a useful quality check: if reviewers frequently disagree, the category definitions need refinement.
The following approaches offer different balances between speed, explainability and contextual accuracy:
| Approach |
Strength |
Limitation |
Suitable use |
| Lexicon-based scoring |
Fast, transparent and inexpensive |
Misses sarcasm, code-switching and political context |
Early exploration and baseline measurement |
| Traditional machine learning |
Works well with labelled local data and interpretable features |
Requires feature engineering and may struggle with new phrases |
Stable monitoring of a defined campaign |
| Multilingual transformer |
Captures context across several languages and scripts |
More expensive, harder to explain and sensitive to training data |
High-quality sentiment and emotion classification |
| Human-assisted coding |
Handles irony, mixed views and emerging political language |
Slower and costly at large scale |
Validation, error analysis and sensitive topics |
The model should be evaluated separately for each major language and sentiment category. Overall accuracy can conceal poor performance on minority languages or neutral posts. Precision, recall and F1 scores are more informative, particularly when negative comments are far more common than positive ones. Manual review of false positives and false negatives can expose politically important errors that a single score would hide.
Reading sentiment trends without overstating them
Imagine that the analysis finds a rise in negative posts about unemployment and household costs during the campaign, while posts about national security generate more positive reactions for a particular alliance. This would indicate that these issues shaped online discourse, not that every voter held the same opinion. Social media users are self-selecting, and political organisations may deliberately amplify selected narratives.
A useful dashboard could show daily sentiment by party, issue, language and region. Volume should be displayed beside sentiment because a small number of highly active accounts can create an apparently large shift. Analysts should also separate original posts from retweets, distinguish verified and unverified accounts where ethically appropriate, and mark major events that may explain sudden changes.
Several findings may appear plausible but require caution. A surge in criticism could reflect a coordinated campaign, a journalist’s post going viral, or a platform controversy rather than a broad change in public belief. Similarly, positive sentiment may come from party workers or influencer networks. Sentiment analysis is strongest as a tool for identifying narratives and questions for further investigation, rather than as a standalone election forecast.
Validation, ethics and responsible reporting
Validation should compare online findings with other evidence, such as opinion polls, news coverage, search trends, constituency-level results and qualitative interviews. These sources measure different things, so agreement is not expected in every case. If online sentiment diverges sharply from polling or voting outcomes, the difference may reveal platform demographics, geographic concentration or organised communication.
Privacy is essential even when posts are publicly visible. Researchers should minimise personal data, avoid publishing usernames unnecessarily and report aggregated findings. Sensitive attributes such as religion, caste, ethnicity and political affiliation require special care. A model should not infer a person’s identity or voting intention from a single post. Data retention, access controls and secure storage should be documented from the beginning.
Australian analysts should also consider the Privacy Act 1988 and the Australian Privacy Principles when adapting this research locally. Political opinions can be sensitive personal information, and combining public posts with location, demographic or customer records may create additional obligations. Australian election communication also requires attention to authorisation rules and misleading political content, even though there is no single federal law that guarantees the factual accuracy of every political claim.
Applying the lessons in Australia
The Indian case is relevant to Australian organisations monitoring public conversation around federal, state or local elections. A project could compare discussion in Sydney, Melbourne, Brisbane, Perth and regional communities, while recognising that each city has distinct media habits and migrant networks. Multilingual analysis matters in suburbs where Hindi, Punjabi, Mandarin, Arabic, Vietnamese or Greek are commonly used alongside English.
Everyday platform behaviour also affects the data. Australians may encounter political content while commuting by train in Melbourne, checking updates during a lunch break in Sydney or following local issues through community groups and podcasts in regional Queensland. X is only one part of this information environment. Reddit, Facebook, TikTok, YouTube comments, online news and talkback discussions may reach different audiences and should be treated as complementary sources rather than interchangeable datasets.
The local market adds another consideration. Australian banks, retailers, universities, media companies and government agencies increasingly use text analytics to understand customer experience and public response. A sentiment model built for Indian election discourse should not be transferred directly to Australian political or commercial language. It needs local annotation, Australian spelling and slang, awareness of Indigenous data governance, and testing across urban and regional contexts.
The most practical outcome is a repeatable analytics framework: define a focused question, collect transparently, preserve multilingual context, validate with humans and communicate uncertainty. Used carefully, sentiment analysis can show which issues dominate public discussion, how narratives travel and where communities express concern. It becomes far more valuable when combined with polling, demographic knowledge, electoral data and responsible governance.
