Guiding Theme A3: Opinion and Sentiment - extrapropositional aspects of discourse


Theme A3 analyzes extra-propositional aspects of meaning: distinguishing facts from non-facts and classifying sentiment associated with an entity (A1) or an event (A2). These aspects of meaning can help identifying entities or facts in a given text that the writer considers important in that he or she conveys an opinion or sentiment towards them. Analysis of modality and sentiment can thus serve as a guide for content extraction for summarization and directly feeds into aspect-based summarization, which aims to highlight different perspectives and opinions about specific contents, or the pros and cons of a situation.

Example PhD Project

A possible PhD project continuing the work in theme A3 focuses on applying sentiment analysis to the extraction of aspects of specific contents in diverse domains to analyze specific goals, intentions, wishes or user needs. This involves the identification of relevant attributes of objects or actions, and their dependencies, across different domains. We will extend current work in this field by addressing novel categories and uses of aspects (e.g. to draw conclusions about general or user-specific desires, intentions or goals underlying aspect-based opinions) and to integrate aspect-based sentiment analysis with domain-specific and common knowledge and knowledge acquisition. As areas of application beyond customer review datasets we envisage politics or journalism, scientific texts, or argumentation.
We aim at applying advanced neural learning methods, including memory networks or graph-based modeling, and will consider transfer learning or zero-shot learning to address multiple domains. As in our prior work, all methods should be transferable to novel languages easily.
The thesis can build on prior work conducted in guiding theme A3 (see below), in particular fine-grained opinion analysis with multi-task learning, discourse-level sentiment inference and anaphora resolution, as well as research conducted in the neighboring themes A2 and A1. Opportunities for collaboration exist with the the projects in area B (NLP for multi-document summarization), as well as area C (e.g. methods on contextual and constraint-based ranking). The work can also be applied to AIPHES multi-document summarization corpora (e.g. Zopf et al. 2016 or Tauchmann et al. 2017) and can feed into research on personalized summarization in area D.

Research results in the first PhD cohort
Deep learning with sentiment inference for discourse-oriented opinion analysis


The current work in theme A3 focuses on fine-grained opinion analysis (FGOA), which encompasses detecting, classifying and representing opinion expressions including holder and target roles, and connecting them to representations of entities and events (A1, A2). We propose a neural model for FGOA. We address data scarcity by embeding this model in a MTL framework and obtain clear performance improvements using semantic role labeling as auxiliary task (Marasović and Frank 2017). Deeper investigations compare different MTL architectures re. complexity, flexibility and interpretability (Marasoviç and Frank, submitted).
To date, FGOA has been strictly confined to the sentence level. We address FGOA from a discourse perspective by resolving abstract anaphors to their antecedents in discourse, so that a sentence-level anaphoric opinion target can be extracted from the preceding context, as in (1). Marasović and Frank (2017) are the first to address the difficult task of abstract anaphora resolution in a neural framework, using a Siamese Network account. The model is trained on artificial training data. This allows us to port our model to German with minimal effort.

(1)    U.S. gun laws enable weapons flow from the U.S. into the hands of Mexican drug chartels. [. . . ] [Mexico's president]Holder [criticized]Opinion [this issue]Target.

Finally, we will address detection of implicit sentiment via inference on explicit sentiment and events that positively or negatively affect entities (Deng and Wiebe 2014). In (2), we can infer that people have negative sentiment towards Chavez through inference based on the explicit opinion expression happy taking as target the embedded clause and the predicate fall expressing a negative effect on its subject Chavez (Deng et al. 2013). Further indicators for implicit sentiment will be rules based on sense-disambiguated modal expressions (cf. Marasović et al. (2016) and the multilingual neural system in Marasović and Frank (2016)).

(2)    I think peoplepos.holder are happypos [ because Chavezneg.affected has fallenneg.event]



Marasović, A. and Frank, A. (2017). SRL4ORL: Improving Opinion Role Labelling using Multi-task Learning with Semantic Role Labeling. In: Learning with Limited Labeled Data Workshop (NIPS) (2017). url:

Marasović, A., Born, L., Opitz, J., and Frank, A. (2017): A Mention-Ranking Model for Abstract Anaphora Resolution. Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing (EMNLP), Copenhagen, Denmark, pp. 221--232.

Marasović, A., Zhou, M., Palmer, A., and Frank, A. (2016): Modal Sense Classification At Large: Paraphrase-Driven Sense Projection, Semantically Enriched Classification Models and Cross-Genre Evaluations. Linguistic Issues in Language Technology, Special issue on Modality in Natural Language Understanding, Stanford, CA., vol. 14 (2), CSLI Publications.

Marasović, A. and Frank, A. (2016): Multilingual Modal Sense Classification using a Convolutional Neural Network. Proceedings of the 1st Workshop on Representation Learning for NLP, Berlin, Germany, pp. 111--120.

Lingjia Deng, Janyce Wiebe, and Yoonjung Choi (2013). Benefactive/Malefactive Event and Writer Attitude Annotation. Annual Meeting of the Association for Computational Linguistics (ACL-2013, short paper).

Lingjia Deng and Janyce Wiebe (2014). Sentiment Propagation via Implicature Constraints. Meeting of the European Chapter of the Association for Computational Linguistics (EACL-2014).

Poster (in German)

A A A | Drucken Print | Impressum Impressum | Sitemap Sitemap | Suche Search | Kontakt Contact | Webseitenanalyse: Mehr Informationen
zum Seitenanfangzum Seitenanfang