TheRiddlerBot
A next step on the ladder towards creative Twitter bots
Iván Guerrero1, Ben Verhoeven2, Francesco Barbieri3, Pedro Martins4,
Rafael Pérez y Pérez5
1Universidad Nacional Autónoma de México, D.F., México
2CLiPS Research Center, University of Antwerp, Belgium
3Universitat Pompeu Fabra, Barcelona, Spain
4CISUC, University of Coimbra, Portugal
5Universidad Autónoma Metropolitana, Cuajimalpa, D.F., México
Abstract
We present a computational model for the generation of a
Twitter bot that aspires to be considered creative by generating
riddles about celebrities and well-known characters. The
riddles are created by combining information from both wellstructured
and poorly-structured information sources. This
model has been implemented as an interactive Twitter bot
(@TheRiddlerBot) that presents its outputs as contests to its
followers, checks the posted answers and replies accordingly.
Lastly, we present a discussion about the main attributes of
a creative Twitter bot, and the remaining work for our bot to
qualify as such.
Introduction
On several social networks, but especially Twitter, a new variety
of users, the bots, are increasingly interacting not only
with human users, but even among themselves. The first
Twitter bots that appeared on the web were considered in the
best case graceful, and sometimes even useful, or helpful,
but they were far from being considered creative. To be
creative usually relates to the generation of something novel
and interesting, not only to oneself, but also to partners
sharing a common background (Mayer 1999). According to
this, a creative activity can be considered a social activity
as well, since the environment evaluates any generational
process to determine if it can be considered truly creative or
not. In this sense, the environment establishes diverse constraints
to any creational process, and the main challenge for
an inventor resides in freeing himself from all these conventions
to create something novel, interesting and yet valuable.
Novel ways of interacting inside social networks have
added new and barely studied constraints to the creative process.
Contests for the generation of micro-stories (Hamid
2014) - 100 words long stories -, similar to Tweet messages,
or the generation of writing maps - writing prompts to
inspire writers - (Maps 2015) have emerged from these new
ways of interaction.
The problem that we tackle in this paper is the design
and implementation of a Twitter bot that can be considered
creative, focusing on missing features in the prevailing bots.
The use of more realistic and diverse knowledge sources
(Twitter, Facebook, Wikipedia, online news sites), evaluative
mechanisms for its own outputs, and the definition of a
purpose which surpasses the generation of pseudo-random
messages, are examples of such omissions. The goal of
our bot is to generate riddles about celebrities, formed
as questions to encourage readers to assert the name of a
famous character.
The rest of the paper is organized as follows. We first
give an overview of the state of the art of Twitter bots, after
which we give a general description of a model to automatically
generate riddles and its implementation in a Twitter
bot (@TheRiddlerBot). We present the results of a questionnaire
where we asked people to evaluate a set of riddles.
We then close with a general discussion of our proposal and
our conclusions.
Related research
We now present relevant research from two different fields:
riddle generation, and automatic Tweet generation. The
existing theories related to the generation of riddles are
not yet complete mainly because their descriptions only
contemplate a subset of riddles (typically those in the
question-answer format). Nevertheless, we present several
approaches that provide relevant features that should be
present in any riddle to be considered as such. Besides,
we describe several first generation Twitter bots, Tweetgenerating
systems that autonomously perform useful and
well-defined services (Veale 2014), that are using Twitter
in diverse ways. We distinguish feeder bots, which create
tons of Tweets for their followers; watcher bots, which are
constantly looking for specific texts to extract information;
and interactive bots, which ask followers for specific ways
of communication and information sharing (Cook 2015).
We describe different Twitter bots as examples of the state
of the art. We will focus on both the creative aspects and
unique features that are already present, as well as missing
features.
Pepicello (Pepicello and Green 1984), among others,
has researched riddles extensively, and described them as
text fragments that employ ordinary language restricted
by semiotic, aesthetic and grammatical artistic constraints.
They argue that ambiguity in these descriptions is a key
Proceedings of the Sixth International Conference on Computational Creativity June 2015 315
aspect of a riddle, and they define three types: phonological
(use of words with the same phonetic code), morphological
(use of words with the same writing) and syntactic (phrases
with different possible interpretations). According to this
work, the goal of a riddle is to confuse the guesser by
utilizing one or more of these ambiguities.
Additionally, Weiner (Weiner and De Palma 1993)
defines a riddle as a language game, initiated by a question,
with the goal to mislead the guesser. They describe two
pragmatic mechanisms for the generation and comprehension
of riddles: accessibility hierarchy and parallelism. The
former relates to categorization, the capability to relate
different concepts to accomplish specific goals. They
describe parallelism as the tendency to remain in the same
cognitive space unless a force makes us change to an
alternative representation. They state that we, as humans,
employ these two mechanisms to generate and comprehend
riddles.
According to this work, there exist two types of concepts
present in every category: context-invariant (what first
comes to our minds) and context-variant (present when a
relevant context appears). They state that a riddle must
bring to our minds the context-invariant information to
mislead the answer of the riddlee. Parallelism, in turn, helps
on to generate false expectations on the part of the guesser.
JAPE - Joke Analysis and Production Engine - (Binsted
1996) is a question-answer riddle generation system.
Herein, several strategies to generate riddles are described:
syllable substitution, word substitution and metathesis. The
first mechanism consists in confusing the syllable in a word
with a similar sounding word; the second, confuses an
entire word with a similar sounding word; the third, reverses
the sounds of two words to suggest a similarity in meaning
between two phrases.
To generate riddles, JAPE uses templates consisting of
‘canned text’ with slots where words or phrases are inserted.
To determine which words are to be incorporated to the
final riddle, the system makes use of predefined schemas,
which establish relationships between words which must
hold to build a joke. These schemas are manually built from
previously known jokes.
In an effort to delineate novel uses of Twitter, Angelina-
5 (Cook and Colton 2014) is a software for the generation
of 3D games that uses a module to evaluate its textures, i.e.
images utilized for decorating walls and ceilings inside the
scenario, in a Twitter account (@angelinasgames). Each
game has a theme, initiated by a word or phrase. Angelina-5
obtains a set of words associated to it from an English
corpus, and uses them to retrieve sound effects, textures,
3D models and fonts to create a game. The bot periodically
Tweets images and asks its followers to associate terms to
it. These terms are collected into a repository to be further
used as tags for the image.
This bot can be classified as a watcher with the goal of
obtaining tags for a tweeted image from the user. The bot
does not have any capabilities for analyzing the information
received, given its very limited function within Angelina-5.
Nevertheless, it is a functional example of how bots can
receive information from humans to enhance the capabilities
of a system, a desirable function to contemplate in our bot.
Flux Capacitor (Veale 2014) is a generator of wellformed
and interesting character arcs (conceptual starting
and ending points for a character inside a narrative). These
character descriptions are defined in terms of properties,
and a well-formed arc contemplates representative changes
by looking for templates (such as XbecomesY ) in Google
n-grams (Brants and Franz 2006). Apart from that, relationships
among properties to describe such states are retrieved
from WordNet (Fellbaum 1998). The output of the bot
serves the MetaphorIsMyBusiness (@MetaphorMagnet)
Twitter bot to generate metaphors related to character twists
in a story.
This bot has several aspects that differentiate it from first
generation bots, such as its capability to deal with massive,
poorly-structured knowledge databases (those lacking
a well-defined format), and its purpose to create outputs surpassing
the generation of pseudo-random messages. Another
aspect of the bot is its high curation coefficient, the
ratio of good outputs to all outputs, since the system contemplates
mechanisms to evaluate its own outputs and filter
those considered with low quality.
General description
We present a model for a Twitter agent with creative behaviors
such as its abilities to utilize real-world, poorlystructured
data sources, to evaluate its own outputs, and to
interact with Twitter users. We describe as well the implementation
of our model in a Twitter bot (@TheRiddlerBot)
that generates creative riddles about fictional or real characters
(e.g. celebrities) using cross-references from different
knowledge bases.
Model description
The model consists of five main modules each subdivided
in three layers (see Figure 1). Each module has a specific
task ranging from the selection of a relevant celebrity, to the
publication of the riddle in Twitter and tracing the answers
of the followers. Besides, a layered structure of the system
provides every module of tools for retrieving additional information
from diverse sources, for processing the information
available, and for evaluating its outputs. Now we describe
the main characteristics of each module and how its
tasks are distributed among the diverse layers of the model.
Character selection module This module initiates by retrieving
a list of celebrity names from diverse knowledge
bases. Some sources may have well-structured information,
such as the Non-Official Characterization (NOC) list (Veale
2015), whereas others may lack this structure, such as
Proceedings of the Sixth International Conference on Computational Creativity June 2015 316
Figure 1: Model architecture
Google News, trending topics from Twitter, or public information
from Facebook. This task resides inside the first
layer of the module, the information retrieval layer. The data
obtained is then passed to the processing layer, where one of
the celebrities is selected according to diverse criteria such
as his public relevance. These criteria give clues about the
current importance of the celebrity due to the events he or
she has recently been involved in. Finally, the evaluation
layer determines if the selected character has been lately
used to generate riddles, in which case it is not suitable for
a new riddle. Once the character selection process finishes,
the name of the celebrity is passed to the next module to look
for as many facts as possible about him.
Feature Extraction Module This second module gathers
attributes about the previously selected character from
both well-structured sources, such as the NOC list, and
poorly-structured sources, such as Wikipedia. Furthermore,
common sense knowledge bases (see the Perception dataset
of the Nodebox project1) serve as repositories for hypernyms
(super categories) of the character’s attributes. These
tasks are performed inside the first layer of the module. All
the information obtained is then passed to the processing
layer, where a subset of features is extracted according to
their uniqueness and interestingness. A subset of features
is considered unique if they describe only one celebrity.
This evaluation is important because a riddle with unique
traits is not always desirable, since it becomes easy to
solve. A riddle is considered interesting when it describes a
character with attributes that altogether represent relevant
traits, but do not provide excessive information so that
the riddle cannot be easily guessed. A set of attributes is
considered relevant when the sum of the n-gram percentage
of its elements, according to the Google N-gram viewer.
An isolated attribute is considered to provide excessive
information when its n-gram percentage is too low, and it
can be considered unique. These values still need to be
determined and further studies must be done to evaluate its
accuracy.
Lastly, the evaluation layer determines if the subset selected
has not been previously used for the same character,
and that the evaluation of the attributes in previous riddles is
acceptable to keep using them. These features are finally
1
http://www.nodebox.net/perception
sent to the next module to extract additional information
from them.
Analogy Generation Module The third module initiates
by gathering information about similar characters according
to the features of the character selected for the riddle. For
this purpose, it uses information available at the NOC list
as well. Then, it retrieves descriptions of analogies for the
generation of relations between characters. We consider
two different types of relations between a character and his
attributes. Direct relations exist between a character and
his features (’Diego Rivera’ lived in ’Mexico’, ’Tequila’ is
produced in ’Mexico’); higher-order relations exist between
a character and a concept related to one of his features
(’Diego Rivera’ lived in the country where ’Tequila’ is
produced). For this last example, we substituted an attribute
by its hypernym to create the relation (’country’ is a
hypernym of ’Mexico’).
The information for the generation of analogies is passed
to the processing layer where such analogies are created according
to the attributes selected for the character. Finally,
the evaluation layer determines if the mixture of attributes
and analogies has been previously used to create riddles of
the same character, in which case the analogies are discarded
and new attributes are analyzed. With the set of features
and analogies complete, the information is now passed to
the next module to convert them into utterances.
Natural Language Generation Module This module
initiates with the retrieval of different types of phrasal
templates for each part of the riddle (initial phrase, clues,
final question). These templates are stored and retrieved
from a repository specially developed for this project.
A phrasal template is a previously-known sentence with
slots to be further filled by specific words (Becker 1975).
Each slot is commonly associated with a part-of-speech
tag which allows to preserve the syntax of the sentence.
Inside the processing layer, the module performs a process
to select one template for each type of sentence. The
selection process begins with the random selection of an
initial-phrase template. Then, several clue templates are
selected preventing than recently utilized templates are
now repeatedly chosen. Lastly, a final-question template is
picked in accordance to the first selected template.
These templates have the purpose of providing the system
with a wider variety of possible generations. Once a template
is selected, the slots are filled with either character’s
attributes, or analogy information.
User Interaction Module The last module extracts a list
of aliases for the character to be guessed. The processing
layer prepares and tweets the riddle, and starts looking for
responses in Twitter, which are compared against the previously
obtained aliases. If there is a match, the riddle is
considered to be finished. Users get points for each correct
answer, which makes this system into a kind of game. When
a wrong answer is detected, the user is notified and encouraged
to try again. The number of incorrect answers of a
Proceedings of the Sixth International Conference on Computational Creativity June 2015 317
riddle can be further employed in the evaluation layer of the
module. They cast light on the difficulty level of the riddle,
and on the interestingness and uniqueness of the attributes
employed to generate it.
System description
We have been incrementally implementing the previous
model in a Twitter bot called ‘TheRiddlerBot’2. We started
out with random character selection, direct relations with
tratis, and Twitter publishing capabilities. By now, we have
incorporated several new features into the system, which
are explained below.
The character selection process retrieves a list of celebrities
from the NOC list, and randomly selects one of them.
If the character has not been used to generate one of the last
riddles, he is passed to the following module. The NOC
list is a matrix where every row contains information about
a famous character and every column is a trait, so every
cell contains the value of a trait for a specific character.
Several additional matrices exist where further information
about specific traits, such as clothing, fictional worlds,
vehicles, weapons..., can be found as part of the project.
Due to its simplicity, we utilize the Pattern package in
Python (De Smedt and Daelemans 2012) for a wide variety
of tasks, from the feature extraction from comma-separatedvalues
files, to the reception and sending of Tweets.
Direct relations between traits and the given character
are obtained from the NOC list as well, whereas common
sense knowledge is obtained from the Perception demo3
to incorporate additional traits to the available character’s
knowledge.
From the list of available traits, three of them are
randomly selected and evaluated to determine its interestingness
and uniqueness (in the current version this
evaluation is not fully implemented yet). The selected
traits are considered unique when we cannot find additional
celebrities inside the NOC list with the same values. To
determine the interestingness of the attributes, we look
for them on the character’s Wikipedia page, and if they
are present, we consider them relevant. If the number of
characters sharing the selected values surpasses a threshold,
they are not considered relevant, in such case a new subset
of features is randomly selected and the generation process
re-initiates. Finally, the evaluation layer looks for similar
subsets for the same character in previous riddles. If this
subset-character pair has been previously used, the feature
extraction process re-initiates until a suitable subset is
found. These features are finally sent to the next module to
determine if additional relations can be obtained.
We retrieve from the additional matrices of the NOC list
project information to generate higher-order relations about
2
Source code available on github:
https://github.com/ivangro/theriddlerbot 3
http://www.nodebox.net/perception
fictional worlds and group affiliations. For this purpose, we
look for characters who share their profession with the previously
depicted character. Then, we filter out those characters
who don’t have any information about their fictional worlds
or group affiliations. From the remaining characters, one is
randomly selected to create an analogy by means of a template.
We randomly depict a template from the repository
available inside the system (see table 1).
Analogy type Template
Fictional <char1>: like <char2> in <world1>
world <char1>: like someone in <world2>
Group <char1>: like <char2> in <group1>
affiliation <char1>: like someone in <group2>
Table 1: Sample analogy templates
An analogy template consist of two parts: concept, and
reinterpretation of the concept, in the form <concept>
: <reinterpretation>. In general, an analogy template
contains redescriptions of the selected character for the
riddle (<char1>), in terms of a second character (<char2>),
and the fictional world or the group affiliation of one of
the characters (<worldi> or <groupi>). For instance, we
can reinterpret the character ‘The Joker’, whose fictional
world is ‘The Dark Knight Rises’, in terms of another
character with the same profession, ‘criminal’. In this case,
we employ ‘Morpheus’, who can be considered a criminal
in ‘The Matrix’, and the first template for fictional worlds,
to state that ‘The Joker’ is like ‘Morpheus’ but in ‘The
Dark Knight Rises’. Besides, using the second template
we obtain that ‘The Joker’ is similar to ’someone’ in ‘The
Matrix’.
Finally, the evaluation layer determines if the mixture of
attributes and analogies has been previously used to create
riddles of the same character, in which case the analogies
would be discarded and new attributes are selected.
The main task of the language generation module is to
represent the attributes and analogies obtained from the
previous step as utterances. Inside the feature extraction
layer, the date of death of the character is retrieved from
his Wikipedia page to determine if he is still alive or not.
This information is further employed to conjugate the verbs
of the generated phrases (a riddle about a deceased person
is written in past tense). Some phrases require additional
information of the character to present a more elaborated
text, for this reason we obtain positive and negative adjectives
describing the character from the NOC list. Inside the
generation layer, we convert features and analogies to text.
For this task we employ three different types of phrasal templates:
introductory templates (see table 2), clue templates
(see table 3), and final question templates (see table 4).
For the clue templates, several attributes are available for
the system to select one of them. The list includes clothing,
opponent, opponent activity, married partner, typical activity,
vehicle and country.
Every template consists of three different types of
Proceedings of the Sixth International Conference on Computational Creativity June 2015 318
Type Template
First person I
Third person Tell me the name of a person that
Table 2: Sample introductory templates
Feature type Template
Group affiliation -be/VB the <char2> of <group1>
analogy -could have belonged to <group2>,
but do/VB not
-be/VB like <char2> but be/VB
not part of <group2>
Fictional world -be/VB similar to someone in
analogy <world2>
-be/VB the <char2> of <world1>
Profession -be/VB <value>
attribute -be/VB <value>, <pos_adj>
yet <neg_adj>
Opponent -do/VB not like <value>
attribute -be/VB definitely not a
close friend of <value>
Hyperonym -be/VB known as <value>
attribute -be/VB <value>, <pos_adj>
yet <neg_adj>
Clothes -have/VB been seen wearing
<value>
Table 3: Sample clue templates
elements: words, verbs (marked with the tag VB), and slots
(<slotf iller>). The verbs are conjugated in accordance
with the type of template depicted for the introductory
phrase. Afterwards, the attributes and analogies selected
in the previous module are converted to text by replacing
the slot fillers of the selected template with the values
associated to the attributes and analogies. To conclude,
the final question template is selected in accordance to the
introductory-phrase template. Once the three phrases are
generated in natural language, they are chunked as a riddle.
The last module tweets the generated riddle to open a new
contest. To determine who wins a contest, we obtain a list
of aliases for the character from his Wikipedia page. Every
time a follower replies a riddle, his answer is obtained to be
compared against each of the available aliases for the character,
and if one of them matches, the contest is declared
finished and a Tweet is published to point out the winner; if
none of the aliases match, a reply to the owner is sent stating
that the answer was not accurate, and the contest continues.
If, after several hours, the riddle had no correct answer, a
Tweet exposing the celebrity is sent, and a new contest begins.
Example of a riddle
Now, we show how to generate a riddle about ‘The Joker’.
Once the character selection module finishes, several attributes
are obtained for the character from a variety of
sources (see Table 5).
Type Template
First person Who might I be?
Third person Who is this?
Table 4: Sample final question templates
Type Value
Hypernym ‘maniac’, ‘madman’, ‘criminal’
Group affiliation ‘The Dark Knight Rises’
Clothes ‘a purple topcoat’,
‘a green wig’
Pos. adjectives ‘playful’, ‘witty’, ‘flamboyant’,
‘cunning’, ‘brilliant’, ‘creative’
Neg. adjectives ‘maniacal’, ‘cruel’, ‘sadistic’,
‘inhuman’
Table 5: Sample attribute and analogy values for ‘The Joker’
From the available attributes, a subset of three traits is
selected (for this example, group, profession, and clothes),
and their corresponding values are sent to the analogy module.
If one of the attributes is suitable to generate analogies,
the process initiates. In this case, the group attribute is used
to create an analogy. We look inside the knowledge base for
characters who share a hypernym (see Table 6).
Character Group affiliation
‘Fagin’ ‘Oliver Twist’
‘John Dillinger’ ‘Public Enemies’
‘Fredo Corleone’ ‘The Godfather’
‘Snake Plissken’ ‘Escape From New York’
‘Morpheus’ ‘The Matrix’
Table 6: Characters sharing a hypernym with ‘The Joker’
With the information obtained from the previous steps,
we randomly select an analogy template (<char1>: like
<char2> in <group1>), and it is encapsulated, with the
rest of the values for the natural language generation.
Here, an introductory template is selected (‘Tell me the
name of a person that’), three clue templates are selected,
one for each of the attributes or analogies employed, and
a final question template is retrieved as well (‘Who is this?’).
The clue template for group is (be/VB the <char2> of
<group1>), for hypernym is (be/VB <value>, <pos_adj>
yet <neg_adj>), and for clothes is (have/VB been seen
wearing <value>). If several values are available for an
attribute, one of the is randomly picked to replace the empty
slot.
Finally, we create the riddle by chunking the three templates
where its slots are replaced with the corresponding
values:
Tell me the name of a person that is the Morpheus of
The Dark Knight Rises, is criminal, playful yet cruel,
has been seen wearing a purple topcoat. Who is this?
Proceedings of the Sixth International Conference on Computational Creativity June 2015 319
Model evaluation
As described above, we save all the posted riddles and their
metadata (number of retweets, favorites, answers, etc.) in
a database. The metadata could be used for the evaluation
of the model if we assume that a riddle with more wrong
answers is harder or that a riddle with a lot of favorites is
better. Unfortunately, our bot is not popular enough yet, so
there is very little interaction. Here are some numbers to
give you an idea. At the time of writing (April 29th 2015),
our bot has 57 followers. Since February 2nd 2015 (date of
implementation of the database) 285 riddles were posted.
Ten different users gave correct answers to 34 riddles in
total.
So we decided to perform a different evaluation. We
asked 86 people to each evaluate five riddles. We first asked
the participants to guess the answer to the riddle. Then,
we presented the correct answer and asked if they knew
the person in question. The participant indicated whether
he considered the quality of the riddle satisfactory and, if
negative, gave us the reason why it wasn’t good.
Figure 2 shows the percentage of correct answers
(15.58%), and the number of known celebrities (54.19%)
once the correct answer was presented.
Figure 2: Results for correct answers and good descriptions
Figure 3 shows the number of riddles considered to have
accurate descriptions of the characters (41.86%). When that
was not the case, the main reason chosen was that the description
was too vague (36.51%). Among the additional
reasons given, the most recurrent was that the character was
already dead and the riddle was written in present (< 1%).
Finally, we present here the top 5 answered riddles,
according to the number of times they appeared and the
number of correct answers given to them.
Tell me the name of a person that can be found in UK,
enjoys robbing from the rich, likes wearing a feathered cap.
(Answer: Robin Hood).
Who is a creator, can be found in Italy, wears a paintstained
smock? (Answer: Michelangelo).
Figure 3: Results about the accuracy of the description
Who is a creative professional, pretty yet superficial,
can be found in USA, enjoys monetizing celebrity status?
(Answer: Paris Hilton).
Who is a religious leader, loves spreading Christianity,
likes wearing sandals? (Answer: Jesus Christ).
Who is the Hermione Granger of The Simpsons, wears
an orange dress, is the Timothy McGee of The Simpsons
Family? (Answer: Lisa Simpson).
Discussion and future development
Relevant results were obtained from applying the questionnaire.
The percentage of known celebrities once the answer
was presented (54.19%) indicates that the process for the
selection of celebrities should be improved. From this
result we realised that almost half of the riddles could not
be correctly answered because people did not have enough
information about the character. One reason for this result is
that owners of the NOC list (the main source for celebrities)
and the riddlees were from different countries, and they did
not have enough information in common.
The percentage of good descriptions of the celebrities
(41.86%) represents our curation coefficient (the ratio of
good outputs to all outputs), and the major cause for our
descriptions to be considered wrong was its vagueness. This
indicates that further work must be done to improve the
interestingness of our riddles (the description of a character
with relevant attributes, but without excessive information
to be easily guessed). Thereby, additional mechanisms to
determine the number of traits to incorporate to a riddle
based on its relevance, might prevent descriptions from
being too vague.
The low number of correct answers (15.58%) suggests
that the complexity of the generated riddles is high. Nevertheless,
by improving the character and trait selection
processes will mitigate this problem.
Proceedings of the Sixth International Conference on Computational Creativity June 2015 320
In general terms, the current version of the system still
lacks selection mechanisms relying on informed decisions.
For instance, the character selection process randomly
picks a celebrity from a list; the feature selection randomly
chooses three character traits, despite of the final evaluation
which determines whether they are good enough or not
to continue with the process; most of the templates are
randomly picked as well, and the values replacing the empty
slots in such templates follow the same track. We consider
that transforming as many random selections as possible
into informed decisions will contribute in an overall increment
of the final quality of the outputs generated by our
bot, and will provide our model of additional traits for it to
be considered creative. A key aspect to distinguish simple
generation from creative generation is the curation coeffi-
cient in the outputs. To increase the number of high-quality
riddles generated by our system several improvements will
take place in the next release of the system.
The work presented here is a first step in building up a
robust, Twitter bot that can be considered creative. For the
next release we still need to improve several aspects related
to intermediate output validation, and mechanisms for the
automatic expansion of the current knowledge bases utilized
by the system.
Despite the fact that several knowledge bases such as
Conceptnet (Liu and Singh 2004) or Facebook, are not part
of the project yet, the current version of the system already
contains fully working mechanisms for information
retrieval, and is still pending to exploit this information to
generate more interesting, high quality riddles.
Conclusions
We have described a computational model to generate
riddles about celebrities. It consists of modules to select
a celebrity, to retrieve relevant traits to describe him, to
generate analogies between his attributes and convert such
descriptions into utterances, and to tweet the generated
riddle and interact with Twitter users by evaluating their
answers. The model presents a subdivision of each module
in layers. The first layer is responsible for all the data
extraction processes; the second, for processing of the information
retrieved; the last, for the evaluation and validation
of the generated outputs. We consider this layered approach
relevant because it provides tools to enrich the intermediate
outputs of every module. It contemplates the retrieval of
additional information, when required, and the validation
of intermediate results to achieve a higher quality in the
outputs. We present an implementation of our model in a
Twitter bot named ‘TheRiddlerBot’. Herein, we introduce
several difficulties emerging from reifying our model, such
as gathering character traits, generating analogies, and
generating natural language utterances.
We consider ‘TheRiddlerBot’ as a creative agent according
to the following considerations. If we describe a creative
bot in terms of its capability to deal with poorly-structured
knowledge to generate something interesting and novel,
we have provided our system with such capabilities. Some
authors on the field consider as essential properties for
an artifact to be considered creative, novelty, quality and
typicality of its outputs (Ritchie 2007). Although similar
riddles can be found widespread over the literature, we
consider that our system generates novel outputs since the
traits employed by our implementation, considering the
incorporation of analogies, make them rare to replicate. We
still need to implement direct and indirect evaluations for
the overall quality of the riddles, but we have sketched in
this document several validation mechanisms to ensure the
overall quality of our outputs. According to our definition
of a riddle, questions to encourage readers to assert the
name of a famous character, we argue that our outputs are
typical examples of this type of queries.
According to Pérez y Pérez (2013, 2014), any output must
be presented in a correct manner (coherence), generate new
knowledge to the reader (interesting), and be considered new
(novelty) to be creative. We verify the coherence of our
riddles particularly in two stages: the analogy and natural
language generation models. The analogy and phrasal
templates provide the system with well-formed structures to
generate complex attributes of a riddle (analogies), and to
generate readable phrases written in natural language. During
the evaluation layer at every module, we validate the
novelty of our riddles, since at every stage of the process
we ensure that the intermediate outputs have not been previously
utilized. Our system considers a riddle to be interesting
looking for the traits to describe a character at his
Wikipedia page, and also detecting that we have not utilized
the same subset on previous riddles. The first validation
gives us clues about the relevance of the traits. If the reader
does not know all the presented information, he will be capable
of learning new qualitites of a celebrity. The second
validation lets the system be certain of the uniqueness of the
employed traits.
Acknowledgements
This research was sponsored by PROSECCO4. We would
like to thank them as well for organizing the code camp on
computational creativity5 in Coimbra, Portugal, where this
research project began and started to grow.
This research was also sponsored by the National Council
of Science and Technology in Mexico (CONACyT), project
number: 181561.
The second author is supported by the FWO Research
Foundation - Flanders.
<references_biblio/>
References
Becker, J. D. 1975. The phrasal lexicon. In Proceedings
of the 1975 Workshop on Theoretical Issues in Natural Language
Processing, TINLAP ’75, 60–63. Stroudsburg, PA,
USA: Association for Computational Linguistics.
4
http://prosecco-network.eu
5
http://codecampcc.dei.uc.pt
Proceedings of the Sixth International Conference on Computational Creativity June 2015 321
Binsted, K. 1996. Machine humour: An implemented model
of puns (PhD thesis). University of Edinburgh.
Brants, T., and Franz, A. 2006. Web 1T 5-gram database,
Version 1. Linguistic Data Consortium.
Cook, M., and Colton, S. 2014. Ludus ex machina: Building
a 3D game designer that competes alongside humans. In
Proceedings of the Fifth International Conference on Computational
Creativity.
Cook, M. 2015. A brief history of the future of twitterbots.
Presented at the PROSECCO Code Camp on Computational
Creativity.
De Smedt, T., and Daelemans, W. 2012. Pattern for python.
Journal of Machine Learning Research 13:2031–2035.
Fellbaum, C. 1998. WordNet: An Electronic Lexical
Database. Cambridge, MA: MIT Press.
Hamid, A. 2014. Just 100 words.
Liu, H., and Singh, P. 2004. Conceptnet: A practical
commonsense reasoning tool-kit. BT Technology Journal
22(4):211–226.
Maps, W. 2015. Writing maps.
Mayer, R. E. 1999. Fifty years of creativity research. Cambridge,
UK: In R.J. Sternberg, Handbook of creativity.
Pepicello, W. J., and Green, T. A. 1984. The language of
riddles: new perspectives. Columbus, USA: Ohio State University
Press.
Pérez y Pérez, R., and Otoniel, O. 2013. A model for evaluating
interestingness in a computer-generated plot. In Proceedings
of the Fourth International Conference on Computational
Creativity.
Pérez y Pérez, R. 2014. The three layers evaluation model
for computer-generated plots. In Proceedings of the Fifth
International Conference on Computational Creativity.
Ritchie, G. 2007. Some empirical criteria for attributing creativity
to a computer program. Minds and Machines 17:76–
99.
Veale, T. 2014. Comming good and breaking bad: Generating
transformative character arcs for use in compelling stories.
In Proceedings of the Fifth International Conference
on Computational Creativity.
Veale, T. 2015. A game of tropes: Exploring the placebo
effect in computational creativity. In Submitted to the International
Conference on Computational Creativity.
Weiner, J. E., and De Palma, P. 1993. Some pragmatic
features of lexical ambiguity and simple riddles. Language
& Communication.
Proceedings of the Sixth International Conference on Computational Creativity June 2015 322