MARL: for IJCAI 2020 submission.

Our paper is published in IJCAI 2020 [1], which is "Retrieve, Program, Repeat: Complex Knowledge Base Question Answering via Alternate Meta-learning". We aim to solve the CQA task [2], which is 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

Now we will talk about how to training and testing our proposed model.
We first clone the project:

git clone https://github.com/DevinJake/MARL.git

, and we could download a project MARL.

1. Experiment environment.

(1). Python = 3.6.4
(2). PyTorch = 1.1.0
(3). TensorFlow = latest
(4). tensorboardX = 2.0
(5). ptan = 0.4
(6). flask = 1.1.1
(7). requests = 2.22.0

2. Accessing knowledge graph.

(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 MARL/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 MARL/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 MARL/data/bfs_data.
Run the project file MARL/BFS/server.py to activate the KG server for retrieval:

python server.py

3. Retriever pre-training.

Based on the edit-distance and the Jaccard similarity, we retrieved the most similar instances for each question. We treated the retrieved instances as the positive samples to pre-train the retriever (which is a DSSM model) to solve the cold-start problem.

(1). Download relevant materials.
Firstly, we need place the following files in the project folder MARL/data/auto_QA_data for pre-training the retriever:
share.question (vocabulary), CSQA_DENOTATIONS_full_944K.json (the file that records the information relevant to all the training questions and is compressed in the Google drive), CSQA_result_question_type_944K.json, CSQA_result_question_type_count944K.json, CSQA_result_question_type_count944k_orderlist.json, and 944k_rangeDict.json (the files that are used to retrieve the support sets).

Also, we need to place a pre-trained model epoch_002_0.394_0.796.dat in the project folder MARL/data/saves/maml_batch8_att=0_newdata2k_1storder_1task, in which the learned word embeddings are stored. The embeddings are used to initialize the question embedding by summing and averaging.

Furthermore, we have processed the training dataset and thus we need to download the file RL_train_TR_new_2k.question and place it in the project folder MARL/data/auto_QA_data/mask_even_1.0%.

We will analyse the question and find the most similar instances from the training dataset by evaluating the edit-distance and the Jaccard similarity as well. The found instances are treated as the positive samples for each question, and are used to pre-train the retriever.

All the materials could be downloaded from the the provided data link.

(2). Pre-train the retriever.
In the project folder MARL/S2SRL, we run the python file to pre-train the retriever:

python retriever_pretrain.py

The program will first automatically create the dataset for training the retriever in the files retriever_question_documents_pair.json (the positive instances for each training question) and retriever_training_samples.json (the training samples for the retriever) in the project folder MARL/data/auto_QA_data.

Also, a model initial_epoch_000_1.000.dat will be automatically created in the folder MARL/data/saves/retriever. The model is used to store the initialized question embedding.

Then, the retriever will be learned by using the above files to accomplish the pre-training. The pre-trained retriever models would also be stored in the project folder MARL/data/saves/retriever.

4. Meta-learner & Retriever joint learning.

(1). Load the pre-trained models.
We have pre-trained a CQA model based on Reinforcement learning, and will further trained this RL-based model by using MAML.
We could download the pre-trained RL model truereward_0.739_29.zip, uncompress it, and place it in the project folder MARL/data/saves/rl_even_TR_batch8_1%.

As mentioned before, we have also pre-trained the retriever.
We have saved a retriever model AdaBound_DocEmbed_QueryEmbed_epoch_140_4.306.zip in the folder MARL/data/saves/retriever, which is the best pre-trained retriever model we got.

We need to download the aforementioned files from the data link, uncompress them, and further put them under the corresponding folders in our project.

(2). Train the MAML model.
In the project folder MARL/S2SRL, we run the python file to train the MAML model:

python train_maml_retriever_joint.py

The trained CQA model and the retriever model would be stored in the folder MARL/data/saves/maml_newdata2k_reptile_retriever_joint.

5. MAML testing.

(1). Load the trained model.
The trained models will be stored in the folder MARL/data/saves/maml_newdata2k_reptile_retriever_joint.
We have saved a trained CQA model net_epoch_016_0.782_0.719.zip and a retriever model retriever_epoch_016_0.785_0.719.zip in this folder, which could lead to the SOTA result.
We could download such models from the data link, uncompress them, and place them 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 the models net_epoch_016_0.782_0.719.dat and retriever_epoch_016_0.785_0.719.dat.

(2). Load the testing dataset.
We also have processed the testing dataset SAMPLE_FINAL_MAML_test.question (which is 1/20 of the full testing dataset) and FINAL_MAML_test.question (which is the full testing dataset), and saved them in the folder MARL/data/auto_QA_data/mask_test.
We could download the files from the data link and put them under the folder `MARL/da

MARL

Install / Use