Review Comment:
I thank the authors for their extensive revision and comments. The article and the repository material have been significantly improved and while I still have few comments/questions, these are rather minor:
- Section 3.1 / my Comment 3: While 3.1 has been revised, I still somewhat miss a very clear objective of Event-centric Subgraph Extraction. I suggest to start Section 3.1 with a sentence such as "Given a target event of interest and a knowledge graph $KG$, our goal is to extract an event-centric knowledge graph from $KG$ which covers relevant information regarding this event. To do so, we propose a link traversal-based method..."
- Definition 7: Two questions: (i) I assume that s' is always an event if $subevent_of(s', n_{start})$ holds, i.e. "$type(s') = event" is redundant? (analogously for o') (ii) By this definition, $n_{start}$ itself is missing from the graph (if not $subevent_of(n_{start}), n_{start})$) which is not intuitive). This is also confirmed by https://github.com/SonyCSLParis/graph_search_framework/blob/main/kg-exam... where dbr:The_French_Revolution is not defined.
- p17, l17: As you confirmed in your comment, you are reusing the method in [8]. This is totally okay but could be made even more explicit, e.g., "...using the transformer-based model by Chanin et al. [8]".
- Fig. 5 shows the result before applying entity linking, correct (that's why "the French Revolution" is a literal)? Can this be clarified in the caption?
- Table 5: I struggle to immediately understand this table. An updated caption should help. I assume the cell numbers refer to the number of events in the KGs, so make that clear (not just "Statistics on all events). Also clarify what "Final" means here.
- 4.2.3: I still think that the study design is not detailed enough: how were the questions created (based on the six question types which presumably served as templates)? How many questions were investigated in your study? I also don't clearly understand how you map question to the event. I assume that your input consists of the question and the KG created about the event in the question? Or was there any kind of automated matching to the event involved?
- Table 15: For several question types, there is frame-based information provided. However, now we know that there on average 28k triples extracted from the text. What part of these triples goes into the prompts?
- Code: The new folder (https://github.com/SonyCSLParis/graph_search_framework/tree/main/kg-example) is good, but it should definitely get a readme file explaining what each of these files are.
Minor:
- My old comment 3: I was asking if, non-sub-event relations, for example, preceding events (e.g., WWI->WWII) are part of the event-centric KG or not. And if yes, if they are considered sub-events.
- Algorithm 1: "see Section 3" is a bit rough.
- p10, l46: Consider briefly extending "candidate nodes are scored and ranked [using/by/through ...] to guide the next iteration"
- p13, l27: "p, p \in R"?
- p 22, l51: 'DBO:ABSTRACT" (and why even upper-cased)?
Typos/very minor:
- p3, l45: The indents of the RQs are somewhat off.
- p7, l37: Dot missing at the end.
- p10, l49: "highlight"
- p15, l10+12: text overflow to the right
- p16, l44: The correct wd prefix is http://www.wikidata.org/entity/ (it is then forwarded to "wiki/")
- Footnote 19 is split over two pages
| 
