Combine topical relevance with visual saliency in annotation selection

Combine topical relevance with visual saliency when one annotated stock chart must both stay on-topic and help explain the plotted history. For example, rank candidate weeks separately by article similarity and by visually salient stock behavior, combine those ranks, and annotate a representative article from the top combined weeks instead of selecting by one feature alone or at random.

A context chart often has two jobs at once: stay aligned with the reader’s article and comment on important visible moments in the data. A combined ranking keeps either objective from dominating completely.

Mechanism: Combined selection places annotations on graphically important periods while preserving a topical connection to the article that triggered the chart.

Evidence: The paper reports that the combined system better balances graphical salience and relevance than the baseline, while the evaluation also shows that single-feature versions optimize their own rating dimensions more strongly in isolation (Hullman et al., 2013).

Notes: The authors also included article-volume information in the full system, but they note that its independent contribution needs further investigation.

• User Goal: Get one chart that is both relevant to the current article and useful for reading the stock history.
• Data: Stock time series paired with dated company articles.
• Chart Setting: Auto-generated annotated line chart embedded with a news article.
• Audience: Readers who need quick context rather than manual searching across many articles.
• Success Criterion: The annotations are both on-topic and placed on graphically important periods.

Break it when: The user’s need is mainly data-driven explanation or mainly content-driven background. Why: The paper suggests weighting saliency higher for data-driven needs and relevance higher for content-driven needs instead of forcing a fixed balance.

Sacrifice: You may not achieve the highest possible score on either relevance alone or saliency alone.
Risk: Fixed feature weights can mismatch the user’s immediate goal.
Mitigation: Reweight the combined ranking toward saliency for data-driven needs and toward relevance for content-driven needs.

Mistake: Optimizing only one feature when the chart must provide both context and visual explanation. Why it fails: The result either drifts off-topic or ignores important turns in the series.

Failure Sign: Either salient weeks remain unannotated or annotated weeks feel off-topic relative to the input article.
Quick Check: Inspect each selected annotation for both topical similarity to the input article and proximity to major stock changes.
Stronger Test: Compare the combined output against a relevance-only version and a saliency-only version and confirm that the combined set does not fail either dimension outright.

• Rank candidate weeks separately by linguistic relevance and visual saliency before making final selections.
• Combine the separate ranks instead of choosing annotations from only one ranking or at random.
• Select one representative article from each top-ranked combined week.
• Adjust the feature weights toward the user’s primary need when the default balance is not appropriate.