[
  {
    "file": "paper_22.txt",
    "start": 169,
    "end": 179,
    "label": "Unsupported claim",
    "user": "Ed",
    "text": " STS tasks"
  },
  {
    "file": "paper_22.txt",
    "start": 635,
    "end": 964,
    "label": "Coherence",
    "user": "Ed",
    "text": "Specifically, Reimers and Gurevych (2019) mainly use the classification objective for an NLI dataset, and Wu et al. (2020) adopt contrastive learning to utilize self-supervision from a large corpus. Yan et al. (2021); Gao et al. (2021) incorporate a parallel corpus such as NLI datasets into their contrastive learning framework."
  },
  {
    "file": "paper_22.txt",
    "start": 1101,
    "end": 1201,
    "label": "Unsupported claim",
    "user": "Ed",
    "text": "One related task is interpretable STS, which aims to predict chunk alignment between two sentences ."
  },
  {
    "file": "paper_22.txt",
    "start": 1291,
    "end": 1313,
    "label": "Format",
    "user": "Ed",
    "text": "(Konopík et al., 2016;"
  },
  {
    "file": "paper_22.txt",
    "start": 1863,
    "end": 1881,
    "label": "Format",
    "user": "Ed",
    "text": " (Li et al., 2020;"
  },
  {
    "file": "paper_22.txt",
    "start": 2408,
    "end": 2746,
    "label": "Coherence",
    "user": "Ed",
    "text": ". To get the solution efficiently, Cuturi (2013) provides a regularizer inspired by a probabilistic theory and then uses Sinkhorn's algorithm. Kusner et al. (2015) relax the problem to get the quadratic-time solution by removing one of the constraints, and Wu et al. (2018) introduce a kernel method to approximate the optimal transport.\n"
  }
]