We work on a pipeline to retrieve and rank absorption-rich user reviews from a large, unlabeled document dump (6+ million reviews in English), in order to allow for the preselection of subsets of the dump that undergo manual annotation. We fine-tuned BERT (Devlin et al., 2018) for a supervised absorption detection task on 16k review sentences absorption-annotated by us (Absorption vs. Nonabsorption), and evaluated it on a held-out dataset of 149 reviews, achieving .75 macro F1 mean (support: 1,011 vs. 3,510 sentences).
Our current focus was to create a model that aggregates sentence level prediction scores on the document level. To this end, BERT’s sentence level absorption probabilities were averaged per review and were used to train a linear regression model on the full corpus to predict Absorption Richness, defined as the proportion of sentences annotated as expressing absorption in a review. Review-level Absorption Richness regression lowers classification error relative to the baseline, defined as the review-level proportion of absorption classifications by taking the argmax of BERT’s logits (Mean Average Errors of .08 vs. .11 and Spearman correlation of .73 vs. .65, respectively).
The increase of the Spearman’s rank correlation coefficient directly expresses that a review ranking by linear regression predictions corresponds more closely to the ground truth ranking than a ranking solely based on BERT. We utilize the regression model in Absorption-Richness-based document filtering, to facilitate the benchmarking and analysis of social reading reviews in our large document dump.
1DH Lab, University of Basel
2Research Institute for Linguistics, Hungarian Academy of Sciences
Organization:1Novartis AG, Basel2University of Zurich3IDSIA, Dalle Molle Institute for Artificial Intelligence, Lugano
Recent improvements in speech technology enable its increasing use in a range of applications, including chatbots, online speech translation and smart home devices, among others. While speech technology already achieves strong results for standardised languages, for languages without orthography, with high regional variation and limited training resources, such as Swiss German, it remains a considerable challenge. A high degree of dialectal variability combined with a lack of standardisation leads to extremely sparse data that decreases the quality of alignments between the acoustic signal and its labels and, therefore, the final accuracy.
To tackle the challenge of speech-to-text for Swiss German, we built a speech recognition system using an adapted Kaldi toolkit recipe on multi-dialectal speech data from the ArchiMob corpus. The system was separately trained on two types of writing in the target texts: a) an approximate acoustic transcription that provides a close correspondence between labels and the acoustic signal and b) a normalised writing that potentially reduces the lexical variability. We find that the system trained on the normalised transcriptions currently achieves better results in word error rate (40.81% vs. 54.39%) but underperforms the system trained on the acoustic transcriptions on the character level (character error rate) (23.19% vs. 22.19%). We investigate possible improvements of both approaches and present the outcomes.
Organization:University of Zurich
Detecting instances of negation in text is crucially important for several applications, yet it is often neglected. Several decades of research in automated negation detection have not yet provided a reliable solution, especially in a multilingual context. Negation scope resolution poses particular challenges since identifying the scope of influence of a negation cue in a sentence requires a deeper level of naturallanguage understanding. Little work has been done on negation scope resolution in languages other than English. Meanwhile, transfer learning is in wide use and large multilingual models are available to the public. This paper explores the feasibility of a cross-lingual transfer-learning approach to negation scope resolution. Preliminary experiments with the Multilingual BERT model and data in English, French, and Spanish show solid results with the highest F1-score 84.73 on zero- shot transfer between English and French.
AbstractNatural Language Processing has seen an explosion of interest in recent years, with many industrial applications relying on key technological developments in the field. However, Natural Language Understanding (NLU) – which requires the machine to get beyond processing strings and involves a deep, semantic level – is particularly challenging due to the pervasive ambiguity of language.In this talk I will present recent research at the Sapienza NLP group on multilingual NLU, including work on new multilingual sense embeddings, and novel neural approaches to word sense disambiguation and semantic role labeling which scale across languages easily and achieve state-of-the-art performance thanks to the integration of deep learning and explicit knowledge.
Organization:Netlive IT AG, Teufen
Organization:1 McGill University, Montreal2 Polytechnic de Montreal
Organization:1 Hieronymus AG, Zurich2 SIA Tilde3 Pangea MT, València
Organization:1Technische Universität Hamburg1Universität Hamburg
Organization:1Karakun AG, Basel2DSwiss AG, Zurich
Organization:1Center for the Study of Language and Society, University of Bern2Department of Psychology, University of Zurich
Organization:University of Applied Sciences of the Grisons
Organization:Schweizerisches Idiotikon, Zurich
Organization:1deepset, Berlin2Bayerische Staatsbibliothek München, Digital Library/Munich Digitization Center
Organization:1Institute for the German Language, Leibniz2Julius-Maximilans-Universität Würzburg
Organization:1ZHAW Zurich University of Applied Sciences2SpinningBytes, Winterthur3Propulsion Academy, Zurich
Organization:1University of Applied Sciences of Southern Switzerland2Lifelike SA, Chiasso
Organization:Berner Fachhochschule, Technik und Informatik
Organization:Deutsches Forschungszentrum für Künstliche Intelligenz (DFKI), Kaiserslautern, Germany
Organization:1University of Zurich2Roche, Basel
Organization:FHNW University of Applied Sciences Northwestern Switzerland
Organization:Zurich University of Applied Sciences
Organization:Swiss National Science Foundation
Organization:1University of Zurich2VetSuisse Bern
Organization:1Contexity AG, Winterthur2GRZ IT Center GmbH, Linz
Organization:1ZHAW Zurich University of Applied Sciences2UNED, Madrid3Euskal Herriko Unibertsitatea/Universidad del Paìs Vasco
Organization:University of Wuerzburg
Organization:1Institute of Computational Linguistics, University of Zurich2University of Zurich
Organization:1HEIG-VD / HES-SO, Yverdon2University of Lausanne
Organization:Hochschule Hannover University of Applied Sciences and Arts
Organization:1IDSIA, Lugano2Stagend, Lugano
Organization:Department of Computational Linguistics, University of Zurich
Organization:FHNW University of Applied Sciences and Arts Northwestern Switzerland