I've tried to use pre-trained model to create paraphrases of sentences,
Even after trying to correlate top_p and top_k the result paraphrases were almost the same as the original sentence.
I would like to get results that look completely different (eg. third person talking about this topic or to point an event that illustrates this topic.)
Below is an example of three sentences that express the same idea in different words:
Looking within is the best way to start solving the challenges in our lives.
People who are looking for solutions to their problems in the world around them will always continue to look for solutions.
The first step in healing our pain lies in our ability to "look in the mirror".
Do you think it possible is to achieve those results with more effort on fine tuning or correlation?
models I've tried:
"ramsrigouthamg/t5_paraphraser": https://huggingface.co/ramsrigouthamg/t5_paraphraser
"Vamsi/T5_Paraphrase_Paws": https://huggingface.co/Vamsi/T5_Paraphrase_Paws
Related
let me introduce the context briefly: I'm fine tuning a generic BERT model for the context of food and beverage. The final goal is a classification task.
To train this model, I'm using a corpus of text gathered from blog posts, articles, magazines etc... that cover the topic.
I am however facing a predicament that I don't know how to handle: specifically, there are sometimes words that either contain a typo, or maybe different accents, but that are semantically the same.
Let me give you an example to briefly illustrate what I mean:
The wine Gewürztraminer is correctly written with the ü, however sometimes you also find it written with just a normal u, or some other times even just Gewurtz. There are several situations like this one.
Now, a human being would obviously know that we're talking exactly about the same thing, but I have absolutely no idea about how BERT would handle these situations. Would it understand that they're the same thing? Would it consider them instead to be completely different words?
I am currently in the process of cleaning my training data, fixing the typos and trying to even out all these inconsistencies, but at this point I'm not even sure if I should do that at all, considering that the text that will need to be classified can potentially contain typos and situations like the one described above.
What would you guys suggest?
I've been working on a sentence transformation task that involves paraphrase identification as a critical step: if we are confident enough that the state of the program (a sentence repeatedly modified) has become a paraphrase of a target sentence, stop transforming. The overall goal is actually to study potential reasoning in predictive models that can generate language prior to a target sentence. The approach is just one specific way of reaching that goal. Nevertheless, I've become interested in the paraphrase identification task itself, as it's received some boost from language models recently.
The problem I run into is when I manipulate sentences from examples or datasets. For example, in this HuggingFace example, if I negate either sequence or change the subject to Bloomberg, I still get a majority "is paraphrase" prediction. I started going through many examples in the MSRPC training set and negating one sentence in a positive example or making one sentence in a negative example a paraphrase of the other, especially when doing so would be a few word edit. I found to my surprise that various language models, like bert-base-cased-finetuned-mrpc and textattack/roberta-base-MRPC, don't change their confidences much on these sorts of changes. It's surprising as these models claim an f1 score of 0.918+. The dataset is clearly missing a focus on negative examples and small perturbative examples.
My question is, are there datasets, techniques, or models that deal well when given small edits? I know that this is an extremely generic question, much more than is typically asked on StackOverflow, but my concern is in finding practical tools. If there is a theoretical technique, then it might not be suitable as I'm in the category of "available tools define your approach" rather than vice-versa. So I hope that the community would have a recommendation on this.
Short answer to the question: yes, they are overfitting. Most of the important NLP data sets are not actually well-crafted enough to test what they claim to test, and instead test the ability of the model to find subtle (and not-so-subtle) patterns in the data.
The best tool I know for creating data sets that help deal with this is Checklist. The corresponding paper, "Beyond Accuracy: Behavioral Testing of NLP models with CheckList" is very readable and goes into depth on this type of issue. They have a very relevant table... but need some terms:
We prompt users to evaluate each capability with
three different test types (when possible): Minimum Functionality tests, Invariance, and Directional Expectation tests... A Minimum Functionality test (MFT), is a collection of simple examples (and labels) to check a
behavior within a capability. MFTs are similar to
creating small and focused testing datasets, and are
particularly useful for detecting when models use
shortcuts to handle complex inputs without actually
mastering the capability.
...An Invariance test (INV) is when we apply
label-preserving perturbations to inputs and expect
the model prediction to remain the same.
A Directional Expectation test (DIR) is similar,
except that the label is expected to change in a certain way. For example, we expect that sentiment
will not become more positive if we add “You are
lame.” to the end of tweets directed at an airline
(Figure 1C).
I haven't been actively involved in NLG for long, so this answer will be a bit more anecdotal than SO's algorithms would like. Starting with the fact that in my corner of Europe, the general sentiment toward peer review requirements for any kind of NLG project are higher by several orders of magnitude compared to other sciences - and likely not without reason or tensor thereof.
This makes funding a bigger challenge, so wherever you are, I wish you luck on that front. I'm not sure of how big of a deal this site is in the niche, but [Ehud Reiter's Blog][1] is where I would start looking into your tooling ideas.
Maybe even reach out to them/him personally, because I can't think of another source that has an academic background and a strong propensity for practical applications of NLG, at least based on the kind of content they've been putting out over the years.
Your background, environment/funding, and seniority level/control you have over the project will eventually compose your vector decision for you. I's just how it goes on the bleeding edge of anything. What I will add, though, is not to limit yourself to a single language or technology in this phase because of those precise reasons you've mentioned. I'd recommend the same in terms of potential open source involvement but if your profile information is accurate, that probably won't happen, no matter what you do and accomplish.
But yeah, in the grand scheme of things, your question is far from too broad, in my view. It identifies a rather unmistakable problem pattern that not all branches of science are as lackadaisical to approach as NLG-adjacent fields seem to be right now. In that regard, it's not broad enough and will need to be promulgated far and wide before community-driven tooling will give you serious options on a micro level.
Blasphemy, sure, but the performance is already stacked against you As for the question potentially being too broad, I'd posit it is not broad enough, so long as we collectively remain in a "oh, I was waiting for you to start doing something about it" phase.
P.S. I'd eliminate any Rust and ECMAScript alternatives prior to looking into Python, blapshemous as this might sound to a 2021 data scientist
. Some might ARight nowccounting forr the ridicule this would receive xou sltrsfx hsbr s fszs drz zhsz s mrnzsl rcrtvidr, sz lrsdz
due to performance easons.
[1]: https://ehudreiter.com/2016/12/18/nlg-vs-templates/
I am a graduate student focusing on ML and NLP. I have a lot of data (8 million lines) and the text is usually badly written and contains so many spelling mistakes.
So i must go through some text cleaning and vectorizing. To do so, i considered two approaches:
First one:
cleaning text by replacing bad words using hunspell package which is a spell checker and morphological analyzer
+
tokenization
+
convert sentences to vectors using tf-idf
The problem here is that sometimes, Hunspell fails to provide the correct word and changes the misspelled word with another word that don't have the same meaning. Furthermore, hunspell does not reconize acronyms or abbreviation (which are very important in my case) and tends to replace them.
Second approache:
tokenization
+
using some embeddings methode (like word2vec) to convert words into vectors without cleaning text
I need to know if there is some (theoretical or empirical) way to compare this two approaches :)
Please do not hesitate to respond If you have any ideas to share, I'd love to discuss them with you.
Thank you in advance
I post this here just to summarise the comments in a longer form and give you a bit more commentary. No sure it will answer your question. If anything, it should show you why you should reconsider it.
Points about your question
Before I talk about your question, let me point a few things about your approaches. Word embeddings are essentially mathematical representations of meaning based on word distribution. They are the epitome of the phrase "You shall know a word by the company it keeps". In this sense, you will need very regular misspellings in order to get something useful out of a vector space approach. Something that could work out, for example, is US vs. UK spelling or shorthands like w8 vs. full forms like wait.
Another point I want to make clear (or perhaps you should do that) is that you are not looking to build a machine learning model here. You could consider the word embeddings that you could generate, a sort of a machine learning model but it's not. It's just a way of representing words with numbers.
You already have the answer to your question
You yourself have pointed out that using hunspell introduces new mistakes. It will be no doubt also the case with your other approach. If this is just a preprocessing step, I suggest you leave it at that. It is not something you need to prove. If for some reason you do want to dig into the problem, you could evaluate the effects of your methods through an external task as #lenz suggested.
How does external evaluation work?
When a task is too difficult to evaluate directly we use another task which is dependent on its output to draw conclusions about its success. In your case, it seems that you should pick a task that depends on individual words like document classification. Let's say that you have some sort of labels associated with your documents, say topics or types of news. Predicting these labels could be a legitimate way of evaluating the efficiency of your approaches. It is also a chance for you to see if they do more harm than good by comparing to the baseline of "dirty" data. Remember that it's about relative differences and the actual performance of the task is of no importance.
Let's say I have a bunch of essays (thousands) that I want to tag, categorize, etc. Ideally, I'd like to train something by manually categorizing/tagging a few hundred, and then let the thing loose.
What resources (books, blogs, languages) would you recommend for undertaking such a task? Part of me thinks this would be a good fit for a Bayesian Classifier or even Latent Semantic Analysis, but I'm not really familiar with either other than what I've found from a few ruby gems.
Can something like this be solved by a bayesian classifier? Should I be looking more at semantic analysis/natural language processing? Or, should I just be looking for keyword density and mapping from there?
Any suggestions are appreciated (I don't mind picking up a few books, if that's what's needed)!
Wow, that's a pretty huge topic you are venturing into :)
There is definitely a lot of books and articles you can read about it but I will try to provide a short introduction. I am not a big expert but I worked on some of this stuff.
First you need to decide whether you are want to classify essays into predefined topics/categories (classification problem) or you want the algorithm to decide on different groups on its own (clustering problem). From your description it appears you are interested in classification.
Now, when doing classification, you first need to create enough training data. You need to have a number of essays that are separated into different groups. For example 5 physics essays, 5 chemistry essays, 5 programming essays and so on. Generally you want as much training data as possible but how much is enough depends on specific algorithms. You also need verification data, which is basically similar to training data but completely separate. This data will be used to judge quality (or performance in math-speak) of your algorithm.
Finally, the algorithms themselves. The two I am familiar with are Bayes-based and TF-IDF based. For Bayes, I am currently developing something similar for myself in ruby, and I've documented my experiences in my blog. If you are interested, just read this - http://arubyguy.com/2011/03/03/bayes-classification-update/ and if you have any follow up questions I will try to answer.
The TF-IDF is a short for TermFrequence - InverseDocumentFrequency. Basically the idea is for any given document to find a number of documents in training set that are most similar to it, and then figure out it's category based on that. For example if document D is similar to T1 which is physics and T2 which is physics and T3 which is chemistry, you guess that D is most likely about physics and a little chemistry.
The way it's done is you apply the most importance to rare words and no importance to common words. For instance 'nuclei' is rare physics word, but 'work' is very common non-interesting word. (That's why it's called inverse term frequency). If you can work with Java, there is a very very good Lucene library which provides most of this stuff out of the box. Look for API for 'similar documents' and look into how it is implemented. Or just google for 'TF-IDF' if you want to implement your own
I've done something similar in the past (though it was for short news articles) using some vector-cluster algorithm. I don't remember it right now, it was what Google used in its infancy.
Using their paper I was able to have a prototype running in PHP in one or two days, then I ported it to Java for speed purposes.
http://en.wikipedia.org/wiki/Vector_space_model
http://www.la2600.org/talks/files/20040102/Vector_Space_Search_Engine_Theory.pdf
Is there a way to classify a particular sentence/paragraph as funny. There are very few pointers as to where one should go further on this.
There is research on this, it's called Computational Humor. It's an interdisciplinary area that takes elements from computational linguistics, psycholinguistics, artificial intelligence, machine learning etc. They are trying to find out what it is that makes stories or jokes funny (e.g. the unexpected connection, or using a taboo topic in a surprising way etc) and apply it to text (either to generate a funny story or to measure the 'funniness' of text).
There are books and articles about it (e.g. by Graeme Ritchie).
Yes, you should use a Training Corpora to build a predictive model able to detect funny sentences. Sometimes this is known as "Sentiment Analysis" in the literature. Take a look at this article about Sentiment Analysis with LingPipe.
If you can use Java, you can use their library (see license matrix). I found it very useful, not exactly in the same context than you.
The only way to pull this off is to get a couple of thousand people (monkeys won't do, sorry) to look through thousands of funny sentences/stories, rate them, and then build some sort of expert system/neural network out of it. Given the problem scope and the subjectivity of it (a thing funny to one person might not be funny - even offensive - to another), I'd say it's an impossible task.
You can use the same technique as spam filters. Instead of spam/non-spam you classify on funny/not-funny. Look into naive bayesian classifiers for more information.
http://en.wikipedia.org/wiki/Naive_Bayesian_classification
Also, try Computational Humor # Google Scholar if you're serious about getting into the field. Sentiment Analysis has been mentioned too, see wikipedia on that.
Of course, this all depends on what your scope and aims are...