Improving Distant Supervised Relation Extraction with Noise Detection Strategy
Round 1
Reviewer 1 Report
The definition of the application goal is clear and well written.
The proposed integrated/new method is coherently applied and results clearly
show its effectiveness.
My only concern is about small issues about the presentation/description.
Some detailed examples are shown as follows:
- line 17: instead of “concentrated” focussed
- line 19: in brackets add the full ref. to NDRE (noise-detection based...)
- line 34: an crucial must be a crucial
- line 46-47: If must be if
- line 64: a instance must be an instance
- line 71: maybe something is missing
- line 206: why the vector is repeated?
- equation (2): softmax (instead of Soft max)
- line 233: largeand?
- line 243: here I have some problem about your definition of the ReLu function.
In which sense you mean “a non linear function”? - line 255: does not (instead doensn’t)
- equation (12): either you display it after line 300, or you add in the texta cross ref to it
- Figure 4: add a short reference to it in the text
- Figure 5: in the caption it would be useful to add a short explanation of single, multiple and whole
- line 424: combat is too strong... solve will be better
- line 424: this paper propose should be this paper proposes.
Revised these issues, the paper will be ready to be more clearly appreciated
Author Response
Thank you for your valuable and detailed suggestions on revision. I carefully corrected the errors in the manuscript. I have revised each item one by one.
I made the following modifications.
Point 1: line 17: instead of “concentrated” focussed 

Response 1: changed ‘concentrated on’ to ‘focussed on’
Point 2: line 19: in brackets add the full ref. to NDRE (noise-detection based...)
Response 2: changed to (NDRE, Noise Detection based Relation Extraction)
Point 3: line 34: an crucial must be a crucial
Response 3: changed ‘an’ to ‘a’
Point 4: line 46-47: If must be if
Response 4: changed capitalization to lowercase
Point 5: line 64: a instance must be an instance
Response 5: changed ‘a’ to ‘an’
Point 6: line 71: maybe something is missing
Response 6: capitalized the first letter
Point 7: line 206: why the vector is repeated?
Response 7: removed the repeated vector
Point 8: equation (2):softmax (instead of Soft max)
Response 8: modified the format of ‘softmax’
Point 9: largeand?
Response 9: deleted ‘and’
Point 10: line 243: here I have some problem about your definition of the ReLu function.In which sense you mean “a non linear function”?
Response 10: relu is a commonly used activation function in neural networks. Its function is to add nonlinear factors.It can replace Sigmoid function and solve the gradient dissipation problem of BP algorithm when optimizing deep neural network to a large extent.
Point 11: line 255: does not (instead doensn’t)
Response 11: Now in line repleaced ‘doesn’t’ with ‘does not’
Point 12: equation (12): either you display it after line 300, or you add in the texta cross ref to it
Response 12: adjusted equation (12) to line 301
Point 13: add a short reference to it in the text
Response 13: Now in line 365, I have gived a reference to Figure 4
Point 14: Figure 5: in the caption it would be useful to add a short explanation of single, multiple and whole
Response 14: Now in line 380, added a short explanation of three testing modes(single, multiple,whole)
Point 15: line 424: combat is too strong... solve will be better
Response 15: Now in line 425, repleaced ‘combat’ with ‘slove’
Point 16: line 424: this paper propose should be this paper proposes.
Response 16: Now in line 425, replaced ‘propose’ with ‘propose’
Reviewer 2 Report
The work is interesting, well-motivated and supported by an appropriate number of relevant and recent references. Quality of English is good and the general structure of the paper is adequate. However, I have some minor comments in order to improve the manuscript:
[MINOR ISSUES]
- Revision of punctuation marks and spaces is mandatory (for example, line 54 “gaining great attention [7-10], Among them” instead of “. Among them”; line 197 “In the following, We first” instead of “, we first”; line 397 “featuresand” instead of “features and”, line 414 “in3.3” instead of “in 3.3”, and so on).
- I recommend enhancing the Figure 1 to better distinguish the three main tool components: The Sentence Encoder, the Noise Detection Strategy and the Bag Aggregation method.
- Include a bibliographic reference related to Xavier initialization.
Author Response
Thank you for your valuable and detailed suggestions on revision. I carefully corrected the errors in the manuscript. I have revised each item one by one and made the following modifications.
Point 1: line 54 “gaining great attention [7-10], Among them”
Response 1: changed the comma to a period
Point 2: Line 177 “In the following, We first”
Response 2: changed the first letter to lowercase
Point 3: line 397 “featuresand”
Response 3: Now in line 399 ,added a space before ‘and’
Point 4: line 414 “in3.3”
Response 4: Now in line 416 , added a space before ‘3.3’
Point 5: line 161 Figure 1
Response 5: I have revised and improved Figure 1 by adding a partition representation to distinguish the steps
Point 6: Include a bibliographic reference related to Xavier initialization
Response 6: added a reference(25) related to Xavier initialization in line 500
25.Glorot X , Bengio Y . Understanding the difficulty of training deep feedforward neural networks[J]. Journal of Machine Learning Research, 2010, 9:249-256.
Author Response File: 
Author Response.docx
Reviewer 3 Report
It seems to me that it is necessary to explain what a web table is earlier than in section 4.1 Is it any website html-code or result of processing many sites?
I see a few mistakes:
Figure 1
Noise Dectection Strategys - mistake in word «Dectection»
397
semantic featuresand
233
dot product result being too largeand
{(h1, t1), (h2, t2)...}h,t - isn’t mistake?
264 (6)
I would like to see an informal explanation of what As means in d(As, rht ) ?
408
Table 3
The correct relation of the current instance is marked in blue, and the wrong relation is marked in red.
- no red and blue in the table
435
url – not valid
Author Response
Thank you for your valuable and detailed suggestions on revision. I carefully corrected the errors in the manuscript. I have revised each item one by one.
I made the following modifications.
- Point 1: It seems to me that it is necessary to explain what a web table is earlier than in section 4.1 Is it any website html-code or result of processing many sites？
Response 1: Web table is High-quality relational tables extracted from Wikipedia, which can be downloaded from http://websail-fe.cs.northwestern.edu/TabEL/#content-code. For the Web table, I added a footnote on page 2.
- Point 2: line 161,Figure 1,Noise Dectection Strategys - mistake in word «Dectection»
Response 2: corrected the spelling mistake in Figure 1
- Point 3: line 397：semantic featuresand
Response 3: Now in 398, added a space before ‘and’
- Point 4: line 233：dot product result being too largeand
Response 4: deleted ‘and’
- Point 5: line185：{(h1, t1), (h2, t2)...}h,t - isn’t mistake?
Response 5: I have changed this set into {(h1,t1), (h2,t2), …}. The original way of describing it was to say that this set was related to (h,t), and I had written an explanation of it in front of the set.
- Point 6: Line 264：Formula (6)
I would like to see an informal explanation of what As means in d(As, rht ) ?
Response 6: A in Eq. 6 is a learnable parameter for dimension transformation of s, and ‘As’ represents the matrix multiplication of A with s. By multiplying A with s, s is transformed into a matrix in the same vector space as R.
- Point 7: Line 408：Table 3 The correct relation of the current instance is marked in blue, and the wrong relation is marked in red. no red and blue in the table
Response 7: In line 412,marked the correct releation in blue and the wrong in red.
- Point 8: line 435：url – not valid
Response 8: Now in 436,The URL link has been revised
Author Response File: Author Response.docx
