What are semantic roles (SRL) in content?

Whether the main topic of the page performs the action in your sentences or merely receives it. The active voice gives the model a complete "who - does what - to what"; the passive voice leaves a hole in that structure.

Context

Why do semantic roles matter to AI models?

A model extracts facts as “who - does what - to what”. A passive sentence with the actor dropped leaves a hole in that structure, which the model has to guess or skip. An active sentence delivers the full set and can be quoted without rewriting.

The second effect is about visibility. When the main topic consistently appears as the actor, the model binds attributes and actions to it - the page starts being “about something” in the entity sense, not merely containing words on the subject. A blurred perspective yields blurred connections.

In practice

What will you see in the report?

You get the share of sentences where the main topic is the actor, plus a list of sentences to rewrite - with a ready active-voice proposal.

This is usually the fastest dimension to act on: the fixes are mechanical and require no new content.

Sample recommendations

Fragments of a report from an audit of an electronics store category page.

Problem: The main topic is the object of the reader’s action: the agent of the sentence is “you”, not the product being described.

Before

In our store you will find a wide choice of wired and wireless models optimised for different game genres.

After

Wired and wireless gaming mice cover every game genre — from FPS, through strategy, to MMO.

Problem: The sentence says what “you can do” instead of what the product does — the agent disappears from the fact.

Before

Thanks to varied sensors you can match the mouse to your needs.

After

Sensors with up to 26,000 DPI match the mouse sensitivity to your play style.

Method

How do we measure semantic roles?

We check the role the page’s central entity plays across sentences: the actor performing an action, the object of that action, or absent. The actor role counts in full, the object role counts half, absence counts for nothing. The score is the sum of collected points divided by the maximum possible, rescaled to ten.

The reference point: a well-written page has its main topic in the actor role in over 70% of the sentences where it appears at all.

How a role counts in a sentence

Role of the central entityValue
Actor (“The bank grants the loan”)full
Object (“The loan is granted by the bank”)half
Absent from the sentencenone
Factors

What raises and what lowers the score?

Raises

  • Active voice with an explicit actor
  • The main topic named directly instead of replaced by a pronoun
  • Concrete verbs (“lowers”, “shortens”) instead of constructions built on “is”
  • A consistent perspective across the text - the same protagonist in the sentences

Lowers

  • Passive voice with the actor dropped (“was implemented”, “is applied”)
  • Subjectless constructions (“one should remember”, “it is worth considering”)
  • Replacing the topic with a pronoun for whole paragraphs
  • A shifting protagonist - product, then customer, then company, with no reason

The most common fix is turning the passive voice active

Instead ofWrite
“The loan is granted for up to 30 years”“The bank grants the loan for up to 30 years”
“The panel was installed”“The crew installs the panel in a single day”
“Servicing should be remembered”“A heat pump needs servicing once a year”
Questions

Frequently asked questions

Is the passive voice always a mistake?

No - the problem starts when the actor is dropped. “The bank makes the decision” is fine; “a decision is made” leaves the model with no information about who makes it.

What if the topic is a service rather than a company?

Then the actor is often the service or product itself: “the policy covers treatment costs”, “the heat pump cuts the bill by 40%”. The point is an unambiguous role for the main topic, not inserting the company name everywhere.

How quickly can this be fixed?

It is usually the cheapest dimension to repair. The fixes are mechanical - reordering a sentence - and need no new content or new data.

Related

Related dimensions