I have tried my hands on many NER tools (OpenNLP, Stanford NER, LingPipe, Dbpedia Spotlight etc).
But what has constantly evaded me is a gazetteer/dictionary based NER system where my free text is matched with a list of pre-defined entity names, and potential matches are returned.
This way I could have various lists like PERSON, ORGANIZATION etc. I could dynamically change the lists and get different extractions. This would tremendously decrease training time (since most of them are based on maximum entropy model so they generally includes tagging a large dataset, training the model etc).
I have built a very crude gazetteer based NER system using a OpenNLP POS tagger, from which I used to take out all the Proper nouns (NP) and then look them up in a Lucene index created from my lists. This however gives me a lot of false positives. For ex. if my lucene index has "Samsung Electronics" and my POS tagger gives me "Electronics" as a NP, my approach would return me "Samsung Electronics" since I am doing partial matches.
I have also read people talking about using gazetteer as a feature in CRF algorithms. But I never could understand this approach.
Can any of you guide me towards a clear and solid approach that builds NER on gazetteer and dictionaries?
I'll try to make the use of gazetteers more clear, as I suspect this is what you are looking for. Whatever training algorithm used (CRF, maxent, etc.) they take into account features, which are most of the time:
tokens
part of speech
capitalization
gazetteers
(and much more)
Actually gazetteers features provide the model with intermediary information that the training step will take into account, without explicitly being dependent on the list of NEs present in the training corpora. Let's say you have a gazetteer about sport teams, once the model is trained you can expand the list as much as you want without training the model again. The model will consider any listed sport team as... a sport team, whatever its name.
In practice:
Use any NER or ML-based framework
Decide what gazetteers are useful (this is maybe the most crucial part)
Affect to each gazetteer a relevant tag (e.g. sportteams, companies, cities, monuments, etc.)
Populate gazetteers with large lists of NEs
Make your model take into account those gazetteers as features
Train a model on a relevant corpus (it should containing many NEs from gazetteers)
Update your list as much as you want
Hope this helps!
You can try this minimal bash Named-Entity Recognizer:
https://github.com/lasigeBioTM/MER
Demo: http://labs.fc.ul.pt/mer/
Related
I have trained a doc2vec (PV-DM) model in gensim on documents which fall into a few classes. I am working in a non-linguistic setting where both the number of documents and the number of unique words are small (~100 documents, ~100 words) for practical reasons. Each document has perhaps 10k tokens. My goal is to show that the doc2vec embeddings are more predictive of document class than simpler statistics and to explain which words (or perhaps word sequences, etc.) in each document are indicative of class.
I have good performance of a (cross-validated) classifier trained on the embeddings compared to one compared on the other statistic, but I am still unsure of how to connect the results of the classifier to any features of a given document. Is there a standard way to do this? My first inclination was to simply pass the co-learned word embeddings through the document classifier in order to see which words inhabited which classifier-partitioned regions of the embedding space. The document classes output on word embeddings are very consistent across cross validation splits, which is encouraging, although I don't know how to turn these effective labels into a statement to the effect of "Document X got label Y because of such and such properties of words A, B and C in the document".
Another idea is to look at similarities between word vectors and document vectors. The ordering of similar word vectors is pretty stable across random seeds and hyperparameters, but the output of this sort of labeling does not correspond at all to the output from the previous method.
Thanks for help in advance.
Edit: Here are some clarifying points. The tokens in the "documents" are ordered, and they are measured from a discrete-valued process whose states, I suspect, get their "meaning" from context in the sequence, much like words. There are only a handful of classes, usually between 3 and 5. The documents are given unique tags and the classes are not used for learning the embedding. The embeddings have rather dimension, always < 100, which are learned over many epochs, since I am only worried about overfitting when the classifier is learned, not the embeddings. For now, I'm using a multinomial logistic regressor for classification, but I'm not married to it. On that note, I've also tried using the normalized regressor coefficients as vector in the embedding space to which I can compare words, documents, etc.
That's a very small dataset (100 docs) and vocabulary (100 words) compared to much published work of Doc2Vec, which has usually used tens-of-thousands or millions of distinct documents.
That each doc is thousands of words and you're using PV-DM mode that mixes both doc-to-word and word-to-word contexts for training helps a bit. I'd still expect you might need to use a smaller-than-defualt dimensionaity (vector_size<<100), & more training epochs - but if it does seem to be working for you, great.
You don't mention how many classes you have, nor what classifier algorithm you're using, nor whether known classes are being mixed into the (often unsupervised) Doc2Vec training mode.
If you're only using known classes as the doc-tags, and your "a few" classes is, say, only 3, then to some extent you only have 3 unique "documents", which you're training on in fragments. Using only "a few" unique doctags might be prematurely hiding variety on the data that could be useful to a downstream classifier.
On the other hand, if you're giving each doc a unique ID - the original 'Paragraph Vectors' paper approach, and then you're feeding those to a downstream classifier, that can be OK alone, but may also benefit from adding the known-classes as extra tags, in addition to the per-doc IDs. (And perhaps if you have many classes, those may be OK as the only doc-tags. It can be worth comparing each approach.)
I haven't seen specific work on making Doc2Vec models explainable, other than the observation that when you are using a mode which co-trains both doc- and word- vectors, the doc-vectors & word-vectors have the same sort of useful similarities/neighborhoods/orientations as word-vectors alone tend to have.
You could simply try creating synthetic documents, or tampering with real documents' words via targeted removal/addition of candidate words, or blended mixes of documents with strong/correct classifier predictions, to see how much that changes either (a) their doc-vector, & the nearest other doc-vectors or class-vectors; or (b) the predictions/relative-confidences of any downstream classifier.
(A wishlist feature for Doc2Vec for a while has been to synthesize a pseudo-document from a doc-vector. See this issue for details, including a link to one partial implementation. While the mere ranked list of such words would be nonsense in natural language, it might give doc-vectors a certain "vividness".)
Whn you're not using real natural language, some useful things to keep in mind:
if your 'texts' are really unordered bags-of-tokens, then window may not really be an interesting parameter. Setting it to a very-large number can make sense (to essentially put all words in each others' windows), but may not be practical/appropriate given your large docs. Or, trying PV-DBOW instead - potentially even mixing known-classes & word-tokens in either tags or words.
the default ns_exponent=0.75 is inherited from word2vec & natural-language corpora, & at least one research paper (linked from the class documentation) suggests that for other applications, especially recommender systems, very different values may help.
I have a collection of over 1 million bodies of text. Within those bodies are multiple entities whose names mimic common stop words and phrases.
This has created issues when tokenizing the data, as there are ~50 entities with the same problem. To counteract this, I've disabled the removal of the matched stop words before their removal. This is fine, but Ideally I'd have a way to differentiate when a token is actually meant to be a stop word vs an entity, since I only care for when it's used as an entity.
Here's a sample excerpt:
A determined somebody slept. Prior to this, A could never be comfortable with the idea of responsibility. It was foreign, something heard about through a story passed down by words of U. As slow as it could be, A began to find meaning in the words of a story.
A and U are entities/nouns in most of their usages here. POS tagging so far has only labelled A as a determiner, and NER either won't tag any instances of the word. Adding the target tags to the NER list will result in every instance being tagged as an entity, which is not the case.
So far I've primarily used the Stanford POS Tagger and SpaCY for NER.
I think you should try to train your own NER model.
You can do this in three steps, as follows:
label a number of documents in your corpus.
You can do this using the spacy-annotator.
train your spacy NER model from scratch.
You can follow the instructions in the spacy docs.
Use the trained model to predict entities in your corpus.
By labelling a good amount of entities at step 1, the model will learn to differentiate between a determiner and an entity.
I would need some help because I don't know what algorithm i could use for the following (I use python) :
Steve is 25 and he buys everyday orange juice
Maria is 23 and she likes to buy smoothies
Steve & Maria tastes are pretty much the same.
Juan is 16 and he only drinks sodas
Juan tastes are not the same as Steve and Maria.
====================================================
I would like to use a matching algorithm that will detect the users who have the same drink preference and a close age. To continue with the example, Steve and Maria would be matched together but not Juan. Which one should I use ?
I agree with #klutt that your task is pretty vague. There are two approaches that come to mind, but not knowing more details about your problem does limit the details I can provide in my answer that would help you. I am interpreting the question as if you are taking in raw text and might want to process more sentences that have very similar semantic and syntactical structure.
An algorithmic approach:
Assuming that your word choices are static in their semantic meaning (Maria is 23 ... Steve is 25), we can parse each sentence and identify tokens like is or and or same and essentially perform lexical analysis on the text... from here, you could continue thinking about how you would go about matching and so forth... but this is rather complicated...
Neural Network approach:
If you are taking in raw text in the form of sentences, it's a problem that's not straight forward to solve using a top-down algorithmic approach.
You could take an approach with neural networks that trains a model to solve your problem, but then again what you seem to be asking is quite complex since there are multiple "facts" within each sentence that are not semantically related. For example, your second sentence identifies that Maria is 23 but at the end of that sentence there is a comparison between Steve and Maria. And your first sentence only identifies Steve as 25.
Even if you chunk raw text into sentences, you would have to have a very fine tuned neural network architecture and a lot of training data to get remotely close to your goal.
Now, both of those solutions are very complex... but if you wanted to create an application that collects this data (via a form or prompt) and puts it into a structured format (like a json or xml object) to organize and store the data in memory (perhaps writing out to a database or file for persistent storage), that might be a good route to go down.
This can serve as a good lesson in how to think about data as well. It is one thing if you have a pool of thousands of sentences, just raw data that you need to organize for quantitative purposes (classic qualitative -> quantitative problems). It is another thing if you are going to be collecting this data. If you are going to be collecting data, having a program that collects and organizes names, ages, and drink preferences (and then organizes that data within certain data structures), then we can talk about matching algorithms.
I will also add here that if you do have structured data, Collaborative filtering (mentioned by Shridhar) is a great starting place.
Collaborative filtering best suits your needs.
In the newer, narrower sense, collaborative filtering is a method of
making automatic predictions (filtering) about the interests of a user
by collecting preferences or taste information from many users
(collaborating). The underlying assumption of the collaborative
filtering approach is that if a person A has the same opinion as a
person B on an issue, A is more likely to have B's opinion on a
different issue than that of a randomly chosen person. For example, a
collaborative filtering recommendation system for television tastes
could make predictions about which television show a user should like
given a partial list of that user's tastes (likes or dislikes).[3]
Note that these predictions are specific to the user, but use
information gleaned from many users. This differs from the simpler
approach of giving an average (non-specific) score for each item of
interest, for example based on its number of votes.
What is the amount of sentences needed to effectively train the CRF for a domain like restauarants (restaurant names, addresses, cusines) or music (artist name, song name genre).
As a point of reference, I believe the CoNLL training data for (location, organization, person, misc) NER has around 14,000 sentences.
It depends a lot of the kind of data you will be tagging and how variable it will be. I've worked on a project also involving restaurant and music domains. In my case we would be handling user queries, which tend to be short and don't present that much variability (particularly for restaurant, but not for music, which is a very noisy domain).
For the restaurant domain, training it with ~2k sentences was fine, but of course, if you can get more data, your model will be much more accurate.
For music, the situation is a little bit more tricky since song/band names can be virtually anything. In this case, only data alone might not be enough to get an acceptable accuracy. In my project we used ~5k for music and many features and some additional post-processing to get things right.
I have got a new task(not traditional) from my client, It is something about machine learning.
As I have never been to "machine learning" except some little Data Mining stuff so I need your help.
My task is to Classify a product present on any Shopping Site, on the basis of gender(whom the product belongs to),agegroup etc, the training data we can have is the product's Title, Keywords(available in the html of the product page), and product description.
I did a lot of R&D , I found Image Recog APIs(cloudsight,vufind) that returned the details of the product image but that did not full fill the need, used google suggestqueries, searched out many machine learning algorithms and finally...
I came to know about the "Decision Tree Learning Algorithm" but cannot figure out, how it is applicable to my problem.
I tried out the "PlayingTennis" dataset but couldn't make the sense what to do.
Can you give me some direction that from where to start this journey? Should I focus on The Decision Tree Learning algorithm or Is there any other algorithm you would suggest I should focus on to categorize the products on the basis of context?
If you say , I would share in detail about what things I searched about to solve my problem.
I would suggest to do the following:
Go through items in your dataset and classify them manually (decide for which gender each item is). Store each decision so that you would be able to somehow link each item in an original dataset with a target class.
Develop an algorithm for converting each item from your dataset into a feature vector. This algorithm should be able to convert each item in your original dataset in a vector of numbers (more about how to do it later).
Convert all your dataset with appropriate classes into a dataset that would look like this:
Feature_1, Feature_2, Feature_3, ..., Gender
value_1, value_2, value_3, ... male
It would be a good decision to store it in CSV file since you would be able to load it and process in different machine learning tools (More about those later).
Load dataset you've created at step 3 in machine learning tool of your choice and try to come up with the best model that can classify items in your dataset by gender.
Store model created at step 4. It will be part of your production system.
Develop a production code that can convert an unclassified product, create feature vector out of it and pass this feature vector to the model you've saved at step 5. The result of this operation should be a predicted gender.
Details
If there too many items (say tens of thousands) in your original dataset it may be impractical to classify them yourself. What you can do is to use Amazon Mechanical Turk to simplify your task. If you are unable to use it (the last time I've checked you had to have a USA address to use it) you can just classify few hundreds of items to start working on your model and classify the rest to improve accuracy of your classification (the more training data you use the better the accuracy, but up to a certain point)
How to extract features from a dataset
If keyword has form like tag=true/false, it's a boolean feature.
If keyword has form like tag=42, it's a numerical one or ordinal. For example it can be price value or price range (0-10, 10-50, 50-100, etc.)
If keyword has form like tag=string_value you can convert it into a categorical value
A class (gender) is simply boolean value 0/1
You can experiment a bit with how you extract your features, since it may influence the result accuracy.
How to extract features from product description
There are different ways to convert a text into a feature vector. Look for TF-IDF algorithms or something similar.
Machine learning tools
You can use one of existing machine learning libraries and hack some code that loads your CSV dataset, trains a model and checks the accuracy, but at first I would suggest to use something like Weka. It has more or less intuitive UI and you can quickly start to experiment with different machine learning algorithms, convert different features in your dataset from string to categories, or from real values to ordinal values, etc. Good thing about Weka is that it has Java API, so you can automate all the process of data conversion, train models programmatically, etc.
What algorithms to choose
I would suggest to use decision tree algorithms like C4.5. It's fast and show good results on wide range of machine learning tasks. Additionally you can use ensemble of classifiers. There are various algorithms that can combine several algorithms like (google for boosting or random forest to find out more) usually they give better results, but work more slowly (since you need to run a single feature vector through several algorithms.
One another trick that you can use to make your algorithm more accurate is to use models that work on different sets of features (say one algorithm uses features extracted from tags and another algorithm uses data extracted from product description). You can then combine them using algorithms like stacking to come up with a final result.
For classification on the basis of features extracted from text, you can try to use Naive Bayes algorithm or SVM. They both show good results in text classification.
Do consider Support Vector Classifier (SVC), or for Google's sake the Support Vector Machine (SVM). If You have a large training set (which I suspect) search for implementations that are "fast" or "scalable".