roomylee / entity-aware-relation-classification

Tensorflow Implementation of "Semantic Relation Classification via Bidirectional LSTM Networks with Entity-aware Attention using Latent Entity Typing" (arXiv 2019)

Hi, I'm Joohong

I am a Machine Learning Researcher at Scatter Lab in South Korea 🇰🇷.

🚀 I’m currently working on ...

• Developing a Korean Open-domain Chatbot Luda.
• Building a Dataset for Korean Language Understanding Evaluation, KLUE.

🔥 I'm interested in ...

• Open-domain Conversational AI
• Language Modeling and Represenation Learning
• Production ML System in the Real World

📫 How to reach me

entity-aware-relation-classification's Issues

questions about paper

Hello,'Figure 4: Visualization of Self Attention' in paper,sentence “the pollutionwas caused
by the shipwreck”,pollution is the effect and shipwreck is the cause,so the relation between two entities is Cause-Effect(e2,e1), but in paper,predicting the Cause-Effect(e1,e2),is it correct? Did I misunderstand?

State-of-the-art accuracy?

Dear authors,

Thanks for your work and for releasing the code.

There is a paper from ACL 2016 Relation Classification via Multi-Level Attention CNNs which claims to reach 88 F-score on the same dataset and still does not use "rich" features. Is there are a reason you don't cite it or compare to its results?

Thanks.

Question about latent type?

Hi,

First, thanks for sharing your work here. I totally enjoyed reading your paper, but I am a bit confused about how the latent type was obtained. Was it computed during the training? Or was it pre-calculated? And what is the latent type vector c in equation 3.11 and 3.12?

Thanks in advance,
Neil

README IMAGES

Meet some problem in logger.py

    str_parse = str(process.communicate()[0]).split("\\n")[-2]

IndexError: list index out of range

question about the result

hello，I have run the code you provided for more than 10 times (no change), and the results are all around 84.0. The highest result is 84.6, which cannot reach 85.2 or even 84.7 in the paper.Do I need to adjust some parameters not mentioned in the paper?
And I wonder if these values mean anything in the code you provide？

vtanh(W[h;p1;p2]+W*[e1;e2]) 85.18%? 84.83% 84.55%

vtanh(W[h;p1;p2;e1;e2]) 85.18% 84.41%