Storytelling with Adjustable Narrator Styles and

Sentiments

Boyang Li, Mohini Thakkar, Yijie Wang, and Mark O. Riedl

School of Interactive Computing, Georgia Institute of Technology

{boyangli, mthakkar, yijiewang, riedl}@gatech.edu

Abstract. Most storytelling systems to date rely on manually coded

knowledge, the cost of which usually restricts such systems to oper-

ate within a few domains where knowledge has been engineered. Open

Story Generation systems are capable of learning knowledge necessary

for telling stories in a given domain. In this paper, we describe a technique

tic and social behavioral knowledge. A user study shows our algorithm

strongly agrees with human judgment on the interestingness, concise-

ness, and sentiments of the generated stories and outperforms existing

algorithms.

1 Introduction

Narrative Intelligence (NI), or the ability to craft, tell, understand, and respond

to stories, is considered a hallmark of human intelligence and an effective commu-

nication method. It follows that Narrative Intelligence is important for Artificial

Intelligence that aims to simulate human intelligence or communicate effectively

with humans. In this paper, we focus on computational NI systems that can

micro-worlds where knowledge is available. For example, an automated story

generator may be told about the characters and environment of Little Red Rid-

ing Hood; that system can tell a large variety of stories about the given set of

characters and topic, but no stories about other characters or topics.

Open Story Generation systems (e.g. [5, 16]) have been proposed in order

to tackle the challenge of generating and telling stories in any domain. Such

systems can learn the needed knowledge for story generation and storytelling

without a priori knowledge engineering about a particular domain. We previ-

ously described an open story generation system, Scheherazade [5], which uses

crowdsourcing to construct a commonsense understanding about how to perform

everyday activities such as going to a restaurant or going to a movie theater.

involves describing the same events from different characters’ perspectives, pos-

sibly with opposing sentiments (cf. [2,18]). As a first step, we tackle the problem

of creating different storytelling styles for Open Story Generation. Style param-

eters are learned from large data sets including the Google N-Gram corpus [8]

and books from Project Gutenburg (www.gutenberg.org). We offer methods

to tune the storytelling with different levels of details, fictional language, and

sentiments. Our user study indicates our algorithm strongly agrees with human

readers’ intuition of linguistic styles and sentiments, and outperforms existing

methods.

Crowdsourcing has been proposed as a means for overcoming the knowledge

bottleneck. Sina et al. [14] use case-based reasoning to modify crowdsourced

semi-structured stories to create alibi for virtual suspects in training. None of

these approaches explicitly model discourse or generate different narration styles.

The work in this paper builds off our previous work on the Scheherazade

crowdsourced exemplar stories describing that situation. As opposed to other

that can be expected to occur, temporal ordering relations between events, and

mutual exclusions between events that create branching alternatives.

The learning of the plot graph proceeds in four steps [4, 5]. After exemplar

stories about a social situation are crowdsourced from Amazon Mechanical Turk

(AMT), the learning starts by creating clusters of sentences of similar semantic

meaning from different exemplar stories. Each cluster becomes an event in the

plot graph. In order to reduce the difficulty in natural language processing,

crowd workers from AMT have been asked to use simple language, i.e., using

one sentence with a single verb to describe one event, avoiding pronouns, etc. The

second and third steps identify the temporal precedences and mutual exclusions

between events. The final step identifies optional events. Story generation in

Focalization in narration refers to telling stories from different viewpoints

(e.g. of an omniscient entity or any story character; cf. [2]), potentially requiring

multiple narration styles. Most computational implementations focus on plot

events [11, 18] instead of linguistic variations. Curveship [10] generates focal-

ized text based on manually coded knowledge. Our work directly addresses the

differ mostly in aspects independent of content, such as in the use of swear

words, exclamation marks and shuttering. Instead of generating from symbolic

representations with precise semantic meaning, we select from existing sentences

that are similar but not strictly synonymous to describe an event (i.e. sentences

may differ in content). We consider parameters directly related to word choices:

degree of details, fictionality, and sentiments.

2 Storytelling with Different Styles and Sentiments

This section describes the process of telling stories in natural language with a

variety of personal styles. The architecture for story generation and communi-

cation is shown in Figure 1. Plot graph learning is typically an offline process

the selected events with personal styles and affects, which we explain below.

Recall that each event in the learned plot graph is a cluster of natural lan-

guage descriptions of similar meaning. Given a generated story (a complete,

linear sequence of events), natural language text is generated by selecting the

sentence from each cluster that best matches the intended narration style. We

describe two criteria for selecting sentences: (1) the interestingness of the text

and (2) the sentiment of the text. We aim to create a diverse set of storytelling

styles that may be suitable for different occasions. For example, some narrators

Fig. 1. The system pipeline.

or story characters may speak very succinctly, whereas others can recall vivid de-

tails. A positive tone may be used if the narrator wants to cheer up the audience;

a negative tone may be suitable for horror stories and so on. We also present

a Viterbi-style algorithm that considers preferences on individual sentences and

workers to provide “interesting” event descriptions. For $1, workers wrote de-

tailed descriptions for each event in a given story; each description may contain

more than one sentences. We allowed workers to interpret “interesting” however

they wanted, though we suggested that they describe characters’ intentions, fa-

cial expressions, and actions. Each worker saw a complete sequence of events

to make sure they understand the story context. We accepted all stories that

describe the events we provide, and did not perform a manual check of interest-

ingness.

story language resembles the language used in fictions. We model the amount of

details with the probability of a sentence in English, since Information Theory

suggests a less likely sentence contains more information (i.e. more details). We

compute the probability of an English word as its frequency in the Google N-

Gram corpus. Due to the large size of the corpus, these frequencies approximate

word probabilities in general English. We compute the probability of a sentence

using the bag-of-word model, where the probability of sentence S containing

words w

times is

Stories may be told with positive or negative sentiment. To detect sentiments

of sentences in each event cluster, we construct a sentiment dictionary called

is considered a unique word. For every word we want to compute sentiment

value for, we consider a neighborhood of 100 words, 50 to its left and 50 to

its right. The target word is at position 0 and denoted as w

its immediate left and right are at position -1 and 1, and so forth. The posi-

tions of these words are included in the index set N. For word w

at position

: g

(j) = exp

−(i − j)

is computed as a weighted average of all kernel functions at position 0:

0

=

(0)

(4)

where s

is the sentiment value from SentiWordNet, i.e. the difference between

the positive and negative polarity. The SentiWordNet value for the target word

has no influence on itself, i.e. 0 /∈ N. As a word can appear multiple times

in different neighborhoods, the final sentiment value for w

is the average over

card(V )

heavily than words with weak sentiments.

We selected a subset of English words that are of interest to our task. The

exemplar stories in two previously crowdsourced social situations—dating at the

movie theater and bank robbery—contain 1001 unique nouns, verbs, adverbs

and adjectives. We selected highly influential adjectives and adverbs from their

direct neighbors, producing a total of 7559 words. We normalize the raw values

produced by smoothing, so that 1 percentile and 99 percentile of the values fall

in the range of [−1, 1], to account for outliers.

Table 1 shows some of the most

positive and most negative sentences. We find the results to reflect the valences

of individual words. Although this approach works most of the time, there are

cases such as sarcasm where the sentiment of a sentence could be the opposite

For each event, we can find individual sentences ranked highest for any criterion

or combinations of criteria using the harmonic mean. However, this selection

does not consider the coherence between sentences and may results in incoherent

texts due to two major problems: (1) previously mentioned objects can suddenly

disappear and previously unmentioned objects can appear, and (2) a sentence

variable generating any observation. That is, the observation indicates preference

over values of the hidden variable. Two, the probabilities of a hidden variable

transiting to the next hidden variable. That is, we have preferences over pairs of

values for adjacent variables.

Our problem is similar as we want to find the highest scored sentence sequence

based on preferences over sentences in each event cluster, and preferences on how

adjacent sentences connect. In this paper, we do not consider connection between

non-adjacent sentences. Specifically, we score the connection between any two

sentences s

from the j

, we need to consider the scores of best sequences

ending in every sentence from the previous cluster c

, in addition to the con-

nection between every sentence from cluster c

and s

. Due to the Markov

in every sentence from every cluster and their scores. For a sequence of n clusters

and m sentences in each cluster, the time and space complexity are O(m

The list of books can be downloaded at http://www.cc.gatech.edu/

bli46/SBG/

list.txt. The resulted dictionary is at: http://www.cc.gatech.edu/

bli46/SBG/

dic.txt.

(seq

, score(seq

)) ← BestSeqEndingIn(s

, c

)

end for

return the highest scored sequence from seq

, seq

, . . . , seq

end function

function BestSeqEndingIn(s

, c

)

for each sentence s

∈ {s

, s

do

end for

best

return (best seq, score(best seq))

end function

Table 3. Statistics of crowdsourced interesting stories

# Stories 20 10

and bank robbery. In addition to the originally crowdsourced exemplar stories,

we crowdsourced interesting stories using procedures described in Section 2.

Some statistics of these stories are shown in Table 3. Some of these sentences can

be seen in prior examples showing least probable and most fictional sentences

for particular clusters (the most probably sentence typically comes from the

original, simplified language exemplars).

factors. The first group includes stories generated from the most interesting

details (MID) criterion, the most probable (MostProb) criterion, and a story

where we use the MID criterion but penalize long sentences. After reading the

stories, participants are asked to select the most interesting story, the most

detailed story and the most concise story. Our hypothesis is that human readers

will select the MID story as containing the most details and the most interesting,

and the MostProb story as the most concise. We set α to 12. The second group of