I would like to apply fine-tuning Bert to calculate semantic similarity between sentences.
I search a lot websites, but I almost not found downstream about this.
I just found STS benchmark.
I wonder if I can use STS benchmark dataset to train a fine-tuning bert model, and apply it to my task.
Is it reasonable?
As I know, there are a lot method to calculate similarity including cosine similarity, pearson correlation, manhattan distance, etc.
How choose for semantic similarity?
In addition, if you're after a binary verdict (yes/no for 'semantically similar'), BERT was actually benchmarked on this task, using the MRPC (Microsoft Research Paraphrase Corpus).
The google github repo https://github.com/google-research/bert includes some example calls for this, see --task_name=MRPC in section Sentence (and sentence-pair) classification tasks.
As a general remark ahead, I want to stress that this kind of question might not be considered on-topic on Stackoverflow, see How to ask. There are, however, related sites that might be better for these kinds of questions (no code, theoretical PoV), namely AI Stackexchange, or Cross Validated.
If you look at a rather popular paper in the field by Mueller and Thyagarajan, which is concerned with learning sentence similarity on LSTMs, they use a closely related dataset (the SICK dataset), which is also hosted by the SemEval competition, and ran alongside the STS benchmark in 2014.
Either one of those should be a reasonable set to fine-tune on, but STS has run over multiple years, so the amount of available training data might be larger.
As a great primer on the topic, I can also highly recommend the Medium article by Adrien Sieg (see here, which comes with an accompanied GitHub reference.
For semantic similarity, I would estimate that you are better of with fine-tuning (or training) a neural network, as most classical similarity measures you mentioned have a more prominent focus on the token similarity (and thus, syntactic similarity, although not even that necessarily). Semantic meaning, on the other hand, can sometimes differ wildly on a single word (maybe a negation, or the swapped sentence position of two words), which is difficult to interpret or evaluate with static methods.
Related
Currently, I'm working on a project where I need to extract the relevant aspects used in positive and negative reviews in real time.
For the notions of more negative and positive, it will be a question of contextualizing the word. Distinguish between a word that sounds positive in a negative context (consider irony).
Here is an example:
Very nice welcome!!! We ate very well with traditional dishes as at home, the quality but also the quantity are in appointment!!!*
Positive aspects: welcome, traditional dishes, quality, quantity
Can anyone suggest to me some tutorials, papers or ideas about this topic?
Thank you in advance.
This task is called Aspect Based Sentiment Analysis (ABSA). Most popular is the format and dataset specified in the 2014 Semantic Evaluation Workshop (Task 5) and its updated versions in the following years.
Overview of model efficiencies over the years:
https://paperswithcode.com/sota/aspect-based-sentiment-analysis-on-semeval
Good source for ressources and repositories on the topic (some are very advanced but there are some more starter friendly ressources in there too):
https://github.com/ZhengZixiang/ABSAPapers
Just from my general experience in this topic a very powerful starting point that doesn't require advanced knowledge in machine learning model design is to prepare a Dataset (such as the one provided for the SemEval2014 Task) that is in a Token Classification Format and use it to fine-tune a pretrained transformer model such as BERT, RoBERTa or similar. Check out any tutorial on how to do fine-tuning on a token classification model like this one in huggingface. They usually use the popular task of Named Entity Recognition (NER) as the example task but for the ABSA-Task you basically do the same thing but with other labels and a different dataset.
Obviously an even easier approach would be to take more rule-based approaches or combine a rule-based approach with a trained sentiment analysis model/negation detection etc., but I think generally with a rule-based approach you can expect a much inferior performance compared to using state-of-the-art models as transformers.
If you want to go even more advanced than just fine-tuning the pretrained transformer models then check out the second and third link I provided and look at some of the machine learning model designs specifically designed for Aspect Based Sentiment Analysis.
Does anyone know how of an accurate tool or method that can be used to compute word embeddings or find similarity among domain-specific words? I'm working on an NLP project that involves computing cosine similarity between technical terms, such as "address" and "socket", but pre-trained models like word2vec aren't giving useful embeddings or accurate cosine similarities because they aren't specific to technical terms. Since the more general-nontechnical meanings of "address" and "socket" aren't similar to one another, these pretrained models aren't giving them sufficiently high similarity scores for the purposes of my project. Would appreciate any advice people would be able to offer. Thank you!
With sufficient data from your specific domain, you can train your own word2vec model - whose resulting word-vectors, being only influenced by your domain data, will be far more reflective of the in-domain meanings.
Similarly, if you have a mixture of data where you have hints that some word uses are for different senses of a polysemous word, you could try preprocessing your text, using those hints, replacing the ambiguous tokens (like say 'address') with a larger number of distinct tokens (like 'address*networking', 'address*delivery', etc). Even with a lot of error in such a process, its results might be sufficient for a specific purpose.
For example, maybe you'd assume all docs of a certain type – like articles from a particular publication – always mean 'address*networking' when they write 'address'. That crude replacement, on just some subset of docs sufficient to collect enough varied examples of 'address*networking' usage, might leave you with a good-enough word-vector for 'address*networking'.
(More generally, deciding which word sense of multiple candidates is meant by a particular word is called "word sense disambiguation", and it might be possible to use other preexisting code for performing that to help preprocess texts - replacing ambiguous tokens with more-speciific stand-ins – before performing word2vec training.)
Even without such assistive pre-processing, there've been a number of research attempts to extend word2vec to better model words with multiple contrasting meanings. Googling for [word2vec polysemy] or [polysemous embeddings] should turn up a bunch of examples.
But I don't know any of those techniques that have become widely-used, or that are explicitly supported by major word2vec libraries, so I can't specifically recommend or show working code for any. I don't know a standard best-practice or off-the-shelf solution – you'd have to treat adopting those ideas from research papers as an R&D project, performing a lot of your own implementation/evaluation to see if any help with your goals.
I want to train a Doc2Vec model with a generic corpus and, then, continue training with a domain-specific corpus (I have read that is a common strategy and I want to test results).
I have all the documents, so I can build and tag the vocab at the beginning.
As I understand, I should train initially all the epochs with the generic docs, and then repeat the epochs with the ad hoc docs. But, this way, I cannot place all the docs in a corpus iterator and call train() once (as it is recommended everywhere).
So, after building the global vocab, I have created two iterators, the first one for the generic docs and the second one for the ad hoc docs, and called train() twice.
Is it the best way or it is a more appropriate way?
If the best, how I should manage alpha and min_alpha? Is it a good decision not to mention them in the train() calls and let the train() manage them?
Best
Alberto
This is probably not a wise strategy, because:
the Python Gensim Doc2Vec class hasn't ever properly supported expanding its known vocabulary after a 1st single build_vocab() call. (Up through at least 3.8.3, such attempts typically cause a Segmentation Fault process crash.) Thus if there are words that are only in your domain-corpus, an initial typical initialization/training on the generic-corpus would leave them out of the model entirely. (You could work around this, with some atypical extra steps, but the other concerns below would remain.)
if there is truly an important contrast between the words/word-senses used in your generic and the different words/word-senses used in your domain corpus, influence of the words from the generic corpus may not be beneficial, diluting domain-relevant meanings
further, any followup training that just uses a subset of all documents (the domain corpus) will only be updating the vectors for that subset of words/word-senses, and the model's internal weights used for further unseen-document inference, in directions that make sense for the domain-corpus alone. Such later-trained vectors may be nudged arbitrarily far out of comparable alignment with other words not appearing in the domain-corpus, and earlier-trained vectors will find themselves no longer tuned in relation to the model's later-updated internal-weights. (Exactly how far will depend on the learning-rate alpha & epochs choices in the followup training, and how well that followup training optimizes model loss.)
If your domain dataset is sufficient, or can be grown with more domain data, it may not be necessary to mix in other training steps/data. But if you think you must try that, the best-grounded approach would be to shuffle all training data together, and train in one session where all words are known from the beginning, and all training examples are presented in balanced, interleaved fashion. (Or possibly, where some training texts considered extra-important are oversampled, but still mixed in with the variety of all available documents, in all epochs.)
If you see an authoritative source suggesting such a "train with one dataset, then another disjoint dataset" approach with the Doc2Vec algorithms, you should press them for more details on what they did to make that work: exact code steps, and the evaluations which showed an improvement. (It's not impossible that there's some way to manage all the issues! But I've seen many vague impressions that this separate-pretraining is straightforward or beneficial, and zero actual working writeups with code and evaluation metrics showing that it's working.)
Update with respect to the additional clarifications you provided at https://stackoverflow.com/a/64865886/130288:
Even with that context, my recommendation remains: don't do this segmenting of training into two batches. It's almost certain to degrade the model compared to a combined training.
I would be interested to see links to the "references in the literature" you allude to. They may be confused or talking about algorithms other than the Doc2Vec ("Paragraph Vectors") algorithm.
If there is any reason to give your domain docs more weight, a better-grounded way would be to oversample them in the combined corpus.
Bu by all means, test all these variants & publish the relative results. If you're exploring shaky hypotheses, I would ignore any advice from StackOverflow-like sources & just run all the variants that your reading of the literature suggest, to see which, if any actually help.
You're right to recognized that the choice of alpha parameters is a murky area that could majorly influence what impact such add-on training has. There's no right answer, so you'll have to search-for and reason-out what might make sense. The inherent issues I've mentioned with such subset-followup-training could make it so that even if you find benefits in some combos, they may be more a product of a lucky combination of data & arbitrary parameters than a generalizable practice.
And: your specific question "if it is better to set such values or not provide them at all" reduces to: "do you want to use the default values, or values set when the model was created, or not?"
Which values might be workable, if at all, for this unproven technique is something that'd need to be experimentally discovered. That is, if you wanted to have comparable (or publishable) results here, I think you'd have to justify from your own novel work some specific strategy for choosing good alpha/epochs and other parameters, rather than adopt any practice merely recommended in a StackOverflow answer.
I am solving a binary text classification problem with corporate filings. Using Doc2Vec embeddings of length 100 with LightGBM is producing great results. However, for this project it would be very valuable to approximate a thematic meaning for at least one of the components. Ideally, this would be a feature ranked with high importance by LightGBM explained anecdotally with a few examples.
Has anyone attempted this, or should interpretation be off the table for a high-dimensional model with this level of complexity?
The individual dimensions of a Doc2Vec representation should not be considered independent, interpretable features. They're only useful in concert with each other, and the exact directions aligned with individual coordinate-axes may not be strongly meaningful in any human-describable sense.
However, neighborhoods of the space may loosely fit describable themes, and certain directions (not specifically parallel with coordinate-axes) may loosely fit semantic themes.
But to characterize those, you might try to find the centroid points of groups-of-related-documents, or discovered clusters, and compare the relative distances/directions between those centroids.
I want to classify text messages into several categories like, "relation building", "coordination", "information sharing", "knowledge sharing" & "conflict resolution". I am using NLTK library to process these data. I would like to know which classifier, in nltk, is better for this particular multi-class classification problem.
I am planning to use Naive Bayes Classification, is it advisable?
Naive Bayes is the simplest and easy to understand classifier and for that reason it's nice to use. Decision Trees with a beam search to find the best classification are not significantly harder to understand and are usually a bit better. MaxEnt and SVM tend be more complex, and SVM requires some tuning to get right.
Most important is the choice of features + the amount/quality of data you provide!
With your problem, I would focus first on ensuring you have a good training/testing dataset and also choose good features. Since you are asking this question you haven't had much experience with machine learning for NLP, so I'd say start of easy with Naive Bayes as it doesn't use complex features- you can just tokenize and count word occurrences.
EDIT:
The question How do you find the subject of a sentence? and my answer are also worth looking at.
Yes, Training a Naive Bayes Classifier for each category and then labeling each message to a class based on which Classifier provides the highest score is a standard first approach to problems like this. There are more sophisticated single class classifier algorithms which you could substitute in for Naive Bayes if you find performance inadequate, such as a Support Vector Machine ( Which I believe is available in NLTK via a Weka plug in, but not positive). Unless you can think of anything specific in this problem domain that would make Naieve Bayes especially unsuitable, its ofen the go-to "first try" for a lot of projects.
The other NLTK classifier I would consider trying would be MaxEnt as I believe it natively handles multiclass classification. (Though the multiple binary classifer approach is very standard and common as well). In any case the most important thing is to collect a very large corpus of properly tagged text messages.
If by "Text Messages" you are referring to actual cell phone text messages these tend to be very short and the language is very informal and varied, I think feature selection may end up being a larger factor in determining accuracy than classifier choice for you. For example, using a Stemmer or Lemmatizer that understands common abbreviations and idioms used, tagging part of speech or chunking , entity extraction, extracting probably relationships between terms may provide more bang than using more complex classifiers.
This paper talks about classifying Facebook status messages based on sentiment, which has some of the same issues, and may provide some insights into this. The links is to a google cache because I'm having problems w/ the original site:
http://docs.google.com/viewer?a=v&q=cache:_AeBYp6i1ooJ:nlp.stanford.edu/courses/cs224n/2010/reports/ssoriajr-kanej.pdf+maxent+classifier+multiple+classes&hl=en&gl=us&pid=bl&srcid=ADGEESi-eZHTZCQPo7AlcnaFdUws9nSN1P6X0BVmHjtlpKYGQnj7dtyHmXLSONa9Q9ziAQjliJnR8yD1Z-0WIpOjcmYbWO2zcB6z4RzkIhYI_Dfzx2WqU4jy2Le4wrEQv0yZp_QZyHQN&sig=AHIEtbQN4J_XciVhVI60oyrPb4164u681w&pli=1