Identifying Negation in the DGS Corpus Graz, 2019-05-03 Marc - PowerPoint PPT Presentation

Identifying Negation 
 in the 
 DGS Corpus Graz, 2019-05-03 Marc Schulder, Thomas Hanke Universität Hamburg {marc.schulder,thomas.hanke}@uni-hamburg.de

Negation Devices 
 in Sign Languages • Negation particles ✔ • Negation content words ✔ • Manual negation morphemes ( ✔ ) • Headshake (( ✔ )) • Facial expression ☹ 3

Negation Particles • Words like “no”, “not”, “without”, etc. • Lexemes are part of core annotation. • Small set of words, easily listed. To Do 4

Negation Content Words • Words like “destroy”, “prevent”, etc. • Large set of words (>1000). • Lists of negation content words available for English (Schulder et al. 2017, 2018-LREC, ...) and German (Schulder et al. 2018-COLING). • Lists can be mapped to new languages using bilingual dictionaries or bilingual 
 word embeddings (Schulder et al. 2018-COLING). To Do 5

Negation Morphemes • Small set of morphemes, 
 e.g. alpha negation. • Restricted set of 
 compatible lexemes. CAN alph cannot • Approach: Inspect all tokens of these lexemes and make sure negation morphemes are annotated as qualifiers. Ongoing... 6

Headshake • Not part of core annotation. • But annotators were asked to add comments about further important observations. • Result: 
 >7000 comments mentioning headshakes. 7

Headshake + Lexeme Emphasise negation Indicate negation NO BRING no not brought 8

Headshake + Phrase HS negates phrase TOGETHER FIT TOGETHER NOT It has nothing to do with each other at all 9

Non-negating Headshake HS indicates negative sentiment ALL OFF - CLOSE TO - CLOSE All of them have been closed down 10

Uses of Headshake • Emphasise existing negation • Negate a word • Negate a phrase • Indicate negative sentiment • Correction • Backchanneling 11

Manual Annotation 
 is slow, so… • Approach 1: 
 Use German translations To Do • Approach 2: 
 Use the visual domain To Do 12

Negation in Translation • Corpus contains German translations • Source is signed communication • Negation in German most likely caused by negation in DGS ➡ If translation contains negation, but DGS contains no negation lemma/morpheme, headshake is likely. 13

Into the Visual Domain: OpenPose (CMU) 14

OpenPose 2017 15

OpenPose 2018 16

But then OpenPose 
 is slow as well… • 3 camera perspectives per recording. • 1 hour recording = 87 hours processing 
 (double-GPU machine) • For our corpus this results in 
 a processing time of 5 ½ years. • 4 months on a High Performance Cluster. 17

Detecting Headshakes in OpenPose Data Track movement of the nose, 
 relative to face contour. 18

Detecting Headshakes in OpenPose Data 1.Run Open Pose. 2.Train a neural net classifier to • detect headshakes in time series data; • determine duration of headshakes. 19

Neural Net Training Challenges • Need annotator comments to train classifier, but time spans of comments are unreliable: • span is for sign, not headshake; • comment combines two observations, 
 e.g. “constructed action + headshake”. ➡ Comments indicate existence of headshake, but not time span. ➡ Translations may fulfil a similar function. 20

Outlook • Lists: Negation particles 
 Negation content words. • Annotation: Negation Morphemes. • Visual Detection: Headshakes. 
 (OpenPose, neural nets, 
 annotator comments, translations) ➡ Compare this “joint effort” with 
 detailed gold standard annotation. 21

Thank you very much 
 for your attention!