SLTC 2020: Programme

Programme for SLTC

Note: all times are CET (Central European Time)

Below the programme you can find a list of all sessions with authors and links to their abstracts, including the keynote speakers and their abstracts.

Workshop day: 25 November

Please see the respective workshops for details:

08:50–13ish RESOURCEFUL
09:00–17:30 NLP4CALL
13:00–17ish Computational Detection of Language Change
13:00–16:00 Applied Swedish NLP

There are some common shared breaks where people can mingle with other workshop participants:

10:00–10:30 Coffee break
12:00–13:00 Lunch
14:30–15:00 Coffee break

SLTC day 1: 26 November

08:50–09:00 Opening remarks
09:00–10:00 Session K1: Keynote – Leon Derczynski (chair: Richard Johansson)
10:00–10:30 Coffee break
10:30–12:00 Session L1: 3 long talks (chair: Shafqat Virk)
12:00–13:00 Lunch break
13:00–14:30 Session L2: 3 long talks (chair: Henrik Björklund)
14:30–14:45 Break
14:45–15:45 Session S1: 4 short talks (chair: Lars Borin)
15:45–16:15 Coffee break
16:15–17:15 Session S2: 4 short talks (chair: Elena Volodina)
17:15–17:30 Break
17:30–18:30 Session K2: Keynote – Vera Demberg (chair: Asad Sayeed)

SLTC day 2: 27 November

09:00–10:30 Session L3: 3 long talks (chair: Sara Stymne)
10:30–11:00 Coffee break
11:00–12:00 Session K3: Keynote – Raquel Fernandez (chair: Staffan Larsson)
12:00–13:00 Lunch break
13:00–14:15 Session S3: 5 short talks (chair: Mehdi Ghanimifard)
14:15–14:30 Break
14:30–15:45 Session S4: 5 short talks (chair: Yvonne Adesam)
15:45–16:15 Coffee break
16:15–17:45 Session L4: 3 long talks (chair: Olof Mogren)
17:45–18:00 Final remarks

Keynote talks

Session K1, 26 Nov 9:00–10:00. Leon Derczynski

Making NLP Greener

We know our modern models are large - often too large - in NLP. We know we have a difficult task ahead in limiting the damage to our climate. This talk examines how and why NLP affects emissions, and presents practical steps at many stages of the processing and development cycle for reducing not only power and hosting bills but also our impact as researchers and practitioners in NLP.

Session K2, 26 Nov 17:30–18:30. Vera Demberg

How "rational" are humans during language production?

The rational speech act model (RSA) aims to explain language comprehension and production in terms of recursive reasoning about speaker and listener. A lot of the evidence for the RSA model comes from language games, where speakers and listeners interact and try to guess what the other is trying to say with an ambiguous referring expression, or to explain why a speaker uses a specific expression. It is however less clear how these results bear out for everyday language usage. In this talk, I will report on several studies that test for rational communication at the discourse level. The first series of experiments explicitly sets out to test predictions of the RSA theory for the production and interpretation of discourse connectives. We find that while our data is in line with RSA predictions in a game-like setting, the effects disappear when moving to more realistic designs. I will discuss the extent to which this null result may be due to the difficulty of computing discourse expectations. Secondly, I will report on a referring expression experiment during driving, where we find that speakers adapt their utterances only at the beginning of the experiment, and then stick to that strategy, without any evidence for fine-grained adaptation to the comprehension difficulty of the comprehender.

Session K3, 27 Nov 11:00–12:00. Raquel Fernandez

Interaction-driven linguistic conventions

Approaches in psycholinguistics and sociolinguistics have emphasised the link between semantic variation and lexical choice and the interaction between speakers – a view which is not common in mainstream Natural Language Processing. In this talk, I will focus on interaction-driven linguistic conventions at the level of communities and within single conversations. I will first present work on analysing semantic variation in online communities of practice, showing that community-specific variation is at play and can have an impact on language processing. I will then move on to two-person dialogue setups where lexical choice is driven by ad-hoc conventions and describe our recent work on generating referring utterances grounded in the conversational and visual context. I argue that computational models of interaction-driven linguistic conventions can help us to better understand the processes underpinning these phenomena in humans, as well as contribute to more robust language technology tools and to user adaptation in dialogue systems.

Accepted presentations

The presentations are divided into long and short talks.

Long talks will be 20 minutes, plus up to 10 minutes for questions.
Short talks will be 10 minutes. Each short talk session is ended with a joint question and answer session of 20 minutes or more.

Thursday 26 November

Session K1, 9:00–10:00. Leon Derczynski

Session K2, 17:30–18:30. Vera Demberg

Friday 27 November

Sesson L3, 9:00–10:30. Long talks

Arezoo Hatefi and Frank Drewes. Document Clustering Using Attentive Hierarchical Document Representation
Felix Morger. LM-inspector: End-to-end Introspection of pre-trained Language Models
Mila Grancharova, Hanna Berg and Hercules Dalianis. Improving Named Entity Recognition and Classiﬁcation in Class Imbalanced Swedish Electronic Patient Records through Resampling

Session K3, 11:00–12:00. Raquel Fernandez

Session S3, 13:00–14:15. Short talks

Hartger Veeman and Ali Basirat. An exploration of the encoding of grammatical gender in word embeddings
Oskar Jerdhaf, Marina Santini, Peter Lundberg, Anette Karlsson and Arne Jönsson. Implant Terms: Focused Terminology Extraction with Swedish BERT - Preliminary Results
Oluwatosin Adewumi, Foteini Simistira Liwicki and Marcus Liwicki. Corpora Compared: The Case of the Swedish Gigaword & Wikipedia Corpora
Ritika Nandi, Geetha Maiya and Priya Kamath. An Empirical Evaluation of various Word Embedding Models for Subjectivity Analysis Tasks
Tim Isbister and Magnus Sahlgren. Why Not Simply Translate? A First Swedish Evaluation Benchmark for Semantic Similarity

Session S4, 14:30–15:45. Short talks

Christina Tånnander and Jens Edlund. Self-perceived preferences of voice and speaking style characteristics in spoken text
Dana Dannells and Shafqat Virk. OCR Error Detection on Historical Text Using Uni and Multi Feature Based Machine Learning Models
Herbert Lange. A Type-Theoretic Approach to Generating Pictures and Descriptions
Johanna Björklund, Frank Drewes and Iris Mollevik. Towards Semantic Representations with a Temporal Dimension
Jonas Sjöbergh and Viggo Kann. Granska API – an Online API for Grammar Checking and Other NLP Services

Session L4, 16:15–17:45. Long talks

Daniel Holmer and Arne Jönsson. Comparing the performance of various Swedish BERT models for classification
Luise Dürlich and Christian Hardmeier. Automatic Recognition and Classification of Errors in Human Translation
Shifei Chen and Ali Basirat. Cross-lingual Word Embeddings beyond Zero-shot Machine Translation