comparatively-finetuning-bert | Comparatively fine-tuning pretrained BERT models | Natural Language Processing library

As for why the tagger doesn't find "accredit" from "accreditation", this is because the scheme .lemma finds the lemma of words, not actually the stems. See the difference between stem and lemma on Wikipedia.

The stem is the part of the word that never changes even when morphologically inflected; a lemma is the base form of the word. For example, from "produced", the lemma is "produce", but the stem is "produc-". This is because there are words such as production and producing In linguistic analysis, the stem is defined more generally as the analyzed base form from which all inflected forms can be formed.

The documentation uses the word "stem", but I do think that the lemma is what is intended here, and getting "accreditation" is the expected behaviour. See the Usage section of the Wikipedia article for "Word stem" for more info. The lemma is the dictionary form of a word, and "accreditation" has a dictionary entry, whereas something like "accredited" doesn't. Whatever you call these things, the point is that there are two distinct concepts, and the tagger gets you one of them, but you are expecting the other one.

As for why the order of the words matters, this is because the tagger tries to analyse your words as "natural language", rather than each one individually. Naturally, word order matters. If you use .lexicalClass, you'll see that it thinks the third word in text2 is an adjective, which explains why it doesn't think its dictionary form is "accredit", because adjectives don't conjugate like that. Note that accredited is an adjective in the dictionary. So "is it a grammar issue?" Exactly.

To remove all non-alpha characters but - between letters, you can use

You could use a regex and extractall:

I would recommend to downgrade your milvus version to a version before the 2.0 release just a week ago. Here is a discussion on that topic: https://github.com/deepset-ai/haystack/issues/2081

This type of problem where you want to extract the customer problem from the original text is called Extractive Summarization and this type of task is solved by Sequence2Sequence models.

The main reason for this type of model being called Sequence2Sequence is because the input and the output of this model would both be text.

I recommend you to use a transformers model called Pegasus which has been pre-trained to predict a masked text, but its main application is to be fine-tuned for text summarization (extractive or abstractive).

This Pegasus model is listed on Transformers library, which provides you with a simple but powerful way of fine-tuning transformers with custom datasets. I think this notebook will be extremely useful as guidance and for understanding how to fine-tune this Pegasus model.

try

There is a paper Masked Language Model Scoring that explores pseudo-perplexity from masked language models and shows that pseudo-perplexity, while not being theoretically well justified, still performs well for comparing "naturalness" of texts.

As for the code, your snippet is perfectly correct but for one detail: in recent implementations of Huggingface BERT, masked_lm_labels are renamed to simply labels, to make interfaces of various models more compatible. I have also replaced the hard-coded 103 with the generic tokenizer.mask_token_id. So the snippet below should work:

You can first replace the dictionary keys in sentence to {} so that you can easily format a string in loop. Then you can use itertools.product to create the Cartesian product of dictionary.values(), so you can simply loop over it to create your desired sentences.

The difference between AutoModel and AutoModelForSequenceClassification model is that AutoModelForSequenceClassification has a classification head on top of the model outputs which can be easily trained with the base model