Automatic Story Generation: Challenges and Attempts

1 Introduction and Motivation

• Earlier surveys covered creativity, interactive intelligence, narrative structures, character building, and machine-learning challenges, but often did not examine those challenges in depth.The survey positions itself as complementary to prior work rather than replacing these specialized reviews.Source: §1 · pp. 1–2
• Automatic storytelling remains difficult because systems must generate logically coherent and interesting narratives, despite recent advances in computational narrative.The paper focuses specifically on story-generation capabilities within computational narrative.Source: §1 · p. 1
• Transformer-based language models motivate an updated survey because many earlier systems relied on planning or predated large-scale neural language models.The paper specifically contrasts newer transformer-driven research with pre-2013 interactive-narrative work and earlier planning-based systems.Source: §1 · p. 1
• The survey addresses a gap in up-to-date coverage by organizing prior work around major story-generation problems and techniques used to solve them.Earlier surveys often focused on particular approaches or sub-problems, while this survey aims to provide a starting point for new researchers.Source: §1 · pp. 1–2
• The survey treats content encoding and comprehension, controllability, and commonsense as foundational concerns for generating on-topic stories with consistent entities and events.Neural models such as GPT-2 generate stories from web-trained language models, making topic maintenance and consistency central challenges.Source: §1 · p. 2

2.1 Controllability in Story Generation

• Controlled generation improves coherence or goal achievement in several systems, but generated stories still show repetition, off-topic content, logical inconsistencies, or limited commonsense inference.Plot Machines outperforms several baselines on human-rated ordering, flow, and repetitiveness, yet its general outlines may limit depth and interestingness.Source: §2.1 · pp. 3–4
• Controllability methods target either desired endings or storylines, with newer transformer-based systems replacing earlier Seq2Seq approaches.Ending-focused examples include Tambwekar et al. (2019) and Wang et al. (2020), while storyline-focused examples include Fan et al. (2018), Yao et al. (2019), and Rashkin et al. (2020).Source: §2.1 · pp. 2, 4
• The main controllability challenges are reduced creativity and interestingness, insufficiently unified evaluation metrics, and unresolved coherent and logical plot generation.The survey notes a trade-off between stronger structural control and narrative creativity, while BLEU, perplexity, and ROUGE are weak for story-specific evaluation.Source: §2.1 · pp. 2, 4

2.2.1 Benchmarks

• Commonsense resources cover general knowledge, causal event relations, contextual explanations, and social reasoning through datasets such as ConceptNet, CommonsenseQA, ATOMIC, GLUCOSE, and SocialIQA.The benchmarks represent knowledge as semantic triples, question-answering tasks, event relations, causal explanations, or social-situation inferences.Source: §2.2.1 · pp. 4–5

2.2.2 Frameworks

• Commonsense integration uses benchmark fine-tuning, causal or quality inference, and knowledge-graph features in model encodings.These approaches include transformed ConceptNet and ATOMIC data, COMeT-style inference, discourse-aware PARA-COMeT, causal plot graphs, and graph-attention knowledge contexts.Source: §2.2.2 · pp. 5–6
• Commonsense-enhanced models can improve logicality, grammaticality, local or global coherence, and selected generation metrics, but no approach guarantees reasonable coherent stories.Sentence-level inference remains limited because story generation often requires paragraph-level context; PARA-COMeT addresses this with discourse-aware and recurrent-memory models.Source: §2.2.2 · pp. 5–6

2.3.1 Characters and Entities

• Entity representations and character affect can improve selected narrative tasks, but multi-character interaction remains underdeveloped.ENGEN improved automated tasks and was preferred on 27 of 50 prompts, while EC-CLF improved emotion faithfulness and content quality but focused on one protagonist.Source: §2.3.1 · pp. 6–7
• Character-affinity, action-plan, and entity-based methods model narrative agents or actions, yet their effects on full plot development and implicit interaction learning remain unclear.The character-affinity model reports affinity evolution rather than generated stories, and SRL + Entity does not establish whether interactions are learned for multi-agent narratives.Source: §2.3.1 · p. 7

Controllability

• Controllability approaches generally target either the story outline or the ending, with transformer-based language models becoming an emerging trend.Source: §3 · p. 8

Commonsense Knowledge

• Combining commonsense integration methods may substantially improve story generation, although no current method ensures reasonable and coherent stories.Approaches include fine-tuning on commonsense datasets, analyzing causality, and incorporating commonsense knowledge graphs into encoders.Source: §3 · p. 8

Characters

• Entity representations can improve plot coherence, while character affect can help navigate generation; extending affect modeling to multiple characters may enhance narratives.Source: §3 · p. 8

Creativity

• Creativity remains relatively underemphasized in generated-text research despite being identified as a less-researched area.Source: §3 · p. 8

Future Research

• Future work should address the creativity–structural-coherence trade-off, weak story evaluation metrics, unresolved plot incoherence, and story-level commonsense inference.The survey specifically calls for metrics capturing creativity and interestingness, more robust unified evaluation, and improved logical coherence.Source: §3 · p. 8

• Table 1 presents a summary of controllability approaches.Source: Table 1 · p. 13
• The table belongs to the appendix section on controllability approaches.Source: Table 1 · p. 13
• The supplied table passage does not state comparative outcomes among the summarized approaches.Source: Table 1 · p. 13