0427 Method Course Presentation

AI in government-citizen communication

From interaction quality to deployment infrastructure

Paper 1

AI language modification and communication quality in government-citizen interaction.

Paper 2

PEARL and label-efficient representation learning for retrieval-based systems.

Today I will introduce two connected papers. The first is a public administration paper about whether AI language modification improves communication quality between government and citizens. The second is a method paper that moves one layer deeper. If we want to deploy AI in these communication systems, we also need reliable retrieval and classification infrastructure when labels are scarce.

Paper 1

Communication quality is not only information transfer

Citizen message

->

AI modification

->

Government response

Potential benefit

clearer information
more polite tone
lower response burden

Core concern

standardized language
muted emotion
weaker urgency signals

The first paper starts from a practical observation. Government-citizen interaction is increasingly text-based. Citizens write online messages, and officials reply through platforms. AI can make this communication clearer and more polite. But communication quality is not only clarity. It also includes respect, empathy, feeling heard, trust, urgency, and emotional signals. So the key question is whether AI improves communication in a deeper way, or whether it only makes the message look polished.

Framework

Two sides, two layers of communication quality

Informational-cognitive

clarity, satisfaction, ease of reply

Expressive-constitutive

politeness, respect, feeling heard, empathy, trust, urgency

The framework separates two kinds of communication quality. The first is informational-cognitive: does the message become easier to understand and easier to respond to? Here the expected effect of AI is quite direct. The second is expressive-constitutive: does the message express respect, emotion, responsiveness, and legitimacy? This is more contested. AI may improve surface politeness, but it may also reduce emotional signals that tell officials a case is urgent or needs empathy.

Research design

Vignette-based survey in China

Participants

220 citizens
214 civil servants

Contexts

requests, inquiries, complaints, suggestions, emergencies

Comparison

original messages vs AI-modified messages

Analysis: paired t-tests and mixed-effects regressions.

The design uses real-world government-citizen communication scenarios from China. The survey includes citizens and civil servants. Citizens evaluated government responses, while civil servants evaluated citizen messages. Each participant saw both original and AI-modified versions across different topics. This within-subject structure is useful because each participant partly serves as their own control. The analysis combines paired t-tests with mixed-effects regressions.

Citizen-side findings

AI improves citizen perceptions across all six dimensions

information clarity
response satisfaction
expressed politeness
feeling heard
empathic satisfaction
trust toward government

The findings are clear on the citizen side. AI-modified government responses were rated higher across all six dimensions, including not only clarity and satisfaction but also feeling heard, empathy, and trust. The important point is that the improvement is not only instrumental. It also appears in relational perceptions. This supports the idea that AI language modification can improve both the informational and relational sides of public communication.

Civil servant-side findings

AI reduces friction without clearly erasing urgency

clarity increased
easiness of reply increased
politeness increased
respectfulness increased
urgency and needed empathy were not significantly reduced

On the civil servant side, AI-modified citizen messages were clearer, easier to reply to, more polite, and more respectful. Importantly, the paper does not find consistent evidence that AI significantly reduces urgency or needed empathy. This means the strongest version of the emotional-flattening concern is not supported here, although the paper still treats it as a context-dependent risk.

Bridge to Paper 2

Better wording is only the front-end layer

Interaction level

clarity, politeness, empathy, trust, burden

System level

routing, retrieval, category matching, similar cases

If nearest neighbors in embedding space are wrong, the system retrieves the wrong cases.

This is where the second paper enters. Once we imagine AI-assisted communication in real systems, the task is not only rewriting a message. The system also needs to retrieve similar cases, route messages, and suggest relevant categories or policies. These tasks often depend on embeddings. If the nearest neighbors in embedding space are wrong, the system retrieves the wrong examples. So the method question becomes: can we improve embedding geometry with only limited labels?

Previous work: SE-VAE

Technical foundation: structured embedding representation

Latent structurelearn compact representations

Semantic alignmentconnect text space and task labels

PEARL extensionreshape fixed embeddings for retrieval

Before PEARL, my earlier SE-VAE work already focused on representation structure. The key connection is that PEARL inherits this concern with embedding geometry, but turns it into a lighter post-processing approach for fixed embeddings and retrieval systems.

Paper 2: PEARL

Label-efficient geometry shaping for fixed embeddings

z_i=f(x_i), y_i ∈ {1, ..., C}

p_c=¹|L_c| ∑_{i ∈ L_c} z_i

s_ic=^{cos(z_i, p_c)}τ

P(c|i)=^exp(s_ic)∑_k=1^C exp(s_ik)

L=CE(y_i, P) + λ ||z'_i - z_i||₂²

z'_i=norm(z_i + αWz_i)

z_iembedding vector for one text

p_caverage vector for class c

s_icsimilarity score to prototype

P(c|i)probability text belongs to class

Ltraining loss: fit label, avoid distortion

z'_irefined vector used for retrieval

Inputfrozen embeddings

Anchorclass prototypes

Outputretrieval geometry

PEARL targets a practical problem. Many systems already use fixed embeddings from large pretrained models. But these embeddings are not always organized in the way a deployment needs. PEARL uses limited labels to form class prototypes, then gently reshapes the embedding space so local neighborhoods become more label-consistent. The point is not to train a new foundation model. It is a lightweight post-processing step for systems where retraining is expensive or impossible.

Results and deployment meaning

PEARL improves early retrieval when labels are scarce

Evidence point	Raw / baseline	PEARL	Gain
Overall low-label neighborhood quality	Raw embeddings	Prototype-aligned embeddings	+25.7% over raw
Against unsupervised post-processing	Centering / L2 / whitening-style fixes	Uses limited labels as anchors	+21.1%+ over strong unsupervised baselines
Purity@1, 300 labels	0.3824	0.5370	+0.1546 points; about +40.4%
Hit@1, 600 labels	0.4386	0.6054	+0.1668 points; about +38.0%
MRR@20, 1200 labels	0.6202	0.7186	+0.0984 points; about +15.9%
Purity@5, 2500 labels	0.4540	0.6359	+0.1819 points; about +40.1%

The main result is that PEARL is strongest in the regime many public-sector deployments actually face: limited labels, fixed embeddings, and reliance on retrieval. It improves local neighborhood quality and early retrieval, especially Hit@1 and MRR. This matters because civil servants often inspect only the first few retrieved cases. When more labels become available, fully supervised methods can become stronger, so PEARL is not claiming to dominate every setting. Its value is as a practical, label-efficient geometry correction.

Closing

One research agenda: evaluate AI at two levels

Interaction level

clarity
politeness
empathy and trust
frontline burden

System level

retrieval quality
label efficiency
drift monitoring
auditability

Good AI governance communication is not only better wording. It also requires systems that retrieve the right cases and adapt with limited supervision.

Together, the two papers form one research agenda. The first paper evaluates what AI changes in the communication experience. The second paper asks how to support that communication system technically when labels are scarce and the environment changes. My final takeaway is that AI in public communication should be judged at both levels: the interaction level and the system level.