Repository files navigation

This is a topic clustering library built with transformer embeddings and analysing cosine similarity between them. The topics are clustered either by kmeans or agglomeratively depending on the use case, and the embeddings are attained after propagating through any of the Transformers present inHuggingFace.The library can be foundhere.

Pytorch

Transformers

Installation is carried out using the pip command as follows:

pipinstallClusterTransformer==0.1

For using inside the Jupyter Notebook or Python IDE:

importClusterTransformer.ClusterTransformerasct

The 'ClusterTransformer_test.py' file contains an example of using the Library in this context.

The steps to operate this library is as follows:

Initialise the class: ClusterTransformer()Provide the input list of sentences: In this case, the quora similar questions dataframe has been taken for experimental purposes.Declare hyperparameters:

• batch_size: Batch size for running model inference
• max_seq_length: Maximum sequence length for transformer to enable truncation
• convert_to_numpy: If enabled will return the embeddings in numpy ,else will keep in torch.Tensor
• normalize_embeddings:If set to True will enable normalization of embeddings.
• neighborhood_min_size:This is used for neighborhood_detection method and determines the minimum number of entries in each cluster
• cutoff_threshold:This is used for neighborhood_detection method and determines the cutoff cosine similarity score to cluster the embeddings.
• kmeans_max_iter: Hyperparameter for kmeans_detection method signifying nnumber of iterations for convergence.
• kmeans_random_state:Hyperparameter for kmeans_detection method signifying random initial state.
• kmeans_no_cluster:Hyperparameter for kmeans_detection method signifying number of cluster.
• model_name:Transformer model name ,any transformer from Huggingface pretrained library

Call the methods:

• ClusterTransfomer.model_inference: For creating the embeddings by running inference through any Transformer library (BERT,Albert,Roberta,Distilbert etc.)Returns a torch.Tensor containing the embeddings.
• ClusterTransformer.neighborhood_detection: For agglomerative clustering from the embeddings created from the model_inference method.Returns a dictionary.
• ClusterTransformer.kmeans_detection:For Kmeans clustering from the embeddings created from the model_inference method.Returns a dictionary.
• ClusterTransformer.convert_to_df: Converts the dictionary from the neighborhood_detection/kmeans_detection methods in a dataframe
• ClusterTransformer.plot_cluster:Used for simple plotting of the clusters for each text topic.

The code steps provided in the tab below, represent all the steps required to be done for creating the clusters. The 'compute_topics' method has the following steps:

• Instantiate the object of the ClusterTransformer
• Specify the transformer name from pretrained transformers
• Specify the hyperparameters
• Get the embeddings from 'model_inference' method
• For agglomerative neighborhood detection use 'neighborhood_detection' method
• For kmeans detection, use the 'kmeans_detection' method
• For converting the dictionary to a dataframe use the 'convert_to_df' method
• For optional plotting of the clusters w.r.t corpus samples, use the 'plot_cluster' method

%%timeimportClusterTransformer.ClusterTransformerascluster_transformerdefcompute_topics(transformer_name):#Instantiate the objectct=cluster_transformer.ClusterTransformer()#Transformer model for inferencemodel_name=transformer_name#Hyperparameters#Hyperparameters for model inferencebatch_size=500max_seq_length=64convert_to_numpy=Falsenormalize_embeddings=False#Hyperparameters for Agglomerative clusteringneighborhood_min_size=3cutoff_threshold=0.95#Hyperparameters for K means clusteringkmeans_max_iter=100kmeans_random_state=42kmeans_no_clusters=8#Sub input data listsub_merged_sent=merged_set[:200]#Transformer (Longformer) embeddingsembeddings=ct.model_inference(sub_merged_sent,batch_size,model_name,max_seq_length,normalize_embeddings,convert_to_numpy)#Hierarchical agglomerative detectionoutput_dict=ct.neighborhood_detection(sub_merged_sent,embeddings,cutoff_threshold,neighborhood_min_size)#Kmeans detectionoutput_kmeans_dict=ct.kmeans_detection(sub_merged_sent,embeddings,kmeans_no_clusters,kmeans_max_iter,kmeans_random_state)#Agglomerative clusteringneighborhood_detection_df=ct.convert_to_df(output_dict)#KMeans clusteringkmeans_df=ct.convert_to_df(output_kmeans_dict)returnneighborhood_detection_df,kmeans_df

Calling the driver code:

%%timeimportmatplotlib.pyplotaspltn_df,k_df=compute_topics('bert-large-uncased')kg_df=k_df.groupby('Cluster').agg({'Text':'count'}).reset_index()ng_df=n_df.groupby('Cluster').agg({'Text':'count'}).reset_index()#Plottingfig,(ax1,ax2)=plt.subplots(1,2,figsize=(15,5))rng=np.random.RandomState(0)s=1000*rng.rand(len(kg_df['Text']))s1=1000*rng.rand(len(ng_df['Text']))ax1.scatter(kg_df['Cluster'],kg_df['Text'],s=s,c=kg_df['Cluster'],alpha=0.3)ax1.set_title('Kmeans clustering')ax1.set_xlabel('No of clusters')ax1.set_ylabel('No of topics')ax2.scatter(ng_df['Cluster'],ng_df['Text'],s=s1,c=ng_df['Cluster'],alpha=0.3)ax2.set_title('Agglomerative clustering')ax2.set_xlabel('No of clusters')ax2.set_ylabel('No of topics')plt.show()

Colab-Demo

Colab-Demo

Kaggle Notebook

Quantum Stat Repository

Cluster Images ( Created With Facebook BART)

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.

MIT