- The objective is to annotate a subset of the entities identified during entity annotation as salient
- Only mentions recognized in entity annotations can be marked up as salient, and by default all mentions are non-salient
- The primary criterion for determining entity salience is when an entity is mentioned in the summary of a document
¶ Direct, indirect and non-nominal mentions
- If an entity is mentioned in the document and is referred to directly using a nominal phrase in both the document and the summary, we consider the summary mention to be direct, and we annotate the mention in the document as salient
- If an entity is identified as a mention in the document, but is not represented by a nominal phrase in the summary, we still mark it as salient in the document. Prominent examples include:
- The summary mentions an adjective corresponding to an NP mention in the text:
- Summary: "... a French trend ..."
- Document: "...[a trend from [France]]..." (also annotate "France" as salient)
- The summary mentions an event as a verb which appears as a noun in the text:
- Summary: "... destroyed Rome ..."
- Document: " ... [the destruction of [Rome]]..." (also annotated the destruction as salient)
- If an entity is mentioned in the document as a non-nominal mention (i.e. as a verb or even a sentence), then it is only annotated as salient if it is already a mention based on the entity annotation guidelines (e.g. it is coreferred to by a pronoun), and it is also mentioned in the summary in some way (directly or indirectly)
- Mentions which are nested as part of a name but are subjectively not actually a salient topic in the document should be regarded as non-salient, for example:
- Summary: Jenna Nicole Mourey, better known as Jenna Marbles, is a very successful American YouTube personality, vlogger, comedian and actress, known for her videos "How To Trick People Into Thinking You're Good Looking" and "How To Avoid Talking To People You Don't Want To Talk To".
- Document: [Her video "How To Trick [People] Into Thinking You're Good Looking"] (the mention "People" is not a salient entity but just part of the titlw, despite the salient mention of the video in the document and summary)
¶ Aggregate and specific mentions
In the following example, "award winning" is mentioned in the summary, but the document contains three specific award mentions - these are then all taken to be salient:
- Summary: Jeanne Moreau was an award winning French actress, singer, screenwriter and director, best known for starring in the films Elevator to the Gallows (1958) and Jules et Jim (1962).
- Salient mentions sanctioned by "award winning" in the document:
- the Cannes Film Festival Award for Best Actress
- the César Award for Best Actress
- the BAFTA Award for Best Foreign Actress for Viva Maria ! ( 1965 )
¶ Dates and date ranges
Mentions of date ranges in a summary do not sanction marking specific dates or years as salient. For example, if a summary mentions "a 4th century monk", but the document lists exact years of birth and death ("324 CE--377 CE"), then those years are not annotated as salient, since the years in the document are only a subset of the range mentioned in the summary.
If a summary makes reference to a kind of entity, for which multiple subsets exist in the document, all subsets are marked as salience. For example:
- Summary: Friends discuss Sam's interest in retirement, whether or not some people they know who are involved in politics are political appointees, and growing basil.
- Can I have [some basil]x? ... I gave him [some basil]y (the set denotations for x and y are non-identical, since x is a hypothetical basil, and y is some basil given to someone else; but since the summary refers only to 'growing basil', both sets are labeled as salient)
If a summary refers to a coordinate NP, and both the coordinate entity and its components are mentioned, both are marked as salient:
- Summary: Two people are playing a game before cooking dinner
- Document: - [We] need to keep playing. - But [I] said [you] have to cook first! ('We' contains 'I' and 'you', the 'two people' in the summary - all three mentions are salient)