Skip to content

k-CNN

📄

k-CNN classifies whether a causal relation holds between two already-marked entity spans by combining two parallel convolutional channels: a knowledge-oriented (K) channel that encodes linguistic cues from the text strictly between the two entities, and a data-oriented (D) channel that encodes the whole sentence with position-relative embeddings.

Introduced by (Li and Mao 2019)1 for binary causal-vs-non-causal relation classification over an already-identified entity pair, k-CNN directly targets Causality Identification — it needs both spans marked ahead of time and does not itself detect or extract them.

Architecture

K-channel (knowledge-oriented): only the lemmatized words strictly between the two entity spans. Convolves with cosine similarity (unit-vector dot product) over 1/2/3-word windows, then max-pools. Filters are pruned by an ANOVA F-test (keeping only filters whose class separation is statistically significant) and the survivors are clustered (k-means, separately for unigrams/bigrams) to control dimensionality.

D-channel (data-oriented): the full sentence, each word embedded together with its relative position to both entity heads (20-d trainable position embeddings). Convolves with a standard dot product + tanh over 3/4-word windows, then max-pools.

Semantic features (optional): WordNet top-level category one-hots for both entities (82-d) plus four FrameNet causal-lexicon scores, summed separately over the before/between/after regions of the sentence relative to the entity pair.

The K-channel, D-channel, and semantic features are concatenated and fed to a small fully connected softmax classifier (dropout 0.4). Word embeddings are the frozen, unit-normalized Levy & Goldberg dependency-based embeddings — not the more commonly substituted word2vec or GloVe embeddings.

Evaluation

(Li and Mao 2019)1 reports macro-averaged F1 (10-fold cross-validation, averaged over 10 seeds) of 92.64 on SemEval-2010 Task 8, 77.86 on CausalTimeBank, and 83.31 on Event StoryLine for the best configuration (both channels plus semantic features). The knowledge-oriented channel — restricting convolution to only the words between the two entities, rather than the whole sentence — is the paper's central contribution over a plain sentence-level CNN baseline.

Strengths and Limitations
Strong on SemEval-2010 Task 8 Competitive F1 without any contextual (transformer) embeddings
Identification-only Needs both entity spans already marked; cannot detect or extract spans itself
Needs markers on negatives too Both channels require two marked entity spans on every example, including non-causal ones — datasets whose causality-identification conversion only marks entities on causal-relation rows (rather than on every candidate pair) can't be fed to k-CNN as-is; see resources/models/kCNN/notes.txt for the concrete blocker hit reproducing it against causalatee's own converted tables
Sensitive to embedding choice Built specifically around Levy-Goldberg dependency-based embeddings, whose original distribution site is no longer reliably available

Citation

@article{li:2019,
  title = {Knowledge-Oriented Convolutional Neural Network for Causal Relation Extraction from Natural Language Texts},
  author = {Li, Pengfei and Mao, Kezhi},
  year = 2019,
  journal = {Expert Syst. Appl.},
  volume = {115},
  pages = {512--523},
  doi = {10.1016/J.ESWA.2018.08.009},
  url = {https://doi.org/10.1016/j.eswa.2018.08.009},
  urldate = {2024-07-19}
}

  1. Li, Pengfei, and Kezhi Mao. 2019. “Knowledge-Oriented Convolutional Neural Network for Causal Relation Extraction from Natural Language Texts.” Expert Syst. Appl. 115: 512–23. https://doi.org/10.1016/J.ESWA.2018.08.009. ↩↩