What is the right approach for multi-label text information extraction/classification
Having texts that describe a caregiver/patient visit : (made-up example)
Mr *** visits the clinic on 02/2/2018 complaining about pain in the
lower back for several days, No pathological findings in the x-ray or
in the blood tests. I suggest Mr *** 5 resting days.
Now, that text can be even in a paragraph size where the only information I care about will be lower back pain and resting days. I have 300-400 different labels but the number of labeled samples can be around 1000-1500 (total) . When I label the text I also mark the relevant words that create the "label" ,here it will be ['pain','lower','back'].
When I just use look-up for those words (or the other 300-400 labels) in other texts I manage to label a larger amount of texts but if the words are written in different patterns such as Ache in the lower back or "lowerback pain" and I've never added that pattern to the look-up table of "lower back pain" I won't find it.
Due to the fact that I can have large paragraph but the only information I need is just 3-4 words, DL/ML models do not manage to learn with that amount of data and a high number of labels.I am wondering if there is a way to use the lookup table as a feature in the training phase or to try other approaches
Related
I have just started a project in NLP. Suppose I have a graph for each word that shows the polarity distribution of sentiments for that word in different sentences. I want to know what I can use to recognize the feelings of new words? Any other use you have in mind I will be happy to share.
I apologize for any possible errors in my writing. Thanks a lot
Assuming you've got some words that have been hand-labeled with positive/negative sentiments, but then you encounter some new words that aren't labeled:
If you encounter the new words totally alone, outside of contexts, there's not much you can do. (Maybe, you could go out to try to find extra texts with those new words, such as vis dictionaries or the web, then use those larger texts in the next approach.)
If you encounter the new words inside texts that also include some of your hand-labeled words, you could try guessing that the new words are most like the words you already know that are closest-to, or used-in-the-same-places. This would leverage what's called "the distributional hypothesis" – words with similar distributions have similar meanings – that underlies a lot of computer natural-language analysis, including word2vec.
One simple thing to try along these lines: across all your texts, for every unknown word U, tally up the counts all neighboring words within N positions. (N could be 1, or larger.) From that, pick the top 5 words occuring most often near the unknown word, and look up your prior labels, and avergae them together (perhaps weighted by the number of occurrences.)
You'll then have a number for the new word.
Alternatively, you could train a word2vec set-of-word-vectors for all of your texts, including the unknown & know words. Then, ask that model for the N most-similar neighbors to your unknown word. (Again, N could be small or large.) Then, from among those neighbors with known labels, average them together (again perhaps weighted by similarity), to get a number for the previously unknown word.
I wouldn't particularly expect either of these techniques to work very well. The idea that individual words can have specific sentiment is somewhat weak given the way that in actual language, their meaning is heavily modified, or even reversed, by the surrounding grammar/context. But in each case these simple calculate-from-neighbors techniqyes are probably better than random guesses.
If your real aim is to calculate the overall sentiment of longer texts, like sentences, paragraphs, reviews, etc, then you should discard your labels of individual words an acquire/create labels for full texts, and apply real text-classification techniques to those larger texts. A simple word-by-word approach won't do very well compared to other techniques – as long as those techniques have plenty of labeled training data.
I have a list of products names and a collection of text generated from random users. I am trying to detect products mentioned in the text while talking into account spelling variation. For example the text
Text = i am interested in galxy s8
Mentions the product samsung galaxy s8
But note the difference in spellings.
I've implemented the following approaches:
1- max tokenized products names and users text (i split words by punctuation and digits so s8 will be tokenized into 's' and '8'. Then i did a check on each token in user's text to see if it is in my vocabulary with damerau levenshtein distance <= 1 to allow for variation in spelling. Once i have detected a sequence of tokens that do exist in the vocabulary i do a search for the product that matches the query while checking the damerau levenshtein distance on each token. This gave poor results. Mainly because the sequence of tokens that exist in the vocabulary do not necessarily represent a product. For example since text is max tokenized numbers can be found in the vocabulary and as such dates are detected as products.
2- i constructed bigram and trigram indicies from the list of products and converted each user text into a query.. but also results weren't so great given the spelling variation
3- i manually labeled 270 sentences and trained a named entity recognizer with labels ('O' and 'Product'). I split the data into 80% training and 20% test. Note that I didn't use the list of products as part of the features. Results were okay.. not great tho
None of the above results achieved a reliable performance. I tried regular expressions but since there are so many different combinations to consider it became too complicated.. Are there better ways to tackle this problem? I suppose ner could give better results if i train more data but suppose there isn't enough training data, what do u think a better solution would be?
If i come up with a better alternative to the ones I've already mentioned, I'll add it to this post. In the meantime I'm open to suggestions
Consider splitting your problem into two parts.
1) Conduct a spelling check using a dictionary of known product names (this is not a NLP task and there should be guides on how to impelement spell check).
2) Once you have done pre-processing (spell checking), use your NER algorithm
It should improve your accuracy.
Suppose you have a set of transcribed customer service calls between customers and human agents, where on average each call's length is 7 minutes. Customers will mostly call because of issues they have with the product. Let's assume that a human can assign one label per axis per call:
Axis 1: What was the problem from the customer's perspective?
Axis 2: What was the problem from the agent's perspective?
Axis 3: Could the agent resolve the customer's issue?
Based on the manually labeled texts you want to train a text classifier that shall predict a label for each call for each of the three axes. But the labeling of recordings takes time and costs money. On the other hand you need a certain amount of training data to get good prediction results.
Given the above assumptions, how many manually labeled training texts would you start with? And how do you know that you need more labeled training texts?
Maybe you've worked on a similar task before and can give some advice.
UPDATE (2018-01-19): There's no right or wrong answer to my question. Ok, ideally, somebody worked on exactly the same task, but that's very unlikely. I'll leave the question open for one more week and then accept the best answer.
This would be tricky to answer but I will try my best based on my experience.
In the past, I have performed text classification on 3 datasets; the number in the bracket indicates how big my dataset was: restaurant reviews (50K sentences), reddit comments (250k sentences) and developer comments from issue tracking systems (10k sentences). Each of them had multiple labels as well.
In each of the three cases, including the one with 10k sentences, I achieved an F1 score of more than 80%. I am stressing on this dataset specifically because I was told by some that the size is less for this dataset.
So, in your case, assuming you have atleast 1000 instances (calls that include conversation between customer and agent) of average 7 minute calls, this should be a decent start. If the results are not satisfying, you have the following options:
1) Use different models (MNB, Random Forest, Decision Tree, and so on in addition to whatever you are using)
2) If point 1 gives more or less similar results, check the ratio of instances of all the classes you have (the 3 axis you are talking about here). If they do not share a good ratio, get more data or try out the different balancing techniques if you cannot get more data.
3) Another way would be to classify them at a sentence level than message or conversation level to generate more data and individual labels for sentences rather than message or the conversation itself.
I seek the most effective and simple way to classify 800k+ scholarly articles as either relevant (1) or irrelevant (0) in relation to a defined conceptual space (here: learning as it relates to work).
Data is: title & abstract (mean=1300 characters)
Any approaches may be used or even combined, including supervised machine learning and/or by establishing features that give rise to some threshold values for inclusion, among other.
Approaches could draw on the key terms that describe the conceptual space, though simple frequency count alone is too unreliable. Potential avenues might involve latent semantic analysis, n-grams, ..
Generating training data may be realistic for up to 1% of the corpus, though this already means manually coding 8,000 articles (1=relevant, 0=irrelevant), would that be enough?
Specific ideas and some brief reasoning are much appreciated so I can make an informed decision on how to proceed. Many thanks!
Several Ideas:
Run LDA and get document-topic and topic-word distributions say (20 topics depending on your dataset coverage of different topics). Assign the top r% of the documents with highest relevant topic as relevant and low nr% as non-relevant. Then train a classifier over those labelled documents.
Just use bag of words and retrieve top r nearest negihbours to your query (your conceptual space) as relevant and borrom nr percent as not relevant and train a classifier over them.
If you had the citations you could run label propagation over the network graph by labelling very few papers.
Don't forget to make the title words different from your abstract words by changing the title words to title_word1 so that any classifier can put more weights on them.
Cluster the articles into say 100 clusters and then choose then manually label those clusters. Choose 100 based on the coverage of different topics in your corpus. You can also use hierarchical clustering for this.
If it is the case that the number of relevant documents is way less than non-relevant ones, then the best way to go is to find the nearest neighbours to your conceptual space (e.g. using information retrieval implemented in Lucene). Then you can manually go down in your ranked results until you feel the documents are not relevant anymore.
Most of these methods are Bootstrapping or Weakly Supervised approaches for text classification, about which you can more literature.
I'm doing a small research project where I should try to split financial news articles headers to positive and negative classes.For classification I'm using SVM approach.The main problem which I see now it that not a lot of features can be produced for ML. News articles contains a lot of Named Entities and other "garbage" elements (from my point of view of course).
Could you please suggest ML features which can be used for ML training? Current results are: precision =0.6, recall=0.8
Thanks
The task is not trivial at all.
The straightforward approach would be to find or create a training set. That is a set of headers with positive news and a set of headers with negative news.
You turn the training set to a TF/IDF representation and then you train a Linear SVM to separate the two classes. Depending on the quality and size of your training set you can achieve something decent - not sure for 0.7 break even point.
Then, to get better results you need to go for NLP approaches. Try use a part-of-speech tagger to identify adjectives (trivial), and then score them using some sentiment DB like SentiWordNet.
There is an excellent overview on Sentiment Analysis by Bo Pang and Lillian Lee you should read:
How about these features?
Length of article header in words
Average word length
Number of words in a dictionary of "bad" words, e.g. dictionary = {terrible, horrible, downturn, bankruptcy, ...}. You may have to generate this dictionary yourself.
Ratio of words in that dictionary to total words in sentence
Similar to 3, but number of words in a "good" dictionary of words, e.g. dictionary = {boon, booming, employment, ...}
Similar to 5, but use the "good"-word dictionary
Time of the article's publication
Date of the article's publication
The medium through which it was published (you'll have to do some subjective classification)
A count of certain punctuation marks, such as the exclamation point
If you're allowed access to the actual article, you could use surface features from the actual article, such as its total length and perhaps even the number of responses or the level of opposition to that article. You could also look at many other dictionaries online such as Ogden's 850 basic english dictionary, and see if bad/good articles would be likely to extract many words from those. I agree that it seems difficult to come up with a long list (e.g. 100 features) of useful features for this purpose.
iliasfl is right, this is not a straightforward task.
I would use a bag of words approach but use a POS tagger first to tag each word in the headline. Then you could remove all of the named entities - which as you rightly point out don't affect the sentiment. Other words should appear frequently enough (if your dataset is big enough) to cancel themselves out from being polarised as either positive or negative.
One step further along, if you still aren't close could be to only select the adjectives and verbs from the tagged data as they are the words that tend to convey the emotion or mood.
I wouldn't be too disheartened in your precision and recall figures though, an F number of 0.8 and above is actually quite good.