AI helps you reading Science

AI generates interpretation videos

AI extracts and analyses the key points of the paper to generate videos automatically

Go Generating

AI Traceability

AI parses the academic lineage of this thesis

Master Reading Tree
Generate MRT

AI Insight

AI extracts a summary of this paper

We conduct systematic experiments to evaluate the quality and applications of the extracted knowledge. Both human and extrinsic evaluations show that ASER is a promising large-scale eventuality knowledge graph with great potential in many downstream tasks

ASER: A Large-scale Eventuality Knowledge Graph

WWW '20: The Web Conference 2020 Taipei Taiwan April, 2020, pp.201-211, (2020)

Cited by: 0|Views197


Understanding human’s language requires complex world knowledge. However, existing large-scale knowledge graphs mainly focus on knowledge about entities while ignoring knowledge about activities, states, or events, which are used to describe how entities or things act in the real world. To fill this gap, we develop ASER (activities, state...More



  • In his conceptual semantics theory, Ray Jackendoff, a Rumelhart Prize1 winner, describes semantic meaning as ‘a finite set of mental primitives and a finite set of principles of mental combination [18]’.
  • ASER contains 64 million edges among eventualities.
  • Table 1 provides a size comparison between ASER and existing eventuality-related knowledge bases.
  • We use patterns from dependency parsing to extract eventualities E’s from unstructured large-scale corpora.
  • All eventualities in ASER are small-dependency graphs, where vertices are the words and edges are the internal dependency relations between these words.
  • Take sentence ‘I have a book’ as an example, we will only select <‘I’, ‘have’, ‘book’> rather than <‘I’, ‘have’> as the valid eventuality, because ‘have’-dobj-‘book’ is a negative dependency edge for pattern ‘s-v’.
  • Used hyper-parameters and other implementation details are as follows: For preprocessing, we first parse all raw corpora with the Stanford Dependency parser, which costs eight days with two 12core Intel Xeon Gold 5118 CPUs. After that, We extract eventualities, build the training instance set, and extract seed relations, which costs two days with the same CPUs. For bootstrapping, Adam optimizer [19] is used and the initial learning rate is 0.001.
  • We ask them to label whether one auto-extracted eventuality phrase can fully and precisely represent the semantic meaning of the original sentence.
  • For both versions of ASER, we randomly select 100 edges per relation type and invite annotators to annotate them using the same way as we annotating the bootstrapping process.
  • We select the Winograd Schema Challenge (WSC) task to evaluate whether the knowledge in ASER can help understand human language.
  • Where ASERR (En, Ep ) indicates the number of edges in ASER that can support that there exist one typed T relation between the eventuality pairs En′ and Ep′ .
  • For each question sentence s, we first extract eventualities that contain the target pronoun Ep and two candidates Ec and select all edges (E1, R, E2) from ASER such that E1 and E2 contains the verb of Ep and Ec respectively and there exists one common word in E1 and E2.
  • From the result in Table 8, we can make the following observations: (1) Pure knowledge-based methods (Knowledge Hunting and ASER) can be helpful, but their help is limited, which is mainly because of their low coverage and the lack of good application methods.
  • We build seed relations among eventualities using unambiguous connectives found from PDTB and use a neural bootstrapping framework to extract more relations.
  • Both human and extrinsic evaluations show that ASER is a promising large-scale eventuality knowledge graph with great potential in many downstream tasks.
  • We choose connectives that are less ambiguous, where more than 90% annotations of each are indicating the same relation, to extract seed relations
  • After ten iterations of bootstrapping, the number of edges grows four times with the decrease of less than 6% accuracy (from 92.3% to 86.5%)
Your rating :