Government Sentiment Analysis: A Complete Guide to Public Opinion Monitoring and Policy Evaluation

In a democratic society, evaluating government performance cannot rely on traditional surveys alone — it requires a real-time grasp of online public opinion. From central ministries to local governments, demand for sentiment analysis systems in Taiwan's public sector keeps growing. This guide explains how agencies use sentiment analysis for policy opinion tracking, satisfaction monitoring and crisis warning, and introduces the security compliance requirements of government procurement (joint supply contracts), as a practical reference for public-sector buyers and policy researchers.

Core Use Cases of Government Sentiment Analysis

Government agencies approach sentiment analysis differently from businesses. A company's core goals are brand management and commercial gain; a government's core goals are understanding public opinion, improving transparency, and responding promptly to social issues. Under this premise, public-sector sentiment analysis shows several typical scenarios:

Policy communication evaluation: when rolling out new policies (social welfare, environmental regulations, transportation projects), governments push heavy media campaigns. Sentiment analysis evaluates whether the message effectively reaches target groups, whether the public understands the policy correctly, and where the main points of contention concentrate — helping authorities adjust communication strategy in time.

Crisis early warning: administrative decisions sometimes trigger unexpectedly strong backlash. With keyword alert rules, a sentiment system can warn while negative opinion is still nascent, letting departments prepare response strategies or proactively clarify before opinion erupts into a full crisis. This early-warning capability is extremely valuable to government PR units.

Long-term satisfaction tracking: traditional satisfaction surveys run quarterly or yearly and, limited by sampling methods, struggle to capture specific groups. As a complement to polling, sentiment analysis provides more timely and diverse opinion samples — especially from the younger generation (20–35), a group severely under-represented in traditional phone surveys.

Methodology for Opinion Monitoring and Policy Evaluation

Effective government sentiment analysis needs a rigorous methodology so results are credible enough to inform policy adjustment. Key methodological elements for building public-sector sentiment capability:

Keyword system design: government keyword design is far more complex than corporate. Beyond agency and policy names, it must cover the nicknames, abbreviations and internet slang the public actually uses (e.g., a policy's critical nickname). A complete keyword system should be co-designed by people who know the policy background and NLP experts, and updated regularly to include emerging issue tags.

資料來源的代表性：政府輿情分析的資料來源應涵蓋不同政治立場的媒體（避免單一立場的資料偏差）、以及不同世代活躍的社群平台。常見的一般性假設是：Facebook 的使用者年齡層偏高、Instagram 與 TikTok 偏年輕、PTT 上的討論較容易被媒體引用，但這些只是待驗證的假設，各平台的受眾結構會隨時間與議題變動，且同一平台內不同社團或看板的組成也差異極大。實務上應以專案自身的資料驗證：先檢視各平台在目標議題上的實際發文與互動組成，再決定各來源在分析中的權重，而不是直接套用刻板印象。若機關的目標族群明確（例如特定年齡層或特定地區居民），更應搭配該族群實際活躍的管道另行盤點。資料的多元性與來源結構的透明說明，是確保民意分析可被檢驗的基礎。

Analysis cadence and triggers: agencies should run a two-tier mechanism — regular reports (weekly or monthly) for long-term satisfaction tracking, and event-triggered instant analysis that starts automatically on major policy announcements, incidents or abnormal sentiment swings, providing an immediate full-picture assessment.

Analyst training: tools only assist — final analysis and interpretation still need professionals with policy background. Agencies should invest in cross-disciplinary talent combining policy understanding with data analysis skills, a profile still scarce in Taiwan's public sector.

Special Considerations for Election Sentiment Analysis

Elections are key moments in Taiwan's political life and the most intensive application scene for sentiment analysis. Election sentiment analysis in Taiwan, however, carries several ethical and legal considerations that deserve special attention.

Candidate volume monitoring: sentiment systems track candidates' online volume, positive/negative ratios and topic heat in real time — highly valuable for campaign strategy and media buying. Note, however, the significant gap between online volume and actual voting intention (the online-active population does not represent the whole electorate); over-reliance on sentiment data can distort strategic judgment.

Disinformation and manipulation detection: in Taiwan's election environment, fabricated disinformation and coordinated opinion manipulation (including bot-account flooding) are serious problems. Advanced sentiment systems detect anomalous propagation patterns — e.g., a topic simultaneously reposted by hundreds of new accounts within a short window — to identify suspicious manipulation and support fact-checking by election authorities and media.

Pre/post-election sentiment comparison: systematically comparing pre-election agenda-setting with post-election opinion trends helps researchers understand the gap between campaign promises and governance evaluation, and which pre-election issues keep generating discussion afterwards — informing policy planning for the next election cycle.

Government Procurement and Security Compliance: The Joint Supply Contract

Taiwan government agencies must procure IT services under the Government Procurement Act, and security-sensitive systems (such as platforms analyzing policy sentiment) face strict security review. Understanding the procurement compliance framework is a key precondition for public-sector sentiment system adoption.

共同供應契約（共契）是政府採購效率化的重要機制，由主辦機關統一辦理招標，各機關可依需求直接簽訂供應契約，免去個別機關重複辦理採購程序的時間與人力成本。現行的共契品項、有效期間、適用機關與訂購方式，應以政府電子採購網及共契資訊系統的公告為準；LargitData 可配合政府採購常見的作業流程，包含規格說明、需求訪談與試用安排，實際的採購途徑請依機關的採購單位評估後決定。

採購程序與現行規定可於政府電子採購網查詢：政府電子採購網。

政府資訊服務採購的資安要求並非單一標準，會依機關的資通安全責任等級（依現行制度分為 A、B、C、D、E 五級，主管機關為數位發展部）、系統的資料分級以及該次採購的規格書而不同。常見會被納入規格的項目包括：資料存放位置的限制、供應商的資訊安全管理制度驗證（例如 ISO 27001）、存取控制（多因素驗證與角色式存取控制）、傳輸加密、稽核留痕與保存期限，以及委外處理個資時的責任分擔條款。這些項目是否為必要條件、門檻如何訂定，須以機關現行的資安規範與該案規格書為準。

針對機敏等級較高的輿情資料，機關較常要求採用私有部署（On-Premise）或專屬環境，而非共享的 SaaS 服務；但這是風險評估的結果而非一體適用的規定，仍應依資料分級與機關政策逐案判斷。實際適用範圍與作業要求，仍應以主管機關最新公告及貴機關（政風、資安與採購單位）認定為準。

在採購規格書的撰寫上，建議政府採購人員參考以下關鍵規格項目：資料來源涵蓋率（應明確列出需涵蓋的平台清單與更新方式）、繁體中文情感分析的驗收方式、資料更新頻率與延遲上限、API 開放程度（確保資料可整合至機關既有的 BI 系統）、系統可用性與服務水準，以及資料匯出與契約終止時的資料處理方式。

其中情感分析的驗收方式特別值得留意：直接在規格書寫一個準確率百分比，往往無法驗收——因為沒有指定測試集、標註準則與計算方式時，任何數字都可以成立。較可行的作法是要求供應商在履約期間，以機關提供的實際樣本（涵蓋各平台與各議題，並包含反諷、否定與混合情感等困難樣本）進行盲測，由機關人員完成標註並計算各類別的精確率與召回率，再以此作為驗收依據；同時約定模型或詞典更新後的重測機制，避免上線後表現退化而無從追究。

Evaluating the Benefits of Government Sentiment Analysis

Agencies need an evaluation framework different from business: the public sector's core value is not profit but governance effectiveness and public trust. Benefit metrics suited to government agencies include:

政策溝通效率提升：導入輿情分析後，政策宣導團隊可望在宣導活動上線後較短時間內取得初步的民眾反應訊號，而不必等待傳統民調的完整結果。實際能多快拿到可用的資料，取決於議題的聲量高低（冷門議題可能數日仍樣本不足）、資料源的更新頻率與延遲、關鍵字體系是否已涵蓋民眾實際使用的說法，以及機關內部的判讀與簽核流程。這類「較快的反饋迴路」有助於縮短宣傳策略的迭代週期，但它取得的是網路上自發性的意見樣本，代表性與民調不同，不能直接互相取代。

Crisis warning success rate: track the share of major sentiment events for which the system issued warnings before eruption, and the average lead time from warning to full outbreak (how many hours ahead). This metric directly reflects the system's real value in crisis prevention.

民意調查成本節省：輿情分析能夠部分替代傳統的電話民調（但不能完全取代），對於部分常規的施政追蹤議題，可考慮以輿情分析降低民調的施作頻率，是否真能減少委外調查費用，取決於原本的民調頻率、議題是否適合以網路自發意見替代，以及機關對代表性的要求。建議政府機關將輿情分析費用與傳統民調費用合併計算，並記錄替代前後的實際支出與研究品質變化，再評估整體的研究效益比。

跨部門資訊共享效益：建立集中化的政府輿情分析平台，讓各部會、地方政府能夠共享同一套基礎資料與分析能力，避免各機關重複建置系統的資源浪費。共契等統一採購機制的設計目的之一即是推動這類跨機關資源共享，可望減少重複建置；實際能節省多少，取決於共用範圍、各機關需求的差異程度，以及既有系統的汰換時程，建議以實際的採購與維運資料檢視。