English. This paper presents BullyFrame, a dataset of cyberbulling interactions collected from WhatsApp conversations in Italian and annotated with FrameNet semantic frames. We will describe the creation of the dataset discussing the problematic aspects found in the annotation process, such as the lack of coverage of FrameNet for the annotation of texts extracted from social media. Finally, we present a preliminary study that describes the relations between the frames and the cyberbullying-related annotation of the original dataset. 1
Italiano. Questo studio presenta BullyFrame, un dataset di conversazioni WhatsApp in italiano contenenti episodi di cyberbullismo e annotate secondo i frame semantici di FrameNet. Verra` descritta la creazione del dataset discutendo gli aspetti problematici incontrati nel processo di annotazione, come ad esempio i limiti di copertura di FrameNet per l’annotazione di testi estratti da social media. Infine, presentiamo uno studio preliminare che descrive le relazioni tra l’annotazione di FrameNet e quella del dataset originale, relativa al cyberbullismo.
The semantic analysis of a text involves the classification of predicates into a set of events, for which it is important to determine who did what, when and where. For example, in the sentence “In 1912, the Titanic hit an iceberg on its first trip across the 1Copyright c 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
Atlantic”, the verb “hit” represents the event,
“Titanic” is the main actor of that event, “1912” and
“Atlantic” indicate when and where it took place,
and so on. The process of extracting the semantic
roles and relations in a sentence is called Semantic
Role Labeling (SRL), and, in the last years, both
resources listing possible events and corpora have
been annotated with this kind of information.
Examples of such datasets are FrameNet
Unfortunately, the vast majority of annotated
datasets relies mainly on newswire and narrative
texts, and their coverage turns out to be inadequate
when it comes to annotate more specific domains,
such as, for instance, football domain
Aside from that, over the last decades, ICT technologies and communication habits underwent profound changes, with the greatest part of text production in the world coming from social networks and being usually written in non-standard language.2 This kind of communication is of fundamental importance, in particular for teenagers’ social life. For instance, according to the last report by the Italian Statistical Institute (ISTAT, 2014) in Italy 82.6 of children aged 11-17 use the mobile phone every day. The use of these new technologies, however, leads also to some undesirable side effects, as the proliferation of hate speech and the digitization of traditional forms of harassment, also known as cyberbullying.
Many studies
In Section 5, we also provide a comparison study to highlight relations between the newlyreleased frame annotation and the existing one regarding the type of cyberbullying expression. Results show that some of them are strictly connected (even when it is not immediate to understand).
Finally, in Section 6 we present Framy, a frame annotation tool that works as a web server and that has been used for annotating BullyFrame. The work presented in this paper spans topics from different research areas. As for the methodology, we deal with issues related to the annotation of Italian texts with FrameNet and frame annotation on social media texts. Then, as case study, we focus on the cyberbullying domain, where we witness a growing interest and a large number of novel works over the last few years.
The FrameNet database is a resource originally developed for the English language that has proven to be largely portable over different languages. This because its frames appear to be mostly language independent, as pointed out by Gilardi and Baker (2018). Nevertheless, some language specific differences can arise both at the level of frames themselves (coarse-grained level) and at the level of frame elements (FEs) (finegrained level) (Lo¨nneker-Rodman, 2007). As an example it is possible to recall the works of Candito et al. (2014) on French, of Ohara (2012) on Japanese and of Subirats and Sato (2004) on Spanish. In all the three languages the creation of a FrameNet-like resource required to add new frames or FEs or modify already existing ones, for instance in French some frames needed to be merged, while others needed to be split into two subframes.
For the Italian language, we rely on
previous researches, carried out at the Universities of
Bologna and Roma Tor Vergata
Gerrard et al. (2017) outline how frame
annotation of texts extracted from social media could
be challenging because of the differences between
social media data and the kind of data on which
FrameNet is built, i.e. edited and well-formed
sentences. For this reason as for today only few
studies annotated social media texts with frame
information
Works on cyberbullying try to detect and
prevent the phenomenon exploiting different
methodologies and techniques. In particular, a dataset
extracting data from Facebook has been developed at
University of Pisa
As for today, at the best of our knowledge, there are not research works that studied the possible interconnections between cyberbullying and frames. 3
For the annotation of the frames related to cyberbullying we use as starting point the dataset from Sprugnoli et al. (2018). The dataset presents a collection of WhatsApp chats written by 12-13 years old students simulating instances of cyberbullying in specific scenarios.
The text of the chats is provided with annotations about i) the role of who is writing (i.e. Victim, Bully, or supporter of one of the two sides) and ii) labels with the type of offense that can be found on each message (in particular, the labels include: Threat or blackmail, General Insult, Body Shame, Sexism, Racism, Curse or Exclusion, Insult Attacking Relatives, Harmless Sexual Talk, Defamation, Sexual Harassment, Defense, Encouragement to the Harassment, and Other).
The dataset consists of 10 chats, for a total of 2192 messages (14,600 tokens) and includes 1,203 cyberbullying expressions, corresponding to 6,000 tokens.
Starting from this, we fully annotated the sentences referring to FrameNet 1.7: the resulting annotation is available for download from the resource website.3 It is released under the Creative Commons Attribution-ShareAlike 4.0 International license.4
A total of 2,458 frames and 2,769 frame element have been annotated on 1,558 sentences. The remaining 1,211 sentences cannot be annotated, 3https://github.com/dhfbk/bullyframe 4https://creativecommons.org/licenses/ by-sa/4.0/ mainly because no corresponding frames can be found (1,180 sentences), or because there was a picture instead (19 sentences), or finally because the messages have been deleted by the user (12 sentences). Table 1 (a) shows statistics on how many frames have been annotated for each sentence. Regarding the coverage, a total of 268 unique frames and 696 unique frame elements have been found in the dataset. Table 1 (b) shows the most frequent frames that have been annotated. Finally, Table 1 (c) shows statistics on how many frame elements are annotated for each frame. 4
In order to investigate possible connections between frames and cyberbullying we annotated all the sentences of the dataset with frames and frame elements referring to the 1.7 version of FrameNet. In each sentence we tried to annotate all the possible evoked frames alongside with their frame elements.
When annotating the sentences we have to face some problems that, due to the nature of this dataset, to the differences between English and Italian, and to the nature of FrameNet itself, is not complete but that is constantly updated and enlarged.
Problematic aspects can be found on three different levels: Frames layer, Frame Elements layer and Frame Evoking Elements layer.
Frames layer: We found that some of the concepts that were evoked by lexical units (LUs) were not present in FrameNet. The missing frames could be: a) Concepts that are new to FrameNet and that are linked to the particular nature of the text. This is the case for instance of frames that occur often in conversations or in oral communication. These concepts are often not present in FrameNet, but frequent in our dataset since it includes interactions between participants and is close to oral communication. For example we found that FrameNet does not have a frame that covers “greetings”, evoked in sentences such as: “Ciao ci sentiamo domani” (Bye, we’ll talk tomorrow) “Hahahah esatto ciao e buon allenamento” (Hahahah, exactly bye and have a good training)
Sentences
Frequency
Frame b) Concepts that are new to FrameNet and that are linked to abusive language and cyberbullying. For example we found that bullies often refer to people’s sexual orientation as an insult such as in: “Crede di essere figo facendo il gay a danza” (He thinks he looks cool acting like a gay when he dances) “Manco fossi gay am I, gay? ) ” (What “Sei cos`ı effemminato che intorno a te ci sono piu` finocchi che in un orto” (You are so effeminate that around you there are more pansies than in a garden) However, a frame that covers this concept is missing in FrameNet. c) Concepts that are new to FrameNet, but that are not specifically linked to the nature of the text nor to abusive language or cyberbullying. For example in FrameNet are missing frames related with ”sports” and similar activities: “Anche tu fai calcio” (You play football as well) “S`ı e tu vai a giocare a rugby” (Yes, and you go play rugby) “Lui non fa danza classica” (He does not do ballet) d) Concepts that are not new to FrameNet corresponding to holes in the FrameNet hierarchy. For example FrameNet has a frame for Silencing, a frame for Becoming silent but it does not have a frame for Being silent.
frames were missing but that it was also possible to find missing FEs.
For example it appears to be missing the FE Reason for the frame Statement, useful for annotating sentences such as: “Lo diciamo per il tuo bene” (We say that for your own sake) here “Per il tuo bene” (For your own sake) expresses the motivation for which the speaker makes his statement and could be labeled as Reason.
Another example can be the frame Ingestion for which a FE Quantity, for annotating the quantity of the ingestibles eaten, appears to be missing. For example, in the sentence: “Non mangiare tanto o diventi ancora piu` obeso” (Do not eat a lot or you will get even fatter) the FE label Quantity would be perfectly fitting for annotating the adverb “tanto” (a lot). Frame-Evoking Elements layer: Problems linked to the fact that in the sentences we tagged we find that not only words or multiword expressions (MWEs) evoke frames but that also other elements. In particular we found that frames can be evoked also by: a) Constructions: For example in the sentences “Di sicuro un cane e` piu` bravo di lui”(A dog is better than him for sure) or “Noi siamo piu` forti di te”(We are stronger than you) the frame Surpassing is evoked by the construction “essere piu` X di Y”(To be Xer than Y)” rather than by a word or a multiword expression.
“Ma tu sei gia` una are already a ) ” (But you the “Pile of Poo” emoji evokes the frame Desirability.
Aside from these three problematic layers, we found that for a considerable amount of messages it was not possible to add any frame annotation because of problems of different nature. More specifically we found that: a) Some messages are only made of punctuation marks, mostly ellipsis, exclamation points and question marks. b) Some messages are made of interjections or discourse markers and it is, thus, not possible to identify any frame evoking element: “Oooooooooooooooooo ” “Ahahahahahahahahahahahahah” c) In some other cases there are sentences that have been split into two or more messages. In these cases it is often possible to find messages in which no frame is evoked, but that constitute a FE of a frame evoked in the bigger sentence that has been split.
For example, the sentence: “Ma noi verremmo con i nostri bei cori” (But we would come with our nice chant) has been split into two different messages “Ma noi verremmo” (But we would come) and “Con i nostri bei cori” (With our nice chants). The first message can be annotated with the frame Arriving while the second message could only be annotated as the Arriving frame element Depictive. The sentence: “Neanche hai capito che e` una citazione di Battiato ” (You didn’t even understand that this is a quote from Battiato) have been split into “Neanche hai capito che e` una citazione”(You didn’t even understand that it is a quote) and “Di Battiato” (From Battiato). In the first message, the LU “capire.v”(understand.v) evokes the frame Awareness, and “Che e` una citazione” (That it is a quote) instantiates its frame element Content, whereas the second message can only be considered as a part of it. d) Some messages contain only affermative and negative expressions, i.e “Yes” or “No”. e) Other messages only repeat a word or a group of words of the previous message or anticipate one word or a group of words that will be part of the subsequent message: “Tu”, “Tu che sei un maschio” (You, You that are a boy) f) Finally there are messages that only aim to correct a word or a letter previously misspelled: “Ai scritto”, “*Hai” (You wrote) “Bravo Bul”, “*Bullo” (Good bully)
A field that is particularly relevant is the semantic field of emotions. We found that FrameNet frames referring to this field have sometimes fuzzy boundaries and that it is sometimes hard to choose a frame over another. Moreover there are also some frames that seem to be missing: for example in FrameNet there is no frame that covers the concept of “Expressing emotions” evoked by LUs such as “weep.v” or “cry.v” or “laugh.v”. Indeed, the first is completely missing in FN, the second is present as evoking Make noise, Communication noise and Vocalization, the third in present only as evoking Make noise. 5
In order to highlight significant relations between
frames and cyberbullying, we compared the frame
annotation with the already existing annotation
regarding the type of cyberbullying expression (see
Section 3). In particular we computed their
correlation using the weighted mutual information.
This kind of evaluation can be useful, for
instance, to predict cyberbullyng conversations
using tools that automatically extract semantic
information with respect to frames, such as SEMAFOR
The results, reported in Table 2, show some interesting outcomes. Most of them are in line with what we could have expected, but some others instead reflect the limitations of FrameNet in the annotation of this kind of interactions. For example we can see that “General insult” is related with frames such as Mental property or Desirability, this well matches with the intuitions that those frames capture respectively expressions which denigrates the interlocutor by referring to his/her lower intelligence, e.g. “Idiota” or “Stupida” (“Idiot”, “Stupid”), or to his/her scarce desirability, e.g. “Sfigato’’ (“Loser/Lame”). The same can be said for the pairs “Treat or Blackmail” - Cause harm and “Insult-BodyShame” - Aesthetics, where the connection between the frame and the cyberbullying type appears to be straightforward. Nevertheless there are also pairs if which the connection is hard to understand. For example “Encouragment to the Harasser” shows a strong relation with the frame Correctness. This is due, once again, to the limitations of FrameNet that lacks of some frames, in this particular case it lacks of a frame for the expressions that indicate a reinforcement of what one of the interlocutors just said such as “Esatto” (“Exactly”) or “Hai ragione” (“You are right”) that are now listed under the frame Correctness.
Bullying annotation Frame wMI Curse or Exclusion Silencing 0.0672 General Insult Desirability 0.0304 General Insult Mental property 0.0227 Encourage Harasser Correctness 0.0177 Curse or Exclusion Desiring 0.0135 Threat or Blackmail Cause harm 0.0127 Discrimination-Sexism Suitability 0.0083 Curse or Exclusion Required event 0.0080 General Insult Silencing 0.0065 Insult-BodyShame Aesthetics 0.0046 The annotation on FrameNet has been performed using a tool called Framy, developed at Fondazione Bruno Kessler and freely available on Github5 under the Apache 2.0 license. It is written in php and needs a MySQL database to work.
The application is optimized for frame semantics annotation, and can be configured to work with every version of FrameNet. After loading the already tokenized text data using the included scripts, a human annotator can select both the lexical unit that evokes the frame and the frame elements relative to the selected words. 7
In this paper, we present and release BullyFrame, an Italian resource consisting in a set of WhatsApp chats with full-text FrameNet annotations. The data, freely accessible on GitHub, increases the availability of resources in Italian. We also discuss how FrameNet lacks certain frames, as it cannot cover some expressions used mainly in the social media language. Finally, we describe Framy, a free tool that supports the manual annotation of texts w.r.t. FrameNet.
In the future, we want to extend this dataset by including other text resources, and extend FrameNet coverage for the social media domain, to deal with informal expressions and emojis.
This work has been supported by the European
Commission project Hatemeter
Stefano Menini, Giovanni Moretti, Michele Corazza, Elena Cabrio, Sara Tonelli, and Serena Villata. 2019. A System to Monitor Cyberbullying based on Message Classification and Social Network Analysis. In Proceedings of the third Workshop on Abusive Language Online, Florence, Italy.
Kyoko Ohara. 2012. Semantic Annotations in Japanese FrameNet: Comparing Frames in Japanese and English. In LREC, pages 1559–1562. Citeseer. Mona O’Moore and Colin Kirkham. 2001. Selfesteem and its relationship to bullying behaviour.
Aggressive behavior, 27(4):269–283.
Martha Palmer, Daniel Gildea, and Paul Kingsbury. 2005. The proposition bank: An annotated corpus of semantic roles. Computational linguistics, 31(1):71–106.