Assister Discovery

Assister discovery is a software that maps user requests to executable commands in Assister Pipeline, illustrated in the following figure. Using Natural Language Understanding coupled with Machine Learning over the Terms and Functions contextual annotations embedded in an application, a discovery can translate a request to the corresponding command. The command will then be executed within the Pipeline. [Link to the full proposal]


Problem Statement and Solution

The research task corresponding to the aforementioned real problem at Assister, can be formally defined as:

How can we map a user request in natural language to a pre-defined executable command?

If we had several mappings from user sentences to the requested commands, we would choose a supervised learning approach. This is not a realistic assumption at the current stage, although we know the possible commands (classes) at each context that can be selected for execution. So, we propose an unsupervised methology to tacke this probelm, as follows:

  1. Representation learning of user requests via word embeddings.
  2. Finding the best embedding of each command description.
  3. Mapping a request to a command with the shortest distance in the embedding space based on a similarity measure.

It is obvious that word/sentence-level embedding algorithm is the key part of the solution. So, we first review the literature for this natural language processing (NLP) task. Then, we describe the two recent state-of-the-art embedding models and how Assister utilizes them to embed the requests and comments. Finally, we explain the common similarity measures and pick a suitable one for Assister to calculate the distance between any pair of embeddings.

Word Embeddings

Undoubtedly, the year 2018 has been an inflection point for NLP after being relatively stationary for a couple of years. Word embedding is definitely one of the most popular representation of document vocabulary with the capability of capturing words' context) in a document, syntactic and semantic similarity, relation between different words, etc. More formally, embeddings are low-dimensional representations of a data point (sample) in a higher-dimensional vector space. In the same way, word embeddings are dense vectorized representations of words in a low-dimensional space. The first neural network-based word embedding model was first proposed by Google in 2013 [1]. Since then, word embedding has received a lot of attention in almost every NLP model in practice. They are very effective, because by translating a word to an embedding one can model the semantic importance of a word in a numeric form and thus perform many mathematical operations on it. To make it more clear, let's take a look at a common example in the literature:

Let $\phi$ be a word embedding mapping $W \rightarrow \mathbb{R}^n$, where $W$ is the word space and $\mathbb{R}^n$ is an $n$-dimensional vector space, then we have:

$\phi(''king'') - \phi(''man'') + \phi(''woman'') = \phi(''queen'')$

It was first introduced by the word2vec [2] model in 2013 that was a great breakthrough. Another fascinating word embedding model was Glove [3] in 2014. Although these two models are powerful, they are context-free in which a single word embedding representation for each word in the vocabulary is generated. So, bank would have the same representation in bank deposit and river bank. Instead, contextual models - including Semi-supervised Sequence Learning (2015) [4], Generative Pre-Training (2018) [5], ELMO (2018) [6], ULMFit (2018) [7] - generate a representation of each word based on the other words in the sentence, so they can capture both a static semantic meaning and a contextualized meaning. For instance, the word apple in the two sentences I like apples and I like Apple macbooks has a different semantic meaning, thus the embedding of this word would have a different vector representation which makes it more powerful for NLP tasks. The two recent state-of-the-art models - USE (2018) [8] and BERT (2018) [9] - use a powerful sequence transductive model for language understanding, called Transformer (2017) [10]. We first review the Transformer model, then describe how the USE and the BERT models take the advantage of using Transformer as a building block.


Today's NLP world benefits from the recent advancements of Deep Learning research. More specifically, Google introduced a novel neural network architecture, called Transformer, in a seminal paper [10] which outperformed many traditional Recurrent Neural Network (RNN) sequence models (like LSTM and GRU). The main advantages of using transformer as a language understanding unit is that (1) it can effectively model the long-term dependencies among words in a temporal word sequence; and (2) its model training phase is efficient by eliminating the sequential dependency on previous words [10].

A transformer is an encoder-decoder architecture model that uses attention mechanisms to forward a complex pattern of the whole sequence to the decoder at once rather than sequentially as depicted in the following figure [source]:


Universal Sequence Encoder (USE)

The USE uses a transformer to provide sentence-level embeddings as easy as it has historically been to look up the embeddings for individual words, $e.g.$ word2vec. The universal sentence encoder is a model that encodes a text into 512-dimensional embeddings. The resulted embeddings can then be used as inputs to NLP tasks such as sentiment classification and textual similarity analysis. Pre-trained word embeddings are considered to be an integral part of modern NLP systems, offering significant improvements over embeddings learned from scratch. A pre-trained and optimized USE model on a variety of data sources is publicly available on TenforFlow Hub. This module is about 800MB and depending on your network speed it might take a whileto load the first time you instantiate it. After that, loading the module should be faster as modules are cached by default. Now, we use the universal sequence encoder's TF Hub module to compute a representation (embedding) for user requests and command descriptions in an online spreadsheet application, over TensorFlow platform.

In [142]:
import tensorflow as tf
import tensorflow_hub as hub
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import re
import os

# Import the USE's TF Hub module
module_url = ""
# Another pre-trained model is ""
module = hub.Module(module_url)

# Compute embeddings for sentences (either of user requests or command descriptions)
requests = ["Format cell A12 as date",
            "Sum the values in column B and store the result in cell C7",
            "Delete cell D20",
            "Delete row 15",
            "Select cells from B16 to E19"]
commands = ["To Format a cell, you need a cell, like B10, and a type, like date",
            "To sum a column, you need a column, like C, and a cell, like A5, to store the result",
            "To delete a cell, you need a cell, like C11",
            "To delete a row, you need a row, like 12",
            "To select cells, you need a start cell, like B7, and an end cell, like D17"]
sentences = requests + commands

# Run the embeddings in a TensorFlow session
with tf.Session() as session:[tf.global_variables_initializer(), tf.tables_initializer()])
    sentence_embeddings =
    for i, sentence_embedding in enumerate(np.array(sentence_embeddings).tolist()):
        print("Sentence: {}".format(sentences[i]))
        print("Embedding size: {}".format(len(sentence_embedding)))
        embedding_short = ", ".join((str(x) for x in sentence_embedding[:3]))
        print("Embedding: [{}, ...]\n".format(embedding_short))
INFO:tensorflow:Saver not created because there are no variables in the graph to restore
I0426 22:25:57.912678 140734900516288] Saver not created because there are no variables in the graph to restore
Sentence: Format cell A12 as date
Embedding size: 512
Embedding: [-0.050871364772319794, -0.0016370579833164811, 0.022183720022439957, ...]

Sentence: Sum the values in column B and store the result in cell C7
Embedding size: 512
Embedding: [-0.03197444975376129, 0.02532351016998291, 0.010689766146242619, ...]

Sentence: Delete cell D20
Embedding size: 512
Embedding: [-0.038657110184431076, 0.049386851489543915, 0.000588558497838676, ...]

Sentence: Delete row 15
Embedding size: 512
Embedding: [-0.04079856723546982, 0.020178884267807007, -0.03739866614341736, ...]

Sentence: Select cells from B16 to E19
Embedding size: 512
Embedding: [-0.024356713518500328, 0.029383739456534386, 0.012827333062887192, ...]

Sentence: To Format a cell, you need a cell, like B10, and a type, like date
Embedding size: 512
Embedding: [-0.032455652952194214, 0.04132832959294319, 0.012323660776019096, ...]

Sentence: To sum a column, you need a column, like C, and a cell, like A5, to store the result
Embedding size: 512
Embedding: [-0.05170544236898422, 0.02925257198512554, -0.007931080646812916, ...]

Sentence: To delete a cell, you need a cell, like C11
Embedding size: 512
Embedding: [0.03696422278881073, 0.019425714388489723, 0.0004833970742765814, ...]

Sentence: To delete a row, you need a row, like 12
Embedding size: 512
Embedding: [-0.07475815713405609, 0.038178637623786926, -0.04855307564139366, ...]

Sentence: To select cells, you need a start cell, like B7, and an end cell, like D17
Embedding size: 512
Embedding: [-0.02912534400820732, 0.047603789716959, 0.00011833444295916706, ...]

Semantic Similarity between Requests and Commands via the USE

The embeddings produced by the Universal Sentence Encoder are approximately normalized. The semantic similarity between any pair of user requests and commands descriptions, which can be simply the inner product of the encodings, could be an informative analysis in Assister discovery. Inner product space is a proper metric in our case, as it satisfies the three well-known axioms: Conjugate symmetry, Linearity property, and Positive-definite.

In [144]:
def plot_similarity(labels, embeddings, rotation):
  inner = np.inner(embeddings[:len(requests)], embeddings[len(requests):])
  g = sns.heatmap(
  g.set_xticklabels(labels[len(requests):], rotation=rotation)
  g.set_title("Semantic Similarity based on Inner Product of Embeddings")

def run_and_plot(session_, input_tensor_, messages_, encoding_tensor):
  message_embeddings_ =, feed_dict={input_tensor_: messages_})
  plot_similarity(messages_, message_embeddings_, 90)

Now, we plot the similarity in a heat map. It is a matrix with No of requests rows and No of commands columns, where each entry $[i, j]$ is colored based on the inner product of the embeddings for user request $i$ and command description $j$. In this test, we use the aforementioned examples in the requests and commands lists, which are matched one-to-one for simplicity. The higher the score at each row (request) is, the more similar the corresponding request will be. When we look at each user request (a row), we expect to have a higher similarity score with its related command, which is in our case the diagonal entries of the heat map. For example, the second request (2nd row) about column summation is more similar to the second command description with distance value $[2, 2] = 0.9$.

In [145]:
similarity_input_placeholder = tf.placeholder(tf.string, shape=(None))
similarity_message_encodings = module(similarity_input_placeholder)
with tf.Session() as session:
  run_and_plot(session, similarity_input_placeholder, sentences, similarity_message_encodings)
INFO:tensorflow:Saver not created because there are no variables in the graph to restore
I0426 22:26:58.279740 140734900516288] Saver not created because there are no variables in the graph to restore

Bidirectional Encoder Representations from Transformer (BERT)

BERT [9] is a method of pre-training language representations in which one can train a general-purpose language understanding model on a relatively large text corpus (like Google or Wikipedia), then use the trained model for downstream NLP tasks. BERT is the first unsupervised, deeply bidirectional contextual system for pre-training natural languages, as opposed to context-free models (like word2vec and Glove) and undirectional/shallowly bidirectional contextual models (like Semi-supervised Sequence Learning, Generative Pre-Training, ELMO, ULMFit). For example, in the sentence I made a bank deposit the unidirectional representation of bank is only based on I made a but not deposit. BERT uses deep bidirectional Transformer to represent bank based on both its left and right context - I made a ... deposit. To overcome the see itself issue, Google's BERT employed masked language modeling, in which it hides 15\% of the words and uses their position information to infer them. An interesting issue with this methodology is that it has a negative impact on convergence time, but it outperforms the state-of-the-art models before convergence.

BERT proposes two different model sizes, containing the number of layers (i.e. Transformer blocks) as $L$, the hidden size as $H$, and the number of self-attention heads as $A$:

  • BERTBASE: $L=12$, $H=768$, $A=12$, Total parameters $=110M$
  • BERTLARGE: $L=24$, $H=1024$, $A=16$, Total parameters $=340M$

A very high-level architecture of BERTBASE looks like this:


Each encoder block encapsulates a sophisticated model architecture. A visual representation of BERT input is as follows:


The four embedding layers are [9]:

  • The input layer is the vector of the sequence tokens along with the special tokens, i.e. [CLS] (first token), [SEP] (sequnce delimiter), and [MASK] (masked words).
  • Token embeddings are the vocabulary IDs for each of the tokens.
  • Sequence Embeddings are just numerics to distinguish between sentences.
  • Transformer Positional Embeddings specify the position of each word in the sequence.

Now, we use the BERT module to compute a representation (embedding) for user requests and command descriptions in our online spreadsheet application. Very nice implementations of pre-trained BERT contextualized word embeddings exist on Google Colab for both TensorFlow platform (here) and Keras platform (here). We tested our example over these codes and we got the similar results as the USE.


[1] Mikolov, T., I. Sutskever, K. Chen, G. Corrado and J. Dean, Distributed Representations of Words and Phrases and their Compositionality (2013)

[2] Mikolov, T., Chen, K., Corrado, G., Dean, J., Sutskever, L. and Zweig, G., Efficient Estimation of Word Representations in Vector Space (2013)

[3] Pennington, J., Socher, R. and Manning, C., Glove: Global vectors for word representation (2014)

[4] Dai, A.M. and Le, Q.V., Semi-supervised sequence learning (2015)

[5] Radford, A., Narasimhan, K., Salimans, T. and Sutskever, I., Improving language understanding by generative pre-training (2018)

[6] Peters, M.E., Neumann, M., Iyyer, M., Gardner, M., Clark, C., Lee, K. and Zettlemoyer, L., Deep contextualized word representations (2018)

[7] Howard, J. and Ruder, S., Universal language model fine-tuning for text classification (2018)

[8] Cer, D., Yang, Y., Kong, S.Y., Hua, N., Limtiaco, N., John, R.S., Constant, N., Guajardo-Cespedes, M., Yuan, S., Tar, C. and Sung, Y.H., Universal sentence encoder (2018)

[9] Devlin, J., Chang, M.W., Lee, K. and Toutanova, K., Bert: Pre-training of deep bidirectional transformers for language understanding (2018)

[10] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A.Gomez, L. Kaiser and I. Polosukhin, Attention Is All You Need (2017)