Code in this folder contains implementation for the CraigslistBargain task in the following paper: [Decoupling Strategy and Generation in Negotiation Dialogues](https://arxiv.org/abs/1808.09637). He He, Derek Chen, Anusha Balakrishnan and Percy Liang. Empirical Methods in Natural Language Processing (EMNLP), 2018. ## Dependencies Python 2.7, PyTorch 0.4. Install `cocoa`: ``` cd ..; python setup.py develop; ``` `pip install -r requirements.txt` ## Dataset All data is on the Codalab [worksheet](https://worksheets.codalab.org/worksheets/0x453913e76b65495d8b9730d41c7e0a0c/). ### Visualize JSON transcripts All dialogues (either generated by self-play or collected from AMT) are in the same [JSON format](../README.md#examples-and-datasets). To visualize the JSON files in HTML, see documentation [here](../README.md#visualize). For CraigslistBargain dialogues, pass in the additional argument: - `--img-path`: path to Craigslist posting images; download on [Codalab](https://worksheets.codalab.org/bundles/0xb93730d80e1c4d4cb4c6bf7c9ebef12f/). ### Collect your own data If you want to collect your own data, read the following steps. ### Scenario generation 1. Schema: `data/craigslist-schema.json`. 2. Scrape Craigslist posts from different categories: ``` cd scraper; for cat in car phone bike electronics furniture housing; do \ scrapy crawl craigslist -o data/negotiation/craigslist_$cat.json -a cache_dir=/tmp/craigslist_cache -a from_cache=False -a num_result_pages=100 -a category=$cat -a image=1; \ done ``` 3. Generate scenarios: ``` PYTHONPATH=. python scripts/generate_scenarios.py --num-scenarios --schema-path data/craigslist-schema.json --scenarios-path data/scenarios.json --scraped-data scraper/data/negotiation --categories furniture housing car phone bike electronics --fractions 1 1 1 1 1 1 --discounts 0.9 0.7 0.5 ``` - `--fractions`: fractions to sample from each category. - `--discounts`: possible targets for the buyer, `discount * listing_price`. ### Set up the website and AMT HITs. See [data collection](../README.md#data-collection) in `cocoa` README. ## Building the bot ### Use the modular approach The modular framework consists of three parts: the parser, the manager, and the generator. #### 1. Build the price tracker. The price tracker recognizes price mentions in an utterance. ``` PYTHONPATH=. python core/price_tracker.py --train-examples-path data/train.json --output ``` #### 2. Parse the training dialogues. Parse both training and validation data. ``` PYTHONPATH=. python parse_dialogue.py --transcripts data/train.json --price-tracker --max-examples -1 --templates-output templates.pkl --model-output model.pkl --transcripts-output data/train-parsed.json PYTHONPATH=. python parse_dialogue.py --transcripts data/dev.json --price-tracker --max-examples -1 --templates-output templates.pkl --model-output model.pkl --transcripts-output data/dev-parsed.json ``` - Parse utterances into coarse dialogue acts using the rule-based parser (`--transcripts-output`). - Learn an n-gram model over the dialogue acts (`--model-output`), which will be used by the **hybrid policy**. - Extract utterance templates (`--templates-output`) for retrieval-based generator. #### 3. Learning the manager. We train a seq2seq model over the coarse dialogue acts using parsed data. ``` mkdir -p mappings/lf2lf; mkdir -p cache/lf2lf; mkdir -p checkpoint/lf2lf; PYTHONPATH=. python main.py --schema-path data/craigslist-schema.json --train-examples-paths data/train-parsed.json --test-examples-paths data/dev-parsed.json \ --price-tracker price_tracker.pkl \ --model lf2lf \ --model-path checkpoint/lf2lf --mappings mappings/lf2lf \ --word-vec-size 300 --pretrained-wordvec '' '' \ --rnn-size 300 --rnn-type LSTM --global-attention multibank_general \ --num-context 2 --stateful \ --batch-size 128 --gpuid 0 --optim adagrad --learning-rate 0.01 \ --epochs 15 --report-every 500 \ --cache cache/lf2lf --ignore-cache \ --verbose ``` #### 4. Finetune the manager with reinforcement learning. Generate self-play dialogues using the above learned policy and run REINFORCE with a given reward function. First, let's generate the training and validation scenarios. We will directly get those from the training and validation data. ``` PYTHONPATH=. python ../scripts/chat_to_scenarios.py --chats data/train.json --scenarios data/train-scenarios.json PYTHONPATH=. python ../scripts/chat_to_scenarios.py --chats data/dev.json --scenarios data/dev-scenarios.json ``` Now, we can run self-play and REINFORCE with a reward function, e.g. `margin`. ``` mkdir checkpoint/lf2lf-margin; PYTHONPATH=. python reinforce.py --schema-path data/craigslist-schema.json \ --scenarios-path data/train-scenarios.json \ --valid-scenarios-path data/dev-scenarios.json \ --price-tracker price_tracker.pkl \ --agent-checkpoints checkpoint/lf2lf/model_best.pt checkpoint/lf2lf/model_best.pt \ --model-path checkpoint/lf2lf-margin \ --optim adagrad --learning-rate 0.001 \ --agents pt-neural pt-neural \ --report-every 500 --max-turns 20 --num-dialogues 5000 \ --sample --temperature 0.5 --max-length 20 --reward margin ``` - `--reward`: `margin` (utility), `fair` (fairness), and `length` (length). - `--agents`: agent types ### Use the end-to-end approach #### 1. Build pretrained word embeddings. First, build the vocabulary. Note that we need the [price tracker](#price-tracker) to bin prices. ``` mkdir -p mappings/seq2seq; PYTHONPATH=. python main.py --schema-path data/craigslist-schema.json --train-examples-paths scr/data/train.json --mappings mappings/seq2seq --model seq2seq --price-tracker price_tracker.pkl --ignore-cache --vocab-only ``` Get the GloVe embedding. ``` wget http://nlp.stanford.edu/data/glove.840B.300d.zip; unzip glove.840B.300d.zip; ``` Filter pretrained embedding for the model vocab. We use separate embeddings for the utterances and the product description specified by `--vocab-type`. ``` PYTHONPATH=. python ../cocoa/neural/embeddings_to_torch.py --emb-file glove.840B.300d.txt --vocab-file mappings/seq2seq/vocab.pkl --output-file mappings/seq2seq/ --vocab-type kb PYTHONPATH=. python ../cocoa/neural/embeddings_to_torch.py --emb-file glove.840B.300d.txt --vocab-file mappings/seq2seq/vocab.pkl --output-file mappings/seq2seq/ --vocab-type utterance ``` #### 2. Train the seq2seq model. ``` mkdir -p cache/seq2seq; mkdir -p checkpoint/seq2seq; PYTHONPATH=. python main.py --schema-path data/craigslist-schema.json --train-examples-paths data/train.json --test-examples-paths data/dev-parsed.json \ --price-tracker price_tracker.pkl \ --model seq2seq \ --model-path checkpoint/seq2seq --mappings mappings/seq2seq \ --pretrained-wordvec mappings/seq2seq/utterance_glove.pt mappings/seq2seq/kb_glove.pt --word-vec-size 300 \ --rnn-size 300 --rnn-type LSTM --global-attention multibank_general \ --enc-layers 2 --dec-layers 2 --num-context 2 \ --batch-size 128 --gpuid 0 --optim adagrad --learning-rate 0.01 \ --report-every 500 \ --epochs 15 \ --cache cache/seq2seq --ignore-cache \ --verbose ``` #### 3. Finetune with RL. See [finetuning](#rl) in the modular framework. We just need to change the the model path to `--model-path checkpoint/seq2seq`. ## Chat with the bot Chat with the bot in the command line interface: ``` PYTHONPATH=. python ../scripts/generate_dataset.py --schema-path data/craigslist-schema.json --scenarios-path data/dev-scenarios.json --results-paths bot-chat-transcripts.json --max-examples 20 --agents cmd --price-tracker price_tracker.pkl --agent-checkpoints "" --max-turns 20 --random-seed --sample --temperature 0.2 ``` Chat with the bot in the web interface: add the bot model to the config file (example: `web/app_params_allsys.json`) and [launch the website](../README.md#web).