Featurised slots with `initial_value` get wrongly embedded in general rules

phoenix · November 6, 2022, 3:02pm

github.com/RasaHQ/rasa

Featurised slots with `initial_value` get wrongly embedded in general rules

opened 10:43AM - 31 Aug 21 UTC

samsucik

type:bug

area:rasa-oss

priority:high area:rasa-oss/ml/policies effort:research/4

### Rasa Open Source version 2.8.3 ### Rasa SDK version _No response_ ### …Rasa X version _No response_ ### Python version 3.8 ### What operating system are you using? OSX ### What happened? We encountered this issue a long time ago and thought we fixed it in #8161\. But looks like we didn't. If a featurised slot has an `initial_value` set, the slot gets included in rules that don't mention this slot explicitly. This is incorrect as those rules are then not applicable in states where the slot has a value different from the initial one. **Findings from initial investigation** _See bellow for an example setup to reproduce this issue._ In #8161, we introduced the `omit_unset_slots` flag, which, when training `RulePolicy`, [is used](https://github.com/RasaHQ/rasa/blob/0d97d427fd342b3f0fab4eaabc3d5169249dbf61/rasa/core/policies/rule_policy.py#L734) to exclude slots that are unset (i.e. set to their initial values) from general rules (i.e. from rules that don't mention those slots and should be applicable regardless of those slots' values). Unfortunately, the intended pattern with `omit_unset_slots` needs further work (and testing). What is going wrong is: 1. During data generation from rules, the generator [removes duplicated trackers](https://github.com/RasaHQ/rasa/blob/0d97d427fd342b3f0fab4eaabc3d5169249dbf61/rasa/shared/core/generator.py#L394) – deduplication is done by [creating tracker states (and caching them)](https://github.com/RasaHQ/rasa/blob/0d97d427fd342b3f0fab4eaabc3d5169249dbf61/rasa/shared/core/generator.py#L108) and subsequently hashing them. In this step, `omit_unset_slots` is switched off. 2. Subsequently, `RulePolicy` trains – [creates its rule lookup](https://github.com/RasaHQ/rasa/blob/0d97d427fd342b3f0fab4eaabc3d5169249dbf61/rasa/core/policies/rule_policy.py#L794), with `omit_unset_slots` switched on. In doing this, trackers need to be turned into states. And, since we have the states cached from the previous step, we [re-use them](https://github.com/RasaHQ/rasa/blob/0d97d427fd342b3f0fab4eaabc3d5169249dbf61/rasa/shared/core/generator.py#L106). And this is where things break – the previously created states incorrectly contain initial values of featurised slots because they were created with `omit_unset_slots=False`. I see a few options for possible fixes: * see if it makes sense to use `omit_unset_slots=True` also during tracker deduplication * don't use the caching * use caching but check that the cached states were created with the same value of `omit_unset_slots` * ... **Example setup to reproduce the issue** Command: <pre>rasa train --augmentation 0 rasa test</pre> The default config. Training data: <pre>version: "2.0" rules: * rule: goodbye steps: * intent: goodbye * action: utter_goodbye nlu: * intent: greet examples: | * hey * hello * hi * hello there * good morning * good evening * moin * hey there * let's go * hey dude * goodmorning * goodevening * good afternoon * intent: goodbye examples: | * cu * good by * cee you later * good night * bye * goodbye * have a nice day * see you around * bye bye * see you later</pre> Test stories: <pre>version: "2.0" stories: * story: goodbye - num_appts slot steps: * slot_was_set: * num_appts: multiple * intent: goodbye * action: utter_goodbye</pre> Domain: <pre>version: "2.0" intents: * greet * goodbye slots: num_appts: type: categorical initial_value: single values: * none * single responses: utter_goodbye: * text: "Bye"</pre> **Definition of done** * <span class="error">[x]</span> Verify that the cause of the issue is as outlined above * <span class="error">[x]</span> Write tests that catch the bug (turns out the tests in #8161 weren't sufficient) * [ ] Decide which way to fix it (involve CSE and Product Management) * [ ] Create followup issue: "Implement the fix, adding more tests if necessary"

I have been following this issue for some time now. Has there been a workaround that anyone has been able to figure out? The action prediction is getting affected in my case since there are some boolean and categorical slots with inital_value that I am unable to do away with. Adding the rules scenario under stories for TED to learn could be a way. But, it would be great if someone could help with a better approach.

phoenix · November 14, 2022, 6:17am

@sonam can you please help me out by tagging someone who would be relevant for this?

porti20 · November 14, 2022, 7:39am

Hi phoenix

With Rasa Release 3.3.0 there was a simmilar fix - maybe Upgrade your Rasa-Version is an option?

github.com/rasahq/rasa

Initial slot values break rule policy by being integrated into all rules by default

opened 08:30AM - 28 Jul 22 UTC

closed 09:40AM - 26 Aug 22 UTC

twerkmeister

type:bug

area:rasa-oss

### Rasa Open Source version 3.2.4 ### Rasa SDK version 3.2.0 ### Py…thon version 3.7 ### What operating system are you using? Linux, OSX ### What happened? Was trying to build a concept bot with some basic task switching logic and noticed I couldn't get certain rules to work. After some bug hunting I found that initial slot values are implicitly integrated into all rules. Thus, if you change the initial value during the conversation, none of your rules will apply anymore: 🥳🎉 ### Minimal example to reproduce: ### nlu ```yaml version: "3.1" nlu: - intent: greet examples: | - hey - hello - hi - hello there - good morning - good evening - moin - hey there - let's go - hey dude - goodmorning - goodevening - good afternoon - intent: goodbye examples: | - cu - good by - cee you later - good night - bye - goodbye - have a nice day - see you around - bye bye - see you later ``` ### rules ```yaml version: "3.1" rules: - rule: Say hi! steps: - intent: greet - action: utter_greet - rule: Say goodbye anytime the user says goodbye steps: - intent: goodbye - action: utter_goodbye ``` ### config: default ### domain ```yaml version: "3.1" intents: - greet - goodbye slots: has_said_hi: type: bool initial_value: false # commenting this out will make the rules work again mappings: - type: from_intent value: true intent: greet responses: utter_greet: - text: "Hey! How are you?" utter_goodbye: - text: "Bye" session_config: session_expiration_time: 60 carry_over_slots_to_new_session: true ``` When you have a look at the rules that are produced with and without initial value for the slot `has_said_hi` you will see the following: ### rules in the trained rule_policy.json (from the model tar.gz) without initial value: ``` "rules": { "[{\"prev_action\": {\"action_name\": \"action_listen\"}, \"user\": {\"intent\": \"goodbye\"}}]": "utter_goodbye", "[{\"prev_action\": {\"action_name\": \"action_listen\"}, \"user\": {\"intent\": \"goodbye\"}}, {\"prev_action\": {\"action_name\": \"utter_goodbye\"}, \"user\": {\"intent\": \"goodbye\"}}]": "action_listen", "[{\"prev_action\": {\"action_name\": \"action_listen\"}, \"user\": {\"intent\": \"greet\"}}]": "utter_greet", "[{\"prev_action\": {\"action_name\": \"action_listen\"}, \"user\": {\"intent\": \"greet\"}}, {\"prev_action\": {\"action_name\": \"utter_greet\"}, \"user\": {\"intent\": \"greet\"}}]": "action_listen" }, ``` ### rules in the trained rule_policy.json (from the model tar.gz) with initial value: ``` "rules": { "[{\"prev_action\": {\"action_name\": \"action_listen\"}, \"slots\": {\"has_said_hi\": [1.0, 0.0]}, \"user\": {\"intent\": \"goodbye\"}}]": "utter_goodbye", "[{\"prev_action\": {\"action_name\": \"action_listen\"}, \"slots\": {\"has_said_hi\": [1.0, 0.0]}, \"user\": {\"intent\": \"goodbye\"}}, {\"prev_action\": {\"action_name\": \"utter_goodbye\"}, \"slots\": {\"has_said_hi\": [1.0, 0.0]}, \"user\": {\"intent\": \"goodbye\"}}]": "action_listen", "[{\"prev_action\": {\"action_name\": \"action_listen\"}, \"slots\": {\"has_said_hi\": [1.0, 0.0]}, \"user\": {\"intent\": \"greet\"}}]": "utter_greet", "[{\"prev_action\": {\"action_name\": \"action_listen\"}, \"slots\": {\"has_said_hi\": [1.0, 0.0]}, \"user\": {\"intent\": \"greet\"}}, {\"prev_action\": {\"action_name\": \"utter_greet\"}, \"slots\": {\"has_said_hi\": [1.0, 0.0]}, \"user\": {\"intent\": \"greet\"}}]": "action_listen" } ``` As you can see the initial value for `has_said_hi` was incorporated into all rules as `[1.0, 0.0]` which is the boolean slot representation for `false`. What you will experience when you train this bot is the following: the greet rule won't work. What happens is that the slot has_said_hi get's filled, and afterwards there are no applicable rules and the bot will reset the previous user message (unsetting the slot again). So you can get a good bye but never get the bot to greet you back. I think it's quite a serious bug, as it basically renders rules almost useless when using slots with initial values and it is super surprising that initial values are causing this. Instead, initial values should not be incorporated into rules. ### Command / Request ```shell `rasa train` with the provided example config `rasa shell -vv` to interact with the bot and see debug info. ``` ### Relevant log output ```shell Bot loaded. Type a message and press enter (use '/stop' to exit): Your input -> hi 2022-07-28 10:27:30 DEBUG rasa.core.lock_store - Issuing ticket for conversation 'a4fa8fc9e981468499d002ae25a30667'. 2022-07-28 10:27:30 DEBUG rasa.core.lock_store - Acquiring lock for conversation 'a4fa8fc9e981468499d002ae25a30667'. 2022-07-28 10:27:30 DEBUG rasa.core.lock_store - Acquired lock for conversation 'a4fa8fc9e981468499d002ae25a30667'. 2022-07-28 10:27:30 DEBUG rasa.core.tracker_store - Could not find tracker for conversation ID 'a4fa8fc9e981468499d002ae25a30667'. 2022-07-28 10:27:30 DEBUG rasa.core.processor - Starting a new session for conversation ID 'a4fa8fc9e981468499d002ae25a30667'. 2022-07-28 10:27:30 DEBUG rasa.core.processor - Policy prediction ended with events '[]'. 2022-07-28 10:27:30 DEBUG rasa.core.processor - Action 'action_session_start' ended with events '[<rasa.shared.core.events.SessionStarted object at 0x7f572869afd0>, ActionExecuted(action: action_listen, policy: None, confidence: None)]'. 2022-07-28 10:27:30 DEBUG rasa.core.processor - Current slot values: has_said_hi: False session_started_metadata: None 2022-07-28 10:27:30 DEBUG rasa.engine.runner.dask - Running graph with inputs: {'__message__': [<rasa.core.channels.channel.UserMessage object at 0x7f56985cb210>]}, targets: ['run_RegexMessageHandler'] and ExecutionContext(model_id='8923f28deae949efb9c00d8482a594b5', should_add_diagnostic_data=False, is_finetuning=False, node_name=None). 2022-07-28 10:27:30 DEBUG rasa.engine.graph - Node 'nlu_message_converter' running 'NLUMessageConverter.convert_user_message'. 2022-07-28 10:27:30 DEBUG rasa.engine.graph - Node 'run_WhitespaceTokenizer0' running 'WhitespaceTokenizer.process'. 2022-07-28 10:27:30 DEBUG rasa.engine.graph - Node 'run_RegexFeaturizer1' running 'RegexFeaturizer.process'. 2022-07-28 10:27:30 DEBUG rasa.engine.graph - Node 'run_LexicalSyntacticFeaturizer2' running 'LexicalSyntacticFeaturizer.process'. 2022-07-28 10:27:30 DEBUG rasa.engine.graph - Node 'run_CountVectorsFeaturizer3' running 'CountVectorsFeaturizer.process'. 2022-07-28 10:27:30 DEBUG rasa.engine.graph - Node 'run_CountVectorsFeaturizer4' running 'CountVectorsFeaturizer.process'. 2022-07-28 10:27:30 DEBUG rasa.engine.graph - Node 'run_DIETClassifier5' running 'DIETClassifier.process'. 2022-07-28 10:27:30 DEBUG rasa.engine.graph - Node 'run_EntitySynonymMapper6' running 'EntitySynonymMapper.process'. 2022-07-28 10:27:30 DEBUG rasa.engine.graph - Node 'run_ResponseSelector7' running 'ResponseSelector.process'. 2022-07-28 10:27:30 DEBUG rasa.nlu.classifiers.diet_classifier - There is no trained model for 'ResponseSelector': The component is either not trained or didn't receive enough training data. 2022-07-28 10:27:30 DEBUG rasa.nlu.selectors.response_selector - Adding following selector key to message property: default 2022-07-28 10:27:30 DEBUG rasa.engine.graph - Node 'run_FallbackClassifier8' running 'FallbackClassifier.process'. 2022-07-28 10:27:30 DEBUG rasa.engine.graph - Node 'domain_provider' running 'DomainProvider.provide_inference'. 2022-07-28 10:27:30 DEBUG rasa.engine.graph - Node 'run_RegexMessageHandler' running 'RegexMessageHandler.process'. 2022-07-28 10:27:30 DEBUG rasa.core.processor - Received user message 'hi' with intent '{'name': 'greet', 'confidence': 1.0}' and entities '[]' 2022-07-28 10:27:30 DEBUG rasa.core.processor - Logged UserUtterance - tracker now has 4 events. 2022-07-28 10:27:30 DEBUG rasa.core.actions.action - Validating extracted slots: has_said_hi 2022-07-28 10:27:30 DEBUG rasa.core.processor - Default action 'action_extract_slots' was executed, resulting in 1 events: SlotSet(key: has_said_hi, value: True) 2022-07-28 10:27:30 DEBUG rasa.engine.runner.dask - Running graph with inputs: {'__tracker__': <rasa.shared.core.trackers.DialogueStateTracker object at 0x7f56c80b69d0>}, targets: ['select_prediction'] and ExecutionContext(model_id='8923f28deae949efb9c00d8482a594b5', should_add_diagnostic_data=False, is_finetuning=False, node_name=None). 2022-07-28 10:27:30 DEBUG rasa.engine.graph - Node 'rule_only_data_provider' running 'RuleOnlyDataProvider.provide'. 2022-07-28 10:27:30 DEBUG rasa.engine.graph - Node 'domain_provider' running 'DomainProvider.provide_inference'. 2022-07-28 10:27:30 DEBUG rasa.engine.graph - Node 'run_MemoizationPolicy0' running 'MemoizationPolicy.predict_action_probabilities'. 2022-07-28 10:27:30 DEBUG rasa.core.policies.memoization - Current tracker state: [state 1] user intent: greet | previous action name: action_listen 2022-07-28 10:27:30 DEBUG rasa.core.policies.memoization - There is no memorised next action 2022-07-28 10:27:30 DEBUG rasa.engine.graph - Node 'run_RulePolicy1' running 'RulePolicy.predict_action_probabilities'. 2022-07-28 10:27:30 DEBUG rasa.core.policies.rule_policy - Current tracker state: [state 0] slots: {'has_said_hi': (1.0, 0.0)} [state 1] user text: hi | previous action name: action_listen | slots: {'has_said_hi': (1.0, 1.0)} 2022-07-28 10:27:30 DEBUG rasa.core.policies.rule_policy - There is no applicable rule. 2022-07-28 10:27:30 DEBUG rasa.core.policies.rule_policy - Current tracker state: [state 0] slots: {'has_said_hi': (1.0, 0.0)} [state 1] user intent: greet | previous action name: action_listen | slots: {'has_said_hi': (1.0, 1.0)} 2022-07-28 10:27:30 DEBUG rasa.core.policies.rule_policy - There is no applicable rule. 2022-07-28 10:27:30 DEBUG rasa.engine.graph - Node 'run_TEDPolicy3' running 'TEDPolicy.predict_action_probabilities'. 2022-07-28 10:27:30 DEBUG rasa.engine.graph - Node 'run_UnexpecTEDIntentPolicy2' running 'UnexpecTEDIntentPolicy.predict_action_probabilities'. 2022-07-28 10:27:30 DEBUG rasa.engine.graph - Node 'select_prediction' running 'DefaultPolicyPredictionEnsemble.combine_predictions_from_kwargs'. 2022-07-28 10:27:30 DEBUG rasa.core.policies.ensemble - Made prediction using user intent. 2022-07-28 10:27:30 DEBUG rasa.core.policies.ensemble - Added `DefinePrevUserUtteredFeaturization(False)` event. 2022-07-28 10:27:30 DEBUG rasa.core.policies.ensemble - Predicted next action using RulePolicy. 2022-07-28 10:27:30 DEBUG rasa.core.processor - Predicted next action 'action_default_fallback' with confidence 0.30. 2022-07-28 10:27:30 DEBUG rasa.core.processor - Policy prediction ended with events '[<rasa.shared.core.events.DefinePrevUserUtteredFeaturization object at 0x7f56982be350>]'. 2022-07-28 10:27:30 DEBUG rasa.core.processor - Action 'action_default_fallback' ended with events '[<rasa.shared.core.events.UserUtteranceReverted object at 0x7f56985a1410>]'. 2022-07-28 10:27:30 DEBUG rasa.core.processor - Current slot values: has_said_hi: False session_started_metadata: None 2022-07-28 10:27:30 DEBUG rasa.core.processor - Predicted next action 'action_listen' with confidence 1.00. 2022-07-28 10:27:30 DEBUG rasa.core.processor - Policy prediction ended with events '[]'. 2022-07-28 10:27:30 DEBUG rasa.core.processor - Action 'action_listen' ended with events '[]'. 2022-07-28 10:27:30 DEBUG rasa.core.lock_store - Deleted lock for conversation 'a4fa8fc9e981468499d002ae25a30667'. Your input -> ``` ```

Greetings

phoenix · November 14, 2022, 8:42am

@porti20, thanks for the help. It would take some time to migrate to Rasa 3.3, do you know of any other workaround?

porti20 · November 14, 2022, 8:55am

If you are looking for a workaround, then work around your initial values . There is no real need for initial values when you could initialize your slots on first usage oder the __init__.py

phoenix · November 14, 2022, 9:11am

Hey @porti20, Consider I have an influential boolean slot “isLoggedIn” and even a categorical slot, so this boolean slot is having an initial value by default and that gets featurised. This goes the same for the categorical slot. So, what could be a workaround for these initial_values?

porti20 · November 14, 2022, 9:37am

I’m not an expert on Rasa 2.x, but strong types like Float or List had no default values in 3.3 if not explicit given. Without an inital value the value ist “None”. That should not trigger the mentioned bug. It would be no supprise to me if this behaviour is the same in 2.x. You could test this easy in a custom action.

Topic		Replies	Views
Rasa x slot bug [Deprecated] Rasa X Community Edition	12	738	July 10, 2019
Slot initial_value Rasa Open Source	3	919	November 10, 2021
Are there good examples of a bot using the Rule Policy with featurized slots? Rasa Open Source	2	667	November 24, 2020
Slot setting Rasa Open Source	3	545	July 14, 2020
SlotSet aborts execution of a rule Rasa Open Source	1	260	June 29, 2022

Featurised slots with `initial_value` get wrongly embedded in general rules

Related Topics