insigen

InsiGEN

A state of the art NLP model to generate insights from a text piece

Features of topic modelling

• Generating a distribution of generalized topics covered in a document/articler
• Extracting contextualized keywords from the text piece
• Generating a summary of the text
• Trained on a corpus of 6000 wikipedia articles for generalized topics
• Can be trained on custom data for more specific topics

How to use the model

• Clone this repository
• Install the dependencies from the requirements.txt
• Basic Usage:

Get a topic distribution

from insigen import insigen
model = insigen()
topic_distribution = model.get_distribution(document)

Important parameters for insigen:

• use_pretrained_embeds: Setting this parameter to False will allow you to train your own embeddings. Further parameters need to be specified for training

• embed_file: This parameter should be used when you've trained your own embeddings. Specify the path to your sentence embeddings.

• dataset_file: This parameter should be used when you've trained your own embeddings. Specify the path to your own dataset.

• embedding_model: (Default = all-mpnet-base-v2) Insigen uses sentence bert models to train it's embeddings. Valid models are:

           all-distilroberta-v1
           all-mpnet-base-v2
           all-MiniLM-L12-v2
           all-MiniLM-L6-v2
  

Important parameters for get_distribution

• document: The text for which the topic distribution is to be generated
• metric: This metric defines how the topics will be found. Can be set to 'threshold', to get all the topics above a similarity threshold. Defaults to 'max'. 'Max' metric gets the top "n" topics
• max_count: This argument should be used with max metric. It specifies the top x amount of topics that get fetched. Defaults to 1.
• threshold: This argument should be used with threshold metric. It specifies the threshold similarity over which all topics will be fetched. Defaults to 0.5.

Get keyword frequency

frequency = model.get_keyword_frequency(document, min_len=2, max_len=3)

#Generate a wordcloud using the frequency
cloud = model.generate_wordcloud(frequency)

Important parameters for get_keyword_frequency

• document: The text for which the keyword frequency is to be generated
• frequency_threshold: minimum frequency of a n-gram to be considered in the keywords (min_len and max_len are also used to adjust the length of n-grams in the text)

Generate Summary

summary = model.generate_summary(article, topic_match=relevant_topic))

# To get a list of topics, use this
#print(model.unique_topics)

Import parameters for generate_summary

• document:The text for which the summary is to be generated
• topic_match: a topic that can match with the text. This adds additional weight to sentences that are more related to the topic. use model.unique_topics to get a list of topics that can match. Defaults to None, in which case weightage to related sentence will not be given.
• topic_weight: Adds weightage to the topic similarity score. Increasing this parameter results to more topic oriented summary. Defaults to 1.
• similarity_weight: Adds weightage to sentence similarity score. Increasing this parameter results in extracting more co-related sentences. Defaults to 1.
• position_weight: Adds weightage to the position of the sentences. Increasing this parameter results to more position oriented summary; i.e Texts present early in the document are given more weightatge. Defaults to 10.
• num_sentences: This specifies the number of sentences that are to be included in the summary. Defaults to 10.

Train on your dataset

embeddings = model.train_embeds(dataset)

Important parameters for train_embeds

• dataset: A pandas dataframe for the dataset to be trained
• batch_size: Batches to divide the dataset into. Defaults to 32.

How does the model work?

Topic Distribution

• Create embedded vectors of labelled training articles

• Find mean embeddings of each topic in the corpus to create topic vectors and create clusters of articles
  

• Use KNN to place new articles in the topic vector cluster
  

• Chunking each article and finding relevant topic from the topic vectors
  

Keyword extraction

• N-grams and keywords are filtered from the text
• Contextually similar keywords to the article are given higher scoring
• A threshold is applied to the filtered list of keywords to get the final list of keywords

Summary Extraction

• The PageRank algorithm is used to create a similarity matrix for sentences in the text
• Additionally, sentences are scored based on their position in the text and their similarity to a relevant topic
• Top N sentences from the similarity matrix are extracted to create a summary.