Back to all articles
Blog post
April 27, 2026

Thai-Language NLP and sentiment analysis: a buyer’s accuracy guide for social listening

Thai is one of the most challenging languages for automated sentiment analysis. It is tonal, scriptless in word boundaries, and rich in particles and honorifics that carry sentiment signals invisible to most NLP (natural language processing) models. With 56 million LINE users, over 44 million TikTok users aged 18+, and tens of millions of active users across Facebook and other platforms, Thailand generates massive volumes of social media content that brands and agencies need to analyse accurately. Yet most global social listening tools achieve materially lower accuracy on Thai content than they do for English — a gap that directly affects the quality of business decisions.

Why Thai defeats standard NLP models

Thai script does not use spaces between words. Unlike English, where word boundaries are visually obvious, Thai text flows continuously, requiring word segmentation as a preprocessing step before any sentiment analysis can begin. Standard NLP libraries trained primarily on space-delimited languages struggle with this fundamental difference.

Thai is a tonal language with five tones. The same syllable spoken with different tones carries different meanings. In written social media, tone markers and context determine meaning, but automated tools often miss these distinctions. Particles like “ค่ะ,” “ครับ,” “นะ,” “จ้ะ,” and “หรอ” modify sentiment and politeness in ways that translation strips away entirely.

Thai social media users also employ extensive romanisation — writing Thai words using the Latin alphabet. “555” (representing Thai laughter, since 5 is pronounced “ha”) is ubiquitous but absent from most NLP training datasets. “Sanook” means fun. “Sabai” means comfortable or well. These romanised expressions carry clear sentiment but are invisible to tools trained only on Thai script.

Academic research underscores the challenge. Studies on Thai sentiment classification show that basic machine learning classifiers typically achieve around 70% accuracy on Thai social media text, while domain-specific fine-tuned models like WangChanBERTa can reach 84–92% accuracy in controlled settings such as hotel reviews or financial news. However, these results are for curated, domain-specific datasets — not the messy, code-switched, romanised content that dominates real-world Thai social media. The practical accuracy of global social listening tools on informal Thai content is likely to sit well below what they report for English, though precise figures depend on the platform, the content mix, and the evaluation methodology.

The consequence is that a sentiment score based on poorly segmented, tone-deaf, particle-ignoring NLP is not just imprecise — it is systematically biased toward misclassification.

The LINE problem compounds NLP challenges

LINE dominates Thailand’s digital communication with 56 million monthly active users — 78.2 percent of the population, according to DataReportal’s Digital 2026 Thailand report. LINE Official Accounts are widely reported to achieve exceptionally high open rates, making them one of the most effective digital communication channels in the country. Yet most social listening platforms cannot monitor LINE at all.

This means Thai social listening is doubly limited: the NLP accuracy on analysable content is below par, and the most important platform in the market is invisible to monitoring tools. The combined effect is that brands relying on global social listening for Thailand are making decisions based on a partial and potentially inaccurate view of public sentiment. [CROSSLINK: Government Social Listening in Thailand: LAO Implementation and Public Sector Results]

How to evaluate Thai NLP accuracy

When evaluating social listening vendors for Thailand, demand a live accuracy test on real Thai content.

Provide 50–100 Thai social media posts including formal Thai, informal Thai with particles, romanised Thai, code-switched Thai-English content, and posts using “555” and other common expressions. Compare the vendor’s sentiment classifications against native Thai speakers’ assessments. This kind of side-by-side evaluation is the only reliable way to judge how a platform handles the specific linguistic features that make Thai difficult.

Ask vendors to disclose their methodology: Are they using off-the-shelf translation followed by English-language NLP? Fine-tuned Thai-language models? Human-in-the-loop verification? The approach matters as much as the headline accuracy number.

Isentia’s Thai-language capabilities

Isentia combines localised Thai NLP with Bangkok-based analyst teams who verify sentiment for cultural context, sarcasm, and informal language. The analysts understand particles, romanisation, regional dialect variations, and the cultural references that define Thai online discourse.

Isentia’s sister company Pulsar provides the data infrastructure, while human verification ensures that the intelligence derived from Thai content is actually accurate. For brands where Thai consumer sentiment directly affects commercial decisions — in a market where approximately 67 percent of internet users make online purchases on a weekly basis — this accuracy is not a nice-to-have. It is a business requirement.

Thailand’s PDPA enforcement is accelerating

Thailand’s PDPA has been fully enforced since June 2022. The PDPC has moved from awareness-building to active enforcement, and the trajectory is clear.

In 2024, the PDPC issued its first major fine — THB 7 million against a major IT product retailer for three charges: failure to appoint a data protection officer, inadequate security measures, and failure to report a data breach to the PDPC. The breach had exposed customer data to criminal call centre gangs. Then, on 1 August 2025, the PDPC announced a further eight administrative fines across five cases involving both public and private entities, bringing cumulative penalties to approximately THB 21.5 million.

Separately, in November 2025, the PDPC ordered World (the digital identity project formerly known as Worldcoin) to halt its iris-scanning operations in Thailand and delete biometric data collected from approximately 1.2 million users. The regulator ruled that collecting sensitive biometric data in exchange for cryptocurrency did not constitute valid consent under the PDPA. The case demonstrates the PDPC’s willingness to act against large-scale data processing operations.

The PDPC’s enforcement infrastructure is also becoming more technology-enabled. The PDPC Eagle Eye, a division within the Office of the PDPC, has launched the PDPC Eagle Eye Crawler — an automated tool that enables continuous monitoring of data breach incidents. This signals a proactive surveillance approach that goes beyond responding to complaints.

For social listening buyers, these developments matter directly. The PDPA does not appear to contain an explicit exemption for publicly available data — the law’s exemptions cover personal/household use, state security, media activities, and parliamentary operations, but do not specifically address publicly available social media content. Many legal practitioners therefore point to legitimate interest as a potentially viable lawful basis for social listening, which requires a documented assessment that the organisational benefit outweighs potential adverse effects on data subjects. However, the PDPC has not issued specific guidance on social listening, so this interpretation should be validated with qualified Thai legal counsel.

Frequently asked questions

Why is Thai particularly challenging for NLP?

Thai lacks word boundary markers (no spaces), is tonal (five tones), uses sentiment-bearing particles, and features extensive romanisation on social media. Academic research consistently identifies Thai as a low-resource language for NLP, with accuracy on informal social media content varying widely depending on the model and approach used.

How should buyers evaluate Thai sentiment analysis accuracy?

Demand a live test: provide 50–100 real Thai social media posts spanning formal, informal, romanised, and code-switched content, and compare the vendor’s classifications against native speakers’ assessments. Ask about methodology — specifically whether the vendor uses Thai-specific NLP models and whether human verification is part of the workflow.

Does Thailand’s PDPA exempt publicly available data?

The PDPA does not contain an explicit exemption for publicly available data. Social listening operations should consider legitimate interest as a potential lawful basis, which requires a documented assessment. We recommend consulting qualified Thai legal counsel, as the PDPC has not yet issued specific guidance on this question.

*Disclaimer: This blog is for informational purposes only and does not constitute legal advice. Thailand’s PDPA regulatory environment continues to evolve, and organisations should consult qualified Thai legal counsel for guidance specific to their circumstances.

Learn more

If you’re interested in how Isentia can support your brand and strategy, simply fill out the form below and one of our specialists will contact you!

Share

Similar articles

object(WP_Post)#7687 (24) { ["ID"]=> int(49040) ["post_author"]=> string(2) "75" ["post_date"]=> string(19) "2026-08-11 02:14:11" ["post_date_gmt"]=> string(19) "2026-08-11 02:14:11" ["post_content"]=> string(14309) "

The advent of LLMs and AI search means that there has been a colossal shift in how audiences are consuming information today,  and a reciprocal shift in how all types of organisations, from government agencies to brands are responding . But while discussion amongst PR and comms pros tends to fall disproportionately on how brands are impacted, the needs of the former are just as keenly felt, and often quite distinctive.

So how can government agencies respond, especially in a time of global flux, when major policy changes need to be communicated and important stakeholders need to be managed? After all, government organisations do care about reputation, much as brands do, but have quite distinctive goals when it comes to ensuring accurate information reaches the right audiences.

There is a new entry point for stakeholders


The era of "Let me Google that" is rapidly fading. Instead of clicking through to official websites, people are asking chatbots for direct answers.  What’s striking is that government agencies often have no visibility on how they’re being talked about in the LLM space, even as it becomes a central channel for messaging and reputation. 

When AI models become the primary gatekeeper, audiences bypass official portals entirely — driving down site traffic and leaving agencies vulnerable to misinformation, negative sentiment, or worse, being left out of the conversation altogether. 

Therefore, the entry point is different. For commercial brands, this shift is profound but in some ways mediated – an FMCG brand, for instance, often discovered through third-party platforms in any case. But for government entities, the stakes are entirely different. As the sole, authoritative source for public information, they need citizens on their websites to get accurate details. Government agencies, especially if they are the authority or regulator in a particular industry or sector, need to make sure audiences know where to go to get the right information.

Different types of government agencies have different considerations

Not all government agencies are alike, and they all have different parameters that are quite non-negotiable for them, just by the way they function. 

  1. Service delivery agencies - these rely heavily on content freshness. They can’t risk outdated sources impacting eligibility or process changes not reaching audiences.
  2. Regulators - these need to transmit authority and trust. Regulators have to make it a priority that they’re amongst the first place audiences go to for information and that the industry they’re regulating does not put them in the shade on the channels stakeholders are actually using
  3. Policy departments - did the AI's account of a policy match what was actually announced? They want to be able to make sure the accuracy of a policy (and ideally, its effectiveness) is translated when audiences search through LLMs.
  4. Local and state authorities want to make sure the services they carry out on behalf of locals are visible too. Much like service providers, there is a question around access and awareness of programmes and regulations, but also an added consideration: not appearing on LLMs discontent amongst those wishing to see a return on taxation and electoral mandates, and give credence to bad actors.

What do government agencies want to get out of this new LLM-mediated landscape? 

Reputation is important, but that exists downstream from maintaining a flow of accurate information. It’s useful for communications teams in government organisations to self reflect and ask themselves the following questions:

  • Where are citizens going to find information about your services if not your website — and do you know what they're being told?
  • If there was a significant policy change or incident in the last twelve months, do you know how it's currently being characterised when someone asks an Al tool about your agency?
  • When you communicate a major service change or policy update, do you have any way of measuring whether it’s surfacing in searches about you?
  • How do you currently understand the gap between what your agency publishes and what citizens actually receive when they search for information?
  • Are there community groups, advocacy organisations, or media outlets shaping perception of your agency - and do you know if that's feeding into what Al models say?

These are gaps they already realise, but they don’t actually know what to do about it – how to manage or measure them. They need a tool that allows them to know this critical piece of information and make informed decisions. 

Ngaire Crawford, Pulsar Group’s Executive Director for AI Strategy says, “It’s easy to disregard LLM reputation as part of a difficult AI landscape or something that is only really relevant to more product-based communication or marketing. Government communications is about social licence, and ensuring the public have access to up to date and accurate information about things that matter to them, the role that Generative AI plays in how a community understand an issue will only continue to grow, and knowing the impact that you can have in that through small shifts in channel strategies or more consistent messaging is a crucial part of the communications toolkit.”  

Lumina AI View: AI visibility for PR & Comms


Lumina AI view is built for communicators who want to understand how their organisation and their competitors are being talked about by various AI models – including ChatGPT, Gemini, and Claude. AI View users get an insight into which sources are being cited, and how they would need to respond as a way of protecting their reputation or making sure correct information about them is being disseminated. 

The tool provides an AI view score — a composite metric ranging from zero to 100, designed to help track brand performance over time and facilitate comparisons against competitors. It is calculated using five weighted factors — sentiment, visibility, authority, dominance and freshness. Beyond the overall score, the platform provides a summary of a brand's AI narrative based on four distinct reputation pillars — direction, performance, integrity and innovation. 

These pillars help users identify exactly which dimension of a brand's reputation is under pressure, offering specific, actionable insights for board presentations or reviews. Ultimately, while the AI view platform provides the necessary intelligence, the strategic decisions regarding how to respond to these insights remain with the organisation.

Spotlight: An Australian Council


This progressive local government council is located in Australia’s leading center for culture and sports.

The council earned an AI view score of 74 reflected by strong reach and authority. Publications like CBD News and its own website are the most cited by LLMs — interestingly, most of them being cited by Claude.

Content freshness scored lower at 48. Their website still carries error pages and annual reports from a few years ago. If a report — one that is seen as an organisation’s most comprehensive and authoritative content, is still being cited even if it’s older, might potentially be in the way of the organisation’s own perception. Which means more work is needed to prevent outdated content from still appearing.

What type of content is showing up?

Most citations for the council originate from government sources, followed by news outlets, reports, and blogs. The domain is evenly split between owned content and content earned from external sources media articles and independent authorities. 

While most are recent, some older articles from major outlets such as The BBC remain visible and may significantly influence how LLMs perceive the council. Owned content typically addresses last year’s budget plans and the council’s latest vision for the city, which LLMs are referencing. Government sources are the major content type, however, external sources have a greater impact on the council’s overall LLM score.

​

The stakes are higher for government agencies

When brands track LLM visibility, they often ask, "Are we shown in a positive light?" or "Are we cited accurately?" For the government, additional questions arise: "Is this information accurate enough for someone to act on?" and "Are we still viewed as more authoritative than what we oversee?" Mistakes can have serious consequences, such as individuals applying for ineligible programs or missing critical deadlines for new initiatives or elections. This can quickly lead to public frustration. It is essential for government communicators to recognize these risks.

If you would like to know more about our Lumina suite, please reach out here and our team will get in touch with for you a quick demo.

" ["post_title"]=> string(77) "Why is tracking government visibility on LLMs different from tracking brands?" ["post_excerpt"]=> string(121) "Government agencies often can't see how AI chatbots describe them. Here's why LLM visibility matters and how to track it." ["post_status"]=> string(7) "publish" ["comment_status"]=> string(4) "open" ["ping_status"]=> string(4) "open" ["post_password"]=> string(0) "" ["post_name"]=> string(46) "latest-reads-ai-visibility-government-agencies" ["to_ping"]=> string(0) "" ["pinged"]=> string(0) "" ["post_modified"]=> string(19) "2026-08-11 02:14:19" ["post_modified_gmt"]=> string(19) "2026-08-11 02:14:19" ["post_content_filtered"]=> string(0) "" ["post_parent"]=> int(0) ["guid"]=> string(32) "https://www.isentia.com/?p=49040" ["menu_order"]=> int(0) ["post_type"]=> string(4) "post" ["post_mime_type"]=> string(0) "" ["comment_count"]=> string(1) "0" ["filter"]=> string(3) "raw" }
Blog
Why is tracking government visibility on LLMs different from tracking brands?

Government agencies often can’t see how AI chatbots describe them. Here’s why LLM visibility matters and how to track it.

object(WP_Post)#9089 (24) { ["ID"]=> int(47963) ["post_author"]=> string(2) "75" ["post_date"]=> string(19) "2026-06-03 02:01:58" ["post_date_gmt"]=> string(19) "2026-06-03 02:01:58" ["post_content"]=> string(5651) "

There is a new frontier where public perception is shaped: Large Language Models. Right now, LLMs are answering critical questions about your organisation. What are they saying? And more importantly, which sources are shaping those answers?

To navigate this landscape, public relations professionals don't need generic tools, but rather technology that speaks their language, and addresses the realities of a changed media and informational landscape.

That is why we're unveiling Lumina AI View, the latest addition to our intelligent suite of AI tools from Isentia. Trained specifically on the workflows and challenges of modern PR & communications, Lumina AI View helps you understand exactly what AI knows about you, and how it learned it.

A new standard for AI visibility

AI View tracks your citation strength and source quality alongside those of your competitors, giving you a clear view of where you hold authority and where you have gaps.

Lumina AI View maps your AI reputation from the ground up, allowing you to:

  • See which sources matter: When tools such as ChatGPT or Gemini discuss your organisation, which outlets do they cite? Track your source footprint over time and view the impact of key target media on how you’re discussed. We measure your citation strength and source quality alongside those of competitors, giving you a clear view of where you have authority and where you have gaps.
  • Gain industry-specific insight: Your competitors get cited from Financial Times and Bloomberg. You get cited on Reddit. Each brings opportunity – and risk. Discover how you measure up against industry standards, and target the sources that actually influence how AI represents you.
  • Catch narrative shifts early: AI responses change when new sources appear, sentiment shifts, or old controversies resurface. Get alerts when citation patterns change suddenly, before they impact the way you’re perceived by stakeholders.

Measure your progress: From media monitoring to full media intelligence

Lumina AI View is built on the principle that insights get stronger with repeated measurement. To help you maintain a clear view of your reputation, our proprietary scoring system provides regular updates that show you:

  • Evolving trends in how sources cite your organisation
  • Competitive standing and benchmark metrics
  • Where models differ in information presented, and sources cited 

Whether you run it weekly, on-demand, or whenever you need a check-in, patterns will emerge, trends will become clear, and you will build a baseline that makes any sudden narrative changes both comprehensible and the prerequisite to action.

Lumina AI View is part of Lumina AI, a comprehensive suite of AI tools built specifically for communicators. Our Lumina suite evolves traditional media monitoring into narrative intelligence, enabling you to truly understand how perceptions form, evolve, and impact your reputation.

Get in touch to register your interest and see what Lumina AI View can do for you.

" ["post_title"]=> string(66) "Introducing Lumina AI View: AI Visibility Built for PR & Comms" ["post_excerpt"]=> string(158) "Lumina AI View, the latest in Isentia's AI suite, is trained on PR & comms workflows to help you understand what AI knows about you — and how it learned it." ["post_status"]=> string(7) "publish" ["comment_status"]=> string(4) "open" ["ping_status"]=> string(4) "open" ["post_password"]=> string(0) "" ["post_name"]=> string(59) "introducing-lumina-ai-view-ai-visibility-built-for-pr-comms" ["to_ping"]=> string(0) "" ["pinged"]=> string(0) "" ["post_modified"]=> string(19) "2026-07-15 03:12:57" ["post_modified_gmt"]=> string(19) "2026-07-15 03:12:57" ["post_content_filtered"]=> string(0) "" ["post_parent"]=> int(0) ["guid"]=> string(32) "https://www.isentia.com/?p=47963" ["menu_order"]=> int(0) ["post_type"]=> string(4) "post" ["post_mime_type"]=> string(0) "" ["comment_count"]=> string(1) "0" ["filter"]=> string(3) "raw" }
Blog
Introducing Lumina AI View: AI Visibility Built for PR & Comms

Lumina AI View, the latest in Isentia’s AI suite, is trained on PR & comms workflows to help you understand what AI knows about you — and how it learned it.

Ready to get started?

Get in touch or request a demo.

Contact us Request a demo