Guiding Theme A2: Identification of complex event structures in discourse
This guiding theme focuses on the analysis of event structures („who did what to whom“) and relations that connect them in discourse, as a basis for content selection in extractive or abstractive multi-document summarization (MDS). We aim at developing robust deep learning techniques in end-to-end architectures that can be applied to event analysis in heterogeneous document sources from different genres and domains. Jointly with the guiding themes A1 and A3, this guiding theme aims to provide important selection criteria for summarization in AIPHES, and to be generalizable to related high-level natural language understanding tasks, such as question-based summarization or story reading comprehension.
Research results of the first Ph.D. cohort
The goal of this thesis is to learn discourse-aware event representations for texts from various domains and to apply them for natural language inference over eventive texts in high-level natural language understanding tasks such as neural abstractive summarization or story reading comprehension.
In a first step, we address two linguistic subtasks that are crucial for understanding events from a discourse perspective: event detection (i.e. detecting what is an event and distinguishing its realis class) and discourse relation sense disambiguation (i.e. how two discourse units referring to events are connected with rhetorical relations, such as ‘contrast’ or ‘causation’). We propose a neural network system enhanced with dependency parse and part-of-speech embeddings that achieves competitive results for event detection (Mihaylov and Frank 2016b). The learned representations can be viewed as neuralized representations of linguistic event annotations that can be directly applied in higher-level natural language understanding tasks. As a second annotation type we address discourse relation sense disambiguation, building on a simple yet powerful method that uses semantic similarity features based on word embeddings (Mihaylov and Frank 2016b). The system achieved the best overall results in the CoNLL shared task 2016. In a large study conducted in Zopf et al. 2018 we could show that – along with other linguistic annotation types such as entity types (A1) and sentiment frames (A3) – discourse relation senses are effective features that capture aspects of the notion of ‘importance’, which in Zopf et al. 2016 serves as a criterion in document summarization.
In Mihaylov et. al. 2017 we then generalize the notion of annotation types to the more general notion of neural representations for ‘linguistic skills’. The skills are learned from specialized corpora for supervised language tasks of different task types (sequence labeling, relation classification or text classification). After learning the knowledge that the learned neural representations incorporate is transferred to novel texts in downstream tasks, in our case, for reading comprehension from stories. We show that the learned ‘linguistic skill representations’ boost the performance of a reading comprehension system (i) early in training and (ii) when training on smaller portions of the original training data, and hence the learned representations prove useful for alleviating the annotation bottleneck for downstream tasks, while enabling powerful end-to-end as opposed to pipelined systems.
Having shown that linguistic knowledge (‘linguistic skills’) can be learned from specialized corpora and applied in downstream tasks, we extend our methods for knowledge transfer to external structured knowledge bases that encode knowledge beneficial for the understanding of events in discourse. In Mihaylov and Frank 2018 we develop a neural story reading comprehension model that integrates external commonsense knowledge, encoded as a key-value memory in neural representations that are interfaced with the question and the document context. The injection of external knowledge improves system performance and at the same time makes the decisions of the system more intuitive and transparent.
In the final part of the thesis, in collaboration with C3, we will combine multi-modal representations learned from images, ‘linguistic skill’ knowledge and structured external knowledge in order to evaluate the understanding of events in tasks such as visual question answering, text summarization, and story reading comprehension.
- PI: Prof. Dr. Anette Frank
- First Cohort PhD student: Todor Mihaylov
- Michael Roth and Anette Frank (2012): Aligning Predicates across Monolingual Comparable Texts using Graph-based Clustering. Proceedings of the 2012 Conference on Empirical Methods in Natural Language Processign (EMNLP), 2012.
- Michael Roth and Mirella Lapata (2015). Context-aware Frame-Semantic Role Labeling. Transactions of the Association for Computational Linguistics, vol. 3, p. 449-460.
- Quang Xuan Do, Yee Seng Chan and Dan Roth (2011): Minimally Supervised Event Causality Identification, Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, p. 294–303.
- Zopf, M., Loza, M. E., and Fürnkranz, J.: Beyond Centrality and Structural Features: Learning Information Importance for Text Summarization. Proceedings of the 20th SIGNLL Conference on Computational Natural Language Learning (CoNLL 2016), 2016.
- Zopf, M., Botschen, T., Falke, T., Marasovic, A., Mihaylov, T., P.V.S., A.,Loza M. E., Fürnkranz, J., Frank, A.: What’s important in a text? An extensive evaluation of linguistic annotations for summarization, Proceedings of The Fifth International Conference on Social Networks Analysis, Management and Security, 2018.
Zopf, Markus ; Botschen, Teresa ; Falke, Tobias ; Heinzerling, Benjamin ; Marasovic, Ana ; Mihaylov, Todor ; P. V. S., Avinesh ; Loza Mencía, Eneldo ; Fürnkranz, Johannes ; Frank, Anette (2018):
What's Important in a Text? An Extensive Evaluation of Linguistic Annotations for Summarization.
Mihaylov, Todor ; Frank, Anette (2018):
Knowledgeable Reader: Enhancing Cloze-Style Reading Comprehension with External Commonsense Knowledge.
In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), In: The 56th Annual Meeting of the Association for Computational Linguistics, Melbourne Australia, 15-20 July 2018, [Online-Edition: http://aclweb.org/anthology/P18-1076],
Mihaylov, Todor ; Balchev, Daniel ; Kiprov, Yasen ; Koychev, Ivan ; Nakov, Preslav (2017):
Large-Scale Goodness Polarity Lexicons for Community Question Answering.
In: SIGIR '17 Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, ACM, Shinjuku, Tokyo, Japan, [Online-Edition: http://delivery.acm.org/10.1145/3090000/3080757/p1185-mihayl...],
Mihaylov, Todor ; Frank, Anette (2017):
Story Cloze Ending Selection Baselines and Data Examination.
In: Proceedings of the Linking Models of Lexical, Sentential and Discourse-level Semantics – Shared Task, DOI: 10.18653/v1/W17-0913,
Mihaylov, Todor ; Kozareva, Zornitsa ; Frank, Anette (2017):
Neural Skill Transfer from Supervised Language Tasks to Reading Comprehension.
abs/1711.03754, [Online-Edition: http://arxiv.org/abs/1711.03754],
Mihaylov, Todor ; Frank, Anette (2016):
AIPHES-HD system at TAC KBP 2016: Neural Event Trigger Detection and Event Type and Realis Disambiguation with Word Embeddings.
In: Proceedings of the TAC Knowledge Base Population (KBP) 2016, [Online-Edition: https://tac.nist.gov/publications/2016/participant.papers/TA...],
Mihaylov, Todor ; Frank, Anette (2016):
Discourse Relation Sense Classification Using Cross-argument Semantic Similarity Based on Word Embeddings.
In: Proceedings of the Twentieth Conference on Computational Natural Language Learning - Shared Task, [Online-Edition: https://aclweb.org/anthology/K/K16/K16-2014.pdf],
Mihaylov, Todor ; Nakov, Preslav (2016):
Hunting for Troll Comments in News Community Forums.
In: Proceedings of the 54rd Annual Meeting of the Association for Computational Linguistics, Association for Computational Linguistics, [Online-Edition: http://www.aclweb.org/anthology/K15-1032],
Mihaylov, Todor ; Nakov, Preslav (2016):
SemanticZ at SemEval-2016 Task 3: Ranking Relevant Answers in Community Question Answering Using Semantic Similarity Based on Fine-tuned Word Embeddings.
In: Proceedings of the 10th International Workshop on Semantic Evaluation (SemEval 2016), [Online-Edition: http://www.aclweb.org/anthology/S16-1136],