=Paper=
{{Paper
|id=Vol-3033/paper32
|storemode=property
|title=Frame Semantics for Social NLP in Italian: Analyzing Responsibility Framing in Femicide News Reports
|pdfUrl=https://ceur-ws.org/Vol-3033/paper32.pdf
|volume=Vol-3033
|authors=Gosse Minnema,Sara Gemelli,Chiara Zanchi,Viviana Patti,Tommaso Caselli,Malvina Nissim
|dblpUrl=https://dblp.org/rec/conf/clic-it/MinnemaGZPCN21
}}
==Frame Semantics for Social NLP in Italian: Analyzing Responsibility Framing in Femicide News Reports==
https://ceur-ws.org/Vol-3033/paper32.pdf
Frame Semantics for Social NLP in Italian:
Analyzing Responsibility Framing in Femicide News Reports
Gosse Minnema1 , Sara Gemelli2 , Chiara Zanchi2 ,
Viviana Patti3 , Tommaso Caselli1 , Malvina Nissim1
1. University of Groningen, The Netherlands
2. University of Pavia, Italy
3. University of Turin, Italy
{g.f.minnema,t.caselli,m.nissim}@rug.nl, chiara.zanchi01@unipv.it,
sara.gemelli01@universitadipavia.it, patti@di.unito.it
Abstract Das et al., 2014), which has seen considerable re-
cent work on English (Swayamdipta et al., 2017;
We propose using a FrameNet-based ap- Yang and Mitchell, 2017; Peng et al., 2018; Jiang
proach for analyzing how socially relevant and Riloff, 2021), there has not been any published
events are framed in media discourses. work on Italian since the EVALITA-2011 shared
Taking femicides as an example, we per- task (Basili et al., 2013). Second, a clear perspec-
form a preliminary investigation on a large tive on how computational frame semantics can be
dataset of news reports and event data cov- useful in real-life applications is still missing.
ering recent femicides in Italy. First, we We aim to advance the practical usability of
revisit the EVALITA 2011 shared task on frame semantics in Italian NLP in two ways. First,
Italian frame labeling, and test a recent we test how well a recently developed multilin-
multilingual frame semantic parser against gual model (LOME, Xia et al. (2021)) for FSP per-
this benchmark. Then, we experiment forms on Italian. For this purpose we use existing
with specializing this model for Italian and data from the EVALITA 2011 campaign, which is
perform a human evaluation to test our the only reference for Italian on FSP, as well as
model’s real-world applicability. We show new “real world” data collected in the context of
how FrameNet-based analyses can help to the socially relevant domain of femicides. Sec-
identify linguistic constructions that back- ond, we show how frame semantics can be used in
ground the agentivity and responsibility of practice to run analysis on real world data. From
femicide perpetrators in Italian news. both efforts, we draw some recommendations for
practical developments in Italian FSP.
1 Introduction
2 Semantic Frames for Events in Society
Frame semantics (Fillmore, 1985; Fillmore, 2006)
is a theory of natural language understanding with Frame semantics assumes that lexical units are
a focus on word meanings (lexical units) and se- points of access to complex conceptual structures:
mantic roles (frame elements). The associated understanding the meaning of a word means to
FrameNet project (Baker et al., 2003) has resulted understand all of the knowledge that is associ-
in an extensive lexicon and annotated corpus im- ated with it. Every semantically loaded lexical
plementing this theory. In the Italian computa- item evokes a frame, a scenario-like unit of en-
tional linguistics community, there has also been cyclopedic knowledge describing the concept as-
considerable work on frame semantics, mostly fo- sociated to it. Frame semantics also describes the
cused on creating FrameNet resources (Tonelli and perspective in which the frame is seen. A clas-
Pianta, 2008; Tonelli et al., 2009; Lenci et al., sical example is that of a commercial transaction
2010; Basili et al., 2017; Brambilla et al., 2020). (Fillmore, 1971), where the same event can be
However the practical usability of frame seman- presented either by foregrounding the buyer (e.g.,
tics for Italian is still largely unexplored. First of “Mary bought a book (from John)”) or the seller
all, on automatic frame semantic parsing (FSP) (e.g., “John sold a book (to Mary)”). Perspec-
(Gildea and Jurafsky, 2002; Baker et al., 2007; tivization can be also related to syntactic construc-
tions: an active sentence (“Mary bought a book”)
Copyright © 2021 for this paper by its authors. Use per-
mitted under Creative Commons License Attribution 4.0 In- and a passive one (“The book has been bought”)
ternational (CC BY 4.0). denote the same event, but make us access it via
two different participants (Meluzzi et al., 2021). emplars containing a single annotated predicate
It has been shown that the variability of linguis- and frame structure. Compared to the English
tic expressions used to describe an event impacts Berkeley FrameNet (BFN), which contains also
the reader’s perception of the event and its social fully annotated documents, the models presented
significance. Previous work in psycholinguistics at FLAIT are impressive (scores up to 80%).
shows that in events involving violence (at any
3.1 LOME experiments
level), the linguistic backgrounding of agents hin-
ders their responsibility and promote victim blam- LOME (Xia et al., 2021) is a recent end-to-end
ing (Huttenlocher et al., 1968; Bohner, 2001; Gray FSP model that reports excellent frame detection
and Wegner, 2009; Zhou et al., 2021; Meluzzi scores on English, and, thanks to its XLM-R en-
et al., 2021). For instance, Te Brömmelstroet coder (Conneau et al., 2020), is the first cross-
(2020) shows that media in the Netherlands fre- lingual FSP model, even though it was trained on
quently report on traffic crashes by foregrounding English data only. Here, we propose several strate-
the more vulnerable participants (e.g., pedestrians gies for adapting LOME to Italian and making
or cyclists), while backgrounding car drivers. A maximum use of the available data.
similar pattern has been observed for news reports Strategies The simplest strategy, LOME-EN, is
of femicides in Italy, where the victim tends to to use the English-trained model in a zero-shot
be foregrounded and the perpetrator backgrounded setup to make predictions for Italian texts. A
(Pinelli and Zanchi, 2021; Meluzzi et al., 2021). downside of this approach is that the model is not
While there have been some proposals to use able to tag the Italian-specific frames that have
frame semantics for analyzing media framing or been created in the IFrameNet project (Basili et
applying it to social media texts (Ziem et al., 2018; al., 2017), which also makes the evaluation on
Brambilla et al., 2019), we are not aware of previ- FLAIT data more challenging. FLAIT contains 10
ous work that applies frame semantics to the study frames that do not currently exist in BFN (7.4%
of linguistic perspectivization of societal issues. of training instances and 6.0% of test instances).
We test this idea and present a preliminary analysis It therefore makes sense to also train LOME on
of how frames and syntactic constructions are used FLAIT directly. In IT-Simple, we only train on
to perspectivize violence in a large corpus of femi- FLAIT data; in IT-Concat, we train on the con-
cide reports in the Italian press. We adopt the data- catenation of FLAIT and the fully annotated doc-
to-text approach to FrameNet analysis (Vossen et uments from BFN; and in IT-Berkeley, we train
al., 2020; Remijnse and Minnema, 2020; Remi- only on FLAIT but initialize the encoder with the
jnse et al., 2021), where structured event metadata parameters of LOME-EN.
is linked to texts referencing real-world events.
A crucial part of this method is defining typical Evaluation For use in real-life applications,
frames, i.e., frames that are hypothesized to con- what truly matters is end-to-end performance, i.e.
ceptualize important aspects of the targeted event from raw texts to the predictions of all predi-
type. For the femicide domain, we selected 15 typ- cate frames and associated roles. Full end-to-end
ical frames;1 some examples are in Table 1. evaluation is impossible in FLAIT since only one
predicate per sentence is annotated. However, we
3 Frame Semantic Parsing for Italian can approximate it by obtaining the full predic-
tions from the models and then evaluate only on
The shared task on Frame Labeling over Italian FLAIT gold predicates. In this way, models are
Texts (FLAIT) at EVALITA 2011 (Basili et al., penalized for missing predicates that should have
2013) introduce the only existing published Ital- been annotated (but not for overgeneration). We
ian FSP models, as well as the only publicly avail- use the SeqLabel metric (Minnema and Nissim,
able corpus for the task on generic texts. As shown 2021) for scoring frame and role label predictions
in Table 2, the FLAIT corpus contains 1,569 an- on a token-by-token basis.
notated sentences, all of which are so-called ex- Additionally, to test LOME against the 2011
1
A BUSING , ATTACK , C AUSATION , C AUSE HARM , models, we reimplement the FLAIT evaluation
C AUSE MOTION , D EAD OR ALIVE , D EATH , E MO - metrics, in which models are asked to predict (i)
TION DIRECTED , E VENT, E XPERIENCE BODILY HARM ,
H IT TARGET, K ILLING , Q UARRELING , R APE , frames given a predicate (Frame Detection [FD]),
U SE FIREARM. (ii) semantic role spans given a frame (Boundary
Frame Description Example
an agent (Killer) actively causes the
KILLING [The man] killed [his wife] English Italian
death of a patient (Victim)
DEATH someone (Protagonist) dies [The woman] died sentences 5,093 0
fulltext
state of someone (Protagonist) being frame instances 29,359 0
DEAD OR ALIVE [She] was found dead
dead or alive sentences 163,801 1,569
an agent (Agent) actively causes a
exemplar
CAUSE HARM [He] stabbed [his girlfriend] frame instances 169,473 1,569
patient (Victim) to be hurt
sentences 168,894 1,569
[The dramatic events] happened total
EVENT an unspecified event (Event) happens frame instances 198,832 1,569
last week
Table 1: Examples of FrameNet frames relevant for describing Table 2: Sentences and annotations
femicides. Semantic role names indicated in italics, lexical in the English and Italian datasets.
units indicated in bold.
frames roles ble 44 we compare LOME against the best system
P R F P R F
from 2011, which is based on a SVM with a tree
EN LOME-EN 0.89 0.70 0.78 0.69 0.59 0.64 kernel (Croce et al., 2013). The most striking re-
LOME-EN 0.63 0.52 0.57 0.63 0.50 0.56 sult is that, on frame prediction, the 2011 winner
IT
IT-Simple -0.14 0.14 -0.01 -0.14 0.16 0.00 is still king, with the LOME-EN and IT-Concat
IT-Concat 0.21 0.14 0.17 0.10 0.08 0.09 models falling short by 0.24 and 0.04 points, re-
IT-Berkeley -0.07 0.17 0.05 0.04 0.12 0.09
spectively. For semantic role prediction, results
Table 3: SeqLabel scores for gold predicates. are mixed: LOME-EN has a modest but consistent
Blue: baseline, green/red: performance deltas improvement on both span (BD) and label (AC)
prediction, while IT-Concat improves on some se-
tups but not on others.
Detection [BD]), or (iii) semantic role labels given
a frame and the role spans (Argument Classifica- 3.2 Evaluating Real-World Performance
tion [AC]).2 We explore how robust are our models when de-
ployed on other data. We focus on frame predic-
Implementation We kept LOME model and
tion only, a task know to be harder to adapt across
training settings the same as described by Xia et
domains (Hartmann et al., 2017)
al. (2021). During testing, we noticed that 56 in-
stances in the FLAIT test set had misspelled frame Femicide annotation We deployed the LOME-
labels,3 causing a large drop in scores. We fixed EN and IT-Concat on a set of femicide news re-
these labels, but since we do not know if the orig- ports (see §4) with typical frames (see §2) in an
inal evaluation script also did this, we report the end-to-end setup (i.e., without predicates as in-
uncorrected scores in our GitHub repository. put). Out of 4,444 frame predictions, the two
models disagreed in 58% of cases. Next, for
Results Sequence labeling performance is re- a subset of 150 conflicts, we manually anno-
ported in Table 3. The zero-shot LOME-EN model tated5 which of the two predictions is better.
achieves an F1 score of 0.57 for frames and 0.56 Table 6 shows that LOME-EN performs much
for roles, substantially less than IT-Concat, which better than IT-Concat, especially on two of the
gets close to scores on English (0.74 F1 on frames, most frequent typical frames (K ILLING and E MO -
0.63 on roles). The other two Italian models have TION DIRECTED ). This is largely due to predi-
mixed results, with improvements on recall but not cate detection: 47% of cases where LOME-EN
on precision. However, IT-Berkeley outperforms is better than IT-Concat are due to IT-Concat not
both LOME-EN and IT-Simple, showing that re- detecting the predicate; in conflicts for predicates
using encoder weights helps boost performance. that both models detected, IT-Concat slightly out-
Turning to EVALITA-style evaluation, in Ta-
4
We only report strict scores for BD and AC. Full tables
2
As we were unable to access the original evaluation with token-based scores are in our GitHub repository.
script, we have attempted to reproduce it as faithfully as pos- 5
Annotation was done by a single annotator, who is also
sible from the description in Basili et al. (2013). one of the co-authors of this paper. Annotation was blind and
3
In these frame names, dashes were used in place of un- randomized, i.e., the annotator had no way to guess which
derscores, e.g. C AUSE - HARM instead of C AUSE HARM. prediction came from which model.
run 1 run 2 run 3 frames
P R F P R F P R F all IFN BFN fcd
FD FLAIT/dev
2011-best 0.81 0.81 0.81 - - - - - - num examples 123 123 113 14
LOME-EN -0.24 -0.24 -0.24 - - - - - - - Simple SVM 0.59 0.59 0.60 0.71
IT-Concat -0.04 -0.04 -0.04 - - - - - - LOME-EN 0.59 0.59 0.65 0.71
BD (strict) IT-Concat 0.85 0.85 0.87 0.93
2011-best 0.67 0.73 0.69 0.67 0.73 0.69 - - - FLAIT/test
LOME-EN 0.10 0.05 0.08 0.02 0.07 0.05 - - - num examples 318 318 299 43
IT-Concat -0.09 -0.06 -0.08 -0.10 -0.06 -0.08 - - - Simple SVM 0.29 0.29 0.30 0.40
AC (strict) LOME-EN 0.57 0.57 0.60 0.60
2011-best 0.48 0.53 0.50 0.51 0.56 0.53 0.70 0.70 0.70 IT-Concat 0.77 0.77 0.76 0.81
LOME-EN -0.01 0.02 0.01 0.09 0.13 0.11 0.16 0.16 0.16 femicides
IT-Concat -0.02 0.00 -0.01 -0.03 0.01 -0.01 0.14 0.14 0.14 num examples 43 43 43 43
Simple SVM 0.14 0.14 0.14 0.14
LOME-EN 0.63 0.63 0.63 0.63
Table 4: EVALITA-2011-style evaluation. As in the original IT-Concat 0.72 0.72 0.72 0.72
task, run 1, 2, and 3 refer to predictions with, resp., no gold
inputs, gold frame inputs, and gold frame and role span inputs. Table 5: Generalizability scores
best prediction and typical frames for femicides (‘fcd’).
EN IT both none
The results show several patterns that are rele-
overall 0.51 0.12 0.12 0.25 vant for real-world usability. First, both LOME
non-null 0.17 0.22 0.44 0.17
by frame models perform as good or better on typical femi-
KILLING 0.70 0.19 0.11 0.00 cide frames compared to other frames, which is
EMOTION D . 0.77 0.05 0.05 0.14 a positive sign for the feasibility of our project.
DEATH 0.33 0.05 0.19 0.42
Furthermore, IT-Concat is clearly the overall best
frame detection model, but only when it already
Table 6: Conflict analysis on the femicides dataset.
knows which predicates to annotate (see above).
‘EN’: LOME-EN; ‘IT’: IT-Concat; ‘both’/‘none’:
However, it is also quite biased towards the FLAIT
both models are equally correct/wrong.
dataset, scoring substantially worse on the test and
femicide datasets compared to the development
performs LOME-EN. We speculate that this might set. By contrast, LOME-EN is very stable across
be explained by the exemplar-style structure of the datasets. The SVM baseline performs surprisingly
FLAIT corpus. well on the development set, but much worse on
the test set and extremely poorly on the femicides
Generalization Table 5 shows frame detection dataset. We interpret this as a sign of the limited
scores on three evaluation sets: the FLAIT devel- coverage of the FLAIT dataset, showing that good
opment set (10% held-out from the training set), performance on the shared task is not necessarily
the FLAIT test set, and the set of cases from our indicative of real-world performance.
femicide annotation experiment in which at least
one of the two models’ predictions was marked as 4 Frame-Based Analysis of Femicide
correct.6 Since we do not have access to the origi-
News
nal FLAIT models, we use a simple linear SVM,7
trained on FLAIT, as an alternative baseline. The
In this section, we provide a concise overview of
task is the same as the FLAIT FD task: the mod-
our initial work on applying frame semantic pars-
els are given the gold predicate and asked to pre-
ing to investigate news coverage of femicides.
dict the frame. Results are split by frame cate-
gory: IFrameNet frames that FLAIT-trained mod-
els can be expected to know (‘IFN’), BFN frames Dataset We perform our analysis on a private
that LOME-EN can be expected to know (‘BFN’), dataset collected by the CRITS research team at
RAI (Radiotelevisione Italiana) and made avail-
6
If the annotator indicated that both predictions for a par- able for use in our project. The dataset contains
ticular predicate were equally good, we randomly selected 2,734 news articles from 31 different Italian news
one of the predictions as the ‘gold’ label.
7
The SVM takes as input a bag-of-bigrams extracted from sources, reporting on 937 femicides perpetrated
a context window of 5 tokens before and after the predicate. between 2015 and 2017, along with structured in-
Figure 1: Typical frame frequencies, split by syntactic construction
formation about these femicides (Belluati, 2021)8 . role, while only 33% express a Killer role. How-
The dataset is unique because it includes rich event ever, instances with a nonverbal construction only
metadata, and contains various news article per express these roles in 40% and 20% of cases, re-
femicide, allowing for investigating variation in spectively, against 71% and 87% in active con-
framing of the same event along different dimen- structions. On the other hand, D EATH expresses
sions, e.g., over time or by news source. a victim-like role (Protagonist) in 79% of cases,
whereas its only role that can encode a perpetrator
Analysis Based on our findings in §3, especially (Explanation) occurs in 14% of cases.
from the human evaluation experiment, we deploy While our analysis is too preliminary to draw
the LOME-EN model to automatically annotate a strong conclusions, our findings are consistent
randomly chosen 200K word subcorpus covering with previous work: agentivity-backgrounding
10% of all events.The frame semantic annotations constructions (especially nonverbal) are very com-
are enriched with dependency parses produced by mon, and semantic roles encoding the victim are
spaCy (Honnibal et al., 2020), which are con- more frequent than those encoding the perpetra-
verted into syntactic construction annotations us- tor. What our frame analysis adds to previous
ing a set of heuristics. work is information about the semantics of the
Figure 1 shows our main results. K ILLING is analyzed constructions. For example, the domi-
by far the most frequent typical frame, followed nance of K ILLING suggests that femicides tend to
by E MOTION DIRECTED and D EATH. Looking be framed as agentive at least on a lexical level,
at syntax, we find that nonverbal constructions, in even if the perpetrator is often backgrounded syn-
which the predicate is expressed by a noun or ad- tactically. On the other hand, non-agentive ways
jective (e.g., “l’omicidio” “the murder”) are dom- of framing the event (D EATH , D EAD OR ALIVE ,
inant in many frames. Instead, verbal:active con- E VENT) are also relatively common, accounting
structions (e.g., “X uccide Y” “X kills Y”) are for 24% of frame instances.
much rarer, as are verbal:passive (e.g., “X è uc-
cisa” “X is killed”) and verbal:unaccusative (e.g., 5 Conclusions
“X è deceduta” “X has died”).
Looking at semantic roles, patterns that vary We took initial steps towards addressing (i) the
greatly depending on frames and constructions. In lack of recent frame semantic parsing models, and
general, semantic roles that are likely to refer to (ii) a missing perspective on how frame semantic
the perpetrator appear to be expressed much less analysis can be applied in practice. We adapted
frequently than those referring to the victim. For the multilingual LOME parser (Xia et al., 2021)
K ILLING, 60% of all instances express a Victim to Italian, tested it against the EVALITA-2011
8
benchmark, and performed experiments to evalu-
The dataset has been collected as an outcome of the
PRIN 2015 research project Rappresentazioni sociali della ate its real-world performance. Furthermore, we
violenza sulle donne: il caso del femminicidio in Italia. hypothesize that frame semantics can be a valu-
able analysis tool for analyzing backgrounding References
(and indirectly, blame attribution) of event partic- Collin F. Baker, Charles J. Fillmore, and Beau Cronin.
ipants, and propose news reports about femicides 2003. The structure of the FrameNet database.
as an example of a domain where this type of anal- International Journal of Lexicography, 16(3):281–
ysis is very socially relevant. 296.
Our results indicate that LOME-based mod- Collin Baker, Michael Ellsworth, and Katrin Erk.
els can achieve acceptable performance, both on 2007. SemEval-2007 task 19: Frame semantic
structure extraction. In Proceedings of the Fourth
the EVALITA benchmark and out-of-domain on International Workshop on Semantic Evaluations
femicide reports, even without a large quantity of (SemEval-2007), pages 99–104, Prague, Czech Re-
training data. We also found that a cross-lingual public, June. Association for Computational Lin-
approach is useful: training on the concatena- guistics.
tion of English and Italian data yields substantial Roberto Basili, Diego De Cao, Alessandro Lenci,
improvements over using only Italian data, and Alessandro Moschitti, and Giulia Venturi. 2013.
even a zero-shot approach with only English data Evalita 2011: The frame labeling over Italian
texts task. In Bernardo Magnini, Francesco Cu-
works quite well. However, our real-world perfor- tugno, Mauro Falcone, and Emanuele Pianta, edi-
mance analysis highlights key limitations of the tors, Evaluation of Natural Language and Speech
Italian data: while models trained on EVALITA Tools for Italian, pages 195–204, Berlin, Heidel-
can achieve good frame detection performance, berg. Springer.
they fail when used ‘end-to-end’, with predicate Roberto Basili, Silvia Brambilla, Danilo Croce, and
identification seemingly the main bottleneck. Fabio Tamburini. 2017. Developing a large scale
FrameNet for Italian: the IFrameNet experience. In
Finally, we performed a preliminary framing Roberto Basili, Malvina Nissim, and Giorgio Satta,
analysis of a large dataset covering femicides in editors, Proceedings of the Fourth Italian Confer-
Italy. While our analysis method is still in very ence on Computational Linguistics (CLiC-it 2017),
early stages, we believe that our initial results Rome, Italy, December 11-13, 2017, volume 2006 of
CEUR Workshop Proceedings. CEUR-WS.org.
demonstrate that frame semantics is meaningful
for analyzing femicides and other social issues, M. Belluati. 2021. Femminicidio. Una lettura tra re-
and that it complements earlier construction-based altà e interpretazione. Biblioteca di testi e studi.
approaches. In the future, we aim to expand our Carocci.
analysis system to make it usable for different so- Gerd Bohner. 2001. Writing about rape: Use of the
cial applications: for example, one could envision passive voice and other distancing text features as an
systems that can help social scientists test specific expression of perceived responsibility of the victim.
British Journal of Social Psychology, 40(4):515–
hypotheses about media reporting, help activists 529.
identify and highlight biased forms of reporting,
or help make journalists more aware of their writ- Silvia Brambilla, Alessio Palmero Aprosio, and Ste-
fano Menini. 2019. BullyFrame: Cyberbullying
ing and its possible social-cognitive effects. meets FrameNet. In Raffaella Bernardi, Roberto
Navigli, and Giovanni Semeraro, editors, Proceed-
ings of the Sixth Italian Conference on Compu-
Acknowledgements tational Linguistics, Bari, Italy, November 13-15,
2019, volume 2481 of CEUR Workshop Proceed-
ings. CEUR-WS.org.
We would like to thank the CRITS department
at RAI for giving us access to the femicides Silvia Brambilla, Danilo Croce, Fabio Tamburini, and
Roberto Basili. 2020. Automatic induction of
dataset. We would also like to thank our col- FrameNet lexical units in Italian. In Johanna Monti,
laborators in the broader responsibility framing Felice Dell’Orletta, and Fabio Tamburini, editors,
research effort that this work is part of: Mar- Proceedings of the Seventh Italian Conference on
ion Bartl, Gaetana Ruggiero, Marco te Brömmel- Computational Linguistics, CLiC-it 2020, Bologna,
Italy, March 1-3, 2021, volume 2769 of CEUR
stroet, and Eva Kwakman. Authors G.M., T.C., Workshop Proceedings. CEUR-WS.org.
and M.N. worked on this paper as part of the
project Framing situations in the Dutch language Alexis Conneau, Kartikay Khandelwal, Naman Goyal,
Vishrav Chaudhary, Guillaume Wenzek, Francisco
(code: VC.GW17.083/6215), funded by the Dutch Guzmán, Edouard Grave, Myle Ott, Luke Zettle-
National Science Organization (NWO). moyer, and Veselin Stoyanov. 2020. Unsupervised
cross-lingual representation learning at scale. In
Proceedings of the 58th Annual Meeting of the Asso- through semi-automatic corpus analysis. In Pro-
ciation for Computational Linguistics, pages 8440– ceedings of the Seventh International Conference
8451, Online, July. Association for Computational on Language Resources and Evaluation (LREC’10),
Linguistics. Valletta, Malta, May. European Language Resources
Association (ELRA).
Danilo Croce, Emanuele Bastianelli, and Giuseppe
Castellucci. 2013. Structured kernel-based learning Chiara Meluzzi, Erica Pinelli, Elena Valvason, and
for the frame labeling over Italian texts. In Eval- Chiara Zanchi. 2021. Responsibility attribution in
uation of Natural Language and Speech Tools for gender-based domestic violence: A study bridging
Italian, pages 195–204, Berlin, Heidelberg. Springer corpus-assisted discourse analysis and readers’ per-
Berlin Heidelberg. ception. Journal of Pragmatics, 185:73–92.
Dipanjan Das, Desai Chen, André F. T. Martins, Gosse Minnema and Malvina Nissim. 2021. Breed-
Nathan Schneider, and Noah A. Smith. 2014. ing Fillmore’s chickens and hatching the eggs: Re-
Frame-semantic parsing. Computational Linguis- combining frames and roles in frame-semantic pars-
tics, 40(1):9–56, March. ing. In Proceedings of the 14th International Con-
ference on Computational Semantics. https:
Charles J. Fillmore. 1971. Subjects, speakers, and //iwcs2021.github.io/proceedings
roles. Synthese, 21(3-4). /iwcs/pdf/2021.iwcs-1.15.pdf.
Charles J. Fillmore. 1985. Frames and the semantics Hao Peng, Sam Thomson, Swabha Swayamdipta, and
of understanding. Quaderni di semantica, 6(2):222– Noah A. Smith. 2018. Learning joint semantic
254. parsers from disjoint data. In Proceedings of the
2018 Conference of the North American Chapter of
Charles J. Fillmore. 2006. Frame semantics. In the Association for Computational Linguistics: Hu-
D. Geeraerts, editor, Cognitive Linguistics: Ba- man Language Technologies, Volume 1 (Long Pa-
sic Readings, pages 373–400. De Gruyter Mouton, pers), pages 1492–1502, New Orleans, Louisiana,
Berlin, Boston. Originally published in 1982. June. Association for Computational Linguistics.
Daniel Gildea and Daniel Jurafsky. 2002. Automatic Erica Pinelli and Chiara Zanchi. 2021. Gender-based
labeling of semantic roles. Computational Linguis- violence in italian local newspapers: How argument
tics, 28(3):245–288. structure constructions can diminish a perpetrator’s
responsibility. Discourse Processes between Reason
Kurt Gray and Daniel M. Wegner. 2009. Moral type- and Emotion: A Post-disciplinary Perspective, page
casting: divergent perceptions of moral agents and 117.
moral patients. Journal of Personality and Social
Levi Remijnse and Gosse Minnema. 2020. Towards
Psychology, 96:505–520.
reference-aware FrameNet annotation. In Proceed-
Silvana Hartmann, Ilia Kuznetsov, Teresa Martin, and ings of the International FrameNet Workshop 2020:
Iryna Gurevych. 2017. Out-of-domain FrameNet Towards a Global, Multilingual FrameNet, pages
semantic role labeling. In Proceedings of the 15th 13–22, Marseille, France, May. European Language
Conference of the European Chapter of the Associa- Resources Association.
tion for Computational Linguistics: Volume 1, Long Levi Remijnse, Marten Postma, and Piek Vossen.
Papers, pages 471–482, Valencia, Spain, April. As- 2021. Variation in framing as a function of temporal
sociation for Computational Linguistics. reporting distance. In Proceedings of the 14th Inter-
national Conference on Computational Semantics.
Matthew Honnibal, Ines Montani, Sofie Van Lan- https://iwcs2021.github.io/proce
deghem, and Adriane Boyd. 2020. spaCy: edings/iwcs/pdf/2021.iwcs-1.22.pdf.
Industrial-strength Natural Language Processing in
Python. Swabha Swayamdipta, Sam Thomson, Chris Dyer, and
Noah A. Smith. 2017. Frame-semantic parsing with
Janellen Huttenlocher, Karen Eisenberg, and Susan softmax-margin segmental rnns and a syntactic scaf-
Strauss. 1968. Comprehension: Relation between fold. CoRR, abs/1706.09528.
perceived actor and logical subject. Journal of Ver-
bal Learning and Verbal Behavior, 7:527–530. Marco Te Brömmelstroet. 2020. Framing systemic
traffic violence: Media coverage of dutch traffic
Tianyu Jiang and Ellen Riloff. 2021. Exploiting defini- crashes. Transportation research interdisciplinary
tions for frame identification. In Proceedings of the perspectives, 5.
16th Conference of the European Chapter of the As-
sociation for Computational Linguistics: Main Vol- Sara Tonelli and Emanuele Pianta. 2008. Frame
ume, pages 2429–2434, Online, April. Association information transfer from English to Italian. In
for Computational Linguistics. Proceedings of the Sixth International Conference
on Language Resources and Evaluation (LREC’08),
Alessandro Lenci, Martina Johnson, and Gabriella Marrakech, Morocco, May. European Language Re-
Lapesa. 2010. Building an Italian FrameNet sources Association (ELRA).
Sara Tonelli, Daniele Pighin, Claudio Giuliano, and
Emanuele Pianta. 2009. Semiautomatic Develop-
ment of FrameNet for Italian. In Proceedings of the
FrameNet Workshop and Masterclass, Co-located
with the Seventh International Workshop on Tree-
banks and Linguistic Theories.
Piek Vossen, Filip Ilievski, Marten Postma, Antske
Fokkens, Gosse Minnema, and Levi Remijnse.
2020. Large-scale cross-lingual language resources
for referencing and framing. In Proceedings of the
12th Language Resources and Evaluation Confer-
ence, pages 3162–3171, Marseille, France, May. Eu-
ropean Language Resources Association.
Patrick Xia, Guanghui Qin, Siddharth Vashishtha,
Yunmo Chen, Tongfei Chen, Chandler May, Craig
Harman, Kyle Rawlins, Aaron Steven White, and
Benjamin Van Durme. 2021. LOME: Large on-
tology multilingual extraction. In Proceedings of
the 16th Conference of the European Chapter of the
Association for Computational Linguistics: System
Demonstrations, pages 149–159, Online, April. As-
sociation for Computational Linguistics.
Bishan Yang and Tom Mitchell. 2017. A joint sequen-
tial and relational model for frame-semantic parsing.
In Proceedings of the 2017 Conference on Empiri-
cal Methods in Natural Language Processing, pages
1247–1256, Copenhagen, Denmark, September. As-
sociation for Computational Linguistics.
Karen Zhou, Ana Smith, and Lillian Lee. 2021. As-
sessing cognitive linguistic influences in the assign-
ment of blame. In Proceedings of the Ninth Inter-
national Workshop on Natural Language Processing
for Social Media, pages 61–69, Online, June. Asso-
ciation for Computational Linguistics.
Alexander Ziem, Christian Pentzold, and Claudia
Fraas. 2018. Medien-Frames als semantische
Frames: Aspekte ihrer methodischen und analytis-
chen Verschränkung am Beispiel der ‘Snowdon-
Affäre’ [Media frames as semantic frames: As-
pects of their methodological and analytical entan-
glement in the example of the Snowdon affair]. In
Detmer Wulf Alexander Ziem, Lars Inderelst, edi-
tor, Frames interdisziplinär: Modelle, Anwendungs-
felder, Methoden, pages 155–184, Düsseldorf. DUP.