Don't agree. LSTM is a very good option for NER. I have personally experimented with simple FFNN and language models to accommodate variable width input and achieved very good results. I will be publishing a paper on this soon.
The system in the article fails at quite simple sentences, such as:
"Rex Tillerson is the secretary of state."
It tags "Rex Tillerson" as ORG instead of PER. If you change "is" to "was" it tags it as PER.
Disclaimer: as you may notice, the tagger is far from being perfect. Some errors are due to the fact that the demo uses a reduced vocabulary (lighter for the API). But not all. It is also very sensible to capital letters, which comes both from the architecture of the model and the training data. For more information about the demo, see here.
"I have found LSTM is little overkill for NER, as your entities are limited and data will not be very huge."
Incorrect. Inducing word representations on a huge corpus, like Common Crawl, can be more effective than manually building a small Gazeeter of named entities.
One can do semi-supervised training with LSTMs, using http://allennlp.org/elmo or similar. That may use huge amounts of data, for which MaxEnt or CRF are not sufficient.
People who want to try it out on some real dataset, can get a sample data for NER from here :
https://dataturks.com/projects/Mohan/Best%20Buy%20E-commerce...