# TopicGPT
TopicGPT integrates the remarkable capabilities of current LLMs such as GPT-3.5 and GPT-4 into topic modelling.
While traditional topic models extract topics as simple lists of top-words, such as ["Lion", "Leopard", "Rhino", "Elephant", "Buffalo"], TopicGPT offers rich and dynamic topic representations that can be intuitively understood, extensively investigated and modified in various ways via a simple text commands in natural language.
More specifically, it provides the following core functionalities:
- Identification of clusters within document-embeddings and top-word extraction
- Generation of informative topic descriptions
- Extraction of detailed information about topics via Retrieval-Augmented-Generation (RAG)
- Comparison of topics
- Splitting and combining of identified topics
- Addition of new topics based on keywords
- Deletion of topics
When directly interacting with TopicGPT via prompting and without explicitly calling functions, an LLM autonomously decides which functionality to use.
## Installation
You can install TopicGPT via [PyPI](https://pypi.org/project/topicgpt/)
```
pip install topicgpt
```
## Further Documentation
You can find detailed documentation of the available classes and functions [here](https://lmu-seminar-llms.github.io/TopicGPT/).
## Example
The following shoort example demonstrates how TopicGPT could be used on a real-world dataset. The Twenty Newsgroups corpus (https://scikit-learn.org/0.19/datasets/twenty_newsgroups.html) is used for this purpose.
Further example-notebooks can be found under examples/ in the repository.
### Load the data
```python
from sklearn.datasets import fetch_20newsgroups
data = fetch_20newsgroups(subset='all', remove=('headers', 'footers', 'quotes')) #download the 20 Newsgroups dataset
corpus = data['data']
corpus = [doc for doc in corpus if doc != ""] #remove empty documents
```
### Initialize the model
Note that an OpenAi API-Key is needed to compute the embeddings and execute the prompts. See https://platform.openai.com/account/api-keys for more details. We select 20 topics in this case since the Twenty Newsgroups corpus comprises documents from 20 different newsgroups. It is also possible to let Hdbscan determine the number of topics automatically.
```python
from topicgpt.TopicGPT import TopicGPT
tm = TopicGPT(
openai_api_key = <your-openai-api-key>,
n_topics = 20 # select 20 topics since the true number of topics is 20
)
```
### Fit the model
The fit-method fits the model. This can take, depending on the size of the dataset and wether embeddings have been provided, from a few minutes to several hours. Especially the computation of the embeddings can take some time.
```python
tm.fit(corpus) # the corpus argument should be of type list[str] where each string represents one document
```
### Inspect the found topics
Obtain an overview over the indentified topics
```python
print(tm.topic_lis)
```
_Output_
```
[Topic 0: Electronics Equipment Sales,
Topic 1: Image Processing,
Topic 2: Gun control,
Topic 3: Online Privacy and Anonymity,
Topic 4: Conflict and Violence.,
Topic 5: Computer Hardware,
Topic 6: Belief and Atheism,
Topic 7: Online Discussions,
Topic 8: Computer Software,
Topic 9: Car Features and Performance,
Topic 10: Encryption and Government,
Topic 11: Technology and Computing.,
Topic 12: Technology and Computing,
Topic 13: Space Exploration,
Topic 14: Motorcycle Riding Techniques,
Topic 15: Technology,
Topic 16: Hockey Games,
Topic 17: Health and Medicine.,
Topic 18: Baseball games and teams.,
Topic 19: Beliefs about Homosexuality.]
```
To obtain more detailed information on each topic, we can call the "print_topics" method:
```python
tm.print_topics()
```
_Output_
```
Topic 0: Electronics Equipment Sales
Topic_description: The common topic of the given words appears to be "electronics and technology".
Various aspects and sub-topics of this topic include:
1. Buying and selling: "offer", "sale", "sell", "price", "buy"
2. Device usage and features: "use", "get", "new", "used", "condition"
3. Technical specifications: "wire", "ground", "power", "circuit", "voltage"
4. Communication and connectivity: "phone", "email", "modem", "wireless", "connection"
5. Accessories and peripherals: "battery", "cable", "manuals", "disk", "monitor"
Top words: ["n't", 'one', 'would', 'use', 'like', 'get', 'new', 'used', 'offer', 'sale']
[...]
```
We can also visualize the resulting clusters to get an overview of the shape and size of the clusters
```
tm.visualize_clusters()
```
### Find out more detailed information about the identified topics
First, we might be interested in knowing what information the space topic (topic 13) contains on the moon landing.
```python
tm.pprompt("Which information related to the keyword 'moon landing' does topic 13 have?")
```
_Output_
```
GPT wants to the call the function: {
"name": "knn_search",
"arguments": "{\n \"topic_index\": 13,\n \"query\": \"moon landing\",\n \"k\": 5\n}"
}
Topic 13, which is related to the keyword "moon landing," has the following information:
1. Document index 258: This document provides an introduction to the solar system and mentions that advancements in rocketry after World War II enabled machines to travel to the Moon and other planets. It highlights that the United States has sent both automated spacecraft and human-crewed expeditions to explore the Moon.
2. Document index 535: This document discusses a $65 million program called the Back to the Moon bill, which aims to encourage private companies to develop lunar orbiters. It mentions that there is a chance of making a lunar mission happen in this decade through this program.
3. Document index 357: This document is a request for more information on a recent newspaper article about the Japanese crashing or crash-landing a package on the Moon. It indicates that the article was vague and unclear.
4. Document index 321: This document speculates about what would have happened if the Soviets had beaten the United States in the Moon race. It suggests that the US would have still performed Moon landings and potentially set up a lunar base. The focus on Mars exploration would have depended on the Soviets' actions.
5. Document index 102: This document mentions the Hiten engineering-test mission, which spent time in a highly eccentric Earth orbit and performed lunar flybys before being inserted into lunar orbit using gravity-assist-like maneuvers. It states that the mission was expected to crash on the Moon eventually.
Please note that the above summaries are based on the content of the documents and may not capture all the information contained within them.
```
From this output we see that an instance of a GPT decided to call the function "knn_search" from the class "TopicPrompting". Indeed some documents on the topic "moon landing" have been found and the model summarizes the relevant information accordingly.
If we want to check, for instance the document with index 102 in topic 13 to learn more about the Hiten engineering-test mission, we can simply do the following:
```python
print(tm.topic_lis[13].documents[535])
```
_Output_
```
Their Hiten engineering-test mission spent a while in a highly eccentric Earth orbit doing lunar flybys, and then was inserted into lunar orbit using some very tricky gravity-assist-like maneuvering. This meant that it would crash on the Moon eventually, since there is no such thing as a stable lunar orbit (as far as anyone knows), and I believe I recall hearing recently that it was about to happen.
```
### Split Topics
We find that topic 6 about Belief and Atheism is a bit general and would thus like to split see what 5 potential subtopics it contains:
```python
tm.pprompt("What are 5 potential subtopics of topic 6")
```
_Output_
```
GPT wants to the call the function: {
"name": "split_topic_kmeans",
"arguments": "{\n \"topic_idx\": 6,\n \"n_clusters\": 5\n}"
}
Here are five potential subtopics of topic 6:
1. Existence of God: This subtopic explores arguments, evidence, proofs, and the existence of God.
2. Atheism vs Theism: This subtopic delves into the concepts of atheism, theism, atheists, and theists, and the debates surrounding these belief systems.
3. Belief and Faith: This subtopic focuses on beliefs, faith, believers, and the roles they play in religious and philosophical contexts.
4. Rationality and Reasoning: This subtopic involves discussions on reason, rationality, logic, and argumentation as they relate to beliefs and debates about religion.
5. Morality and Ethics: This subtopic explores moral values, ethics, morals, and how they influence and are influenced by religious beliefs and debates.
These subtopics were inferred from the words provided and the descriptions of the topics in the corpus used for topic modeling.
```
Based on those results we decide to not actually split the topic into 5 subtopics but rather just 3 - One on "religious faith", one on "atheism" and one on "ethics and philosophy".
We also want to actually make those changes to the existing topics, so we tell TopicGPT to do the changes inplace.
```python
tm.pprompt("Please split topic 6 into subtopics based on the keywords 'religious faith', 'atheism' and 'ethics and philosophy'. Do this inplace.")
```
_Output_
```
I have split topic 6 into three subtopics based on the provided keywords: "religious faith", "atheism", and "ethics and philosophy". Here are the new subtopics:
1. Topic 21: Religious Beliefs
- This topic encompasses the general theme of "Religion and Belief".
- Subtopics and aspects of this topic include:
- Faith and Beliefs: Believing, belief, true, evidence, truth.
- Religious Practices: Praying, heretic, bible, church.
- Religious Communities: Christian, Muslim, atheist.
- Interpretation and Understanding: Interpretations, interpretation, understanding, context.
- Morality and Ethics: Good, wrong, punishment, violence.
2. Topic 22: Existence of God
- This topic focuses on "Atheism and Belief".
- Subtopics and aspects of this topic include:
- Arguments and Evidence: Atheists provide evidence for their beliefs.
- Existence of God: Atheists question the existence of God.
- Religious Beliefs: Atheists challenge religious beliefs.
- Science and Reason: Atheists rely on science and reason.
- Morality and Ethics: Atheists discuss morality and ethics.
3. Topic 23: Ethics and Morality
- This topic revolves around "Morality and Ethics".
- Subtopics and aspects of this topic include:
- Moral reasoning: "Argument," "reasoning," "logic," "fallacy," "premises."
- Ethical principles: "Right," "wrong," "immoral," "morals," "ethical."
- Moral judgments: "Moral," "ethical," "morality," "judgments," "values."
- Human behavior: "Homosexual," "sexual," "promiscuous," "behavior," "sex."
- Scientific perspective: "Science," "scientific," "evidence," "empirical," "studies."
Please note that these subtopics are based on the keywords provided and may not cover all aspects of the original topic.
```
We see that TopicGPT performed the splitting as intended. However, the names and descriptions of the topics got changed slightly in order to optimally fit to the documents that actually constitute the topics based on the keywords. (If you like you can rename them at any time).
### Combining topics
Since topics 15 ("Hockey Games") and 17 ("Baseball games and teams") are both about sports, we want to combine them into a single topic.
```python
tm.pprompt("Please combine topics 15 and 17. Do this inplace.")
```
_Output_
```
GPT wants to the call the function: {
"name": "combine_topics",
"arguments": "{\n \"topic_idx_lis\": [15, 17],\n \"inplace\": true\n}"
}
The topics 15 and 17 have been combined into a new topic called "Sports". This topic includes aspects and sub-topics related to sports such as team and players, games and seasons, performance and skills, fans and audience, and statistics and records. Some of the common words found in this topic include "team," "players," "hockey," "baseball," "game," "games," "season," "playoffs," "good," "better," "win," "hit," "score," "fans," "series," "watch," "fan," "stats," "record," "pts," and "career".
```
## Limitations and Caveats
It is important to note that, as a model built on top of inherently stochastic LLMs and all their shortcomings, TopicGPT has several limitations and shortcomings as well. LLMs are Machine Learning models and as such, they are not perfect at solving the intended tasks; They may be useful because they are correct reasonably often, but they can always fail. The following list is not complete, but may provide useful information on what may go wrong when using TopicGPT:
- **Hallucination**: LLMs are well known for yielding incorrect but coherent and plausible answers that seem convincing but are actually just made up. Although we tried to minimize this undesired behavior through carefully designing the used prompts, we found that TopicGPT may hallucinate (especially) with respect to the following aspects:
- Making up, distorting or misinterpreting content of documents retrieved via knn-search.
- Incorrectly naming and describing topics based on top-words. Specifically, the model can identify topics that seem coherent and reasonable although the corresponding documents are not actually related.
- **Unsdesired Behaviour**: When using the "prompt" or "pprompt" function, TopicGPT may not call the function you intended it to call. This can be alleviated by explicitly telling the model which function to use or directly calling the function yourself. It sometimes also tires to call invalid functions or functions with invalid arguments.
- **Stoachasticity**: The behavior of TopicGPT is not deterministic and exhibits some randomness. There is always some probability that certain actions do not work as intended at the first try because some components of the LLM do not function as desired. Simply trying again should mostly help with those issues.
- On the other hand, TopicGPT may also be overly cautious and report that no relevant information has been found or no topic exists that matches a certain keyword even though it does. This could be caused by designing prompts to prevent massive occurrence of falsely positive results.
Note that using GPT-4 in TopicGPT can help to significantly alleviate issues with hallucination.
- **Erroneous embeddings**: The document- and word-embeddings used in TopicGPT may not always reflect the actual semantics of the texts correctly. More specifically, the embeddings sometimes reflect, for instance, grammatical or orthographical aspects such that clusters based on those aspects may be identified.
- **Size of the dataset**: TopicGPT might fail when the dataset is too small (less than 1000 documents). This is because then the identified topics might become very small and noisy. The RAG aspect will also likely not work as intended. Datasets of more than 10,000 documents are recommended. Note that the processing of very large datasets might not fit into the main memory of your computer.
## Tips and tricks for prompting TopicGPT
When using the "pprompt" or "prompt" function, TopicGPT can behave differently than intended. To alleviate those issues some simple tricks can help:
- Explicitly tell the model which function it should use and which parameters to select. (Sometimes the model simply cannot know what you expect it to do.) For example, instead of using ```tm.pprompt("What are the subtopic of topic 13?")```, use something like ```tm.pprompt("What are the subtopic of topic 13? Please use the function that uses the k-means algorithm to split the topic. Use a parameter of k = 5 and do this inplace")```
- Just ask the same prompt again. Since TopicGPT is a stochastic system, calling the same function with the same argument again might yield a different functionality to be used or a different result.
- If this doesn't help, you can also directly call the function you want to use from the TopicPrompting class. In the example above you could do ```tm.topic_prompting.split_topic_kmeans(topic_idx = 13, n_clusters = 5, inplace = True)```. Note that all functions the model can call can also be called directly.
- In case of hallucination of facts it may help to use GPT-4 for TopicGPT
## How TopicGPT works
TopicGPT is centrally built on top of text embeddings and the prompting mechanisms obtained via LLMs and provided by the OpenAI API. Please also see the section [References](#references) for more details on the models and ideas used in TopicGPT.
### Embeddings
When no embeddings are provided, TopicGPT automatically computes the embeddings of the documents of the provided corpus and also of the vocabulary that is extracted from the corpus. This happens after the fit-method is called.
The class ```GetEmbeddingsOpenAI``` is used for this purpose.
### Clustering
In order to identify topics among the documents, TopicGPT reduces the dimensionality of the document embeddings via UMAP and then uses Hdbscan to identify the clusters. Dimensionality reduction is necessary since the document embeddings are of very high dimensionality and thus the curse of dimensionality would make it very difficult, if not impossible, to identify the clusters.
When not specifying the number of topics in the ```Topic GPT``` class, Hdbscan is used to automatically determine the number of topics. If the number of topics is specified, agglomerative clustering is used on top of the clusters identified by HDBSCAN.
The class ```Clustering``` is used for this purpose.
### Extraction of Top-Words
After the clusters have been identified, TopicGPT extracts the top-words of each topic. This is done via two different methods:
- **Tf-idf**: The tf-idf method is based on the idea that words that occur frequently in a topic but rarely in other topics are good indicators for the topic. The top-words are thus the words with the highest tf-idf scores.
- **Centroid similarity**: The centroid similarity method is based on the idea that the words that are closest to the centroid of a topic are good indicators for the topic. The top-words are thus the words that are closest to the centroid of the topic.
Note that the Tf-idf heuristic was introduced for the BerTopic Model (Grootendorst, Maarten. "BERTopic: Neural topic modeling with a class-based TF-IDF procedure." arXiv preprint arXiv:2203.05794 (2022)) and a similar idea to the centroid similarity method is used in Top2Vec (Angelov, Dimo. "Top2vec: Distributed representations of topics." arXiv preprint arXiv:2008.09470 (2020)).
Topword extraction is performed with help of the class ```ExtractTopWords```.
### Describing and naming topics
In the next step, all topics are provided with a short name and a description. This is done via prompting an LLM provided by OpenAI with around 500 top-words of each topic. The LLM then generates a short name and a description for each topic.
The class ```TopwordEnhancement``` is used for this purpose.
Note that computation of Embeddings, Extraction of Top-Words and Describing and Naming Topics are all performed when calling the ```fit``` method of the ```TopicGPT``` class.
### Prompting of TopicGPT
When formalizing a prompt via the ```pprompt``` or ```prompt``` function, TopicGPT uses the following steps:
1. The prompt, together with basic model- and corpus-information, is sent to an LLM provided by OpenAI. The LLM then decides which function of the ```TopicPrompting``` class to call. The LLM also decides which arguments to use for the function.
2. The function is called according to the information by the LLM. The full result of the function will be returned to the user.
3. Parts of the results of the function are returned to the LLM. The LLM then generates a short answer of the original prompt with help of the function result and returns it to the user.
## References
The following models, software packages and ideas are central for TopicGPT:
- **UMAP**: The Uniform Manifold Approximation and Projection for Dimension Reduction algorithm is used for reducing the dimensionality of document- and word embeddings (McInnes, Leland, John Healy, and James Melville. "Umap: Uniform manifold approximation and projection for dimension reduction." arXiv preprint arXiv:1802.03426 (2018).)
- **HDBSCAN**: Hierarchical density based clustering is used to identify the clusters among the dimensionality reduced topics (McInnes, Leland, John Healy, and Steve Astels. "hdbscan: Hierarchical density based clustering." J. Open Source Softw. 2.11 (2017): 205.)
- **Agglomerative Clustering**: The agglomerative clustering functionality from sklearn is used to combine topics in case the identified number of clusters exeeds the number of topics specified by the user (Pedregosa, Fabian, et al. "Scikit-learn: Machine learning in Python." the Journal of machine Learning research 12 (2011): 2825-2830., https://scikit-learn.org/stable/modules/generated/sklearn.cluster.AgglomerativeClustering.html)
- **Topword extraction**: Even though the corresponding packages are not directly used, the topword extraction methods used for this package are based on very similar ideas as found in the BerTopic Model (Grootendorst, Maarten. "BERTopic: Neural topic modeling with a class-based TF-IDF procedure." arXiv preprint arXiv:2203.05794 (2022)) in the case of the tf-idf method and in Top2Vec for the centroid-similarity method (Angelov, Dimo. "Top2vec: Distributed representations of topics." arXiv preprint arXiv:2008.09470 (2020)).
- **LLMs from the GPT family**: Some references for the models for computing embeddings and answering the prompts include:
- Brown, Tom B., et al. “Language Models are Few-Shot Learners.” Advances in Neural Information Processing Systems 33 (2020).
- Radford, Alec, et al. “GPT-4: Generative Pre-training of Transformers with Discrete Latent Variables.” arXiv preprint arXiv:2302.07413 (2023).
- Radford, Alec, et al. “Improving Language Understanding by Generative Pre-Training.” URL: https://s3-us-west-2.amazonaws.com/openai-assets/research-covers/language-unsupervised/language_understanding_paper.pdf. [6]
- Radford, Alec, et al. “Language Models are Unsupervised Multitask Learners.” OpenAI Blog 1.8 (2019): 9. [7]
Raw data
{
"_id": null,
"home_page": "",
"name": "topicgpt",
"maintainer": "",
"docs_url": null,
"requires_python": "",
"maintainer_email": "",
"keywords": "Topic Modelling,GPT,LLM,OpenAI,Retrieval Augmented Generation,Chat-GPT,GPT-3,GPT-4",
"author": "Arik Reuter",
"author_email": "arik_reuter@gmx.de",
"download_url": "https://files.pythonhosted.org/packages/13/2c/e6b19ee0f0ad1e60d4ef94293391f1b7bdffcd66d11af09c1b9a12165bfe/topicgpt-1.0.0.tar.gz",
"platform": null,
"description": "# TopicGPT\r\nTopicGPT integrates the remarkable capabilities of current LLMs such as GPT-3.5 and GPT-4 into topic modelling. \r\n\r\nWhile traditional topic models extract topics as simple lists of top-words, such as [\"Lion\", \"Leopard\", \"Rhino\", \"Elephant\", \"Buffalo\"], TopicGPT offers rich and dynamic topic representations that can be intuitively understood, extensively investigated and modified in various ways via a simple text commands in natural language. \r\n\r\nMore specifically, it provides the following core functionalities: \r\n- Identification of clusters within document-embeddings and top-word extraction\r\n- Generation of informative topic descriptions \r\n- Extraction of detailed information about topics via Retrieval-Augmented-Generation (RAG)\r\n- Comparison of topics\r\n- Splitting and combining of identified topics\r\n- Addition of new topics based on keywords\r\n- Deletion of topics\r\n \r\nWhen directly interacting with TopicGPT via prompting and without explicitly calling functions, an LLM autonomously decides which functionality to use.\r\n\r\n## Installation\r\n\r\nYou can install TopicGPT via [PyPI](https://pypi.org/project/topicgpt/)\r\n\r\n```\r\npip install topicgpt\r\n```\r\n\r\n## Further Documentation\r\n\r\nYou can find detailed documentation of the available classes and functions [here](https://lmu-seminar-llms.github.io/TopicGPT/).\r\n\r\n\r\n## Example \r\n\r\nThe following shoort example demonstrates how TopicGPT could be used on a real-world dataset. The Twenty Newsgroups corpus (https://scikit-learn.org/0.19/datasets/twenty_newsgroups.html) is used for this purpose. \r\n\r\nFurther example-notebooks can be found under examples/ in the repository.\r\n\r\n### Load the data\r\n\r\n```python\r\nfrom sklearn.datasets import fetch_20newsgroups\r\n\r\ndata = fetch_20newsgroups(subset='all', remove=('headers', 'footers', 'quotes')) #download the 20 Newsgroups dataset\r\ncorpus = data['data'] \r\n\r\ncorpus = [doc for doc in corpus if doc != \"\"] #remove empty documents\r\n```\r\n### Initialize the model \r\n\r\nNote that an OpenAi API-Key is needed to compute the embeddings and execute the prompts. See https://platform.openai.com/account/api-keys for more details. We select 20 topics in this case since the Twenty Newsgroups corpus comprises documents from 20 different newsgroups. It is also possible to let Hdbscan determine the number of topics automatically. \r\n\r\n```python \r\nfrom topicgpt.TopicGPT import TopicGPT\r\n\r\ntm = TopicGPT(\r\n openai_api_key = <your-openai-api-key>,\r\n n_topics = 20 # select 20 topics since the true number of topics is 20 \r\n)\r\n```\r\n\r\n### Fit the model \r\n\r\nThe fit-method fits the model. This can take, depending on the size of the dataset and wether embeddings have been provided, from a few minutes to several hours. Especially the computation of the embeddings can take some time. \r\n\r\n```python \r\ntm.fit(corpus) # the corpus argument should be of type list[str] where each string represents one document\r\n```\r\n\r\n### Inspect the found topics\r\n\r\nObtain an overview over the indentified topics\r\n```python\r\nprint(tm.topic_lis)\r\n```\r\n_Output_\r\n```\r\n[Topic 0: Electronics Equipment Sales,\r\n Topic 1: Image Processing,\r\n Topic 2: Gun control,\r\n Topic 3: Online Privacy and Anonymity,\r\n Topic 4: Conflict and Violence.,\r\n Topic 5: Computer Hardware,\r\n Topic 6: Belief and Atheism,\r\n Topic 7: Online Discussions,\r\n Topic 8: Computer Software,\r\n Topic 9: Car Features and Performance,\r\n Topic 10: Encryption and Government,\r\n Topic 11: Technology and Computing.,\r\n Topic 12: Technology and Computing,\r\n Topic 13: Space Exploration,\r\n Topic 14: Motorcycle Riding Techniques,\r\n Topic 15: Technology,\r\n Topic 16: Hockey Games,\r\n Topic 17: Health and Medicine.,\r\n Topic 18: Baseball games and teams.,\r\n Topic 19: Beliefs about Homosexuality.]\r\n```\r\nTo obtain more detailed information on each topic, we can call the \"print_topics\" method: \r\n\r\n```python\r\ntm.print_topics()\r\n```\r\n_Output_\r\n```\r\nTopic 0: Electronics Equipment Sales\r\n\r\nTopic_description: The common topic of the given words appears to be \"electronics and technology\". \r\n\r\nVarious aspects and sub-topics of this topic include:\r\n1. Buying and selling: \"offer\", \"sale\", \"sell\", \"price\", \"buy\"\r\n2. Device usage and features: \"use\", \"get\", \"new\", \"used\", \"condition\"\r\n3. Technical specifications: \"wire\", \"ground\", \"power\", \"circuit\", \"voltage\"\r\n4. Communication and connectivity: \"phone\", \"email\", \"modem\", \"wireless\", \"connection\"\r\n5. Accessories and peripherals: \"battery\", \"cable\", \"manuals\", \"disk\", \"monitor\"\r\nTop words: [\"n't\", 'one', 'would', 'use', 'like', 'get', 'new', 'used', 'offer', 'sale']\r\n\r\n[...]\r\n```\r\nWe can also visualize the resulting clusters to get an overview of the shape and size of the clusters\r\n```\r\ntm.visualize_clusters()\r\n```\r\n\r\n### Find out more detailed information about the identified topics\r\n\r\nFirst, we might be interested in knowing what information the space topic (topic 13) contains on the moon landing. \r\n\r\n```python \r\ntm.pprompt(\"Which information related to the keyword 'moon landing' does topic 13 have?\")\r\n```\r\n\r\n_Output_\r\n```\r\nGPT wants to the call the function: {\r\n \"name\": \"knn_search\",\r\n \"arguments\": \"{\\n \\\"topic_index\\\": 13,\\n \\\"query\\\": \\\"moon landing\\\",\\n \\\"k\\\": 5\\n}\"\r\n}\r\nTopic 13, which is related to the keyword \"moon landing,\" has the following information:\r\n\r\n1. Document index 258: This document provides an introduction to the solar system and mentions that advancements in rocketry after World War II enabled machines to travel to the Moon and other planets. It highlights that the United States has sent both automated spacecraft and human-crewed expeditions to explore the Moon.\r\n\r\n2. Document index 535: This document discusses a $65 million program called the Back to the Moon bill, which aims to encourage private companies to develop lunar orbiters. It mentions that there is a chance of making a lunar mission happen in this decade through this program.\r\n\r\n3. Document index 357: This document is a request for more information on a recent newspaper article about the Japanese crashing or crash-landing a package on the Moon. It indicates that the article was vague and unclear.\r\n\r\n4. Document index 321: This document speculates about what would have happened if the Soviets had beaten the United States in the Moon race. It suggests that the US would have still performed Moon landings and potentially set up a lunar base. The focus on Mars exploration would have depended on the Soviets' actions.\r\n\r\n5. Document index 102: This document mentions the Hiten engineering-test mission, which spent time in a highly eccentric Earth orbit and performed lunar flybys before being inserted into lunar orbit using gravity-assist-like maneuvers. It states that the mission was expected to crash on the Moon eventually.\r\n\r\nPlease note that the above summaries are based on the content of the documents and may not capture all the information contained within them.\r\n```\r\n\r\nFrom this output we see that an instance of a GPT decided to call the function \"knn_search\" from the class \"TopicPrompting\". Indeed some documents on the topic \"moon landing\" have been found and the model summarizes the relevant information accordingly. \r\n\r\nIf we want to check, for instance the document with index 102 in topic 13 to learn more about the Hiten engineering-test mission, we can simply do the following:\r\n\r\n```python\r\nprint(tm.topic_lis[13].documents[535])\r\n```\r\n_Output_\r\n```\r\nTheir Hiten engineering-test mission spent a while in a highly eccentric Earth orbit doing lunar flybys, and then was inserted into lunar orbit using some very tricky gravity-assist-like maneuvering. This meant that it would crash on the Moon eventually, since there is no such thing as a stable lunar orbit (as far as anyone knows), and I believe I recall hearing recently that it was about to happen.\r\n```\r\n\r\n### Split Topics\r\nWe find that topic 6 about Belief and Atheism is a bit general and would thus like to split see what 5 potential subtopics it contains: \r\n\r\n```python\r\ntm.pprompt(\"What are 5 potential subtopics of topic 6\")\r\n```\r\n\r\n_Output_\r\n\r\n```\r\nGPT wants to the call the function: {\r\n \"name\": \"split_topic_kmeans\",\r\n \"arguments\": \"{\\n \\\"topic_idx\\\": 6,\\n \\\"n_clusters\\\": 5\\n}\"\r\n}\r\n\r\nHere are five potential subtopics of topic 6:\r\n\r\n1. Existence of God: This subtopic explores arguments, evidence, proofs, and the existence of God.\r\n\r\n2. Atheism vs Theism: This subtopic delves into the concepts of atheism, theism, atheists, and theists, and the debates surrounding these belief systems.\r\n\r\n3. Belief and Faith: This subtopic focuses on beliefs, faith, believers, and the roles they play in religious and philosophical contexts.\r\n\r\n4. Rationality and Reasoning: This subtopic involves discussions on reason, rationality, logic, and argumentation as they relate to beliefs and debates about religion.\r\n\r\n5. Morality and Ethics: This subtopic explores moral values, ethics, morals, and how they influence and are influenced by religious beliefs and debates.\r\n\r\nThese subtopics were inferred from the words provided and the descriptions of the topics in the corpus used for topic modeling.\r\n\r\n```\r\nBased on those results we decide to not actually split the topic into 5 subtopics but rather just 3 - One on \"religious faith\", one on \"atheism\" and one on \"ethics and philosophy\".\r\nWe also want to actually make those changes to the existing topics, so we tell TopicGPT to do the changes inplace. \r\n\r\n```python\r\ntm.pprompt(\"Please split topic 6 into subtopics based on the keywords 'religious faith', 'atheism' and 'ethics and philosophy'. Do this inplace.\")\r\n```\r\n_Output_\r\n\r\n```\r\nI have split topic 6 into three subtopics based on the provided keywords: \"religious faith\", \"atheism\", and \"ethics and philosophy\". Here are the new subtopics:\r\n\r\n1. Topic 21: Religious Beliefs\r\n - This topic encompasses the general theme of \"Religion and Belief\".\r\n - Subtopics and aspects of this topic include:\r\n - Faith and Beliefs: Believing, belief, true, evidence, truth.\r\n - Religious Practices: Praying, heretic, bible, church.\r\n - Religious Communities: Christian, Muslim, atheist.\r\n - Interpretation and Understanding: Interpretations, interpretation, understanding, context.\r\n - Morality and Ethics: Good, wrong, punishment, violence.\r\n\r\n2. Topic 22: Existence of God\r\n - This topic focuses on \"Atheism and Belief\".\r\n - Subtopics and aspects of this topic include:\r\n - Arguments and Evidence: Atheists provide evidence for their beliefs.\r\n - Existence of God: Atheists question the existence of God.\r\n - Religious Beliefs: Atheists challenge religious beliefs.\r\n - Science and Reason: Atheists rely on science and reason.\r\n - Morality and Ethics: Atheists discuss morality and ethics.\r\n\r\n3. Topic 23: Ethics and Morality\r\n - This topic revolves around \"Morality and Ethics\".\r\n - Subtopics and aspects of this topic include:\r\n - Moral reasoning: \"Argument,\" \"reasoning,\" \"logic,\" \"fallacy,\" \"premises.\"\r\n - Ethical principles: \"Right,\" \"wrong,\" \"immoral,\" \"morals,\" \"ethical.\"\r\n - Moral judgments: \"Moral,\" \"ethical,\" \"morality,\" \"judgments,\" \"values.\"\r\n - Human behavior: \"Homosexual,\" \"sexual,\" \"promiscuous,\" \"behavior,\" \"sex.\"\r\n - Scientific perspective: \"Science,\" \"scientific,\" \"evidence,\" \"empirical,\" \"studies.\"\r\nPlease note that these subtopics are based on the keywords provided and may not cover all aspects of the original topic.\r\n```\r\n\r\nWe see that TopicGPT performed the splitting as intended. However, the names and descriptions of the topics got changed slightly in order to optimally fit to the documents that actually constitute the topics based on the keywords. (If you like you can rename them at any time). \r\n\r\n### Combining topics\r\n\r\nSince topics 15 (\"Hockey Games\") and 17 (\"Baseball games and teams\") are both about sports, we want to combine them into a single topic.\r\n\r\n```python\r\ntm.pprompt(\"Please combine topics 15 and 17. Do this inplace.\")\r\n```\r\n_Output_\r\n\r\n```\r\nGPT wants to the call the function: {\r\n \"name\": \"combine_topics\",\r\n \"arguments\": \"{\\n \\\"topic_idx_lis\\\": [15, 17],\\n \\\"inplace\\\": true\\n}\"\r\n}\r\nThe topics 15 and 17 have been combined into a new topic called \"Sports\". This topic includes aspects and sub-topics related to sports such as team and players, games and seasons, performance and skills, fans and audience, and statistics and records. Some of the common words found in this topic include \"team,\" \"players,\" \"hockey,\" \"baseball,\" \"game,\" \"games,\" \"season,\" \"playoffs,\" \"good,\" \"better,\" \"win,\" \"hit,\" \"score,\" \"fans,\" \"series,\" \"watch,\" \"fan,\" \"stats,\" \"record,\" \"pts,\" and \"career\".\r\n```\r\n\r\n## Limitations and Caveats\r\n\r\nIt is important to note that, as a model built on top of inherently stochastic LLMs and all their shortcomings, TopicGPT has several limitations and shortcomings as well. LLMs are Machine Learning models and as such, they are not perfect at solving the intended tasks; They may be useful because they are correct reasonably often, but they can always fail. The following list is not complete, but may provide useful information on what may go wrong when using TopicGPT:\r\n\r\n- **Hallucination**: LLMs are well known for yielding incorrect but coherent and plausible answers that seem convincing but are actually just made up. Although we tried to minimize this undesired behavior through carefully designing the used prompts, we found that TopicGPT may hallucinate (especially) with respect to the following aspects:\r\n - Making up, distorting or misinterpreting content of documents retrieved via knn-search. \r\n - Incorrectly naming and describing topics based on top-words. Specifically, the model can identify topics that seem coherent and reasonable although the corresponding documents are not actually related.\r\n\r\n- **Unsdesired Behaviour**: When using the \"prompt\" or \"pprompt\" function, TopicGPT may not call the function you intended it to call. This can be alleviated by explicitly telling the model which function to use or directly calling the function yourself. It sometimes also tires to call invalid functions or functions with invalid arguments.\r\n\r\n- **Stoachasticity**: The behavior of TopicGPT is not deterministic and exhibits some randomness. There is always some probability that certain actions do not work as intended at the first try because some components of the LLM do not function as desired. Simply trying again should mostly help with those issues. \r\n - On the other hand, TopicGPT may also be overly cautious and report that no relevant information has been found or no topic exists that matches a certain keyword even though it does. This could be caused by designing prompts to prevent massive occurrence of falsely positive results. \r\n Note that using GPT-4 in TopicGPT can help to significantly alleviate issues with hallucination.\r\n\r\n- **Erroneous embeddings**: The document- and word-embeddings used in TopicGPT may not always reflect the actual semantics of the texts correctly. More specifically, the embeddings sometimes reflect, for instance, grammatical or orthographical aspects such that clusters based on those aspects may be identified.\r\n\r\n- **Size of the dataset**: TopicGPT might fail when the dataset is too small (less than 1000 documents). This is because then the identified topics might become very small and noisy. The RAG aspect will also likely not work as intended. Datasets of more than 10,000 documents are recommended. Note that the processing of very large datasets might not fit into the main memory of your computer.\r\n\r\n\r\n## Tips and tricks for prompting TopicGPT\r\nWhen using the \"pprompt\" or \"prompt\" function, TopicGPT can behave differently than intended. To alleviate those issues some simple tricks can help: \r\n\r\n- Explicitly tell the model which function it should use and which parameters to select. (Sometimes the model simply cannot know what you expect it to do.) For example, instead of using ```tm.pprompt(\"What are the subtopic of topic 13?\")```, use something like ```tm.pprompt(\"What are the subtopic of topic 13? Please use the function that uses the k-means algorithm to split the topic. Use a parameter of k = 5 and do this inplace\")```\r\n\r\n- Just ask the same prompt again. Since TopicGPT is a stochastic system, calling the same function with the same argument again might yield a different functionality to be used or a different result. \r\n\r\n- If this doesn't help, you can also directly call the function you want to use from the TopicPrompting class. In the example above you could do ```tm.topic_prompting.split_topic_kmeans(topic_idx = 13, n_clusters = 5, inplace = True)```. Note that all functions the model can call can also be called directly.\r\n\r\n- In case of hallucination of facts it may help to use GPT-4 for TopicGPT\r\n\r\n\r\n## How TopicGPT works\r\n\r\nTopicGPT is centrally built on top of text embeddings and the prompting mechanisms obtained via LLMs and provided by the OpenAI API. Please also see the section [References](#references) for more details on the models and ideas used in TopicGPT.\r\n\r\n### Embeddings\r\nWhen no embeddings are provided, TopicGPT automatically computes the embeddings of the documents of the provided corpus and also of the vocabulary that is extracted from the corpus. This happens after the fit-method is called. \r\n\r\nThe class ```GetEmbeddingsOpenAI``` is used for this purpose.\r\n\r\n### Clustering\r\nIn order to identify topics among the documents, TopicGPT reduces the dimensionality of the document embeddings via UMAP and then uses Hdbscan to identify the clusters. Dimensionality reduction is necessary since the document embeddings are of very high dimensionality and thus the curse of dimensionality would make it very difficult, if not impossible, to identify the clusters.\r\n\r\nWhen not specifying the number of topics in the ```Topic GPT``` class, Hdbscan is used to automatically determine the number of topics. If the number of topics is specified, agglomerative clustering is used on top of the clusters identified by HDBSCAN. \r\n\r\nThe class ```Clustering``` is used for this purpose.\r\n\r\n### Extraction of Top-Words\r\n\r\nAfter the clusters have been identified, TopicGPT extracts the top-words of each topic. This is done via two different methods:\r\n- **Tf-idf**: The tf-idf method is based on the idea that words that occur frequently in a topic but rarely in other topics are good indicators for the topic. The top-words are thus the words with the highest tf-idf scores. \r\n- **Centroid similarity**: The centroid similarity method is based on the idea that the words that are closest to the centroid of a topic are good indicators for the topic. The top-words are thus the words that are closest to the centroid of the topic.\r\n\r\nNote that the Tf-idf heuristic was introduced for the BerTopic Model (Grootendorst, Maarten. \"BERTopic: Neural topic modeling with a class-based TF-IDF procedure.\" arXiv preprint arXiv:2203.05794 (2022)) and a similar idea to the centroid similarity method is used in Top2Vec (Angelov, Dimo. \"Top2vec: Distributed representations of topics.\" arXiv preprint arXiv:2008.09470 (2020)).\r\n\r\nTopword extraction is performed with help of the class ```ExtractTopWords```.\r\n\r\n### Describing and naming topics\r\n\r\nIn the next step, all topics are provided with a short name and a description. This is done via prompting an LLM provided by OpenAI with around 500 top-words of each topic. The LLM then generates a short name and a description for each topic.\r\n\r\nThe class ```TopwordEnhancement``` is used for this purpose.\r\n\r\n\r\nNote that computation of Embeddings, Extraction of Top-Words and Describing and Naming Topics are all performed when calling the ```fit``` method of the ```TopicGPT``` class.\t\r\n\r\n### Prompting of TopicGPT\r\n\r\nWhen formalizing a prompt via the ```pprompt``` or ```prompt``` function, TopicGPT uses the following steps: \r\n\r\n1. The prompt, together with basic model- and corpus-information, is sent to an LLM provided by OpenAI. The LLM then decides which function of the ```TopicPrompting``` class to call. The LLM also decides which arguments to use for the function.\r\n2. The function is called according to the information by the LLM. The full result of the function will be returned to the user.\r\n3. Parts of the results of the function are returned to the LLM. The LLM then generates a short answer of the original prompt with help of the function result and returns it to the user.\r\n\r\n\r\n## References\r\n\r\nThe following models, software packages and ideas are central for TopicGPT: \r\n\r\n- **UMAP**: The Uniform Manifold Approximation and Projection for Dimension Reduction algorithm is used for reducing the dimensionality of document- and word embeddings (McInnes, Leland, John Healy, and James Melville. \"Umap: Uniform manifold approximation and projection for dimension reduction.\" arXiv preprint arXiv:1802.03426 (2018).)\r\n\r\n- **HDBSCAN**: Hierarchical density based clustering is used to identify the clusters among the dimensionality reduced topics (McInnes, Leland, John Healy, and Steve Astels. \"hdbscan: Hierarchical density based clustering.\" J. Open Source Softw. 2.11 (2017): 205.)\r\n\r\n- **Agglomerative Clustering**: The agglomerative clustering functionality from sklearn is used to combine topics in case the identified number of clusters exeeds the number of topics specified by the user (Pedregosa, Fabian, et al. \"Scikit-learn: Machine learning in Python.\" the Journal of machine Learning research 12 (2011): 2825-2830., https://scikit-learn.org/stable/modules/generated/sklearn.cluster.AgglomerativeClustering.html)\r\n\r\n- **Topword extraction**: Even though the corresponding packages are not directly used, the topword extraction methods used for this package are based on very similar ideas as found in the BerTopic Model (Grootendorst, Maarten. \"BERTopic: Neural topic modeling with a class-based TF-IDF procedure.\" arXiv preprint arXiv:2203.05794 (2022)) in the case of the tf-idf method and in Top2Vec for the centroid-similarity method (Angelov, Dimo. \"Top2vec: Distributed representations of topics.\" arXiv preprint arXiv:2008.09470 (2020)). \r\n\r\n- **LLMs from the GPT family**: Some references for the models for computing embeddings and answering the prompts include:\r\n - Brown, Tom B., et al. \u201cLanguage Models are Few-Shot Learners.\u201d Advances in Neural Information Processing Systems 33 (2020).\r\n - Radford, Alec, et al. \u201cGPT-4: Generative Pre-training of Transformers with Discrete Latent Variables.\u201d arXiv preprint arXiv:2302.07413 (2023).\r\n - Radford, Alec, et al. \u201cImproving Language Understanding by Generative Pre-Training.\u201d URL: https://s3-us-west-2.amazonaws.com/openai-assets/research-covers/language-unsupervised/language_understanding_paper.pdf. [6]\r\n - Radford, Alec, et al. \u201cLanguage Models are Unsupervised Multitask Learners.\u201d OpenAI Blog 1.8 (2019): 9. [7]\r\n",
"bugtrack_url": null,
"license": "MIT",
"summary": "A package for integrating LLMs like GPT-3.5 and GPT-4 into topic modelling",
"version": "1.0.0",
"project_urls": null,
"split_keywords": [
"topic modelling",
"gpt",
"llm",
"openai",
"retrieval augmented generation",
"chat-gpt",
"gpt-3",
"gpt-4"
],
"urls": [
{
"comment_text": "",
"digests": {
"blake2b_256": "12b5a3c1df04b714c26969f819451c0d47a1e77e7bb9f96e83cc6ff16b5d7141",
"md5": "26b51f5760543a9cfe7d4af9d56def02",
"sha256": "ead2eca9a77045c7d7e75117658123672253ff058eacad5542f979d79eab9b52"
},
"downloads": -1,
"filename": "topicgpt-1.0.0-py3-none-any.whl",
"has_sig": false,
"md5_digest": "26b51f5760543a9cfe7d4af9d56def02",
"packagetype": "bdist_wheel",
"python_version": "py3",
"requires_python": null,
"size": 42289,
"upload_time": "2023-09-08T10:00:02",
"upload_time_iso_8601": "2023-09-08T10:00:02.495815Z",
"url": "https://files.pythonhosted.org/packages/12/b5/a3c1df04b714c26969f819451c0d47a1e77e7bb9f96e83cc6ff16b5d7141/topicgpt-1.0.0-py3-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "132ce6b19ee0f0ad1e60d4ef94293391f1b7bdffcd66d11af09c1b9a12165bfe",
"md5": "5a3958a848ddd49eb188217b2cc94b33",
"sha256": "fe85138f228d0ae30be5cfee2b7a550a92855169a754ce3a1fdbea9024fd6726"
},
"downloads": -1,
"filename": "topicgpt-1.0.0.tar.gz",
"has_sig": false,
"md5_digest": "5a3958a848ddd49eb188217b2cc94b33",
"packagetype": "sdist",
"python_version": "source",
"requires_python": null,
"size": 51257,
"upload_time": "2023-09-08T10:00:04",
"upload_time_iso_8601": "2023-09-08T10:00:04.492409Z",
"url": "https://files.pythonhosted.org/packages/13/2c/e6b19ee0f0ad1e60d4ef94293391f1b7bdffcd66d11af09c1b9a12165bfe/topicgpt-1.0.0.tar.gz",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2023-09-08 10:00:04",
"github": false,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"lcname": "topicgpt"
}
JSON