Comments (3)
This is a known limitation with how the parser is currently formulated: each tree must contain a non-terminal node at the root. In your example above, the "part-of-speech" tag PU is the root node, which breaks that assumption.
The way I usually work around this is by wrapping all trees with a "TOP" symbol for the root node, and then disabling the "strip_top" flag when the treebank is known to contain examples that don't have a root non-terminal label.
from self-attentive-parser.
It looks like you're using the English ELMo embeddings to learn a parser for Chinese, which works poorly due to the language mismatch. I recommend that you use the Chinese version of BERT instead. Here is the command needed for training the model: https://github.com/nikitakit/self-attentive-parser/blob/master/EXPERIMENTS.md#chinese-models
It also looks like our dataset sizes don't match. See here for CTB data preparation scripts. My data sizes are 17544 train / 352 dev / 348 test.
from self-attentive-parser.
This is a known limitation with how the parser is currently formulated: each tree must contain a non-terminal node at the root. In your example above, the "part-of-speech" tag PU is the root node, which breaks that assumption.
The way I usually work around this is by wrapping all trees with a "TOP" symbol for the root node, and then disabling the "strip_top" flag when the treebank is known to contain examples that don't have a root non-terminal label.
Great! Thank you very much! That is very helpful.
However, When I tried to train the model on CTB data as the following comand:
python src/main.py train --use-elmo --model-path-base models/ctb_elmo --num-layers 4 --epochs 500
I got a result:
epoch 134 batch 112/112 processed 2,985,118 batch-loss 6.4597 grad-norm 17.0195 epoch-elapsed 0h05m52s total-elapsed 13h54m52s
dev-fscore (Recall=84.26, Precision=85.77, FScore=85.01, CompleteMatch=19.32) dev-elapsed 0h00m15s total-elapsed 13h55m07s
Terminating due to lack of improvement in dev fscore.
and this score is less than your "91.69 F1 on CTB 5.1 test set in benepar_zh"
my CTB data size:
- train data: 22277
- dev data: 1761
can you provide parameter settings in detail?
from self-attentive-parser.
Related Issues (20)
- How are F1 scores calculated? HOT 1
- what version of protobuf is suitable for this tool? HOT 1
- import benepar fails with torch error
- Error loading German model HOT 4
- Need help understanding the labels of the parser model
- Error reading EVALB results.
- Spacy and Berkeley parser Multi-processing HOT 3
- How to train a benepar on Ontonotes 5.0 (CONLL 2012) dataset?
- Different in parse results compared to the demo page HOT 3
- benepar_en2
- Parse quality on long sentences
- ```._.labels``` doesn't work for spans with length of one HOT 4
- Serializing the output
- Sentences with ; not split into different clauses
- Mark as unmaintained? HOT 2
- training data for benepar_zh2
- Parse pretokenized sentences?
- Cannot generate WSJ data HOT 1
- How to solve the unresognized model identifier
- The name of the benepar models (at least for french) are not updated.
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from self-attentive-parser.