vliu15 / elmo-kmeans

GPU-accelerated Topic Analysis pipeline

ELMo Embeddings for Clustering

To convert text datasets into clusters based on topic. Currently in progress for benchmarking NVIDIA Rapids cuML/pyGDF for performance.

Requirements:

• Python3 (>=3.6 for AllenNLP)
• AllenNLP
• TensorFlow
• NumPy
• SKLearn (for clustering)
• Torch (for AllenNLP)
• SciPy (for SKLearn stop words)

Install

To install the necessary dependencies:

apt-get install python3-pip
pip3 install allennlp
pip3 install tensorflow-gpu
pip3 install numpy
pip3 install sklearn
pip3 install torch
# pip3 install scipy

Usage

To generate sentence embeddings, make sure that the sentences.txt file is formatted as such:

<sentence/transcription>
<sentence/transcription>
# and so on...

Run python3 main.py with the following options:

• --mode embed to embed the sentence file
• --mode sif to enhance sentence embeddings with SIF
• --mode cluster to cluster embeddings
• --mode project to reduce dimensionality for visualization
• --mode metadata to write metadata file
• --mode tensorboard to create TensorBoard files

Adjust runtime flags in main.py

A couple auxiliary files:

• Run sh clean.sh to convert transcriptions to lower case and remove stop words

Run

To run inside a Docker container:

docker build -t elmo-embeddings .
docker exec -it elmo-embeddings /bin/bash
python3.6 main.py

Outputs

An output folder will be created in the current directory containing:

• embeddings.npy: a NumPy array of sentence embeddings (NumPy arrays) in binary format
• embeddings_sif.npy: a NumPy array of sentence embeddings after SIF
• embeddings_pc.npy: a NumPy array of sentence embeddings after PCA
• embeddings_ts.npy: a NumPy array of sentence embeddings after t-SNE
• km_labels.json: a list of cluster labels generated by KMeans
• metadata.tsv: metadata of sentence labels for visualization

Other nested output folders:

• tensorboard: for TensorBoard output logs
• kmeans: for clustered sentences with kmeans
• trimmed: for another copy of embeddings/sentences with specified clusters removed
• hierarchy: for clustered sentences with hierarchical kmeans

GPU-acceleration:

• Allow ELMo to use GPU for embedding (UPDATE: GPU speedup by as much as 4x)
• Utilize NVIDIA Rapids cuML to GPU-accelerate clustering (by ~10x)

TO-DO

• Preprocess transcriptions
• Embed each sentence with ELMo
• Enhance embeddings with SIF
• Cluster using SKLearn KMeans (optional: hierarchically)
• Find optimal k using elbow method and silhouette scores (optional)
• Reduce dimensionality for visualization (PCA, t-SNE)
• Run in TensorBoard
• Conclusions