I am working on automating task flow of application using text based Natural Language Processing.
It is something like chatting application where the user can type in the text area. At same time python code interprets what user wants and it performs the corresponding action.
Application has commands/actions like:
Create Task
Give Name to as t1
Add time to task
Connect t1 to t2
The users can type in chat (natural language). It will be like a general English conversation, for example:
Can you create a task with name t1 and assign time to it. Also, connect t1 to t2
I could write a rule drive parser, but it would be limited to few rules only.
Which approach or algorithm can I use to solve this task?
How can I map general English to command or action?
I think the best solution would be to use an external service like API.ai or wit.ai. You can create a free account and then you can map certain texts to so-called 'intents'.
These intents define the main actions of your system. You can also define 'entities' that would capture, for instance, the name of the task. Please have a look at these tools. I'm sure they can handle your use case.
I think your issue is related to Rule-based system (Wiki).
You need to two basic components in core of project like this:
1- Role base:
list of your roles.
2- Inference engine:
infers information or takes action based on the interaction of input and the rule base.
spacy is python approach that I think it will help you. (More information).
You may want to try nltk. This is an excellent library for NLP and comes with a handy book to get you started. I think you may find chapter 8 helpful for finding sentence structure, and chapter 7 useful for figuring out what your user is requesting the bot to do. I would recommend you read the entire thing if you have more than a passing interest in NLP, as most of it is quite general and can be applied outside of NLTK.
What you are describing is a general problem with quite a few possible solutions. Your business requirements, which we do not know, are going to heavily influence the correct approach.
For example, you will need to tokenize the natural language input. Should you use a rules-based approach, or a machine learning one? Maybe both? Let's consider your input string:
Can you create a task with name t1 and assign time to it. Also, connect t1 to t2
Our system might tokenize this input in the following manner:
Can you [create a task] with [name] [t1] and [assign] [time] to it. Also, [connect] [t1] to [t2]
The brackets indicate semantic information, entirely without structure. Does the structure matter? Do you need to know that connect t1 is related to t2 in the text itself, or can we assume that it is because all inputs are going to follow this structure?
If the input will always follow this structure, and will always contain these kinds of semantics, you might be able to get away with parsing this using regular expressions and feeding prebuilt methods.
If the input is instead going to be true natural language (ie, you are building a siri or alexa competitor) then this is going to be wildly more complex, and you aren't going to get a useful answer in a SO post like this. You would instead have a few thousand SO posts ahead of you, assuming you have sufficient familiarity with both linguistics and computer science to allow you to approach the problem systematically.
Lets say text is "Please order a pizza for me" or "May I have a cab booking from uber"
Use a good library like nltk and parse these sentences. As social English is generally grammatically incorrect, you might have to train your parser with your custom broken English corpora. Next, These are the steps you have to follow to get an idea about what a user wants.
Find out the full stop's in a paragraph, keeping in mind the abbreviations, lingos like ...., ??? etc.
Next find all the verbs and noun phrases in individual sentences can be done through POS(part of speech tagging) by different libraries.
After that the real work starts, My approach would be to create a graph of verbs where similar verbs are close to each other and dissimilar verbs are very far off.
Lets say you have words like arrange, instruction , command, directive, dictate which are closer to order. So if your user writes any one of the above verbs in their text , your algorithm will identify that user really means to imply order. you can also use edges of that graph to specify the context in which the verb was used.
Now, you have to assign action to this verb "order" based on the noun phrase which were parsed in the original sentence.
This is just a high level explanation of this algorithm, it has many problems which needs serious considerations, some of them are listed below.
Finding similarity index between root_verb and the given verb in very short time.
New words who doesn't have an entry in the graph. A possible approach is to update your graph by searching google for this word, find a context from the pages on which it was mentioned and find an appropriate place for this new word in the graph.
Similarity indexes of misspelled words with proper verbs or nouns.
If you want to build a more sophisticated model, you can construct graph for every part of speech and can select appropriate words from each graph to form sentences in response to the queries. Above mentioned graph is meant for Verb Part of speech.
Although, #whrrgarbl is right. It seems like you do not want to train a bot.
So, then to handle language input variations(lexical, semantic..) you would need a pre-trained bot which you can customize(or may be just add rules according to your need).
The easiest business oriented solution is Amazon Lex. There is a free preview program too.
Another option would be to use Google's Parsey McParseface(a pre-trained English parser, there is support for 40 languages) and integrate it with a chat-framework. Here is a link to a python repo, where the author claims to have made the installation and training process convenient.
Lastly, this provides a comparison of various chatbot platforms.
Related
I have done a lot of research on how to create chat bots (the responding part) however I can't find a way to make it more advanced. For example, I keep seeing NLTK reflections but I want to know if there are more advanced methods in NLTK (or other modules) that allow me to create a learning bot, smart bot or even an AI but I am struggling in finding modules, tutorials or documentation that help with getting started and proceeding that way. Reflections don't always work well like responding in context unless you have many lines of code pre-written for content which is inefficient and may not always be accurate. Note: I don't want to be spoon fed, I just want to be pointed in the right direction of stuff that I can do and look at.
a solution would be
e.g. user asks: "who is your favourite actor?"
bot replies with: "Brad Pitt"
(only though of Brad because of the ad astra advertisements xD)
Below is the code that I am trying to stay away from.
pairs = [
[
r"my name is (.*)",
["Hello %1, How are you today ?",]
],
[
r"what is your name ?",
["My name is Chatty and I'm a chatbot ?",]
],
[
r"how are you ?",
["I'm doing good\nHow about You ?",]
],
[
r"sorry (.*)",
["Its alright","Its OK, never mind",]
],
[
r"i'm (.*) doing good",
["Nice to hear that","Alright :)",]
]```
There are two main styles of conversational agents: retrieval and generative. The regex code you show can be thought of as a very simple retrieval model. More complicated retrieval models classify user input with a classifier (at this point, almost always a neural network). Generative models treat the input to output mapping as a machine translation problem, and use techniques from NMT, neural machine translation.
Some resources:
The new verrsion of Speech and Language processing, by Dan Jurafsky and James H. Martin has three chapters on bots/question-answering
If you are trying to build something practical, then you should use a mature library. IMO, Rasa has been leading the pack for a few years at this point
The code you want to stay away from used to be the very beginning of cahtbots (Eliza: https://blog.infermedica.com/introduction-to-chatbots-in-healthcare/). A good starting point is a full dialogue system. You could use for example the trindikit for python, which is basically a dialogue manager. Furthermore, you need to implement some sort of common sense reasoning database (e.g. compare Erik T Mueller: Commonsense Reasoning - An event Calculus based approach). Normally, most chatbots are focussed on a specific domain (product questions, recommender etc.), so you need to fix exactly what intentions may provoke which speech acts, classify and model them accordingly (LSTM for calssification, Bayes for production).
Upon all this you either have to build a surface realisation system or use canned text as templates, which is growing work when your domain expands more and more.
I am working on a customer relations chatbot. The user can input either a greeting, inital_query or a query related to a product. The initial query is when the user gives their user_id to the chatbot. This is done to filter results from the database.
I created a few training examples to help the chatbot classify initial_query from the others. But the problem is the chatbot is not able to recognize a user_id as an entity if it is not specified in the training data. for example
## intent:initial_query
- My name is [Karthik](name) and my user ID is [0234](UserID)
this is one such example for initial_query. Here the userId specified is 0234. but the database contains many more users with unique userIds for each user and it is not possible for me to add all the ids into the training example.
What should I do to make the bot understand when a user id is specified? I saw somewhere that lookup tables can be used. But when I tried using lookup tables, it still did not recognize ids not part of the training examples.
This is the link I used to try lookup tables in my code.
intent_entity_featurizer_regex does not seem to work for me. I am stuck here as this is a crucial part of the bot. If lookup tables is not the best solution to this problem I am also open to other ideas.
Thank you
I'm going to get a bad wrap for always saying you Need more training data, but I would imagine thet is playing a part here as well.
I believe you have a few possible courses of action:
Provide more training data, I've never seen a good intent with fewer than 10 training examples. This number increases with every possible permutation of an intent as well as with more similar intents.
Use a pre-built entity recognizer like Duckling or spaCy. They won't necessarily know that 1234 is a userId, but they can auto extract numbers.
If you are using new_crf with Rasa then it is important to realize that it is actually learning the pattern of utterances and recognizes entities by what is around that entity rather than the actual value.
Also you could use regex with Rasa, but the regex featurizer isn't just a lookup tool. It adds a flag to the CRF whether or not the token matches that pattern. Given this it still needs sufficient training data to learn that that token is important for that entity.
I have the following search:
titles = Title.objects.filter(title__icontains=search)
If this is a search to find:
Thomas: Splish, Splash, Splosh
I can type in something like "Thomas" or "Thomas: Splish, Splash, Splosh" and it will work.
However, if I type in something like "Thomas Splash", it will not work. How would I improve the search to do something like that (also note that if we split on words, the comma and other non-alphanumerics should be ignored -- for example, the split words should not be "Thomas:", "Splish," but rather "Thomas", "Splish", etc.
This kind of search is starting to push the boundaries of django and the ORM. Once it gets to this level of complexity I always switch over to a system that is built entirely for search. I dig lucene, so I usually go for ElasticSearch or Solr
Keep in mind that full text searching is a subsystem all unto itself, but can really add a lot of value to your site.
As Django models are using database queries there is not much magic you can do.
You could split your search by non-alphanumeric chars and search objects containing all words but this will not be smart and efficient.
If you want something really smart maybe you should check out haystack:
http://haystacksearch.org/
Vocabulary that I am using:
nounphrase -- A short phrase that refers to a specific person, place, or idea. Examples of different nounphrases include "Barack Obama", "Obama", "Water Bottle", "Yellowstone National Park", "Google Chrome web browser", etc.
category -- The semantic concept defining which nounphrases belong to it and which ones do not. Examples of categories include, "Politician", "Household items", "Food", "People", "Sports teams", etc. So, we would have that "Barack Obama" belongs to "Politician" and "People" but does not belong to "Food" or "Sports teams".
I have a very lage unlabeled NLP dataset consisting of millions of nounphrases. I would like to use Freebase to label these nounphrases. I have a mapping of Freebase types to my own categories. What I need to do is download every single examples for every single Freebase type that I have.
The problem that I face is that need to figure out how to structure this type of query. At a high level, the query should ask Freebase "what are all of the examples of topic XX?" and Freebase should respond with "here's a list of all examples of topic XX." I would be very grateful if someone could give me the syntax of this query. If it can be done in Python, that would be awesome :)
The basic form of the query (for a person, for example) is
[{
"type":"/people/person",
"name":None,
"/common/topic/alias":[],
"limit":100
}]
There's documentation available at http://wiki.freebase.com/wiki/MQL_Manual
Using freebase.mqlreaditer() from the Python library http://code.google.com/p/freebase-python/ is the easiest way to cycle through all of these. In this case, the "limit" clause determines the chunk size used for querying, but you'll get each result individually at the API level.
BTW, how do you plan to disambiguate Jack Kennedy the president, from the hurler, from the football player, from the book, etc, etc http://www.freebase.com/search?limit=30&start=0&query=jack+kennedy You may want to consider capturing additional information from Freebase (birth & death dates, book authors, other types assigned, etc) if you'll have enough context to be able to use it to disambiguate.
Past a certain point, it may be easier and/or more efficient to work from the bulk data dumps rather than the API http://wiki.freebase.com/wiki/Data_dumps
Edit - here's a working Python program which assumes you've got a list of type IDs in a file called 'types.txt':
import freebase
f = file('types.txt')
for t in f:
t=t.strip()
q = [{'type':t,
'mid':None,
'name':None,
'/common/topic/alias':[],
'limit':500,
}]
for r in freebase.mqlreaditer(q):
print '\t'.join([t,r['mid'],r['name']]+r['/common/topic/alias'])
f.close()
If you make the query much more complex, you'll probably want to lower the limit to keep from running into timeouts, but for a simple query like this, boosting the limit above the default of 100 will make it more efficient by querying in bigger chunks.
The general problem described here is called Entity Linking in natural language processing.
Unabashed self plug:
See our book chapter on the topic for an introduction and an approach to perform large scale entity linking.
http://cs.jhu.edu/~delip/entity_linking.pdf
#deliprao
I'm going back over an old project where I added preprocessor functionality to Essence' and I realised that my previous solution of writing a domain specific language and associated lexer/parser was overkill.
Instead I just need to be able to embed dynamic language code into the file, isolate it at runtime, eval and insert the results. In other words very similar to the PHP model of inserting dynamic code into HTML. I'd rather not use PHP as Python is much easier to distribute as part of a larger project (IronPython or Jython)
So the question goes, how best to implement something like the following:
<code>Python goes here</code>
Lots of essence <code>Python</code> prime code goes here
I don't want to have to alter the structure of the Essence' file (if I remove all the code blocks everything left should be able to be syntactically correct. It needs to be able to insert text in place of a code block like PHP.
Finally security wise I'm not bothered about code injection, as it would be the user themselves choosing the file to execute although if there were security benefits to one model over another with no extra costs that would obviously be good.
Cheers in advance
Your best bet is to use one of the already made (and battle tested) Templating Engines. The two big ones that I've used are Mako, and Cheetah. They allow you to embed code right in the page, and are mostly used as the View in an MVC architecture.
If you feel that using one of those engines is overkill for your project, here is a small tutorial on how to implement basic templates yourself. Keep in mind that the example will need to be modified to suit your particular project/needs.