5 ms·
Sequence Tagging with Tensorflow
- gajju3588 9y agoNice article. Explains the whole things really well. I have found LSTM is little overkill for NER, as your entities are limited and data will not be very huge. CRF and Max-Ent seems to better options. 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%20NER%20dataset https://dataturks.com/projects/Mohan/Best%20Buy%20E-commerce...
- Lindrian 9y agoDon'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.
- melzarei 9y agoDisclaimer: 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.
- bravura 9y ago"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.
- pacala 9y agoOne can do semi-supervised training with LSTMs, using http://allennlp.org/elmo http://allennlp.org/elmo or similar. That may use huge amounts of data, for which MaxEnt or CRF are not sufficient.
- mark_l_watson 9y agoI had thought of doing the same thing but POS tagging is already “solved” in some sense by OpenNlp and the Stanford NLP libraries. I think of using deep learning for problems that don’t already have good solutions. Nice paper, and I look forward to reading the example code.
- bravura 9y agoThis does named-entity recognition tagging, which is harder. Also POS results are still not that good on out-of-vocabulary words, so using a word representation that includes sub-word ngrams (like FB Fasttext) can improve results on OOV words.
- imh 9y agoI may be misunderstanding things. Are pronouns not tagged as people? "Obama was the president" vs "I was the president" or "He was the president"
- Lindrian 9y agoTypically not. Sometimes they are tagged as nominal mentions.