Use nearest neighbors to predict text classification with fasttext - deep-learning

I may misunderstood how fasttext/deep-learning work for classification, I would like to take in consideration nearest neighbors to predict labels. The aim of my work is to predict label with synonyms.
I train a big dataset with fasttext:
fasttext supervised -input data/spam_status.txt -output models/sem -lr 1.0 -wordNgrams 1 -epoch 25
Where spam_status.txt use a regexp to labelize message containing the word "skype":
__label__skype i dont have skype __NUMBER__ sorry
__label__skype skype
__label__skype si ta un skype si
__label__skype i will give u my skype
__label__skype pv ici no skype
__label__skype skype
And plenty of other messages, with other labels, or "ok" if nothing is found.
Nearest neighbors of "skype" are (with fasttext nn models/sem.bin):
email
viber
emaill
skp
This is excellent, fasttext give me good similar words. But if I ask a prediction:
fasttext predict-prob ./models/sem.bin -
donne moi ton skype
__label__skype 1.00001
donne moi ton viber
__label__ok 1.00001
donne moi ton emaill
__label__ok 1.00001
Why NN is not taken in consideration here?

Because you trained the model with examples where ONLY messages with the word "skype" have the label Skype. Therefore messages with words like "email" and "Viber" are labelled "ok."
Your first pass taught you that you should re-label. Using a regex to label data is always going to cause problems like this. You can now at least re-label any of the messages with "email" or "Viber" as "__label__skype" so it will learn that pattern. However, it will probably not get you anything better than just using a regex as a classifier, because the model will learn the pattern going in: if it has one of the words from this short list, label it "Skype" otherwise label it "ok".
Your get better results by sounding a few hours manually labelling data instead of using a regex.

Related

Model suggestion: Keyword spotting

I want to predict the occurrences of the word "repeat" in a speech as well as the word's approximate duration. For this task, I'm planning to build a Deep Learning model. I've around 50 positive as well as 50 negative utterances (I couldn't collect more).
Initially I've searched for any pretrained models for keyword spotting, but I couldn't get a good one.
Then I tried Speech Recognition models (Deep Speech), but it couldn't predict the exact repeat words as my data follows Indian accent. Also, I've thought that going for ASR models for this task would be a over-killing one.
Now, I've split the entire audio into chunk of 1 secs with 50% overlapping and tried a binary audio classification in each chunk that is whether the chunk has the word "repeat" or not. For building the classification model, I calculated the MFCC features and build a sequence model on the top of it. Nothing seems to work for me.
If anyone already worked with this kind of task, please provide me with a correct method/resources to build a DL model for this task. Thanks in advance!

Data PreProcessing for BERT (base-german)

I am working on a sentiment analysis solution with BERT to analyze tweets in german. My training dataset of is a class of 1000 tweets, which have been manually annotated into the classes neutral, positive and negative.
The dataset with 10.000 tweets is quite unevenly distributed:
approx.
3000 positive
2000 negative
5000 neutral
the tweets contain formulations with #names, https links, numbers, punctuation marks, smileys like :3 :D :) etc..
The interesting thing is, if I remove them with the following code during Data Cleaning, the F1 score gets worse. Only the removal of https links (if I do it alone) leads to a small improvement.
# removing the punctuation and numbers
def remove_punct(text):
text = re.sub(r'http\S+', '', text) # removing links
text = re.sub(r'#\S+', '', text) # removing referencing on usernames with #
text = re.sub(r':\S+', '', text) # removing smileys with : (like :),:D,:( etc)
text = "".join([char for char in text if char not in string.punctuation])
text = re.sub('[0-9]+', '', text)
return text
data['Tweet_clean'] = data['Tweet'].apply(lambda x: remove_punct(x)) # extending the dataset with the column tweet_clean
data.head(40)
also steps like stop words removal or lemmitazation lead more to a deterioration. Is this because I do something wrong or can the model BERT actually handle such values?
A second question is:
I found other records that were also manually annotated, but these are not tweets and the structure of the sentences and language use is different. Would you still recommend to add these records to my original?
There are about 3000 records in German.
My last question:
Should I reduce the class sizes to the size of the smallest unit and thus balance?
BERT can handle punctuation, smileys etc. Of course, smileys contribute a lot to sentiment analysis. So, don't remove them. Next, it would be fair to replace #mentions and links with some special tokens, because the model will probably never see them again in the future.
If your model is designed for tweets, I suggest that you fine-tune BERT with additional corpus, and after fine-tune with Twitter corpus. Or do it simultaneously. More training samples is generally better.
No, it is better to use class weights instead of downsampling.
Based on this paper (By Adam Ek, Jean-Philippe Bernardy and Stergios Chatzikyriakidis), BERT models outperform BiLSTM in terms of better generalizing to punctuations. Looking at the experiments' results in the paper, I say keep the punctuations.
I couln't find anything solid for smiley faces; However, after doing some experiments with the HuggingFace API, I didn't notice much difference with/without smiley faces.

How is the self-attention mechanism in Transformers able to learn how the words are related to each other?

Given the sentence The animal didn't cross the street because it was too tired, how the self-attention is able to map with a higher score the word aninal intead of the word street ?
I'm wondering if that might be a consequence of the word embedding vectors fed into the network, that some how already encapsulate some degree of distance among the words.
Word Embeddings are first added to Positional Encoding which adds information about the word's position in the sequence. Then through each Encoder stack(6 to be precise), the Embeddings undergo multiple transformations and are refined to form better representations before they are passed on to the decoder.
The modification to the Embeddings as it passes through the Encoder Stack is learnable. Sometimes it may appear that some Attention-Heads at the top Stack are doing something that may look like coreference resolution which you pointed out in your example. Attending more to the word "animal" simply results in better representation than attending to "street".
How do we know which representations are better? The one that minimizes the loss or produces a better output of course!

wordnet on different text?

I am new to nltk, and I find wordnet functionality pretty useful. It gives synsets, hypernyms, similarity, etc. But however it fails to give similarity between locations like 'Delhi'-'Hyderabad' obviously as these words are not in the wordnet corpus.
So, I would like to know, if somehow I can update the wordnet corpus OR create wordnet over a different corpus e.g. Set of pages extracted from wikipedia related to travel? If at all we can create wordnet over different corpus, then what would be the format, steps to do the same, any limitations?
Please can you point me to links that describe the above concerns. I have searched the internet, googled, read portions of nltk book, but I don't have a single hint to above question.
Pardon me, if the question sounds completely ridiculous.
For flexibility in measuring the semantic similarity of very specific terms like Dehli or Hyderabad, what you want is not something hand-crafted like WordNet, but an automatically-learned similarity measure from a very large database. These are statistical similarity approaches. Of course, you want to avoid having to train such a model on data yourself...
Thus one thing that may be useful is the Google Distance (wikipedia, original paper). It seems fairly simple to implement such a measure in a language like R (code), and the original paper reports 87% agreement with WordNet.
The similarity measures in Wordnet work as expected because Wordnet measures semantic similarity. In that sense, both are cities, so they are very similar. What you are looking for is probably called geographic similarity.
delhi = wn.synsets('Delhi', 'n')[0]
print delhi.definition()
# a city in north central India
hyderabad = wn.synsets('Hyderabad', 'n')[0]
print hyderabad.definition()
# a city in southern Pakistan on the Indus River
delhi.wup_similarity(hyderabad)
# 0.9
melon = wn.synsets('melon', 'n')[0]
delhi.wup_similarity(melon)
# 0.3
There is a Wordnet extension, called Geowordnet. I kind of had the same problem as you at one point and tried to unify Wordnet with some of its extensions: wnext. Hope that helps.

Best method to train Tesseract 3.02

i'm wondering what is the best method to train Tesseract (kind of text/TIFF and so on) for a particular kind of documents, with these particularities:
the structure and main text of the documents is always the same
the only things that change are 5 alphanumeric codes (THIS ARE THE REAL IMPORTANT THING TO DETECT!)
Some of thes codes are bold
At the moment I used standard trained datas, I detect the entire text and I extrapolate the codes with some regular expressions.
It's okay, but I've got errors sometimes, for example:
0 / O
L / I / 1
Please someone knowns some "tricks" to improve precision?
Thanks!
during training part of Tesseract, you have to make a file manually to give to the engine in order to specify ambiguous characters.
For more information look at the "unicharambigs" part of the Tesseract documentation.
Best Regards.