My question relates directly to the text summarization task. I know that there are a bunch of implementations with RNN networks (especially LSTMs) that use sentence-level attention to extract salient sentences of the source, using an attentive LSTM decoder. I have been digging into it to see if this is possible with the Transformer's networks, specifically the Transformer's decoder part, but do not really have an idea how to get this incorporated.
Look, for example, in terms of LSTM decoder, the LSTM encoder can produce contextualized encodings for the sentences that are in the source, then the last hidden state is passed to the LSTM decoder, and at each decoding timestep, the decoder attends to the source sentences (encoded by the encoder) to get an attention score over those. Finally, these scores and the sentence hidden states are combined to form the context vector, which is further processed with the decoder hidden state to predict the right sentence to pick up.
Assume that I have obtained the sentence encodings using Transformer's encoder, I'm just wondering how I can relate the scenario happening in LSTM network's decoder part to the Transformer's network decoder side.
Also, a question, how are these networks (both LSTM and Transformer's) that use sentence-level attention instead of word-level attention trained?
Update: the behaviour that I intended to achieve is as follows: I want the Transformer's decoder gets in sentences (instead of tokens which is then regarded as abstractive summarization), compute attention on the source sentences considering the partial summary that has been selected in prior timesteps, and then give a probability distribution over the source sentences denoting how much it is probable that a sentence is being copied into the target. So to make it explicit, I'm looking for an extractive summarizer with decoder.
Related
Imagine that you have a corpus in which some lines have just one word, so there is no context around some of the words. In this situation how does Fasttext perform to provide embeddings for these single words? Note that the frequency of some of these words are one and there is no cut-off to get rid of them.
There's no way to train a context_word -> target_word skip-gram pair for such words (in either 'context' or 'target' roles), so such words can't receive trained representations. Only texts with at least 2 tokens contribute anything to word2vec or FastText word-vector training.
(One possible exception: FastText in its 'supervised classification' mode might be able to make use of, and train vectors for, such words, because then even single words can be used to predict the known-label of training texts.)
I suspect that such corpuses will still result in the model counting the word in its initial vocabulary-discovery scan, and thus it will be allocated a vector (if it appears at least min_count times), and that vector will receive the usual small-random-vector initialization. But the word-vector will receive no further training – so when you request the vector back after training, it will be of low-quality, with the only meaningful contributions coming from any char n-grams shared with other words that received real training.
You should consider any text-breaking process that results in single-word texts as buggy for the purposes of FastText. If those single-word texts come from another meaningful context where they were once surrounded by other contextual words, you should change your text-breaking process to work in larger chunks that retain that context.
Also note: it's rare for min_count=1 to be a good idea for word-vector models, at least when the training text is real natural-language material where word-token frequencies roughly follow Zipf's law. There will be many, many 1-occurrence (or few-occurrence) words, but with just one to a few example usage contexts, not likely representing the true breadth and subtleties of that word's real usages, it's nearly impossible for such words to receive good vectors that generalize to other uses of those same words elsewhere.
Training good vectors require a variety of usage examples, and just one or a few examples will practically be "noise" compared to the tens-to-hundreds of examples of other words' usage. So keeping these rare words, instead of dropping them like a default min_count=5 (or higher in larger corpuses) would do, tends to slow training, slow convergence ("settling") of the model, and lower the quality of the other more-frequent word vectors at the end – due to the significant-but-largely-futile efforts of the algorithm to helpfully position these many rare words.
First of all, I know this question is kind of off-topic, but I have already tried to ask elsewhere but got no response.
Adding a UNK token to the vocabulary is a conventional way to handle oov words in tasks of NLP. It is totally understandable to have it for encoding, but what's the point to have it for decoding? I mean you would never expect your decoder to generate a UNK token during prediction, right?
Depending on how you preprocess your training data, you might need the UNK during training. Even if you use BPE or other subword segmentation, OOV can appear in the training data, usually some weird UTF-8 stuff, fragments of alphabets, you are not interested in at all, etc.
For example, if you take WMT training data for English-German translation, do BPE and take the vocabulary, you vocabulary will contain thousands of Chinese characters that occur exactly once in the training data. Even if you keep them in the vocabulary, the model has no chance to learn anything about them, not even to copy them. It makes sense to represent them as UNKs.
Of course, what you usually do at the inference time is that you prevent the model predict UNK tokens, UNK is always incorrect.
I have used it one time in the following situation:
I had a preprocessed word2vec(glove.6b.50d.txt) and I was outputting an embedded vector, in order to transform it into a word I used cosine similarity based on all vectors in the word2vec if the most similar vector was the I would output it.
Maybe I'm just guessing it here, but what I think might happen under the hoods is that it predicts based on previous words(e.g. it predicts the word that appeared 3 iterations ago) and if that word is the neural net outputs it.
I have a question about CBOW prediction. Suppose my job is to use 3 surrounding words w(t-3), w(t-2), w(t-1)as input to predict one target word w(t). Once the model is trained and I want to predict a missing word after a sentence. Does this model only work for a sentence with four words which the first three are known and the last is unknown? If I have a sentence in 10 words. The first nine words are known, can I use 9 words as input to predict the last missing word in that sentence?
Word2vec CBOW mode typically uses symmetric windows around a target word. But it simply averages the (current in-training) word-vectors for all words in the window to find the 'inputs' for the prediction neural-network. Thus, it is tolerant of asymmetric windows – if there are fewer words are available on either side, fewer words on that side are used (and perhaps even zero on that side, for words at the front/end of a text).
Additionally, during each training example, it doesn't always use the maximum-window specified, but some random-sized window up-to the specified size. So for window=5, it will sometimes use just 1 on either side, and other times 2, 3, 4, or 5. This is done to effectively overweight closer words.
Finally and most importantly for your question, word2vec doesn't really do a full-prediction during training of "what exact word does the model say should be heat this target location?" In either the 'hierarchical softmax' or 'negative-sampling' variants, such an exact prediction can be expensive, requiring calculations of neural-network output-node activation levels proportionate to the size of the full corpus vocabulary.
Instead, it does the much-smaller number-of-calculations required to see how strongly the neural-network is predicting the actual target word observed in the training data, perhaps in contrast to a few other words. In hierarchical-softmax, this involves calculating output nodes for a short encoding of the one target word – ignoring all other output nodes encoding other words. In negative-sampling, this involves calculating the one distinct output node for the target word, plus a few output nodes for other randomly-chosen words (the 'negative' examples).
In neither case does training know if this target word is being predicted in preference over all other words – because it's not taking the time to evaluate all others words. It just looks at the current strength-of-outputs for a real example's target word, and nudges them (via back-propagation) to be slightly stronger.
The end result of this process is the word-vectors that are usefully-arranged for other purposes, where similar words are close to each other, and even certain relative directions and magnitudes also seem to match human judgements of words' relationships.
But the final word-vectors, and model-state, might still be just mediocre at predicting missing words from texts – because it was only ever nudged to be better on individual examples. You could theoretically compare a model's predictions for every possible target word, and thus force-create a sort of ranked-list of predicted-words – but that's more expensive than anything needed for training, and prediction of words like that isn't the usual downstream application of sets of word-vectors. So indeed most word2vec libraries don't even include any interface methods for doing full target-word prediction. (For example, the original word2vec.c from Google doesn't.)
A few versions ago, the Python gensim library added an experimental method for prediction, [predict_output_word()][1]. It only works for negative-sampling mode, and it doesn't quite handle window-word-weighting the same way as is done in training. You could give it a try, but don't be surprised if the results aren't impressive. As noted above, making actual predictions of words isn't the usual real goal of word2vec-training. (Other more stateful text-analysis, even just large co-occurrence tables, might do better at that. But they might not force word-vectors into interesting constellations like word2vec.)
In the traditional "one-hot" representation of words as vectors you have a vector of the same dimension as the cardinality of your vocabulary. To reduce dimensionality usually stopwords are removed, as well as applying stemming, lemmatizing, etc. to normalize the features you want to perform some NLP task on.
I'm having trouble understanding whether/how to preprocess text to be embedded (e.g. word2vec). My goal is to use these word embeddings as features for a NN to classify texts into topic A, not topic A, and then perform event extraction on them on documents of topic A (using a second NN).
My first instinct is to preprocess removing stopwords, lemmatizing stemming, etc. But as I learn about NN a bit more I realize that applied to natural language, the CBOW and skip-gram models would in fact require the whole set of words to be present --to be able to predict a word from context one would need to know the actual context, not a reduced form of the context after normalizing... right?). The actual sequence of POS tags seems to be key for a human-feeling prediction of words.
I've found some guidance online but I'm still curious to know what the community here thinks:
Are there any recent commonly accepted best practices regarding punctuation, stemming, lemmatizing, stopwords, numbers, lowercase etc?
If so, what are they? Is it better in general to process as little as possible, or more on the heavier side to normalize the text? Is there a trade-off?
My thoughts:
It is better to remove punctuation (but e.g. in Spanish don't remove the accents because the do convey contextual information), change written numbers to numeric, do not lowercase everything (useful for entity extraction), no stemming, no lemmatizing.
Does this sound right?
I've been working on this problem myself for some time. I totally agree with the other answers, that it really depends on your problem and you must match your input to the output that you expect.
I found that for certain tasks like sentiment analysis it's OK to remove lot's of nuances by preprocessing, but e.g. for text generation, it is quite essential to keep everything.
I'm currently working on generating Latin text and therefore I need to keep quite a lot of structure in the data.
I found a very interesting paper doing some analysis on that topic, but it covers only a small area. However, it might give you some more hints:
On the Role of Text Preprocessing in Neural Network Architectures: An Evaluation Study on Text Categorization and Sentiment Analysis
by Jose Camacho-Collados and Mohammad Taher Pilehvar
https://arxiv.org/pdf/1707.01780.pdf
Here is a quote from their conclusion:
"Our evaluation highlights the importance of being consistent in the preprocessing strategy employed across training and evaluation data. In general a simple tokenized corpus works equally or better than more complex preprocessing techniques such as lemmatization or multiword grouping, except for a dataset corresponding to a specialized domain, like health, in which sole tokenization performs poorly. Addi- tionally, word embeddings trained on multiword- grouped corpora perform surprisingly well when applied to simple tokenized datasets."
So many questions. The answer to all of them is probably "depends". It needs to be considered the classes you are trying to predict and the kind of documents you have. It's not the same to try to predict authorship (then you definitely need to keep all kinds of punctuation and case so stylometry will work) than sentiment analysis (where you can get rid of almost everything but have to pay special attention to things like negations).
I would say apply the same preprocessing to both ends. The surface forms are your link so you can't normalise in different ways. I do agree with the point Joseph Valls makes, but my impression is that most embeddings are trained in a generic rather than a specific manner. What I mean is that the Google News embeddings perform quite well on various different tasks and I don't think they had some fancy preprocessing. Getting enough data tends to be more important. All that being said -- it still depends :-)
What is Conditional Random Field?
How does exactly Conditional Random Field identify proper names as person, organization, or place in a structured or unstructured text?
For example: This product is ordered by StackOverFlow Inc.
What does Conditional Random Field do to identify StackOverFlow Inc. as an organization?
A CRF is a discriminative, batch, tagging model, in the same general family as a Maximum Entropy Markov model.
A full explanation is book-length.
A short explanation is as follows:
Humans annotate 200-500K words of text, marking the entities.
Humans select a set of features that they hope indicate entities. Things like capitalization, or whether the word was seen in the training set with a tag.
A training procedure counts all the occurrences of the features.
The meat of the CRF algorithm search the space of all possible models that fit the counts to find a pretty good one.
At runtime, a decoder (probably a Viterbi decoder) looks at a sentence and decides what tag to assign to each word.
The hard parts of this are feature selection and the search algorithm in step 4.
Well to understand that you got to study a lot of things.
For start
Understand the basic of markov and bayesian networks.
Online course available in coursera by daphne coller
https://class.coursera.org/pgm/lecture/index
CRF is a special type of markov network where we have observation and hidden states.
The objective is to find the best State Assignment to the unobserved variables also known as MAP problem.
Be Prepared for a lot of probability and Optimization. :-)