Hierarchical unsupervised and semi-supervised topic models for sparse count data with CorEx

Overview

Anchored CorEx: Hierarchical Topic Modeling with Minimal Domain Knowledge

Correlation Explanation (CorEx) is a topic model that yields rich topics that are maximally informative about a set of documents. The advantage of using CorEx versus other topic models is that it can be easily run as an unsupervised, semi-supervised, or hierarchical topic model depending on a user's needs. For semi-supervision, CorEx allows a user to integrate their domain knowledge via "anchor words." This integration is flexible and allows the user to guide the topic model in the direction of those words. This allows for creative strategies that promote topic representation, separability, and aspects. More generally, this implementation of CorEx is good for clustering any sparse binary data.

If you use this code, please cite the following:

Gallagher, R. J., Reing, K., Kale, D., and Ver Steeg, G. "Anchored Correlation Explanation: Topic Modeling with Minimal Domain Knowledge." Transactions of the Association for Computational Linguistics (TACL), 2017.

Getting Started

Install

Python code for the CorEx topic model can be installed via pip:

pip install corextopic

Example Notebook

Full details on how to retrieve and interpret output from the CorEx topic model are given in the example notebook. Below we describe how to get CorEx running as an unsupervised, semi-supervised, or hierarchical topic model.

Running the CorEx Topic Model

Given a doc-word matrix, the CorEx topic model is easy to run. The code follows the scikit-learn fit/transform conventions.

import numpy as np
import scipy.sparse as ss
from corextopic import corextopic as ct

# Define a matrix where rows are samples (docs) and columns are features (words)
X = np.array([[0,0,0,1,1],
              [1,1,1,0,0],
              [1,1,1,1,1]], dtype=int)
# Sparse matrices are also supported
X = ss.csr_matrix(X)
# Word labels for each column can be provided to the model
words = ['dog', 'cat', 'fish', 'apple', 'orange']
# Document labels for each row can be provided
docs = ['fruit doc', 'animal doc', 'mixed doc']

# Train the CorEx topic model
topic_model = ct.Corex(n_hidden=2)  # Define the number of latent (hidden) topics to use.
topic_model.fit(X, words=words, docs=docs)

Once the model is trained, we can get topics using the get_topics() function.

topics = topic_model.get_topics()
for topic_n,topic in enumerate(topics):
    # w: word, mi: mutual information, s: sign
    topic = [(w,mi,s) if s > 0 else ('~'+w,mi,s) for w,mi,s in topic]
    # Unpack the info about the topic
    words,mis,signs = zip(*topic)    
    # Print topic
    topic_str = str(topic_n+1)+': '+', '.join(words)
    print(topic_str)

Similarly, the most probable documents for each topic can be accessed through the get_top_docs() function.

top_docs = topic_model.get_top_docs()
for topic_n, topic_docs in enumerate(top_docs):
    docs,probs = zip(*topic_docs)
    topic_str = str(topic_n+1)+': '+', '.join(docs)
    print(topic_str)

Summary files and visualizations can be outputted from vis_topic.py.

from corextopic import vis_topic as vt
vt.vis_rep(topic_model, column_label=words, prefix='topic-model-example')

Choosing the Number of Topics

Each topic explains a certain portion of the total correlation (TC). We can access the topic TCs through the tcs attribute, as well as the overall TC (the sum of the topic TCs) through the tc attribute. To determine how many topics we should use, we can look at the distribution of tcs. If adding additional topics contributes little to the overall TC, then the topics already explain a large portion of the information in the documents. If this is the case, then we likely do not need more topics in our topic model. So, as a general rule of thumb continue adding topics until the overall TC plateaus.

We can also restart the CorEx topic model from several different initializations. This allows CorEx to explore different parts of the topic space and potentially find more informative topics. If we want to follow a strictly quantitative approach to choosing which of the multiple topic model runs we should use, then we can choose the topic model that has the highest TC (the one that explains the most information about the documents)

Semi-Supervised Topic Modeling

Using Anchor Words

Anchored CorEx allows a user integrate their domain knowledge through "anchor words." Anchoring encourages (but does not force) CorEx to search for topics that are related to the anchor words. This helps us find topics of interest, enforce separability of topics, and find aspects around topics.

If words is initialized, then it is easy to use anchor words:

topic_model.fit(X, words=words, anchors=[['dog','cat'], 'apple'], anchor_strength=2)

This anchors "dog" and "cat" to the first topic, and "apple" to the second topic. The anchor_strength is the relative amount of weight given to an anchor word relative to all the other words. For example, if anchor_strength=2, then CorEx will place twice as much weight on the anchor word when searching for relevant topics. The anchor_strength should always be set above 1. The choice of anchor_strength beyond that depends on the size of the vocabulary and the task at hand. We encourage users to experiment with anchor_strength to find what is useful for their own purposes.

If words is not initialized, we can anchor by specifying the column indices of the document-term matrix that we wish to anchor on. For example,

topic_model.fit(X, anchors=[[0, 2], 1], anchor_strength=2)

anchors the words of columns 0 and 2 to the first topic, and word 1 to the second topic.

Anchoring Strategies

There are a number of strategies we can use with anchored CorEx. Below we provide just a handful of examples.

  1. Anchoring a single set of words to a single topic. This can help promote a topic that did not naturally emerge when running an unsupervised instance of the CorEx topic model. For example, we might anchor words like "snow," "cold," and "avalanche" to a topic if we supsect there should be a snow avalanche topic within a set of disaster relief articles.
topic_model.fit(X, words=words, anchors=[['snow', 'cold', 'avalanche']], anchor_strength=4)
  1. Anchoring single sets of words to multiple topics. This can help find different aspects of a topic that may be discussed in several different contexts. For example, we might anchor "protest" to three topics and "riot" to three other topics to understand different framings that arise from tweets about political protests.
topic_model.fit(X, words=words, anchors=['protest', 'protest', 'protest', 'riot', 'riot', 'riot'], anchor_strength=2)
  1. Anchoring different sets of words to multiple topics. This can help enforce topic separability if there appear to be "chimera" topics that are not well-separated. For example, we might anchor "mountain," "Bernese," and "dog" to one topic and "mountain," "rocky," and "colorado" to another topic to help separate topics that merge discussion of Bernese Mountain Dogs and the Rocky Mountains.
topic_model.fit(X, words=words, anchors=[['bernese', 'mountain', 'dog'], ['mountain', 'rocky', 'colorado']], anchor_strength=2)

The example notebook details other examples of using anchored CorEx. We encourage domain experts to experiment with other anchoring strategies that suit their needs.

Note: when running unsupervised CorEx, the topics are returned and sorted according to how much total correlation they each explain. When running anchored CorEx, the topics are not sorted by total correlation, and the first n topics will correspond to the n anchored topics in the order given by the model input.

Hierarchical Topic Modeling

Building a Hierarchical Topic Model

For the CorEx topic model, topics are latent factors that can be expressed or not in each document. We can use the matrices of these topic expressions as input for another layer of the CorEx topic model, yielding a hierarchical topic model.

# Train the first layer
topic_model = ct.Corex(n_hidden=100)
topic_model.fit(X)

# Train successive layers
tm_layer2 = ct.Corex(n_hidden=10)
tm_layer2.fit(topic_model.labels)

tm_layer3 = ct.Corex(n_hidden=1)
tm_layer3.fit(tm_layer2.labels)

Visualizations of the hierarchical topic model can be accessed through vis_topic.py.

vt.vis_hierarchy([topic_model, tm_layer2, tm_layer3], column_label=words, max_edges=300, prefix='topic-model-example')

Technical notes

Binarization of Documents

For speed reasons, this version of the CorEx topic model works only on binary data and produces binary latent factors. Despite this limitation, our work demonstrates CorEx produces coherent topics that are as good as or better than those produced by LDA for short to medium length documents. However, you may wish to consider additional preprocessing for working with longer documents. We have several strategies for handling text data.

  1. Naive binarization. This will be good for documents of similar length and especially short- to medium-length documents.

  2. Average binary bag of words. We split documents into chunks, compute the binary bag of words for each documents and then average. This implicitly weights all documents equally.

  3. All binary bag of words. Split documents into chunks and consider each chunk as its own binary bag of words documents.This changes the number of documents so it may take some work to match the ids back, if desired. Implicitly, this will weight longer documents more heavily. Generally this seems like the most theoretically justified method. Ideally, you could aggregate the latent factors over sub-documents to get 'counts' of latent factors at the higher layers.

  4. Fractional counts. This converts counts into a fraction of the background rate, with 1 as the max. Short documents tend to stay binary and words in long documents are weighted according to their frequency with respect to background in the corpus. This seems to work Ok on tests. It requires no preprocessing of count data and it uses the full range of possible inputs. However, this approach is not very rigorous or well tested.

For the python API, for 1 and 2, you can use the functions in vis_topic to process data or do the same yourself. Naive binarization is specified through the python api with count='binarize' and fractional counts with count='fraction'. While fractional counts may be work theoretically, their usage in the CorEx topic model has not be adequately tested.

Single Membership of Words in Topics

Also for speed reasons, the CorEx topic model enforces single membership of words in topics. If a user anchors a word to multiple topics, the single membership will be overriden.

References

If you use this code, please cite the following:

Gallagher, R. J., Reing, K., Kale, D., and Ver Steeg, G. "Anchored Correlation Explanation: Topic Modeling with Minimal Domain Knowledge." Transactions of the Association for Computational Linguistics (TACL), 2017.

See the following papers if you interested in how CorEx works generally beyond sparse binary data.

Discovering Structure in High-Dimensional Data Through Correlation Explanation, Ver Steeg and Galstyan, NIPS 2014.

Maximally Informative Hierarchical Representions of High-Dimensional Data, Ver Steeg and Galstyan, AISTATS 2015.

Owner
Greg Ver Steeg
Research professor at USC
Greg Ver Steeg
A PyTorch implementation of paper "Learning Shared Semantic Space for Speech-to-Text Translation", ACL (Findings) 2021

Chimera: Learning Shared Semantic Space for Speech-to-Text Translation This is a Pytorch implementation for the "Chimera" paper Learning Shared Semant

Chi Han 43 Dec 28, 2022
An attempt to map the areas with active conflict in Ukraine using open source twitter data.

Live Action Map (LAM) An attempt to use open source data on Twitter to map areas with active conflict. Right now it is used for the Ukraine-Russia con

Kinshuk Dua 171 Nov 21, 2022
Findings of ACL 2021

Assessing Dialogue Systems with Distribution Distances [arXiv][code] We propose to measure the performance of a dialogue system by computing the distr

Yahui Liu 16 Feb 24, 2022
An ultra fast tiny model for lane detection, using onnx_parser, TensorRTAPI, torch2trt to accelerate. our model support for int8, dynamic input and profiling. (Nvidia-Alibaba-TensoRT-hackathon2021)

Ultra_Fast_Lane_Detection_TensorRT An ultra fast tiny model for lane detection, using onnx_parser, TensorRTAPI to accelerate. our model support for in

steven.yan 121 Dec 27, 2022
American Sign Language (ASL) to Text Converter

Signterpreter American Sign Language (ASL) to Text Converter Recommendations Although there is grayscale and gaussian blur, we recommend that you use

0 Feb 20, 2022
Conversational-AI-ChatBot - Intelligent ChatBot built with Microsoft's DialoGPT transformer to make conversations with human users!

Conversational AI ChatBot Intelligent ChatBot built with Microsoft's DialoGPT transformer to make conversations with human users! In this project? Thi

Rajkumar Lakshmanamoorthy 6 Nov 30, 2022
Text editor on python to convert english text to malayalam(Romanization/Transiteration).

Manglish Text Editor This is a simple transiteration (romanization ) program which is used to convert manglish to malayalam (converts njaan to ഞാൻ ).

Merin Rose Tom 1 May 11, 2022
Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS)

TOPSIS implementation in Python Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) CHING-LAI Hwang and Yoon introduced TOPSIS

Hamed Baziyad 8 Dec 10, 2022
Easy, fast, effective, and automatic g-code compression!

Getting to the meat of g-code. Easy, fast, effective, and automatic g-code compression! MeatPack nearly doubles the effective data rate of a standard

Scott Mudge 97 Nov 21, 2022
KR-FinBert And KR-FinBert-SC

KR-FinBert & KR-FinBert-SC Much progress has been made in the NLP (Natural Language Processing) field, with numerous studies showing that domain adapt

5 Jul 29, 2022
Simple text to phones converter for multiple languages

Phonemizer -- foʊnmaɪzɚ The phonemizer allows simple phonemization of words and texts in many languages. Provides both the phonemize command-line tool

CoML 762 Dec 29, 2022
Production First and Production Ready End-to-End Keyword Spotting Toolkit

Production First and Production Ready End-to-End Keyword Spotting Toolkit

223 Jan 02, 2023
Finally decent dictionaries based on Wiktionary for your beloved eBook reader.

eBook Reader Dictionaries Finally, decent dictionaries based on Wiktionary for your beloved eBook reader. Dictionaries Catalan 🚧 Ελληνικά (help welco

Mickaël Schoentgen 163 Dec 31, 2022
Constituency Tree Labeling Tool

Constituency Tree Labeling Tool The purpose of this package is to solve the constituency tree labeling problem. Look from the dataset labeled by NLTK,

张宇 6 Dec 20, 2022
A Lightweight NLP Data Loader for All Deep Learning Frameworks in Python

LineFlow: Framework-Agnostic NLP Data Loader in Python LineFlow is a simple text dataset loader for NLP deep learning tasks. LineFlow was designed to

TofuNLP 177 Jan 04, 2023
Global Rhythm Style Transfer Without Text Transcriptions

Global Prosody Style Transfer Without Text Transcriptions This repository provides a PyTorch implementation of AutoPST, which enables unsupervised glo

Kaizhi Qian 193 Dec 30, 2022
Fidibo.com comments Sentiment Analyser

Fidibo.com comments Sentiment Analyser Introduction This project first asynchronously grab Fidibo.com books comment data using grabber.py and then sav

Iman Kermani 3 Apr 15, 2022
2021海华AI挑战赛·中文阅读理解·技术组·第三名

文字是人类用以记录和表达的最基本工具,也是信息传播的重要媒介。透过文字与符号,我们可以追寻人类文明的起源,可以传播知识与经验,读懂文字是认识与了解的第一步。对于人工智能而言,它的核心问题之一就是认知,而认知的核心则是语义理解。

21 Dec 26, 2022
A2T: Towards Improving Adversarial Training of NLP Models (EMNLP 2021 Findings)

A2T: Towards Improving Adversarial Training of NLP Models This is the source code for the EMNLP 2021 (Findings) paper "Towards Improving Adversarial T

QData 17 Oct 15, 2022