Our paper is published in JWS 2020 [1], which is 'Less is more: Data-efficient complex question answering over knowledge bases'. We aim to solve the CQA task [2], targeting at answering factual questions through complex inferring over a realistic-sized KG of millions of entities.
We could learn the details of the CQA dataset here.
All the materials required for running the KG sever, training the model, and testing in this task could be downloaded from the data link.
We should follow the folder structure in the data link, and place the files in the corresponding location under the data
folder.
Following this README, we will instruct how to use the relevant data from the data link.
The questions in the CQA could be categorized into seven groups.
The typical examples of these seven question types are displayed in the following table.
Question Type | Question | KB artifacts | Action Sequence | Answer |
---|---|---|---|---|
Simple | Which administrative territory is Danilo Ribeiro an inhabitant of? | E1: Danilo Ribeiro R1: country of citizenship T1: administrative territory |
Select(E1, R1, T1) | Brazil |
Logical | Which administrative territories are twin towns of London but not Bern? | E1: London E2: Bern R1: twinned adm. body T1: administrative territory |
Select(E1, R1, T1) Diff(E2, R1, T1) |
Sylhet, Tokyo, Podgorica, Phnom Penh, Delhi, Los Angeles, Sofia, New Delhi, ... |
Quantitative | Which sports teams have min number of stadia or architectural structures as their home venue? | R1: home venue T1: sports team T2: stadium T3: architectural structure |
SelectAll(T1, R1, T2) SelectAll(T1, R1, T3) ArgMin() |
Detroit Tigers, Drbak-Frogn IL, Club Sport Emelec, Chunichi Dragons, ... |
Comparative | Which buildings are a part of lesser number of architectural structures and universities than Midtown Tower? | E1: Midtown Tower R1: part of T1: building T2: architectural structure T3: university |
SelectAll(T1, R1, T2) SelectAll(T1, R1, T3) LessThan(E1) |
Amsterdam Centraal, Hospital de Sant Pau, Budapest Western Railway Terminal, El Castillo, ... |
Verification | Is Alda Pereira-Lemaitre a citizen of France and Emmelsbull-Horsbull? | E1: Alda Pereira-Lemaitre E2: France E3: Emmelsbull-Horsbull R1: country of citizenship T1: administrative territory |
Select(E1, R1, T1) Bool(E2) Bool(E3) |
YES and NO respectively |
Quantitative Count | How many assemblies or courts have control over the jurisdiction of Free Hanseatic City of Bremen? | E1: Bremen R1: applies to jurisdiction T1: deliberative assembly T2: court |
Select(E1, R1, T1) Union(E1, R1, T2) Count() |
2 |
Comparative Count | How many art genres express more number of humen or concepts than floral painting? | E1: floral painting R1: depicts T1: art genre T2: human t3: concept |
SelectAll(T1, R1, T2) SelectAll(T1, R1, T3) GreaterThan(E1) Count() |
8 |
We first clone the project:
git clone https://github.com/DevinJake/NS-CQA.git
, and we could download a project NS-CQA
.
Now we will talk about how to training and testing our proposed model on CQA dataset.
(1). Python = 3.6.4
(2). PyTorch = 1.1.0
(3). TensorFlow = 1.14.0
(4). tensorboardX = 2.1
(5). ptan = 0.4
(6). flask = 1.1.2
(7). requests = 2.24.0
(1). Assign the IP address and the port number for the KG server.
Manually assign the IP address and the port number in the file of the project NS-CQA/BFS/server.py
.
Insert the host address and the post number for your server in the following line of the code:
app.run(host='**.***.**.**', port=####, use_debugger=True)
Manually assign the IP address and the port number in the file of the project NS-CQA/S2SRL/SymbolicExecutor/symbolics.py
.
Insert the host address and the post number for your server in the following three lines in the symbolics.py
:
content_json = requests.post("http://**.***.**.**:####/post", json=json_pack).json()
(2). Run the KG server.
Download the bfs data bfs_data.zip
from the provided data link.
We need to uncompress the file bfs_data.zip
and copy the three pkl files into the project folder NS-CQA/data/bfs_data
.
Run the project file NS-CQA/BFS/server.py
to activate the KG server for retrieval:
python server.py
(1). Load the pre-trained models. By using a breadth-first-search (BFS) algorithm, we generated pseudo-gold action sequences for a tiny subset of questions and pre-trained the model by Teacher Forcing with the help of these pairs of questions and action sequences. Therefore, we will further train the neural generator by using Reinforcement learning.
We could download the pre-trained RL model pre_bleu_0.956_43.zip
, uncompress it, and place it in the project folder NS-CQA/data/saves/pretrained
.
If you want to pre-train the models by yourself, you could:
- Download the 'data\auto_QA_data\mask_even_1.0%\PT_train_INT.question' and 'data\auto_QA_data\mask_even_1.0%\PT_train_INT.action' from the data link. Put them under the folder 'mask_even_1.0%'. The files are the pseudo-gold annotations that were formed by using the BFS algorithm.
- Run python file 'S2SRL\train_crossent.py'.
- Under the folder 'data\saves\pretrained' you could find the models, which are trained under the teacher-forcing paradigm. The models are named following the format 'pre_bleu_TestingScore_numberOfEpochs.dat'. Normally we chose the model with the highest bleu testing score as the pre-trained model for following RL training. Or you could choose whatever model you like.
- The performance of the chosen model might have a tiny difference between the uploaded pre-trained model 'pre_bleu_0.956_43.dat'. We recommend you to choose 'pre_bleu_0.956_43.dat' for re-implementation.
We also provided the code for the BFS algorithm. If you are interested, you could:
-
Download the 'data\official_downloaded_data\10k\train_10k.zip' from the data link. Uncompress the zip file and put them under the folder 'train_10k'. The files are the provided 10K samples of the CQA dataset.
-
Under the folder
NS-CQA/S2SRL/SymbolicExecutor
, we run the python file to search for the pseudo-gold annotations for thesimple
type questions by using the BFS algorithm:
python auto_symbolic_simple.py
The pseudo-gold annotations would be stored in the file NS-CQA/data/annotation_logs/jws_simple_auto.log
.
Similarly, we could get jws_logical_auto.log
and jws_count_auto.log
by running auto_symbolic_logical.py
and auto_symbolic_count.py
.
We provide the BFS codes for these three types of the questions for demonstration.
(2). Train the neural generator.
In the project folder NS-CQA/S2SRL
, we run the python file to train the MAML model:
python train_scst_cher.py
The trained neural generator and the corresponding action memory would be stored in the folder NS-CQA/data/saves/rl_cher
.
(1). Load the trained model.
The trained models will be stored in the folder NS-CQA/data/saves/rl_cher
.
We have saved a trained CQA model epoch_022_0.793_0.730.zip
in this folder, which could leads to the SOTA result.
We could download such model from the data link, uncompress it, and place it under the corresponding project folder.
When testing the model, we could choose a best model from all the models that we have trained, or simply use our trained model epoch_022_0.793_0.730.dat
.
(2). Load the testing dataset.
We also have processed the testing dataset SAMPLE_FINAL_INT_test.question
(which is 1/20 of the full testing dataset) and FINAL_INT_test.question
(which is the full testing dataset), and saved them in the folder NS-CQA/data/auto_QA_data/mask_test
.
We could download the files from the data link and put them under the folder NS-CQA/data/auto_QA_data/mask_test
in the project.
(3). Test.
In the project file NS-CQA/S2SRL/data_test.py
, we could change the parameters to meet our requirement.
In the command line:
sys.argv = ['data_test.py', '-m=epoch_022_0.793_0.730.dat', '--cuda', '-p=final_int',
'--n=rl_cher', '--att=0', '--lstm=1',
'--int', '-w2v=300', '--beam_search']
, we could change the following settings.
If we want to use the entire testing dataset to get the QA result, we should set -p=final_int
.
Otherwise, we could set -p=sample_final_int
to test on the subset of the testing dataset to get an approximation testing result with less time.
Based on our observation, the testing results of the entire testing dataset are always better than those of the subset.
If we want to use the models stored in the named folder NS-CQA/data/saves/rl_cher
, we set --n=rl_cher
.
If we want to use our saved CQA model net_***.dat
in the named folder to test the questions, we set -m=net_***.dat
.
After setting, we run the file NS-CQA/S2SRL/data_test.py
to generate the action sequence for each testing question:
python data_test.py
We could find the generated action sequences in the folder where the model is in (for instance NS-CQA/data/saves/rl_cher
), which is stored in the file final_int_predict.actions
or sample_final_int_predict.actions
.
(4). Calculate the result.
Firstly, we should download the files CSQA_ANNOTATIONS_test_INT.json
from the data link and put it into the folder NS-CQA/data/auto_QA_data/
of the project, which is used to record the ground-truth answers of each question.
After generating the actions, we could use them to compute the QA result.
For example, we use the saved models to predict actions for the testing questions, and therefore generate a file NS-CQA/data/saves/rl_cher/final_int_predict.actions
to record the generated actions for all the testing questions.
Then in the file NS-CQA/S2SRL/SymbolicExecutor/calculate_sample_test_dataset.py
, we set the parameters as follows.
This line of code
calculate_RL_or_DL_result('rl_cher', withint=True)
will finally call the function transMask2Action()
to compute the accuracy of the actions stored in the file NS-CQA/data/saves/rl_cher/final_int_predict.actions
.
We could change the path of the generated file in the following line in the function transMask2Action()
if we want:
with open(json_path, 'r') as load_f, \
open("../../data/saves/rl_cher/final_int_predict.actions", 'r') as predict_actions, \
open(question_path, 'r') as RL_test:
Then we run the file NS-CQA/S2SRL/SymbolicExecutor/calculate_sample_test_dataset.py
to get the final result:
python calculate_sample_test_dataset.py
The result will be stored in the file NS-CQA/data/auto_QA_data/test_result/rl_cher.txt
.
Now we will talk about how to training and testing our proposed model on WebQSP(WebQuestions Semantic Parsest) dataset.
(1). Python = 3.6.4
(2). PyTorch = 1.1.0
(3). TensorFlow = 1.14.0
(4). tensorboardX = 2.1
(5). ptan = 0.4
(6). flask = 1.1.2
(7). requests = 2.24.0
(1). Assign the IP address and the port number for the KG server. Manually assign the IP address and the port number in the file of the project NS-CQA/webqspUtil/server.py Insert the host address and the post number for your server in the following line of the code:
app.run(host='**.***.**.**', port=####, use_debugger=True)
Manually assign the IP address and the port number in the file of the project NS-CQA/S2SRL/SymbolicExecutor/symbolics_webqsp_novar.py
.
Insert the host address and the post number for your server in the following three lines in the symbolics.py
:
content_json = requests.post("http://**.***.**.**:####/post", json=json_pack).json()
(2). Run the KG server.
The kg we used is the subgraph about webQSP of freebase, you can download the file webQSP_freebase_subgraph.zip
form the provided data link,
the final path is : data/webqsp_data/webQSP_freebase_subgraph.json
.
Run the project file NS-CQA/webqspUtil/server.py
to activate the KG server:
python server.py
(1). Prepare the data.
The WebQSP dataset can be download from the data link. We process the original dataset to
train and test in our experiment.
The processed data we use can be download from the data link. WE could place the processed
pre-train data PT_train.question
, PT_train.action
, PT_test.question
, PT_test.action
in the folder NS-CQA/data/webqsp_data/mask
,
place vocab file share.webqsp.question
and final data final_webqsp_train_RL.json
, final_webqsp_test_RL.json
in NS-CQA/data/webqsp_data
,
(2). Load the pre-trained models.
You can run the project file NS-CQA/S2SRL/train_crossent_webqsp.py
to train your own pre-train model:
python train_crossent_webqsp.py
or download the pre-trained RL model epoch_030_0.995_0.966.dat, uncompress it,
and place it in the project folder NS-CQA/data/webqsp_saves/pretrained
.
(2). Train the neural generator.
In the project folder NS-CQA/S2SRL
, we run the python file to train the model:
python train_scst_cher_webqsp.py
The trained neural generator and the corresponding action memory would be stored in the folder NS-CQA/data/webqsp_saves/rl_cher
.
(1). Load the trained model.
The trained models will be stored in the folder NS-CQA/data/webqsp_saves/rl_cher
.
We have saved a trained CQA model epoch_000_0.838_0.000.dat
in this folder, which could leads to the SOTA result.
We could download such model from the data link, uncompress it,
and place it under the corresponding project folder.
When testing the model, we could choose a best model from all the models that we have trained, or simply use our trained model epoch_000_0.838_0.000.dat
.
(2). Load the testing dataset.
You can find final_webqsp_test_RL.json
in NS-CQA/data/webqsp_data
as our test data.
(3). Test.
In the project file NS-CQA/S2SRL/data_test_RL_webqsp.py
, we could change the parameters to meet our requirement.
In the command line:
sys.argv = ['data_test_RL_webqsp.py', '-m=epoch_000_0.838_0.000.dat', '-p=rl', '--n=rl_cher', '--att=0', '-w2v=300', '--lstm=1']
, we could change the following settings.
If we want to use the models stored in the named folder NS-CQA/data/saves/rl_cher
, we set --n=rl_cher
.
If we want to use our saved WebQSP model net_***.dat
in the named folder to test the questions, we set -m=net_***.dat
.
After setting, we run the file NS-CQA/S2SRL/data_test_RL_webqsp.py
to generate the action sequence for each testing question:
python data_test_RL_webqsp.py
We could find the generated action sequences in the folder where the model is in (for instance NS-CQA/data/webqsp_saves/rl_cher
),
which is stored in the file rl_predict.actions
.
[1]. Yuncheng Hua, Yuan-Fang Li, Guilin Qi, Wei Wu, Jingyao Zhang, and Daiqing Qi. 2020. Less is more: Data-efficient complex question answering over knowledge bases. In Journal of Web Semantics 65 (2020): 100612.
[2]. Amrita Saha, Vardaan Pahuja, Mitesh M Khapra, Karthik Sankaranarayanan, and Sarath Chandar. 2018. Complex sequential question answering: Towards learning to converse over linked question answer pairs with a knowledge graph. In ThirtySecond AAAI Conference on Artificial Intelligence.
Hua, Y., Li, Y. F., Qi, G., Wu, W., Zhang, J., & Qi, D. (2020). Less is more: Data-efficient complex question answering over knowledge bases. Journal of Web Semantics, 65, 100612.