Bert-MLM_arXiv-MP-class_arXiv

This is a sentence-transformers model: It maps sentences & paragraphs to a 768 dimensional dense vector space and can be used for tasks like clustering or semantic search.

The model is specifically designed to compute similarities of short mathematical texts.

Usage (Sentence-Transformers)

Using this model becomes easy when you have sentence-transformers installed:

pip install -U sentence-transformers

Then you can use the model like this:

from sentence_transformers import SentenceTransformer
sentences = ["In this paper we show how to compute the $\\Lambda_{\\alpha}$ norm, $\\alpha\\ge 0$, using the dyadic grid. This result is a consequence of the description of the Hardy spaces $H^p(R^N)$ in terms of dyadic and special atoms.",
             "We show that a determinant of Stirling cycle numbers counts unlabeled acyclic single-source automata. The proof involves a bijection from these automata to certain marked lattice paths and a sign-reversing involution to evaluate the determinant."]

model = SentenceTransformer('math-similarity/Bert-MLM_arXiv-MP-class_arXiv')
embeddings = model.encode(sentences)
print(embeddings)

Usage (HuggingFace Transformers)

Without sentence-transformers, you can use the model like this: First, you pass your input through the transformer model, then you have to apply the right pooling-operation on-top of the contextualized word embeddings.

from transformers import AutoTokenizer, AutoModel
import torch


#Mean Pooling - Take attention mask into account for correct averaging
def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output[0] #First element of model_output contains all token embeddings
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)


# Sentences we want sentence embeddings for
sentences = ["In this paper we show how to compute the $\\Lambda_{\\alpha}$ norm, $\\alpha\\ge 0$, using the dyadic grid. This result is a consequence of the description of the Hardy spaces $H^p(R^N)$ in terms of dyadic and special atoms.",
             "We show that a determinant of Stirling cycle numbers counts unlabeled acyclic single-source automata. The proof involves a bijection from these automata to certain marked lattice paths and a sign-reversing involution to evaluate the determinant."]

# Load model from HuggingFace Hub
tokenizer = AutoTokenizer.from_pretrained('math-similarity/Bert-MLM_arXiv-MP-class_arXiv')
model = AutoModel.from_pretrained('math-similarity/Bert-MLM_arXiv-MP-class_arXiv')

# Tokenize sentences
encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')

# Compute token embeddings
with torch.no_grad():
    model_output = model(**encoded_input)

# Perform pooling. In this case, mean pooling.
sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])

print("Sentence embeddings:")
print(sentence_embeddings)

---

Background

Intended uses

Our model is intended to be used as a sentence and short paragraph encoder for mathematical texts. Given an input text, it outputs a vector which captures the semantic information. The sentence vector may be used for information retrieval, clustering or sentence similarity tasks.

By default, input text longer than 256 word pieces is truncated.

Training procedure

Domain-adaption

We use the domain-adapted math-similarity/Bert-MLM_arXiv model. Please refer to the model card for more detailed information about the domain-adaption procedure.

Pooling

We add a mean-pooling layer on top of the domain-adapted model.

Fine-tuning

We fine-tune the model using a cosine-similarity objective. Formally, it computes the vectors u = model(sentence_A) and v = model(sentence_B) and measures the cosine-similarity between the two. By default, it minimizes the following loss: ||input_label - cos_score_transformation(cosine_sim(u,v))||_2, with MSE as loss function.

We use pairs of concenated title+abstract texts from arXiv as fine-tuning dataset and model semantic similarity with their classificattion codes. Two texts are defined as similar, if they share their classification code. Otherwise, they are defined as semantically dissimilar. We only use texts with a singular classification code in this dataset. The training set contains 43.572 text pairs and the evaluation set contains 5.447 pairs. See the fine-tuning notebook for more information.

We do not include the fine-tuning dataset directly. It can be recreated using its generation notebook and datasets/math-similarity/arXiv-metadata-oai-snapshot-111.

Citing & Authors

This model is an additional resource for the CICM'24 submission On modelling similarity of short mathematical texts.