Implicit Discourse Relation Recognition (IDRR) remains a challenging task due to the requirement for deep semantic understanding in the absence of explicit discourse markers. A further limitation is that existing methods only predict relations without providing any supporting explanations. Recent advances in large language models (LLMs) have shown strong reasoning capabilities in both deep language understanding and natural language explanation generation. In this work, we propose a simple yet effective approach to distill the reasoning capabilities of LLMs into lightweight IDRR models to improve both performance and interpretability. Specifically, we first prompt an LLM to generate explanations for each training instance conditioned on its gold label. Then, we introduce a novel classification-generation framework that jointly performs relation prediction and explanation generation, and train it with the additional supervision of LLM-generated explanations. Our framework is plug-and-play, enabling easy integration with most existing IDRR models. Experimental results on PDTB demonstrate that our approach significantly improves IDRR performance, while human evaluation further confirms that the generated explanations enhance model interpretability. Furthermore, we validate the generality of our approach on sentiment classification and natural language inference
Figures & tables
Figure 1: Our lightweight IDRR model distills LLM reasoning to predict discourse relations and generate interpretable explanations. Arg1 and Arg2 are two arguments.
Figure 2: The prompt template for explanation generation. An explanation consists of two parts: the restatement of two arguments, and a rationale of their discourse relationship. S means two arguments.
Figure 3: Our proposed classification-generation framework. For ease of understanding, we use the IDRR task as an example to illustrate our framework. It first predicts the discourse relation y~ , which is then used as extra input to generate the explanation e~ .
Method
Acc (%)
F1 (%)
ChatGPT
50.24
44.09
LLaMA-3B + IDRR gt
65.11
57.11
LLaMA-8B + IDRR gt
70.36
61.01
T5 + IDRR gt w/o Exp
64.05
55.92
T5 + IDRR gt
64.63
56.65
T5 + IDRR ct w/o Exp
63.67
52.39
Table 1: Comparison with baselines. IDRR gt and IDRR ct means that IDRR is formulated as a generation task and a classification task, respectively. Here w/ Exp and w/o Exp indicate whether the explanation generation task is included or omitted.
Figure 4: Quality of generated explanations. (a) Compared with e-INFERSENT, our method achieves a higher average score of 4.20 out of 5. (b) The distribution of scores assigned through human evaluation. (c) The distribution of scores for each aspect.
Figure 5: Feature Importance Agreement (Left) and Robustness Equivalence (Right). Con denotes the proportion of label–explanation alignment. Random and Important indicate that randomly selected or top-ranked important features are occluded, respectively.
Example and Predictions
Explanation
Arg1 : Instead, the rally only paused for about 25 minutes and then steamed forward as institutions resumed buying. Arg2 : The market closed minutes after reaching its high for the day. Relationship : Temporal EIDRR : Temporal✓ w/o explanation : Comparison ✗
The first sentence describes the slow start of a rally, which then moved forward with institutions resuming buying. The second sentence mentions the closing price of the market, which was just after reaching its highest point for the day. The temporal relationship is evident because the second event (the market closing after reaching its highest point) happened after the first event (the rally slowing and then moving forward). This sequence of events establishes the temporal connection between the two sentences.
Arg1 : We view the market here as going through a relatively normal cycle Arg2 : We continue to feel that the stock market is still the place to be for long-term appreciation. Relationship : Expansion EIDRR : Expansion✓ w/o explanation : Contingency ✗
The first sentence states that the market is seen as undergoing a relatively normal cycle. The second sentence provides additional information about the market’s continued positive outlook for long-term appreciation. The second sentence expands on the idea introduced in the first sentence by giving more detail about the market’s outlook and the reasons behind it. Thus, the relationship between the two sentences is one of expansion.
Arg1 : It would have been too late to think about on Friday. Arg2 : We had to think about it ahead of time. Relationship : Contingency EIDRR : Contingency✓ w/o explanation : Comparison ✗
The first sentence indicates a situation where thinking about something would have been too late on Friday. The second sentence states that there was a need to think about it before it could be decided. The contingency relationship is demonstrated here because the need for thinking ahead of time (second sentence) is contingent upon the situation described in the first sentence. In other words, the need for thinking ahead is based on the situation being discussed.
Arg1 : Tokyo stocks closed off a significant but less-than-alarming 1.8 % on thin volume Arg2 : Hong Kong stocks declined 6.5 % in orderly trading Relationship : Comparison EIDRR : Comparison✓ w/o explanation : Expansion ✗
The first sentence states that Tokyo stocks closed lower by 1.8% on a narrow volume, which is less than analgamating, but less than analgamating one. The second sentence mentions that Hong Kong stocks declined by 6.5% in stable trading. These two sentences form a comparison by contrasting the closing performance of Tokyo stocks with the performance of HK stocks in different markets. The comparison highlights the difference in performance between the two stocks in terms of their closing value and their performance in different markets.
Table 2: Examples of IDRR explanations. More examples are in Appendix E and on Github (all test instances).