We introduce “extreme summarization”, a new single-document summarization task which does not favor extractive strategies and calls for an abstractive modeling approach. The idea is to create short, one-sentence news summary answering the question “What article about?”. collect real-world, large-scale dataset this by harvesting online articles from British Broadcasting Corporation (BBC). propose novel model conditioned on article’s topics based entirely convolutional neural networks....

Shashi Narayan, Shay B. Cohen, Mirella Lapata. Proceedings of the 2018 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). 2018.

Stock movement prediction is a challenging problem: the market highly stochastic, and we make temporally-dependent predictions from chaotic data. We treat these three complexities present novel deep generative model jointly exploiting text price signals for this task. Unlike case with discriminative or topic modeling, our introduces recurrent, continuous latent variables better treatment of stochasticity, uses neural variational inference to address intractable posterior inference. also...

Meaning Representation (AMR) is a semantic representation for natural language that embeds annotations related to traditional tasks such as named entity recognition, role labeling, word sense disambiguation and co-reference resolution. We describe transition-based parser AMR parses sentences left-to-right, in linear time. further propose test-suite assesses specific subtasks are helpful comparing parsers, show our competitive with the state of art on LDC2015E86 dataset it outperforms...

This article deals with adversarial attacks towards deep learning systems for Natural Language Processing (NLP), in the context of privacy protection. We study a specific type attack: an attacker eavesdrops on hidden representations neural text classifier and tries to recover information about input text. Such scenario may arise situations when computation network is shared across multiple devices, e.g. some representation computed by user's device sent cloud-based model. measure ability...

Marco Damonte, Shay B. Cohen. Proceedings of the 2019 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). 2019.

It is well-known that abstractive summaries are subject to hallucination—including material not supported by the original text. While can be made hallucination-free limiting them general phrases, such would fail very informative. Alternatively, one try avoid hallucinations verifying any specific entities in summary appear text a similar context. This approach taken our system, Herman. The system learns recognize and verify quantity (dates, numbers, sums of money, etc.) beam-worth produced...

We present and study the contribution-selection algorithm (CSA), a novel for feature selection. The is based on multiperturbation shapley analysis (MSA), framework that relies game theory to estimate usefulness. iteratively estimates usefulness of features selects them accordingly, using either forward selection or backward elimination. It can optimize various performance measures over unseen data such as accuracy, balanced error rate, area under receiver-operator-characteristic curve....

We present a family of priors over probabilistic grammar weights, called the shared logistic normal distribution. This extends partitioned distribution, enabling factored covariance between probabilities different derivation events in grammar, providing new way to encode prior knowledge about an unknown grammar. describe variational EM algorithm for learning based on this priors. then experiment with unsupervised dependency induction and show significant improvements using our model both...

We propose a new sentence simplification task (Split-and-Rephrase) where the aim is to split complex into meaning preserving sequence of shorter sentences. Like simplification, splitting-and-rephrasing has potential benefiting both natural language processing and societal applications. Because sentences are generally better processed by NLP systems, it could be used as preprocessing step which facilitates improves performance parsers, semantic role labellers machine translation systems. It...

Marco Damonte, Shay B. Cohen. Proceedings of the 2018 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). 2018.

Most extractive summarization methods focus on the main body of document from which sentences need to be extracted. However, gist may lie in side information, such as title and image captions are often available for newswire articles. We propose explore information context single-document summarization. develop a framework composed hierarchical encoder an attention-based extractor with attention over information. evaluate our model large scale news dataset. show that consistently outperforms...

The goal of medical relation extraction is to detect relations among entities, such as genes, mutations and drugs in texts. Dependency tree structures have been proven useful for this task. Existing approaches leverage off-the-shelf dependency parsers obtain a syntactic or forest the text. However, domain, low parsing accuracy may lead error propagation downstream pipeline. In work, we propose novel model which treats structure latent variable induces it from unstructured text an end-to-end...

Afonso Mendes, Shashi Narayan, Sebastião Miranda, Zita Marinho, André F. T. Martins, Shay B. Cohen. Proceedings of the 2019 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). 2019.

We introduce an open-domain neural semantic parser which generates formal meaning representations in the style of Discourse Representation Theory (DRT; Kamp and Reyle 1993). propose a method transforms Structures (DRSs) to trees develop structure-aware model decomposes decoding process into three stages: basic DRS structure prediction, condition prediction (i.e., predicates relations), referent variables). Experimental results on Groningen Meaning Bank (GMB) show that our outperforms...