Part of my series of notes from NAACL-HLT 2019 in Minneapolis.
- local (e.g. skipgram) vs. non-local (e.g. LDA) approaches for word embeddings
- graphs to capture relation between word & its context
- vertices as words, edge if co-occur in a context
- (isn’t this what GloVe does kinda?)
- edge weights according to non-local info (co-occurrence in documents)
- sampling strategy – prioritise local context
Factorising AMR generation through syntax
- paper here
- generate text from abstract meaning representation (AMR) graph
- AMR underspecifies some things, i.e. syntax
- core idea: predict syntax first, then text conditioned on both AMR and syntax tree
- some experiments on how much syntax is in AMR anyway
- answer: not too much
Now this is an architecture diagram I can get behind.
- want to use seq2seq, but AMR and parse tree are both graphs…
- sequencify them first
- some tweaks to ensure predict well-formed constituency trees
- copy mechanism (this is getting pretty standard)
- experiments: it works
- side effect: can generate paraphrases by sampling different syntax trees
- this is a cool idea
Crowdsourced Frame Disambiguation
- the speaker really does not believe discrete annotations, ever
- causes of annotator disagreement:
- workers are human and therefore flawed
- sentence / data issues
- target semantics - are these even appropriate?
- CrowdTruth – metrics etc.
- annotation disagreement is signal NOT noise!
- FrameNet – dataset of crowd-annotated frames
- then he kicked the cable – “as you can’t see here…”
Inoculation by Fine-Tuning
- paper here
- NLP system = training dataset + model architecture
- failures may be due to dataset or model weaknesses
- challenge datasets that break models trained on some original dataset
- inoculation: fine-tune models on some challenge data to better understand why they fail
- possible outcomes:
- data weakness – fine-tuning a little bit completely closes the gap on original vs. challenge dataset, original training data was missing something
- model weakness – doesn’t close gap, something up with the model
- artifacts / other – damages outcome on both original and challenge, something else is going on
- example from NLI – word overlap & spelling error challenge sets
- word overlap: dataset weakness (fine-tuning closes the gap)
- spelling errors: model weakness (gap doesn’t close even upon fine-tuning)
- also other examples, e.g. SQuAD
- see some other outcomes, learn things about dataset issues etc.
- method can be applied to any train/test distribution mismatch, transfer tasks, etc.
- I really liked this talk! clear readable slides, nice simple idea
Word Polysemy Aware Document Representation
- background: context sensitive / topic modeling, deep learning, doc2vec, weighted averaging
- focus on simple compositional methods
- works well for sentence embedding but not larger documents / multiple topics
- averaging vs. partition averaging
- cluster word embeddings
- average within cluster and concatenate across clusters
- improve method by taking into account word meanings in context, various other modelling improvements
- SOTA NLI models are getting quite good
- but do models reason like we think they do?
- e.g. quantitative reasoning
- introduce datasets – sourced from the wild and constructed in a controlled way