-doc embeds via add comp model ->bc simple -compute cos(docs, queries) ->works q well -comp sim b/w doc/queries acc to cos() -!{...} -better to use cf not -doc+query via diff langs -!+helps monoling lang's as well -docs+queries live in diff langs -...is res. w/bridging lang? ->eng well conn=>other lang -...does eng bridging lang help? -..>Y, shared task in cross-ling. IR -!not so huge anymore, just a nice toy app. ->=this simple model -...- -idea, you can do more adv. approaches using embeddings -+another appl I like is word embedding in diff langs, rank words in issued word/source lang query -+pretty much all word embeddings evaluated on this -:1 interesting thing, typ2 models can outperform {...} -...improvement not spectacular -§Q:"#?" -§A:Not so many/not so bad -(...)70% acc -one of probs in sub-community is don't have ways to [...eval?] -eng-de quite close as lang. pair -ppl pub contradict results ->in every paper diff results/scorings -people improve 95%=>95.4%, another has 95.6 ->unconvinced good x-ling. eval -further q's what to do w/x-ling embeddings -...phrase-level{&c} -instead of conclusion here is summary slide -..."(Intrinsic)..." ---Q&A--- -disc this week: what to be other task, reach other domains and see if something could be done -...x-ling lexicon attainment -...more like Q to whole community -...should do more interesting questions like cross-ling.{...} -study on using NMT embeds to improve sem sim for eng -...each model 0.5 on sim lex -!not sure how compare -...1 prob is need Pearl data(?) -...good Q, have to think about -§as curiosity, when remove need for parallel corpus, have you tried triling model? -=>1 paper reported enrichment with other model, comb with other langs -...e(x) as more quality supervision -...unsure abt context but would e(x) -§has anyone ever put into generative model? -...see what spits out in Spanish -=>talking abt comb 2, -work on Bayesian approach -topic model, embeddings, -!unaware of {...} -typ4 popular published results? -=>think have plenty papers using this type methodology in work -egframework[other/own] -..love'.'{...} -§these words still compared against when many ideas sim-in fact same main principle-hugely resembles LSI -=>learn word vecs, -...compare skip-gram vecs with LSI -higher for skip-gram -2 sides same coin 'imo' -...roots in psych as LSI dev in psych -representation of brain -...same idea ! implemented same way -§any essence learning research -=>Platonic way? -...essence of word: king minus man + woman = queen -lang independent space -main assumptions of biling embedding space -cat in corresp embed space -...same pt in space interling.

Relational vector logic model pre-tendent upon essence

-all similar words are family relations -essence of word