English. In the domain of Natural Language Processing (NLP), the interest in figurative language is enhanced, especially in the last few years, thanks to the amount of linguistic data provided by web and social networks. Figurative language provides a non-literary sense to the words, thus the utterances require several interpretations disclosing the play of signification. In order to individuate different meaning levels in case of ironic texts detection, it is necessary a computational model appropriated to the complexity of rhetorical artifice. In this paper we describe our rulebased system of irony detection as it has been presented to the SENTIPOLC task of EVALITA 2016, where we ranked third on twelve participants.
Italiano. Nell’ambito del Natural Language Processing (NLP) l’interesse per il linguaggio figurativo è particolarmente aumentato negli ultimi anni, grazie alla quantità d’informazione linguistica messa a disposizione dal web e dai social network. Il linguaggio figurativo conferisce alle parole un senso che va oltre quello letterale, pertanto gli enunciati richiedono interpretazioni plurivoche che possano svelare i giochi di significato del discorso. Nel caso specifico del riconoscimento automatico di un testo ironico, infatti, determinare la presenza di diversi gradi di significazione esige un modello computazionale adeguato alla complessità dell’artificio retorico. In questo articolo descriviamo il nostro sistema “rule-based” dedito al riconoscimento dell’ironia che ha partecipato al task SENTIPOLC di EVALITA 2016, nel quale ci siamo classificati terzi su dodici partecipanti. the length of tweets is limited (140 characters), users are encouraged to use some creative devices in order to communicate their opinions. In particular they express their emotions or feelings through some morphosyntactic elements or conventional expedients, such as: emoticons, hashtags, heavy punctuation, etc. It seems that these elements represent a substitution of typical gestures and tones of oral communication. In this research we used some linguistic features, frequently found in ironic tweets, as referent points to create the rules of our irony detection system in Italian tweets.
The results we gained are promising and reveal the features considered can be good ironic clues to identify ironic texts.
In the following section we synthetically describe the state of art about irony detection. In the third and fourth sections we present our approach, describing the linguistic resources used and data processing. The fifth section contains the description of linguistic features, and finally in the sixth section we present the results obtained in SENTIPOLC evaluation. 2
Although the difficulties of research, it is evident in the literature an attempt to understand this linguistic phenomenon and develop some computational models to detect or generate irony.
In the 90s Lessard and Levison (1992, 1993)1 and Binsted and Ritchie (1994, 1997)2 developed the first joke generators and recently Stock and Strapparava (2006) realized HAHAcronym, a system designed to generate and re-analyze the acronyms, considering semantic opposition and rhythm criteria.
The research described by Utsumi (1996) was
one of the first approaches to automatic irony
processing, even though it was too abstract for a
computational framework. In 2009, Veale and
Hao noted that English figurative comparisons
(as X as Y) are often used to express ironic
opinions, especially when the marker “about” is
present (about as X as Y). Recently, Reyes et al.
(2013) produced a multidimensional model for
detecting irony on Twitter based on four
conceptual features: signatures (pointedness,
counterfactuality, and temporal compression),
unexpectedness (temporal imbalance and contextual
imbalance), style and emotional scenarios
Karoui et al. (2015) focused on the presence of
negation markers as well as on both implicit and
explicit opposition in French ironic tweets.
Moreover, this research highlights the
importance of surface traits in ironic texts, such as:
punctuation marks (González-Ibáñez et al.,
2011), sequence or combination of exclamation
and question marks
Carvalho et al. (2009) distinguished eight “clues” for irony detection in some comments (each consisting of about four sentences) from a Portuguese online newspaper. Their attention focused on positive comments because in a previous research they showed that positive sentences are more subjected to irony and it is more difficult to recognize their true polarity. So the idea is to identify the irony in apparently positive sentences that require the presence of at least one positive adjective or noun in a window of four words. Carvalho et al. (2009) based their model on both oral and gestural “clues” of irony, such as: emoticons, heavy punctuation, quotation marks, onomatopoeic expressions for laughter and positive interjections and, on the other hand, on specific morphosyntactic constructions, such as: the diminutive form of NE, the demonstrative determiners before NE, the pronoun “tu” specifi3Reyes et al. (2013: 249). 4Karoui et al. (2015). 5Karoui et al. (2015). 6Karoui et al. (2015). cally referred or embedded in the morphology of the verb “ser”.
Our work proposes an adaptation for some of these clues, increased by other surface features, to Italian irony detection in Twitter. 3
Approaching the detection of irony in tweets means to understand how people, especially net users, make irony. We try to approach this hard work by analyzing the corpus of tweets and identifying possible ironic clues. Once identified, surface features common to ironic tweets are inserted as binary rules in our system.
Our rule-based system, written in Perl, finds ironic features (described in section 5) in tweets and consequently distinguishes the ironic ones from the non-ironic.
In the following sections we describe resources used, data processing, ironic clues and the results obtained in the EVALITA 2016 SENTIPOLC task. 4
For this research we used a corpus of tweets
provided by SENTIPOLC organizers
For the analysis and processing of Italian tweets we used some linguistic resources available online, such as: • •
Sentiment Lexicon LOD (Linked Open Data). Developed by the Institute for Computational Linguistics “A. Zampolli”, it contains 24.293 lexical entries annotated with positive/negative/neutral polarity.
Morph-it!
A tweet is composed of different essential elements for linguistic analysis, as interjections and emoticons. We therefore developed a lexicon of interjections and a list of emoticons described summarily below: • •
The interjections, extracted from Morphit! and Treccani7, are manually annotated with their polarity. The annotation has been developed with the support of Vocabolario Treccani, while the sentiment lexicon has been used to label improper interjections (see Table 1).
The emoticons, extracted from Wikipedia, are subdivided in EMOPOS, EMONEG and EMOIRO, according to the classification of Di Gennaro et al. (2014) and Wikipedia description8, especially for the ironic annotation (see Table 2).
evviva
urrà complimenti congratulazioni
mah macché bah puah
boh
mhm
chissà
beh
=) =] :D (-: [-: (-; [-;
:-> :) :-) (; ;)
:[ =( :-( :'( :-/ :/ :-> :\> :/
=/ =\ :L =L :S
^^ ^.^ :P xP ^3^ ^L^ ^_^
^-^ ^w^
Incoming file processed by our system has been
previously lemmatized and syntactically
annotated by TreeTagger
Nevertheless, before syntactic analysis, we applied the rules of substitution and elimination of some textual elements, in order to clean up the texts and avoid hampering the process of POStagging and lemmatization of TreeTagger. In particular: the label EMOPOS replaces positive emoticons; 7http://www.treccani.it 8Wikipedia version of the 6th of June. the characters of url are removed.
This method allows us to clean up the texts from those characters that may hinder the analysis of data and ironic clues retrieval. 5
In section 2 we have presented the research of Carvalho et al. (2009) which demonstrated how the most productive patterns (with a precision from 45% to 85%) are the ones related to orality and gesture, as emoticons or expressions for laughter. Based on this analysis, we try to recognize ironic tweets with a system designed to find ironic clues into the texts. Some of these clues are adapted to Italian language from Portuguese, while some other features are individuated during the analysis of the tweets.
All of these features are used as binary rules in our system to classify the texts in ironic and nonironic. 5.1
Ameka (1992)9 describes the interjections as “relatively conventionalized vocal gestures which express a speaker’s mental state, action or attitude or reaction to a situation”. These linguistic elements are used as simple ways to communicate user’s feelings or moods.
In previous researches interjections were represented as good humor clues. Kreuz and Caucci (2007) tried to determine if specific lexical factors might suggest the interpretation of a statement as sarcastic. They demonstrated with a test that the presence of interjections is a good predictor for the readers. They provided a group of students with some extracts from various works, a part of which originally contained the word “sarcastically”. Students were able to classify correctly the extracts where the word “sarcastically” was deleted thanks to the interjections.
Carvalho et al. (2009) noted that positive interjections has very often an ironical use in apparently positive utterances.
Taking into consideration these precedent researches, we consider improper and proper interjections annotated with positive polarity (see Table 1 in section 4.1). Improper interjections are 9Lindbladh (2015: 1). usually followed by exclamations or question marks, which suggest a rising intonation (“sicuro!”), whereas proper ones (or onomatopoeic expressions) are sometimes added to the phrase without any punctuation characters (“ah dimenticavo”, “ah comunque”). 5.2
The adjective or pronoun “che” can be used with exclamatory intention in expressions such as “che ridere”, “che educato”, “che sorpresa”. Like interjections, these expressions are used as marks to express user’s emotions and their ironic intent. 5.3
The use of pronoun “tu” and its morphological
inflection of the verb “essere” expresses a high
degree of proximity between the user and the
person it refers to
In the training set we note how disjunctive conjunctions (“o”, “oppure”) are used to introduce an alternative between two propositions or concepts which may belong to very different semantic domains (for example: In televisione stamattina: i cartoni animati o Mario Monti.[…]). This strange combination of ideas surprises the readers and suggests them a possible ironic interpretation of the message. 5.5
Onomatopoeic expressions for laughter (the most
diffused are “ahah”, “hehe” and “ihih”) are
usually used in humorous texts
Users utilize emoticons to show their facial expressions as well as their emotions in the texts. Tavosanis (2010) presents a macro-classification of emoticons: expressive, decorative/pleasant and of morphosyntactic substitution, which stand for a word or a whole phrase.
In our research we only consider expressive
emoticons which add information about the
user’s mood. In particular we focus on the ironic
emoticons, those which express joking or ironic
intention (see section 4.1). We have
distinguished EMOIRO from EMOPOS because
positive emoticons
Hashtag is a special element in the syntax of tweets used to connect those ones containing the same keywords (which may be a part of the speech) or phrases as #mobbastaveramenteperò.
The user communicates through hashtags several information about events, people they refers to and the topic of message. We focus on hashtags that may suggest to the readers an ironic connotation of the message as #lol and #ironia, and on others that we extracted from ironic tweets in the training set: #stranezze, #Ahahahahah, #benecosì, etc. 5.8
It seems that regional expressions are utilized by users in ironic texts to underline their own mood and emotions. In particular, common constructions deriving from local use may be: “annamo bene”, “namo bene” and “ce” followed by the verb (e.g. “ce vuole”, “ce sta”, “ce potrebbe”), as in this ironic tweet: “@zdizoro t'appassionerà sapè che nel prossimo governo #Monti ce potrebbe rimanè MaryStar Gelmini, come n'incrostazione”. 5.9
We focus on the use of quotation marks as a sign for the readers to interpret non-literally the content of text. In fact, in the social networks these elements are frequently used to underline the possible different meanings of the word between quotation marks, and emphasize the ironic content. 5.10
In web communication the punctuation plays an
important role in the expression of the emotions
and feelings. Several researches
Our system is evaluated on the SENTIPOLC
official test data composed of 3000 tweets and the
values of precision, recall and average F-score
are calculated using the evaluation tool provided
by the organizers
Rank 1 2 3 4 5 6 7 8 9 10 11 12
F-score In this paper we have described our computational model based on linguistic features which have proven to be good clues for the identification of ironic texts. Nonetheless, in future works we plan to examine in depth semantic inconsistencies and ambiguities, amusing wordplay and rhymes that may surprise the reader. In conclusion, we think that a good detection of irony is possible if all the levels of linguistic analysis are considered.