When users of a task-oriented dialogue system say "book the same one for Tuesday" or "what about that clinic?", the system's natural-language-understanding component must map the utterance to a complete logical-form (LF) meaning representation in which the anaphoric expressions are bound to the intended discourse referents. The source paper (Cohen, Galescu, and Shvo, 2023, the Eva collaborative dialogue system) states that its neural semantic parser does not incorporate anaphoric expressions and that finding and resolving coreference candidates is research it has not undertaken. Prior work shows that meaning-representation-level coreference resolution exists in restricted settings (SQL over temporally structured data, conversational knowledge-base question answering, formal travel commands), that much of the field sidesteps the problem via utterance rewriting or coreference preprocessing, and that work through 2026 still reports poor cross-domain generalization of coreference resolution in task-based dialogue. The open problem is therefore not whether any system integrates coreference at the LF level but whether such integration can be made robust: a neural semantic-parsing NLU component that resolves anaphora and coreference inside expressive dialogue LFs (encoding speech acts — the communicative acts an utterance performs, such as requesting or informing — embedded beliefs and goals, and modal operators), across anaphora types and domains (and, ideally, languages), with robustness demonstrated empirically rather than assumed, in comparison with the pipeline and rewriting alternatives that dominate current practice.

Background

Task-oriented dialogue systems must map each user utterance to a machine-readable meaning. The prevalent industrial scheme, intent-plus-slot filling, is too weak to express compositional content, so research systems map utterances to logical forms (LFs): formal expressions that encode both the utterance's propositional content and, in expressive frameworks such as the Eva system studied here, the speech act performed (request, inform, question) and operators over beliefs, goals, and intentions of several dialogue participants. Semantic parsing is the task of mapping natural language to such formal representations with neural models. Anaphora denotes linguistic expressions (e.g., pronouns, demonstratives, definite descriptions like "that clinic") whose interpretation is borrowed from an antecedent elsewhere in the discourse; coreference resolution is the task of deciding which discourse entity such an expression refers to. In multi-turn dialogue every other utterance contains such expressions, so an NLU component that cannot resolve them produces incomplete or wrong LFs.

The problem is a recognized subfield, context-dependent semantic parsing (CDSP): mapping an utterance to a meaning representation (MR) using dialogue context. Zettlemoyer and Collins (2009) learned context-dependent sentence-to-logical-form mappings; Suhr, Iyer, and Artzi (2018) mapped context-dependent sentences to executable formal queries over a flight-booking domain, encoding contextual reuse by copying from previous LFs. Two systems resolve coreference explicitly inside the representation: Chen and Bunescu (2019) generate MR templates containing special REF tokens that a pointer network (a decoder that emits positions of its input sequence rather than fixed vocabulary words) replaces with entities copied from the previous MR, for queries over temporally structured data, and Shen et al. (2019) built a pointer-equipped conversational knowledge-base question-answering parser designed to resolve coreference in conversations. The survey of Li, Qu, and Haffari (2020) maps this landscape and lists open problems, including which contextual information parsers should draw on and phenomena beyond ellipsis and coreference. For task-oriented dialogue specifically, the Dialogue Meaning Representation (DMR) of Hu et al. (2022) defines a graph representation that can express coreference across utterances and releases a 70k-utterance annotated corpus, but representation design is distinct from a parser that robustly resolves the references. In the era of large language models (LLMs), Lin, Xia, and Fang (2023) compared four paradigms for contextual utterances (parse with utterance history, parse with a reference program, parse then resolve, rewrite then parse) and found rewriting the utterance before parsing to be the most promising paradigm — evidence that LF-level resolution is being avoided rather than solved; Jain and Lapata (2023) improved conversational parsing with dynamic context graphs; a 2025 hybrid pipeline (Papakostas et al.) runs rule-based coreference resolution as preprocessing before AMR (abstract meaning representation) graph parsing, a pipeline rather than LF-integrated design; and 2026 work (Ijurco and de Lacalle; ReacTOD) still reports that coreference resolution in task-based dialogue generalizes poorly across domains, with symbolic validators enforcing only coreference consistency of dialogue-state updates.

The source is Eva (Cohen, Galescu, and Shvo, arXiv:2302.09646; v1 2023 titled "A Planning-Based Explainable Collaborative Dialogue System", v3 2024 retitled "An Explainable Collaborative Dialogue System using a Theory of Mind"), a planning-based collaborative dialogue system whose LFs are formulas of a modal logic of belief, goals, and intentions, with speech acts treated as planned actions. Its NLU is a neural sequence-to-sequence semantic parser (a fine-tuned CodeT5 encoder, CodeT5 being a pre-trained code-oriented language model) trained Overnight-style, i.e., on canonical utterances paired with LFs that are then paraphrased into natural language. The paper's Limitations section lists "semantic parsing of natural language that incorporates anaphoric expressions" among the components needing improvement, and the equality-reasoning section notes that although Eva's declarative equality mechanism could represent and resolve co-referential and anaphoric references — including asking the user to disambiguate — how to find coreference candidates and resolve them "is a topic of much research, but not one we have concentrated on to date". Eva's LFs are intensional (equality reasoning may not cross belief/goal operators), which makes referent binding inside the LF semantically delicate — a referent can be shared between agents' goals without the agents sharing beliefs about identity.

Problem Statement

Build and validate a neural semantic-parsing natural-language-understanding component that resolves anaphoric expressions and coreference within the logical-form (meaning-representation) level itself: given a dialogue context and an utterance containing anaphoric or coreferential expressions, the parser must output a complete logical form in which each such expression is bound to the intended discourse referent from context, or, only when the referent is not uniquely determined from context, to an explicit request for disambiguation, rather than delegating reference resolution to a separate coreference preprocessor or to an utterance-rewriting front end. The target setting is expressive task-oriented dialogue, where logical forms encode speech acts and operators over agents' beliefs, goals, and intentions, so referent binding must respect the LF's own semantics (e.g., equality that cannot cross intensional operators). The objective is robustness: coverage of the required anaphora types — pronominal, demonstrative, definite-description, bridging (reference to an entity associated with, but not identical to, a previously mentioned one), one-anaphora such as "the same one", and event reference — as well as generalization across dialogue domains (with additional language coverage desirable but not required), and degradation at longer dialogue distances no worse than at short ones. The claim of robust LF-level resolution should be empirically supported rather than assumed; relevant evidence would include LF exactness and reference-resolution accuracy on context-dependent parsing benchmarks with LF-level reference annotations, and comparison under matched conditions with pipeline (coreference resolution before parsing) and rewrite-then-parse designs, the dominant alternatives in current practice.

References

  1. Philip R. Cohen, Lucian Galescu, Maayan Shvo. A Planning-Based Explainable Collaborative Dialogue System (arXiv:2302.09646v1, 2023; v3 2024 retitled An Explainable Collaborative Dialogue System using a Theory of Mind). DOI: 10.48550/arXiv.2302.09646. https://arxiv.org/abs/2302.09646
  2. Zhuang Li, Lizhen Qu, Gholamreza Haffari. Context Dependent Semantic Parsing: A Survey. COLING 2020, pages 2509-2521. DOI: 10.18653/v1/2020.coling-main.226 (arXiv:2011.00797). https://aclanthology.org/2020.coling-main.226/
  3. Charles Chen, Jr., Razvan Bunescu. Context Dependent Semantic Parsing over Temporally Structured Data. NAACL-HLT 2019, pages 3576-3585. DOI: 10.18653/v1/N19-1360. https://aclanthology.org/N19-1360/
  4. Alane Suhr, Srinivasan Iyer, Yoav Artzi. Learning to Map Context-Dependent Sentences to Executable Formal Queries. NAACL-HLT 2018. DOI: 10.18653/v1/N18-1203. https://doi.org/10.18653/v1/n18-1203
  5. Luke Zettlemoyer, Michael Collins. Learning Context-Dependent Mappings from Sentences to Logical Form. ACL-IJCNLP 2009, pages 976-984. https://aclanthology.org/P09-1110/
  6. Tao Shen, Xiubo Geng, Tao Qin, Daya Guo, Duyu Tang, Nan Duan, Guodong Long, Daxin Jiang. Multi-Task Learning for Conversational Question Answering over a Large-Scale Knowledge Base. arXiv:1910.05069, 2019. DOI: 10.48550/arXiv.1910.05069. https://arxiv.org/abs/1910.05069
  7. Xiangkun Hu, Junqi Dai, Hang Yan, Yi Zhang, Qipeng Guo, Xipeng Qiu, Zheng Zhang. Dialogue Meaning Representation for Task-Oriented Dialogue Systems. arXiv:2204.10989, 2022. DOI: 10.48550/arXiv.2204.10989. https://arxiv.org/abs/2204.10989
  8. Parag Jain, Mirella Lapata. Conversational Semantic Parsing using Dynamic Context Graphs. arXiv:2305.06164, 2023. DOI: 10.48550/arXiv.2305.06164. https://arxiv.org/abs/2305.06164
  9. Kevin Lin, Patrick Xia, Hao Fang. Few-Shot Adaptation for Parsing Contextual Utterances with LLMs. arXiv:2309.10168, 2023. DOI: 10.48550/arXiv.2309.10168. https://arxiv.org/abs/2309.10168
  10. Christos Papakostas, Christos Troussas, Akrivi Krouska, Cleo Sgouropoulou. A Hybrid Neuro-Symbolic Pipeline for Coreference Resolution and AMR-Based Semantic Parsing. Information 16(7):529, 2025. DOI: 10.3390/info16070529. https://doi.org/10.3390/info16070529
  11. Oier Ijurco, Oier Lopez de Lacalle. Reasoning over Object Descriptions Improves Coreference Resolution in Task-Based Dialogue Systems. arXiv:2604.27850, 2026. DOI: 10.48550/arXiv.2604.27850. https://arxiv.org/abs/2604.27850
  12. Yanjun Lin, Zimo Xiao, Kartik Natarajan, Mahesh Sankaranarayanan, Niraj Nawanit, Rakshit Parashar, Austin Zhang, Karthik Konaraddi, Rishita Mote, Wei Niu. ReacTOD: Bounded Neuro-Symbolic Agentic NLU for Zero-Shot Dialogue State Tracking. arXiv:2605.19077, 2026. DOI: 10.48550/arXiv.2605.19077. https://arxiv.org/abs/2605.19077
  13. Sagar Sapkota, Mohammad Saqib Hasan, Mubarak Shah, Santu Karmaker. Multi-Party Conversational Agents: A Survey. arXiv:2505.18845, 2025. DOI: 10.48550/arXiv.2505.18845. https://arxiv.org/abs/2505.18845
  14. Rikunari Sagara, Koichiro Terao, Naoto Iwahashi. BeliefNest: A Joint Action Simulator for Embodied Agents with Theory of Mind. arXiv:2505.12321, 2025. DOI: 10.48550/arXiv.2505.12321. https://arxiv.org/abs/2505.12321
  15. Amogh Mannekote. Towards a Neural Era in Dialogue Management for Collaboration: A Literature Survey. arXiv:2307.09021, 2023. DOI: 10.48550/arXiv.2307.09021. https://arxiv.org/abs/2307.09021