ROUGE_L score of summarization/t5 is very lower than that of paper.

[takahiro971]

🐛 Bug

Information

I try to use summarization/t5 in examples. The ROUGE_1 and ROUGE_2 is equals to that of google's paper. But, only ROUGE_L is very low!

ROUGE_1: paper=41.12 | my result=40.48 (almost equal)
ROUGE_2: paper=19.56 | my result=18.59 (almost equal)
ROUGE_L: paper=38.35 | my result=28.22 (very low ?)

Model I am using (Bert, XLNet ...): T5

Language I am using the model on (English, Chinese ...): English

The problem arises when using:

• [ ] the official example scripts: (give details below)
• [ ] my own modified scripts: (give details below)

The tasks I am working on is:

• [ ] an official GLUE/SQUaD task: (give the name)
• [ ] my own task or dataset: (give details below)

To reproduce

Steps to reproduce the behavior:

1. 2. 3.

Expected behavior

Environment info

• transformers version:
• Platform: Cent OS 7 (64bit)
• Python version: 3.7
• PyTorch version (GPU?):
• Tensorflow version (GPU?):
• Using GPU in script?:
• Using distributed or parallel set-up in script?:

[takahiro971]

So, I investigated the google's code, then I found:

https://github.com/google-research/text-to-text-transfer-transformer/blob/master/t5/evaluation/metrics.py#L76

I think that they uses not "rougeL" but "rougeLsum". And also, they says: "# Add newlines between sentences so that rougeLsum is computed correctly."

https://github.com/google-research/text-to-text-transfer-transformer/blob/master/t5/evaluation/metrics.py#L82

So, I tried the following hacks:

$ g log -p
commit 11bd4a086438b100c47e5e2b7e8696fcd67e94d1
Author: Takahiro Ito <65151988+takahiro971@users.noreply.github.com>
Date:   Tue Jun 9 14:35:19 2020 +0900

    スコア計算の不具合を修正

diff --git a/examples/summarization/t5/evaluate_cnn.py b/examples/summarization/t5/evaluate_cnn.py
index d2d6ee9..e1db944 100644
--- a/examples/summarization/t5/evaluate_cnn.py
+++ b/examples/summarization/t5/evaluate_cnn.py
@@ -44,17 +44,27 @@ def generate_summaries(lns, output_file_path, model_size, batch_size, device):

 def calculate_rouge(output_lns, reference_lns, score_path):
     score_file = Path(score_path).open("w")
-    scorer = rouge_scorer.RougeScorer(["rouge1", "rouge2", "rougeL"], use_stemmer=True)
+    scorer = rouge_scorer.RougeScorer(["rouge1", "rouge2", "rougeL", "rougeLsum"], use_stemmer=True)
     aggregator = scoring.BootstrapAggregator()

+    # copy from
+    # https://github.com/google-research/text-to-text-transfer-transformer/blob/master/t5/evaluation/metrics.py#L80
+    def _prepare_summary(summary):
+        # Make sure the summary is not bytes-type
+        # Add newlines between sentences so that rougeLsum is computed correctly.
+        summary = summary.replace(" . ", " .\n")
+        return summary
+
     for reference_ln, output_ln in zip(reference_lns, output_lns):
+        reference_ln = _prepare_summary(reference_ln)
+        output_ln = _prepare_summary(output_ln)
         scores = scorer.score(reference_ln, output_ln)
         aggregator.add_scores(scores)

     result = aggregator.aggregate()
     score_file.write(
-        "ROUGE_1: \n{} \n\n ROUGE_2: \n{} \n\n ROUGE_L: \n{} \n\n".format(
-            result["rouge1"], result["rouge2"], result["rougeL"]
+        "ROUGE_1: \n{} \n\n ROUGE_2: \n{} \n\n ROUGE_L: \n{} \n\n ROUGE_Lsum: \n{} \n\n".format(
+            result["rouge1"], result["rouge2"], result["rougeL"], result["rougeLsum"]
         )
     )

, and I got a score (37.94), near paper score. Note that: the above my code shows both "rougeL" and "rougeLsum".

Question: Why don't your code use "rougeLsum" ?

https://github.com/huggingface/transformers/blob/master/examples/summarization/t5/evaluate_cnn.py#L47

I'm sorry, I'm not good at English. I hope some kind people fix this and create PR, thanks.

Best,

[takahiro971]

P.S. the above hack is based on 41a1d27cdefd6417c298518198f99e3b8431a5c0:

$ gglv
* commit 11bd4a086438b100c47e5e2b7e8696fcd67e94d1 (HEAD, master)
| Author: Takahiro Ito <65151988+takahiro971@users.noreply.github.com>
| Date:   Tue Jun 9 14:35:19 2020 +0900
| 
|     スコア計算の不具合を修正
|  
* commit 41a1d27cdefd6417c298518198f99e3b8431a5c0 (origin/master, origin/HEAD)
| Author: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
| Date:   Mon Jun 8 21:22:37 2020 -0400

[takahiro971]

Sorry, I accidentally closed issue ...

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.