Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

BERTopic: The Future of Topic Modeling

pinecone.io

Post reply on HN
Print view View original

21–29 of 29 posts

Re: BERTopic: The Future of Topic Modeling

#21

Post by chaxor »

a-dub wrote: post #16
uoaei wrote: post #14

Earlier quoted context omitted.

Is that not exactly what's happening in TFA?

if TFA means "the forememtioned article", i don't think so. i'm not convinced that the clusters found and the frequencies in those clusters would be the same as what LDA computes with gibbs sampling or the variational calculus method would find. but i must admit it's been a while since i've played with this stuff. if TFA is some other method, i am unfamiliar and would like to know more. in my experience, while it's t…

I think he might mean term frequency analysis?

Re: BERTopic: The Future of Topic Modeling

#22

Post by chaxor »

chaxor wrote: post #21
a-dub wrote: post #16

Earlier quoted context omitted.

if TFA means "the forememtioned article", i don't think so. i'm not convinced that the clusters found and the frequencies in those clusters would be the same as what LDA computes with gibbs sampling or the variational calculus method would find. but i must admit it's been a while since i've played with this stuff. if TFA is some other method, i am unfamiliar and would like to know more. in my experience, while it's t…

I think he might mean term frequency analysis?

Nevermind. I should have read what was said.

Re: BERTopic: The Future of Topic Modeling

#23

Post by lmeyerov »

victorianoi wrote: post #3

take a look at Graphext ( https://www.graphext.com ) it automatically creates the clustering embeddings using BERT for you + great visualization libraries to interpret the clusters :D it took us 5 years to build the product

Latest pygraphistry has this flow for free and OSS, just `pip install graphistry[umap-learn]` or, for transformers, `pip install graphistry[ai]` :)

And per the article, with pluggable sentence transformers -> UMAP automatically as part of the auto-featurization: graphistry.nodes(accounts_df).umap().plot() :)

We haven't published tutorials yet, just been using with some fraud/cyber/misinfo/genomics/sales/gov/etc teams (including with RAPIDS gpu accel), so cool to see excitement here already! Till then, it should work out-of-the-box with no parameters, and then all sorts of fun things to tune: https://github.com/graphistry/pygraphistry/blob/21fad42412cc...

Re: BERTopic: The Future of Topic Modeling

#24

Post by jll29 »

vosper wrote: post #13
visarga wrote: post #11

Next step: automatically naming clusters using few-shot GPT-3. Cluster naming is a non-trivial problem.

I've run into this problem at a previous employer. Do you know if anyone's working on it?

"topic labeling" papers: https://scholar.google.com/scholar?hl=en&as_sdt=0%2C5&q=%22t...

Re: BERTopic: The Future of Topic Modeling

#25

Post by ahoho »

Color me skeptical on BERTopic. Without human validation, I'm not convinced that it's an improvement over existing methods.

I'm an author on a recent paper about automated topic model evaluation [1], and we found that current metrics do not line up with human judgements as well as previously thought. To my knowledge, BERTopic has only been evaluated on these automated metrics.

For datasets of under a few hundred thousand documents, Mallet (LDA estimated with Gibbs sampling) can produce stable, high-quality outputs in minutes on a laptop [2]. Even larger datasets remain tractable, although depending on your use case you may be better off subsampling.

It's possible that I've missed something, but I'm not clear on what benefits BERTopic has that existing methods do not. I don't mean to be overly negative---it has a nice API and the approach seems reasonable---I'm just wondering what's really new here.

[1]: https://proceedings.neurips.cc/paper/2021/hash/0f83556a305d7... [2]: https://mimno.github.io/Mallet/ [3]: https://maartengr.github.io/BERTopic/faq.html#why-are-the-re...

Re: BERTopic: The Future of Topic Modeling

#26

Post by uoaei »

a-dub wrote: post #16
uoaei wrote: post #14

Earlier quoted context omitted.

Is that not exactly what's happening in TFA?

if TFA means "the forememtioned article", i don't think so. i'm not convinced that the clusters found and the frequencies in those clusters would be the same as what LDA computes with gibbs sampling or the variational calculus method would find. but i must admit it's been a while since i've played with this stuff. if TFA is some other method, i am unfamiliar and would like to know more. in my experience, while it's t…

Well, they're not supposed to be the same clusters. The reason people develop new methods is to surpass the old ones.

I'm just saying that the method described in the link seems to be exactly what you are describing: using document embedding vectors as input to soft clustering mechanisms akin to LDA. Of course it does not interface perfectly with the theoretical underpinnings of LDA because those are quite constrained to tf-idf (generally count-based) inputs.

As an aside, "TFA" translates to "the fucking article" and is a reference to the classic Internet acronym "RTFM" standing for "read the fucking manual". Both are passive-aggressive-cum-colloquial ways to imply that answers are in places you would expect to find them, if only you go to read the source.

Re: BERTopic: The Future of Topic Modeling

#27

Post by victorianoi »

m1sta_ wrote: post #17
victorianoi wrote: post #3

take a look at Graphext ( https://www.graphext.com ) it automatically creates the clustering embeddings using BERT for you + great visualization libraries to interpret the clusters :D it took us 5 years to build the product

I'm not going to contact sales. I have no problem paying though. Part of your value proposition is saving people time, but your sales model is time expensive.

You are totally right! We are actually making the pricing public next week. We will keep the free version plus starting at $30/month for individual accounts and $150 for team accounts ;)

Re: BERTopic: The Future of Topic Modeling

#28

Post by jasonys »

victorianoi wrote: post #3

take a look at Graphext ( https://www.graphext.com ) it automatically creates the clustering embeddings using BERT for you + great visualization libraries to interpret the clusters :D it took us 5 years to build the product

I think you should open source the core part of the HTML visualization algos and if ppl like it, they may consider paying for the premium version. I don’t feel like ppl want to move their analysis workflow to yet another platform without trying it enough in their existing workflows (e.g Jupyter/Colab/Databricks notebooks)

Re: BERTopic: The Future of Topic Modeling

#29

Post by a-dub »

uoaei wrote: post #26
a-dub wrote: post #16

Earlier quoted context omitted.

if TFA means "the forememtioned article", i don't think so. i'm not convinced that the clusters found and the frequencies in those clusters would be the same as what LDA computes with gibbs sampling or the variational calculus method would find. but i must admit it's been a while since i've played with this stuff. if TFA is some other method, i am unfamiliar and would like to know more. in my experience, while it's t…

Well, they're not supposed to be the same clusters. The reason people develop new methods is to surpass the old ones. I'm just saying that the method described in the link seems to be exactly what you are describing: using document embedding vectors as input to soft clustering mechanisms akin to LDA. Of course it does not interface perfectly with the theoretical underpinnings of LDA because those are quite constraine…

i'm pretty sure the method mentioned in the article finds single topic assignments where LDA finds mixtures of topics.

hierarchical in LDA refers to the stacked multinomial nature of the model over word counts, documents and topics.

hierarchical in bertopic means assuming and finding a hierarchical relationship between the topics themselves at cluster time.

they use the same word, but appear very different things, at least to me.

Post reply on HN