This is a small chatbot created in order to try out AIML markup language for dialog modeling. This chatbot attempts to imitate a cocktail adviser, it tries to make an intelligible advice based on the information the user provides.

You don't need to install it, just make sure you have python3. Before running chatbotty also check if you have the following python3 modules installed.

• nltk (Wordnet, SnowballStemmer)
• numpy

Once you installed the above start with ./chatbotty.py, or if your default python is python2 python3 chatbotty.py

As mentioned above, the chatbot attempts to imitate a cocktail adviser and a casual chatbot. The chat topics are of course limited to just one, cocktails. In order to get an advice about which cocktail to drink the user has to ask the chatbot about it by basically saying "I want ..." or whatever suits the conversation context. After that the chatbot analyzes user query and returns the recipe for the cocktail that best matches the query.

There are two similarity strategies used in order to analyze the query TFIDF and WORDNET. By default the chatbot uses TFIDF but you can change it during conversation by issuing the following commands: Hanna use wordnet, Hanna use tfidf. You can check which strategy is used currently by asking Hanna what do you use? (available after greeting). The chatbot uses a small cocktails database cocktails.xml to run its similarity comparisons. The database contains cocktail descriptions, history and some trivia facts. Once the user asks/replies to system with "I want ..." or any other phrase suitable in context the system takes the query and runs similarity analysis. For example:

$ ./chatbotty.py 
Loading cocktail_brains.aiml... done (0.29 seconds)
Hi there! I am Hanna and I shall be your favourite cocktail adviser :)
What's on your mind?
> hi
Hola!
> How are you?
Doing fine.
> I want a cocktail
Cocktails! I love them! What will you drink?
> I want something blue
I found the one you might like! "Mai Tai"
 The Mai Tai is an alcoholic cocktail based on rum, Curacao liqueur, and lime juice, 
 associated with Polynesian-style settings. (...)
> awesome thanks
You are welcome.

Obviously you don't have to straight away ask the chatbot for a cocktail advice but the final result of the conversation is the recipe.

The chatbot uses cocktails.xml file as a "database" where we keep cocktail descriptions, facts and mixing recipes. Every database cocktail entry must have the following fields in order to work as intended:

<cocktail name="MY COCKTAIL">
 <description>...</description>
 <history>...</history>
 <trivia>...</trivia>
 <comments>...</comments>
 <ingredients>...</ingredients>
 <mixing>...</mixing>
</cocktail>

If you want to check how many cocktails are there you can ask the chatbot how many cocktails do you know? or something of this sort. If you want to see the list of cocktails ask what cocktails do you know?. Moreover you can ask about specific cocktail what do you know about Appletini? or do you know the history of Appletini?. You can even ask tell me something funny about Appletini. The chatbot will query the database and return the requested field.

I shall omit TFIDF description since you can read about it on Wikipedia. WORDNET however is a custom similarity search. Before running wordnet comparisons the user query gets expanded with wordnet adjective definitions. Every adjective in user query is expanded with most semantically close adjectives according to wordnet, non-contentful (there is a filter) words are removed. For example: I want something blue is converted into clear intermediate green similar blue sky. The database is called with all the available fields. We iterate through every word in the expanded query and match it against the cocktail database. The number of matched words is divided by the number of words in the field and multiplied by the multiplier. Every cocktail entry has several different fields in the database <description>, <history>, <ingredients>, .... If the match is found in "description" it is multiplied by 2 since the words from "description" field have more weight than the ones from "history", every field has a multiplier. When we calculate the scores for every cocktail in the database we return the one with the highest value. WORDNET strategy is not as accurate as TFIDF but it certainly suffers less because of a small query or a small database (tbh I think it just needs some tuning).