Skip to main content
ViceWire does not reduce an article to one broad positive or negative sentiment label. Directional analytics are attached to individual event objects and sit alongside event family, stage status, subtype, metadata, and affected business surface.

​
Two layers

​
1) Event-attached directional interpretation

Directional interpretation answers two simple questions:
  1. Could this event be favorable for the entity?
  2. Could this event be adverse for the entity?
Directional interpretation is attached to the event object rather than assigned once at the article level for two reasons:

Family-specific derivation

Direction is derived within the logic of the specific event family.For example, a favorable legal outcome is derived differently from a favorable competitive position.

Mixed articles stay intact

One article can contain a favorable demand signal for iPhone while also containing an unfavorable geopolitical development such as armed conflict. Storing direction inside each event object keeps those implications separate instead of flattening them into a single document-level view.
These directional fields are part of the event-extraction layer. They capture how a specific event may matter for an entity within the logic of that event family, not just the overall tone of the article.

​
2) FinBERT

Each event object can also carry a separate finbert layer. FinBERT is an open-source finance-specific NLP model. In ViceWire, finbert is computed from event-linked evidence associated with the confirmed event family. The model produces an aggregate tone for that event-linked evidence and also derives supporting diagnostics from the same evidence set. This means finbert should be read as an event-linked tone layer rather than a broad document-level sentiment label. finbert is organized into:
​
tone
tone stores the aggregate FinBERT output for the event-linked evidence.It is best interpreted as the overall tone of the evidence tied to that event, not as a broad document-level sentiment label.
​
tone_distribution
tone_distribution stores the percentage makeup of the scored event-linked sentences across negative, neutral, and positive classifications, plus the total number of scored sentences.
​
tone_extremes
tone_extremes stores the strongest single-sentence probability observed for each label across the scored event-linked evidence.It is best interpreted as the most extreme negative, neutral, and positive signal present in the event-linked evidence.
Overall, finbert should be used in conjunction with event-attached directional interpretation, not as a replacement for it.

​
Event context and mixed sentiment implications

Text tone and event meaning are not the same thing, for example:
  • A partnership can read as neutral in tone when the article is largely descriptive, while still carrying a meaningful favorable implication if the partnership enables product capability or distribution.
That is why ViceWire does not rely on a single document-level sentiment label or a simple positive-versus-negative event tag.

​
Simplified Example

In early 2026, articles reported that Apple was planning a major Siri revamp into a built-in conversational AI chatbot across iPhone, iPad, and Mac. The event is framed as part of Apple’s attempt to recover ground in generative AI after a weak Apple Intelligence rollout and to make Siri more competitive as a core product capability. In one article, Google and OpenAI appear as competitor references because the article presents them as the main external benchmarks and pressure points. OpenAI is positioned as a direct AI threat through ChatGPT and its broader platform ambitions, while Google is framed both as a competitive benchmark through Gemini and as a technical dependency because Apple is said to be relying heavily on Google-developed models for the Siri upgrade. This article touches on multiple L1 event families. If we focus solely on the Competition & Market Structure, the competitive read is mixed. To capture that, the directional layer asks two simple questions:
  1. Does the article explicitly support a materially favorable competition or market-structure implication for Apple? (material_favorable_implication_present)
  2. Does the article explicitly support a materially adverse competition or market-structure implication for Apple? (material_adverse_implication_present)
In this example, the model answers true to both questions. At the same time, FinBERT gives the event-linked evidence a negative tone reading. Both outputs are plausible because the article frames Apple as behind in the AI race, which supports the adverse side. But it also describes Apple taking a concrete step to accelerate Siri and Apple Intelligence capabilities, which supports a favorable strategic interpretation within the competitive landscape.
Example Payload: Apple Siri Gemini Integration
{
  "event_family": "competition_market_structure",
  "event_family_metadata": {
    "material_favorable_implication_present": true,
    "material_adverse_implication_present": true,
    "competitor_reference": ["Google", "OpenAI"]
  },
  "finbert": {
    "tone": {
      "negative": 0.37,
      "neutral": 0.31,
      "positive": 0.32,
      "label": "negative"
    },
    "tone_distribution": {
      "negative_sentence_share": 0.43,
      "neutral_sentence_share": 0.29,
      "positive_sentence_share": 0.29,
      "sentence_count": 7
    },
    "tone_extremes": {
      "top_negative_sentence_prob": 0.96,
      "top_neutral_sentence_prob": 0.87,
      "top_positive_sentence_prob": 0.78
    }
  }
}

Example Payload: Apple Siri Gemini Integration

See the full Apple Siri Gemini Integration Payload Example

​
The core idea

Directional layers in ViceWire are additive. They help interpret an event, but they do not replace event family, stage, subtype, metadata, or affected business surface. They should be read as structured interpretation layers, not as guaranteed predictors of market reaction.