MetaSLU

This project explores the joint learning of multi-task in the few-shot setting, which is the code of the paper Learning to Bridge Metric Spaces: Few-shot Joint Learning of Intent Detection and Slot Filling.

We conduct experiments with two most common tasks in dialog system (Spoken Language Understanding): Intent Detection and Slot Filling.

Result

FewJoint1.0

• 1-shot

Models	Intent Acc.	Slot F1	Joint Acc
SepProto	66.35	27.24	10.92
JointProto	58.52	29.49	9.64
LD-Proto	67.70	27.73	13.7
LD-Proto + TR	67.63	34.06	16.98
ConProm (Ours)	65.26	33.09	16.32
ConProm+TR (Ours)	65.73	37.97	19.57
JointTransfer	41.83	26.89	12.27

• 5-shot

Models	Intent Acc.	Slot F1	Joint Acc
SepProto	75.64	36.08	15.93
JointProto	70.93	39.47	14.48
LD-Proto	78.29	39.88	22.91
LD-Proto + TR	75.75	51.62	27.59
ConProm (Ours)	78.05	39.4	24.18
ConProm+TR (Ours)	75.54	50.28	28.69
JointTransfer	57.5	29	18.81

FewJoint1.1

• 1-shot

Models	Intent Acc.	Slot F1	Joint Acc
SepProto	68.29	33.16	15.71
JointProto	59.41	35.51	15.81
LD-Proto	69.08	34.53	19.44
LD-Proto + TR	67.64	45.20	22.69
ConProm (Ours)	66.51	39.07	21.09
ConProm+TR (Ours)	75.13	55.15	32.54
JointTransfer	45.22	28.98	14.79

• 5-shot

Models	Intent Acc.	Slot F1	Joint Acc
SepProto	76.42	40.08	20.40
JointProto	66.94	40.08	17.95
LD-Proto	77.70	41.14	22.23
LD-Proto + TR	77.01	53.23	27.94
ConProm (Ours)	77.93	43.68	25.07
ConProm+TR (Ours)	77.07	57.11	34.44
JointTransfer	62.08	33.11	23.53

Get Started

Requirement

python >= 3.6
pytorch >= 0.4.1
pytorch_pretrained_bert >= 0.6.1
allennlp >= 0.8.2
pytorch-nlp

Step1: Prepare BERT embedding:

• Download the pytorch bert model, or convert tensorflow param by yourself, or click here: password: atma

• Set BERT path in the file ./scripts/smp_ConProm.sh to your setting:

bert_base_uncased=/your_dir/chinese_L-12_H-768_A-12/
bert_base_uncased_vocab=/your_dir/chinese_L-12_H-768_A-12/vocab.txt

Step2: Prepare data

• For snips, FewJoint1.0 and FewJoint1.1, you can get them in data here.
• Set test, train, dev data file path in ./scripts/smp_ConProm.sh to your setting.

For simplicity, you only need to set the root path for data as follow:

base_data_dir=/your_dir/your_data_dir/result/SMP_data_slu_1/

Step3: Train and test the main model

• Build a folder to collect running log

mkdir result

• Execute cross-evaluation script with two params:
  • [gpu id]
  • [dataset name]
    • snips or smp

Example for 1-shot Snips:

source ./scripts/snips_ConProm.sh 0 snips

Example for 1-shot Smp:

source ./scripts/smp_ConProm.sh 0 smp

To run 5-shots experiments, use ./scripts/snips_ConProm_5.sh

Step4: Check your predict results (if do_predict is set to True)

• For 1-shot and SMP, you can find the results in /your_dir/your_data_dir/result/SMP_data_slu_1/smp.spt_s_1.q_s_16.ep_50.cross_id_0/your_model_name_dir
• For 5-shot and SMP, you can find the results in /your_dir/your_data_dir/result/SMP_data_slu_5/smp.spt_s_5.q_s_8.ep_50.cross_id_0/your_model_name_dir
• For 1-shot or 5-shot about snips, you can find the results in /data/snips according to your model's name

You should read the script carefully to get the model's name

Model for Other Setting

We also provide scripts of four model settings as follows:

• ConProm + TR
• LD-Proto
• LD-Proto + TR
• JointProto
• SepProto (just modify the param task as intent or slot_filling)

You can find their corresponding scripts in ./scripts/ with the same usage as above.

After you execute the script, you can view the specific parameters in the result logs.

Project Architecture

Root

• the project contains three main parts:
  • models: the neural network architectures
  • scripts: running scripts for cross evaluation
  • tools: some tools scrips for get batch score from result files
  • utils: auxiliary or tool function files
  • main.py: the entry of few-shot models

models

• Main Model
  • Sequence Labeler (few_shot_seq_labeler.py): a framework that integrates modules below to perform sequence labeling.
  • Text Classifier (few_shot_text_classifier.py): a framework that integrates modules below to perform text classifier.
  • Joint Slot Filling and Intent Detection (few_shot_slu.py): a framework that integrates modules below to perform joint slot filling and intent detection.
  • Model with Fine-tune (normal_slu.py): a framework that integrates modules below to perform classical supervised model with support set.
• Modules
  • Embedder Module (context_embedder_base.py): modules that provide embeddings.
  • Emission Module (emission_scorer_base.py): modules that compute emission scores.
  • Transition Module (transition_scorer.py): modules that compute transition scores.
  • Similarity Module (similarity_scorer_base.py): modules that compute similarities for metric learning based emission scorer.
  • Output Module (seq_labeler.py, conditional_random_field.py): output layer with normal mlp or crf.
  • Scale Module (scale_controller.py): a toolkit for re-scale and normalize logits.

utils

• utils contains assistance modules for:
  • data processing (data_loader.py, preprocessor.py),
  • constructing model architecture (model_helper.py),
  • controlling training process (trainer.py),
  • controlling testing process (tester.py),
  • controllable parameters definition (opt.py),
  • device definition (device_helper)
  • config (config.py).

tools

• get batch scores from result files(count_score.py)
  • usage

   python count_score.py [snip/smp] [result_file_path - without $cross_id.m_seed_$seed]

  • example

   python count_score.py smp ../data/smp_release_new/smp.spt_s_1.q_s_16.ep_50.cross_id_0/ds_SepProto.intent.bert.ga_4.bs_4.ep_4.lr_0.00001.up_lr_0.001.dec_sms.DATA.smp.shots_1.cross_id_

• get semantic acc score for SepProto(count_semantic_acc_for_sep_proto.py)
  • usage

   python count_semantic_acc_for_sep_proto.py [snip/smp] [result_file_path - without $cross_id.m_seed_$seed]

  • example

   python count_semantic_acc_for_sep_proto.py smp ../data/smp_release_new/smp.spt_s_1.q_s_16.ep_50.cross_id_0/ds_SepProto.intent.bert.ga_4.bs_4.ep_4.lr_0.00001.up_lr_0.001.dec_sms.DATA.smp.shots_1.cross_id_

• get slot acc from result files(cal_sentence_level_slot_acc.py)
  • usage

   python cal_sentence_level_slot_acc.py [snip/smp] [result_file_path - without $cross_id.m_seed_$seed]

  • example

   python cal_sentence_level_slot_acc smp ../data/smp_release_new/smp.spt_s_1.q_s_16.ep_50.cross_id_0/ds_SepProto.intent.bert.ga_4.bs_4.ep_4.lr_0.00001.up_lr_0.001.dec_sms.DATA.smp.shots_1.cross_id_

• format batch scores as (format_res.py)
  • usage
    • copy the result scores get from count_score.py in test.txt
    • run follow scripts:

     python format_res

    • get formatted result scores in res.txt

Important Parameters (you can get more detailed information in /utils/opt.py)

Task:

• intent for Intent Detection
• slot_filling for Slot Filling
• slu for Joint Slot Filling and Intent Detection

Few-Shot Models

• slu_model_type: few-shot model type
• proto_merge_type: ProtoMerge methods
• split_metric: the type of SplitMetric module
• proto_update: the method for merging ProtoMerge methods
• slu_regular: the type of regular loss(use metric learning method - contrastive learning)
• margin: the margin in Metric Learning
• with_homo: add contrastive loss between prototypes in sampe task
• seed: seed
• epoch: epoch of training
• decoder: the type of decoder in Slot Filling
  • sms: simple sequence decoder
  • rule: use simple transition rule decoder