Monday 5th - Monash Caulfield B2.14
|9:00||Tutorial 1a: NP Bayes|
|10:45||Tutorial 1b: NP Bayes|
|13:30||Tutorial 2a: Succinct data struct.|
|15:45||Tutorial 2b: Succinct data struct.|
|16:45||End of session|
Tuesday 6th - Monash Caulfield B2.14
|Session 1: Opening & Invited talk|
|9:15||Invited talk: Mark Steedman||On Distributional Semantics|
|Session 2: Translation (Chair: Stephen Wan)|
|10:45||Presentation: Kyo Kageura, Martin Thomas, Anthony Hartley, Masao Utiyama, Atsushi Fujita, Kikuko Tanabe and Chiho Toyoshima||Supporting Collaborative Translator Training: Online Platform, Scaffolding and NLP|
|11:10||Presentation : Nitika Mathur, Trevor Cohn and Timothy Baldwin||Improving Human Evaluation of Machine Translation|
|11:25||Paper: Cong Duy Vu Hoang, Reza Haffari and Trevor Cohn||Improving Neural Translation Models with Linguistic Factors|
|11:40||Presentation : Daniel Beck, Lucia Specia and Trevor Cohn||Exploring Prediction Uncertainty in Machine Translation Quality Estimation|
|11:55||CLEF eHealth 2017 Shared tasks|
|Session 3a: Invited talk (Chair: Hanna Suominen)|
|13:15||Invited talk: Hercules Dalianis||HEALTH BANK: A Workbench for Data Science Applications in Healthcare|
|Session 3b: Health (Chair: Hanna Suominen)|
|14:00||Presentation : Raghavendra Chalapathy, Ehsan Zare Borzeshi and Massimo Piccardi||An Investigation of Recurrent Neural Architectures for Drug Name Recognition|
|14:15||Paper: Hamed Hassanzadeh, Anthony Nguyen and Bevan Koopman||Evaluation of Medical Concept Annotation Systems on Clinical Records|
|14:30||Paper: Mahnoosh Kholghi, Lance De Vine, Laurianne Sitbon, Guido Zuccon and Anthony Nguyen||The Benefits of Word Embeddings Features for Active Learning in Clinical Information Extraction|
|14:45||Presentation : Rebecka Weegar and Hercules Dalianis||Mining Norwegian pathology reports: A research proposal|
|15:00||Paper: Pin Huang, Andrew MacKinlay and Antonio Jimeno||Syndromic Surveillance using Generic Medical Entities on Twitter|
|15:15||Paper: Yufei Wang, Stephen Wan and Cecile Paris||The Role of Features and Context on Suicide Ideation Detection|
|Session 4: Relation & Information extraction (Chair: Andrew MacKinlay)|
|16:00||Presentation : Dat Quoc Nguyen and Mark Johnson||Modeling topics and knowledge bases with embeddings|
|16:15||Paper: Zhuang Li, Lizhen Qu, Qiongkai Xu and Mark Johnson||Unsupervised Pre-training With Seq2Seq Reconstruction Loss for Deep Relation Extraction Models|
|16:30||Presentation : Hanieh Poostchi, Ehsan Zare Borzeshi and Massimo Piccardi||PersoNER: Persian Named-Entity Recognition|
|16:45||Paper: Nagesh C. Panyam, Karin Verspoor, Trevor Cohn and Rao Kotagiri||ASM Kernel: Graph Kernel using Approximate Subgraph Matching for Relation Extraction|
|17:00||Paper: Gitansh Khirbat, Jianzhong Qi and Rui Zhang||N-ary Biographical Relation Extraction using Shortest Path Dependencies|
|17:15||End of session|
Wednesday 7th - Monash Caulfield B2.14
|Session 5: Invited talk & Shared task (Chair: Diego Molla)|
|9:00||Invited talk: Steven Bird||Getting started with an Australian language|
|Andrew Chisholm, Ben Hachey and Diego Mollá||Overview of the 2016 ALTA Shared Task: Cross-KB Coreference|
|Gitansh Khirbat, Jianzhong Qi and Rui Zhang||Disambiguating Entities Referred by Web Endpoints using Tree Ensembles|
|S. Shivashankar, Yitong Li and Afshin Rahimi||Filter and Match Approach to Pair-wise Web URI Linking|
|Cheng Yu, Bing Chu, Rohit Ram, James Aichinger, Lizhen Qu and Hanna Suominen||Pairwise FastText Classifier for Entity Disambiguation|
|Session 6: Short-papers & posters (Chair: Karin Verspoor)|
|10:45||Short-paper lightning talks|
|Aditya Joshi, Vaibhav Tripathi, Pushpak Bhattacharyya, Mark Carman, Meghna Singh, Jaya Saraswati and Rajita Shukla||How Challenging is Sarcasm versus Irony Classification?: A Study With a Dataset from English Literature|
|Ming Liu, Gholamreza Haffari and Wray Buntine||Learning cascaded latent variable models for biomedical text classification|
|Bo Han, Antonio Jimeno Yepes, Andrew MacKinlay and Lianhua Chi||Temporal Modelling of Geospatial Words in Twitter|
|Antonio Jimeno Yepes and Andrew MacKinlay||NER for Medical Entities in Twitter using Sequence to Sequence Neural Networks|
|Dat Quoc Nguyen, Mark Dras and Mark Johnson||An empirical study for Vietnamese dependency parsing|
|Will Radford, Ben Hachey, Bo Han and Andy Chisholm||:telephone::person::sailboat::whale::okhand: ; or “Call me Ishmael” – How do you translate emoji?|
|Xavier Holt, Will Radford and Ben Hachey||Presenting a New Dataset for the Timeline Generation Problem|
|Session 7: Applications (Chair: Trevor Cohn)|
|13:35||Paper: Hafsah Aamer, Bahadorreza Ofoghi and Karin Verspoor||Syndromic Surveillance through Measuring Lexical Shift in Emergency Department Chief Complaint Texts|
|13:50||Paper: Rui Wang, Wei Liu and Chris McDonald||Featureless Domain-Specific Term Extraction with Minimal Labelled Data|
|14:05||Presentation : Ehsan Shareghi||Unbounded and Scalable Smoothing for Language Modeling|
|14:30||Paper: Shunichi Ishihara||An Effect of Background Population Sample Size on the Performance of a Likelihood Ratio-based Forensic Text Comparison System: A Monte Carlo Simulation with Gaussian Mixture Model|
|14:45||Presentation: Oliver Adams, Shourya Roy and Raghu Krishnapuram||Distributed Vector Representations for Unsupervised Automatic Short Answer Grading|
|15:00||Paper: Andrei Shcherbakov, Ekaterina Vylomova and Nick Thieberger||Phonotactic Modeling of Extremely Low Resource Languages|
|15:15||Presentation: Oliver Adams, Adam Makarucha, Graham Neubig, Steven Bird and Trevor Cohn||Cross-Lingual Word Embeddings for Low-Resource Language Modeling|
|Session 8: Closing|
|16:00||Awards for best paper and best presentation|
|16:25||End of session|
The central problem in open domain-question answering from text is the problem of entailment. Given enough text, the answer is almost certain to be there, but is likely to be expressed in a different form from the one the question suggest-either in a paraphrase, or in a sentence that entails or implies the answer.
We cannot afford to bridge this gap by open-ended theorem-proving search. Instead we need a semantics for natural language that directly supports common-sense inference, such as that arriving somewhere implies subsequently being there, and invading a country implies attacking it. We would like this semantics to be compatible with traditional logical operator semantics including quantification, negation and tense, so that not being there implies not having arrived, and not attacking implies not invading.
There have been many attempts to build such a semantics of content words by hand, from the generative semantics of the '60s to WordNet and other resources of the present. The '60s saw attempts based on generative semantics, while more recently, they have engendered WordNet and other computational resources. However, such systems have remained incomplete and language-specific in comparison to the vastness of human common-sense reasoning. One consequence has been renewed interest in the idea of treating the semantics as "hidden", to be discovered through machine learning, an idea that has its origins in the "semantic differential" of Osgood, Suci, and Tannenbaum in the '50s.
There are two distinct modern approaches to the problem of data-driven or "distributional" semantics. The first, which I will call "collocational", is the direct descendant of the semantic differential. In its most basic form, the meaning of a word is taken to be a vector in a space whose dimensions are defined by the lexicon of the language, and whose magnitude is defined by counts of those lexical items within a fixed window over the string (although in practice the dimensionality is reduced and the relation to frequency less direct). Crucially, semantic composition is defined in terms of linear algebraic operations, notably vector addition.
A second "denotational" approach defines the meaning of a word in terms of the entities that it is predicated over and the ensembles of predications over entities of the same types, obtained by machine-reading with wide coverage parsers. (Names or designators in text are generally used as a proxy for the entities themselves.) Semantic composition can then be defined as an applicative system using logical opertors such as quantifiers and negation, as in traditional formal semantics.
The talk reviews recent work in both collocation- and denotation- based distributional semantics, and asks for each what dimensions of meaning are actually being represented. It argues that the two approaches are largely orthogonal on these dimensions. Collocational representations are good for representing ambiguity, with linear algebraic composition most effective at disambiguation and representing distributional similarity. Denotational representations represent something more like a traditional compositional semantics, but one in which the primitive relations correspond to those of a hidden language of logical form representing paraphrase and common-sense entailment directly.
To make this point, the talk discusses recent work in which collocational distributional representations such as embeddings have been used as proxies for semantic features in models such as LSTM, to guide disambiguation during parsing, while a lexicalized denotation-based distributional semantics is used to support inference of entailment. I will show that this hybrid approach can be applied with a number of parsing models, including transition-based and supertagging, to support entailment-based QA with denotation-based distributional representations. I will discuss work at Edinburgh and elsewhere in which the semantics of paraphrases is represented by a single cluster identifier, and where common-sense inference (derived from a learned entailment graph) is built into the lexicon and projected by syntactic derivation, rather than delegated to a later stage of inference. The method can be applied cross-linguistically, in support of machine translation. Ongoing work extends the method to extract multi-word items, light-verb constructions, and an aspect-based semantics for temporal/causal entailment, and to the creation and interrogation of Knowledge Graphs and Semantic Nets via natural language.
Healthcare has many challenges in form of monitoring and predicting adverse events as healthcare associated infections or adverse drug events. All this can happen while treating a patient at the hospital for her disease. The research question is: When and how many adverse events have occurred, how can one predict them? Nowadays all information is contained in the electronic patient records and are written both in structured form and in unstructured free text. This talk will describe the data used for our research in HEALTH BANK - Swedish Health Record Research Bank containing over 2 million patient records from 2007-2014. Topics are detection of symptoms, diseases, body parts and drugs from Swedish electronic patient record text, including deciding on the certainty of a symptom or disease and detecting adverse (drug) events. Future research are detecting early symptoms of cancer and de-identification of electronic patient records for secondary use.
At least a dozen Australian indigenous languages are still being learnt by children as their first language. These children have limited access to western-style education and often gain only limited proficiency in English. The languages are effectively unwritten, as there are no naturally occurring contexts where people would need to write the language. The same situation is repeated around the world, where remote communities do not write their language and do not acquire the national language, and government and NGO employees who work with these communities must learn to speak an unwritten language without the help of written resources. In this presentation I will report on early experiences working with Kunwinjku, a polysynthetic language spoken by 1,200 people in western Arnhem Land, leading to several open research questions in the area of tools for adult learners of unwritten languages.