The name finder has the concept of "adaptive data" in the feature generation. The feature generators can remember things from previous sentences and use it to generate features based on it. Usually that can help with the recognition rate if you have names that are repeated. You can tweak this to your data, or just pass in the entire document.
Jörn On Fri, Aug 26, 2016 at 3:25 PM, Damiano Porta <damianopo...@gmail.com> wrote: > Hi! > Yes I can train a good model (sure It will takes a lot of time), i have 30k > resumes. So the "data" isnt a problem. > I thought about many things, i am also creating a custom features > generator, with dictionary too (for names) and regex for Birthday, then > the machine learning will look at their contexts. > So now i need to separate the sentences to create a custom model. > At this point i will not try with one per line CV. > > Il 26/Ago/2016 15:10, "Russ, Daniel (NIH/CIT) [E]" <dr...@mail.nih.gov> ha > scritto: > > Hi Damiano, > I am not sure that the NameFinder will be effective as-is for you. Do > you have training data (and I mean a lot of training data)? You need to > consider what feature are useful in your case. You might consider a > feature such as line number on the page (since people tend to put their > name on the top or second line), maybe the font-size. You can add a > dictionary of common names and have a feature “inDictionary”. You will have > to use your domain knowledge to help you here. > > For birthday you may want to consider using regex to pick out dates. > Then look at the context around the date (words before/after, remove > graduated or if another date just before) or maybe years before present > year (if you are looking at resumes, you probably won’t find any 5 year > olds or 200 year olds. > > Daniel Russ, Ph.D. > Staff Scientist, Office of Intramural Research > Center for Information Technology > National Institutes of Health > U.S. Department of Health and Human Services > 12 South Drive > Bethesda, MD 20892-5624 > > On Aug 26, 2016, at 5:57 AM, Damiano Porta <damianopo...@gmail.com<mailto: > damianopo...@gmail.com>> wrote: > > Hi Daniel! > > Thank you so much for your opinion. > It makes perfectly sense. But i am still a bit confused about the length of > the sentences. > In a resume there are many names, dates etc etc. So my doubt is regarding > the structure of the sentences because they follow specific patterns > sometimes. > > For example i need to extract the personal name, (Who wrote the resume) the > Birthday etc etc. > > As You know there are many names and dates inside a resume so i thought > about to write the entire resume as sentence to also train the "position" > less or more of the entities. If i "decompose" all the resume into > sentences i will lose this information. No? > > Damiano > > Il 25/Ago/2016 16:26, "Russ, Daniel (NIH/CIT) [E]" <dr...@mail.nih.gov > <mailto:dr...@mail.nih.gov>> ha > scritto: > > Hi Damiano, > > Everyone can feel feel to correct my ignorance but I view the the > name finder as follows. > > I look at it as walking down the sentence and classifying words as > “NOT IN NAME” until I hit the start of a name than it is “START NAME”, > Followed by “STILL IN NAME” until “NOT IN NAME”. Take the sentence “Did > John eat the stew”. Starting with the first word in the sentence decide > what are the odds that the first word starts a name (given that it is the > first word happens to be “Did” in a sentence, with a capital but not all > caps) starts a person’s name. Then go to then next word in the sentence. > If the first word was not in a name, what are the odds that the second word > starts a name (given that the previous word did not start a name, the word > starts with a capital (but not all capital), the word is John, and the > previous word is “Did”). If it decides that we are starting a name at > “John”, we are now looking for the end. What are the odds that “eat” is > part of the name given that [“Did”: was not part of the name, was > capitalized] and that [“John”: was the first word in the name, was > capitalized]. You are essentially classifying [Did <- OTHER] [John > <-START] [eat<-OTHER] [the<-OTHER] [stew<-OTHER]. If it was “Did John > Smith eat the stew”. You would have [Did <- OTHER] [John > <-START][Smith<-IN] [eat<-OTHER] [the<-OTHER] [stew<-OTHER]. There are > other features other than just word, previous word, and the shape (first > letter capitalized, all letters capitalized). I think the name finder uses > part of speech also. > > > So you see that it is not a name lookup table, but dependent on the > previous classification of words earlier in the sentence. Therefore, you > must have sentences. Does that help? > Daniel > > > Daniel Russ, Ph.D. > Staff Scientist, Office of Intramural Research > Center for Information Technology > National Institutes of Health > U.S. Department of Health and Human Services > 12 South Drive > Bethesda, MD 20892-5624 > > On Aug 25, 2016, at 9:55 AM, Damiano Porta <damianopo...@gmail.com<mailto: > damianopo...@gmail.com><mailto: > damianopo...@gmail.com<mailto:damianopo...@gmail.com>>> wrote: > > Hello everybody! > > Could someone explain why should I separate each sentence of my documents > to train my models? > My documents are like resume/cv and the sentences can be very different. > For example a sentence could also be : > > 1. Name: John > 2. Surname: travolta > > Etc etc > So my question is. What is the problem if i train ny models > (namefinder,tokenizer) with the complete resume/cv one per line? > > Could It be a problem? > In this case when i will like to tokenize the resume and doing the NER i > will simply pass the complete resume text skiping the "sentences detection" > process. > > Thanks for your opinion in advance! > > Best > Damiano >
