/tensorflow-nlp

Building blocks for NLP and Text Generation in TensorFlow 2.x / 1.x

Primary LanguageJupyter NotebookMIT LicenseMIT

	Code has been run on Google Colab, thanks Google for providing computational resources

Contents

  • Natural Language Processing（自然语言处理）

    • Text Classification（文本分类）

      • IMDB（English Data）

         Abstract:
 
 1. we show the classic ML model (tfidf + logistic regression) is able to reach 89.6%

    which is decent for its simplicity, efficiency and low-cost

 2. we show FastText model is able to reach 90% accuracy

 3. we show cnn-based model is able to improve the accuracy to 91.7%

 4. we show rnn-based model is able to improve the accuracy to 92.6%

 5. we show pretrained model (bert) is able to improve the accuracy to 94%

 6. we show pretrained model (roberta) is able to improve the accuracy to 94.7%

 7. we use back-translation, label smoothing, cyclical lr as training helpers

    • Text Matching（文本匹配）

      • SNLI（English Data）

         Abstract:

 1. we show dam (lots of interact) is able to reach 85.3% accuracy

 2. we show pyramid (rnn + image processing) is able to improve the accuracy to 87.1%

 3. we show esim (rnn + lots of interact) is able to improve the accuracy to 87.4%

 4. we show re2 (rnn + lots of interact + residual) is able to improve the accuracy to 88.3%

 5. we show bert (pretrained model) is able to improve the accuracy to 90.4%

 6. we show roberta (pretrained model) is able to improve the accuracy to 91.1%

 7. we use label smoothing and cyclical lr as training helpers

      • 微众银行智能客服（Chinese Data）

         Abstract:

 1. we show esim, pyramid, re2 are able to reach 82.5% ~ 82.9% accuracy (very close)

 2. we show re2 is able to be improved to 83.8% by using cyclical lr and label smoothing

 3. we show bert (pretrained model) is able to further improve the accuracy to 84.75%

 4. we show char-level re2 actually performs better than word-level re2 on this dataset

    • Spoken Language Understanding（对话理解）

      • ATIS（English Data）

    • Generative Dialog（生成式对话）

      • Large-scale Chinese Conversation Dataset

    • Multi-turn Dialogue Rewriting（多轮对话改写）

      • 20k 腾讯 AI 研发数据（Chinese Data）

         Highlight:
 
 1. our implementation of rnn-based pointer network reaches 60% exact match without bert

    which is higher than other implementations using bert

    e.g. (https://github.com/liu-nlper/dialogue-utterance-rewriter) 57.5% exact match

 2. we show how to deploy model in java production

 3. we explain this task can be decomposed into two stages (extract keywords & recombine query)

    the first stage is fast (tagging) and the second stage is slow (autoregressive generation)

    for the first stage, we show birnn extracts keywords at 79.6% recall and 42.6% exact match

    then we show bert is able to improve this extraction task to 93.6% recall and 71.6% exact match
 
 4. we find that we need to predict the intent as well (whether to rewrite the query or not)

    in other words, whether to trigger the rewriter or not at the first place

    we have finetuned a bert to jointly predict intent and extract the keywords
    
    the result is: 97.9% intent accuracy; 90.2% recall and 64.3% exact match for keyword extraction

    • Semantic Parsing（语义解析）

      • Facebook AI Research Data（English Data）

         Highlight:
 
 our implementation of pointer-generator reaches 80.3% exact match on testing set

 which is higher than all the results of the original paper including rnng (78.5%)

 we further improve exact match to 81.1% by adding more embeddings (char & contextual)

 (https://aclweb.org/anthology/D18-1300)

    • Question Answering（问题回答）

      • bAbI（Engish Data）

    • Text Processing Tools（文本处理工具）

  • Knowledge Graph（知识图谱）

  • Recommender System（推荐系统）

    • Movielens 1M（English Data）

Text Classification

└── finch/tensorflow2/text_classification/imdb
	│
	├── data
	│   └── glove.840B.300d.txt          # pretrained embedding, download and put here
	│   └── make_data.ipynb              # step 1. make data and vocab: train.txt, test.txt, word.txt
	│   └── train.txt  		     # incomplete sample, format <label, text> separated by \t 
	│   └── test.txt   		     # incomplete sample, format <label, text> separated by \t
	│   └── train_bt_part1.txt  	     # (back-translated) incomplete sample, format <label, text> separated by \t
	│
	├── vocab
	│   └── word.txt                     # incomplete sample, list of words in vocabulary
	│	
	└── main              
		└── attention_linear.ipynb   # step 2: train and evaluate model
		└── attention_conv.ipynb     # step 2: train and evaluate model
		└── fasttext_unigram.ipynb   # step 2: train and evaluate model
		└── fasttext_bigram.ipynb    # step 2: train and evaluate model
		└── sliced_rnn.ipynb         # step 2: train and evaluate model
		└── sliced_rnn_bt.ipynb      # step 2: train and evaluate model

Text Matching

└── finch/tensorflow2/text_matching/snli
	│
	├── data
	│   └── glove.840B.300d.txt       # pretrained embedding, download and put here
	│   └── download_data.ipynb       # step 1. run this to download snli dataset
	│   └── make_data.ipynb           # step 2. run this to generate train.txt, test.txt, word.txt 
	│   └── train.txt  		  # incomplete sample, format <label, text1, text2> separated by \t 
	│   └── test.txt   		  # incomplete sample, format <label, text1, text2> separated by \t
	│
	├── vocab
	│   └── word.txt                  # incomplete sample, list of words in vocabulary
	│	
	└── main              
		└── dam.ipynb      	  # step 3. train and evaluate model
		└── esim.ipynb      	  # step 3. train and evaluate model
		└── ......


└── finch/tensorflow2/text_matching/chinese
	│
	├── data
	│   └── make_data.ipynb           # step 1. run this to generate char.txt and char.npy
	│   └── train.csv  		  # incomplete sample, format <text1, text2, label> separated by comma 
	│   └── test.csv   		  # incomplete sample, format <text1, text2, label> separated by comma
	│
	├── vocab
	│   └── cc.zh.300.vec             # pretrained embedding, download and put here
	│   └── char.txt                  # incomplete sample, list of chinese characters
	│   └── char.npy                  # saved pretrained embedding matrix for this task
	│	
	└── main              
		└── pyramid.ipynb      	  # step 2. train and evaluate model
		└── esim.ipynb      	  # step 2. train and evaluate model
		└── ......

Spoken Language Understanding

└── finch/tensorflow2/spoken_language_understanding/atis
	│
	├── data
	│   └── glove.840B.300d.txt           # pretrained embedding, download and put here
	│   └── make_data.ipynb               # step 1. run this to generate vocab: word.txt, intent.txt, slot.txt 
	│   └── atis.train.w-intent.iob       # incomplete sample, format <text, slot, intent>
	│   └── atis.test.w-intent.iob        # incomplete sample, format <text, slot, intent>
	│
	├── vocab
	│   └── word.txt                      # list of words in vocabulary
	│   └── intent.txt                    # list of intents in vocabulary
	│   └── slot.txt                      # list of slots in vocabulary
	│	
	└── main              
		└── bigru_clr.ipynb               # step 2. train and evaluate model
		└── bigru_self_attn_clr.ipynb     # step 2. train and evaluate model
		└── bigru_clr_crf.ipynb           # step 2. train and evaluate model

Generative Dialog

└── finch/tensorflow1/free_chat/chinese_lccc
	│
	├── data
	│   └── LCCC-base.json           	# raw data downloaded from external
	│   └── LCCC-base_test.json         # raw data downloaded from external
	│   └── make_data.ipynb           	# step 1. run this to generate vocab {char.txt} and data {train.txt & test.txt}
	│   └── train.txt           		# processed text file generated by {make_data.ipynb}
	│   └── test.txt           			# processed text file generated by {make_data.ipynb}
	│
	├── vocab
	│   └── char.txt                	# list of chars in vocabulary for chinese
	│   └── cc.zh.300.vec			# fastText pretrained embedding downloaded from external
	│   └── char.npy			# chinese characters and their embedding values (300 dim)	
	│	
	└── main
		└── lstm_seq2seq_train.ipynb    # step 2. train and evaluate model
		└── lstm_seq2seq_export.ipynb   # step 3. export model
		└── lstm_seq2seq_infer.ipynb    # step 4. model inference
		└── transformer_train.ipynb     # step 2. train and evaluate model
		└── transformer_export.ipynb    # step 3. export model
		└── transformer_infer.ipynb     # step 4. model inference

  • Task: Chinese Conversation Dataset

      Training Data: 2000000, Testing Data: 19008

    • Data

    • Model: RNN Seq2Seq + Attention

    • Model: Transformer

    • If you want to deploy model in Java production

       └── FreeChatInference
 	│
 	├── data
 	│   └── transformer_export/
 	│   └── char.txt
 	│   └── libtensorflow-1.14.0.jar
 	│   └── tensorflow_jni.dll
 	│
 	└── src              
 		└── ModelInference.java

      • <Notebook> Java Inference

      • If you don't know the input and output node names in Java, you can call:

         !saved_model_cli show --dir ../model/xxx/1587959473/ --tag_set serve --signature_def serving_default

        which will display the node names:

         The given SavedModel SignatureDef contains the following input(s):
 inputs['history'] tensor_info:
 	dtype: DT_INT32
 	shape: (-1, -1, -1)
 	name: history:0
 inputs['query'] tensor_info:
 	dtype: DT_INT32
 	shape: (-1, -1)
 	name: query:0
 The given SavedModel SignatureDef contains the following output(s):
 outputs['output'] tensor_info:
 	dtype: DT_INT32
 	shape: (-1, -1)
 	name: Decoder/decoder/transpose_1:0
 Method name is: tensorflow/serving/predict

Semantic Parsing

└── finch/tensorflow2/semantic_parsing/tree_slu
	│
	├── data
	│   └── glove.840B.300d.txt     	# pretrained embedding, download and put here
	│   └── make_data.ipynb           	# step 1. run this to generate vocab: word.txt, intent.txt, slot.txt 
	│   └── train.tsv   		  	# incomplete sample, format <text, tokenized_text, tree>
	│   └── test.tsv    		  	# incomplete sample, format <text, tokenized_text, tree>
	│
	├── vocab
	│   └── source.txt                	# list of words in vocabulary for source (of seq2seq)
	│   └── target.txt                	# list of words in vocabulary for target (of seq2seq)
	│	
	└── main
		└── lstm_seq2seq_tf_addons.ipynb           # step 2. train and evaluate model
		└── ......

Knowledge Graph Inference

└── finch/tensorflow2/knowledge_graph_completion/wn18
	│
	├── data
	│   └── download_data.ipynb       	# step 1. run this to download wn18 dataset
	│   └── make_data.ipynb           	# step 2. run this to generate vocabulary: entity.txt, relation.txt
	│   └── wn18  		          	# wn18 folder (will be auto created by download_data.ipynb)
	│   	└── train.txt  		  	# incomplete sample, format <entity1, relation, entity2> separated by \t
	│   	└── valid.txt  		  	# incomplete sample, format <entity1, relation, entity2> separated by \t 
	│   	└── test.txt   		  	# incomplete sample, format <entity1, relation, entity2> separated by \t
	│
	├── vocab
	│   └── entity.txt                  	# incomplete sample, list of entities in vocabulary
	│   └── relation.txt                	# incomplete sample, list of relations in vocabulary
	│	
	└── main              
		└── distmult_1-N.ipynb    	# step 3. train and evaluate model

Knowledge Graph Tools

Knowledge Base Question Answering

  • Rule-based System（基于规则的系统）

    For example, we want to answer the following questions:

     	宝马是什么?  /  what is BMW?
     	我想了解一下宝马  /  i want to know about the BMW
     	给我介绍一下宝马  /  please introduce the BMW to me
 	宝马这个牌子的汽车怎么样?  /  how is the car of BMW group?
     	宝马如何呢?  /  how is the BMW?
     	宝马汽车好用吗?  /  is BMW a good car to use?
     	宝马和奔驰比怎么样?  /  how is the BMW compared to the Benz?
     	宝马和奔驰比哪个好?  /  which one is better, the BMW or the Benz?
     	宝马和奔驰比哪个更好?  /  which one is even better, the BMW or the Benz?

Question Answering

└── finch/tensorflow1/question_answering/babi
	│
	├── data
	│   └── make_data.ipynb           		# step 1. run this to generate vocabulary: word.txt 
	│   └── qa5_three-arg-relations_train.txt       # one complete example of babi dataset
	│   └── qa5_three-arg-relations_test.txt	# one complete example of babi dataset
	│
	├── vocab
	│   └── word.txt                  		# complete list of words in vocabulary
	│	
	└── main              
		└── dmn_train.ipynb
		└── dmn_serve.ipynb
		└── attn_gru_cell.py

Text Processing Tools

Recommender System

└── finch/tensorflow1/recommender/movielens
	│
	├── data
	│   └── make_data.ipynb           		# run this to generate vocabulary
	│
	├── vocab
	│   └── user_job.txt
	│   └── user_id.txt
	│   └── user_gender.txt
	│   └── user_age.txt
	│   └── movie_types.txt
	│   └── movie_title.txt
	│   └── movie_id.txt
	│	
	└── main              
		└── dnn_softmax.ipynb
		└── ......

Multi-turn Dialogue Rewriting

└── finch/tensorflow1/multi_turn_rewrite/chinese/
	│
	├── data
	│   └── make_data.ipynb         # run this to generate vocab, split train & test data, make pretrained embedding
	│   └── corpus.txt		# original data downloaded from external
	│   └── train_pos.txt		# processed positive training data after {make_data.ipynb}
	│   └── train_neg.txt		# processed negative training data after {make_data.ipynb}
	│   └── test_pos.txt		# processed positive testing data after {make_data.ipynb}
	│   └── test_neg.txt		# processed negative testing data after {make_data.ipynb}
	│
	├── vocab
	│   └── cc.zh.300.vec		# fastText pretrained embedding downloaded from external
	│   └── char.npy		# chinese characters and their embedding values (300 dim)	
	│   └── char.txt		# list of chinese characters used in this project 
	│	
	└── main              
		└── baseline_lstm_train.ipynb
		└── baseline_lstm_export.ipynb
		└── baseline_lstm_predict.ipynb

  • Task: 20k 腾讯 AI 研发数据（Chinese Data）

     data split as: training data (positive): 18986, testing data (positive): 1008

 Training data = 2 * 18986 because of 1:1 Negative Sampling

    • <Text File>: Full Data

    • <Notebook>: Make Data & Vocabulary & Pretrained Embedding

        There are six incorrect data and we have deleted them

    • Model: RNN Seq2Seq + Attention + Multi-hop Memory

    • Model: RNN Pointer Networks

      • TensorFlow 1

        Pointer Net returns probability distribution, therefore no need to do softmax again in beam search

        Go to beam search source code, replace this line

         step_log_probs = nn_ops.log_softmax(logits)

        with this line

         step_log_probs = math_ops.log(logits)

    • If you want to deploy model

      • Python Inference（基于 Python 的推理）

      • Java Inference（基于 Java 的推理）

        • <Notebook> Inference

           └── MultiDialogInference
 	│
 	├── data
 	│   └── baseline_lstm_export/
 	│   └── char.txt
 	│   └── libtensorflow-1.14.0.jar
 	│   └── tensorflow_jni.dll
 	│
 	└── src              
 		└── ModelInference.java

    • Despite End-to-End, this problem can also be decomposed into two stages

      • Stage 1 (Fast). Detecting the (missing or referred) keywords from the context

        which is a sequence tagging task with sequential complexity O(1)

      • Stage 2 (Slow). Recombine the keywords with the query based on language fluency

        which is a sequence generation task with sequential complexity O(N)

         For example, for a given query: "买不起" and the context: "成都房价是多少 不买就后悔了成都房价还有上涨空间"

 First retrieve the keyword "成都房" from the context which is very important

 Then recombine the keyword "成都房" with the query "买不起" which becomes "买不起成都房"

      • For Stage 1 (sequence tagging for retrieving the keywords), the experiment results are:

    • However, there is still a practical problem to prefine whether the query needs to be rewritten or not

      • if not, we just simply skip the rewriter and pass the query to the next stage

      • there are actually three situations needs to be classified

        • 0: the query does not need to be rewritten because it is irrelevant to the context

           你喜欢五月天吗	超级喜欢阿信	中午出去吃饭吗

        • 1: the query needs to be rewritten

           你喜欢五月天吗	超级喜欢阿信	你喜欢他的那首歌 -> 你喜欢阿信的那首歌

        • 2: the query does not need to be rewritten because it already contains enough information

           你喜欢五月天吗	超级喜欢阿信	你喜欢阿信的那首歌

      • therefore, we aim for training the model to jointly predict:

        • intent: three situations {0, 1, 2} whether the query needs to be rewritten or not

        • keyword extraction: extract the missing or referred keywords in the context

      • <Text File>: Positive Data Example

      • <Text File>: Negative Data Example

      • <Notebook> BERT (chinese_base)

        -> Intent: 97.9% accuracy

        -> Keyword Extraction: 90.2% recall   80.7% precision   64.3% exact match