Skip to content

Latest commit

 

History

History
175 lines (122 loc) · 10.9 KB

proposal.md

File metadata and controls

175 lines (122 loc) · 10.9 KB
Raw

Making an Impact with Python Natural Language Processing Tools

Description

Do your tweets get lost in the shuffle? Would you like to predict a tweet's impact before you hit send? Python now has all the tools to make this possible. Several Python packages for machine learning and natural language processing have reached "critical mass" and can now be combined to perform these and other powerful natural language processing tasks. This tutorial will teach you how.

Audience

Amateur and professional data scientists who want to learn about a powerful combination of python tools and techniques for natural language processing

Objectives

Attendees will build a python module that can determine the best time of day to tweet on a particular subject. While building this tool, attendees will become familiar with the most powerful combination of python packages for performing state-of-the-art natural language processing.

Detailed Abstract

Prerequisites

Students who have experience writing python scripts or modules and are familiar with the string manipulation and formatting capabilities built into python will have the necessary skill to complete this tutorial.

In addition, any students who are familiar with linear algebra, and basic statistics concepts (like probability and variance) will be able to grasp the mathematics behind the tools assembled during the tutorial, but this is not required. Likewise, familiarity with scikit-learn and pandas would be helpful, but not necessary.

Also, students who are familiar with git and GitHub will be able to follow along with the logistics of the workshop sessions more quickly and spend more time developing their NLP pipeline.

Python Development Environment

Students will need iPython, NLTK, scipy, scikit-learn, and Pandas installed on their laptops to run the examples in this tutorial and build the tweet impact predictor tool. Students can install these requirements with:

pip install -r https://github.com/totalgood/pycon-2016-nlp-tutorial/master/requirements.txt

In addition, students have the option of installing a python twitter API client rather than utilizing the preprocessed collection of twitter feeds provided with the course material.

Overview

Participants will develop a natural language processing pipeline for tweets in three modules.

The first section of the pipeline will be a natural language feature extractor and normalizer based on python builtin modules collections, string, and re. The Pandas DataFrame data structure will also be introduced.

The second section will utilize scikit-learn and numpy to simplify the feature set to a manageable number of features. It will find optimal combinations of reduced numbers of features that provide the greatest information about the subject matter of the tweets being processed.

The third section of the pipeline will assemble a training set based on tweet statistics not contained in the natural language content of the tweets. These statistics will be combine with the natural language features to classify tweets according to their popularity (number of favorites), and reach (number of potential viewers due to retweets). A neural net will be trained to predict tweet impact (popularity and reach) based on the time of day and day of week as well as the tweet text NLP features.

Finally an Advanced section will provide attendees with the tools and resources necessary to further develop their tweet prediction pipeline.

Outline

Introduction (15 min)

  • Logistics (restrooms, WiFi, classroom etiquette)
  • Agenda & Schedule (4 sessions, 4 workshops)
  • Interesting NLP applications
    • Behavior modification with MMORPG Troll-police
    • Sports and Financial news natural language generation
  • State of the art NLP
    • gensim word vector math
      • teaser: "king" - "man" + "woman" = "queen"

Feature Extraction with Python (30 min)

  • str.split to quickly extract words from a tweet
  • collections.Counter to count word occurrences
  • Explore regular expressions in a text adventure
    • Text Adventure games vs. Choose Your Own Adventure books
    • Python Regular Expressions vs. Memoryless Regular Expressions
    • re.split to more accurately extract words (tokens)
    • nltk stemmers
    • nltk part-of-speech tagging
    • nltk word root parsers
    • nltk stop word filters
  • pandas.Series and pandas.DataFrame
    • analogy to builtin collections.OrderedDict
    • use for storing word vectors
  • np.linalg.norm and np.dot to efficiently normalize word counts and frequencies
  • sklearn...TfidfVectorizer to efficiently store (sparse) normalized word frequencies
  • np.linalg.norm, np.dot to compute "distances" between tweets
  • sklearn.cluster to group similar tweets

Workshop: Feature Extraction Pipeline (20 min)

Students will use the tools provided in the presentation to build a python function capable of processing 10's of thousands of tweets in a few minutes to produce meaningful clusters based on tweet content.

Feature Reduction (30 min)

  • Feature Reduction
    • Calculating entropy (information value) with numpy
    • sklearn..PCA Principle Component Analysis
      • how it works (overview of matrix algebra)
      • where it works best
      • what to watch out for
      • apply to tweet TFIDF to reduce vocabulary
  • Plotting and Exploring
    • scipy scatter matrix plots
      • visulizing natural language feature vectors
      • projecting/slicing
    • json.dumps of TFIDF matrices for d3.js matrix visualizations
    • using python to manipulate nested dicts to create json required for interactive d3.js force-directed graphs

Workshop: Feature Reduction (15 min)

Attendees will use the tools provided simplify the natural language feature set extracted from their twitter feeds. They will use scikit-learn to identify more informative clusters and patterns than was possible in the previous workshop.

Supervised Learning (30 min)

  • Extracting numerical statistics about tweets
    • pandas.DataFrame .group_by and .hist
      • time-of-day, day-of-week, day-of-quarter, day-of-year, month-of year
    • Following the trail of retweets
      • Model after Python builtin os.walk
    • Favorites/Likes
      • numpy.corrcoeff and numpy.cov to correlate
        • numerical metrics
        • tweet subject
        • twitter ID
      • Identify influential "likers"
  • Modeling
    • Use builtin random.sample to compose test and training data sets
    • np.linalg.norm and np.dot to calculate tweet similarity
      • sklearn.cluster.KMeans to classify tweets by topic
      • numpy.linalg.norm similarity to cluster means to score tweets
      • label tweets using pandas.DataFrame.get ([]) sql-like queries to threshold score
    • sklearn..Lasso for efficient linear regression (p-norm, cosine-distance, supremum distance)
  • Measuring Model Performance
    • ConfusionMatrix
      • sensitivity
      • specificity
  • sklearn.lda Linear Discriminant Analysis (use topic labels above)
    • show why model performance is improved relative to PCA alone

Workshop: Supervised Learning (20 min)

Attendees will mine the Twitter API and data sets provided to compute the numerical statistics and assign scores to each tweet. Attendees will build a ConfusionMatrix class that inherits pandas.DataFrame and adds a from_labels method to injest scored/labeled data. Attendees will also add accuracy and specificity methods to their class and combine them to create a custom performance metric that targets their individual performance goals for their tweet predictor. They will balance the likelihood of a great tweet and the likelihood of a dud tweet. Attendees will use sklearn.lda to reduce dimensions further and generalize the model. Finally attendees will compare their model performance metric with and without the LDA pipeline element to confirm improved performance.

Advanced NLP (30 min)

Attendees will be introduced to recent advances in NLP and resources to help them explore further.

  • Adding another dimension: word order
    • Scale Space Processing as alternative to the orderless "Bag of Words" approach
    • pandas.DataFrame.rolling_window to perform 1-D convolution
    • matplotlib.pyplot.pcolor + pandas.DataFrame = heatmap of twitter streams
  • Neural Networks
    • pybrain2 convolutional neural network to classify tweets
  • Word Vectors
    • Explanation of skip-grams
    • Utilizing Google's well-trained Word2Vec model
    • Example Word Vector "math"

Workshop: Level Up Your Tweet Predictor (20 min)

Attendees will add scale-space processing to their tweet predictor and plot a topic heatmap of their twitter stream. Students will be provided iPython notebooks to help them incorporate the other 2 advanced features into their pipeline on their own.

More Info

All material will be accompanied by iPython notebooks and provided in open-source (MIT-licensed) GitHub repositories. Data sets will be prepossessed and compressed to simplify participant environment setup. A sequence of git tags and branches will provide an "answer key" for workshop activities, to allow students to continue moving forward.

Notes

Hobson is a passionate advocate for Python and open source. Hobson has spoken about natural language processing on numerous occasions and has a track record of successfully teaching novices to use python for natural language processing. Hobson served for years as a mentor for Georgia Institute of Technology grad students in Machine Learning and is currently mentoring SlideRule students. Hobson's talks are very interactive and engage participants individually throughout a tutorial or presentation by soliciting their ideas and provoking their critical thinking.

Rob Ludwick was introduced to python by a friend and was never the same again. He's worked for large and small companies in the past, and most recently at Talentpair, where he enjoys the process of matching employees to jobs.

Speaker Bio

Hobson has nearly two decades of engineering and teaching experience in robotics, autonomous systems, data science, and natural language processing. Hobson has relied on Python as his language of choice for Data Science and Natural Language Processing for companies like Squishy Media, Building Energy, Sharp Labs, Total Good, Hack Oregon, Pellego, Intel Labs, and Talentpair, as well as numerous open source projects.