Neural Machine Translation: Using Open-NMT for training a translation model

A complete guide to learn translations between any language pairs This blog is aimed at providing a step by step tutorial to learn to generate translations from a given language to any target language. The methodology we use for the task at hand is entirely motivated by an open source library a pyTorch implementation of which is available in python language, called Open-NMT (Open-Source Neural Machine Translation). It is designed to be research-friendly for deep learning enthusiasts to implement their ideas in the domain of machine translation, summarization, image to text conversion, morphology, etc. Though quite a lot of efficient translation systems by Google Translate, Microsoft, etc are existent, they are either not open-source or are closed under restrictive licence. Other libraries like models are existent for the purpose, but as research code. tensorflow-seq2seq is not only open-source but also provides greatly documented, modular and readable code for fast training and efficient performance of the models. Open-NMT We elaborate further, a detailed guide for setting up the library and using the toolkit for training your own custom translation system. This blog deals with generating translation for Hindi language from given English text. A Brief Overview of the Architecture of Open-NMT Open-NMT is based on the research by Guillaume Klein et al, found . here According to the Paper, the following details are revealed about its architecture: OpenNMT is a complete library for training and deploying neural machine translation models. The system is successor to seq2seq-attn developed at Harvard, and has been completely rewritten for ease of efficiency, readability, and generalizability. It includes vanilla NMT models along with support for attention, gating, stacking, input feeding, regularization, beam search and all other options necessary for state-of-the-art performance. The main system is implemented in the Lua/Torch mathematical framework, and can be easily be extended using Torch’s internal standard neural network components. It has also been extended by Adam Lerer of Facebook Research to support Python/PyTorch framework, with the same API. Setup of Required Modules The chief package required for training your custom translation system is essentially pyTorch, in which the Open-NMT models have been implemented. The priliminary step, of course is to clone the repository: OpenNMT-py git clone http //github. /OpenNMT/OpenNMT- OpenNMT- s: com py cd py Here’s a requirements.txt file to gather all the required packages: six tqdm torch>= git+https: future 0.4 .0 //github.com/pytorch/text Since PyTorch is has been continuously evolving, we recommend forking the PyTorch 0.4 version to ensure a stable performance of the code base. Run the following command to automatically gather pre-requisite dependencies: pip -r requirements.txt install Gather the Datasets The dataset comprises of a parallel corpus of source and target language files containing one sentence per line such that each tokens are separated by a space. For our tutorial, we use a parallel corpora of English and Hindi sentences stored in separate files. The data is gathered from various sources and combined. The data is then re-arranged so as to create a set of files as follows: : Training file containing 10000 English (Source Language) src-train.txt : Training file containing 10000 Hindi (Target Language) sentencestgt-train.txt : Validation data consisting of 1000 English (Source Language) sentencessrc-val.txt : Validation data consisting of 1000 Hindi (Target Language) sentencestgt-val.txt : Test Evaluation data consisting of 1000 English (Source Language) sentencessrc-test.txt : Test Evaluation data consisting of 1000 Hindi (Target Language) sentences sentencestgt-test.txt All of the above files are placed in the /data directory. We have used a limited amount of data for explanation and experimentation in this tutorial. However it is recommended to use a large corpora with millions of sentences to ensure a vast vocabulary of unique words for better learning and a . NOTE: close-to-human-translation Validation data is employed to evaluate the model at each step to identify the convergence point. It should contain a maximum of 5000 sentences typically. Here’s a sample showing how the text data is arranged in the corresponding files: Source Files : They also bring out Tamil weekly newspapers. They are hard — working people most them work labourers. Tamil films are also shown cinema halls. There are quite large Malayalees living here. Target Files :तमिल भाषा में वे अनेक समाचार पत्र व पत्रिकाएं भी निकालते हैं . ये लोग काफी परिश्रमी हैं , अधिकांश लोग मजदूरी करते हैं . स्थानीय सिनेमा हालों में तमिल चलचित्रों का प्रदर्शन अक्सर किया जाता है . मलयालम लोगों की बहुत बडी संख्या है . a number of a and of as in the local a number of Pre-Processing Text Data Execute the following command for pre-processing the training and validation data and extract features for training and generate vocabulary files for the model. python preprocess.py -train_src data/src-train.txt -train_tgt data/tgt-train.txt -valid_src data/src-val.txt -valid_tgt data/tgt-val.txt -save_data data/demo python preprocess -train_src data/src-train -train_tgt data/tgt-train -valid_src data/src-val -valid_tgt data/tgt-val -save_data data/demo .py .txt .txt .txt .txt Training the Translator model The chief command for trianing is really simple to use. Essentially, it takes as input, a data file and a save file. A summary of the default model used is as follows: NMTModel( (encoder): RNNEncoder( (embeddings): Embeddings( (make_embedding): Sequential( (emb_luts): Elementwise( ( ): Embedding( , , padding_idx= ) ) ) ) (rnn): LSTM( , , num_layers= , dropout= ) ) (decoder): InputFeedRNNDecoder( (embeddings): Embeddings( (make_embedding): Sequential( (emb_luts): Elementwise( ( ): Embedding( , , padding_idx= ) ) ) ) (dropout): Dropout(p= ) (rnn): StackedLSTM( (dropout): Dropout(p= ) (layers): ModuleList( ( ): LSTMCell( , ) ( ): LSTMCell( , ) ) ) (attn): GlobalAttention( (linear_in): Linear(in_features= , out_features= , bias= ) (linear_out): Linear(in_features= , out_features= , bias= ) (softmax): Softmax() (tanh): Tanh() ) ) (generator): Sequential( ( ): Linear(in_features= , out_features= , bias= ) ( ): LogSoftmax() ) ) 0 20351 500 1 500 500 2 0.3 0 20570 500 1 0.3 0.3 0 1000 500 1 500 500 500 500 False 1000 500 False 0 500 20570 True 1 train.py - python /demo -save_model demo-model data data The above command will run a default model, comprising a two layer LSTM possessing 500 hidden units for both, the encoder and the decoder. To specify utilization of your GPU for training, specify the -gpuid argument in the above command (say -gpuid 1 for specifying usage og GPU 1). Typically, the default model goes on till 100000 epochs, such that a check-point is saved after every 5000 epochs. So if your model converges and the validation accuracy reaches a stable point earlier, you can stop further training and use the previously saved checkpoint. Translate your own Data The following command may be executed to perform an inference step on unseen text in the Source language (English) and generate corresponding translations predicted: python translate -model demo-model_XYZ -src data/src-test -output pred -replace_unk -verbose .py .pt .txt .txt This will generate the translated output and store the predictions into a file namedpred.txt. The model was trained for 10000 epochs on a NVIDIA GEFORCE 2GB GPU. Training on a CPU will require a very high computational cost, hence it is recommended to use a high end GPU for training the model with a large amount of data at a faster rate. Sample Predictions of the Model Shown below, are couple of examples for the Hindi translations generated for corresponding English sentences after training the model. Trees are capable absorbing more carbon dioxide, thus maintaining equilibrium the air composition of of in PREDICTED : पेडों में कार्बन डाईआक्साइड के बुरे लोग इस प्रकार पेड - पौधे का प्रयोग करने के लिए मौजूद हैं. He has hope that the gods will good things the heavens, upon them throw from PREDICTED :वे उमीद है कि वे घर से कुछ नहीं बची हैं The Buddhist temple , the Dalai Lama Palace and dispensary Tibet are tourist attractions here of PREDICTED :यहां का बौद्ध मंदिर दलाई लामा का आवास तथा तिब्बती औषधालय स्थानिय लोगो में मिलता है . He lets the hair grow long. PREDICTED : वह अपने बढा लेता है . As seen above, the predictions are not good enough yet with less training data to be used for any real time translation. For performing translations close to real world, the model has to be trained on on a large vocabulary and about a million sentences, which will parallely involve a lot of computational cost in terms of hardware requirements and training time. Evaluate your Trained Model Bilingual Evaluation Understudy Score The Bilingual Evaluation Understudy Score, or BLEU Score, refers to an evaluation metric for the purpose of evaluating Machine Translation Systems by comparing a generated sentence to a reference sentence. A perfect match in this comparison results in a BLEU score of 1.0, whereas a complete mismatch results in a BLEU score of 0.0. The BLEU Score is a universally adapted metric for evaluating translation models as it is independent of language, simple to interpret and has high correlation with manual evaluation. The BLEU score was proposed in a research conducted by Kishore Papineni, et al. “ “. BLEU: a Method for Automatic Evaluation of Machine Translation The BLEU SCore is generated after counting n-grams in the candidate translation matching with the n-grams in the reference text. Word order is not taken into account in this comparison. So how would we define a n-gram? Let’s say a 1-gram or uni-gram would mean each individual token and a bi-gram would represent each pair of word. The Code for calculating BLEU Scores, given your predicted candidate file and a reference file given in the GitHub repository, the link to whic is provided at the end of the blog. Here’s how we run the code to evaluate the model: python calculatebleu .py "pred.txt" "tgt-test.txt" Where pred.txt is our candidate predicted translation file and tgt-test.txt is the file containing the actual translations in the target language. Since our data vocabulary generated with 10k sentences consists only of a few thousand words, the BLEU Score we get on our prediction is quite poor (0.025). : Since our primary aim is focussed at elaborating on the usage of Open-NMT, we use only a small dataset, which is why the evaluation of our predicted translation results is a poor BLEU Score. A BLEU Score around 0.5 implies a decent translation. Increase the training vocabulary manifold by adding several thousands of more examples to improve the score. NOTE However Open-NMT allows us to train our own custom translator models between any pair of languages and is very convenient to use. The Code for generating the BLEU Score and the datasets used in training our model have been provided . here If you have any queries or suggestions, I would love to hear about it. Please write to me at abhishek.narayanan@dataturks.com.