How does SharP use soft and hard prompts for low resource fake news detection?

Fluent synthetic text is not the same as useful evidence

The paper targets a specific setting: only a small labeled set is available for training, while evaluation follows a fixed and unchanged test protocol. The motivation is practical. Manual fact-checking is costly and ambiguous, and emerging topics and domains are poorly covered by existing benchmarks, so a detector cannot count on a large, well-matched labeled corpus.

The authors argue that existing augmentation routes share a weakness in this setting. Synonym substitution, generative adversarial networks, and LLM-based generation can produce grammatically correct and fluent samples, but the text is often overly generic. It lacks the task-specific relevance needed to separate real from fake news, because deceptive stylistic patterns and subtle linguistic cues are exactly what generic paraphrasing tends to smooth away.

A second gap concerns timing. Most augmentation methods do not consider where the detector is currently weak during training, so the synthetic data may not land in the input regions where the model is uncertain or error-prone. SharP is designed around both gaps: make generation task-aligned, and aim it at the classifier's uncertain region.

Hard prompts constrain the task while soft prompts adapt to the data

SharP's dual prompt mechanism splits the guidance job in two. The hard prompt is a set of discrete instruction tokens that provides structured guidance and factual constraints. The soft prompt is a learnable continuous prefix that introduces semantic patterns learned from the data. In the implementation, the input text is concatenated with the hard prompt, and the soft prompt vector is added to the resulting embedding before the LLM produces its hidden representation and decoded output.

This division matters because the two components fail differently. A hard prompt alone can specify the task but cannot adapt to a particular dataset's style. A soft prompt alone can fit the data but has no explicit instruction about what the generated text must preserve. Combining them lets the generator stay anchored to the news domain while still shifting toward the patterns the classifier actually sees.

The paper treats the LLM as a tool adapted to produce task-specific synthetic samples, not as a general-purpose paraphraser. Each generated sample inherits the label of its source instance, so augmentation expands the labeled pool without introducing new label decisions.

Uncertainty selects which samples are worth generating

SharP runs as an iterative active learning loop. At each round, the classifier scores unlabeled instances, and predictive entropy measures how uncertain it is. Entropy is computed from the two class posteriors, and instances whose entropy meets or exceeds a threshold form the boundary sample set. Those instances are queried for labels, either from human annotators or an auxiliary labeling model, and then passed to the augmentation module.

The classifier itself combines two feature streams. A CNN over BERT word embeddings captures word-level text representation, while a bidirectional GRU over emotion features captures sentiment flow; the two are fused by element-wise multiplication. The emotion features come from five lexicon categories, including eight NRC Emotion features, two sentiment features, ten moral features, two imageability features, and roughly 350 hyperbole words.

The augmented set is generated from the queried batch, and the labeled pool is updated with both the queried originals and their synthetic counterparts. This is why the boundary focus is not cosmetic: the synthetic samples are produced precisely where the current classifier is least confident, which the authors describe as helping clarify subtle distinctions between real and fake content.

Two objectives balance source fidelity against classifier feedback

During pre-training, the soft prompt is learned with an objective that rewards semantic alignment: the cosine similarity between the feature representation of the original input and that of the generated output, plus a regularization term. This keeps generated text close to the source data rather than drifting into unrelated content.

During fine-tuning, the objective becomes a weighted sum of two losses. The first is the same alignment loss. The second is a squared error between the classifier's score on the generated output and a target score, which pushes generation toward samples that are useful for refining the decision boundary. The weights on the two losses are tunable, and the paper reports that models using only a single loss performed worse under all three sample conditions than models combining both.

The authors' interpretation is that this dual objective yields samples that are both domain-consistent and helpful for sharpening decision boundaries. That is a claim about the design rationale supported by the ablation-style comparison, not a separate measurement of each generated sentence's quality.

Prompt choice and generation volume change the size of the gain

The sensitivity analysis evaluates hyperparameters at training data proportions of 5%, 10%, 30%, and 50%. Hard prompt category mattered most when training data was small: differences between categories were minimal with large training sets, but certain categories performed significantly worse in the small-data regime. The authors hypothesize that the classifier prefers particular text styles, noting that a more aggressive-toned prompt outperformed a more declarative one, possibly because fake news often uses a provocative style. Hard prompt 1, using simpler and more straightforward expressions, achieved the best or joint-best accuracy in most settings and became the default.

Soft prompt length showed an inverted pattern: performance improved as initialization length grew, peaked, then declined. The authors attribute the decline to an overly long soft prompt occupying a large share of the concatenated input and hindering the model's understanding of the task objective.

Generated text volume behaved similarly, rising briefly to a peak and then declining continuously. When training data was relatively low, the peak arrived later, which the authors explain as added diversity helping more when data is scarce, while excessive semantically similar augmented text hurts once data is sufficient.

What the evidence supports and where it stops

The paper's central claim is that dual prompt-guided, boundary-aware augmentation improves detection precision and overall classification accuracy in low resource settings by supplying diverse, high-quality samples that preserve original labels. The reported support comes from experiments under the fixed test protocol described in the paper, with hyperparameter analysis across the four training data proportions.

Scope conditions are built into the design. The method assumes queried instances can be labeled, whether by annotators or an auxiliary labeling model, and it assumes the classifier's entropy is a usable signal for locating the boundary. The augmentation inherits source labels, so label noise in the queried batch would propagate into synthetic samples. The authors also position the framework as model-independent and adaptable to various LLM architectures, and they list extension to other text classification tasks, multilingual data, distribution shifts, and multimodal inputs as future work rather than demonstrated results.

The practical priority that follows is a comparison, not a blanket recommendation. When labeled data is very scarce, the choice of hard prompt category and the number of generated samples matter enough to change accuracy, and the peak generation volume shifts later as data shrinks. SharP's advantage is therefore best read as conditional on tuning those two knobs to the resource level, with the dual objective kept in place rather than reduced to a single loss.

Sources used in this answer