Yi Zhu et al. provide a general framework for training subword-informed word representations by varying the following components:

• subword segmentation methods;
• subword embeddings and position embeddings;
• composition functions;

For the whole framework architecture and more details, please refer to the reference.

There are 4 segmentation methods, 3 possible ways of embedding subwords, 3 ways of enhancing with position embeddings, and 3 different composition functions.

Here is a full table of different options and their labels:

For example, sms.wwppmtxatt means they use CHIPMUNK as segmentation, insert word token into the subword sequence, enhance with additive position embedding, and use multi-head self-attention as composition function.

Taking the word dishonestly as an example, with different segmentation methods, the word will be segmented into the following subword sequence:

• ChipMunk: (<dis, honest, ly>) + (PREFIX, ROOT, SUFFIX) - Supervised segmentation
• Morfessor: (<dishonest, ly>) - Unsupervised segmentation
• BPE (10k merge ops): (<dish, on, est, ly>) - Unsupervised segmentation
• Character n-gram (from 3 to 6): (<di, dis, ... , ly>, <dis, ... ,tly>, <dish, ... , stly>, <disho, ... , estly>) - Unsupervised segmentation

where < and > are word start and end markers.

After the segmentation, we will obtain a subword sequence S for each segmentation method, and another morphortactic tag sequence T for sms.

We can embed the subword sequence S directly into subword embedding sequence by looking up in the subword embedding matrix, or insert a word token (ww) into S before embedding, i.e. for sms it will be (<dis, honest, ly>, <dishonestly>).

Then we can enhance the subword embeddings with additive (pp) or elementwise (mp) multiplication.

For sms, we can also embed the concatenation of the subword and its morphortactic tags (wp): (<dis:PREFIX, honest:ROOT, ly>:SUFFIX). And <dishonest>:WORD will be inserted if we choose ww. Note that position embeddings are not applicable to wp as a kind of morphological position information has already been provided.

Data is obtained from IndicCorp Monolingual corpus from here. They have scrapped the data from multiple sources like news, magazines, books, etc. The data consists of 41M sentences and around 719M sentences. I create a vocabulary of 157898 words which occured more than 100 times in the corpus. This vocabulary can be found in toy_data/gu/gu.sent.1m.5.word. I couldn't train the model on entire data because of lack of resources. I tested my models on a subset of 5L sentences.

The data was preprocessed to remove punctuations and normalize the text. The code for that can be found in code\preprocess.ipynb and code\preprocess_2.ipynb notebooks.

Specialized files were needed to run different segmentation methods.

A BPE trained embedding model was needed for our language. This was obtained from here and can be found at toy_data/gu/gu.wiki.bpe.vs10000.model in this repo.

A model was trained for getting the morfessor embeddings following the instruction found here. The resulting segments were processed and saved at the file toy_data/gu/gu.sent.1m.5.morf

The segments were directly prepared as the language model was trained.

I didn't have a supervised data for Gujarati for this segmentation and this wasn't prepared.

The models were trained by running ./run.sh gu SEGM_METH COMP_FUNC LR BS where SEGM_METHOD is the segmentation method(can be bpe, morf, charn). COMP_FUNC is the function used for composition mentioned in the table above. It can be add, att, mtxatt. Postion embedding information can be mentioned here as well. wwppatt would mean subword embedding with word(ww), with addition of postion embeddings(pp) and single self-attention(att) as a composition function. LR is the learning rate. BS is the batch-size. An example of the train configurations can be seen as follows:

./run.sh gu bpe add 0.01 1000

Models were evaluated on a word-similarity task. In this task, model predicted embeddings are used to find the similarity between two words and this similarity is compared with rated similarity. These word pairs are previously labelled for their similarity scores. The labelled data for Gujarati word similarity task was obtained from here. This dataset needs correction and I am focussing on that as of now. A cosine similarity was measured between the word embeddings and scaled to 10 as our model predicted thresholds.

The code for evaluating these models can be found in code/evaulate.ipynb notebook.

Results of some of the models can be seen as follows:

• A Systematic Study of Leveraging Subword Information for Learning Word Representations. Yi Zhu, Ivan Vulić, and Anna Korhonen. In Proc. of NAACL 2019.