I'm Melanie. I work to design and implement scalable applications for natural language processing (NLP) and machine learning (ML). I'm currently at WayBlazer, where I focus on generating personalized travel recommendations from natural language queries.
Turns out I actually did get a lot of stuff done, who would've thought. Lists are good; I recommend you make one yourself.
Also everyone has been asking about which tech blogs to keep an eye on recently, so I finally decided to make, you might have guessed it, another list! Here's the shortlist.
See below for more details on previous projects.
Edit distances and sequence alignment
During my senior year, I finally took a class on advanced C++. Surprisingly enough, it didn't seem nearly as hard as the first one I had to struggle through a few years ago, and I ended up having a lot of fun with it. As final project, I decided to work on edit distances, and implemented the Wagner–Fischer algorithm as an instance of dynamic programming. Later on, I expanded the project to also cover the Needleman-Wunsch algorithm for global sequence alignment.
Semantic Role Labeling using linear-chain CRF
My very last undergrad project for a class on advanced language modeling, where we discussed the theoretical foundations of hidden Markov models, the Viterbi and EM algorithm, log-linear models, maximum entropy models (MEMMs), and as well as conditional random fields (CRFs).
String to semantic graph alignment
For my undergrad thesis, I started working on semantic parsing: the problem of mapping natural language strings to meaning representations. In order to train a semantic parser for English into Abstract Meaning Representation, we first need to know which phrases in the input sentence invoked which concepts in the corresponding AMR graph. The project aimed at building an English/AMR aligner to solve this task automatically.
Origin Thesis Code
Semantic dependency graph parsing
For a class on semantic dependency graph parsing, I wrote a short script that computes statistics for semantic dependency graphs and generates plots for the distribution of words per indegree and outdegree. As final project, I submitted a comprehensive review on Abstract Meaning Representation (AMR), a set of English sentences paired with simple, readable semantic representations.
Research internship at Textkernel
In 2014, I was a research intern at Textkernel, where we explored new methods of improving resume parsing for multi-lingual documents. In order to extract structured information in the form of specific phrases like name or address, we adopted the probabilistic conditional random fields (CRF) framework. In addition, we experimented with a novel approach that integrates continuous vector representations of words as input features for such a model.
Report Paper Interview
Word meaning in context
For a really great class on distributional semantics, I presented a paper on ‘Measuring Distributional Similarity in Context’ (Dinu and Lapata, 2003). In a nutshell, they attempt to model the intuition that word meaning is represented as a probability distribution over a set of latent senses, and thus modulated by context. They employ two different models: the first based on non-negative matrix factorization (NMF), and the second implementing Latent Dirichlet Allocation (LDA).
I studied abroad and learned some linguistics:
Consider an example where a zombie has died and been reanimated, and John drowns him.
Presentation slides may or may not help to understand what is going on.
I took some classes on psycholinguistics, where I presented a range of interesting papers, including ‘Expectation-based syntactic comprehension’ (Levy, 2008), and ‘Dependency Locality Theory’ (DLT) (Gibson, 2000). Check out the slides below!