update readme
Browse files
README.md
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
This checkpoint is a variant of monot5 (T5 pointwise re-ranking model).
|
| 2 |
+
Specifically, we fuse the "P2Q (i.e. doc2query)" and "Rank (i.e. passage ranking)" to learn the **discriminative** view (Rank) and **geneartive** view (P2Q).
|
| 3 |
+
|
| 4 |
+
We found that under the specific **mixing ratio** of these two task, the effectiveness of passage re-ranking improves on par with monot5-3B models.
|
| 5 |
+
Hence, you can try to do both the task with this checkpoint by the following input format:
|
| 6 |
+
- P2Q: Document: <here is a document or a passage> Translate Document to Query:
|
| 7 |
+
- Rank: Query: <here is a query> Document: <here is a document or a passage> Relevant:
|
| 8 |
+
which the outputs will be like:
|
| 9 |
+
- P2Q: "<relevant query of the given text>"
|
| 10 |
+
- Rank: "true" or "false"
|
| 11 |
+
(Noted that we usually use the logit values of true/false tokens from T5 reranker as our query-passage relevant scores)
|
| 12 |
+
|
| 13 |
+
Check our SIGIR2021 short paper: https://dl.acm.org/doi/10.1145/3404835.3463048
|