thu-coai / ccm Goto Github PK

This project is a tensorflow implement of our work, CCM (Commonsense Conversational Model).

License: Apache License 2.0

Python 100.00%

ccm's Introduction

Commonsense Knowledge Aware Conversation Generation with Graph Attention

Introduction

Commonsense knowledge is vital to many natural language processing tasks. In this paper, we present a novel open-domain conversation generation model to demonstrate how large-scale commonsense knowledge can facilitate language understanding and generation. Given a user post, the model retrieves relevant knowledge graphs from a knowledge base and then encodes the graphs with a static graph attention mechanism, which augments the semantic information of the post and thus supports better understanding of the post. Then, during word generation, the model attentively reads the retrieved knowledge graphs and the knowledge triples within each graph to facilitate better generation through a dynamic graph attention mechanism, as shown in Figure 1.

This project is a tensorflow implement of our work, CCM.

Dependencies

Python 2.7
Numpy
Tensorflow 1.3.0

Quick Start

Dataset

Commonsense Conversation Dataset contains one-turn post-response pairs with the corresponding commonsense knowledge graphs. Each pair is associated with some knowledge graphs retrieved from ConceptNet. We have applied some filtering rules to retain high-quality and useful knowledge graphs.

Please download the Commonsense Conversation Dataset to data directory.
Train

python main.py

The model will achieve the expected performance after 20 epochs.
Test

python main.py --is_train False

You can test the model using this command. The statistical result and the text result will be output to the 'test.res' file and the 'test.log' file respectively.

Details

Training

You can change the model parameters using:

--units xxx 				the hidden units
--layers xxx 				the number of RNN layers
--batch_size xxx 			batch size to use during training 
--per_checkpoint xxx 			steps to save and evaluate the model
--train_dir xxx				training directory

Evaluation

Paper

Hao Zhou, Tom Yang, Minlie Huang, Haizhou Zhao, Jingfang Xu, Xiaoyan Zhu.
Commonsense Knowledge Aware Conversation Generation with Graph Attention.
IJCAI-ECAI 2018, Stockholm, Sweden.

Please kindly cite our paper if this paper and the code are helpful.

Acknowlegments

Thanks for the kind help of Prof. Minlie Huang and Prof. Xiaoyan Zhu. Thanks for the support of my teammates.

License

Apache License 2.0

ccm's People

Contributors

Stargazers

Watchers

Forkers

cyzhangathit pkulzb tonydeep yifei87 neuqmiao haozijie tk1363704 qss2012 nlpr-ss sungjinlees wangxuekui preke shinichr liyazheng zxsted liuchang97 jufengada chenmoshushi zhongyunuestc happyyolanda awesome-text-generation hins helicqin memozhu sixingwu yaoyiran bellamn foreseez litetoooooom eric11eca glin93 huanglinshan luomuqinghan xiehuateng kl-source ylqi fw339wj 643245611 awa121 hatleon embeddedsamurai jiangchenglin521 nbgao cuiqingyao panda0881 auscenery shark-3 sdwivedi mengruwg jmmunozr xrosliang tuxchow fodexiaozhijiagai hodge-ge xiaoanshi d294270681 chrisbenka

ccm's Issues

Unable to download dataset

I have tried both the normal and ftp version of download listed here http://coai.cs.tsinghua.edu.cn/hml/dataset/#commonsense

The download doesn't proceed after 274KB

Thanks

ConceptNet and input data

Hi, I am very interested in your work, but I have some doubts about where is the input data and how do you use the Conceptnet? Thanks for sharing!

Can't get the dataset

Hi, tuxcow , I can't get the dataset from the download link. Any help, thanks.

Can't get the dataset

Hi tuxcow , I can't get the dataset from the download link, could you provide the new download address? Or provide sample dataset? Thank you anyway~

Not the author but I'm working on extending this paper for my master's thesis so I've done some work to decode the input data and be able to recreate it so I think I can provide some insight.

Match_triples are the triples where an entity from the post and an entity from the response appear in the same commonsense knowledge triple.

Match_index is the list of response entities matched with post entities in the following format: the first index is the number of the entity in post_triples that matches the current word (the list is the length of the response), the second index is the index of the entity that matches in all_entities. [-1,-1] is appended if the response word is not an entity or the entity doesn't match anything in the post.

Post_triples is a list of the entities that appear in the post, with 0 representing an entity is not found in the list of entities and >0 indicating an index of entities starting with 1 and incrementing each time a new entity is found.

all_entities is a list of all the matching entities on the other end of a csk triple for the entities found in the post.

response_triples is either -1 if a word is not an entity or the entity doesn't match a triple in csk or the index of the matched triple where the current word is an entity and that entity is part of a triple in both the post and response.

all_triples is simply a list of all the matched triples between the post and response entities.

Below is the script I've written to recreate the training data. It seems to output extra entities in all_entities for some reason, at least more than what the authors found but that shouldn't break anything I think. If you find a bug in the script please let me know so I can update it on my end.

import json

test = {"post": ["you", "mean", "the", "occupation", "that", "did", "happen", "?"], "response": ["no", "i", "mean", "the", "fighting", "invasion", "that", "the", "military", "made", "so", "many", "purple", "hearts", "for", "in", "anticipation", "for", "that", "we", "have", "n't", "used", "up", "to", "this", "day", "."]}
f = open('resource.txt')
data = json.load(f)
f.close()

data['postEntityToCSKTripleIndex'] = {}
data['postEntityToOtherCSKTripleEntities'] = {}
index = 0
for triple in data['csk_triples']:
    firstEntity = triple.split(',')[0]
    secondEntity = triple.split(',')[2].strip()
    if(not firstEntity in data['postEntityToCSKTripleIndex']):
        data['postEntityToCSKTripleIndex'][firstEntity] = []
    data['postEntityToCSKTripleIndex'][firstEntity].append(index)
    if(not secondEntity in data['postEntityToCSKTripleIndex']):
        data['postEntityToCSKTripleIndex'][secondEntity] = []
    data['postEntityToCSKTripleIndex'][secondEntity].append(index)

    if (not firstEntity in data['postEntityToOtherCSKTripleEntities']):
        data['postEntityToOtherCSKTripleEntities'][firstEntity] = []
    data['postEntityToOtherCSKTripleEntities'][firstEntity].append(data['dict_csk_entities'][secondEntity])
    if (not secondEntity in data['postEntityToOtherCSKTripleEntities']):
        data['postEntityToOtherCSKTripleEntities'][secondEntity] = []
    data['postEntityToOtherCSKTripleEntities'][secondEntity].append(data['dict_csk_entities'][firstEntity])
    index += 1

data['indexToCSKTriple'] = {v: k for k,v in data['dict_csk_triples'].items()}

post_triples = []
all_triples = []
all_entities = []

post = test['post']
index = 0
for word in post:
    try:
        entityIndex = data['dict_csk_entities'][word]
        index += 1
        post_triples.append(index)
        all_triples.append(data['postEntityToCSKTripleIndex'][word])
        all_entities.append(data['postEntityToOtherCSKTripleEntities'][word])
    except:
        post_triples.append(0)
test['post_triples'] = post_triples
test['all_triples'] = all_triples
test['all_entities'] = all_entities

response_triples = []
match_index = []
match_triples = []
for word in test['response']:
    try:
        found = False
        entityIndex = data['dict_csk_entities'][word]
        for index,entitiesList in enumerate(test['all_entities']):
            for subindex,entity in enumerate(entitiesList):
                if(entity == entityIndex):
                    match_index.append([index+1,subindex])
                    response_triples.append(test['all_triples'][index][subindex])
                    match_triples.append(test['all_triples'][index][subindex])
                    found = True
                    break
        if not found:
            response_triples.append(-1)
            match_index.append([-1,-1])
    except:
        response_triples.append(-1)
        match_index.append([-1,-1])

test['response_triples'] = response_triples
test['match_index'] = match_index
test['match_triples'] = match_triples
print(str(test))

Originally posted by @andrewtackett in https://github.com/tuxchow/ccm/issues/3#issuecomment-461907771

Thanks,
Serena