Implicit Discourse Relation Recognition (IDRR)

• Implicit Discourse Relation Recognition (IDRR) is the task of inferring discourse relations between text spans without overt connectives by analyzing semantics, syntax, and pragmatics.
• Hierarchical and multi-label models, including approaches like LDSGM and DiscoPrompt, leverage multi-level sense prediction and connective path strategies to manage ambiguity.
• State-of-the-art methods incorporate diverse corpora and modalities while addressing challenges such as label hierarchy, domain shift, and annotation inconsistency to drive future research.

Searching arXiv for recent and foundational work on implicit discourse relation recognition to ground the article in published papers. Implicit Discourse Relation Recognition (IDRR) is the task of identifying the discourse relation between two text spans, typically Arg1 and Arg2, when there is no overt discourse connective such as because, however, or moreover. In the Penn Discourse TreeBank (PDTB) formulation, the input is an argument pair and the output is a discourse sense label; unlike explicit discourse relation recognition, the model cannot rely on a connective as a trigger and must instead infer the relation from semantics, pragmatics, syntax, and cross-argument interaction. This makes IDRR substantially harder than explicit relation recognition, and recent work increasingly treats that difficulty as a consequence not only of missing lexical markers but also of label hierarchy, ambiguity, disagreement, domain shift, and evaluation instability (Xiang et al., 2022, Yung et al., 26 Mar 2025, Costa et al., 28 Aug 2025).

1. Task definition and label structure

In PDTB-style annotation, a discourse relation connects two arguments, and the standard top-level senses are Temporal, Contingency, Comparison, and Expansion. A common formalization is the classification mapping

$(Arg1, Arg2) \rightarrow y$

where $y$ is the discourse sense label (Cai, 7 Mar 2025). The same basic formulation underlies both early PDTB systems and later prompt-based, hierarchical, and multi-label models (Xiang et al., 2022).

The sense inventory is hierarchical rather than flat. The survey literature describes PDTB senses in three levels—class, type, and subtype—while more recent work operationalizes this hierarchy in different ways. LDSGM treats multi-level IDRR as conditional label sequence generation over top-level labels, second-level labels, and connectives as the third-level labels (Wu et al., 2021). HArch predicts probability distributions across all three PDTB 3.0 sense levels and, on DiscoGeM 2.0, uses 4 level-1, 17 level-2, and 28 level-3 senses after projecting missing lower-level senses (Costa et al., 28 Aug 2025). Prompt-based path prediction methods likewise encode the hierarchy explicitly: DiscoPrompt predicts paths such as Comparison -> Contrast -> however rather than isolated labels (Chan et al., 2023).

A persistent issue is that the single-label formulation does not fully match the phenomenon being modeled. Multiple papers emphasize that implicit discourse relations are often ambiguous, weakly signaled, or underdetermined, and that a single relation may legitimately express multiple senses (Yung et al., 26 Mar 2025, Costa et al., 28 Aug 2025). This motivates both multi-level modeling and multi-label or distributional prediction.

2. Corpora, annotation practices, and supervision regimes

The PDTB remains the central resource for IDRR, and its versions anchor much of the field’s experimental practice. The survey identifies PDTB 1.0, 2.0, and 3.0 as the principal annotation resources, with PDTB 3.0 adding revised sense hierarchy and many more intra-sentential relations (Xiang et al., 2022, Cai, 7 Mar 2025). CoNLL-2015 and CoNLL-2016 introduced PDTB-based shallow discourse parsing benchmarks with standardized evaluation settings (Xiang et al., 2022).

Recent work has broadened the corpus landscape in ways that expose phenomena underrepresented in Wall Street Journal-style data. DiscoGeM is central to this shift. A multi-task, multi-label model trained exclusively on DiscoGeM treats discourse annotation as a distribution over senses rather than a single gold label and reports that, for 302 relations annotated in both DiscoGeM and PDTB 3.0, the PDTB reference label matched DiscoGeM’s majority label only 92 times (30.5%), highlighting annotation mismatch across corpora (Costa et al., 2024). DiscoGeM 2.0 extends this line with multilingual, parallel, multi-label annotation in English, German, French, and Czech, with each relation annotated by at least 10 crowdworkers and converted into a multi-label sense distribution (Costa et al., 28 Aug 2025).

Dialogue and speech have required further departures from standard newswire settings. Edina-DR, built from the Edina self-dialogue corpus, contains 27,998 discourse relation pairs extracted by a weak-supervision pipeline and is explicitly designed for open-domain dialogue, where turns are short, informal, and topic-shifting (Ma et al., 2019). MM-IDR constructs multilingual multimodal datasets for English, French, and Spanish from TED/TEDx talks by exploiting translation explicitation and aligning extracted argument spans with audio, yielding paired text + audio examples for each discourse instance (Ruby et al., 4 Feb 2026).

These differences in corpus design are not incidental. They directly affect label inventories, class balance, ambiguity, and what counts as a correct prediction. Several papers therefore treat annotation philosophy itself as part of the IDRR problem rather than as mere background metadata (Yung et al., 26 Mar 2025, Yung et al., 26 Feb 2026).

3. Methodological development

Early and mid-period neural work focused on richer discourse representations for single-label PDTB classification. SeMDER introduced a neural recognizer with semantic memory, using a shallow convolutional encoder for surface representation, an attention-based semantic encoder over an external memory matrix, and a final recognizer that combines surface and semantic representations. On PDTB 2.0 top-level classification, SeMDER reported an average improvement of 2.56% F1 over state-of-the-art baselines, with especially large gains for Temporal (Zhang et al., 2016). DER then pushed multi-granularity representation further by combining subword, word, contextualized ELMo, sentence-level stacked encoders, and bi-attention over argument pairs; it reported the first result above 48% accuracy in 11-way classification and above 50% F1 in 4-way classification on PDTB 2.0 (Bai et al., 2018). The memory-augmented extension “Memorizing All” stored learned representations and relation labels for all training instances and reported 60.69% accuracy and 52.19% F1 on the 4-way task, described as the first time the 4-way setting exceeded 60% accuracy (Bai et al., 2019).

A second line of work made hierarchical dependence explicit. LDSGM reformulated multi-level IDRR as conditional label sequence generation, with a label attentive encoder, a GRU-based top-down decoder, and an auxiliary reverse-order decoder used during training for mutual learning (Wu et al., 2021). GOLF then combined a discourse relation encoder, staircase classifier, global hierarchy-aware contrastive loss, and local hierarchy-aware contrastive loss; with RoBERTa-large it reported 74.21 top-level F1, 60.11 second-level F1, and 20.66 connective F1 on PDTB 3.0, together with improved multi-level consistency (Jiang et al., 2022).

Prompt-based IDRR reshaped the task around connective or path prediction. PCP framed fine-grained IDRR as masked connective prediction with RoBERTa and showed that connective prediction outperformed direct label prediction, reporting 67.79 macro-F1 and 73.80 accuracy on PDTB 2.0 top-level classification with RoBERTa-large (Zhou et al., 2022). DiscoPrompt replaced flat label prediction with hierarchical path prediction, using T5 to predict top-level label, second-level sense, and representative connective jointly; for T5-large on PDTB 2.0 Ji it reported 70.84 F1 and 75.65 accuracy at the top level, and 49.03 F1 and 64.58 accuracy at the second level (Chan et al., 2023). TEPrompt supplemented the main Discourse Relation Recognition prompt with auxiliary prompt tasks—Sense Semantics Classification and Annotated Connective Prediction—and with RoBERTa reported 75.51 Acc / 72.26 F1 on PDTB 3.0, surpassing ConnPrompt under the same conditions (Xiang et al., 2023). AdaptPrompt replaced manually designed discrete prompts and answer words with continuous prompt learning, virtual answer tokens, and connective knowledge distillation, reporting 76.12 Acc / 71.79 F1 with RoBERTa on PDTB v3.0 top-level classification (Wang et al., 2023). PLSE used explicit discourse pairs mined from Gigaword for prompt-based connective pre-training and a mutual-information objective for Global Logical Semantics Learning, reporting 71.40 F1 / 75.43 Acc on PDTB 2.0 top-level and 50.11 F1 / 64.00 Acc on second-level classification (Wang et al., 2023). A later prompt-based line replaced manual verbalizers with learned hierarchical prototypes, using instance-to-instance, instance-to-prototype, and prototype-to-prototype losses to structure the verbalizer space (Long et al., 2024).

A third development is the move from single-label classification to distributional prediction. A RoBERTa-based multi-task model trained on DiscoGeM predicts sense distributions for all three PDTB levels, using Mean Absolute Error (MAE) for multi-label learning and Cross-Entropy (CE) for single-label learning; on DiscoGeM it reported 0.299 ± 0.002, 0.446 ± 0.003, and 0.523 ± 0.002 JS distance for levels 1–3, and 65.89 ± 1.35, 55.99 ± 1.73, and 50.82 ± 1.26 weighted F1 for single-label prediction (Costa et al., 2024). HArch extends this by predicting distributions jointly across levels and feeding lower-level predictions into higher-level classification heads, arguing that hierarchical conditioning helps especially at finer granularity (Costa et al., 28 Aug 2025).

4. Ambiguity, disagreement, and evaluation

Ambiguity is not an edge case in IDRR; it is a central empirical property. Cross-domain synthetic-data work emphasizes that real implicit relations are often subtle and underspecified, and that in real text the relation can be ambiguous or underdetermined (Yung et al., 26 Mar 2025). Multilingual multi-label work argues that the single-label formulation is inadequate because annotators often disagree and because a single relation may legitimately express multiple senses (Costa et al., 28 Aug 2025).

This view is borne out quantitatively in human-label studies. On DiscoGeM 1.5, Level 3 shows an average of 6.3 distinct labels per instance, an agreement rate of 0.404, and entropy of 0.86. A manual analysis of 100 samples categorized disagreement as 35% agreement, 31% inherent ambiguity, 5% unidentified cues, 10% sentence complexity, 11% annotation error, and 8% task design (Yung et al., 26 Feb 2026). Under these conditions, models trained to predict full label distributions are reported to yield more stable predictions than annotator-specific models, especially at finer granularity (Yung et al., 26 Feb 2026).

Fine-grained diagnosis on PDTB 3.0 reinforces the same conclusion. RoBERTa-large reported 53.07 macro-F1 on level-2 classification, outperforming BERT-large (43.82) and ALBERT-xlarge (33.58), yet the analysis showed that some senses remain difficult even with ample data. Exp.Level-of-detail, for example, has 3,361 instances but only about 45–58 F1 depending on the model, while very small classes such as Comp.Similarity (31 instances) and Exp.Disjunction (30 instances) are hard largely because of data sparsity (Cai, 7 Mar 2025). The same study found strong reliance on surface cues such as to-infinitives, negation, same subject, and similar syntactic pattern, and documented hierarchy inconsistency in which a model is often correct at level 1 but wrong at level 2 (Cai, 7 Mar 2025).

Evaluation itself is unstable. Cross-domain work explicitly warns that statistical significance matters, that macro-F1 may be misleading for rare classes, that different treatments of multi-label instances reduce comparability, and that cross-dataset annotation mismatch—expert PDTB annotation versus crowdsourced DiscoGeM annotation—can hinder adaptation (Yung et al., 26 Mar 2025). This makes direct comparison of headline numbers across PDTB, CoNLL, DiscoGeM, dialogue, and multimodal datasets methodologically fragile.

5. Domain shift, multilinguality, and multimodality

IDRR generalization across domains is difficult. A direct test of LLM-based synthetic augmentation for cross-domain IDRR trained a base RoBERTa classifier on PDTB 3.0 and generated synthetic target-domain samples from unlabeled raw texts in the DiscoGeM 1.5 subcorpora EP, WK, and NV. The headline result was negative: none of the synthetic-data variants consistently outperform the PDTB baseline, and the one consistent finding was that confusion-screen and combi-screen variants performed significantly worse than the baseline (Yung et al., 26 Mar 2025). The authors concluded that LLM-generated discourse continuations were not useful for cross-domain IDRR in that setup and that there was no clear advantage over the simpler pseudo-labeling baseline (Yung et al., 26 Mar 2025).

Dialogue provides a different form of domain shift. Edina-DR treats adjacent utterances or turns as arguments and classifies only the four top-level PDTB relations. On this corpus, DER (PDTB) dropped to Accuracy 0.61, F1 0.51, showing a substantial mismatch between formal text and dialogue, whereas DER (Edina-DR) reached Accuracy 0.80, F1 0.76, and DER+Dialogue (Edina-DR) improved further to Accuracy 0.81, F1 0.77 with p = 0.006 (Ma et al., 2019). The result is not merely that dialogue is another benchmark; it is that argument structure, feature inventories, and relation distributions change materially in conversational settings.

Multilingual IDRR introduces both data scarcity and hierarchy transfer. HArch is presented as the first multilingual and multi-label model for IDRR and is evaluated on DiscoGeM 2.0, a multilingual parallel corpus with 13,063 implicit discourse relations across English (5,847), German (2,588), French (2,628), and Czech (2,000) (Costa et al., 28 Aug 2025). In the multilingual setting, XLM-RoBERTa-HArch performed best among HArch encoder variants and consistently outperformed few-shot prompted GPT-4o and Llama-4-Maverick (Costa et al., 28 Aug 2025). Zero-shot cross-lingual work with hierarchical prototypes likewise aligned source and target prototypes using contrastive learning and reported improvements over both vanilla fine-tuning and earlier zero-shot transfer baselines across six languages in TED-MDB (Long et al., 2024).

Speech adds another axis of variation. MM-IDR argues that coherence is inherently multimodal and proposes a Qwen2-Audio-based classifier integrating text, audio, prosody, and audio pooling statistics for English, French, Spanish, and Arabic. Across languages, text-only models outperform audio-only models, but multimodal fusion sometimes helps, especially when text supervision is limited or multilingual transfer is used (Ruby et al., 4 Feb 2026). The reported pattern is that audio alone is weak, while audio/prosody can provide gains in selected settings rather than a universal improvement (Ruby et al., 4 Feb 2026).

6. Interpretability, explanation, and current directions

A longstanding limitation of IDRR systems is that they usually output only a label. EIDRR addresses this by prompting an LLM to generate explanations for each training instance conditioned on its gold label and then training a joint classification-generation model with a shared RoBERTa encoder and T5 decoder. On PDTB 2.0, the full model reported 73.14% Accuracy and 68.01% Macro-F1, improving over EIDRR w/o Exp at 71.80 / 66.87 (Wang et al., 25 Feb 2026). Human evaluation compared EIDRR with e-INFERSENT and reported average explanation scores of 4.20 versus 1.71, with 93% of EIDRR explanations scoring 3 or higher (Wang et al., 25 Feb 2026). In this line of work, explanation is treated as an additional supervisory signal rather than merely post hoc text generation.

Across papers, LLMs play sharply differentiated roles. Few-shot prompting with GPT-4o and Llama-4-Maverick is outperformed by task-specific fine-tuning in multilingual multi-label IDRR (Costa et al., 28 Aug 2025), and LLM-generated discourse continuations do not provide reliable gains for cross-domain IDRR (Yung et al., 26 Mar 2025). By contrast, LLM-generated natural language explanations can improve both performance and interpretability when distilled into a smaller classifier-generator (Wang et al., 25 Feb 2026). This suggests that, within current IDRR pipelines, LLMs are more effective as sources of supervision than as direct relation predictors or synthetic instance generators.

Future directions are increasingly defined by this mixture of representational, annotational, and evaluative constraints. The survey emphasizes interaction-boosted representation learning, synthetic implicit corpus refinement, and joint relation recognition and discourse parsing (Xiang et al., 2022). Fine-grained evaluation argues for targeted augmentation of truly low-resource senses and for stronger use of hierarchy (Cai, 7 Mar 2025). Cross-domain work suggests future generation strategies should produce more ambiguous examples and notes that pseudo-labeling may be more promising because it preserves real-text ambiguity (Yung et al., 26 Mar 2025). Multimodal work proposes incorporating the visual modality and better-controlled fusion across languages and relation types (Ruby et al., 4 Feb 2026). Taken together, these directions indicate that IDRR is no longer defined solely as a sentence-pair classification problem; it is increasingly treated as a structured inference problem shaped by hierarchy, ambiguity, heterogeneous annotation, and the limits of available supervision.