Multlingual Abstractive Event Extraction for the Real World

1. This work constructs an event extraction dataset but does not provide detailed data analysis, such as the distribution of event types across different languages, argument distribution, document count, and entity distribution.
2. The experimental section is insufficient and needs additional experiments to verify the effectiveness of the proposed dataset.
3. Table 4 shows an uneven distribution of event types. Is this factor considered in the experiments?
4. Table 2 presents the overall results, but how do the results vary across different event types and languages?
5. Supervised experiments based on large language models are crucial, and the authors should focus on this aspect in the methodology section.

This paper contributes to event extraction by defining a schema for abstractive event extraction and creating a high-quality dataset. However, from my perspective, it has several weaknesses.

1. First, while the defined schema is general, it lacks specificity. For example, arguments such as "entity," "group_1," and "group_2" are extracted without a precise argument type, which may limit practicality in real-world applications. A more useful approach could involve defining argument types as an open-set extraction task, where argument types are inferred from the context rather than using general labels.

2. Second, the authors discuss some challenges in current work, such as Entity Normalization and Linking. However, not all challenges are thoroughly addressed in this paper. For instance, event coreference resolution is mentioned but not actually covered, which raises reader expectations without fully addressing the challenge.

3. Finally, as a potential regular paper at ICLR, the methodology feels relatively weak aside from the dataset contribution. The framework design is straightforward, which diminishes the overall impact of the paper's contributions.

1. The key claim of the paper is to have a new tool for real-world EE, and thus propose the new Abstractive event extraction problem. However, the author lack the evidence to prove the how these new paradigm of task is more helpful than classical definition of event extraction. The discussion on different use cases should be included.

2. I think the claimed "novel/new" Abstractive event extraction problem is simply the previous event extraction annotation paradigm plus a "entity-linking" filtering on argument lists. For example, in Figure 1, the entities "two communties" can be and should be disambuigated by havinge an entity-linker after the traditional event extraction system. And this practice has been used in the information extraction communities for a long time. For the two event or a single event case in Figure 1, "event coreference system" are designed for this purpose. For these point of view, I do not see the claimed "Abstractive event extraction" is a new task.

3. Several of the design choice are arguable:

• Why single event for each writing is enough? If we consider diverse potential downstream applications, such as event interactions, plot understanding. Single event is far from useful.
• It is unclear to me whether intermediate annotations SHOULD be included or excluded for annotation. I think in many previous efforts, like ACE. These intermediate annotations is very important for (a.) guarantee the annotation quality (b.) evaluating step-by-step performance. In this paper, I do not see how missing these annotations influence the annotation quality and the corresponding IAA. I think this paper neglect the potential ambuiguity done by ACLED dataset.

Didn't really show much why Lemonade is useful and stands out among other such datasets, or show how ZEST is a performant system for AEE, as the results are not a prevailing sweep and the baseline LLaMAX is not a widely recognized system for such a task.