BERT with a masked language model head was trained on ~20,000 recipes from the AllRecipes dataset and YouCook II video annotations. The unlabelled dataset was created in the script fine-tune.py by combining allrecipes.json with the annotations in youucookii_annotations_trainval.json.
The co-occurence table of softmax probabilities found at bert_co-occurrence.csv was then clustered using KMeans. All the nouns in EPIC_noun_classes.csv and all the verbs in EPIC_train_action_labels.csv are accounted for.
To produce the table at bert_co-occurrence.csv, for each verb X and each noun Y, the average of a Bert MLM prediction for both "now let me show you how to X the [MASK]." and "now let me show you how to [MASK] the Y." were found.
- For clustering:
python3 -m lib.clustering.cluster_embeds {n_clusters} {noun_cap}- Example:
python3 lib.clustering.cluster_embeds 15 50 - Outputs: Logging statements of each cluster, some stats, and the nouns in each cluster.
- Example:
- For masked language modeling:
python3 -m cli.mlm {text} {results_length}- Example:
python3 -m cli.mlm "Next, we want to [MASK] the salmon into thin fillets." 3 - Outputs: Top k words most likely to fill in [MASK]
- Example:
- For masked language prediction:
python3 -m cli.mlm_prob {text} {word}- Example:
python3 -m cli.mlm_prob "Next, we want to [MASK] the salmon into thin fillets." "cut" - Outputs: softmax probability of target word filling in for [MASK] token.
- Example:
The full co-occurrence table can be found here. Below is an excerpt from the larger csv:
| pan | pan:dust | tap | plate | knife | bowl | spoon | cupboard | |
|---|---|---|---|---|---|---|---|---|
| take | 219 | 2 | 0 | 402 | 312 | 202 | 360 | 9 |
| put | 260 | 1 | 3 | 403 | 302 | 224 | 323 | 6 |
| open | 4 | 0 | 347 | 1 | 1 | 5 | 0 | 636 |
| close | 5 | 0 | 275 | 0 | 0 | 3 | 0 | 381 |
| wash | 223 | 0 | 7 | 263 | 219 | 128 | 187 | 1 |
| cut | 0 | 0 | 0 | 2 | 1 | 0 | 0 | 0 |
| mix | 91 | 0 | 0 | 0 | 1 | 3 | 8 | 0 |
| pour | 3 | 1 | 0 | 3 | 0 | 2 | 1 | 2 |
The BERT probability co-occurence table can be found here.
Below are heatmaps displaying a small sample of all the verb and noun co-occurences both in EpicKitchen, and extracted using BERT with a masked language model head. For the BERT heatmap, a noun/verbs softmax probability was obtained by averaging the output of the sentence "now let me show you how to [MASK] the [MASK].", where the [MASK] token probabilities were predicted for the verb and noun, respectively.
- Cluster 2: All things you can 'open' and 'close'
- Cluster 3 and Cluster 4: Smaller non-food tools
- Cluster 9: Ingredients you can 'roll', 'knead'
- Cluster 1: Pretty large cluster of all food items; no non-food nouns (aside from 'boiler')
coherance_score = sum(counts for 5 most common verbs) / len(all verb tokens in the cluster)
Kmeans with k = 15
Embeddings were clustered with k = 15.
The 'Cluster coherence score' is defined as:
sum(counts for 5 most common verbs) / len(all verb tokens in the cluster).
'Top Verbs' list the 5 most frequent verbs attached to the clustered nouns in EpicKitchens and their intra-cluster frequencies.
- handle
- stereo
- napkin
- ladder
- towel
- finger
- chair
- sticker
- tortilla
- spice
- coffee
- egg
- shell:egg
- waffle
- coriander
- crisp
- heart
- cereal
- tofu
- cream
- top
- berry
- mayonnaise
- breadcrumb
- pan
- sugar
- grass:lemon
- honey
- milk
- thyme
- biscuit
- alarm
- smoothie
- herb
- shirt
- soap
- hob
- lettuce
- almond
- syrup
- ketchup
- cake
- powder:coconut
- corn
- cover
- pie
- cucumber
- tea
- tarragon
- oatmeal
- banana
- leaf:mint
- cherry
- squash
- turmeric
- apple
- dust
- pepper
- pineapple
- flour
- pith
- bar:cereal
- flame
- mocha
- rim
- hummus
- sandwich
- switch
- risotto
- garni:bouquet
- cinnamon
- powder:washing
- bread
- nut:pine
- ginger
- butter
- mesh
- flake:chilli
- salt
- jambalaya
- pancake
- casserole
- pear
- base
- lemon
- melon
- fruit
- croissant
- skin
- mustard
- kettle
- tail
- avocado
- pesto
- leaf
- boiler
- yeast
- wrap
- aubergine
- coconut
- blueberry
- popcorn
- vinegar
- potato
- fishcakes
- raisin
- cumin
- oregano
- foil
- parsley
- chocolate
- seed
- part
- peach
- rubber
- rest
- ring:onion
- tahini
- pepper:cayenne
- sprout
- grape
- chilli
- yoghurt
- mouse
- ingredient
- paste:garlic
- cloth
- basil
- content
- mint
- butter:peanut
- cheese
- lead
- omelette
- rosemary
- carrot
- oat
- asparagus
- timer
- wine
- trousers
- jar
- door:kitchen
- bottle
- drawer
- package
- envelope
- container
- lid
- book
- box
- window
- capsule
- knife
- trouser
- opener:bottle
- spatula
- brush
- masher
- shaker:pepper
- backpack
- whetstone
- holder:filter
- grater
- squeezer:lime
- phone
- control:remote
- fork
- juicer:lime
- chopstick
- hand
- utensil
- presser
- tap
- peeler:potato
- remover:spot
- filter
- tablet
- floor
- desk
- shelf
- scale
- sheets
- table
- tray
- light
- grill
- rack:drying
- tablecloth
- spot
- wall
- rug
- plate
- stand
- wire
- paper
- board:chopping
- mat
- tongs
- button
- recipe
- soup
- gravy
- kitchen
- food
- mixture
- salad
- dressing:salad
- sauce
- salmon
- beef
- noodle
- fish
- shrimp
- pasta
- sausage
- chicken
- bacon
- tuna
- turkey
- ball
- dough
- roll
- boxer
- basket
- supplement
- slicer
- cup
- lime
- cap
- drink
- poster
- liquid
- water
- freezer
- mat:sushi
- air
- clip
- juice
- pan:dust
- nesquik
- stick:crab
- tab
- driver:screw
- microwave
- sink
- candle
- sponge
- clothes
- ladle
- toaster
- wrap:plastic
- olive
- leftover
- heater
- machine:washing
- tube
- muffin
- liquid:washing
- leek
- cork
- glass
- apron
- time
- caper
- quorn
- hat
- form
- can
- oven
- artichoke
- cd
- oil
- tissue
- fridge
- dish:soap
- rubbish
- bag
- beer
- lamp
- funnel
- bowl
- whisk
- bin
- stock
- cupboard
- sleeve
- scrap
- heat
- sock
- celery
- tomato
- courgette
- broccoli
- cabbage
- garlic
- spinach
- fire
- pizza
- pea
- dumpling
- vegetable
- onion
- onion:spring
- bean:green
- pot
- nutella
- kiwi
- cooker:slow
- kale
- curry
- paella
- mushroom
- burger:tuna
- rice
- meat
- salami
- knob
- plug
- processor:food
- chip
- power
- mortar
- jug
- straw
- glove:oven
- spoon
- support
- cutlery
- instruction
- paprika
- dishwasher
- pin:rolling
- pestle
- slipper
- label
- coke
- watch
- fan:extractor
- guard:hand
- grinder
- rinse
- maker:coffee
- colander
- cutter:pizza
- blender
- lighter
- scissors
- strainer
- alcohol
- towel:kitchen
- Use masked language model to disambiguate verbs not in our vocab
- Have probability threshold required to accept it?
- https://arxiv.org/pdf/1904.01766v2.pdf
- Use structure of
now let me show you how to [MASK] the [MASK]to extract verb/noun relations
- Use structure of
- Use either clusters or distance from verbs as prior assumptions about the actions allowed to certain nouns
- Use knowledge of a known object being able to be used in a certain way to guide interpretation of foreign object
- Fine-tuned BERT produces clusters with lower 'cluster coherence' scores, but much more relevant masked model predictions
- Specifically: standard
bert-base-uncasedLM is better at clustering appliances vs. non-appliances - Domain-specific recipe text likely brings these embeddings closer due to the more-frequent co-occurrence, as opposed to Wikipedia data
- Unlike with the MLM, we don't provide clustering task with a specific semantic relationship (i.e. similar verbs) to use.
- Specifically: standard
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent