ANAPHORA RESOLUTION: THE STATE OF THE ART

Paper
LLM Failure ModesNLP and Linguistics

The paper is an introduction to anaphora resolution offering a brief survey of the major works in the field.

Introduction. Anaphora resolution is a complicated problem in Natural Language Processing and has attracted the attention of many researchers. The approaches developed - traditional (from purely syntactic ones to highly semantic and pragmatic ones), alternative (statistic, uncertainty-reasoning etc.) or knowledge-poor, offer only approximate solutions. The etymology of the term "anaphora" goes back to Ancient Greek with “anaphora” (αναφορα) being a compound word consisting of the separate words ανα − back, upstream, back in an upward direction and φορα - the act of carrying and denoted the act of carrying back upstream. For Computational Linguists embarking upon research in the field of anaphor resolution, I strongly recommend as a primer Graham Hirst's book "Anaphora in natural language understanding" (Hirst 1981) which may seem a bit dated in that it does not include developments in the 80's and the 90's, but which provides an excellent survey of the theoretical work on anaphora and of the early computational approaches and is still very useful reading.

Discussion / Conclusion. Against the background of growing interest in the field, it seems that insufficient attention has been paid to the evaluation of the systems developed. Even though the number of works reporting extensively on evaluation in anaphora resolution is increasing (Aone & Bennet 1996; Azzam et al. 1998; Baldwin 1997; Gaizauskas & Humphreys 1996; Lappin & Leass 1994, Mitkov & Stys 1997, Mitkov et al. 1998), the forms of evaluation that have been proposed are not sufficient or perspicuous. It is felt, however, that evaluation in anaphora resolution needs further attention. Measuring the success rate of an anaphora resolution system in terms of "recall" and "precision" is undoubtedly an important (and consistent) step in assessing the efficiency of anaphora resolution approaches, but as we have already pointed out, they cannot be seen as distinct measures for robust systems. In addition, it appears that they alone 10The above definitions refer to the resolution process only and not to the process of identification of anaphors which of course will have "recall" and "process" defined differently. Automatic referential links in corpora is a highly attractive research task that will definitely need further attention in the future.

Synthesis notes from this paper's topics 8

This paper has no direct semantic matches yet — here are notes from the topic(s) it's filed under.

Lines of inquiry this paper opens 24

Research framings built by reading the notes related to this paper — the questions it feeds into.

Do single-axis benchmarks adequately measure deployment-ready agent capability? Why doesn't LLM knowledge translate to improved reasoning outputs? How does explicit structure improve LLM reasoning on complex tasks? Should agent evaluation measure execution quality beyond task success? What makes reasoning models struggle with ill-posed and unfamiliar problems? Do base models contain latent reasoning capabilities that training can activate? Why does separating planning from execution enhance reasoning performance? When do additional thinking tokens stop improving reasoning performance? Do language models truly understand their encoded knowledge or merely retrieve patterns? What determines whether visible reasoning traces improve task performance? How do pretraining data composition and task diversity shape reasoning? Do reasoning traces accurately represent what models actually compute? Does semantic diversity improve reasoning quality in reinforcement learning? What determines how neural networks organize their internal knowledge representations? Why does self-revision increase model confidence despite degraded accuracy? How does policy entropy collapse limit reasoning RL scaling? How does silent agreement emerge and undermine deliberation in multi-agent systems? Can scaling alone produce compositional generalization in neural networks?