09-22-2021-jam.md

• Benchmarking the Combinatorial Generalizability of Complex Query Answering on Knowledge Graphs, compositionality

  • "including 301 different queries types", this could be used for reference when designing new linguistic phenomenon that is not covered by current evaluation datasets

• ReaSCAN: Compositional Reasoning in Language Grounding, nips2021 compositionality

  • This paper analyzes the limitations of a previously proposed benchmark gSCAN, and build upon gSCAN a new dataset that requires compositional understanding and reasoning over entities and relations
  • Section 6 of this paper is worth reading, since it analyzes different types of compositional generalization types

• A Plug-and-Play Method for Controlled Text Generation, emnlp2021 text generation plug-and-play

  • given a topic or keyword, a shift distribution is added to the language model's step-wise distribution
  • there are many details ignored here, and should be reading into the paper itself

• MIRRORWIC: On Eliciting Word-in-Context Representations from Pretrained Language Models, emnlp2021 word embedding

  • Word-in-Context task accomplished through representations from the pre-trained language model
  • a contrastive learning approach again!