Why AI is not enough anymore

Sebastian Welter

AI has found its way into the daily business in enterprises and institutions. However, the majority of AI projects still fail in the end, even though they are sometimes very successful in POCs and Pilots. Especially in NLP: Bot projects, document classification projects – mature technology – often fails in production.

We tend to forget about the most significant challenges when excelling in AI – integration, maintainability and especially the organization and the people involved in it.

This talk will shed some light on the organizational obstacles in enterprises, challenges evolving through managerial and project expectations and the mistakes that happen when we don’t take the people involved serious.

It will also touch on the struggles of IT integration and maintainability and why a data scientist needs to be an architect, a DevOps specialist, a UI developer, a business analyst, a coach, an integration developer – or needs to make sure someone in the project has most of those capabilities.

Be prepared for some painful and joyful stories about AI projects in the wild.

Organizations

Accenture

Posters

Scientific Track

NFinBERT—A Number-Aware Language Model for Financial Disclosures (Hao-Lun Lin, Jr-Shian Wu, Yu-Shiang Huang, Ming-Feng Tsai and Chuan-Ju Wang)
Old is Not Always Gold—Early Identification of Milestone Patents Employing Network Flow Metrics (Manajit Chakraborty and Fabio Crestani)
Neural Argumentation Mining on Essays and Microtexts with Contextualized Word Embeddings (Mohammad Yeghaneh Abkenar, Prof. Manfred Stede and Prof. Stephan Oepen)

Applied Track

FLIE with Rules (Ela Pustulka-Hunt, Thomas Hanne and Lucia de Espona)
Let your subject-matter experts train AI—an example for document review using PDFs (Juan Miguel Cejuela)
CareerCoach—Automatic Knowledge Extraction and Recommender Systems for Personalized Re- and Upskilling suggestions (Albert Weichselbraun, Roger Waldvogel, Philipp Kuntschik, Alexander van Schie and Andreas Fraefel)
Use-Case Driven Tokenization for Non-Standard Corpora (Niclas Bodenmann)
Understanding structural information in scanned documents (Holger Keibel)
SST-BERT—Semantic Shift Tracing by Clustering in BERT-based Embedding Spaces (Vani Kanjirangat, Sandra Mitrović, Alessandro Antonucci and Fabio Rinaldi)
From Normalized Swiss German to Standard German with Synthetic Parallel Data (Benjamin Suter, Larissa Schmidt and Josef Novak)
Question classification in German dialogues (Alexandros Paramythis and Eleni Adamantidou)
A complete End-to-End Human Resources system backed by NLP and Deep Learning (Alejandro Jesús Castañeira Rodríguez)
Keyword spotting and complete transcriptions for spoken Swiss German (Sonja Huclova, Michelle Tschumi, Claus Hintermeier, Lukas Peter and Thomas Wäspe)
Neural Machine Translation in Academic Contexts (Alice Delorme Benites and Fernando Benites)
Document Similarity Analysis for Clinical Summary Documents in Pharma (Stephan Zähringer, Marco Zeller, Raul Mora and Hans-Peter Güllich)

Automatic Document Parsing with Weak Supervision from Spreadsheets

Johannes Rausch, Susie Xi Rao, Ce Zhang, Peter Egger

Automated processing of electronic documents is a common downstream ML task in industry and research. Here, the lack of structures in formats such as PDF files or scanned documents remains a major obstacle. In practice, extensive engineering and ad-hoc code are required to recover document structures, e.g., for headings, tables, or nested figures.
ML-based systems require large amounts of labeled training data.

In DocParser, Rausch et al. [2021] have shown a robust way to parse complete document structures from rendered inputs. This problem is solved through a novel weak supervision approach that automatically generates training data from structured LaTeX source files of readily available scientific articles.

However, the application in new target domains, e.g., the business context, is hindered by the mismatch between source and target document layouts. Furthermore, weak supervision often cannot be applied, if business documents are generated from different source files, e.g., office software.
As a result, producing sufficient training data sets demands high human labeling efforts.

We contribute a new system to automatically generate weakly labeled data for ML-based document parsers in real-world settings. We achieve this by developing a weak annotation scheme that can utilize source file types such as spreadsheets. In a preliminary case study, we are able to generate 12,000 annotated samples from Excel files. We demonstrate that training our system on this weakly-annotated data allows accurate parsing of structured data in real-world office documents. Previous findings in DocParser suggests that pretraining systems using such weak supervision scheme can lead to a significant reduction of labeling complexity. (Acknowledgement: We thank Mr. Livio Kaiser for supporting this project during his master thesis.)

Organizations

ETH Zürich

Emotions in the Parliament: Lexical Emotion Analysis of Parliamentarian Speech Transcriptions

Egon Werlen, Christof Imhof, Per Bergamin

Politics is emotional. So far, relatively few studies investigated the emotional content in parliamentary speeches. In this study, we analysed emotional valence and arousal of German and French speeches of a Swiss cantonal parliament and whether we can use them to predict the membership of parliamentarians to one of two groups: those who won more of the votings than others.

The emotional text analysis showed that these speeches are indeed emotional. However, the results regarding the predictions were mixed. Arousal and language showed no effects and valence was only partially successful as a predictor.

Further reading: original paper

Organizations

Swiss Distance University of Applied Sciences

Swiss Voice Project: Swiss German Speech Synthesis

Julian Mäder

At the ETH Media Technology Center, we started the Swiss Voice Project to research the technical possibilities required to build Swiss German voice assistants, with a focus on text-to-speech models. Swiss German is a low-resource language with a dialect continuum, for which only little data and no standardized written form is available. This makes building text-to-speech models challenging, since state-of-the-art models for languages such as English and German are data-driven and rely on large data sets.

The Swiss Voice project tries to address those challenges with the following contributions.
In this talk we will highlight the key findings and outcomes from this project. We introduce the first annotated parallel corpus of spoken Swiss German across 8 different dialects, with Swiss German transcripts and parallel standard German reference translations.

This data set will enable the NLP community to research and evaluate Swiss German speech synthesis and the use of High German resources for transfer learning. Based on our data set we trained powerful deep learning models for machine translation and speech synthesis and deployed the Swiss Voice REST API, which takes High German text as input and produces audio in all 8 supported dialects. Building on our API and the Speech Recognition Technology from recapp AG, we created the first voice assistant prototype that understands and speaks Swiss German.

The jointly developed prototype answers everyday requests, such as reading news headlines and weather forecasts in any of the 8 dialects currently available.

Organizations

ETH Zürich

Speaker Recognition

Jan Deriu, Amin Moghaddam, Malgorzata Anna Ulasik, Katsiaryna Mlynchyk, Mark Cieliebak

When we transfer insights from the research setting to a real-world application, problems often arise, which never occurred in the clean lab setting.

In this talk, we showcase this problem in the context of speaker diarization or speaker recognition, i.e., the task of segmenting a recorded conversation into “Who spoke when?”. Potential use-cases include the creation of transcripts of interviews, or for authentication of speakers.

One major issue is that it is not essential in the lab setting if the change of one speaker to the other is precise to the tenth of a second. However, in the real-world setting, this is crucial. For instance, when we want to get an automatic transcript of an interview, one-tenth of a second might shift a couple of words to the wrong speaker, which leads to the software being perceived as low quality.

This talk aims to showcase how we tackled this issue using speech processing and linguistic features. Furthermore, we show how the setting in which diarization is trained in a lab setting differs from the actual use-case and how this can be leveraged (e.g., in our use-case, we have information from an ASR system, and we know the number of speakers).

Organizations

Zürich University of Applied Sciences
SpinningBytes AG

How We Crowdsourced a Large Swiss German Speech-to-Text Dataset: “Die Schweizer Dialektsammlung”

Manuela Hürlimann, Michel Plüss, Marc Cuny, Alla Stöckli, Malgorzata Anna Ulasik, Manfred Vogel, Mark Cieliebak

Speech-to-Text is essential for many innovative applications such as automated customer support or automatic transcription of meetings. While this is a solved problem for languages such as English, Spanish or Chinese, there are currently no solutions for Swiss German because of the lack of suitable training data.

Our project “Die Schweizer Dialektsammlung” (“The Swiss Dialect Collection”) creates such a large training dataset using crowdsourcing by prompting users to speak a provided Standard German text in their dialect. Our goal is to collect at least 2’000 hours of Swiss German audio – equivalent to approximately 650’000 samples. We will discuss how we organised such a large-scale endeavour, and how we motivated the Swiss population to take part.

We will discuss the following aspects of the project:

Demonstrate the web application for data collection, based on the Mozilla CommonVoice open source code base.
Present our strategy for reaching and addressing various target groups so as to motivate them to participate in the data collection. We will show the different strategies (including gamification/rewards) and messages and evaluate their success.
Provide insights into the collected data (e.g. geographical distribution, quality) and how high-quality Speech-to-Text models for Swiss German can be trained on the collected data.

Organizations

ZHAW
FHNW
SpinningBytes AG

Automatic Classification of Service Desk Tickets

Michael Zemp, Don Tuggener, Stephen Cryan, Mark Cieliebak

Manual routing of support tickets (or service desk tickets) within a large organization is difficult, prone to error, and requires a lot of domain-specific knowledge. Even service desk agents struggle with finding the affected services and support groups. To help these central support units assign tickets, statistical models and neural networks can be trained based on actual historical support cases.

We present a solution implemented for the Julius Baer Group, a medium-sized internationally active private bank. The enterprise has multiple service desks for different sort of inquiries (business or general) and regions. They accept re-quests from the organization through different channels: phone, mail or intranet forms. Fortunately, they all share the same data model as well as a common ticketing system.

Results: Using the historic tickets, text classification with deep learning has shown a 92% validation accuracy in predicting the affected service and therefore enabling an initial routing of the tickets. Despite poor text quality and mixed languages, a macro F1-score of 0.75 is achieved over more than 300 classes. Especially preprocessing has shown a significant impact on how individual classes are being predicted. There is even more automation potential, as the tickets can also be classified into requests and incidents with a 93% accuracy. Finding the responsible support group itself that can resolve an incident has shown promising results, yet the available data itself lacks the needed information. It still succeeds to present a good selection of responsible groups, out of which the top 4 can resolve the issue with a probability of 96%.

Methods: Different neural networks such as CNNs, LSTMs and Transformers have been tried with different configurations as well as different tokenization techniques. Overall, trained word embeddings in combination with convolutional layers achieve the best results in accuracy and macro F1-score. Further pretraining the language model of RoBERTa, a state-of-the-art transformer-based ma-chine learning technique, outperformed the best CNN by 0.018 in macro F1-score. However, when it comes to real-world application and usability, their predictions suffer from the low text quality and lack the domain-specific terms.

Real-World Application: During development, the practicability was evaluated frequently by experienced service desk agents, who were provided the prediction features for their daily work. The developed models are used now on a daily basis on the production environment. One measurable benefit is that quality managers have less work in terms of adjusting tickets and can instead focus on their other tasks. The support group predictions, which are implemented in the live system as well, help service desk agents to route the tickets to the corresponding services and support groups.

Based on these successes, there is already the requirement to start fully automated labeling of inquiries which the bank receives through channels like email and intranet form.

Organizations

ZHAW
Julius Bär

Blackbox Testing for Conversational Systems

David Klaper

Conversational systems in production vary widely in quality. Testing such systems is the most important way to ensure quality. This talk focuses on raising awareness for testing conversational systems with insights from practical experience.

Testing NLP systems can be tricky. Interpreting the underlying models is difficult and in some cases they are not in our control. This hinders building a conversational solution that improves over time without breaking the existing understanding. We would like to share our approach to this problem and outline our key learnings from voice projects over various use cases. It focuses on testing without looking inside the model, so-called blackbox testing.

A central lesson is that blackbox testing can reveal meaningful insights into the model and training data. Suddenly failing tests are not just a nuisance but also a chance to find structural problems. Thus, there are some strategies for analyzing test failures and fixing them the right way.

Then, conversational systems can be tested with different purposes and levels. We show what tests we have right now and what additional test types we consider useful for the future.

Finally, conversational tests can serve as a crucial communication tool between developers, project managers and business experts. In contrast to testing other kinds of software, it is much more natural to formulate tests in a language easily understandable for a wider range of stakeholders than just developers.

Organizations

Paixon GmbH

Swiss Parliaments Corpus, an Automatically Aligned Swiss German Speech to Standard German Text Corpus

Michel Plüss, Lukas Neukom, Christian Scheller, Manfred Vogel

We present the Swiss Parliaments Corpus (SPC), an automatically aligned Swiss German speech to Standard German text corpus. The corpus is based on publicly available data of a Bernese parliament and consists of 293 hours of data. It was created using a novel forced sentence alignment procedure and an alignment quality estimator, which can be used to trade off corpus size and quality.

We trained Automatic Speech Recognition (ASR) models as baselines on different subsets of the data and achieved a Word Error Rate (WER) of 0.278 and a BLEU score of 0.586 on the SPC test set. The corpus is freely available for download.

Further reading: original paper

Organizations

FHNW

SST-BERT: Semantic Shift Tracing by Clustering in BERT-based Embedding Spaces

Vani Kanjirangat, Sandra Mitrović, Alessandro Antonucci, Fabio Rinaldi

Lexical semantic change detection (also known as semantic shift tracing) is the task of identifying words that have changed their meaning over time. The task of unsupervised semantic shift tracing in SemEval2020 is particularly challenging.

This task comprised two subtasks; The first one refers to a binary clustering task, where we have to identify whether a word gains or loses a sense over the time period. The second subtask was a ranking task, to measure the degree of lexical semantic change. Given the unsupervised setup, in this work, we propose to identify clusters among different occurrences of each target word, considering these as representatives of different word meanings or senses. As such, disagreements in the obtained clusters naturally allow quantifying the level of semantic shift per each target word in the four given target languages (German, English, Swedish, and Latin). To leverage this idea, clustering is performed on contextualized (BERT-based) embeddings of word occurrences. The ranking task is done using Shannon-Jensen distance.

The obtained results show that our approach performs well both measured separately (per language) and overall, where we surpass the provided SemEval baselines. The approach would be interesting with respect to any language, especially to understand the linguistic properties and variations of the words within the language and how they have evolved over time.

Original Paper: Kanjirangat, V., Mitrovic, S., Antonucci, A., & Rinaldi, F. (2020, December). SST-BERT at SemEval-2020 Task 1: Semantic Shift Tracing by Clustering in BERT-based Embedding Spaces. In Proceedings of the Fourteenth Workshop on Semantic Evaluation (pp. 214-221).

Code: Github Repo

Organizations

Dalle Molle Institute for Artificial Intelligence USI-SUPSI

What’s Hot in the News Today?

Claudia Schulz

To stay tuned about the daily news, you probably browse your favourite news outlet every morning. However, not everything you read is equally “hot”. Some of the articles you come across will be covered by other news providers too. These are the “hot topics” everyone will be discussing throughout the day. So which headlines are “hot”, i.e. most covered, among the thousands of articles published in print media as well as online every day?

To answer this question, we evaluated various off-the-shelf document clustering and topic extraction algorithms and libraries, all resulting in poor performance. We thus created our own pipeline to determine daily hot topics, using a mixture of state-of-the-art neural models and old-school NLP techniques to tackle the document clustering and text summarisation tasks.

In addition to performance results, this presentation covers challenges encountered, such as dealing with articles in multiple different languages, and lessons learned.

Organizations

Argus Data Insights

Re-Evaluating GermEval17 Using German Pre-Trained Language Models

Matthias Assenmacher, Alessandra Corvonato, Christian Heumann

The lack of a commonly used benchmark data set (collection) such as (Super-)GLUE (Wang et al., 2018, 2019) for the evaluation of non-English pre-trained language models is a severe shortcoming of current English-centric NLP-research. It concentrates a large part of the research on English, neglecting the uncertainty when transferring conclusions found for the English language to other languages.

We evaluate the performance of the German and multilingual BERT-based models currently available via the huggingface transformers library on the four tasks of the GermEval17 workshop. We compare them to pre-BERT architectures (Wojatzki et al.,2017; Schmitt et al., 2018; Attia et al., 2018) as well as to an ELMo-based architecture (Biesialska et al., 2020) and a BERT-based approach (Guhr et al., 2020) .

The observed improvements are put in relation to those for similar tasks and similar models (pre-BERT vs. BERT-based) for the English language in order to draw tentative conclusions about whether the observed improvements are transferable to German or potentially other related languages.

Further reading: original paper

Organizations

Ludwig-Maximilians-Universität München

Boosting the Quality of Experience of a Leadership Training System with Natural Language Processing

Oscar Lithgow-Serrano, Denis Broggini, Giancarlo Corti, Luca Chiarabini, Daniele Puccinelli, Fabio Rinaldi, Andrea Laus

SkillGym is a computer-based training system designed to improve communication skills by walking the end-user through a sequence of videos related to specific management situations.

As part of the Innosuisse-funded project “Boosting the SkillGym Quality of Experience with Artificial Intelligence” (BOOST), our team is developing a novel technological solution that leverages voice-based interaction and natural language processing to boost SkillGym’s Quality of Experience. This has presented various demanding challenges.

Here we present the practical path that we have followed to achieve our goal.
At each simulation step, the user is offered a selection of hints on what to say to the system. The hints are carefully crafted to move the simulation along based on the available pre-recorded video segments. The user is free to express herself in her own words.

In general terms, the objective is to select the pre-recorded video segment whose content provides the best fit as a response to the user utterance.
The freedom afforded to the user requires the system to leverage a multi-stage approach to sort out the practical difficulties of dealing with free natural language interaction.

The tasks addressed include the detection of unsuitable utterances, the detection and completion of incomplete utterances, the detection and correction of spurious words introduced by the voice-to-text transcription, and fine-grained utterance classification.

The implemented solutions range from simple heuristics based on grammatical features to fine-tuned deep learning models with some hybrid approaches in between.

Organizations

University of Applied Sciences of Southern Switzerland
Lifelike SA

NOVEL2GRAPH: Knowledge Graphs and Dynamic Embeddings for Narrative Text Understanding

Vani Kanjirangat, Alessandro Antonucci, Simone Mellace

Narrative Text Understanding aims the extraction and processing of aspects such as entity relations, semantic relatedness and temporal variations in the text. We focus on a specific class of narratives, the literary text, which offers a challenging benchmark due to the inherent complexity of language usage and the hurdles in gathering supervised information.

We propose an unsupervised approach to identify the possible characters and their relationships in the novels, which can be further used to model knowledge graphs and study the temporal evolutions of characters. The knowledge graph relations are identified using transformer based clustering techniques, while the dynamic evolutions are modelled using temporal embeddings. The initial pre-processing, character identifications and dealiasing utilize standard natural language processing techniques.

Preliminary tests and empirical analysis show that the clustering approach can successfully detect semantic relations between the characters, while the dynamic embeddings can depict meaningful temporal evolutions. Overall, we provide a challenging and unexploited benchmark for natural language understanding, which can be easily generalized to different domains. The tool is freely available to the community on Github.

Original Papers: Vani K, Simone Mellace, and Alessandro Antonucci. Temporal embeddings and transformer models for narrative text understanding. In Proceedings of Text2Story – Third Workshop on Narrative Extraction From Texts, 2020.
Simone Mellace, Vani K and Alessandro Antonucci. Relation Clustering in Narrative Knowledge Graphs. Proceedings of the first AI4Narratives Workshop, Yokohama, Japan, January 2021.

Organizations

Dalle Molle Institute for Artificial Intelligence USI-SUPSI

GPT3 – A Practitioner’s Guide

Pius von Däniken

The release of OpenAI’s GPT3 in early 2020 set a new bar for the size and capabilities of neural language models.
In this short introduction to GPT3 we will look at its capabilities beyond text generation and investigate its performance on some common NLP tasks.

We will also address other topics we consider important to practitioners, such as pricing, terms of
service, and the app approval process.

Organizations

ZHAW

COVID-19 Twitter Monitor: Aggregating and Visualizing COVID-19 Related Trends in Social Media

Joseph Cornelius, Tilia Ellendorff, Lenz Furrer, Fabio Rinaldi

Social media plays an increasingly important role for the Swiss general public in the acquisition of information and the forming of political opinions[A][B]. Thus, at times of the SARS-CoV-2 pandemic, platforms such as Twitter offer an essential source of information for the public, as they often provide the most up-to-date information due to their real-time nature. Within the NLP community, a pool of methods and tools already exists to process data from social media. The importance of this research area is shown by numerous conferences and workshops, like for example the Social Media Mining for Health shared task (SMM4H) [1]. However, to make the latest methods identified by the research community accessible to a non-expert audience, a user-friendly interface is required.

In our paper “COVID-19 Twitter Monitor: aggregating and visualizing COVID-19 related trends in social media” [2] presented at the 5th SMM4H, we introduced our platform that aggregates methods for analyzing COVID-19-related tweets and provides an interactive visualization of the results. By combining the different methods visually and interactively, interconnections among the obtained results can be revealed. Our contribution is intended to enrich the Swiss discussion about possibilities to facilitate the retrieval of COVID-19-related information from social media.

[A] Social Media Survey 2020
[B] Article in “Netzwoche”

[1] Klein, A., Alimova, I., Flores, I., Magge, A., Miftahutdinov, Z., Minard, A. L., … & Gonzalez, G. (2020, December). Overview of the fifth social media mining for health applications (# smm4h) shared tasks at coling 2020. In Proceedings of the Fifth Social Media Mining for Health Applications Workshop & Shared Task (pp. 27-36).
[2] Cornelius, J., Ellendorff, T., Furrer, L., & Rinaldi, F. (2020, December). COVID-19 Twitter Monitor: Aggregating and visualizing COVID-19 related trends in social media. In Proceedings of the Fifth Social Media Mining for Health Applications Workshop & Shared Task (pp. 1-10).

Organizations

Dalle Molle Institute for Artificial Intelligence USI-SUPSI
Department of Computational Linguistics, University of Zürich

Semi-supervised Machine Learning for Speech Recognition

Lukas Neukom, Christian Scheller

Current state-of-the-art models for speech recognition require large amounts of transcribed audio data to attain good performance. Recently, semi-supervised learning has been used to drastically improve the performance by using large unlabeled datasets in combination with a small amount of labeled data.

In this talk we will present two such approaches used for automatic speech recognition. First, we introduce a noisy student training method that uses iterative self-training with data augmentation. Second, we present wav2vec that proposes self-supervised pre-training on unlabeled data followed by supervised fine-tuning. We will then show how these techniques can be used to improve Swiss German Speech-To-Text models.

Organizations

FHNW

Spot The Bot: A Robust and Efficient Framework for the Evaluation of Conversational Dialogue Systems

Jan Deriu

The lack of time-efficient and reliable evaluation methods hamper the development of conversational dialogue systems (chatbots). Evaluations requiring humans to converse with chatbots are time and cost-intensive, put high cognitive demands on the human judges, and yield low-quality results.

In this work, we introduce Spot The Bot, a cost-efficient and robust evaluation framework that replaces human-bot conversations with conversations between bots. Human judges then only annotate for each entity in a conversation whether they think it is human or not (assuming there are humans participants in these conversations). These annotations then allow us to rank chatbots regarding their ability to mimic the conversational behavior of humans.

Since we expect that all bots are eventually recognized as such, we incorporate a metric that measures which chatbot can uphold human-like behavior the longest, i.e., Survival Analysis. This metric has the ability to correlate a bot’s performance to certain of its characteristics (e.g., fluency or sensibleness), yielding interpretable results. The comparably low cost of our framework allows for frequent evaluations of chatbots during their evaluation cycle.

We empirically validate our claims by applying Spot The Bot to three domains, evaluating several state-of-the-art chatbots, and drawing comparisons to related work. The framework is released as a ready-to-use tool.

Organizations

ZHAW

Battle of NLP Ideas

Brought to you by the Data Innovation Alliance

Explore your innovative ideas with fellow NLP enthusiasts and get a chance to turn your ideas into an Innosuisse-funded project! In this interactive session we will guide you through the initial steps of project ideation. It is also an opportunity to broaden your network and find partners for collaboration.

Organizations

Data Innovation Alliance

From research to production: Enabling real world applications with Conversational AI

Zenodia Charpy, Adam Henryk Grzywaczewski

In today’s NLP landscape, new papers and models are being published, and SOTA benchmarks beaten almost every week. However, the gap between research initiatives and production is vast and growing.

This talk will focus on the process of maturing research projects preparing them for production deployment. In particular we will highlight how this process is being applied to NVIDIA conversational AI stack including tools such as Nemo and Jarvis.

The described pipeline is composed of the following activities:

Graduating top performing research initiatives/models
Optimization for deployment (maximize throughput minimizing latency)
Packaging in enterprise ready SDKs maintaining the flexibility to bring your own data (BYOD) and model interoperability (allowing users to swap out models in production pipelines)

This process is the foundation for multitude of NVIDIA Software Products including NVIDIA Jarvis (platform for multimodal conversational AI), NVIDIA Clara Guardian (enabling smart hospitals and care home) or NVIDIA Maxine (supporting video conferencing through features such as Neural Image Compression, or Image Super Resolution).

Organizations

Nvidia

A Guided Tour Through the Transformers Landscape

Lewis Tunstall

Since their introduction in 2017, Transformers have become the de facto standard for tackling a wide range of NLP tasks in both academia and industry. However, with new architectures being released every week and a trend towards training 100-billion parameter models, it can be tricky to figure out which Transformer is best suited for your application.

In this talk I’ll present an overview of the Transformers landscape and discuss how Hugging Face’s open-source ecosystem is helping thousands of researchers and developers train these models effectively and share them with the wider community.

With an eye towards the future, I’ll introduce an ambitious new project called BigScience, which aims to bring hundreds of researchers together to develop an open-source large language model that can be used as a shared resource for the scientific community.

Organizations

Hugging Face

Predicting COVID-19 cases using Reddit posts and other online resources

Felix Drinkall, Janet Pierrehumbert

This paper evaluates the ability to predict COVID-19 caseloads in local areas using the text of geographically specific subreddits, in conjunction with other features.

The problem is constructed as a binary classification task on whether the caseload change exceeds a threshold or not. We find that including Reddit features, alongside other informative resources, improves the models’ performance in predicting COVID-19 cases.

On top of this, we show that exclusive use of Reddit features can act as a strong alternative data source for predicting a short-term rise in caseload due to its strong performance and the fact that it is readily available and updates instantaneously.

Further reading: original paper

Organizations

University of Oxford

Demos

Let your subject-matter experts train AI—an example for document review using PDFs (Juan Miguel Cejuela)
Abtippen war gestern – jetzt wird getögglt (David Imseng and Damian Murezzan)
PangeaMasker—adaptive anonymiser (Abel Meneses-Abad, Oscar Luis Garcell Martínez, Mercedes García-Martínez, Amando Estela, Laurent Bié, Aleix Cerdà, Ernesto Rodríguez and Manuel Herranz)
impresso—Integrating NLP Tools for Media Monitoring 200 Years of Historical Newspapers (Simon Clematide, Marten During and Maud Ehrmann)
Interscriber—Automatic Interview Transcription (Alla Stöckli, Nicola Good, Katsiaryna Mlynchyk, Pius von Däniken, Jan Deriu, Malgorzata Anna Ulasik, Nicolas Ganz, Flurin Gishamer, Amin K. Moghaddam, Don Tuggener, Marc Cuny, Iuliia Parfenova, Mark Cieliebak and Philippe Schläpfer)
Digital transformation in legal industry—showcased on Swiss Data Protection Act (DPA) (Philipp Thomann, Jacqueline Stählin and Christian Laux)

Multimodal Simultaneous Machine Translation

Lucia Specia

Simultaneous machine translation (SiMT) aims to translate a continuous input text stream into another language with the lowest latency and highest quality possible. Therefore, translation has to start with an incomplete source text, which is read progressively, creating the need for anticipation.

In this talk I will present work where we seek to understand whether the addition of visual information can compensate for the missing source context. We analyse the impact of different multimodal approaches and visual features on state-of-the-art SiMT frameworks, including fixed and dynamic policy approaches using reinforcement learning. Our results show that visual context is helpful and that visually-grounded models based on explicit object region information perform the best. Our qualitative analysis illustrates cases where only the multimodal systems are able to translate correctly from English into gender-marked languages, as well as deal with differences in word order, such as adjective-noun placement between English and French.

Organizations

Imperial College London

Programme

Monday 14.6

Tuesday 15.6

Wednesday 16.6