Changing the narrative perspective: From deictic to anaphoric point of view

https://doi.org/10.1016/j.ipm.2021.102559Get rights and content

Highlights

  • Introduce the new task of changing the narrative perspectives of a text.

  • Develop an end-to-end system for deictic to anaphoric point of view (PoV) change.

  • Design a neural architecture for mention selection, trained on coreference data.

  • Introduce a dataset with different types of narratives annotated for PoV change.

  • Evaluations show that the output text is generally fluent and non-ambiguous.

Abstract

We introduce the task of changing the narrative point of view, where characters are assigned a narrative perspective that is different from the one originally used by the writer. The resulting shift in the narrative point of view alters the reading experience and can be used as a tool in fiction writing or to generate types of text ranging from educational to self-help and self-diagnosis. We introduce a benchmark dataset containing a wide range of types of narratives annotated with changes in point of view from deictic (first or second person) to anaphoric (third person) and describe a pipeline for processing raw text that relies on a neural architecture for mention selection. Evaluations on the new benchmark dataset show that the proposed architecture substantially outperforms the baselines by generating mentions that are less ambiguous and more natural.

Section snippets

Introduction and motivation

The narrative point of view (PoV), also called the narrative perspective, is the position from which the events in a story are observed and communicated (van Peer & Chatman, 2001). There are three types of narrative perspective, mirroring the use of personal pronouns: first, second, and third person. The narrative point of view influences the readers’ perspective-taking and their involvement in the characters’ thoughts and feelings. In fiction, the most common narrative perspectives are the

Related work

Wiebe (1994) addressed the problem of tracking the psychological point of view in third-person fictional narrative text, which is about recognizing sentences that convey the thoughts, perceptions, and inner states of characters and mapping these psychological perspectives to the corresponding character in the story. Mani (2013) provided a systematic overview of computational methods applied to the processing of core narratological concepts, such as narrative perspective, narrative threads, and

Task definition and important aspects

We consider as input a text document in a prototypical PoV setting where at most one discourse entity has the 1st person PoV, at most one entity has the 2nd person PoV, and any number of entities have the 3rd person PoV.1 Given a focus entity E that has PoV p, the

The PoV change system pipeline

To change the narrative perspective from 1st/2nd to 3rd PoV, the raw text is processed using the pipeline of 4 steps shown below. As a running example, we will use the text below, where the focus entity (to be indexed with 2, and later named Nick) has the 1st person PoV (as expressed through the pronoun I) and the task is to convert it to a 3rd person PoV:

  • “Phil returns to Boston, to his job. I drive to the city every other week, to work a night or two at Pine Street, to see Emily”. (Flynn, 2005)

A neural mention selection model

Given a mention position t in the text for an entity E, the mention selection task is to select from the set S(E) of possible strings for that entity the best string (noun phrase or pronoun) to use as a mention Mt in that context. Using the same example as in the previous section, where the focus entity E is the one indexed with 2, and assuming that it was already mentioned 5 times before the sentences in this example, then the current mention position would be 6, as shown below:

  • [Phil]1 returns

Datasets for PoV change in English

Training deep learning models, such as the neural mention selection architecture that we propose for this task, requires a large number of training examples. However, manually annotating the corresponding large number of documents with PoV changes is prohibitive. Instead, we propose to use existing coreference resolution corpora to create training examples for the mention selection part of the PoV change task, as described in Section 6.1 below. Separately, we also create a benchmark dataset

Experimental evaluation

The training portion of the CoNLL-2012 corpus is used to train two LSTM-based models: a model that uses only the token-level LSTMs, and a model that uses both the token-level and mention-level LSTMs. The size of the hidden state of LSTM cell is set to 50, whereas the size of the hidden layer of the fully connected network that computes the ranking score is set to 100. We use N=50 words and K=10 mentions for both the left and right contexts. If there are less than N words or K mentions

Conclusion and future work

We introduced the task of changing the narrative perspectives and proposed a pipeline architecture where a mention selection model is trained on data extracted from a corpus annotated with coreference chains. We annotated a benchmark dataset with point of view changes, using text from a wide range of types of narratives. Evaluations on a newly created PoV benchmark dataset showed that the proposed architecture substantially outperforms the baselines and comes close to the manual annotation in

Acknowledgments

We would like to thank Dana Simionescu for the annotation of the PoV change dataset and Edmond Chang for providing connections to research on the use of point of view in literature and virtual worlds. In its initial stages, the project benefited from numerous discussions with Pooya Taherkhani. Finally, we would like to thank the anonymous reviewers for their detailed and constructive feedback.

References (79)

  • JorgeA.M. et al.

    Information processing & management journal special issue on narrative extraction from texts (text2story): Preface

    Information Processing & Management

    (2019)
  • AsimovI.

    Nobody here but –

    (1953)
  • AtwoodM.

    The handmaid’s tale

    (1998)
  • BelzA. et al.

    Generating referring expressions in context: The GREC task evaluation challenges

  • BissE.

    Notes from no man’s land

    (2009)
  • BlackD.

    Why can i see my avatar? Embodied visual engagement in the third-person video game

    Games and Culture

    (2017)
  • BohmanJ.

    The importance of the second person: Interpretation, practical knowledge, and normative attitudes

  • BrunyéT.T. et al.

    When You and I share perspectives: Pronouns modulate perspective taking during narrative comprehension

    Psychology Science

    (2009)
  • BrysonB.

    A walk in the woods

    (2010)
  • CenterF.C.M.

    Fulton county medical center main page

    (2018)
  • ChoiferA.

    A new understanding of the first-person and third-person perspectives

    Philosophical Papers

    (2018)
  • ChuJ.

    The water that falls on you from nowhere

    (2013)
  • ClarkE. et al.

    Neural text generation in stories using entity representations as context

  • CollobertR. et al.

    Natural language processing (almost) from scratch

    Journal of Machine Learning Research

    (2011)
  • DevlinJ. et al.

    BERT: Pre-training of deep bidirectional transformers for language understanding

  • DozatT. et al.

    Deep biaffine attention for neural dependency parsing

  • Eisenberg, J. D., & Finlayson, M. A. (2016). Automatic identification of narrative diegesis and point of view. In...
  • EvansM.T.

    Narcissistic abuse - you’re damned if you do, damned if you don’t

    (2015)
  • Ferreira, T. C., Moussallem, D., Kádár, Á., Wubben, S., & Krahmer, E. (2018). NeuralREG: An end-to-end approach to...
  • FludernikM.

    Second person fiction: Narrative “you” as addressee and/or protagonist

    AAA: Arbeiten aus Anglistik und Amerikanistik

    (1993)
  • FlynnN.

    Another bullshit night in suck city

    (2005)
  • FriedrichA. et al.

    Discourse-sensitive automatic identification of generic expressions

  • Fu, Z., Tan, X., Peng, N., Zhao, D., & Yan, R. (2018). Style transfer in text: Exploration and evaluation. In...
  • GardentC. et al.

    The webNLG challenge: Generating text from RDF data

  • GastN.G. et al.

    Optimal generation and storage scheduling in the presence of renewable forecast uncertainties

    (2013)
  • GroszB.J. et al.

    Centering: A framework for modeling the local coherence of discourse

    Computational Linguistics

    (1995)
  • HarriganP. et al.

    Second person: Role-playing and story in games and playable media

    (2007)
  • HartungF. et al.

    Taking perspective: Personal pronouns affect experiential aspects of literary reading

    PLoS One

    (2016)
  • HeylighenF. et al.

    Formality of language: Definition, measurement and behavioral determinantsTechnical report

    (1999)
  • HochreiterS. et al.

    Long short-term memory

    Neural Computation

    (1997)
  • HonnibalM. et al.

    Spacy

    (2017)
  • HovyE. et al.

    Ontonotes: The 90% solution

  • JaegerT.F.

    Redundancy and reduction: speakers manage syntactic information density

    (2010)
  • JhamtaniH. et al.

    Shakespearizing modern language using copy-enriched sequence to sequence models

  • JiY. et al.

    Dynamic entity representations in neural language models

  • Cited by (0)

    View full text