Learning to Tell Tales: A Data-driven Approach to Story Generation
McIntyre, Neil and Lapata, Mirella

Article Structure

Abstract

Computational story telling has sparked great interest in artificial intelligence, partly because of its relevance to educational and gaming applications.

Introduction

Recent years have witnessed increased interest in the use of interactive language technology in educational and entertainment applications.

The Story Generator

As common in previous work (e. g., Shim and Kim 2002), we assume that our generator operates in an interactive context.

Story Ranking

We have so far described most modules of our story generator, save one important component, namely the story ranker.

Experimental Setup

In this section we present our experimental setup for assessing the performance of our story generator.

Results

Our results are summarized in Table 3 which lists the average human ratings for the three systems.

Conclusions and Future Work

In this paper we proposed a novel method to computational story telling.

Topics

knowledge base

Appears in 14 sentences as: knowledge base (14)
In Learning to Tell Tales: A Data-driven Approach to Story Generation
  1. The generator next constructs several possible stories involving these entities by consulting a knowledge base containing information about dogs and ducks (e. g., dogs bark, ducks swim) and their interactions (e.g., dogs chase ducks, ducks love dogs).
    Page 2, “The Story Generator”
  2. Although we are ultimately searching for the best overall story at the document level, we must also find the most suitable sentences that can be generated from the knowledge base (see Figure 4).
    Page 3, “The Story Generator”
  3. The space of possible stories can increase dramatically depending on the size of the knowledge base so that an exhaustive tree search becomes computationally prohibitive.
    Page 3, “The Story Generator”
  4. As mentioned earlier our generator has access to a knowledge base recording entities and their interactions.
    Page 3, “The Story Generator”
  5. In our experiments this knowledge base was created using the RASP relational parser (Briscoe and Carroll, 2002).
    Page 3, “The Story Generator”
  6. The knowledge base described above can only inform the generation system about relationships on the sentence level.
    Page 3, “The Story Generator”
  7. Our sentence planner aggregates together information from the knowledge base , without however generating referring expressions.
    Page 4, “The Story Generator”
  8. Next, we select verbs from the knowledge base that take the words duck and dog as their subject (e.g., bark, run, fly).
    Page 5, “The Story Generator”
  9. We select an object for bark, by retrieving from the knowledge base the set of objects it co-occurs with.
    Page 5, “The Story Generator”
  10. In default of generating all of these exhaustively, our system utilizes the MI scores from the knowledge base to guide the
    Page 5, “The Story Generator”
  11. Instead, it decides deterministically how to generate a story on the basis of the most likely predicate-argument and predicate-predicate counts in the knowledge base .
    Page 7, “Experimental Setup”

See all papers in Proc. ACL 2009 that mention knowledge base.

See all papers in Proc. ACL that mention knowledge base.

Back to top.

highest score

Appears in 6 sentences as: highest score (3) highest scoring (3)
In Learning to Tell Tales: A Data-driven Approach to Story Generation
  1. Story generation amounts to traversing the tree and selecting the nodes with the highest score
    Page 3, “The Story Generator”
  2. Once we reach the required length, the highest scoring story is presented to the user.
    Page 3, “The Story Generator”
  3. To evaluate which system configuration was best, we asked two human evaluators to rate (on a 1—5 scale) stories produced in the following conditions: (a) score the candidate stories using the interest function first and then coherence (and vice versa), (b) score the stories simultaneously using both rankers and select the story with the highest score .
    Page 7, “Experimental Setup”
  4. We also examined how best to prune the search space, i.e., by selecting the highest scoring stories, the lowest scoring one, or simply at random.
    Page 7, “Experimental Setup”
  5. The results showed that the evaluators preferred the version of the system that applied both rankers simultaneously and maintained the highest scoring stories in the beam.
    Page 7, “Experimental Setup”
  6. Story creation amounts to traversing the tree and selecting the nodes with the highest score .
    Page 8, “Conclusions and Future Work”

See all papers in Proc. ACL 2009 that mention highest score.

See all papers in Proc. ACL that mention highest score.

Back to top.

co-occurrence

Appears in 4 sentences as: co-occurrence (4)
In Learning to Tell Tales: A Data-driven Approach to Story Generation
  1. Our generator operates over predicate-argument and predicate-predicate co-occurrence statistics gathered from corpora.
    Page 2, “Introduction”
  2. A fragment of the action graph is shown in Figure 3 (for simplicity, the edges in the example are weighted with co-occurrence frequencies).
    Page 4, “The Story Generator”
  3. As explained earlier, our generator produces stories stochastically, by relying on co-occurrence frequencies collected from the training corpus.
    Page 5, “Story Ranking”
  4. The second one creates a story randomly without taking any co-occurrence frequency into account.
    Page 7, “Experimental Setup”

See all papers in Proc. ACL 2009 that mention co-occurrence.

See all papers in Proc. ACL that mention co-occurrence.

Back to top.

language model

Appears in 4 sentences as: language model (4)
In Learning to Tell Tales: A Data-driven Approach to Story Generation
  1. And Knight and Hatzivassiloglou (1995) use a language model for selecting a fluent sentence among the vast number of surface realizations corresponding to a single semantic representation.
    Page 2, “Introduction”
  2. The top-ranked candidate is selected for presentation and verbalized using a language model interfaced with RealPro (Lavoie and Rambow, 1997), a text generation engine.
    Page 2, “Introduction”
  3. Since we do not know a priori which of these parameters will result in a grammatical sentence, we generate all possible combinations and select the most likely one according to a language model .
    Page 4, “The Story Generator”
  4. We used the SRI toolkit to train a trigram language model on the British National Corpus, with interpolated Kneser—Ney smoothing and perplexity as the scoring metric for the generated sentences.
    Page 4, “The Story Generator”

See all papers in Proc. ACL 2009 that mention language model.

See all papers in Proc. ACL that mention language model.

Back to top.

generation process

Appears in 3 sentences as: generation process (3)
In Learning to Tell Tales: A Data-driven Approach to Story Generation
  1. Although they have limited vocabulary and non-elaborate syntax, they nevertheless present challenges at almost all stages of the generation process .
    Page 2, “Introduction”
  2. the story generation process as a tree (see Figure 2) whose levels represent different story lengths.
    Page 3, “The Story Generator”
  3. So, at each choice point in the generation process , e. g., when selecting a verb for an entity or a frame for a verb, we consider the N best alternatives assuming that these are most likely to appear in a good story.
    Page 5, “The Story Generator”

See all papers in Proc. ACL 2009 that mention generation process.

See all papers in Proc. ACL that mention generation process.

Back to top.

PDF available from www.aclweb.org.
Send us an email