Update README.md
Browse files
README.md
CHANGED
|
@@ -84,9 +84,9 @@ The cited-paper encoder was trained jointly with the query-talk encoder under a
|
|
| 84 |
Each talk *Ti* and paper *Rj* is encoded into embeddings *fT(Ti)* and *fR(Rj)*.
|
| 85 |
Their dot-product similarity *sij = fT(Ti)路fR(Rj)* is optimized using a sigmoid-based binary loss supporting multiple positives per query:
|
| 86 |
|
| 87 |
-
|
| 88 |
L = - \sum_i [y_i \log \sigma(s_i) + (1 - y_i)\log(1 - \sigma(s_i))]
|
| 89 |
-
|
| 90 |
|
| 91 |
Negatives are sampled in-batch from other talk鈥損aper pairs.
|
| 92 |
Before training, a **domain adaptation stage** aligned each talk with its own paper鈥檚 abstract to adapt to scientific and spoken-language data.
|
|
|
|
| 84 |
Each talk *Ti* and paper *Rj* is encoded into embeddings *fT(Ti)* and *fR(Rj)*.
|
| 85 |
Their dot-product similarity *sij = fT(Ti)路fR(Rj)* is optimized using a sigmoid-based binary loss supporting multiple positives per query:
|
| 86 |
|
| 87 |
+
$$
|
| 88 |
L = - \sum_i [y_i \log \sigma(s_i) + (1 - y_i)\log(1 - \sigma(s_i))]
|
| 89 |
+
$$
|
| 90 |
|
| 91 |
Negatives are sampled in-batch from other talk鈥損aper pairs.
|
| 92 |
Before training, a **domain adaptation stage** aligned each talk with its own paper鈥檚 abstract to adapt to scientific and spoken-language data.
|