Dialogflow

Removing feature requests again

Users, of course, should not only be able to add recipe preferences but, when recognizing, for example, there are few or no recipes to their liking that meet the preferences they expressed, should also be able to remove them again to look for other options. They should, for example, be able to say 1. “I wish to remove the preference for vegetables”, 2. “Forget about excluding broccoli”, and 3. “Can you delete the cuisine preference?”. Of course, they could already get rid of these filters by expressing the opposite, and say, for example, “I would like recipes with broccoli”. However, instead of simply removing the constraint to exclude broccoli, that would also introduce a preference for recipes with broccoli (it is useful to better understand this point to check again how the removal of conflicting requests has been implemented in the ingredient_hierarchies.pl file).

It is worthwhile to briefly analyze the user expression examples above in more detail. You should note that in the first two expressions, the user only mentions an ingredient (type) but not the type of feature or filter (if you recall, the first expression would result in a filter ingredienttype='vegetables' and the second in a filter excludeingredient='broccoli'). In other words, the user mentions a value for a specific type of filter but not the type. In contrast, in the third expression, the user only indicates a feature or filter type (cuisine) but the specific cuisine the preference is about. For this last expression, things get even more complex, as the request might be about excluding a cuisine (e.g., excludecuisine='chinese') rather than including the cuisine. In other words, the request might be about the excludecuisine feature and not about the cuisine feature (your Dialogflow agent will not be able to deal with this kind of subtlety as it does not keep track of the feature requests that have been made; it will be up to your MARBEL Dialog management agent to handle this).

To be able to make sense of user requests asking to remove a value or a filter type, you should add two new intents to your Dialogflow agent. Call the first one deleteFilterValue and the second one deleteParameter (for the latter intent, the name stems from the fact that Dialogflow calls the entity types that have been supplied by the end-user parameters). Making these intents robust, i.e. for ensuring that the potential endless variety of user expressions will be recognized by your agent, you should add training phrases for all the different entity types and parameters/features/filters that a user could ask the agent to delete.

You should annotate the training phrases for deleteFilterValue with existing entity types but use different parameter names to indicate that the values captures should be removed from the feature requests stored in the agent’s conversational memory. You tag the entity in your training phrases as you did before (e.g. using @ingredient) but then you should change the parameter name to something of the form entitynameDel (we suggest adding Del at the end of each parameter that you already have to facilitate relating them to each other). So, for example, you would add new parameter names for ingredientDel, cuisineDel, durationDel, etc.

For the deleteParameter intent, you should introduce a new entity type called @filtertype (with corresponding parameter name) and add as entry entities all the features you have introduced, such as ingredient, cuisine, etc. You may also wish to consider any synonyms that a user might use to refer to any of these features. Again, make sure to include a sufficient number of training phrases to make recognition of this intent robust.

Expressing that you are done

A user should also be able to say it does not want to provide the agent with more information to go on (which does not mean the agent should accept that in all cases!). As we have implemented things thus far, we would expect a user to respond negatively to an agent request to elaborate on what they are looking for:

A: Can you elaborate on what you're aiming for in your recipe?

U: I don't want to add anything else.

Recall that we used the agent intent ackFilter to implement the first move in this fragment as part of the a21featureRequest pattern. If a user complies with the agent’s request, that would trigger another a21featureRequest pattern. If a user does not comply but rather responds negatively, we can also view that as a disconfirmation move. We argue that we can reuse that intent to capture negative responses to the agent’s request. A user, for that matter, could also simply respond to the agent’s move by saying “No, I won’t” or something similar. We implemented this intent before, but we should extend it to now also cover the variety of user expressions a user could use to say they do not want to add anything else. So, you should add training phrases to make sure these expressions are covered too.

Prolog and Patterns

A user does not want to specify more recipe constraints

The pattern that we want to implement to enable a user to express they do not want to provide more information is the following:

(A: Can you elaborate on what you're aiming for in your recipe?)

U: I don't want to add anything else.

A: OK. Here is a list of recipes that you choose from. [Alternative: Sorry, there are still too many recipes left to show them all. Please add more preferences.]

The first move has been put between brackets because, as explained above, it does not belong to the pattern that we want to implement but to the a21featureRequest pattern. The new pattern is named a21noMoreFilters, as it fits best in the A2 category of Request patterns in Moore and Arar’s taxonomy. It starts with a disconfirmation move of the user and continues with an agent intent that either does or does not grant the user’s request.

But what would granting the user’s request mean? The idea is to somewhat lift our previous restriction of only showing pictures of the list of filtered recipes when there are no more than 15 recipes left. The agent can, for example, show the list of filtered recipes if there are a 100 recipes or fewer left when a user indicates they do not want to provide more information. It does not seem reasonable to show many more recipes, as that does not seem to align well with the ambition to develop a conversational first agent. In other words, our agent should not grant the request if the recipe list still is more than a 100 recipes long.

The agent can then allow the user to see the list with the pictureGranted response, or deny with pictureNotGranted. If the agent grants the user's request, the pattern should end with the agent updating its memory using '[agent, update(['show'='true'])]'. This triggers the showing of recipes by updating the agent's memory.

In responses.pl we have to make two rules, for 'grantPicture' and 'pictureNotGranted'. The recipesFiltered/1 predicate provides a list of the filtered recipes. Fill in the two rules with the corresponding conditions.

Implement the a21noMoreFilters pattern in the patterns.pl file.

Add this predicate to the responses.pl file in the indicated spot.

% used in a21removeKeyFromMemory for handling deleteParameter. 
text(removedSpecificFilter(DelFilter), Txt) :- 	convert_to_string(DelFilter, Str1), 
	string_concat("I removed the ", Str1, Str2), 
	string_concat(Str2, " requirement.", Txt).

featureInquiry intent

Multiple text predicates take the argument feature inquiry in responses.pl. Each feature inquiry response is different based on how many recipes there are left to filter.

First Rule for Large Number of Recipes

We want to prompt the user for more specific preferences if there are too many recipes. The rule will be triggered when the length of the list of recipes is greater than 890.

• Start with the text/2 predicate, where the first argument is featureInquiry.

• The second argument will be the response text "What kind of recipe would you like?".

• Use the recipesFiltered/1 predicate to get the list of recipes.

• Use the length/2 predicate to determine the number of recipes.

• Set a condition that the length L must be greater than 890.

Second Rule for a Moderate Number of Recipes

If there's a moderate number of recipes (more than 15 but fewer than 891), we ask the user to add more preferences unless a certain memory key-value pair is set.

• Repeat the initial steps of the first rule.

• The response text will change to "What other preference would you like to add?".

• Set the length condition to be less than 891 and greater than 15.

• Include a condition using not/1 to check that the memory does not have the key-value pair 'show', 'true'.

Third Rule for No Recipes

If there are no recipes after filtering, we want to inform the user to remove some requirements.

• Use the recipesFiltered/1 predicate again.

• The response text will be "There are no recipes, please remove more requirements.".

• Check that the length L is equal to 0 using L=0.

Fourth Rule for a Small Number of Recipes or a Show Command

This rule handles two scenarios: either there's a small number of recipes, or the user has explicitly asked to show the recipes despite the number.

• Start with the text/2 predicate for featureInquiry.

• The response text will be "Here are some recipes that fit your requirements.".

• Use the recipesFiltered/1 predicate to get the filtered list.

• Create a condition that checks if the length L is less than 16 but more than 0, or if the memory key-value pair 'show', 'true' exists. This condition is composed of two parts joined by a semicolon ;, representing a logical OR.

Visuals

What you do visually for this capability is up to you.

Test it Out

Now Run your Conversational Agent again and try to remove some of the recipe constraints after you added them.