Post-hoc explanation methods are an important tool for increasing model
transparency for users. Unfortunately, the currently used methods for
attributing token importance often yield diverging patterns. In this work, we
study potential sources of disagreement across methods from a linguistic
perspective. We find that different methods systematically select different
classes of words and that methods that agree most with other methods and with
humans display similar linguistic preferences. Token-level differences between
methods are smoothed out if we compare them on the syntactic span level. We
also find higher agreement across methods by estimating the most important
spans dynamically instead of relying on a fixed subset of size $k$. We
systematically investigate the interaction between $k$ and spans and propose an
improved configuration for selecting important tokens.

通过研究不同解释方法对词汇重要性的选择差异，我们发现选择不同类别的词汇会导致方法的分歧，而与其他方法和人类的最一致的方法更倾向于选择类似的语言偏好，通过在句法范围级别上比较可以减少方法间的差异，通过动态预估最重要的范围而不是依赖于固定的部分，方法间的一致性较高，我们还提出了一种改进的配置来选择重要的词汇。

句法跨度偏好在事后解释异议中的作用

The Role of Syntactic Span Preferences in Post-Hoc Explanation  Disagreement

Instruction-tuned Language Models (LMs) are widely used by users to address
various problems with task-specific prompts. Constraints associated with the
context window length and computational costs encourage the development of
compressed prompts. Existing methods rely heavily on training embeddings, which
are designed to accommodate multiple token meanings. This presents challenges
in terms of interpretability, a fixed number of embedding tokens, reusability
across different LMs, and inapplicability when interacting with black-box APIs.
This study proposes prompt compression with reinforcement learning (PCRL), a
novel discrete prompt compression method that addresses these issues. PCRL
employs a computationally efficient policy network that directly edits prompts.
The PCRL training approach can be flexibly applied to various types of LMs, as
well as decoder-only and encoder-decoder architecture, and can be trained
without gradient access to LMs or labeled data. PCRL achieves an average
reduction of 24.6% in token count across various instruction prompts while
preserving performance. Further, we demonstrate that the learned policy can be
transferred to larger LMs, and through various analyses, we aid the
understanding of token importance within prompts.

本研究提出了一种基于强化学习的离散提示压缩方法 (PCRL)，通过使用计算效率高的策略网络直接编辑提示，实现对各种类型的语言模型以及解码器和编码器 - 解码器架构的灵活应用，同时减少平均令牌数量 24.6%，且保持性能不变。此外，通过多种分析，揭示了提示中令牌重要性的理解。