Verb-centered sentiment analysis provides a fine-grained perspective on the polar impact a situation has on its participants. It is good for a person to be honored for e.g. some achievement and we know in advance that such an achievement is expected to have positive polarity. Also, there is a positive relation between the honorer and the honoree. We introduce an Italian verb lexicon that specifies the polar relations a verb expresses and the efects and expectations that the roles of the verb bear. A polar verb often expresses a polar relation like in favour or against. Given the sentence Lei lo critica (Eng. She criticizes him), we conclude that the referent of lei (Eng. she) is (in a situation-specific way) against the referent of lo (Eng. him). We can fix this directed relation by qualifying the subject (the agent) of the verb as the source and the direct object (patient or theme) as the target. We 2. Related Work do not claim that all of these relations are intentional, i.e. express a deliberate polar opinion of the source towards There are quite a number of approaches to Italian senti- the target. Our ultimate goal is to enable (transitive) senment analysis, be it for social media [1], stance detection timent inference on the basis of our lexical specifications. [2] or in the context of aspect-based sentiment analysis Take the sentence She is against the contract because it [3]. Recently, deep learning approaches have conquered prevents a solution. From this we can directly learn that the field, but there are lexicon-based approaches as well. a) she is intentionlly against the contract and that b) the To the best of our knowledge, our lexicon is the first for contract is (non-intentionally directed) against a solution. Italian sentiment inference based on a verb resource. In Only the presence of b) allows us to draw the transitve the past, we have created resources and prototypes e.g. inference that she is in favour of a solution. The underly[4], [5] and [6] for German sentiment inference1 and we ing rule is: if X is against Y and Y is against Z, then X is in favour of Z. Thus, the source and target can be animate or non-animate. If it is animate, than the relation might indicate an intentional polar relation originating from the source towards the target. A non-animate source just counts as a cause. Some polar verbs do not cast a polar relation, but yield a positive or negative perspectivation. For instance, the intransitive usage of the verb sofrire (Eng. sufer) like in Lui sofre (Eng. He suefrs) just tells us that the referent
A sentence is not just positive or negative, it moreover sheds a (polar) light on its participants. Given a sentence like I mourn my freedom (It. Piango la mia libertà), a negative efect is cast on the subject and we expect a positive filler of the direct object. Furthermore, we might conclude that the subject is in favour of the object. We have created an Italian verb lexicon comprising 307 verbs with up to 3 frames per verb in order to enable such kind of inferences. Such a lexicon might be useful for stance detection or even for lexicon induction (e.g. from the object expectation of mourn we can conclude that libertà is a positive concept).
In this short paper, we discuss the ingredients of our lexicon both from a practical and theoretical point of view. of lui is experiencing a negative efect. A polar efect (oov) cases. Here deep learning (e.g. word embeddings) is a positive or negative connotation associated with a might actually be useful, since the similarity of oov verbs verb role (comparable to good/bad for efects of [ 13]). A to verbs that have been seen in the training phase could verb expressing a polar relation might also cast efects, be exploited. but not in any case. In Lei lo punisce (Eng. she punishes him), besides the negative relation there is also a negative efect casted onto lo (Eng. him). However in Lui sofre 4. Corpus: Data Format del suo capo (Eng. he sufers from his boss) although a negative relation from Lui towards suo capo holds, there We have created a json version of our resource and placed is no negative efect on it. This becomes even clearer if it at Zenodo3. For each of the 307 verbs4 a dictionary is we replace the animate filler by a non-animate one like in given with the verb lemma, all (subcategorization) frames, Lui sofre per questa situazione (Eng. He sufers from the the verb polarity, the relations expressed, the efects and situation). It makes no sense to claim a negative efect expections and one or more sample sentences. For inon the situation. stance the verb interrompere (Eng. e.g. cut of) (only
In sum: our annotation scheme for polar verbs speci- one frame is shown) is described in Fig. 1. The English ifes which roles of a verb do have a positive or negative {"verb": "interrompere", (directed) relation given an afirmative usage of the verb. "frame": ["Subj", "Obj"], Also, we specify whether a verb casts a positive or nega- "polarity": "NEG", tive efect on its roles. "relation":[("Subj","against","Obj")]
So far in our discussion, we have used dependency ""eexfpfeeccttast"i:on[(s"":Ob[j]",, "neg")], labels or grammatical functions in order to refer to the "example": "Hanno interrotto gli aiuti roles of a verb. We claim, for instance, that the subject umanitari"} of criticare (Eng. criticise) is against the direct object.
Even if we assume some normalization, namely passive Figure 1: Sample entry for interrompere voice normalization, this might lead to problems. In both cases, The famous chef cooks and The rice cooks there is a subject, but in the first case it realises an agent in the translation of the sample sentence is They have cut of second a theme. Although in principle semantic role la- humanitarian aid. According to the verb frame, a subject belling (SRL) might be regarded a better basis for such and an object are required. Given this, a negative efect specification, we have found in [ 14] that at least for Ger- is cast on the object (humanitarian aid) and an against man the available SRL approaches are not suited. Among relation is drawn to the object. In the case of non-overt others, the problem is that the agent role is assigned not pronouns the arc of the finite verb is annotated, see the exclusively to animate objects (or metonomically used against relation in Fig. 2. inanimate objects) but also to plain physical objects. For example, the semantic role labeller InVeRo [15] using Some roles of a verb seem to carry a polar expectation. VerbAtlas2 assigns the theme role to both (German) sub- For instance if we hear that somebody is sufering from jects: a) Vergeltung droht ihm (Eng. retribution threatens something, we immediately know that this something is him) and b) Er (animate) droht ihm (Engl. He threatens (regarded as) negative. We have integrated expectations him). Only in b) a negative relation should hold (i.e. He in our lexicon, since it might be useful for e.g. lexicon against him). For our German model, we sticked with de- induction. pendency labels and specified coarse-grained selectional restrictions for each verb frame. We then used an ani- 5. Further Details and Restrictions macy classifier, see [ 16], in order to select the right verb frame. For Italian, the quality of semantic labeling has to First of all, verb disambiguation cannot be totally rebe evaluated in order to come to a final decision about placed by subcategorization checks. Take an English the role inventory. This is future work. example: the two sentences a) he won the race and b) he
Detecting polar relations and efects in sentences on won the trophy have the same subcategorization frame the basis of a lexicon (and a parser) might be regarded (also VerbAtlas assigns the same semantic roles), but the antiquated in the age of deep learning. However, there is (polar) meaning is diferent due to verb ambiguity. In a) no need to learn these patterns from annotated data. If he receives a positive efect, but has no relation to the the trigger conditions of a verb (frame) are met, then the race, while in b) we can conclude that he is in favour of relations and efects can safely be asserted. Nevertheless, lexicon-based approaches sufer from out-of-vocabulary
the trophy. This is to say that using our resource in a system demands additional functionalities.
Some verbs need further checks for them to get triggered. Take for instance the verb impedire (Eng. impede). If someone impedes something negative (a struggle), this is good of her (a positive efect and negative relation), but if it is positive (e.g. a solution), it is bad of her (a negative efect and - again - a negative relation). We can determine polar restrictions on the basis of a traditional polarity lexicon. But sometimes also sentiment composition might be needed as in He stopped the unjustified promotion of the minister . Here unjustified promotion is negative according to a simple compostition rule that says that a negative adjective followed by a positive noun gives a negative phrase. Our manually created Italian polarity lexicon5 (3,247 lemmas) can be used for such sentiment composition. Nouns and adjectives have been specified on the basis of the appraisal theory [17]: a word might be positive or negative wrt. to one of three dimensions: moral, emotion, appreciation. This allows even to distinguish between cases where a negative efect depends on the appraisal type of the filler. sick and cheating are both negative, but the former is an appreciation while the latter is morally negative adjective. While the sentence she admires her sick colleague proves that she is compassionate, the sentence she admires her cheating colleague makes her a morally questionable person. Table 1 shows the distribution of word and polarities in our polarity lexicon.
PoS noun adjectives adverbs positive 370 421 76 negative 1242 1060 78
Many Italian polarity lexicons6 specifying the word level polarity of nouns, adjectives and verbs are available.
6https://www.ai-lc.it/en/afective-lexica-and-other-resourcesfor-italian However, there is no fine-grained lexicon at the verb level comparable with ours.
In this paper, we have described an Italian verb lexicon and the theory behind it. The lexicon specifies for each Italian verb the underlying relation (in favour or against) and for each verb argument the efects and expectations (if any). The resource has been released together with a traditional polaritiy lexicon (nouns, adjectives). The verb lexicon can be used for sentiment inference, finding out who is in favour of or against who or what. This is useful e.g. for the profiling of political parties: are they in favour or against some topic. Also, the polar perspective which we call efects can be determined: to answer the question whether a text is positive or negative towards an animate or inimate object. This enables to determine the stance of a text author, but also can be used in order to find out how a person or a product is framed e.g. in a particular text or text collection.
We like to thank Michi Amsler for the coordination of the development of our lexical resources for German, French and Italian. He also was involved in the annotation of the German lexicon. - task a (short paper), EVALITA Evaluation of ceedings of the 51st Annual Meeting of the AssoNLP and Speech Tools for Italian - December 17th, ciation for Computational Linguistics (Volume 2: 2020 (2020). URL: https://api.semanticscholar.org/ Short Papers), Association for Computational LinCorpusID:229292804. guistics, Sofia, Bulgaria, 2013, pp. 120–125. URL: [3] P. Basile, D. Croce, V. Basile, M. Polignano, et al., https://aclanthology.org/P13-2022.
Overview of the evalita 2018 aspect-based senti- [14] M. Klenner, A. Göhring, Semantic role labeling for ment analysis task (absita), in: EVALITA Evalu- sentiment inference: A case study, in: Proceedings ation of NLP and Speech Tools for Italian, CEUR, of the 18th Conference on Natural Language Pro2018, pp. 1–10. cessing (KONVENS 2022), KONVENS 2022 Orga[4] M. Klenner, S. Tron, M. Amsler, N. Hollenstein, The nizers, Potsdam, Germany, 2022, pp. 144–149. URL: Detection and Analysis of Bi-polar Phrases and Po- https://doi.org/10.5167/uzh-220537. larity Conflicts, De Gruyter, 2015, pp. 161–172. URL: [15] S. Conia, R. Orlando, F. Brignone, F. Cecconi, R. Navhttps://doi.org/10.5167/uzh-99629. igli, InVeRo-XL: Making cross-lingual Semantic [5] M. Klenner, M. Amsler, Sentiframes: A resource for Role Labeling accessible with intelligible verbs and verb-centered German sentiment inference, in: Pro- roles, in: Proceedings of the 2021 Conference ceedings of the Tenth International Conference on on Empirical Methods in Natural Language ProLanguage Resources and Evaluation (LREC 2016), cessing: System Demonstrations, ACL, Online and European Language Resources Association (ELRA), Punta Cana, Dominican Republic, 2021, pp. 319–328.
Paris, France, 2016. URL: https://aclanthology.org/2021.emnlp-demo.36.
[6] M. Klenner, D. Tuggener, S. Clematide, Stance de- doi:10.18653/v1/2021.emnlp-demo.36.
tection in Facebook posts of a German right-wing [16] M. Klenner, A. Göhring, Animacy denoting german
party, in: LSDSem 2017/LSD-Sem Linking Models nouns: Annotation and classification, in:
Proceedof Lexical, Sentential and Discourse-level Seman- ings of the Language Resources and Evaluation
Contics, 2017. URL: https://doi.org/10.5167/uzh-136567. ference, European Language Resources Association
[7] M. Klenner, Sentiment inference for gender pro- (ELRA), Marseille, France, 2022, pp. 1360–1364. URL:
ifling, in: Proceedings of the 4th Conference on https://aclanthology.org/2022.lrec-1.145.
Language, Data and Knowledge, in press, Viena, [17] J. R. Martin, P. R. R. White, Appraisal in English,
Austria, 2023. Palgrave, London, 2005.
[8] A. Neviarouskaya, H. Prendinger, M. Ishizuka,
Semantically distinct verb classes involved in
sentiment analysis, in: IADIS AC (
A data-driven investigation, in: Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Association for Computational Linguistics, Berlin, Germany, 2016, pp. 311–321. URL: https://aclanthology.
org/P16-1030. doi:10.18653/v1/P16-1030. [11] K. Park, Z. Pan, J. Joo, Who blames or endorses whom? entity-to-entity directed sentiment extraction in news text, in: Findings of the Association for Computational Linguistics: ACLIJCNLP 2021, Association for Computational Linguistics, Online, 2021, pp. 4091–4102. URL: https: //aclanthology.org/2021.findings-acl.358. doi: 10.
18653/v1/2021.findings-acl.358. [12] B. Zhang, D. Ding, L. Jing, How would stance detection techniques evolve after the launch of Chat
GPT?, CoRR abs/1207.0016 (2023). [13] L. Deng, Y. Choi, J. Wiebe, Benefactive/malefactive
event and writer attitude annotation, in: ProAppendix: Verbs and Frame Size (#) verb ricompensare ridere rifiutare rimproverare ringraziare rinunciare rischiare risolvere rispettare ritirare rivelarsi rovesciare sacrificare salvare sbagliare scappare schiacciare sciupare scomparire sconvolgere scoraggiare scuotere separarsi sfuggire sgridare smentire snobbare sofrire sopprimere sospendere sottomettere sperare stancare stimare stravolgere subire suicidarsi temere tradire trascurare trufare ubbidire urtare vegliare vincere verb riconoscere ridurre rimpiangere rinchiudersi rinnovare riposare risentire risparmiare ristabilire riuscire rompere rubare salutare sanzionare sbarazzare scatenare scioccare scombussolare scompigliare scoppiare scordare sedurre sfidare sgomberare smarrire smettere soddisfare sopportare sorridere sospettare sottrarsi spezzare stigmatizzare strappare stuprare suggerire suscitare tormentare trascinare trasgredire turbare uccidere vantare vietare violare 2 3 1 1 1 2 3 1 4 1 2 1 1 1 1 2 1 1 2 2 1 1