BERT Reviews & Product Details

Its ability to capture contextual nuances in language is outstanding & allowing for more accurate and context-aware natural language understanding also. Its bidirectional approach and pre-training on extensive datasets contribute to its versatility across a spectrum of NLP tasks, making it a powerful tool in the field. Review collected by and hosted on G2.com.

Its computational intensity, requiring substantial resources for training and inference. Also, it struggle with out-of-vocabulary words and might not handle long-range dependencies as effectively. Despite these limitations, ongoing research and advancements aim to address and mitigate these challenges in future models. Review collected by and hosted on G2.com.

I have been using BERT for the last 3 months now, I give like precise and to-point answers to my daily activities, and as a chatbot, it gives completely relevant information like a mentor available for 24/7. I highly recommend this to everyone. I'm saving lots of time and effort using BERT. Review collected by and hosted on G2.com.

About the interface, as a Google product, it should look more classy. The information can be made more like human, as till now also it looks like machine generated. Review collected by and hosted on G2.com.

It's very easy to use and it have so many resources around it online that anyone get a very good grasp on it even without any background knowledge a out transformers.

Apart from ease of use it is also pretrained and we just need to fine tune as per our own task.

Also fine tuning is also pretty straightforward so yeah overall experience is really nice. Review collected by and hosted on G2.com.

There are only few things like it being computationally costly and like many other transformers it's mostly a black box when we try to understand why it gave out certain results.

Also since we are moving into age of AI, the token limitation in BERT actually makes its capabilities very limited. Review collected by and hosted on G2.com.

- Great for tasks where bidirectional context is required, as opposed to GPT models where the context is unidirectional. Suitable for question-answering, analyzing small paragraphs of words, etc.

- Output is more trustworthy as compared to GPT models.

- Open source

- Easy to fine-tune for domain-specific applications as long as enough data is available. Review collected by and hosted on G2.com.

- It is extremely computationally expensive to build and deploy, especially to produce a quality output.

- Balancing the context window takes a lot of trial and error.

- With the arrival of GPT models, the lack of long context, i.e., limited context, is more noticeable than ever.

- Not suitable for large documents which require broader context.

- (not limited to BERT) A bit of a black box once implemented.

- Not a good choice for tasks where text has to be generated. Review collected by and hosted on G2.com.

It is best situated for the random searches that we do on a search engine and have to go through multiple pages to build our understanding. But with the new BERT engine it has become so efficient to look for queries and questions also in terms of seeking other text information. Review collected by and hosted on G2.com.

Some of times the responses and like a general statement and we don't get exactly what we are looking Review collected by and hosted on G2.com.

An open source product by Google. Very easy to implement and work with. It is very flexible to customise for any specific tasks that is very helpful for a developer. It helps us in our day to day works with NLP. Review collected by and hosted on G2.com.

It takes a lot of time to train the model. Hence Computationaly costly and need high end machines. High memory consumption is also there. Review collected by and hosted on G2.com.

Regarding bert , It is the first model which i tried for context based embedding.

I best thing about BERT is that it is simple to understand and retraining ,finetunning of it and support about it is available.Also there are 3 to 4 generalised english model is available Review collected by and hosted on G2.com.

As compare to distillbert it is heavy in size and bulky in nature as same thing of BERT can be possible with distillbert. Review collected by and hosted on G2.com.

* BERT generates high-quality texts by understanding the context around a word. I found good performance on document retrieval, and Question Answering.

* Finetuning BERT on custom data (or transfer learning) is very simple and gives good results. BERT inference is also faster than GPT.

* BERT has an extensive community and good support. Almost everyone around me has used BERT. Review collected by and hosted on G2.com.

In my experience with BERT, I think it still needs improvements:

* I found that BERT fine-tuning does not work well with large-scale datasets (e.g PILE)

* Its domain knowledge is constricted. It does not know much about domains such as healthcare, and education.

Hence, BERT can be considered enough for simple tasks, however, for complex tasks (e.g. open-ended generation, language translation etc.), it needs improvement.

I trust it's newer version will accommodate for major fixes. Wish them luck, Review collected by and hosted on G2.com.

Worked on a use-case for detecting toxicity in prompts and their respective completions. BERT worked effectively providing us a very high accuracy of upto 92% for correct detections. Review collected by and hosted on G2.com.

Can try including more classes except for the 6: toxic, severe-toxic, obscene, threat, insult and identity_hate. Some useful recommended classes: gender_bias, ethnicity_bias etc. Review collected by and hosted on G2.com.

We are using BERT for personalized marketing campaigns to address the customer concern or questions about good and services, which improve customer services.

In order to send customer emails with content that is pertinent to their interests or to target customer with ads for products they are likely to be interest in. Review collected by and hosted on G2.com.

I don't think there are drawbacks of it but certainly at the first I encountered issues like the fact that BERT is a blackbox model which makes it difficult to always understand how it generate its predictions.

The cost of computing was high for both deployment and training, as a result it might be difficult to understand why BERT makes certain predicts.

Biases can be their as the large dataset is used to train BERT. Review collected by and hosted on G2.com.