---
license: cc-by-nc-4.0
library_name: gensim
tags:
- word2vec
- word-embeddings
- gensim
- earnings-calls
- finance
- nlp
---

# Earnings-Call Word2Vec

Word2Vec word embeddings (gensim, CBOW, 300 dimensions, window 5) trained on the
question-and-answer portions of quarterly earnings-call transcripts. The repository
holds one model per year (2022-2025); each year continues training from the previous
year's model. The 2025 bigram and trigram phrase models are included so that new text
can be preprocessed into the same multi-word tokens that appear in the vocabulary
(multi-word terms are joined with an underscore, e.g. `supply_chain`).

These embeddings are general purpose. Measuring corporate culture (Li, Mai, Shen, and
Yan, 2021) is one downstream use; the vectors can be applied to any task over
earnings-call or similar business text.

## Files

| File | Contents |
|------|----------|
| `word2vec_YYYY.kv` (+ `.kv.vectors.npy`) | KeyedVectors for year `YYYY` (2022-2025): word vectors + vocabulary |
| `phrases_bigram_2025.mod` | gensim Phrases model that joins two-word phrases |
| `phrases_trigram_2025.mod` | gensim Phrases model that joins three-word phrases |

## Usage

```python
from huggingface_hub import hf_hub_download
from gensim.models import KeyedVectors

# KeyedVectors.load needs the .vectors.npy file alongside the .kv file
kv_path = hf_hub_download("maifeng/earnings-call-word2vec", "word2vec_2025.kv")
hf_hub_download("maifeng/earnings-call-word2vec", "word2vec_2025.kv.vectors.npy")
kv = KeyedVectors.load(kv_path)
kv.most_similar("innovation")
```

To tokenize new text the same way before looking up vectors, apply the phrase models:

```python
from huggingface_hub import hf_hub_download
from gensim.models.phrases import Phrases

bigram = Phrases.load(hf_hub_download("maifeng/earnings-call-word2vec", "phrases_bigram_2025.mod"))
trigram = Phrases.load(hf_hub_download("maifeng/earnings-call-word2vec", "phrases_trigram_2025.mod"))
tokens = trigram[bigram["the supply chain was disrupted".split()]]
```

## Method and reference

The training pipeline (preprocessing, phrase detection, Word2Vec, seed-word dictionary
expansion) is available as the [`lmsy_w2v_rfs`](https://github.com/maifeng/lmsy_w2v_rfs)
package.

> Kai Li, Feng Mai, Rui Shen, and Xinyan Yan, "Measuring Corporate Culture Using Machine
> Learning," *The Review of Financial Studies*, 2021. DOI:
> [10.1093/rfs/hhaa079](https://doi.org/10.1093/rfs/hhaa079).

## License

Released under [CC-BY-NC-4.0](https://creativecommons.org/licenses/by-nc/4.0/):
non-commercial reuse and redistribution are permitted with attribution to the
reference above.