Research Topic:

Explainable Phrasal Reasoning (EPR)

​

EPR is a phrase-oriented, weakly supervised, interpretable NLI reasoning scheme that combines neural networks and fuzzy logic to achieve interpretable and high-performance natural language inference and automatic interpretation generation, thus advancing the progress of neural symbolic hybrid AI.

• Objective: To address the "black box" problem and weak interpretability of deep learning NLI models, enabling models to provide observable reasoning processes that explain "why sentences are implied, contradictory, or neutral".

• Method framework:

  • First, the sentence's phrases are automatically detected (instead of words, as phrases are more meaningful semantic units).

  • Align phrases in premise and hypothesis sentences using embedding similarity.

  • NLI labeling (Entailment, Contradiction, Neutral) is performed on phrases aligned using a neural network, while unaligned phrases are also treated with special labels.

  • Finally, fuzzy logic is used to summarize phrase-level labels into overall sentence labels. The entire EPR model supports end-to-end weakly supervised training—using only sentence labels, without manual annotation of phrase-level labels.

• Innovation and Advantages:

  • It can provide explicit, phrase-based logical explanations for NLI decisions, which are easy for humans to understand.

  • Phrase-level reasoning training can be completed with weak supervision without the need for fine-grained manual annotation.

  • The model is almost differentiable everywhere, and backpropagation can be performed directly using gradients, which is highly efficient.

  • The proposed inference results can also improve automatic text interpretation generation (e.g., a 2-point increase in BLEU score on the e-SNLI dataset), reaching a new performance level.

• Experimental results:

  • Compared with traditional, pre-trained and baseline methods, it significantly outperforms the state-of-the-art (SOTA) in phrase reasoning F-score.

  • Models with fewer parameters (such as T5-small) combined with EPR logic achieve better interpretation and generation results than large models plus manual annotation.

  • Validated on mainstream datasets (SNLI, MNLI), the inference performance, text interpretation, and sentence-level classification are all well-balanced.